diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.h b/llvm/lib/Target/AMDGPU/AMDGPU.h
--- a/llvm/lib/Target/AMDGPU/AMDGPU.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.h
@@ -317,6 +317,9 @@
 void initializeSIPostRABundlerPass(PassRegistry&);
 extern char &SIPostRABundlerID;
 
+void initializeGCNCreateVOPDPass(PassRegistry &);
+extern char &GCNCreateVOPDID;
+
 void initializeAMDGPUUnifyDivergentExitNodesPass(PassRegistry&);
 extern char &AMDGPUUnifyDivergentExitNodesID;
 
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
--- a/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -22,11 +22,13 @@
 #include "AMDGPUTargetTransformInfo.h"
 #include "GCNIterativeScheduler.h"
 #include "GCNSchedStrategy.h"
+#include "GCNVOPDUtils.h"
 #include "R600.h"
 #include "R600TargetMachine.h"
 #include "SIMachineFunctionInfo.h"
 #include "SIMachineScheduler.h"
 #include "TargetInfo/AMDGPUTargetInfo.h"
+#include "Utils/AMDGPUBaseInfo.h"
 #include "llvm/Analysis/CGSCCPassManager.h"
 #include "llvm/CodeGen/GlobalISel/CSEInfo.h"
 #include "llvm/CodeGen/GlobalISel/IRTranslator.h"
@@ -278,6 +280,12 @@
                          cl::desc("Enable s_delay_alu insertion"),
                          cl::init(true), cl::Hidden);
 
+// Enable GFX11+ VOPD
+static cl::opt<bool>
+    EnableVOPD("amdgpu-enable-vopd",
+               cl::desc("Enable VOPD, dual issue of VALU in wave32"),
+               cl::init(true), cl::Hidden);
+
 // Option is used in lit tests to prevent deadcoding of patterns inspected.
 static cl::opt<bool>
 EnableDCEInRA("amdgpu-dce-in-ra",
@@ -383,6 +391,7 @@
   initializeSIPreAllocateWWMRegsPass(*PR);
   initializeSIFormMemoryClausesPass(*PR);
   initializeSIPostRABundlerPass(*PR);
+  initializeGCNCreateVOPDPass(*PR);
   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
   initializeAMDGPUAAWrapperPassPass(*PR);
   initializeAMDGPUExternalAAWrapperPass(*PR);
@@ -909,6 +918,8 @@
     DAG->addMutation(ST.createFillMFMAShadowMutation(DAG->TII));
     DAG->addMutation(createIGroupLPDAGMutation());
     DAG->addMutation(createSchedBarrierDAGMutation());
+    if (isPassEnabled(EnableVOPD, CodeGenOpt::Less))
+      DAG->addMutation(createVOPDPairingMutation());
     return DAG;
   }
 
@@ -1385,6 +1396,8 @@
 }
 
 void GCNPassConfig::addPreEmitPass() {
+  if (isPassEnabled(EnableVOPD, CodeGenOpt::Less))
+    addPass(&GCNCreateVOPDID);
   addPass(createSIMemoryLegalizerPass());
   addPass(createSIInsertWaitcntsPass());
 
diff --git a/llvm/lib/Target/AMDGPU/CMakeLists.txt b/llvm/lib/Target/AMDGPU/CMakeLists.txt
--- a/llvm/lib/Target/AMDGPU/CMakeLists.txt
+++ b/llvm/lib/Target/AMDGPU/CMakeLists.txt
@@ -100,6 +100,7 @@
   AMDGPUUnifyDivergentExitNodes.cpp
   AMDGPUUnifyMetadata.cpp
   R600MachineCFGStructurizer.cpp
+  GCNCreateVOPD.cpp
   GCNDPPCombine.cpp
   GCNHazardRecognizer.cpp
   GCNILPSched.cpp
@@ -109,6 +110,7 @@
   GCNPreRAOptimizations.cpp
   GCNRegPressure.cpp
   GCNSchedStrategy.cpp
+  GCNVOPDUtils.cpp
   R600AsmPrinter.cpp
   R600ClauseMergePass.cpp
   R600ControlFlowFinalizer.cpp
diff --git a/llvm/lib/Target/AMDGPU/GCNCreateVOPD.cpp b/llvm/lib/Target/AMDGPU/GCNCreateVOPD.cpp
new file mode 100644
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/GCNCreateVOPD.cpp
@@ -0,0 +1,182 @@
+//===- GCNCreateVOPD.cpp - Create VOPD Instructions ----------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// Combine VALU pairs into VOPD instructions
+/// Only works on wave32
+/// Has register requirements, we reject creating VOPD if the requirements are
+/// not met.
+/// shouldCombineVOPD mutator in postRA machine scheduler puts candidate
+/// instructions for VOPD back-to-back
+///
+//
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "GCNSubtarget.h"
+#include "GCNVOPDUtils.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "SIInstrInfo.h"
+#include "Utils/AMDGPUBaseInfo.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/StringMap.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineOperand.h"
+#include "llvm/Support/Casting.h"
+#include "llvm/Support/Debug.h"
+#include <utility>
+
+#define DEBUG_TYPE "gcn-create-vopd"
+STATISTIC(NumVOPDCreated, "Number of VOPD Insts Created.");
+
+using namespace llvm;
+
+namespace {
+
+class GCNCreateVOPD : public MachineFunctionPass {
+private:
+public:
+  static char ID;
+  const GCNSubtarget *ST = nullptr;
+
+  GCNCreateVOPD() : MachineFunctionPass(ID) {}
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesCFG();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+  StringRef getPassName() const override {
+    return "GCN Create VOPD Instructions";
+  }
+
+  bool doReplace(const SIInstrInfo *SII,
+                 std::pair<MachineInstr *, MachineInstr *> &Pair) {
+    auto *FirstMI = Pair.first;
+    auto *SecondMI = Pair.second;
+    unsigned Opc1 = FirstMI->getOpcode();
+    unsigned Opc2 = SecondMI->getOpcode();
+    int NewOpcode = AMDGPU::getVOPDFull(AMDGPU::getVOPDOpcode(Opc1),
+                                        AMDGPU::getVOPDOpcode(Opc2));
+    assert(NewOpcode != -1 &&
+           "Should have previously determined this as a possible VOPD\n");
+
+    auto VOPDInst = BuildMI(*FirstMI->getParent(), FirstMI,
+                            FirstMI->getDebugLoc(), SII->get(NewOpcode))
+                        .setMIFlags(FirstMI->getFlags() | SecondMI->getFlags());
+    VOPDInst.add(FirstMI->getOperand(0))
+        .add(SecondMI->getOperand(0))
+        .add(FirstMI->getOperand(1));
+
+    switch (Opc1) {
+    case AMDGPU::V_MOV_B32_e32:
+      break;
+    case AMDGPU::V_FMAMK_F32:
+      VOPDInst.add(FirstMI->getOperand(2));
+      VOPDInst.add(FirstMI->getOperand(3));
+      break;
+    case AMDGPU::V_FMAAK_F32:
+      VOPDInst.add(FirstMI->getOperand(2));
+      VOPDInst.add(FirstMI->getOperand(3));
+      break;
+    default:
+      VOPDInst.add(FirstMI->getOperand(2));
+      break;
+    }
+
+    VOPDInst.add(SecondMI->getOperand(1));
+
+    switch (Opc2) {
+    case AMDGPU::V_MOV_B32_e32:
+      break;
+    case AMDGPU::V_FMAMK_F32:
+      VOPDInst.add(SecondMI->getOperand(2));
+      VOPDInst.add(SecondMI->getOperand(3));
+      break;
+    case AMDGPU::V_FMAAK_F32:
+      VOPDInst.add(SecondMI->getOperand(2));
+      VOPDInst.add(SecondMI->getOperand(3));
+      break;
+    default:
+      VOPDInst.add(SecondMI->getOperand(2));
+      break;
+    }
+
+    VOPDInst.copyImplicitOps(*FirstMI);
+    VOPDInst.copyImplicitOps(*SecondMI);
+
+    LLVM_DEBUG(dbgs() << "VOPD Fused: "; VOPDInst->dump();
+               dbgs() << " from\tX: "; Pair.first->dump(); dbgs() << "\tY: ";
+               Pair.second->dump(); dbgs() << "\n");
+    FirstMI->eraseFromParent();
+    SecondMI->eraseFromParent();
+    ++NumVOPDCreated;
+    return true;
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override {
+    if (skipFunction(MF.getFunction()))
+      return false;
+    ST = &MF.getSubtarget<GCNSubtarget>();
+    if (!AMDGPU::hasVOPD(*ST) || !ST->isWave32())
+      return false;
+    LLVM_DEBUG(dbgs() << "CreateVOPD Pass:\n");
+
+    const SIInstrInfo *SII = ST->getInstrInfo();
+    bool Changed = false;
+
+    SmallVector<std::pair<MachineInstr *, MachineInstr *>> ReplaceCandidates;
+
+    for (auto &MBB : MF) {
+      auto MII = MBB.begin(), E = MBB.end();
+      while (MII != E) {
+        auto *FirstMI = &*MII;
+        MII = next_nodbg(MII, MBB.end());
+        if (MII == MBB.end())
+          break;
+        if (FirstMI->isDebugInstr())
+          continue;
+        auto *SecondMI = &*MII;
+        unsigned Opc = FirstMI->getOpcode();
+        unsigned Opc2 = SecondMI->getOpcode();
+        auto FirstCanBeVOPD = AMDGPU::getCanBeVOPD(Opc);
+        auto SecondCanBeVOPD = AMDGPU::getCanBeVOPD(Opc2);
+        std::pair<MachineInstr *, MachineInstr *> Pair;
+
+        if (FirstCanBeVOPD.X && SecondCanBeVOPD.Y)
+          Pair = {FirstMI, SecondMI};
+        else if (FirstCanBeVOPD.Y && SecondCanBeVOPD.X)
+          Pair = {SecondMI, FirstMI};
+        else
+          continue;
+        // checkVOPDRegConstraints cares about program order, but doReplace
+        // cares about X-Y order in the constituted VOPD
+        if (llvm::checkVOPDRegConstraints(*SII, *FirstMI, *SecondMI)) {
+          ReplaceCandidates.push_back(Pair);
+          ++MII;
+        }
+      }
+    }
+    for (auto &Pair : ReplaceCandidates) {
+      Changed |= doReplace(SII, Pair);
+    }
+
+    return Changed;
+  }
+};
+
+} // namespace
+
+char GCNCreateVOPD::ID = 0;
+
+char &llvm::GCNCreateVOPDID = GCNCreateVOPD::ID;
+
+INITIALIZE_PASS(GCNCreateVOPD, DEBUG_TYPE, "GCN Create VOPD Instructions",
+                false, false)
diff --git a/llvm/lib/Target/AMDGPU/GCNVOPDUtils.h b/llvm/lib/Target/AMDGPU/GCNVOPDUtils.h
new file mode 100644
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/GCNVOPDUtils.h
@@ -0,0 +1,32 @@
+//===- GCNVOPDUtils.h - GCN VOPD Utils  ------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This file contains the AMDGPU DAG scheduling
+/// mutation to pair VOPD instructions back to back. It also contains
+//  subroutines useful in the creation of VOPD instructions
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_LIB_TARGET_AMDGPU_VOPDUTILS_H
+#define LLVM_LIB_TARGET_AMDGPU_VOPDUTILS_H
+
+#include "llvm/CodeGen/MachineScheduler.h"
+
+namespace llvm {
+
+class SIInstrInfo;
+
+bool checkVOPDRegConstraints(const SIInstrInfo &TII,
+                             const MachineInstr &FirstMI,
+                             const MachineInstr &SecondMI);
+
+std::unique_ptr<ScheduleDAGMutation> createVOPDPairingMutation();
+
+} // namespace llvm
+
+#endif // LLVM_LIB_TARGET_AMDGPU_VOPDUTILS_H
diff --git a/llvm/lib/Target/AMDGPU/GCNVOPDUtils.cpp b/llvm/lib/Target/AMDGPU/GCNVOPDUtils.cpp
new file mode 100644
--- /dev/null
+++ b/llvm/lib/Target/AMDGPU/GCNVOPDUtils.cpp
@@ -0,0 +1,212 @@
+//===- GCNVOPDUtils.cpp - GCN VOPD Utils  ------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file This file contains the AMDGPU DAG scheduling
+/// mutation to pair VOPD instructions back to back. It also contains
+//  subroutines useful in the creation of VOPD instructions
+//
+//===----------------------------------------------------------------------===//
+
+#include "GCNVOPDUtils.h"
+#include "AMDGPUSubtarget.h"
+#include "GCNSubtarget.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "SIInstrInfo.h"
+#include "llvm/ADT/STLExtras.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineOperand.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/MacroFusion.h"
+#include "llvm/CodeGen/ScheduleDAG.h"
+#include "llvm/CodeGen/ScheduleDAGMutation.h"
+#include "llvm/CodeGen/TargetInstrInfo.h"
+#include "llvm/MC/MCInst.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "gcn-vopd-utils"
+
+bool llvm::checkVOPDRegConstraints(const SIInstrInfo &TII,
+                                   const MachineInstr &FirstMI,
+                                   const MachineInstr &SecondMI) {
+  const MachineFunction *MF = FirstMI.getMF();
+  const GCNSubtarget &ST = MF->getSubtarget<GCNSubtarget>();
+  const SIRegisterInfo *TRI = dyn_cast<SIRegisterInfo>(ST.getRegisterInfo());
+  const MachineRegisterInfo &MRI = MF->getRegInfo();
+  const unsigned NumVGPRBanks = 4;
+  // Literals also count against scalar bus limit
+  SmallVector<const MachineOperand *> UniqueLiterals;
+  auto addLiteral = [&](const MachineOperand &Op) {
+    for (auto &Literal : UniqueLiterals) {
+      if (Literal->isIdenticalTo(Op))
+        return;
+    }
+    UniqueLiterals.push_back(&Op);
+  };
+  SmallVector<Register> UniqueScalarRegs;
+  assert([&]() -> bool {
+    for (auto MII = MachineBasicBlock::const_iterator(&FirstMI);
+         MII != FirstMI.getParent()->instr_end(); ++MII) {
+      if (&*MII == &SecondMI)
+        return true;
+    }
+    return false;
+  }() && "Expected FirstMI to precede SecondMI");
+  // Cannot pair dependent instructions
+  for (const auto &Use : SecondMI.uses())
+    if (Use.isReg() && FirstMI.modifiesRegister(Use.getReg()))
+      return false;
+
+  struct ComponentInfo {
+    ComponentInfo(const MachineInstr &MI) : MI(MI) {}
+    Register Dst, Reg0, Reg1, Reg2;
+    const MachineInstr &MI;
+  };
+  ComponentInfo CInfo[] = {ComponentInfo(FirstMI), ComponentInfo(SecondMI)};
+
+  for (ComponentInfo &Comp : CInfo) {
+    switch (Comp.MI.getOpcode()) {
+    case AMDGPU::V_FMAMK_F32:
+      // cannot inline the fixed literal in fmamk
+      addLiteral(Comp.MI.getOperand(2));
+      Comp.Reg2 = Comp.MI.getOperand(3).getReg();
+      break;
+    case AMDGPU::V_FMAAK_F32:
+      // cannot inline the fixed literal in fmaak
+      addLiteral(Comp.MI.getOperand(3));
+      Comp.Reg1 = Comp.MI.getOperand(2).getReg();
+      break;
+    case AMDGPU::V_FMAC_F32_e32:
+    case AMDGPU::V_DOT2_F32_F16:
+    case AMDGPU::V_DOT2_F32_BF16:
+      Comp.Reg1 = Comp.MI.getOperand(2).getReg();
+      Comp.Reg2 = Comp.MI.getOperand(0).getReg();
+      break;
+    case AMDGPU::V_CNDMASK_B32_e32:
+      UniqueScalarRegs.push_back(AMDGPU::VCC_LO);
+      Comp.Reg1 = Comp.MI.getOperand(2).getReg();
+      break;
+    case AMDGPU::V_MOV_B32_e32:
+      break;
+    default:
+      Comp.Reg1 = Comp.MI.getOperand(2).getReg();
+      break;
+    }
+
+    Comp.Dst = Comp.MI.getOperand(0).getReg();
+
+    const MachineOperand &Op0 = Comp.MI.getOperand(1);
+    if (Op0.isReg()) {
+      if (!TRI->isVectorRegister(MRI, Op0.getReg())) {
+        if (!is_contained(UniqueScalarRegs, Op0.getReg()))
+          UniqueScalarRegs.push_back(Op0.getReg());
+      } else
+        Comp.Reg0 = Op0.getReg();
+    } else {
+      if (!TII.isInlineConstant(Comp.MI, 1))
+        addLiteral(Op0);
+    }
+  }
+
+  if (UniqueLiterals.size() > 1)
+    return false;
+  if ((UniqueLiterals.size() + UniqueScalarRegs.size()) > 2)
+    return false;
+
+  // check port 0
+  if (CInfo[0].Reg0 && CInfo[1].Reg0 &&
+      CInfo[0].Reg0 % NumVGPRBanks == CInfo[1].Reg0 % NumVGPRBanks)
+    return false;
+  // check port 1
+  if (CInfo[0].Reg1 && CInfo[1].Reg1 &&
+      CInfo[0].Reg1 % NumVGPRBanks == CInfo[1].Reg1 % NumVGPRBanks)
+    return false;
+  // check port 2
+  if (CInfo[0].Reg2 && CInfo[1].Reg2 &&
+      !((CInfo[0].Reg2 ^ CInfo[1].Reg2) & 0x1))
+    return false;
+  if (!((CInfo[0].Dst ^ CInfo[1].Dst) & 0x1))
+    return false;
+
+  LLVM_DEBUG(dbgs() << "VOPD Reg Constraints Passed\n\tX: "; FirstMI.dump();
+             dbgs() << "\n\tY: "; SecondMI.dump(); dbgs() << "\n");
+  return true;
+}
+
+/// Check if the instr pair, FirstMI and SecondMI, should be scheduled
+/// together. Given SecondMI, when FirstMI is unspecified, then check if
+/// SecondMI may be part of a fused pair at all.
+static bool shouldScheduleVOPDAdjacent(const TargetInstrInfo &TII,
+                                       const TargetSubtargetInfo &TSI,
+                                       const MachineInstr *FirstMI,
+                                       const MachineInstr &SecondMI) {
+  const SIInstrInfo &STII = static_cast<const SIInstrInfo &>(TII);
+  unsigned Opc2 = SecondMI.getOpcode();
+  auto SecondCanBeVOPD = AMDGPU::getCanBeVOPD(Opc2);
+
+  // One instruction case
+  if (!FirstMI)
+    return SecondCanBeVOPD.Y;
+
+  unsigned Opc = FirstMI->getOpcode();
+  auto FirstCanBeVOPD = AMDGPU::getCanBeVOPD(Opc);
+
+  if (!((FirstCanBeVOPD.X && SecondCanBeVOPD.Y) ||
+        (FirstCanBeVOPD.Y && SecondCanBeVOPD.X)))
+    return false;
+
+  return checkVOPDRegConstraints(STII, *FirstMI, SecondMI);
+}
+
+/// Adapts design from MacroFusion
+/// Puts valid candidate instructions back-to-back so they can easily
+/// be turned into VOPD instructions
+/// Greedily pairs instruction candidates. O(n^2) algorithm.
+struct VOPDPairingMutation : ScheduleDAGMutation {
+  ShouldSchedulePredTy shouldScheduleAdjacent; // NOLINT: function pointer
+
+  VOPDPairingMutation(
+      ShouldSchedulePredTy shouldScheduleAdjacent) // NOLINT: function pointer
+      : shouldScheduleAdjacent(shouldScheduleAdjacent) {}
+
+  void apply(ScheduleDAGInstrs *DAG) override {
+    const TargetInstrInfo &TII = *DAG->TII;
+    const GCNSubtarget &ST = DAG->MF.getSubtarget<GCNSubtarget>();
+    if (!AMDGPU::hasVOPD(ST) || !ST.isWave32()) {
+      LLVM_DEBUG(dbgs() << "Target does not support VOPDPairingMutation\n");
+      return;
+    }
+
+    std::vector<SUnit>::iterator ISUI, JSUI;
+    for (ISUI = DAG->SUnits.begin(); ISUI != DAG->SUnits.end(); ++ISUI) {
+      const MachineInstr *IMI = ISUI->getInstr();
+      if (!shouldScheduleAdjacent(TII, ST, nullptr, *IMI))
+        continue;
+      if (!hasLessThanNumFused(*ISUI, 2))
+        continue;
+
+      for (JSUI = ISUI + 1; JSUI != DAG->SUnits.end(); ++JSUI) {
+        if (JSUI->isBoundaryNode())
+          continue;
+        const MachineInstr *JMI = JSUI->getInstr();
+        if (!hasLessThanNumFused(*JSUI, 2) ||
+            !shouldScheduleAdjacent(TII, ST, IMI, *JMI))
+          continue;
+        if (fuseInstructionPair(*DAG, *ISUI, *JSUI))
+          break;
+      }
+    }
+    LLVM_DEBUG(dbgs() << "Completed VOPDPairingMutation\n");
+  }
+};
+
+std::unique_ptr<ScheduleDAGMutation> llvm::createVOPDPairingMutation() {
+  return std::make_unique<VOPDPairingMutation>(shouldScheduleVOPDAdjacent);
+}
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -3974,6 +3974,14 @@
   int Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0);
   int Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src1);
   int Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src2);
+  int Src3Idx = -1;
+  if (Src0Idx == -1) {
+    // VOPD V_DUAL_* instructions use different operand names.
+    Src0Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0X);
+    Src1Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vsrc1X);
+    Src2Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::src0Y);
+    Src3Idx = AMDGPU::getNamedOperandIdx(Opcode, AMDGPU::OpName::vsrc1Y);
+  }
 
   // Make sure the number of operands is correct.
   const MCInstrDesc &Desc = get(Opcode);
@@ -4247,9 +4255,9 @@
     // Only look at the true operands. Only a real operand can use the constant
     // bus, and we don't want to check pseudo-operands like the source modifier
     // flags.
-    for (int OpIdx : {Src0Idx, Src1Idx, Src2Idx}) {
+    for (int OpIdx : {Src0Idx, Src1Idx, Src2Idx, Src3Idx}) {
       if (OpIdx == -1)
-        break;
+        continue;
       const MachineOperand &MO = MI.getOperand(OpIdx);
       if (usesConstantBus(MRI, MO, MI.getDesc().OpInfo[OpIdx])) {
         if (MO.isReg()) {
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -2948,6 +2948,27 @@
   let ValueCols = [["1"]];
 }
 
+def VOPDComponentTable : GenericTable {
+  let FilterClass = "VOPD_Component";
+  let CppTypeName = "VOPDComponentInfo";
+  let Fields = ["BaseVOP", "VOPDOp", "CanBeVOPDX"];
+  let PrimaryKey = ["BaseVOP"];
+  let PrimaryKeyName = "getVOPDComponentHelper";
+}
+
+def VOPDPairs : GenericTable {
+  let FilterClass = "VOPD_Base";
+  let CppTypeName = "VOPDInfo";
+  let Fields = ["Opcode", "OpX", "OpY"];
+  let PrimaryKey = ["Opcode"];
+  let PrimaryKeyName = "getVOPDOpcodeHelper";
+}
+
+def getVOPDInfoFromComponentOpcodes : SearchIndex {
+  let Table = VOPDPairs;
+  let Key = ["OpX", "OpY"];
+}
+
 include "SIInstructions.td"
 
 include "DSInstructions.td"
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -465,6 +465,14 @@
 LLVM_READONLY
 bool getMAIIsGFX940XDL(unsigned Opc);
 
+struct CanBeVOPD {
+  bool X;
+  bool Y;
+};
+
+LLVM_READONLY
+CanBeVOPD getCanBeVOPD(unsigned Opc);
+
 LLVM_READONLY
 const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t BitsPerComp,
                                                   uint8_t NumComponents,
@@ -477,6 +485,12 @@
 LLVM_READONLY
 int getMCOpcode(uint16_t Opcode, unsigned Gen);
 
+LLVM_READONLY
+unsigned getVOPDOpcode(unsigned Opc);
+
+LLVM_READONLY
+int getVOPDFull(unsigned OpX, unsigned OpY);
+
 void initDefaultAMDKernelCodeT(amd_kernel_code_t &Header,
                                const MCSubtargetInfo *STI);
 
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -277,6 +277,18 @@
   uint16_t Opcode;
 };
 
+struct VOPDComponentInfo {
+  uint16_t BaseVOP;
+  uint16_t VOPDOp;
+  bool CanBeVOPDX;
+};
+
+struct VOPDInfo {
+  uint16_t Opcode;
+  uint16_t OpX;
+  uint16_t OpY;
+};
+
 #define GET_MTBUFInfoTable_DECL
 #define GET_MTBUFInfoTable_IMPL
 #define GET_MUBUFInfoTable_DECL
@@ -293,6 +305,10 @@
 #define GET_VOPC64DPPTable_IMPL
 #define GET_VOPC64DPP8Table_DECL
 #define GET_VOPC64DPP8Table_IMPL
+#define GET_VOPDComponentTable_DECL
+#define GET_VOPDComponentTable_IMPL
+#define GET_VOPDPairs_DECL
+#define GET_VOPDPairs_IMPL
 #include "AMDGPUGenSearchableTables.inc"
 
 int getMTBUFBaseOpcode(unsigned Opc) {
@@ -394,6 +410,19 @@
   return Info ? Info->is_gfx940_xdl : false;
 }
 
+CanBeVOPD getCanBeVOPD(unsigned Opc) {
+  const VOPDComponentInfo *Info = getVOPDComponentHelper(Opc);
+  if (Info)
+    return {Info->CanBeVOPDX, 1};
+  else
+    return {0, 0};
+}
+
+unsigned getVOPDOpcode(unsigned Opc) {
+  const VOPDComponentInfo *Info = getVOPDComponentHelper(Opc);
+  return Info ? Info->VOPDOp : ~0u;
+}
+
 // Wrapper for Tablegen'd function.  enum Subtarget is not defined in any
 // header files, so we need to wrap it in a function that takes unsigned
 // instead.
@@ -401,6 +430,11 @@
   return getMCOpcodeGen(Opcode, static_cast<Subtarget>(Gen));
 }
 
+int getVOPDFull(unsigned OpX, unsigned OpY) {
+  const VOPDInfo *Info = getVOPDInfoFromComponentOpcodes(OpX, OpY);
+  return Info ? Info->Opcode : -1;
+}
+
 namespace IsaInfo {
 
 AMDGPUTargetID::AMDGPUTargetID(const MCSubtargetInfo &STI)
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.gather4.dim.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.gather4.dim.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.gather4.dim.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.gather4.dim.ll
@@ -1,7 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -o - %s | FileCheck -check-prefix=GFX6 %s
 ; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -o - %s | FileCheck -check-prefix=GFX10NSA %s
-; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -o - %s | FileCheck -check-prefix=GFX10NSA %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -amdgpu-enable-vopd=0 -o - %s | FileCheck -check-prefix=GFX10NSA %s
 
 define amdgpu_ps <4 x float> @gather4_2d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %s, float %t) {
 ; GFX6-LABEL: gather4_2d:
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2d.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2d.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2d.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2d.ll
@@ -94,14 +94,14 @@
 ; GFX11-LABEL: load_2d_v4f32_xyzw_tfe:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    v_mov_b32_e32 v7, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    s_mov_b32 s0, s2
 ; GFX11-NEXT:    s_mov_b32 s1, s3
-; GFX11-NEXT:    v_mov_b32_e32 v8, v7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v7
-; GFX11-NEXT:    v_mov_b32_e32 v10, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v7
+; GFX11-NEXT:    v_mov_b32_e32 v10, v7
+; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    s_mov_b32 s2, s4
 ; GFX11-NEXT:    s_mov_b32 s3, s5
 ; GFX11-NEXT:    s_mov_b32 s4, s6
@@ -109,10 +109,8 @@
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v7
-; GFX11-NEXT:    v_mov_b32_e32 v1, v8
-; GFX11-NEXT:    v_mov_b32_e32 v2, v9
-; GFX11-NEXT:    v_mov_b32_e32 v3, v10
-; GFX11-NEXT:    v_mov_b32_e32 v4, v11
+; GFX11-NEXT:    v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, v8 :: v_dual_mov_b32 v4, v11
 ; GFX11-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v7, v4, s[10:11]
@@ -184,14 +182,14 @@
 ; GFX11-LABEL: load_2d_v4f32_xyzw_tfe_lwe:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    v_mov_b32_e32 v7, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    s_mov_b32 s0, s2
 ; GFX11-NEXT:    s_mov_b32 s1, s3
-; GFX11-NEXT:    v_mov_b32_e32 v8, v7
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v7
-; GFX11-NEXT:    v_mov_b32_e32 v10, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v7
+; GFX11-NEXT:    v_mov_b32_e32 v10, v7
+; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    s_mov_b32 s2, s4
 ; GFX11-NEXT:    s_mov_b32 s3, s5
 ; GFX11-NEXT:    s_mov_b32 s4, s6
@@ -199,10 +197,8 @@
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v7
-; GFX11-NEXT:    v_mov_b32_e32 v1, v8
-; GFX11-NEXT:    v_mov_b32_e32 v2, v9
-; GFX11-NEXT:    v_mov_b32_e32 v3, v10
-; GFX11-NEXT:    v_mov_b32_e32 v4, v11
+; GFX11-NEXT:    v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, v8 :: v_dual_mov_b32 v4, v11
 ; GFX11-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v7, v4, s[10:11]
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.a16.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.a16.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.a16.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.a16.ll
@@ -1,7 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 %s
 ; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10PLUS,GFX10 %s
-; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11 %s
+; RUN: llc -global-isel -mtriple=amdgcn-mesa-mesa3d -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10PLUS,GFX11 %s
 
 define amdgpu_ps <4 x float> @load_2darraymsaa_v4f32_xyzw(<8 x i32> inreg %rsrc, i16 %s, i16 %t, i16 %slice, i16 %fragid) {
 ; GFX9-LABEL: load_2darraymsaa_v4f32_xyzw:
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.2darraymsaa.ll
@@ -98,14 +98,13 @@
 ; GFX11-LABEL: load_2darraymsaa_v4f32_xyzw_tfe:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    v_mov_b32_e32 v9, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v8, v3
-; GFX11-NEXT:    v_mov_b32_e32 v10, v9
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v8, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v9
-; GFX11-NEXT:    v_mov_b32_e32 v12, v9
 ; GFX11-NEXT:    v_mov_b32_e32 v13, v9
+; GFX11-NEXT:    v_mov_b32_e32 v12, v9
+; GFX11-NEXT:    v_mov_b32_e32 v10, v9
 ; GFX11-NEXT:    s_mov_b32 s0, s2
 ; GFX11-NEXT:    s_mov_b32 s1, s3
 ; GFX11-NEXT:    s_mov_b32 s2, s4
@@ -115,10 +114,8 @@
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v9
-; GFX11-NEXT:    v_mov_b32_e32 v1, v10
-; GFX11-NEXT:    v_mov_b32_e32 v2, v11
-; GFX11-NEXT:    v_mov_b32_e32 v3, v12
-; GFX11-NEXT:    v_mov_b32_e32 v4, v13
+; GFX11-NEXT:    v_dual_mov_b32 v2, v11 :: v_dual_mov_b32 v3, v12
+; GFX11-NEXT:    v_dual_mov_b32 v1, v10 :: v_dual_mov_b32 v4, v13
 ; GFX11-NEXT:    image_load v[0:4], v[5:8], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA_ARRAY unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v9, v4, s[10:11]
@@ -194,14 +191,13 @@
 ; GFX11-LABEL: load_2darraymsaa_v4f32_xyzw_tfe_lwe:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    v_mov_b32_e32 v9, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v8, v3
-; GFX11-NEXT:    v_mov_b32_e32 v10, v9
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v8, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v9
-; GFX11-NEXT:    v_mov_b32_e32 v12, v9
 ; GFX11-NEXT:    v_mov_b32_e32 v13, v9
+; GFX11-NEXT:    v_mov_b32_e32 v12, v9
+; GFX11-NEXT:    v_mov_b32_e32 v10, v9
 ; GFX11-NEXT:    s_mov_b32 s0, s2
 ; GFX11-NEXT:    s_mov_b32 s1, s3
 ; GFX11-NEXT:    s_mov_b32 s2, s4
@@ -211,10 +207,8 @@
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v9
-; GFX11-NEXT:    v_mov_b32_e32 v1, v10
-; GFX11-NEXT:    v_mov_b32_e32 v2, v11
-; GFX11-NEXT:    v_mov_b32_e32 v3, v12
-; GFX11-NEXT:    v_mov_b32_e32 v4, v13
+; GFX11-NEXT:    v_dual_mov_b32 v2, v11 :: v_dual_mov_b32 v3, v12
+; GFX11-NEXT:    v_dual_mov_b32 v1, v10 :: v_dual_mov_b32 v4, v13
 ; GFX11-NEXT:    image_load v[0:4], v[5:8], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA_ARRAY unorm tfe lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v9, v4, s[10:11]
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.a16.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.a16.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.a16.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.a16.ll
@@ -113,12 +113,12 @@
 ; GFX11-NEXT:    s_mov_b32 s0, s2
 ; GFX11-NEXT:    s_mov_b32 s2, s4
 ; GFX11-NEXT:    s_mov_b32 s4, s6
+; GFX11-NEXT:    v_mov_b32_e32 v6, v5
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_lshl_b32 s8, s0, 16
-; GFX11-NEXT:    v_mov_b32_e32 v6, v5
 ; GFX11-NEXT:    v_mov_b32_e32 v7, v5
-; GFX11-NEXT:    v_mov_b32_e32 v8, v5
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v5
+; GFX11-NEXT:    v_mov_b32_e32 v8, v5
 ; GFX11-NEXT:    v_and_or_b32 v10, 0xffff, v0, v1
 ; GFX11-NEXT:    v_and_or_b32 v11, 0xffff, v2, s8
 ; GFX11-NEXT:    s_mov_b32 s1, s3
@@ -126,10 +126,8 @@
 ; GFX11-NEXT:    s_mov_b32 s5, s7
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v5
-; GFX11-NEXT:    v_mov_b32_e32 v1, v6
-; GFX11-NEXT:    v_mov_b32_e32 v2, v7
-; GFX11-NEXT:    v_mov_b32_e32 v3, v8
-; GFX11-NEXT:    v_mov_b32_e32 v4, v9
+; GFX11-NEXT:    v_dual_mov_b32 v2, v7 :: v_dual_mov_b32 v1, v6
+; GFX11-NEXT:    v_dual_mov_b32 v3, v8 :: v_dual_mov_b32 v4, v9
 ; GFX11-NEXT:    image_load v[0:4], v[10:11], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm a16 tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v5, v4, s[10:11]
@@ -211,12 +209,12 @@
 ; GFX11-NEXT:    s_mov_b32 s0, s2
 ; GFX11-NEXT:    s_mov_b32 s2, s4
 ; GFX11-NEXT:    s_mov_b32 s4, s6
+; GFX11-NEXT:    v_mov_b32_e32 v6, v5
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_lshl_b32 s8, s0, 16
-; GFX11-NEXT:    v_mov_b32_e32 v6, v5
 ; GFX11-NEXT:    v_mov_b32_e32 v7, v5
-; GFX11-NEXT:    v_mov_b32_e32 v8, v5
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v5
+; GFX11-NEXT:    v_mov_b32_e32 v8, v5
 ; GFX11-NEXT:    v_and_or_b32 v10, 0xffff, v0, v1
 ; GFX11-NEXT:    v_and_or_b32 v11, 0xffff, v2, s8
 ; GFX11-NEXT:    s_mov_b32 s1, s3
@@ -224,10 +222,8 @@
 ; GFX11-NEXT:    s_mov_b32 s5, s7
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v5
-; GFX11-NEXT:    v_mov_b32_e32 v1, v6
-; GFX11-NEXT:    v_mov_b32_e32 v2, v7
-; GFX11-NEXT:    v_mov_b32_e32 v3, v8
-; GFX11-NEXT:    v_mov_b32_e32 v4, v9
+; GFX11-NEXT:    v_dual_mov_b32 v2, v7 :: v_dual_mov_b32 v1, v6
+; GFX11-NEXT:    v_dual_mov_b32 v3, v8 :: v_dual_mov_b32 v4, v9
 ; GFX11-NEXT:    image_load v[0:4], v[10:11], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm a16 tfe lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v5, v4, s[10:11]
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.load.3d.ll
@@ -95,16 +95,15 @@
 ;
 ; GFX11-LABEL: load_3d_v4f32_xyzw_tfe:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v8, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v8, 0
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    s_mov_b32 s0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v9, v8
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    s_mov_b32 s1, s3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v8
-; GFX11-NEXT:    s_mov_b32 s1, s3
+; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_mov_b32_e32 v9, v8
 ; GFX11-NEXT:    s_mov_b32 s2, s4
 ; GFX11-NEXT:    s_mov_b32 s3, s5
 ; GFX11-NEXT:    s_mov_b32 s4, s6
@@ -112,10 +111,8 @@
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-NEXT:    v_mov_b32_e32 v1, v9
-; GFX11-NEXT:    v_mov_b32_e32 v2, v10
-; GFX11-NEXT:    v_mov_b32_e32 v3, v11
-; GFX11-NEXT:    v_mov_b32_e32 v4, v12
+; GFX11-NEXT:    v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v1, v9
+; GFX11-NEXT:    v_dual_mov_b32 v3, v11 :: v_dual_mov_b32 v4, v12
 ; GFX11-NEXT:    image_load v[0:4], v[5:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v8, v4, s[10:11]
@@ -188,16 +185,15 @@
 ;
 ; GFX11-LABEL: load_3d_v4f32_xyzw_tfe_lwe:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v8, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v8, 0
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    s_mov_b32 s0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v9, v8
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    s_mov_b32 s1, s3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v8
-; GFX11-NEXT:    s_mov_b32 s1, s3
+; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_mov_b32_e32 v9, v8
 ; GFX11-NEXT:    s_mov_b32 s2, s4
 ; GFX11-NEXT:    s_mov_b32 s3, s5
 ; GFX11-NEXT:    s_mov_b32 s4, s6
@@ -205,10 +201,8 @@
 ; GFX11-NEXT:    s_mov_b32 s6, s8
 ; GFX11-NEXT:    s_mov_b32 s7, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-NEXT:    v_mov_b32_e32 v1, v9
-; GFX11-NEXT:    v_mov_b32_e32 v2, v10
-; GFX11-NEXT:    v_mov_b32_e32 v3, v11
-; GFX11-NEXT:    v_mov_b32_e32 v4, v12
+; GFX11-NEXT:    v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v1, v9
+; GFX11-NEXT:    v_dual_mov_b32 v3, v11 :: v_dual_mov_b32 v4, v12
 ; GFX11-NEXT:    image_load v[0:4], v[5:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm tfe lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v8, v4, s[10:11]
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.sample.g16.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.sample.g16.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.sample.g16.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.sample.g16.ll
@@ -1,6 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
-; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
 
 define amdgpu_ps <4 x float> @sample_d_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, half %dsdh, half %dsdv, float %s) {
 ; GFX10-LABEL: sample_d_1d:
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.store.2d.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.store.2d.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.store.2d.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.image.store.2d.ll
@@ -622,8 +622,7 @@
 ;
 ; GFX11-LABEL: image_store_f32_dmask_1111:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v1, s10
-; GFX11-NEXT:    v_mov_b32_e32 v2, s11
+; GFX11-NEXT:    v_dual_mov_b32 v1, s10 :: v_dual_mov_b32 v2, s11
 ; GFX11-NEXT:    s_mov_b32 s0, s2
 ; GFX11-NEXT:    s_mov_b32 s1, s3
 ; GFX11-NEXT:    s_mov_b32 s2, s4
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.interp.inreg.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.interp.inreg.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.interp.inreg.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.interp.inreg.ll
@@ -42,9 +42,8 @@
 ; GCN-NEXT:    lds_param_load v2, attr2.x wait_vdst:15
 ; GCN-NEXT:    lds_param_load v3, attr3.x wait_vdst:15
 ; GCN-NEXT:    s_mov_b32 exec_lo, s3
-; GCN-NEXT:    v_mov_b32_e32 v4, s0
-; GCN-NEXT:    v_mov_b32_e32 v5, s1
-; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GCN-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s1
+; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
 ; GCN-NEXT:    v_interp_p10_f32 v6, v0, v4, v0 wait_exp:3
 ; GCN-NEXT:    v_interp_p10_f32 v7, v1, v4, v1 wait_exp:2
 ; GCN-NEXT:    v_interp_p10_f32 v8, v2, v4, v2 wait_exp:1
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.intersect_ray.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.intersect_ray.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.intersect_ray.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/llvm.amdgcn.intersect_ray.ll
@@ -228,10 +228,8 @@
 ;
 ; GFX11-LABEL: image_bvh_intersect_ray_vgpr_descr:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v18, v0
-; GFX11-NEXT:    v_mov_b32_e32 v19, v1
-; GFX11-NEXT:    v_mov_b32_e32 v15, v2
-; GFX11-NEXT:    v_mov_b32_e32 v16, v3
+; GFX11-NEXT:    v_dual_mov_b32 v18, v0 :: v_dual_mov_b32 v19, v1
+; GFX11-NEXT:    v_dual_mov_b32 v15, v2 :: v_dual_mov_b32 v16, v3
 ; GFX11-NEXT:    v_mov_b32_e32 v17, v4
 ; GFX11-NEXT:    s_mov_b32 s1, exec_lo
 ; GFX11-NEXT:  .LBB6_1: ; =>This Inner Loop Header: Depth=1
@@ -346,19 +344,18 @@
 ;
 ; GFX11-LABEL: image_bvh_intersect_ray_a16_vgpr_descr:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v13, v0
+; GFX11-NEXT:    v_dual_mov_b32 v13, v0 :: v_dual_mov_b32 v14, v1
 ; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v5
-; GFX11-NEXT:    v_mov_b32_e32 v14, v1
-; GFX11-NEXT:    v_mov_b32_e32 v15, v2
-; GFX11-NEXT:    v_mov_b32_e32 v16, v3
+; GFX11-NEXT:    v_dual_mov_b32 v15, v2 :: v_dual_mov_b32 v16, v3
+; GFX11-NEXT:    v_dual_mov_b32 v17, v4 :: v_dual_lshlrev_b32 v2, 16, v5
 ; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v7
-; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v5
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_3) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v6
-; GFX11-NEXT:    v_mov_b32_e32 v17, v4
 ; GFX11-NEXT:    s_mov_b32 s1, exec_lo
 ; GFX11-NEXT:    v_and_or_b32 v4, 0xffff, v7, v2
 ; GFX11-NEXT:    v_and_or_b32 v5, 0xffff, v1, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
 ; GFX11-NEXT:    v_and_or_b32 v6, 0xffff, v8, v3
 ; GFX11-NEXT:  .LBB7_1: ; =>This Inner Loop Header: Depth=1
 ; GFX11-NEXT:    v_readfirstlane_b32 s4, v9
@@ -470,12 +467,9 @@
 ;
 ; GFX11-LABEL: image_bvh64_intersect_ray_vgpr_descr:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v19, v0
-; GFX11-NEXT:    v_mov_b32_e32 v20, v1
-; GFX11-NEXT:    v_mov_b32_e32 v21, v2
-; GFX11-NEXT:    v_mov_b32_e32 v16, v3
-; GFX11-NEXT:    v_mov_b32_e32 v17, v4
-; GFX11-NEXT:    v_mov_b32_e32 v18, v5
+; GFX11-NEXT:    v_dual_mov_b32 v19, v0 :: v_dual_mov_b32 v20, v1
+; GFX11-NEXT:    v_dual_mov_b32 v21, v2 :: v_dual_mov_b32 v16, v3
+; GFX11-NEXT:    v_dual_mov_b32 v17, v4 :: v_dual_mov_b32 v18, v5
 ; GFX11-NEXT:    s_mov_b32 s1, exec_lo
 ; GFX11-NEXT:  .LBB8_1: ; =>This Inner Loop Header: Depth=1
 ; GFX11-NEXT:    v_readfirstlane_b32 s4, v12
@@ -595,18 +589,17 @@
 ;
 ; GFX11-LABEL: image_bvh64_intersect_ray_a16_vgpr_descr:
 ; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v14, v0
+; GFX11-NEXT:    v_dual_mov_b32 v14, v0 :: v_dual_mov_b32 v15, v1
 ; GFX11-NEXT:    v_lshrrev_b32_e32 v0, 16, v6
-; GFX11-NEXT:    v_mov_b32_e32 v15, v1
-; GFX11-NEXT:    v_mov_b32_e32 v16, v2
-; GFX11-NEXT:    v_mov_b32_e32 v17, v3
+; GFX11-NEXT:    v_dual_mov_b32 v16, v2 :: v_dual_mov_b32 v17, v3
 ; GFX11-NEXT:    v_lshrrev_b32_e32 v1, 16, v8
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 16, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(SKIP_2) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX11-NEXT:    v_lshlrev_b32_e32 v3, 16, v7
-; GFX11-NEXT:    v_mov_b32_e32 v18, v4
+; GFX11-NEXT:    v_dual_mov_b32 v18, v4 :: v_dual_lshlrev_b32 v3, 16, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v19, v5
 ; GFX11-NEXT:    v_and_or_b32 v4, 0xffff, v8, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_and_or_b32 v5, 0xffff, v1, v0
 ; GFX11-NEXT:    v_and_or_b32 v6, 0xffff, v9, v3
 ; GFX11-NEXT:    s_mov_b32 s1, exec_lo
@@ -710,35 +703,31 @@
 ; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x34
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 2, v0
+; GFX11-NEXT:    s_mov_b32 s8, 2.0
 ; GFX11-NEXT:    s_mov_b32 s9, 0x40400000
 ; GFX11-NEXT:    s_mov_b32 s12, 0x40c00000
-; GFX11-NEXT:    s_mov_b32 s8, 2.0
 ; GFX11-NEXT:    s_mov_b32 s11, 0x40a00000
 ; GFX11-NEXT:    s_mov_b32 s10, 4.0
 ; GFX11-NEXT:    s_mov_b32 s14, 0x41000000
 ; GFX11-NEXT:    s_mov_b32 s13, 0x40e00000
 ; GFX11-NEXT:    v_mov_b32_e32 v6, s12
-; GFX11-NEXT:    v_mov_b32_e32 v7, s13
-; GFX11-NEXT:    v_mov_b32_e32 v8, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s14 :: v_dual_mov_b32 v7, s13
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s4
-; GFX11-NEXT:    v_mov_b32_e32 v1, s5
-; GFX11-NEXT:    v_mov_b32_e32 v2, s6
-; GFX11-NEXT:    v_mov_b32_e32 v3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
 ; GFX11-NEXT:    s_mov_b32 s6, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v4
 ; GFX11-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
 ; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, v4
 ; GFX11-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v3, vcc_lo
-; GFX11-NEXT:    s_mov_b32 s7, 1.0
 ; GFX11-NEXT:    flat_load_b32 v9, v[0:1]
 ; GFX11-NEXT:    flat_load_b32 v10, v[2:3]
-; GFX11-NEXT:    v_mov_b32_e32 v0, s6
-; GFX11-NEXT:    v_mov_b32_e32 v3, s9
-; GFX11-NEXT:    v_mov_b32_e32 v1, s7
-; GFX11-NEXT:    v_mov_b32_e32 v2, s8
-; GFX11-NEXT:    v_mov_b32_e32 v4, s10
-; GFX11-NEXT:    v_mov_b32_e32 v5, s11
+; GFX11-NEXT:    s_mov_b32 s7, 1.0
+; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v3, s9
+; GFX11-NEXT:    v_dual_mov_b32 v1, s7 :: v_dual_mov_b32 v2, s8
+; GFX11-NEXT:    v_dual_mov_b32 v5, s11 :: v_dual_mov_b32 v4, s10
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    image_bvh_intersect_ray v[0:3], [v9, v10, v[0:2], v[3:5], v[6:8]], s[0:3]
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -865,29 +854,26 @@
 ; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x34
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v4, 2, v0
-; GFX11-NEXT:    s_mov_b32 s9, 0x42004600
 ; GFX11-NEXT:    s_mov_b32 s8, 2.0
+; GFX11-NEXT:    s_mov_b32 s9, 0x42004600
 ; GFX11-NEXT:    s_mov_b32 s10, 0x44004700
 ; GFX11-NEXT:    s_mov_b32 s11, 0x45004800
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s4
-; GFX11-NEXT:    v_mov_b32_e32 v1, s5
-; GFX11-NEXT:    v_mov_b32_e32 v2, s6
-; GFX11-NEXT:    v_mov_b32_e32 v3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-NEXT:    v_dual_mov_b32 v2, s6 :: v_dual_mov_b32 v3, s7
 ; GFX11-NEXT:    s_mov_b32 s6, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v4
 ; GFX11-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
 ; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v2, v4
 ; GFX11-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, 0, v3, vcc_lo
 ; GFX11-NEXT:    s_mov_b32 s7, 1.0
 ; GFX11-NEXT:    flat_load_b32 v6, v[0:1]
 ; GFX11-NEXT:    flat_load_b32 v7, v[2:3]
-; GFX11-NEXT:    v_mov_b32_e32 v0, s6
-; GFX11-NEXT:    v_mov_b32_e32 v3, s9
-; GFX11-NEXT:    v_mov_b32_e32 v1, s7
-; GFX11-NEXT:    v_mov_b32_e32 v2, s8
-; GFX11-NEXT:    v_mov_b32_e32 v4, s10
-; GFX11-NEXT:    v_mov_b32_e32 v5, s11
+; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v3, s9
+; GFX11-NEXT:    v_dual_mov_b32 v1, s7 :: v_dual_mov_b32 v2, s8
+; GFX11-NEXT:    v_dual_mov_b32 v5, s11 :: v_dual_mov_b32 v4, s10
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[0:2], v[3:5]], s[0:3] a16
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -978,34 +964,29 @@
 ; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x24
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x34
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX11-NEXT:    s_mov_b32 s8, 2.0
+; GFX11-NEXT:    s_mov_b32 s7, 1.0
 ; GFX11-NEXT:    s_mov_b32 s6, 0
 ; GFX11-NEXT:    s_mov_b32 s9, 0x40400000
 ; GFX11-NEXT:    s_mov_b32 s12, 0x40c00000
-; GFX11-NEXT:    s_mov_b32 s8, 2.0
-; GFX11-NEXT:    s_mov_b32 s7, 1.0
 ; GFX11-NEXT:    s_mov_b32 s11, 0x40a00000
 ; GFX11-NEXT:    s_mov_b32 s10, 4.0
 ; GFX11-NEXT:    s_mov_b32 s14, 0x41000000
 ; GFX11-NEXT:    s_mov_b32 s13, 0x40e00000
-; GFX11-NEXT:    v_mov_b32_e32 v3, s9
 ; GFX11-NEXT:    v_mov_b32_e32 v6, s12
-; GFX11-NEXT:    v_mov_b32_e32 v4, s10
-; GFX11-NEXT:    v_mov_b32_e32 v5, s11
-; GFX11-NEXT:    v_mov_b32_e32 v7, s13
-; GFX11-NEXT:    v_mov_b32_e32 v8, s14
+; GFX11-NEXT:    v_dual_mov_b32 v8, s14 :: v_dual_mov_b32 v3, s9
+; GFX11-NEXT:    v_dual_mov_b32 v4, s10 :: v_dual_mov_b32 v7, s13
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s4
+; GFX11-NEXT:    v_dual_mov_b32 v5, s11 :: v_dual_mov_b32 v0, s4
 ; GFX11-NEXT:    v_mov_b32_e32 v1, s5
 ; GFX11-NEXT:    s_mov_b32 s4, 0xb36211c7
 ; GFX11-NEXT:    s_movk_i32 s5, 0x102
 ; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX11-NEXT:    v_mov_b32_e32 v10, s5
+; GFX11-NEXT:    v_dual_mov_b32 v10, s5 :: v_dual_mov_b32 v9, s4
 ; GFX11-NEXT:    v_add_co_u32 v0, vcc_lo, v0, v2
 ; GFX11-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
-; GFX11-NEXT:    v_mov_b32_e32 v9, s4
 ; GFX11-NEXT:    flat_load_b32 v11, v[0:1]
-; GFX11-NEXT:    v_mov_b32_e32 v0, s6
-; GFX11-NEXT:    v_mov_b32_e32 v1, s7
+; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
 ; GFX11-NEXT:    v_mov_b32_e32 v2, s8
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    image_bvh64_intersect_ray v[0:3], [v[9:10], v11, v[0:2], v[3:5], v[6:8]], s[0:3]
@@ -1131,11 +1112,9 @@
 ; GFX11-NEXT:    s_mov_b32 s7, 1.0
 ; GFX11-NEXT:    s_mov_b32 s10, 0x44004700
 ; GFX11-NEXT:    s_mov_b32 s11, 0x45004800
-; GFX11-NEXT:    v_mov_b32_e32 v3, s9
-; GFX11-NEXT:    v_mov_b32_e32 v4, s10
-; GFX11-NEXT:    v_mov_b32_e32 v5, s11
+; GFX11-NEXT:    v_dual_mov_b32 v3, s9 :: v_dual_mov_b32 v4, s10
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s4
+; GFX11-NEXT:    v_dual_mov_b32 v5, s11 :: v_dual_mov_b32 v0, s4
 ; GFX11-NEXT:    v_mov_b32_e32 v1, s5
 ; GFX11-NEXT:    s_mov_b32 s4, 0xb36211c6
 ; GFX11-NEXT:    s_movk_i32 s5, 0x102
@@ -1145,8 +1124,7 @@
 ; GFX11-NEXT:    v_add_co_ci_u32_e32 v1, vcc_lo, 0, v1, vcc_lo
 ; GFX11-NEXT:    v_mov_b32_e32 v6, s4
 ; GFX11-NEXT:    flat_load_b32 v8, v[0:1]
-; GFX11-NEXT:    v_mov_b32_e32 v0, s6
-; GFX11-NEXT:    v_mov_b32_e32 v1, s7
+; GFX11-NEXT:    v_dual_mov_b32 v0, s6 :: v_dual_mov_b32 v1, s7
 ; GFX11-NEXT:    v_mov_b32_e32 v2, s8
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    image_bvh64_intersect_ray v[0:3], [v[6:7], v8, v[0:2], v[3:5]], s[0:3] a16
diff --git a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
--- a/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
+++ b/llvm/test/CodeGen/AMDGPU/atomic_optimizations_local_pointer.ll
@@ -212,10 +212,9 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB0_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s3, s3
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
+; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 ; GFX1132-NEXT:    s_mul_i32 s3, s3, 5
-; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-NEXT:    v_mov_b32_e32 v2, s3
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s3
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_add_rtn_u32 v1, v1, v2
@@ -454,11 +453,10 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB1_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s2, s2
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX1132-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1132-NEXT:    s_mul_i32 s2, s0, s2
 ; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-NEXT:    v_mov_b32_e32 v2, s2
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s2
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_add_rtn_u32 v1, v1, v2
@@ -1569,8 +1567,8 @@
 ; GFX1132-NEXT:    s_mul_hi_u32 s7, s2, s5
 ; GFX1132-NEXT:    s_mul_i32 s5, s2, s5
 ; GFX1132-NEXT:    s_add_i32 s7, s7, s6
-; GFX1132-NEXT:    v_mov_b32_e32 v0, s5
-; GFX1132-NEXT:    v_mov_b32_e32 v1, s7
+; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s7
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_add_rtn_u64 v[0:1], v3, v[0:1]
@@ -1878,10 +1876,9 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB7_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s3, s3
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
+; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(SALU_CYCLE_1)
 ; GFX1132-NEXT:    s_mul_i32 s3, s3, 5
-; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-NEXT:    v_mov_b32_e32 v2, s3
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s3
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_sub_rtn_u32 v1, v1, v2
@@ -2124,11 +2121,10 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB8_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    s_bcnt1_i32_b32 s2, s2
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX1132-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1132-NEXT:    s_mul_i32 s2, s0, s2
 ; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
-; GFX1132-NEXT:    v_mov_b32_e32 v2, s2
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, s2
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_sub_rtn_u32 v1, v1, v2
@@ -3265,8 +3261,8 @@
 ; GFX1132-NEXT:    s_mul_hi_u32 s7, s2, s5
 ; GFX1132-NEXT:    s_mul_i32 s5, s2, s5
 ; GFX1132-NEXT:    s_add_i32 s7, s7, s6
-; GFX1132-NEXT:    v_mov_b32_e32 v0, s5
-; GFX1132-NEXT:    v_mov_b32_e32 v1, s7
+; GFX1132-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX1132-NEXT:    v_dual_mov_b32 v0, s5 :: v_dual_mov_b32 v1, s7
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_sub_rtn_u64 v[0:1], v3, v[0:1]
@@ -5037,8 +5033,7 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB18_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    v_mov_b32_e32 v0, 5
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
-; GFX1132-NEXT:    v_mov_b32_e32 v2, 0
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_max_rtn_i64 v[0:1], v2, v[0:1]
@@ -5648,8 +5643,7 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB20_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    v_mov_b32_e32 v0, 5
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
-; GFX1132-NEXT:    v_mov_b32_e32 v2, 0
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_min_rtn_i64 v[0:1], v2, v[0:1]
@@ -6256,8 +6250,7 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB22_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    v_mov_b32_e32 v0, 5
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
-; GFX1132-NEXT:    v_mov_b32_e32 v2, 0
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_max_rtn_u64 v[0:1], v2, v[0:1]
@@ -6864,8 +6857,7 @@
 ; GFX1132-NEXT:    s_cbranch_execz .LBB24_2
 ; GFX1132-NEXT:  ; %bb.1:
 ; GFX1132-NEXT:    v_mov_b32_e32 v0, 5
-; GFX1132-NEXT:    v_mov_b32_e32 v1, 0
-; GFX1132-NEXT:    v_mov_b32_e32 v2, 0
+; GFX1132-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v2, 0
 ; GFX1132-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX1132-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX1132-NEXT:    ds_min_rtn_u64 v[0:1], v2, v[0:1]
diff --git a/llvm/test/CodeGen/AMDGPU/dual-source-blend-export.ll b/llvm/test/CodeGen/AMDGPU/dual-source-blend-export.ll
--- a/llvm/test/CodeGen/AMDGPU/dual-source-blend-export.ll
+++ b/llvm/test/CodeGen/AMDGPU/dual-source-blend-export.ll
@@ -8,17 +8,16 @@
 ; GCN-NEXT:    s_mov_b32 s1, exec_lo
 ; GCN-NEXT:    s_wqm_b32 exec_lo, exec_lo
 ; GCN-NEXT:    s_mov_b32 m0, s0
-; GCN-NEXT:    v_mov_b32_e32 v2, v0
+; GCN-NEXT:    v_mbcnt_lo_u32_b32 v7, -1, 0
 ; GCN-NEXT:    lds_param_load v3, attr1.x wait_vdst:15
 ; GCN-NEXT:    lds_param_load v4, attr1.y wait_vdst:15
 ; GCN-NEXT:    lds_param_load v5, attr1.z wait_vdst:15
 ; GCN-NEXT:    lds_param_load v6, attr1.w wait_vdst:15
-; GCN-NEXT:    v_mbcnt_lo_u32_b32 v7, -1, 0
-; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GCN-NEXT:    v_mbcnt_hi_u32_b32 v7, -1, v7
-; GCN-NEXT:    v_and_b32_e32 v7, 1, v7
-; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_4) | instid1(VALU_DEP_4)
+; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GCN-NEXT:    v_dual_mov_b32 v2, v0 :: v_dual_and_b32 v7, 1, v7
 ; GCN-NEXT:    v_cmp_eq_u32_e32 vcc_lo, 0, v7
+; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
 ; GCN-NEXT:    v_interp_p10_f32 v8, v4, v2, v4 wait_exp:2
 ; GCN-NEXT:    v_interp_p10_f32 v10, v5, v2, v5 wait_exp:1
 ; GCN-NEXT:    v_interp_p10_f32 v9, v6, v2, v6
@@ -32,9 +31,8 @@
 ; GCN-NEXT:    v_mov_b32_dpp v4, v4 dpp8:[1,0,3,2,5,4,7,6]
 ; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GCN-NEXT:    v_mov_b32_dpp v6, v6 dpp8:[1,0,3,2,5,4,7,6]
-; GCN-NEXT:    v_cndmask_b32_e32 v3, v4, v5, vcc_lo
-; GCN-NEXT:    v_cndmask_b32_e32 v4, v5, v4, vcc_lo
-; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GCN-NEXT:    v_dual_cndmask_b32 v3, v4, v5 :: v_dual_cndmask_b32 v4, v5, v4
+; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GCN-NEXT:    v_cndmask_b32_e32 v5, v2, v6, vcc_lo
 ; GCN-NEXT:    v_cndmask_b32_e32 v2, v6, v2, vcc_lo
 ; GCN-NEXT:    v_mov_b32_dpp v4, v4 dpp8:[1,0,3,2,5,4,7,6]
diff --git a/llvm/test/CodeGen/AMDGPU/flat-scratch-svs.ll b/llvm/test/CodeGen/AMDGPU/flat-scratch-svs.ll
--- a/llvm/test/CodeGen/AMDGPU/flat-scratch-svs.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-scratch-svs.ll
@@ -47,8 +47,7 @@
 ; GFX11-SDAG-LABEL: soff1_voff1:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
 ; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 4
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-SDAG-NEXT:    v_add3_u32 v0, 4, s0, v0
@@ -64,13 +63,11 @@
 ; GFX11-GISEL-LABEL: soff1_voff1:
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -135,11 +132,10 @@
 ; GFX11-SDAG-LABEL: soff1_voff2:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_lshlrev_b32 v0, 1, v0
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v2, 2 :: v_dual_mov_b32 v3, 4
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-SDAG-NEXT:    v_add3_u32 v0, 4, s0, v0
 ; GFX11-SDAG-NEXT:    scratch_store_b8 v0, v1, off offset:1 dlc
 ; GFX11-SDAG-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -154,13 +150,11 @@
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -225,10 +219,8 @@
 ; GFX11-SDAG-LABEL: soff1_voff4:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v2, 2 :: v_dual_mov_b32 v3, 4
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-SDAG-NEXT:    s_add_i32 s0, s0, 4
 ; GFX11-SDAG-NEXT:    scratch_store_b8 v0, v1, s0 offset:1 dlc
@@ -244,13 +236,11 @@
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -315,8 +305,7 @@
 ; GFX11-SDAG-LABEL: soff2_voff1:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
 ; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 4
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-SDAG-NEXT:    s_lshl_b32 s0, s0, 1
@@ -334,14 +323,12 @@
 ; GFX11-GISEL-LABEL: soff2_voff1:
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    s_lshl_b32 s0, s0, 1
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -408,13 +395,11 @@
 ; GFX11-SDAG-LABEL: soff2_voff2:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_lshlrev_b32 v0, 1, v0
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v2, 2 :: v_dual_mov_b32 v3, 4
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-SDAG-NEXT:    s_lshl_b32 s0, s0, 1
-; GFX11-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; GFX11-SDAG-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instid1(SALU_CYCLE_1)
 ; GFX11-SDAG-NEXT:    v_add3_u32 v0, 4, s0, v0
 ; GFX11-SDAG-NEXT:    scratch_store_b8 v0, v1, off offset:1 dlc
 ; GFX11-SDAG-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -429,14 +414,12 @@
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    s_lshl_b32 s0, s0, 1
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -503,10 +486,8 @@
 ; GFX11-SDAG-LABEL: soff2_voff4:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v3, 4 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-SDAG-NEXT:    s_lshl_b32 s0, s0, 1
 ; GFX11-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
@@ -524,14 +505,12 @@
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    s_lshl_b32 s0, s0, 1
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -596,9 +575,8 @@
 ; GFX11-SDAG-LABEL: soff4_voff1:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v4, 4
 ; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 2
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v4, 4
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-SDAG-NEXT:    s_lshl_b32 s0, s0, 2
 ; GFX11-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
@@ -616,14 +594,12 @@
 ; GFX11-GISEL-LABEL: soff4_voff1:
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    s_lshl_b32 s0, s0, 2
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -690,8 +666,7 @@
 ; GFX11-SDAG-LABEL: soff4_voff2:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_lshlrev_b32 v0, 1, v0
 ; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
 ; GFX11-SDAG-NEXT:    v_mov_b32_e32 v4, 4
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
@@ -712,14 +687,12 @@
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    s_lshl_b32 s0, s0, 2
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
@@ -786,10 +759,8 @@
 ; GFX11-SDAG-LABEL: soff4_voff4:
 ; GFX11-SDAG:       ; %bb.0: ; %bb
 ; GFX11-SDAG-NEXT:    s_load_b32 s0, s[0:1], 0x24
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-SDAG-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v2, 2
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v3, 4 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-SDAG-NEXT:    s_lshl_b32 s0, s0, 2
 ; GFX11-SDAG-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
@@ -807,14 +778,12 @@
 ; GFX11-GISEL:       ; %bb.0: ; %bb
 ; GFX11-GISEL-NEXT:    s_load_b32 s0, s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v2, 1
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v3, 4
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v2, 1 :: v_dual_mov_b32 v3, 4
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-GISEL-NEXT:    s_lshl_b32 s0, s0, 2
 ; GFX11-GISEL-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-GISEL-NEXT:    v_add_nc_u32_e64 v1, s0, 4
-; GFX11-GISEL-NEXT:    v_add_nc_u32_e32 v0, v1, v0
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 2 :: v_dual_add_nc_u32 v0, v1, v0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v2, off offset:1 dlc
 ; GFX11-GISEL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-GISEL-NEXT:    scratch_store_b8 v0, v1, off offset:2 dlc
diff --git a/llvm/test/CodeGen/AMDGPU/flat-scratch.ll b/llvm/test/CodeGen/AMDGPU/flat-scratch.ll
--- a/llvm/test/CodeGen/AMDGPU/flat-scratch.ll
+++ b/llvm/test/CodeGen/AMDGPU/flat-scratch.ll
@@ -58,10 +58,8 @@
 ; GFX11-NEXT:    s_mov_b32 s1, s0
 ; GFX11-NEXT:    s_mov_b32 s2, s0
 ; GFX11-NEXT:    s_mov_b32 s3, s0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-NEXT:    s_clause 0x3
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], off offset:64
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], off offset:48
@@ -175,10 +173,8 @@
 ; GFX11-PAL-NEXT:    s_mov_b32 s1, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s2, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s3, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-PAL-NEXT:    s_clause 0x3
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], off offset:64
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], off offset:48
@@ -239,10 +235,8 @@
 ; GFX11-NEXT:    s_mov_b32 s1, s0
 ; GFX11-NEXT:    s_mov_b32 s2, s0
 ; GFX11-NEXT:    s_mov_b32 s3, s0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-NEXT:    s_clause 0x3
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:48
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:32
@@ -313,10 +307,8 @@
 ; GFX11-PAL-NEXT:    s_mov_b32 s1, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s2, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s3, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-PAL-NEXT:    s_clause 0x3
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:48
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:32
@@ -685,9 +677,8 @@
 ;
 ; GFX11-LABEL: store_load_vindex_kernel:
 ; GFX11:       ; %bb.0: ; %bb
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 15
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_sub_nc_u32_e32 v2, 4, v0
 ; GFX11-NEXT:    scratch_store_b32 v0, v1, off offset:4 dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -748,9 +739,8 @@
 ;
 ; GFX11-PAL-LABEL: store_load_vindex_kernel:
 ; GFX11-PAL:       ; %bb.0: ; %bb
-; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 15
-; GFX11-PAL-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-PAL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-PAL-NEXT:    v_sub_nc_u32_e32 v2, 4, v0
 ; GFX11-PAL-NEXT:    scratch_store_b32 v0, v1, off offset:4 dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -815,10 +805,9 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_and_b32_e32 v1, 15, v0
+; GFX11-NEXT:    v_dual_mov_b32 v2, 15 :: v_dual_and_b32 v1, 15, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v2, 15
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 2, v1
 ; GFX11-NEXT:    scratch_store_b32 v0, v2, s32 dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -871,10 +860,9 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_and_b32_e32 v1, 15, v0
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, 15 :: v_dual_and_b32 v1, 15, v0
 ; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 15
-; GFX11-PAL-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-PAL-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v1, 2, v1
 ; GFX11-PAL-NEXT:    scratch_store_b32 v0, v2, s32 dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1035,10 +1023,8 @@
 ; GFX11-NEXT:    s_mov_b32 s1, s0
 ; GFX11-NEXT:    s_mov_b32 s2, s0
 ; GFX11-NEXT:    s_mov_b32 s3, s0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-NEXT:    s_clause 0x3
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], off offset:272
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], off offset:288
@@ -1164,10 +1150,8 @@
 ; GFX11-PAL-NEXT:    s_mov_b32 s1, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s2, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s3, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-PAL-NEXT:    s_clause 0x3
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], off offset:272
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], off offset:288
@@ -1237,10 +1221,8 @@
 ; GFX11-NEXT:    s_mov_b32 s1, s0
 ; GFX11-NEXT:    s_mov_b32 s2, s0
 ; GFX11-NEXT:    s_mov_b32 s3, s0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-NEXT:    s_clause 0x3
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:256
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:272
@@ -1319,10 +1301,8 @@
 ; GFX11-PAL-NEXT:    s_mov_b32 s1, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s2, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s3, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-PAL-NEXT:    s_clause 0x3
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:256
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], s32 offset:272
@@ -1764,8 +1744,7 @@
 ;
 ; GFX11-LABEL: store_load_vindex_small_offset_kernel:
 ; GFX11:       ; %bb.0: ; %bb
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 15
+; GFX11-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    scratch_load_b32 v3, off, off offset:4 glc dlc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    v_sub_nc_u32_e32 v2, 0x104, v0
@@ -1859,8 +1838,7 @@
 ;
 ; GFX11-PAL-LABEL: store_load_vindex_small_offset_kernel:
 ; GFX11-PAL:       ; %bb.0: ; %bb
-; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 15
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-PAL-NEXT:    scratch_load_b32 v3, off, off offset:4 glc dlc
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-PAL-NEXT:    v_sub_nc_u32_e32 v2, 0x104, v0
@@ -1927,9 +1905,8 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_and_b32_e32 v1, 15, v0
+; GFX11-NEXT:    v_dual_mov_b32 v2, 15 :: v_dual_and_b32 v1, 15, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-NEXT:    scratch_load_b32 v3, off, s32 glc dlc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 2, v1
@@ -1993,9 +1970,8 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_and_b32_e32 v1, 15, v0
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, 15 :: v_dual_and_b32 v1, 15, v0
 ; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-PAL-NEXT:    scratch_load_b32 v3, off, s32 glc dlc
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v1, 2, v1
@@ -2095,10 +2071,8 @@
 ; GFX11-NEXT:    s_mov_b32 s1, s0
 ; GFX11-NEXT:    s_mov_b32 s2, s0
 ; GFX11-NEXT:    s_mov_b32 s3, s0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo
 ; GFX11-NEXT:    s_movk_i32 vcc_lo, 0x4010
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo offset:16
@@ -2234,10 +2208,8 @@
 ; GFX11-PAL-NEXT:    s_mov_b32 s1, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s2, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s3, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo
 ; GFX11-PAL-NEXT:    s_movk_i32 vcc_lo, 0x4010
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo offset:16
@@ -2317,10 +2289,8 @@
 ; GFX11-NEXT:    s_mov_b32 s1, s0
 ; GFX11-NEXT:    s_mov_b32 s2, s0
 ; GFX11-NEXT:    s_mov_b32 s3, s0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo
 ; GFX11-NEXT:    s_add_i32 vcc_lo, s32, 0x4010
 ; GFX11-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo offset:16
@@ -2441,10 +2411,8 @@
 ; GFX11-PAL-NEXT:    s_mov_b32 s1, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s2, s0
 ; GFX11-PAL-NEXT:    s_mov_b32 s3, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, s1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, s2
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, s2 :: v_dual_mov_b32 v3, s3
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo
 ; GFX11-PAL-NEXT:    s_add_i32 vcc_lo, s32, 0x4010
 ; GFX11-PAL-NEXT:    scratch_store_b128 off, v[0:3], vcc_lo offset:16
@@ -2871,8 +2839,7 @@
 ;
 ; GFX11-LABEL: store_load_vindex_large_offset_kernel:
 ; GFX11:       ; %bb.0: ; %bb
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 15
+; GFX11-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_movk_i32 vcc_lo, 0x4004
 ; GFX11-NEXT:    scratch_load_b32 v3, off, off offset:4 glc dlc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -2968,8 +2935,7 @@
 ;
 ; GFX11-PAL-LABEL: store_load_vindex_large_offset_kernel:
 ; GFX11-PAL:       ; %bb.0: ; %bb
-; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 15
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-PAL-NEXT:    s_movk_i32 vcc_lo, 0x4004
 ; GFX11-PAL-NEXT:    scratch_load_b32 v3, off, off offset:4 glc dlc
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0)
@@ -3037,9 +3003,8 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_and_b32_e32 v1, 15, v0
+; GFX11-NEXT:    v_dual_mov_b32 v2, 15 :: v_dual_and_b32 v1, 15, v0
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-NEXT:    s_add_i32 vcc_lo, s32, 0x4004
 ; GFX11-NEXT:    scratch_load_b32 v3, off, s32 offset:4 glc dlc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -3107,9 +3072,8 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_and_b32_e32 v1, 15, v0
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v2, 15 :: v_dual_and_b32 v1, 15, v0
 ; GFX11-PAL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-PAL-NEXT:    s_add_i32 vcc_lo, s32, 0x4004
 ; GFX11-PAL-NEXT:    scratch_load_b32 v3, off, s32 offset:4 glc dlc
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0)
@@ -3190,8 +3154,7 @@
 ;
 ; GFX11-LABEL: store_load_large_imm_offset_kernel:
 ; GFX11:       ; %bb.0: ; %bb
-; GFX11-NEXT:    v_mov_b32_e32 v0, 13
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0x3000
+; GFX11-NEXT:    v_dual_mov_b32 v0, 13 :: v_dual_mov_b32 v1, 0x3000
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-NEXT:    scratch_store_b32 off, v0, off offset:4 dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3285,8 +3248,7 @@
 ;
 ; GFX11-PAL-LABEL: store_load_large_imm_offset_kernel:
 ; GFX11-PAL:       ; %bb.0: ; %bb
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, 13
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 0x3000
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, 13 :: v_dual_mov_b32 v1, 0x3000
 ; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-PAL-NEXT:    scratch_store_b32 off, v0, off offset:4 dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3344,8 +3306,7 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v0, 13
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0x3000
+; GFX11-NEXT:    v_dual_mov_b32 v0, 13 :: v_dual_mov_b32 v1, 0x3000
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-NEXT:    scratch_store_b32 off, v0, s32 offset:4 dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3406,8 +3367,7 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, 13
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 0x3000
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, 13 :: v_dual_mov_b32 v1, 0x3000
 ; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 15
 ; GFX11-PAL-NEXT:    scratch_store_b32 off, v0, s32 offset:4 dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3601,8 +3561,7 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 15
-; GFX11-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_mov_b32 v2, 0
 ; GFX11-NEXT:    scratch_store_b64 v0, v[1:2], off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    scratch_load_b64 v[0:1], v0, off glc dlc
@@ -3647,8 +3606,7 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 15
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_mov_b32 v2, 0
 ; GFX11-PAL-NEXT:    scratch_store_b64 v0, v[1:2], off dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-PAL-NEXT:    scratch_load_b64 v[0:1], v0, off glc dlc
@@ -3698,8 +3656,7 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 15
-; GFX11-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_mov_b32 v2, 0
 ; GFX11-NEXT:    scratch_store_b64 v0, v[1:2], off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    scratch_load_b64 v[0:1], v0, off glc dlc
@@ -3744,8 +3701,7 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 15
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 15 :: v_dual_mov_b32 v2, 0
 ; GFX11-PAL-NEXT:    scratch_store_b64 v0, v[1:2], off dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-PAL-NEXT:    scratch_load_b64 v[0:1], v0, off glc dlc
@@ -3797,8 +3753,7 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-NEXT:    v_mov_b32_e32 v2, 2
+; GFX11-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
 ; GFX11-NEXT:    v_mov_b32_e32 v3, 3
 ; GFX11-NEXT:    scratch_store_b96 v0, v[1:3], off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3847,8 +3802,7 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 2
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
 ; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, 3
 ; GFX11-PAL-NEXT:    scratch_store_b96 v0, v[1:3], off dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3904,10 +3858,8 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-NEXT:    v_mov_b32_e32 v2, 2
-; GFX11-NEXT:    v_mov_b32_e32 v3, 3
-; GFX11-NEXT:    v_mov_b32_e32 v4, 4
+; GFX11-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
+; GFX11-NEXT:    v_dual_mov_b32 v3, 3 :: v_dual_mov_b32 v4, 4
 ; GFX11-NEXT:    scratch_store_b128 v0, v[1:4], off dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    scratch_load_b128 v[0:3], v0, off glc dlc
@@ -3958,10 +3910,8 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 1
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v2, 2
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v3, 3
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v4, 4
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_mov_b32 v2, 2
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v3, 3 :: v_dual_mov_b32 v4, 4
 ; GFX11-PAL-NEXT:    scratch_store_b128 v0, v[1:4], off dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-PAL-NEXT:    scratch_load_b128 v[0:3], v0, off glc dlc
@@ -4109,8 +4059,7 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_add_nc_u32_e32 v0, 0xfffff000, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 1
+; GFX11-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_add_nc_u32 v0, 0xfffff000, v0
 ; GFX11-NEXT:    scratch_store_b8 v0, v1, off offset:-129 dlc
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    scratch_load_u8 v0, v0, off offset:-129 glc dlc
@@ -4167,8 +4116,7 @@
 ; GFX11-PAL:       ; %bb.0: ; %bb
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-PAL-NEXT:    v_add_nc_u32_e32 v0, 0xfffff000, v0
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 1
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v1, 1 :: v_dual_add_nc_u32 v0, 0xfffff000, v0
 ; GFX11-PAL-NEXT:    scratch_store_b8 v0, v1, off offset:-129 dlc
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-PAL-NEXT:    scratch_load_u8 v0, v0, off offset:-129 glc dlc
@@ -4243,8 +4191,7 @@
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    scratch_load_b128 v[0:3], off, off offset:3024 glc dlc
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, 16
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0x810
+; GFX11-NEXT:    v_dual_mov_b32 v0, 16 :: v_dual_mov_b32 v1, 0x810
 ; GFX11-NEXT:    ;;#ASMSTART
 ; GFX11-NEXT:    ; use v0
 ; GFX11-NEXT:    ;;#ASMEND
@@ -4344,8 +4291,7 @@
 ; GFX11-PAL-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-PAL-NEXT:    scratch_load_b128 v[0:3], off, off offset:3024 glc dlc
 ; GFX11-PAL-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v0, 16
-; GFX11-PAL-NEXT:    v_mov_b32_e32 v1, 0x810
+; GFX11-PAL-NEXT:    v_dual_mov_b32 v0, 16 :: v_dual_mov_b32 v1, 0x810
 ; GFX11-PAL-NEXT:    ;;#ASMSTART
 ; GFX11-PAL-NEXT:    ; use v0
 ; GFX11-PAL-NEXT:    ;;#ASMEND
diff --git a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
--- a/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
+++ b/llvm/test/CodeGen/AMDGPU/llc-pipeline.ll
@@ -377,6 +377,7 @@
 ; GCN-O1-NEXT:        Branch Probability Basic Block Placement
 ; GCN-O1-NEXT:        Insert fentry calls
 ; GCN-O1-NEXT:        Insert XRay ops
+; GCN-O1-NEXT:        GCN Create VOPD Instructions
 ; GCN-O1-NEXT:        SI Memory Legalizer
 ; GCN-O1-NEXT:        MachinePostDominator Tree Construction
 ; GCN-O1-NEXT:        SI insert wait instructions
@@ -665,6 +666,7 @@
 ; GCN-O1-OPTS-NEXT:        Branch Probability Basic Block Placement
 ; GCN-O1-OPTS-NEXT:        Insert fentry calls
 ; GCN-O1-OPTS-NEXT:        Insert XRay ops
+; GCN-O1-OPTS-NEXT:        GCN Create VOPD Instructions
 ; GCN-O1-OPTS-NEXT:        SI Memory Legalizer
 ; GCN-O1-OPTS-NEXT:        MachinePostDominator Tree Construction
 ; GCN-O1-OPTS-NEXT:        SI insert wait instructions
@@ -955,6 +957,7 @@
 ; GCN-O2-NEXT:        Branch Probability Basic Block Placement
 ; GCN-O2-NEXT:        Insert fentry calls
 ; GCN-O2-NEXT:        Insert XRay ops
+; GCN-O2-NEXT:        GCN Create VOPD Instructions
 ; GCN-O2-NEXT:        SI Memory Legalizer
 ; GCN-O2-NEXT:        MachinePostDominator Tree Construction
 ; GCN-O2-NEXT:        SI insert wait instructions
@@ -1258,6 +1261,7 @@
 ; GCN-O3-NEXT:        Branch Probability Basic Block Placement
 ; GCN-O3-NEXT:        Insert fentry calls
 ; GCN-O3-NEXT:        Insert XRay ops
+; GCN-O3-NEXT:        GCN Create VOPD Instructions
 ; GCN-O3-NEXT:        SI Memory Legalizer
 ; GCN-O3-NEXT:        MachinePostDominator Tree Construction
 ; GCN-O3-NEXT:        SI insert wait instructions
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fdot2.f32.bf16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fdot2.f32.bf16.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fdot2.f32.bf16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.fdot2.f32.bf16.ll
@@ -8,13 +8,12 @@
 ; GFX11-LABEL: test_llvm_amdgcn_fdot2_f32_bf16_clamp:
 ; GFX11:       ; %bb.0: ; %entry
 ; GFX11-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    s_load_b32 s6, s[6:7], 0x0
 ; GFX11-NEXT:    s_load_b32 s2, s[2:3], 0x0
 ; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s6
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s6
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_dot2_f32_bf16 v0, s2, s3, v0 clamp
 ; GFX11-NEXT:    global_store_b32 v1, v0, s[0:1]
@@ -38,13 +37,12 @@
 ; GFX11-LABEL: test_llvm_amdgcn_fdot2_f32_bf16_no_clamp:
 ; GFX11:       ; %bb.0: ; %entry
 ; GFX11-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    s_load_b32 s6, s[6:7], 0x0
 ; GFX11-NEXT:    s_load_b32 s2, s[2:3], 0x0
 ; GFX11-NEXT:    s_load_b32 s3, s[4:5], 0x0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s6
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s6
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_dot2_f32_bf16 v0, s2, s3, v0
 ; GFX11-NEXT:    global_store_b32 v1, v0, s[0:1]
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.dim.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.dim.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.dim.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.dim.ll
@@ -124,17 +124,14 @@
 ;
 ; GFX11-LABEL: load_1d_tfe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v6, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v6
-; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v10, v6
+; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_mov_b32_e32 v7, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v6
-; GFX11-NEXT:    v_mov_b32_e32 v1, v7
-; GFX11-NEXT:    v_mov_b32_e32 v2, v8
-; GFX11-NEXT:    v_mov_b32_e32 v3, v9
-; GFX11-NEXT:    v_mov_b32_e32 v4, v10
+; GFX11-NEXT:    v_dual_mov_b32 v2, v8 :: v_dual_mov_b32 v1, v7
+; GFX11-NEXT:    v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v10
 ; GFX11-NEXT:    image_load v[0:4], v5, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v6, v4, s[8:9]
@@ -231,17 +228,14 @@
 ;
 ; GFX11-LABEL: load_1d_lwe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v6, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v6
-; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v10, v6
+; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_mov_b32_e32 v7, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v6
-; GFX11-NEXT:    v_mov_b32_e32 v1, v7
-; GFX11-NEXT:    v_mov_b32_e32 v2, v8
-; GFX11-NEXT:    v_mov_b32_e32 v3, v9
-; GFX11-NEXT:    v_mov_b32_e32 v4, v10
+; GFX11-NEXT:    v_dual_mov_b32 v2, v8 :: v_dual_mov_b32 v1, v7
+; GFX11-NEXT:    v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v10
 ; GFX11-NEXT:    image_load v[0:4], v5, s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v6, v4, s[8:9]
@@ -377,18 +371,15 @@
 ;
 ; GFX11-LABEL: load_2d_tfe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v7, 0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v7
-; GFX11-NEXT:    v_mov_b32_e32 v10, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v7
+; GFX11-NEXT:    v_mov_b32_e32 v10, v7
+; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v7
-; GFX11-NEXT:    v_mov_b32_e32 v1, v8
-; GFX11-NEXT:    v_mov_b32_e32 v2, v9
-; GFX11-NEXT:    v_mov_b32_e32 v3, v10
-; GFX11-NEXT:    v_mov_b32_e32 v4, v11
+; GFX11-NEXT:    v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, v8 :: v_dual_mov_b32 v4, v11
 ; GFX11-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v7, v4, s[8:9]
@@ -528,19 +519,15 @@
 ;
 ; GFX11-LABEL: load_3d_tfe_lwe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v8, 0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v9, v8
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v8, 0
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_mov_b32_e32 v9, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-NEXT:    v_mov_b32_e32 v1, v9
-; GFX11-NEXT:    v_mov_b32_e32 v2, v10
-; GFX11-NEXT:    v_mov_b32_e32 v3, v11
-; GFX11-NEXT:    v_mov_b32_e32 v4, v12
+; GFX11-NEXT:    v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v1, v9
+; GFX11-NEXT:    v_dual_mov_b32 v3, v11 :: v_dual_mov_b32 v4, v12
 ; GFX11-NEXT:    image_load v[0:4], v[5:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_3D unorm tfe lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v8, v4, s[8:9]
@@ -680,19 +667,15 @@
 ;
 ; GFX11-LABEL: load_cube_lwe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v8, 0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v9, v8
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v8, 0
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_mov_b32_e32 v9, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-NEXT:    v_mov_b32_e32 v1, v9
-; GFX11-NEXT:    v_mov_b32_e32 v2, v10
-; GFX11-NEXT:    v_mov_b32_e32 v3, v11
-; GFX11-NEXT:    v_mov_b32_e32 v4, v12
+; GFX11-NEXT:    v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v1, v9
+; GFX11-NEXT:    v_dual_mov_b32 v3, v11 :: v_dual_mov_b32 v4, v12
 ; GFX11-NEXT:    image_load v[0:4], v[5:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_CUBE unorm lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v8, v4, s[8:9]
@@ -828,18 +811,15 @@
 ;
 ; GFX11-LABEL: load_1darray_tfe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v7, 0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v7
-; GFX11-NEXT:    v_mov_b32_e32 v10, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v7
+; GFX11-NEXT:    v_mov_b32_e32 v10, v7
+; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v7
-; GFX11-NEXT:    v_mov_b32_e32 v1, v8
-; GFX11-NEXT:    v_mov_b32_e32 v2, v9
-; GFX11-NEXT:    v_mov_b32_e32 v3, v10
-; GFX11-NEXT:    v_mov_b32_e32 v4, v11
+; GFX11-NEXT:    v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, v8 :: v_dual_mov_b32 v4, v11
 ; GFX11-NEXT:    image_load v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D_ARRAY unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v7, v4, s[8:9]
@@ -979,19 +959,15 @@
 ;
 ; GFX11-LABEL: load_2darray_lwe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v8, 0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v9, v8
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v8, 0
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_mov_b32_e32 v9, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-NEXT:    v_mov_b32_e32 v1, v9
-; GFX11-NEXT:    v_mov_b32_e32 v2, v10
-; GFX11-NEXT:    v_mov_b32_e32 v3, v11
-; GFX11-NEXT:    v_mov_b32_e32 v4, v12
+; GFX11-NEXT:    v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v1, v9
+; GFX11-NEXT:    v_dual_mov_b32 v3, v11 :: v_dual_mov_b32 v4, v12
 ; GFX11-NEXT:    image_load v[0:4], v[5:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_ARRAY unorm lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v8, v4, s[8:9]
@@ -1131,19 +1107,15 @@
 ;
 ; GFX11-LABEL: load_2dmsaa_both:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v8, 0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v9, v8
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v8, 0
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_mov_b32_e32 v9, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-NEXT:    v_mov_b32_e32 v1, v9
-; GFX11-NEXT:    v_mov_b32_e32 v2, v10
-; GFX11-NEXT:    v_mov_b32_e32 v3, v11
-; GFX11-NEXT:    v_mov_b32_e32 v4, v12
+; GFX11-NEXT:    v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v1, v9
+; GFX11-NEXT:    v_dual_mov_b32 v3, v11 :: v_dual_mov_b32 v4, v12
 ; GFX11-NEXT:    image_load v[0:4], v[5:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA unorm tfe lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v8, v4, s[8:9]
@@ -1287,20 +1259,16 @@
 ;
 ; GFX11-LABEL: load_2darraymsaa_tfe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v9, 0
-; GFX11-NEXT:    v_mov_b32_e32 v8, v3
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v9, 0 :: v_dual_mov_b32 v8, v3
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v10, v9
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v9
-; GFX11-NEXT:    v_mov_b32_e32 v12, v9
 ; GFX11-NEXT:    v_mov_b32_e32 v13, v9
+; GFX11-NEXT:    v_mov_b32_e32 v12, v9
+; GFX11-NEXT:    v_mov_b32_e32 v10, v9
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v9
-; GFX11-NEXT:    v_mov_b32_e32 v1, v10
-; GFX11-NEXT:    v_mov_b32_e32 v2, v11
-; GFX11-NEXT:    v_mov_b32_e32 v3, v12
-; GFX11-NEXT:    v_mov_b32_e32 v4, v13
+; GFX11-NEXT:    v_dual_mov_b32 v2, v11 :: v_dual_mov_b32 v3, v12
+; GFX11-NEXT:    v_dual_mov_b32 v1, v10 :: v_dual_mov_b32 v4, v13
 ; GFX11-NEXT:    image_load v[0:4], v[5:8], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D_MSAA_ARRAY unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v9, v4, s[8:9]
@@ -1436,18 +1404,15 @@
 ;
 ; GFX11-LABEL: load_mip_1d_lwe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v7, 0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
+; GFX11-NEXT:    v_dual_mov_b32 v7, 0 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v7
-; GFX11-NEXT:    v_mov_b32_e32 v10, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v7
+; GFX11-NEXT:    v_mov_b32_e32 v10, v7
+; GFX11-NEXT:    v_mov_b32_e32 v8, v7
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v7
-; GFX11-NEXT:    v_mov_b32_e32 v1, v8
-; GFX11-NEXT:    v_mov_b32_e32 v2, v9
-; GFX11-NEXT:    v_mov_b32_e32 v3, v10
-; GFX11-NEXT:    v_mov_b32_e32 v4, v11
+; GFX11-NEXT:    v_dual_mov_b32 v2, v9 :: v_dual_mov_b32 v3, v10
+; GFX11-NEXT:    v_dual_mov_b32 v1, v8 :: v_dual_mov_b32 v4, v11
 ; GFX11-NEXT:    image_load_mip v[0:4], v[5:6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_1D unorm lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v7, v4, s[8:9]
@@ -1587,19 +1552,15 @@
 ;
 ; GFX11-LABEL: load_mip_2d_tfe:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v8, 0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v2
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v9, v8
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v8, 0
+; GFX11-NEXT:    v_dual_mov_b32 v7, v2 :: v_dual_mov_b32 v6, v1
 ; GFX11-NEXT:    v_mov_b32_e32 v11, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v8
+; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_mov_b32_e32 v9, v8
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v8
-; GFX11-NEXT:    v_mov_b32_e32 v1, v9
-; GFX11-NEXT:    v_mov_b32_e32 v2, v10
-; GFX11-NEXT:    v_mov_b32_e32 v3, v11
-; GFX11-NEXT:    v_mov_b32_e32 v4, v12
+; GFX11-NEXT:    v_dual_mov_b32 v2, v10 :: v_dual_mov_b32 v1, v9
+; GFX11-NEXT:    v_dual_mov_b32 v3, v11 :: v_dual_mov_b32 v4, v12
 ; GFX11-NEXT:    image_load_mip v[0:4], v[5:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v8, v4, s[8:9]
@@ -1996,15 +1957,12 @@
 ;
 ; GFX11-LABEL: load_1d_tfe_V4_dmask3:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v5, 0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v5
+; GFX11-NEXT:    v_dual_mov_b32 v4, v0 :: v_dual_mov_b32 v5, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v7, v5
 ; GFX11-NEXT:    v_mov_b32_e32 v8, v5
-; GFX11-NEXT:    v_mov_b32_e32 v0, v5
-; GFX11-NEXT:    v_mov_b32_e32 v1, v6
-; GFX11-NEXT:    v_mov_b32_e32 v2, v7
-; GFX11-NEXT:    v_mov_b32_e32 v3, v8
+; GFX11-NEXT:    v_mov_b32_e32 v6, v5
+; GFX11-NEXT:    v_dual_mov_b32 v0, v5 :: v_dual_mov_b32 v3, v8
+; GFX11-NEXT:    v_dual_mov_b32 v1, v6 :: v_dual_mov_b32 v2, v7
 ; GFX11-NEXT:    image_load v[0:3], v4, s[0:7] dmask:0x7 dim:SQ_RSRC_IMG_1D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v5, v3, s[8:9]
@@ -2089,13 +2047,11 @@
 ;
 ; GFX11-LABEL: load_1d_tfe_V4_dmask2:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v4, 0
-; GFX11-NEXT:    v_mov_b32_e32 v3, v0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v4
+; GFX11-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_mov_b32 v4, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v6, v4
+; GFX11-NEXT:    v_mov_b32_e32 v5, v4
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v4
-; GFX11-NEXT:    v_mov_b32_e32 v1, v5
-; GFX11-NEXT:    v_mov_b32_e32 v2, v6
+; GFX11-NEXT:    v_dual_mov_b32 v2, v6 :: v_dual_mov_b32 v1, v5
 ; GFX11-NEXT:    image_load v[0:2], v3, s[0:7] dmask:0x6 dim:SQ_RSRC_IMG_1D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v4, v2, s[8:9]
@@ -2174,11 +2130,9 @@
 ;
 ; GFX11-LABEL: load_1d_tfe_V4_dmask1:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v3, 0
-; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    v_dual_mov_b32 v2, v0 :: v_dual_mov_b32 v3, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v4, v3
-; GFX11-NEXT:    v_mov_b32_e32 v0, v3
-; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    v_dual_mov_b32 v0, v3 :: v_dual_mov_b32 v1, v4
 ; GFX11-NEXT:    image_load v[0:1], v2, s[0:7] dmask:0x8 dim:SQ_RSRC_IMG_1D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v3, v1, s[8:9]
@@ -2257,11 +2211,9 @@
 ;
 ; GFX11-LABEL: load_1d_tfe_V2_dmask1:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v3, 0
-; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    v_dual_mov_b32 v2, v0 :: v_dual_mov_b32 v3, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v4, v3
-; GFX11-NEXT:    v_mov_b32_e32 v0, v3
-; GFX11-NEXT:    v_mov_b32_e32 v1, v4
+; GFX11-NEXT:    v_dual_mov_b32 v0, v3 :: v_dual_mov_b32 v1, v4
 ; GFX11-NEXT:    image_load v[0:1], v2, s[0:7] dmask:0x8 dim:SQ_RSRC_IMG_1D unorm tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v3, v1, s[8:9]
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.a16.dim.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.a16.dim.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.a16.dim.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.a16.dim.ll
@@ -1,7 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
-; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
 
 define amdgpu_ps <4 x float> @sample_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, half %s) {
 ; GFX9-LABEL: sample_1d:
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.d16.dim.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.d16.dim.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.d16.dim.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.d16.dim.ll
@@ -117,8 +117,7 @@
 ; GFX11-NEXT:    s_wqm_b32 exec_lo, exec_lo
 ; GFX11-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v5, v4
-; GFX11-NEXT:    v_mov_b32_e32 v2, v4
-; GFX11-NEXT:    v_mov_b32_e32 v3, v5
+; GFX11-NEXT:    v_dual_mov_b32 v2, v4 :: v_dual_mov_b32 v3, v5
 ; GFX11-NEXT:    s_and_b32 exec_lo, exec_lo, s14
 ; GFX11-NEXT:    image_sample v[2:3], v[0:1], s[0:7], s[8:11] dmask:0x1 dim:SQ_RSRC_IMG_2D tfe d16
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -199,15 +198,24 @@
 ; GFX9-NEXT:    v_mov_b32_e32 v1, v5
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: image_sample_c_d_1d_v2f16_tfe:
-; GFX10PLUS:       ; %bb.0: ; %main_body
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v5, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v4, v1
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, v0
-; GFX10PLUS-NEXT:    image_sample_c_d v[0:1], [v5, v4, v2, v3], s[0:7], s[8:11] dmask:0x3 dim:SQ_RSRC_IMG_1D tfe d16
-; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0)
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: image_sample_c_d_1d_v2f16_tfe:
+; GFX10:       ; %bb.0: ; %main_body
+; GFX10-NEXT:    v_mov_b32_e32 v5, v0
+; GFX10-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-NEXT:    v_mov_b32_e32 v4, v1
+; GFX10-NEXT:    v_mov_b32_e32 v1, v0
+; GFX10-NEXT:    image_sample_c_d v[0:1], [v5, v4, v2, v3], s[0:7], s[8:11] dmask:0x3 dim:SQ_RSRC_IMG_1D tfe d16
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: image_sample_c_d_1d_v2f16_tfe:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, 0
+; GFX11-NEXT:    v_mov_b32_e32 v1, v0
+; GFX11-NEXT:    image_sample_c_d v[0:1], [v5, v4, v2, v3], s[0:7], s[8:11] dmask:0x3 dim:SQ_RSRC_IMG_1D tfe d16
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    ; return to shader part epilog
 main_body:
   %tex = call {<2 x half>,i32} @llvm.amdgcn.image.sample.c.d.1d.v2f16i32.f32.f32(i32 3, float %zcompare, float %dsdh, float %dsdv, float %s, <8 x i32> %rsrc, <4 x i32> %samp, i1 false, i32 1, i32 0)
   %tex.vec = extractvalue {<2 x half>, i32} %tex, 0
@@ -313,20 +321,33 @@
 ; GFX9-NEXT:    v_mov_b32_e32 v2, v5
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: image_sample_b_2d_v3f16_tfe:
-; GFX10PLUS:       ; %bb.0: ; %main_body
-; GFX10PLUS-NEXT:    s_mov_b32 s12, exec_lo
-; GFX10PLUS-NEXT:    s_wqm_b32 exec_lo, exec_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v5, v2
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v4, v1
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, v0
-; GFX10PLUS-NEXT:    s_and_b32 exec_lo, exec_lo, s12
-; GFX10PLUS-NEXT:    image_sample_b v[0:2], v[3:5], s[0:7], s[8:11] dmask:0x7 dim:SQ_RSRC_IMG_2D tfe d16
-; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0)
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: image_sample_b_2d_v3f16_tfe:
+; GFX10:       ; %bb.0: ; %main_body
+; GFX10-NEXT:    s_mov_b32 s12, exec_lo
+; GFX10-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX10-NEXT:    v_mov_b32_e32 v3, v0
+; GFX10-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-NEXT:    v_mov_b32_e32 v5, v2
+; GFX10-NEXT:    v_mov_b32_e32 v4, v1
+; GFX10-NEXT:    v_mov_b32_e32 v1, v0
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX10-NEXT:    image_sample_b v[0:2], v[3:5], s[0:7], s[8:11] dmask:0x7 dim:SQ_RSRC_IMG_2D tfe d16
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: image_sample_b_2d_v3f16_tfe:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    s_mov_b32 s12, exec_lo
+; GFX11-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX11-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_mov_b32 v0, 0
+; GFX11-NEXT:    v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v0
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX11-NEXT:    image_sample_b v[0:2], v[3:5], s[0:7], s[8:11] dmask:0x7 dim:SQ_RSRC_IMG_2D tfe d16
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    ; return to shader part epilog
 main_body:
   %tex = call {<3 x half>,i32} @llvm.amdgcn.image.sample.b.2d.v3f16i32.f32.f32(i32 7, float %bias, float %s, float %t, <8 x i32> %rsrc, <4 x i32> %samp, i1 false, i32 1, i32 0)
   %tex.vec = extractvalue {<3 x half>, i32} %tex, 0
@@ -438,20 +459,33 @@
 ; GFX9-NEXT:    v_mov_b32_e32 v2, v5
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: image_sample_b_2d_v4f16_tfe:
-; GFX10PLUS:       ; %bb.0: ; %main_body
-; GFX10PLUS-NEXT:    s_mov_b32 s12, exec_lo
-; GFX10PLUS-NEXT:    s_wqm_b32 exec_lo, exec_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v5, v2
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v4, v1
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, v0
-; GFX10PLUS-NEXT:    s_and_b32 exec_lo, exec_lo, s12
-; GFX10PLUS-NEXT:    image_sample_b v[0:2], v[3:5], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D tfe d16
-; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0)
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: image_sample_b_2d_v4f16_tfe:
+; GFX10:       ; %bb.0: ; %main_body
+; GFX10-NEXT:    s_mov_b32 s12, exec_lo
+; GFX10-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX10-NEXT:    v_mov_b32_e32 v3, v0
+; GFX10-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-NEXT:    v_mov_b32_e32 v5, v2
+; GFX10-NEXT:    v_mov_b32_e32 v4, v1
+; GFX10-NEXT:    v_mov_b32_e32 v1, v0
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX10-NEXT:    image_sample_b v[0:2], v[3:5], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D tfe d16
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: image_sample_b_2d_v4f16_tfe:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    s_mov_b32 s12, exec_lo
+; GFX11-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX11-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_mov_b32 v0, 0
+; GFX11-NEXT:    v_dual_mov_b32 v5, v2 :: v_dual_mov_b32 v4, v1
+; GFX11-NEXT:    v_mov_b32_e32 v1, v0
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX11-NEXT:    image_sample_b v[0:2], v[3:5], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D tfe d16
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    ; return to shader part epilog
 main_body:
   %tex = call {<4 x half>,i32} @llvm.amdgcn.image.sample.b.2d.v4f16i32.f32.f32(i32 15, float %bias, float %s, float %t, <8 x i32> %rsrc, <4 x i32> %samp, i1 false, i32 1, i32 0)
   %tex.vec = extractvalue {<4 x half>, i32} %tex, 0
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.dim.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.dim.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.dim.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.dim.ll
@@ -104,17 +104,14 @@
 ; GFX11:       ; %bb.0: ; %main_body
 ; GFX11-NEXT:    s_mov_b32 s14, exec_lo
 ; GFX11-NEXT:    s_wqm_b32 exec_lo, exec_lo
-; GFX11-NEXT:    v_mov_b32_e32 v6, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v6
-; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v10, v6
+; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_mov_b32_e32 v7, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v6
-; GFX11-NEXT:    v_mov_b32_e32 v1, v7
-; GFX11-NEXT:    v_mov_b32_e32 v2, v8
-; GFX11-NEXT:    v_mov_b32_e32 v3, v9
-; GFX11-NEXT:    v_mov_b32_e32 v4, v10
+; GFX11-NEXT:    v_dual_mov_b32 v2, v8 :: v_dual_mov_b32 v1, v7
+; GFX11-NEXT:    v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v10
 ; GFX11-NEXT:    s_and_b32 exec_lo, exec_lo, s14
 ; GFX11-NEXT:    image_sample v[0:4], v5, s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_1D tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -344,18 +341,30 @@
 ; GFX6789-NEXT:    s_waitcnt vmcnt(0)
 ; GFX6789-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: sample_1d_tfe_adjust_writemask_12:
-; GFX10PLUS:       ; %bb.0: ; %main_body
-; GFX10PLUS-NEXT:    s_mov_b32 s12, exec_lo
-; GFX10PLUS-NEXT:    s_wqm_b32 exec_lo, exec_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, v0
-; GFX10PLUS-NEXT:    s_and_b32 exec_lo, exec_lo, s12
-; GFX10PLUS-NEXT:    image_sample v[0:2], v3, s[0:7], s[8:11] dmask:0x3 dim:SQ_RSRC_IMG_1D tfe
-; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0)
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: sample_1d_tfe_adjust_writemask_12:
+; GFX10:       ; %bb.0: ; %main_body
+; GFX10-NEXT:    s_mov_b32 s12, exec_lo
+; GFX10-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX10-NEXT:    v_mov_b32_e32 v3, v0
+; GFX10-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-NEXT:    v_mov_b32_e32 v1, v0
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX10-NEXT:    image_sample v[0:2], v3, s[0:7], s[8:11] dmask:0x3 dim:SQ_RSRC_IMG_1D tfe
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: sample_1d_tfe_adjust_writemask_12:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    s_mov_b32 s12, exec_lo
+; GFX11-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX11-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_mov_b32 v0, 0
+; GFX11-NEXT:    v_mov_b32_e32 v1, v0
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX11-NEXT:    image_sample v[0:2], v3, s[0:7], s[8:11] dmask:0x3 dim:SQ_RSRC_IMG_1D tfe
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    ; return to shader part epilog
 main_body:
   %v = call {<4 x float>,i32} @llvm.amdgcn.image.sample.1d.v4f32i32.f32(i32 15, float %s, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 1, i32 0)
   %res.vec = extractvalue {<4 x float>,i32} %v, 0
@@ -396,18 +405,30 @@
 ; GFX6789-NEXT:    s_waitcnt vmcnt(0)
 ; GFX6789-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: sample_1d_tfe_adjust_writemask_24:
-; GFX10PLUS:       ; %bb.0: ; %main_body
-; GFX10PLUS-NEXT:    s_mov_b32 s12, exec_lo
-; GFX10PLUS-NEXT:    s_wqm_b32 exec_lo, exec_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, v0
-; GFX10PLUS-NEXT:    s_and_b32 exec_lo, exec_lo, s12
-; GFX10PLUS-NEXT:    image_sample v[0:2], v3, s[0:7], s[8:11] dmask:0xa dim:SQ_RSRC_IMG_1D tfe
-; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0)
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: sample_1d_tfe_adjust_writemask_24:
+; GFX10:       ; %bb.0: ; %main_body
+; GFX10-NEXT:    s_mov_b32 s12, exec_lo
+; GFX10-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX10-NEXT:    v_mov_b32_e32 v3, v0
+; GFX10-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-NEXT:    v_mov_b32_e32 v1, v0
+; GFX10-NEXT:    v_mov_b32_e32 v2, v0
+; GFX10-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX10-NEXT:    image_sample v[0:2], v3, s[0:7], s[8:11] dmask:0xa dim:SQ_RSRC_IMG_1D tfe
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: sample_1d_tfe_adjust_writemask_24:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    s_mov_b32 s12, exec_lo
+; GFX11-NEXT:    s_wqm_b32 exec_lo, exec_lo
+; GFX11-NEXT:    v_dual_mov_b32 v3, v0 :: v_dual_mov_b32 v0, 0
+; GFX11-NEXT:    v_mov_b32_e32 v1, v0
+; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    s_and_b32 exec_lo, exec_lo, s12
+; GFX11-NEXT:    image_sample v[0:2], v3, s[0:7], s[8:11] dmask:0xa dim:SQ_RSRC_IMG_1D tfe
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    ; return to shader part epilog
 main_body:
   %v = call {<4 x float>,i32} @llvm.amdgcn.image.sample.1d.v4f32i32.f32(i32 15, float %s, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 1, i32 0)
   %res.vec = extractvalue {<4 x float>,i32} %v, 0
@@ -546,17 +567,14 @@
 ; GFX11:       ; %bb.0: ; %main_body
 ; GFX11-NEXT:    s_mov_b32 s14, exec_lo
 ; GFX11-NEXT:    s_wqm_b32 exec_lo, exec_lo
-; GFX11-NEXT:    v_mov_b32_e32 v6, 0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
-; GFX11-NEXT:    v_mov_b32_e32 v7, v6
-; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_dual_mov_b32 v5, v0 :: v_dual_mov_b32 v6, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v9, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v10, v6
+; GFX11-NEXT:    v_mov_b32_e32 v8, v6
+; GFX11-NEXT:    v_mov_b32_e32 v7, v6
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v6
-; GFX11-NEXT:    v_mov_b32_e32 v1, v7
-; GFX11-NEXT:    v_mov_b32_e32 v2, v8
-; GFX11-NEXT:    v_mov_b32_e32 v3, v9
-; GFX11-NEXT:    v_mov_b32_e32 v4, v10
+; GFX11-NEXT:    v_dual_mov_b32 v2, v8 :: v_dual_mov_b32 v1, v7
+; GFX11-NEXT:    v_dual_mov_b32 v3, v9 :: v_dual_mov_b32 v4, v10
 ; GFX11-NEXT:    s_and_b32 exec_lo, exec_lo, s14
 ; GFX11-NEXT:    image_sample v[0:4], v5, s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_1D lwe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -1614,8 +1632,7 @@
 ; GFX11:       ; %bb.0: ; %main_body
 ; GFX11-NEXT:    v_mov_b32_e32 v11, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v11
-; GFX11-NEXT:    v_mov_b32_e32 v9, v11
-; GFX11-NEXT:    v_mov_b32_e32 v10, v12
+; GFX11-NEXT:    v_dual_mov_b32 v9, v11 :: v_dual_mov_b32 v10, v12
 ; GFX11-NEXT:    image_sample_c_d_o v[9:10], v[0:15], s[0:7], s[8:11] dmask:0x4 dim:SQ_RSRC_IMG_2D_ARRAY tfe
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-NEXT:    v_mov_b32_e32 v0, v9
@@ -1678,17 +1695,28 @@
 ; GFX6789-NEXT:    v_mov_b32_e32 v2, v11
 ; GFX6789-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: sample_c_d_o_2darray_V2_tfe:
-; GFX10PLUS:       ; %bb.0: ; %main_body
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v9, 0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v10, v9
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v11, v9
-; GFX10PLUS-NEXT:    image_sample_c_d_o v[9:11], v[0:15], s[0:7], s[8:11] dmask:0x6 dim:SQ_RSRC_IMG_2D_ARRAY tfe
-; GFX10PLUS-NEXT:    s_waitcnt vmcnt(0)
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, v9
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, v10
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, v11
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: sample_c_d_o_2darray_V2_tfe:
+; GFX10:       ; %bb.0: ; %main_body
+; GFX10-NEXT:    v_mov_b32_e32 v9, 0
+; GFX10-NEXT:    v_mov_b32_e32 v10, v9
+; GFX10-NEXT:    v_mov_b32_e32 v11, v9
+; GFX10-NEXT:    image_sample_c_d_o v[9:11], v[0:15], s[0:7], s[8:11] dmask:0x6 dim:SQ_RSRC_IMG_2D_ARRAY tfe
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v0, v9
+; GFX10-NEXT:    v_mov_b32_e32 v1, v10
+; GFX10-NEXT:    v_mov_b32_e32 v2, v11
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: sample_c_d_o_2darray_V2_tfe:
+; GFX11:       ; %bb.0: ; %main_body
+; GFX11-NEXT:    v_mov_b32_e32 v9, 0
+; GFX11-NEXT:    v_mov_b32_e32 v10, v9
+; GFX11-NEXT:    v_mov_b32_e32 v11, v9
+; GFX11-NEXT:    image_sample_c_d_o v[9:11], v[0:15], s[0:7], s[8:11] dmask:0x6 dim:SQ_RSRC_IMG_2D_ARRAY tfe
+; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_mov_b32_e32 v2, v11
+; GFX11-NEXT:    v_dual_mov_b32 v0, v9 :: v_dual_mov_b32 v1, v10
+; GFX11-NEXT:    ; return to shader part epilog
 main_body:
   %v = call {<2 x float>, i32} @llvm.amdgcn.image.sample.c.d.o.2darray.v2f32i32.f32.f32(i32 6, i32 %offset, float %zcompare, float %dsdh, float %dtdh, float %dsdv, float %dtdv, float %s, float %t, float %slice, <8 x i32> %rsrc, <4 x i32> %samp, i1 0, i32 1, i32 0)
   %v.vec = extractvalue {<2 x float>, i32} %v, 0
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.encode.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.encode.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.encode.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.encode.ll
@@ -1,6 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX10 %s
-; RUN: llc -march=amdgcn -mcpu=gfx1100 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX11 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX11 %s
 
 define amdgpu_ps <4 x float> @sample_d_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, half %dsdh, half %dsdv, float %s) {
 ; GFX10-LABEL: sample_d_1d:
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.image.sample.g16.ll
@@ -1,6 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
-; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10 %s
 
 define amdgpu_ps <4 x float> @sample_d_1d(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, half %dsdh, half %dsdv, float %s) {
 ; GFX10-LABEL: sample_d_1d:
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.interp.inreg.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.interp.inreg.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.interp.inreg.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.interp.inreg.ll
@@ -42,9 +42,8 @@
 ; GCN-NEXT:    lds_param_load v2, attr2.x wait_vdst:15
 ; GCN-NEXT:    lds_param_load v3, attr3.x wait_vdst:15
 ; GCN-NEXT:    s_mov_b32 exec_lo, s3
-; GCN-NEXT:    v_mov_b32_e32 v4, s0
-; GCN-NEXT:    v_mov_b32_e32 v5, s1
-; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_3) | instid1(VALU_DEP_4)
+; GCN-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s1
+; GCN-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_3) | instid1(VALU_DEP_4)
 ; GCN-NEXT:    v_interp_p10_f32 v6, v0, v4, v0 wait_exp:3
 ; GCN-NEXT:    v_interp_p10_f32 v7, v1, v4, v1 wait_exp:2
 ; GCN-NEXT:    v_interp_p10_f32 v8, v2, v4, v2 wait_exp:1
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.intersect_ray.ll
@@ -62,19 +62,16 @@
 ;
 ; GFX11-LABEL: image_bvh_intersect_ray_a16:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-NEXT:    s_lshr_b32 s2, s7, 16
 ; GFX11-NEXT:    s_lshr_b32 s3, s5, 16
-; GFX11-NEXT:    v_mov_b32_e32 v2, s4
+; GFX11-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v7, s1
 ; GFX11-NEXT:    s_pack_ll_b32_b16 s2, s3, s2
 ; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s5, s7
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s3
 ; GFX11-NEXT:    s_pack_ll_b32_b16 s4, s6, s8
-; GFX11-NEXT:    v_mov_b32_e32 v3, s3
-; GFX11-NEXT:    v_mov_b32_e32 v4, s2
-; GFX11-NEXT:    v_mov_b32_e32 v5, s4
-; GFX11-NEXT:    v_mov_b32_e32 v6, s0
-; GFX11-NEXT:    v_mov_b32_e32 v7, s1
+; GFX11-NEXT:    v_dual_mov_b32 v4, s2 :: v_dual_mov_b32 v5, s4
 ; GFX11-NEXT:    s_mov_b32 s15, s12
 ; GFX11-NEXT:    s_mov_b32 s14, s11
 ; GFX11-NEXT:    s_mov_b32 s13, s10
@@ -137,20 +134,17 @@
 ;
 ; GFX11-LABEL: image_bvh64_intersect_ray_a16:
 ; GFX11:       ; %bb.0: ; %main_body
-; GFX11-NEXT:    v_mov_b32_e32 v0, s3
-; GFX11-NEXT:    v_mov_b32_e32 v6, s0
-; GFX11-NEXT:    s_lshr_b32 s0, s8, 16
+; GFX11-NEXT:    v_dual_mov_b32 v0, s3 :: v_dual_mov_b32 v1, s4
+; GFX11-NEXT:    v_dual_mov_b32 v2, s5 :: v_dual_mov_b32 v7, s1
 ; GFX11-NEXT:    s_lshr_b32 s3, s6, 16
-; GFX11-NEXT:    v_mov_b32_e32 v7, s1
-; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
 ; GFX11-NEXT:    s_pack_ll_b32_b16 s1, s6, s8
-; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s7, s9
-; GFX11-NEXT:    v_mov_b32_e32 v1, s4
-; GFX11-NEXT:    v_mov_b32_e32 v2, s5
-; GFX11-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-NEXT:    v_mov_b32_e32 v4, s0
-; GFX11-NEXT:    v_mov_b32_e32 v5, s3
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_4) | instid1(SALU_CYCLE_1)
+; GFX11-NEXT:    v_dual_mov_b32 v6, s0 :: v_dual_mov_b32 v3, s1
+; GFX11-NEXT:    s_lshr_b32 s0, s8, 16
 ; GFX11-NEXT:    v_mov_b32_e32 v8, s2
+; GFX11-NEXT:    s_pack_ll_b32_b16 s0, s3, s0
+; GFX11-NEXT:    s_pack_ll_b32_b16 s3, s7, s9
+; GFX11-NEXT:    v_dual_mov_b32 v4, s0 :: v_dual_mov_b32 v5, s3
 ; GFX11-NEXT:    s_mov_b32 s15, s13
 ; GFX11-NEXT:    s_mov_b32 s14, s12
 ; GFX11-NEXT:    s_mov_b32 s13, s11
@@ -226,11 +220,9 @@
 ; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x34
-; GFX11-NEXT:    v_mov_b32_e32 v4, 4.0
-; GFX11-NEXT:    v_mov_b32_e32 v5, 0x40a00000
-; GFX11-NEXT:    v_mov_b32_e32 v6, 0
-; GFX11-NEXT:    v_mov_b32_e32 v7, 1.0
+; GFX11-NEXT:    v_dual_mov_b32 v5, 0x40a00000 :: v_dual_mov_b32 v6, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v8, 2.0
+; GFX11-NEXT:    v_dual_mov_b32 v4, 4.0 :: v_dual_mov_b32 v7, 1.0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    v_add_co_u32 v0, s4, s4, v2
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
@@ -323,20 +315,19 @@
 ; GFX11-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x34
-; GFX11-NEXT:    v_mov_b32_e32 v4, 1.0
-; GFX11-NEXT:    v_mov_b32_e32 v5, 2.0
+; GFX11-NEXT:    v_dual_mov_b32 v4, 1.0 :: v_dual_mov_b32 v5, 2.0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_add_co_u32 v0, s4, s4, v2
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, s5, 0, s4
 ; GFX11-NEXT:    v_add_co_u32 v2, s4, s6, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, null, s7, 0, s4
 ; GFX11-NEXT:    flat_load_b32 v6, v[0:1]
 ; GFX11-NEXT:    flat_load_b32 v7, v[2:3]
-; GFX11-NEXT:    v_mov_b32_e32 v0, 0x46004200
 ; GFX11-NEXT:    v_mov_b32_e32 v1, 0x47004400
+; GFX11-NEXT:    v_dual_mov_b32 v0, 0x46004200 :: v_dual_mov_b32 v3, 0
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0x48004500
-; GFX11-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    image_bvh_intersect_ray v[0:3], [v6, v7, v[3:5], v[0:2]], s[0:3] a16
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
@@ -422,14 +413,10 @@
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x34
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0x41000000
-; GFX11-NEXT:    v_mov_b32_e32 v3, 0x40400000
-; GFX11-NEXT:    v_mov_b32_e32 v4, 4.0
-; GFX11-NEXT:    v_mov_b32_e32 v5, 0x40a00000
-; GFX11-NEXT:    v_mov_b32_e32 v6, 0
-; GFX11-NEXT:    v_mov_b32_e32 v7, 1.0
-; GFX11-NEXT:    v_mov_b32_e32 v8, 2.0
-; GFX11-NEXT:    v_mov_b32_e32 v9, 0xb36211c7
-; GFX11-NEXT:    v_mov_b32_e32 v10, 0x102
+; GFX11-NEXT:    v_dual_mov_b32 v3, 0x40400000 :: v_dual_mov_b32 v4, 4.0
+; GFX11-NEXT:    v_dual_mov_b32 v5, 0x40a00000 :: v_dual_mov_b32 v6, 0
+; GFX11-NEXT:    v_dual_mov_b32 v8, 2.0 :: v_dual_mov_b32 v9, 0xb36211c7
+; GFX11-NEXT:    v_dual_mov_b32 v10, 0x102 :: v_dual_mov_b32 v7, 1.0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    v_add_co_u32 v0, s4, s4, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
@@ -513,12 +500,9 @@
 ; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x24
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x34
-; GFX11-NEXT:    v_mov_b32_e32 v2, 0x48004500
-; GFX11-NEXT:    v_mov_b32_e32 v3, 0
-; GFX11-NEXT:    v_mov_b32_e32 v4, 1.0
-; GFX11-NEXT:    v_mov_b32_e32 v5, 2.0
-; GFX11-NEXT:    v_mov_b32_e32 v6, 0xb36211c6
-; GFX11-NEXT:    v_mov_b32_e32 v7, 0x102
+; GFX11-NEXT:    v_dual_mov_b32 v2, 0x48004500 :: v_dual_mov_b32 v5, 2.0
+; GFX11-NEXT:    v_dual_mov_b32 v4, 1.0 :: v_dual_mov_b32 v7, 0x102
+; GFX11-NEXT:    v_dual_mov_b32 v6, 0xb36211c6 :: v_dual_mov_b32 v3, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    v_add_co_u32 v0, s4, s4, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mov.dpp8.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mov.dpp8.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mov.dpp8.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.mov.dpp8.ll
@@ -1,7 +1,7 @@
 ; RUN: llc -global-isel=0 -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10PLUS %s
 ; RUN: llc -global-isel=1 -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10PLUS %s
-; RUN: llc -global-isel=0 -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10PLUS %s
-; RUN: llc -global-isel=1 -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10PLUS %s
+; RUN: llc -global-isel=0 -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10PLUS %s
+; RUN: llc -global-isel=1 -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10PLUS %s
 
 ; GFX10PLUS-LABEL: {{^}}dpp8_test:
 ; GFX10PLUS: v_mov_b32_e32 [[SRC:v[0-9]+]], s{{[0-9]+}}
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.permlane64.ll
@@ -11,9 +11,8 @@
 ; GFX11-NEXT:    s_clause 0x1
 ; GFX11-NEXT:    s_load_b32 s2, s[0:1], 0x2c
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_permlane64_b32 v0, v0
 ; GFX11-NEXT:    global_store_b32 v1, v0, s[0:1]
@@ -28,9 +27,8 @@
 ; GFX11-LABEL: test_i:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v0, 0x63
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT:    v_dual_mov_b32 v0, 0x63 :: v_dual_mov_b32 v1, 0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_permlane64_b32 v0, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v1, v0, s[0:1]
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.sendmsg.rtn.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.sendmsg.rtn.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.sendmsg.rtn.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.sendmsg.rtn.ll
@@ -7,9 +7,8 @@
 ; GFX11-SDAG:       ; %bb.0:
 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-SDAG-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(MSG_RTN_GET_DOORBELL)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-SDAG-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-SDAG-NEXT:    s_endpgm
@@ -18,9 +17,8 @@
 ; GFX11-GISEL:       ; %bb.0:
 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(MSG_RTN_GET_DOORBELL)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-GISEL-NEXT:    global_store_b32 v1, v0, s[0:1]
 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-GISEL-NEXT:    s_endpgm
@@ -34,9 +32,8 @@
 ; GFX11-SDAG:       ; %bb.0:
 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-SDAG-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(MSG_RTN_GET_DDID)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-SDAG-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-SDAG-NEXT:    s_endpgm
@@ -45,9 +42,8 @@
 ; GFX11-GISEL:       ; %bb.0:
 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(MSG_RTN_GET_DDID)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-GISEL-NEXT:    global_store_b32 v1, v0, s[0:1]
 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-GISEL-NEXT:    s_endpgm
@@ -63,8 +59,7 @@
 ; GFX11-NEXT:    s_sendmsg_rtn_b64 s[2:3], sendmsg(MSG_RTN_GET_TMA)
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
@@ -80,8 +75,7 @@
 ; GFX11-NEXT:    s_sendmsg_rtn_b64 s[2:3], sendmsg(MSG_RTN_GET_REALTIME)
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
@@ -95,9 +89,8 @@
 ; GFX11-SDAG:       ; %bb.0:
 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-SDAG-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(MSG_RTN_SAVE_WAVE)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-SDAG-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-SDAG-NEXT:    s_endpgm
@@ -106,9 +99,8 @@
 ; GFX11-GISEL:       ; %bb.0:
 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(MSG_RTN_SAVE_WAVE)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-GISEL-NEXT:    global_store_b32 v1, v0, s[0:1]
 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-GISEL-NEXT:    s_endpgm
@@ -124,8 +116,7 @@
 ; GFX11-NEXT:    s_sendmsg_rtn_b64 s[2:3], sendmsg(MSG_RTN_GET_TBA)
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
@@ -139,9 +130,8 @@
 ; GFX11-SDAG:       ; %bb.0:
 ; GFX11-SDAG-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-SDAG-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(0, 0, 0)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-SDAG-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-SDAG-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-SDAG-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-SDAG-NEXT:    s_endpgm
@@ -150,9 +140,8 @@
 ; GFX11-GISEL:       ; %bb.0:
 ; GFX11-GISEL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
 ; GFX11-GISEL-NEXT:    s_sendmsg_rtn_b32 s2, sendmsg(0, 0, 0)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-GISEL-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-GISEL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-GISEL-NEXT:    global_store_b32 v1, v0, s[0:1]
 ; GFX11-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-GISEL-NEXT:    s_endpgm
@@ -168,8 +157,7 @@
 ; GFX11-NEXT:    s_sendmsg_rtn_b64 s[2:3], 99999
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.mulo.ll b/llvm/test/CodeGen/AMDGPU/llvm.mulo.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.mulo.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.mulo.ll
@@ -81,9 +81,8 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v4, v0 :: v_dual_mov_b32 v5, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v4, v2, 0
 ; GFX11-NEXT:    v_mad_u64_u32 v[6:7], null, v4, v3, 0
 ; GFX11-NEXT:    v_mad_u64_u32 v[9:10], null, v5, v2, 0
@@ -226,9 +225,8 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v0
-; GFX11-NEXT:    v_mov_b32_e32 v5, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v4, v0 :: v_dual_mov_b32 v5, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v4, v2, 0
 ; GFX11-NEXT:    v_mad_u64_u32 v[6:7], null, v4, v3, 0
 ; GFX11-NEXT:    v_mad_u64_u32 v[9:10], null, v5, v2, 0
@@ -251,19 +249,18 @@
 ; GFX11-NEXT:    v_subrev_co_ci_u32_e32 v10, vcc_lo, 0, v7, vcc_lo
 ; GFX11-NEXT:    v_cmp_gt_i32_e32 vcc_lo, 0, v5
 ; GFX11-NEXT:    v_add3_u32 v1, v1, v9, v8
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT:    v_cndmask_b32_e32 v6, v6, v2, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v5, v7, v10, vcc_lo
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_cndmask_b32_e32 v6, v6, v2, vcc_lo
 ; GFX11-NEXT:    v_ashrrev_i32_e32 v2, 31, v1
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_sub_co_u32 v4, vcc_lo, v6, v4
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_4)
 ; GFX11-NEXT:    v_subrev_co_ci_u32_e32 v7, vcc_lo, 0, v5, vcc_lo
 ; GFX11-NEXT:    v_cmp_gt_i32_e32 vcc_lo, 0, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_mov_b32_e32 v3, v2
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_cndmask_b32_e32 v5, v5, v7, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v4, v6, v4, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v5, v5, v7 :: v_dual_cndmask_b32 v4, v6, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_cmp_ne_u64_e32 vcc_lo, v[4:5], v[2:3]
 ; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc_lo
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
@@ -558,23 +555,21 @@
 ; GFX11-NEXT:    s_addc_u32 s6, 0, s6
 ; GFX11-NEXT:    s_sub_u32 s9, s4, s2
 ; GFX11-NEXT:    s_subb_u32 s10, s6, 0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s9
+; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_3) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_mov_b32 v1, s9 :: v_dual_mov_b32 v0, s10
 ; GFX11-NEXT:    s_cmp_lt_i32 s1, 0
-; GFX11-NEXT:    v_mov_b32_e32 v0, s10
 ; GFX11-NEXT:    s_cselect_b32 vcc_lo, -1, 0
 ; GFX11-NEXT:    s_cmp_lt_i32 s3, 0
 ; GFX11-NEXT:    v_cndmask_b32_e32 v2, s4, v1, vcc_lo
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cndmask_b32_e32 v0, s6, v0, vcc_lo
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_sub_co_u32 v3, vcc_lo, v2, s0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-NEXT:    v_subrev_co_ci_u32_e32 v1, vcc_lo, 0, v0, vcc_lo
 ; GFX11-NEXT:    s_cselect_b32 vcc_lo, -1, 0
 ; GFX11-NEXT:    s_add_i32 s1, s8, s7
 ; GFX11-NEXT:    s_mul_i32 s0, s0, s2
 ; GFX11-NEXT:    s_add_i32 s1, s1, s5
-; GFX11-NEXT:    v_cndmask_b32_e32 v1, v0, v1, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v0, v1 :: v_dual_cndmask_b32 v0, v2, v3
 ; GFX11-NEXT:    s_ashr_i32 s4, s1, 31
 ; GFX11-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
 ; GFX11-NEXT:    s_mov_b32 s5, s4
@@ -641,9 +636,8 @@
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_ashrrev_i64 v[5:6], 2, v[4:5]
 ; GFX11-NEXT:    v_cmp_ne_u64_e32 vcc_lo, v[5:6], v[0:1]
-; GFX11-NEXT:    v_mov_b32_e32 v0, v4
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT:    v_mov_b32_e32 v1, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v3
 ; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc_lo
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
 bb:
@@ -696,15 +690,12 @@
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_and_b32_e32 v7, 0x3fffffff, v1
-; GFX11-NEXT:    v_mov_b32_e32 v6, v0
+; GFX11-NEXT:    v_dual_mov_b32 v6, v0 :: v_dual_and_b32 v7, 0x3fffffff, v1
 ; GFX11-NEXT:    v_lshlrev_b64 v[4:5], 2, v[0:1]
 ; GFX11-NEXT:    v_alignbit_b32 v3, v1, v0, 30
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_cmp_ne_u64_e32 vcc_lo, v[6:7], v[0:1]
-; GFX11-NEXT:    v_mov_b32_e32 v0, v4
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT:    v_mov_b32_e32 v1, v3
+; GFX11-NEXT:    v_dual_mov_b32 v0, v4 :: v_dual_mov_b32 v1, v3
 ; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, vcc_lo
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
 bb:
diff --git a/llvm/test/CodeGen/AMDGPU/mad_64_32.ll b/llvm/test/CodeGen/AMDGPU/mad_64_32.ll
--- a/llvm/test/CodeGen/AMDGPU/mad_64_32.ll
+++ b/llvm/test/CodeGen/AMDGPU/mad_64_32.ll
@@ -32,8 +32,7 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_i64_i32 v[0:1], null, v5, v4, v[2:3]
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
@@ -70,8 +69,7 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_i64_i32 v[0:1], null, v5, v4, v[2:3]
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
@@ -108,8 +106,7 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v5, v4, v[2:3]
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
@@ -146,8 +143,7 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v5, v4, v[2:3]
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
@@ -254,8 +250,7 @@
 ; GFX11-NEXT:    v_ashrrev_i32_e32 v15, 31, v1
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_u64_u32 v[9:10], null, v14, v1, v[7:8]
-; GFX11-NEXT:    v_mov_b32_e32 v7, v10
-; GFX11-NEXT:    v_mov_b32_e32 v10, v8
+; GFX11-NEXT:    v_dual_mov_b32 v7, v10 :: v_dual_mov_b32 v10, v8
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_mad_u64_u32 v[11:12], null, v0, v15, v[9:10]
 ; GFX11-NEXT:    v_mad_i64_i32 v[9:10], null, v1, v14, 0
@@ -310,8 +305,7 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_i64_i32 v[0:1], null, v5, v4, v[2:3]
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
@@ -405,8 +399,7 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v4, v1
-; GFX11-NEXT:    v_mov_b32_e32 v5, v0
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_mov_b32 v5, v0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
 ; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v5, v4, v[2:3]
 ; GFX11-NEXT:    v_ashrrev_i32_e32 v5, 31, v5
@@ -494,14 +487,12 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v3, v2
-; GFX11-NEXT:    v_mov_b32_e32 v2, v0
+; GFX11-NEXT:    v_dual_mov_b32 v3, v2 :: v_dual_mov_b32 v2, v0
 ; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v2, v3, v[4:5]
-; GFX11-NEXT:    v_and_b32_e32 v5, 1, v6
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
-; GFX11-NEXT:    v_mov_b32_e32 v4, v1
+; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v2, v3, v[4:5]
+; GFX11-NEXT:    v_dual_mov_b32 v4, v1 :: v_dual_and_b32 v5, 1, v6
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_u64_u32 v[1:2], null, v5, v3, v[4:5]
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %trunc.lhs = and i64 %arg0, 8589934591
@@ -550,10 +541,9 @@
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    v_mov_b32_e32 v6, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v6, v2, v[4:5]
-; GFX11-NEXT:    v_and_b32_e32 v4, 1, v3
-; GFX11-NEXT:    v_mov_b32_e32 v3, v1
+; GFX11-NEXT:    v_dual_mov_b32 v3, v1 :: v_dual_and_b32 v4, 1, v3
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
 ; GFX11-NEXT:    v_mad_u64_u32 v[1:2], null, v6, v4, v[3:4]
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
@@ -631,9 +621,8 @@
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    v_mad_u64_u32 v[2:3], null, v1, v0, v[0:1]
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
-; GFX11-NEXT:    v_mov_b32_e32 v0, v2
-; GFX11-NEXT:    v_mov_b32_e32 v1, v3
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_mov_b32 v0, v2 :: v_dual_mov_b32 v1, v3
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %tmp4 = lshr i64 %arg0, 32
   %tmp5 = and i64 %arg0, 4294967295
@@ -697,14 +686,13 @@
 ; GFX11-NEXT:    s_load_b64 s[2:3], s[0:1], 0x2c
 ; GFX11-NEXT:    s_load_b64 s[4:5], s[0:1], 0x34
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    s_mul_i32 s6, s2, s3
 ; GFX11-NEXT:    s_mul_hi_u32 s3, s2, s3
 ; GFX11-NEXT:    s_add_u32 s2, s6, s4
 ; GFX11-NEXT:    s_addc_u32 s3, s3, s5
 ; GFX11-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_mov_b32 v1, s3
 ; GFX11-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
@@ -935,9 +923,8 @@
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
-; GFX11-NEXT:    v_mov_b32_e32 v6, v1
-; GFX11-NEXT:    v_mov_b32_e32 v7, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_4)
+; GFX11-NEXT:    v_dual_mov_b32 v6, v1 :: v_dual_mov_b32 v7, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_3)
 ; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v7, v2, v[4:5]
 ; GFX11-NEXT:    v_mul_lo_u32 v3, v7, v3
 ; GFX11-NEXT:    v_mul_lo_u32 v2, v6, v2
diff --git a/llvm/test/CodeGen/AMDGPU/mad_u64_u32.ll b/llvm/test/CodeGen/AMDGPU/mad_u64_u32.ll
--- a/llvm/test/CodeGen/AMDGPU/mad_u64_u32.ll
+++ b/llvm/test/CodeGen/AMDGPU/mad_u64_u32.ll
@@ -15,13 +15,6 @@
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    v_mad_u64_u32 v[0:1], null, v0, v1, v[2:3]
 ; GFX10-NEXT:    ; return to shader part epilog
-;
-; GFX11-LABEL: mad_i32_vvv:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v3, v1
-; GFX11-NEXT:    v_mov_b32_e32 v4, v0
-; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v4, v3, v[2:3]
-; GFX11-NEXT:    ; return to shader part epilog
   %mul = mul i32 %a, %b
   %add = add i32 %mul, %c
   %cast = bitcast i32 %add to float
@@ -51,13 +44,6 @@
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    v_mad_u64_u32 v[0:1], null, v0, v1, 42
 ; GFX10-NEXT:    ; return to shader part epilog
-;
-; GFX11-LABEL: mad_i32_vvc:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v2, v1
-; GFX11-NEXT:    v_mov_b32_e32 v3, v0
-; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v3, v2, 42
-; GFX11-NEXT:    ; return to shader part epilog
   %mul = mul i32 %a, %b
   %add = add i32 %mul, 42
   %cast = bitcast i32 %add to float
@@ -76,13 +62,6 @@
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    v_mad_u64_u32 v[0:1], null, v0, v1, 0x12d687
 ; GFX10-NEXT:    ; return to shader part epilog
-;
-; GFX11-LABEL: mad_i32_vvi:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v2, v1
-; GFX11-NEXT:    v_mov_b32_e32 v3, v0
-; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v3, v2, 0x12d687
-; GFX11-NEXT:    ; return to shader part epilog
   %mul = mul i32 %a, %b
   %add = add i32 %mul, 1234567
   %cast = bitcast i32 %add to float
@@ -143,13 +122,6 @@
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    v_mad_u64_u32 v[0:1], null, v0, v1, s[0:1]
 ; GFX10-NEXT:    ; return to shader part epilog
-;
-; GFX11-LABEL: mad_i32_vvs:
-; GFX11:       ; %bb.0:
-; GFX11-NEXT:    v_mov_b32_e32 v2, v1
-; GFX11-NEXT:    v_mov_b32_e32 v3, v0
-; GFX11-NEXT:    v_mad_u64_u32 v[0:1], null, v3, v2, s[0:1]
-; GFX11-NEXT:    ; return to shader part epilog
   %mul = mul i32 %a, %b
   %add = add i32 %mul, %c
   %cast = bitcast i32 %add to float
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-agent.ll
@@ -119,26 +119,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -256,26 +254,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -402,30 +398,28 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -562,34 +556,32 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -691,10 +683,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_unordered_store:
@@ -703,10 +695,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -807,10 +799,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_monotonic_store:
@@ -819,10 +811,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -935,12 +927,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_release_store:
@@ -949,12 +941,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1067,12 +1059,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_seq_cst_store:
@@ -1081,12 +1073,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1187,10 +1179,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_monotonic_atomicrmw:
@@ -1199,10 +1191,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1322,8 +1314,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1338,8 +1329,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1458,12 +1448,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_release_atomicrmw:
@@ -1472,12 +1462,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1609,8 +1599,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1627,8 +1616,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1768,8 +1756,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1786,8 +1773,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1921,14 +1907,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acquire_ret_atomicrmw:
@@ -1937,14 +1923,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2083,8 +2069,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2093,6 +2078,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acq_rel_ret_atomicrmw:
@@ -2101,8 +2087,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2111,6 +2096,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2249,8 +2235,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2259,6 +2244,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_seq_cst_ret_atomicrmw:
@@ -2267,8 +2253,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2277,6 +2262,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2390,11 +2376,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_monotonic_monotonic_cmpxchg:
@@ -2403,11 +2388,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2540,10 +2524,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2557,10 +2539,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2691,13 +2671,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_release_monotonic_cmpxchg:
@@ -2706,13 +2685,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2857,10 +2835,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -2876,10 +2852,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3031,10 +3005,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3050,10 +3022,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3193,10 +3163,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3210,10 +3178,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3351,10 +3317,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3368,10 +3332,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3521,10 +3483,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3540,10 +3500,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3695,10 +3653,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3714,10 +3670,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3869,10 +3823,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3888,10 +3840,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4043,10 +3993,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4062,10 +4010,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4217,10 +4163,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4236,10 +4180,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4391,10 +4333,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4410,10 +4350,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4565,10 +4503,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4584,10 +4520,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4739,10 +4673,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4758,10 +4690,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4906,13 +4836,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_monotonic_monotonic_ret_cmpxchg:
@@ -4921,13 +4850,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5076,15 +5004,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acquire_monotonic_ret_cmpxchg:
@@ -5093,15 +5020,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5253,15 +5179,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_release_monotonic_ret_cmpxchg:
@@ -5270,15 +5195,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5439,10 +5363,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5450,6 +5372,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acq_rel_monotonic_ret_cmpxchg:
@@ -5458,10 +5381,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5469,6 +5390,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5629,10 +5551,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5640,6 +5560,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_seq_cst_monotonic_ret_cmpxchg:
@@ -5648,10 +5569,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5659,6 +5578,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5807,15 +5727,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_monotonic_acquire_ret_cmpxchg:
@@ -5824,15 +5743,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5981,15 +5899,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acquire_acquire_ret_cmpxchg:
@@ -5998,15 +5915,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6167,10 +6083,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6178,6 +6092,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_release_acquire_ret_cmpxchg:
@@ -6186,10 +6101,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6197,6 +6110,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6357,10 +6271,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6368,6 +6280,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acq_rel_acquire_ret_cmpxchg:
@@ -6376,10 +6289,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6387,6 +6298,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6547,10 +6459,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6558,6 +6468,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_seq_cst_acquire_ret_cmpxchg:
@@ -6566,10 +6477,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6577,6 +6486,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6737,10 +6647,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6748,6 +6656,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_monotonic_seq_cst_ret_cmpxchg:
@@ -6756,10 +6665,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6767,6 +6674,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6927,10 +6835,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6938,6 +6844,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acquire_seq_cst_ret_cmpxchg:
@@ -6946,10 +6853,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6957,6 +6862,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7117,10 +7023,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7128,6 +7032,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_release_seq_cst_ret_cmpxchg:
@@ -7136,10 +7041,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7147,6 +7050,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7307,10 +7211,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7318,6 +7220,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_acq_rel_seq_cst_ret_cmpxchg:
@@ -7326,10 +7229,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7337,6 +7238,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7497,10 +7399,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7508,6 +7408,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_seq_cst_seq_cst_ret_cmpxchg:
@@ -7516,10 +7417,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7527,6 +7426,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7646,26 +7546,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -7783,26 +7681,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -7935,32 +7831,30 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -8103,36 +7997,34 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -8234,10 +8126,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_unordered_store:
@@ -8246,10 +8138,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8350,10 +8242,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_monotonic_store:
@@ -8362,10 +8254,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8478,12 +8370,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_release_store:
@@ -8492,12 +8384,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8610,12 +8502,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_seq_cst_store:
@@ -8624,12 +8516,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8730,10 +8622,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_monotonic_atomicrmw:
@@ -8742,10 +8634,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8863,8 +8755,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -8878,8 +8769,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -8997,12 +8887,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_release_atomicrmw:
@@ -9011,12 +8901,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -9146,8 +9036,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9163,8 +9052,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9301,8 +9189,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9318,8 +9205,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9457,8 +9343,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
@@ -9466,6 +9351,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acquire_ret_atomicrmw:
@@ -9474,8 +9360,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
@@ -9483,6 +9368,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -9626,8 +9512,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9637,6 +9522,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acq_rel_ret_atomicrmw:
@@ -9645,8 +9531,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9656,6 +9541,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -9799,8 +9685,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9810,6 +9695,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_seq_cst_ret_atomicrmw:
@@ -9818,8 +9704,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9829,6 +9714,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -9942,11 +9828,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_monotonic_monotonic_cmpxchg:
@@ -9955,11 +9840,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10090,10 +9974,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10106,10 +9988,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -10239,13 +10119,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_release_monotonic_cmpxchg:
@@ -10254,13 +10133,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10403,10 +10281,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10421,10 +10297,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10573,10 +10447,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10591,10 +10463,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10731,10 +10601,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10747,10 +10615,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -10885,10 +10751,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10901,10 +10765,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -11051,10 +10913,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11069,10 +10929,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11221,10 +11079,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11239,10 +11095,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11391,10 +11245,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11409,10 +11261,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11561,10 +11411,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11579,10 +11427,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11731,10 +11577,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11749,10 +11593,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11901,10 +11743,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11919,10 +11759,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12071,10 +11909,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12089,10 +11925,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12241,10 +12075,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12259,10 +12091,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12406,13 +12236,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -12421,13 +12250,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12582,16 +12410,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acquire_monotonic_ret_cmpxchg:
@@ -12600,16 +12427,15 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12761,15 +12587,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_release_monotonic_ret_cmpxchg:
@@ -12778,15 +12603,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12953,10 +12777,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12965,6 +12787,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -12973,10 +12796,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12985,6 +12806,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13151,10 +12973,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13163,6 +12983,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -13171,10 +12992,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13183,6 +13002,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13337,16 +13157,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_monotonic_acquire_ret_cmpxchg:
@@ -13355,16 +13174,15 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13519,16 +13337,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acquire_acquire_ret_cmpxchg:
@@ -13537,16 +13354,15 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13713,10 +13529,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13725,6 +13539,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_release_acquire_ret_cmpxchg:
@@ -13733,10 +13548,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13745,6 +13558,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13911,10 +13725,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13923,6 +13735,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -13931,10 +13744,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13943,6 +13754,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14109,10 +13921,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14121,6 +13931,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -14129,10 +13940,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14141,6 +13950,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14307,10 +14117,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14319,6 +14127,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -14327,10 +14136,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14339,6 +14146,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14505,10 +14313,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14517,6 +14323,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -14525,10 +14332,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14537,6 +14342,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14703,10 +14509,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14715,6 +14519,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_release_seq_cst_ret_cmpxchg:
@@ -14723,10 +14528,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14735,6 +14538,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14901,10 +14705,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14913,6 +14715,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -14921,10 +14724,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14933,6 +14734,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -15099,10 +14901,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -15111,6 +14911,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -15119,10 +14920,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -15131,6 +14930,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll
@@ -119,26 +119,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] slc dlc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_load_0:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] slc dlc
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -273,10 +271,10 @@
 ; GFX11-WGP-NEXT:    v_add_co_u32 v0, s0, s0, v0
 ; GFX11-WGP-NEXT:    v_add_co_ci_u32_e64 v1, null, s1, 0, s0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] slc dlc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v0, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_load_1:
@@ -288,10 +286,10 @@
 ; GFX11-CU-NEXT:    v_add_co_u32 v0, s0, s0, v0
 ; GFX11-CU-NEXT:    v_add_co_ci_u32_e64 v1, null, s1, 0, s0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] slc dlc
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v0, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -411,26 +409,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2 glc slc dlc
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_store_0:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2 glc slc dlc
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -559,31 +555,31 @@
 ; GFX11-WGP-LABEL: flat_nontemporal_store_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
-; GFX11-WGP-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-WGP-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-WGP-NEXT:    v_add_co_u32 v0, s0, s2, v0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[1:2]
 ; GFX11-WGP-NEXT:    v_add_co_ci_u32_e64 v1, null, s3, 0, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2 glc slc dlc
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_store_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
-; GFX11-CU-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-CU-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-CU-NEXT:    v_add_co_u32 v0, s0, s2, v0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[1:2]
 ; GFX11-CU-NEXT:    v_add_co_ci_u32_e64 v1, null, s3, 0, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2 glc slc dlc
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-singlethread.ll
@@ -119,26 +119,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -256,26 +254,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -393,26 +389,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -530,26 +524,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -651,10 +643,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_unordered_store:
@@ -663,10 +655,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -767,10 +759,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_store:
@@ -779,10 +771,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -883,10 +875,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_store:
@@ -895,10 +887,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -999,10 +991,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_store:
@@ -1011,10 +1003,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1115,10 +1107,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_atomicrmw:
@@ -1127,10 +1119,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1231,10 +1223,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_atomicrmw:
@@ -1243,10 +1235,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1347,10 +1339,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_atomicrmw:
@@ -1359,10 +1351,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1463,10 +1455,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_atomicrmw:
@@ -1475,10 +1467,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1579,10 +1571,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_atomicrmw:
@@ -1591,10 +1583,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1711,12 +1703,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_ret_atomicrmw:
@@ -1725,12 +1717,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1848,12 +1840,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_ret_atomicrmw:
@@ -1862,12 +1854,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1985,12 +1977,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_ret_atomicrmw:
@@ -1999,12 +1991,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2118,11 +2110,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_monotonic_cmpxchg:
@@ -2131,11 +2122,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2249,11 +2239,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_monotonic_cmpxchg:
@@ -2262,11 +2251,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2380,11 +2368,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_monotonic_cmpxchg:
@@ -2393,11 +2380,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2511,11 +2497,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_monotonic_cmpxchg:
@@ -2524,11 +2509,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2642,11 +2626,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_monotonic_cmpxchg:
@@ -2655,11 +2638,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2773,11 +2755,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_acquire_cmpxchg:
@@ -2786,11 +2767,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2904,11 +2884,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_acquire_cmpxchg:
@@ -2917,11 +2896,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3035,11 +3013,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_acquire_cmpxchg:
@@ -3048,11 +3025,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3166,11 +3142,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_acquire_cmpxchg:
@@ -3179,11 +3154,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3297,11 +3271,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_acquire_cmpxchg:
@@ -3310,11 +3283,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3428,11 +3400,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_seq_cst_cmpxchg:
@@ -3441,11 +3412,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3559,11 +3529,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_seq_cst_cmpxchg:
@@ -3572,11 +3541,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3690,11 +3658,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_seq_cst_cmpxchg:
@@ -3703,11 +3670,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3821,11 +3787,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_seq_cst_cmpxchg:
@@ -3834,11 +3799,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3952,11 +3916,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_seq_cst_cmpxchg:
@@ -3965,11 +3928,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4107,13 +4069,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_monotonic_ret_cmpxchg:
@@ -4122,13 +4083,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4268,13 +4228,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_monotonic_ret_cmpxchg:
@@ -4283,13 +4242,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4429,13 +4387,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_monotonic_ret_cmpxchg:
@@ -4444,13 +4401,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4590,13 +4546,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_monotonic_ret_cmpxchg:
@@ -4605,13 +4560,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4751,13 +4705,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_monotonic_ret_cmpxchg:
@@ -4766,13 +4719,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4912,13 +4864,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_acquire_ret_cmpxchg:
@@ -4927,13 +4878,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5073,13 +5023,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_acquire_ret_cmpxchg:
@@ -5088,13 +5037,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5234,13 +5182,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_acquire_ret_cmpxchg:
@@ -5249,13 +5196,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5395,13 +5341,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_acquire_ret_cmpxchg:
@@ -5410,13 +5355,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5556,13 +5500,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_acquire_ret_cmpxchg:
@@ -5571,13 +5514,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5717,13 +5659,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_monotonic_seq_cst_ret_cmpxchg:
@@ -5732,13 +5673,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5878,13 +5818,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acquire_seq_cst_ret_cmpxchg:
@@ -5893,13 +5832,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6039,13 +5977,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_release_seq_cst_ret_cmpxchg:
@@ -6054,13 +5991,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6200,13 +6136,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_acq_rel_seq_cst_ret_cmpxchg:
@@ -6215,13 +6150,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6361,13 +6295,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_seq_cst_seq_cst_ret_cmpxchg:
@@ -6376,13 +6309,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6502,26 +6434,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6639,26 +6569,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6776,26 +6704,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6913,26 +6839,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -7034,10 +6958,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_unordered_store:
@@ -7046,10 +6970,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7150,10 +7074,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_store:
@@ -7162,10 +7086,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7266,10 +7190,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_store:
@@ -7278,10 +7202,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7382,10 +7306,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_store:
@@ -7394,10 +7318,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7498,10 +7422,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_atomicrmw:
@@ -7510,10 +7434,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7614,10 +7538,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_atomicrmw:
@@ -7626,10 +7550,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7730,10 +7654,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_atomicrmw:
@@ -7742,10 +7666,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7846,10 +7770,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_atomicrmw:
@@ -7858,10 +7782,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7962,10 +7886,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_atomicrmw:
@@ -7974,10 +7898,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8094,12 +8018,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_ret_atomicrmw:
@@ -8108,12 +8032,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8231,12 +8155,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_ret_atomicrmw:
@@ -8245,12 +8169,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8368,12 +8292,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_ret_atomicrmw:
@@ -8382,12 +8306,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8501,11 +8425,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_monotonic_cmpxchg:
@@ -8514,11 +8437,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8632,11 +8554,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_monotonic_cmpxchg:
@@ -8645,11 +8566,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8763,11 +8683,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_monotonic_cmpxchg:
@@ -8776,11 +8695,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8894,11 +8812,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_monotonic_cmpxchg:
@@ -8907,11 +8824,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9025,11 +8941,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_monotonic_cmpxchg:
@@ -9038,11 +8953,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9156,11 +9070,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_acquire_cmpxchg:
@@ -9169,11 +9082,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9287,11 +9199,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_acquire_cmpxchg:
@@ -9300,11 +9211,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9418,11 +9328,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_acquire_cmpxchg:
@@ -9431,11 +9340,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9549,11 +9457,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_acquire_cmpxchg:
@@ -9562,11 +9469,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9680,11 +9586,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_acquire_cmpxchg:
@@ -9693,11 +9598,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9811,11 +9715,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_seq_cst_cmpxchg:
@@ -9824,11 +9727,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9942,11 +9844,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_seq_cst_cmpxchg:
@@ -9955,11 +9856,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10073,11 +9973,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_seq_cst_cmpxchg:
@@ -10086,11 +9985,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10204,11 +10102,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
@@ -10217,11 +10114,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10335,11 +10231,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
@@ -10348,11 +10243,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10490,13 +10384,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -10505,13 +10398,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10651,13 +10543,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
@@ -10666,13 +10557,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10812,13 +10702,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_monotonic_ret_cmpxchg:
@@ -10827,13 +10716,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10973,13 +10861,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -10988,13 +10875,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11134,13 +11020,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -11149,13 +11034,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11295,13 +11179,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
@@ -11310,13 +11193,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11456,13 +11338,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_acquire_ret_cmpxchg:
@@ -11471,13 +11352,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11617,13 +11497,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_acquire_ret_cmpxchg:
@@ -11632,13 +11511,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11778,13 +11656,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -11793,13 +11670,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11939,13 +11815,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -11954,13 +11829,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12100,13 +11974,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -12115,13 +11988,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12261,13 +12133,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -12276,13 +12147,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12422,13 +12292,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_release_seq_cst_ret_cmpxchg:
@@ -12437,13 +12306,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12583,13 +12451,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -12598,13 +12465,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12744,13 +12610,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -12759,13 +12624,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-system.ll
@@ -119,26 +119,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -256,26 +254,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -404,30 +400,28 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -566,34 +560,32 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -695,10 +687,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_unordered_store:
@@ -707,10 +699,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -811,10 +803,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_monotonic_store:
@@ -823,10 +815,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -941,12 +933,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_release_store:
@@ -955,12 +947,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1075,12 +1067,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_seq_cst_store:
@@ -1089,12 +1081,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1195,10 +1187,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_monotonic_atomicrmw:
@@ -1207,10 +1199,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1332,8 +1324,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1348,8 +1339,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1470,12 +1460,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_release_atomicrmw:
@@ -1484,12 +1474,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1625,8 +1615,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1643,8 +1632,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1788,8 +1776,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1806,8 +1793,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1943,14 +1929,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acquire_ret_atomicrmw:
@@ -1959,14 +1945,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2109,8 +2095,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2119,6 +2104,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acq_rel_ret_atomicrmw:
@@ -2127,8 +2113,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2137,6 +2122,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2279,8 +2265,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2289,6 +2274,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_seq_cst_ret_atomicrmw:
@@ -2297,8 +2283,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2307,6 +2292,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2420,11 +2406,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_monotonic_monotonic_cmpxchg:
@@ -2433,11 +2418,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2572,10 +2556,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2589,10 +2571,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2725,13 +2705,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_release_monotonic_cmpxchg:
@@ -2740,13 +2719,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2895,10 +2873,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -2914,10 +2890,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3073,10 +3047,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3092,10 +3064,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3237,10 +3207,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3254,10 +3222,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3397,10 +3363,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3414,10 +3378,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3571,10 +3533,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3590,10 +3550,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3749,10 +3707,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3768,10 +3724,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3927,10 +3881,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3946,10 +3898,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4105,10 +4055,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4124,10 +4072,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4283,10 +4229,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4302,10 +4246,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4461,10 +4403,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4480,10 +4420,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4639,10 +4577,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4658,10 +4594,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4817,10 +4751,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4836,10 +4768,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -4984,13 +4914,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_monotonic_monotonic_ret_cmpxchg:
@@ -4999,13 +4928,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5156,15 +5084,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acquire_monotonic_ret_cmpxchg:
@@ -5173,15 +5100,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5335,15 +5261,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_release_monotonic_ret_cmpxchg:
@@ -5352,15 +5277,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5525,10 +5449,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5536,6 +5458,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acq_rel_monotonic_ret_cmpxchg:
@@ -5544,10 +5467,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5555,6 +5476,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5719,10 +5641,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5730,6 +5650,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_seq_cst_monotonic_ret_cmpxchg:
@@ -5738,10 +5659,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -5749,6 +5668,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5899,15 +5819,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_monotonic_acquire_ret_cmpxchg:
@@ -5916,15 +5835,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6075,15 +5993,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acquire_acquire_ret_cmpxchg:
@@ -6092,15 +6009,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6265,10 +6181,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6276,6 +6190,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_release_acquire_ret_cmpxchg:
@@ -6284,10 +6199,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6295,6 +6208,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6459,10 +6373,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6470,6 +6382,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acq_rel_acquire_ret_cmpxchg:
@@ -6478,10 +6391,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6489,6 +6400,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6653,10 +6565,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6664,6 +6574,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_seq_cst_acquire_ret_cmpxchg:
@@ -6672,10 +6583,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6683,6 +6592,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6847,10 +6757,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6858,6 +6766,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_monotonic_seq_cst_ret_cmpxchg:
@@ -6866,10 +6775,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -6877,6 +6784,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7041,10 +6949,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7052,6 +6958,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acquire_seq_cst_ret_cmpxchg:
@@ -7060,10 +6967,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7071,6 +6976,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7235,10 +7141,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7246,6 +7150,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_release_seq_cst_ret_cmpxchg:
@@ -7254,10 +7159,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7265,6 +7168,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7429,10 +7333,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7440,6 +7342,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_acq_rel_seq_cst_ret_cmpxchg:
@@ -7448,10 +7351,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7459,6 +7360,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7623,10 +7525,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7634,6 +7534,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_seq_cst_seq_cst_ret_cmpxchg:
@@ -7642,10 +7543,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -7653,6 +7552,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -7772,26 +7672,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -7909,26 +7807,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -8063,32 +7959,30 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -8233,36 +8127,34 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -8364,10 +8256,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_unordered_store:
@@ -8376,10 +8268,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8480,10 +8372,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_monotonic_store:
@@ -8492,10 +8384,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8610,12 +8502,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_release_store:
@@ -8624,12 +8516,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8744,12 +8636,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_seq_cst_store:
@@ -8758,12 +8650,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -8864,10 +8756,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_monotonic_atomicrmw:
@@ -8876,10 +8768,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8999,8 +8891,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9014,8 +8905,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9135,12 +9025,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_release_atomicrmw:
@@ -9149,12 +9039,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -9288,8 +9178,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9305,8 +9194,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9447,8 +9335,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9464,8 +9351,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9605,8 +9491,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
@@ -9614,6 +9499,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acquire_ret_atomicrmw:
@@ -9622,8 +9508,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
@@ -9631,6 +9516,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -9778,8 +9664,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9789,6 +9674,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acq_rel_ret_atomicrmw:
@@ -9797,8 +9683,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9808,6 +9693,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -9955,8 +9841,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9966,6 +9851,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_seq_cst_ret_atomicrmw:
@@ -9974,8 +9860,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -9985,6 +9870,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -10098,11 +9984,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_monotonic_monotonic_cmpxchg:
@@ -10111,11 +9996,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10248,10 +10132,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10264,10 +10146,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -10399,13 +10279,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_release_monotonic_cmpxchg:
@@ -10414,13 +10293,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10567,10 +10445,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10585,10 +10461,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10741,10 +10615,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10759,10 +10631,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10901,10 +10771,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10917,10 +10785,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -11057,10 +10923,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -11073,10 +10937,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -11227,10 +11089,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11245,10 +11105,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11401,10 +11259,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11419,10 +11275,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11575,10 +11429,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11593,10 +11445,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11749,10 +11599,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11767,10 +11615,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11923,10 +11769,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -11941,10 +11785,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12097,10 +11939,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12115,10 +11955,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12271,10 +12109,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12289,10 +12125,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12445,10 +12279,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12463,10 +12295,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -12610,13 +12440,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -12625,13 +12454,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12788,16 +12616,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acquire_monotonic_ret_cmpxchg:
@@ -12806,16 +12633,15 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12969,15 +12795,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_release_monotonic_ret_cmpxchg:
@@ -12986,15 +12811,14 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13165,10 +12989,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13177,6 +12999,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -13185,10 +13008,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13197,6 +13018,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13367,10 +13189,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13379,6 +13199,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -13387,10 +13208,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13399,6 +13218,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13555,16 +13375,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_monotonic_acquire_ret_cmpxchg:
@@ -13573,16 +13392,15 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13739,16 +13557,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acquire_acquire_ret_cmpxchg:
@@ -13757,16 +13574,15 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13937,10 +13753,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13949,6 +13763,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_release_acquire_ret_cmpxchg:
@@ -13957,10 +13772,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13969,6 +13782,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14139,10 +13953,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14151,6 +13963,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -14159,10 +13972,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14171,6 +13982,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14341,10 +14153,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14353,6 +14163,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -14361,10 +14172,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14373,6 +14182,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14543,10 +14353,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14555,6 +14363,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -14563,10 +14372,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14575,6 +14382,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14745,10 +14553,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14757,6 +14563,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -14765,10 +14572,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14777,6 +14582,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -14947,10 +14753,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14959,6 +14763,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_release_seq_cst_ret_cmpxchg:
@@ -14967,10 +14772,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -14979,6 +14782,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -15149,10 +14953,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -15161,6 +14963,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -15169,10 +14972,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -15181,6 +14982,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -15351,10 +15153,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -15363,6 +15163,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -15371,10 +15172,8 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -15383,6 +15182,7 @@
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-volatile.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-volatile.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-volatile.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-volatile.ll
@@ -67,28 +67,26 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_load_0:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -170,10 +168,10 @@
 ; GFX11-WGP-NEXT:    v_add_co_ci_u32_e64 v1, null, s1, 0, s0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v0, s2
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_load_1:
@@ -186,10 +184,10 @@
 ; GFX11-CU-NEXT:    v_add_co_ci_u32_e64 v1, null, s1, 0, s0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1] glc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s3 :: v_dual_mov_b32 v0, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -261,28 +259,26 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2 dlc
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_store_0:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2 dlc
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -357,33 +353,33 @@
 ; GFX11-WGP-LABEL: flat_nontemporal_store_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s1
-; GFX11-WGP-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-WGP-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-WGP-NEXT:    v_add_co_u32 v0, s0, s2, v0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[1:2]
 ; GFX11-WGP-NEXT:    v_add_co_ci_u32_e64 v1, null, s3, 0, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2 dlc
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_nontemporal_store_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s1
-; GFX11-CU-NEXT:    s_delay_alu instid0(VALU_DEP_3)
+; GFX11-CU-NEXT:    s_delay_alu instid0(VALU_DEP_2)
 ; GFX11-CU-NEXT:    v_add_co_u32 v0, s0, s2, v0
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[1:2]
 ; GFX11-CU-NEXT:    v_add_co_ci_u32_e64 v1, null, s3, 0, s0
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2 dlc
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -455,28 +451,26 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_volatile_workgroup_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -543,12 +537,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_volatile_workgroup_release_store:
@@ -557,11 +551,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32 %in, i32* %out) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-wavefront.ll
@@ -119,26 +119,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -256,26 +254,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -393,26 +389,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -530,26 +524,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -651,10 +643,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_unordered_store:
@@ -663,10 +655,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -767,10 +759,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_store:
@@ -779,10 +771,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -883,10 +875,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_store:
@@ -895,10 +887,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -999,10 +991,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_store:
@@ -1011,10 +1003,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1115,10 +1107,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_atomicrmw:
@@ -1127,10 +1119,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1231,10 +1223,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_atomicrmw:
@@ -1243,10 +1235,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1347,10 +1339,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_atomicrmw:
@@ -1359,10 +1351,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1463,10 +1455,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_atomicrmw:
@@ -1475,10 +1467,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1579,10 +1571,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_atomicrmw:
@@ -1591,10 +1583,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1711,12 +1703,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_ret_atomicrmw:
@@ -1725,12 +1717,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1848,12 +1840,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_ret_atomicrmw:
@@ -1862,12 +1854,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1985,12 +1977,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_ret_atomicrmw:
@@ -1999,12 +1991,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2118,11 +2110,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_monotonic_cmpxchg:
@@ -2131,11 +2122,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2249,11 +2239,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_monotonic_cmpxchg:
@@ -2262,11 +2251,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2380,11 +2368,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_monotonic_cmpxchg:
@@ -2393,11 +2380,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2511,11 +2497,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_monotonic_cmpxchg:
@@ -2524,11 +2509,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2642,11 +2626,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_monotonic_cmpxchg:
@@ -2655,11 +2638,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2773,11 +2755,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_acquire_cmpxchg:
@@ -2786,11 +2767,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2904,11 +2884,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_acquire_cmpxchg:
@@ -2917,11 +2896,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3035,11 +3013,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_acquire_cmpxchg:
@@ -3048,11 +3025,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3166,11 +3142,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_acquire_cmpxchg:
@@ -3179,11 +3154,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3297,11 +3271,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_acquire_cmpxchg:
@@ -3310,11 +3283,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3428,11 +3400,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_seq_cst_cmpxchg:
@@ -3441,11 +3412,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3559,11 +3529,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_seq_cst_cmpxchg:
@@ -3572,11 +3541,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3690,11 +3658,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_seq_cst_cmpxchg:
@@ -3703,11 +3670,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3821,11 +3787,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_seq_cst_cmpxchg:
@@ -3834,11 +3799,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3952,11 +3916,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_seq_cst_cmpxchg:
@@ -3965,11 +3928,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4107,13 +4069,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_monotonic_ret_cmpxchg:
@@ -4122,13 +4083,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4268,13 +4228,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_monotonic_ret_cmpxchg:
@@ -4283,13 +4242,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4429,13 +4387,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_monotonic_ret_cmpxchg:
@@ -4444,13 +4401,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4590,13 +4546,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_monotonic_ret_cmpxchg:
@@ -4605,13 +4560,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4751,13 +4705,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_monotonic_ret_cmpxchg:
@@ -4766,13 +4719,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4912,13 +4864,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_acquire_ret_cmpxchg:
@@ -4927,13 +4878,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5073,13 +5023,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_acquire_ret_cmpxchg:
@@ -5088,13 +5037,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5234,13 +5182,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_acquire_ret_cmpxchg:
@@ -5249,13 +5196,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5395,13 +5341,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_acquire_ret_cmpxchg:
@@ -5410,13 +5355,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5556,13 +5500,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_acquire_ret_cmpxchg:
@@ -5571,13 +5514,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5717,13 +5659,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_monotonic_seq_cst_ret_cmpxchg:
@@ -5732,13 +5673,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5878,13 +5818,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acquire_seq_cst_ret_cmpxchg:
@@ -5893,13 +5832,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6039,13 +5977,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_release_seq_cst_ret_cmpxchg:
@@ -6054,13 +5991,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6200,13 +6136,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_acq_rel_seq_cst_ret_cmpxchg:
@@ -6215,13 +6150,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6361,13 +6295,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_seq_cst_seq_cst_ret_cmpxchg:
@@ -6376,13 +6309,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6502,26 +6434,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6639,26 +6569,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6776,26 +6704,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6913,26 +6839,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -7034,10 +6958,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_unordered_store:
@@ -7046,10 +6970,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7150,10 +7074,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_store:
@@ -7162,10 +7086,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7266,10 +7190,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_release_store:
@@ -7278,10 +7202,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7382,10 +7306,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_store:
@@ -7394,10 +7318,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7498,10 +7422,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_atomicrmw:
@@ -7510,10 +7434,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7614,10 +7538,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_atomicrmw:
@@ -7626,10 +7550,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7730,10 +7654,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_release_atomicrmw:
@@ -7742,10 +7666,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7846,10 +7770,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_rel_atomicrmw:
@@ -7858,10 +7782,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7962,10 +7886,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_atomicrmw:
@@ -7974,10 +7898,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8094,12 +8018,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_ret_atomicrmw:
@@ -8108,12 +8032,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8231,12 +8155,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_rel_ret_atomicrmw:
@@ -8245,12 +8169,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8368,12 +8292,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_ret_atomicrmw:
@@ -8382,12 +8306,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8501,11 +8425,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_monotonic_cmpxchg:
@@ -8514,11 +8437,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8632,11 +8554,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_monotonic_cmpxchg:
@@ -8645,11 +8566,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8763,11 +8683,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_release_monotonic_cmpxchg:
@@ -8776,11 +8695,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8894,11 +8812,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_rel_monotonic_cmpxchg:
@@ -8907,11 +8824,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9025,11 +8941,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_monotonic_cmpxchg:
@@ -9038,11 +8953,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9156,11 +9070,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_acquire_cmpxchg:
@@ -9169,11 +9082,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9287,11 +9199,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_acquire_cmpxchg:
@@ -9300,11 +9211,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9418,11 +9328,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_release_acquire_cmpxchg:
@@ -9431,11 +9340,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9549,11 +9457,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_rel_acquire_cmpxchg:
@@ -9562,11 +9469,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9680,11 +9586,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_acquire_cmpxchg:
@@ -9693,11 +9598,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9811,11 +9715,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_seq_cst_cmpxchg:
@@ -9824,11 +9727,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9942,11 +9844,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_seq_cst_cmpxchg:
@@ -9955,11 +9856,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10073,11 +9973,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_release_seq_cst_cmpxchg:
@@ -10086,11 +9985,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10204,11 +10102,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_rel_seq_cst_cmpxchg:
@@ -10217,11 +10114,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10335,11 +10231,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
@@ -10348,11 +10243,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10490,13 +10384,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -10505,13 +10398,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10651,13 +10543,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
@@ -10666,13 +10557,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10812,13 +10702,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -10827,13 +10716,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10973,13 +10861,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -10988,13 +10875,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11134,13 +11020,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_acquire_ret_cmpxchg:
@@ -11149,13 +11034,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11295,13 +11179,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_acquire_ret_cmpxchg:
@@ -11310,13 +11193,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11456,13 +11338,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_release_acquire_ret_cmpxchg:
@@ -11471,13 +11352,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11617,13 +11497,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -11632,13 +11511,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11778,13 +11656,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -11793,13 +11670,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11939,13 +11815,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -11954,13 +11829,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12100,13 +11974,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -12115,13 +11988,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12261,13 +12133,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_release_seq_cst_ret_cmpxchg:
@@ -12276,13 +12147,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12422,13 +12292,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_acq_relc_seq_cst_ret_cmpxchg:
@@ -12437,13 +12306,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12583,13 +12451,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -12598,13 +12465,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-workgroup.ll
@@ -119,26 +119,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -256,26 +254,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -401,28 +397,26 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -557,31 +551,29 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -683,10 +675,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_unordered_store:
@@ -695,10 +687,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -799,10 +791,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_monotonic_store:
@@ -811,10 +803,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -924,12 +916,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_release_store:
@@ -938,11 +930,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1052,12 +1044,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_seq_cst_store:
@@ -1066,11 +1058,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -1171,10 +1163,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_monotonic_atomicrmw:
@@ -1183,10 +1175,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1299,8 +1291,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1314,11 +1305,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1428,12 +1419,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_release_atomicrmw:
@@ -1442,11 +1433,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1568,8 +1559,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1585,12 +1575,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1712,8 +1702,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -1729,12 +1718,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -1854,13 +1843,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acquire_ret_atomicrmw:
@@ -1869,12 +1858,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2004,8 +1993,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2013,6 +2001,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acq_rel_ret_atomicrmw:
@@ -2021,13 +2010,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2157,8 +2146,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2166,6 +2154,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_seq_cst_ret_atomicrmw:
@@ -2174,13 +2163,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -2294,11 +2283,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_monotonic_monotonic_cmpxchg:
@@ -2307,11 +2295,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2437,10 +2424,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2453,12 +2438,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2581,13 +2565,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_release_monotonic_cmpxchg:
@@ -2596,12 +2579,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2736,10 +2718,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -2754,13 +2734,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -2895,10 +2874,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -2913,13 +2890,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3045,10 +3021,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3061,12 +3035,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3192,10 +3165,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3208,12 +3179,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3348,10 +3318,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3366,13 +3334,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3507,10 +3474,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3525,13 +3490,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3666,10 +3630,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3684,13 +3646,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3825,10 +3786,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -3843,13 +3802,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -3987,13 +3945,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_monotonic_monotonic_ret_cmpxchg:
@@ -4002,13 +3959,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4154,14 +4110,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acquire_monotonic_ret_cmpxchg:
@@ -4170,13 +4125,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4325,15 +4279,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_release_monotonic_ret_cmpxchg:
@@ -4342,14 +4295,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4504,16 +4456,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acq_rel_monotonic_ret_cmpxchg:
@@ -4522,14 +4473,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4684,16 +4634,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_seq_cst_monotonic_ret_cmpxchg:
@@ -4702,14 +4651,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -4855,14 +4803,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_monotonic_acquire_ret_cmpxchg:
@@ -4871,13 +4818,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5023,14 +4969,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acquire_acquire_ret_cmpxchg:
@@ -5039,13 +4984,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5200,16 +5144,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_release_acquire_ret_cmpxchg:
@@ -5218,14 +5161,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5380,16 +5322,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acq_rel_acquire_ret_cmpxchg:
@@ -5398,14 +5339,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5560,16 +5500,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_seq_cst_acquire_ret_cmpxchg:
@@ -5578,14 +5517,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5740,16 +5678,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_monotonic_seq_cst_ret_cmpxchg:
@@ -5758,14 +5695,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -5920,16 +5856,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acquire_seq_cst_ret_cmpxchg:
@@ -5938,14 +5873,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6100,16 +6034,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_release_seq_cst_ret_cmpxchg:
@@ -6118,14 +6051,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6280,16 +6212,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_acq_rel_seq_cst_ret_cmpxchg:
@@ -6298,14 +6229,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6460,16 +6390,15 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_seq_cst_seq_cst_ret_cmpxchg:
@@ -6478,14 +6407,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -6605,26 +6533,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_unordered_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6742,26 +6668,24 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_monotonic_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -6883,28 +6807,26 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acquire_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -7030,30 +6952,28 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_load_b32 v2, v[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_seq_cst_load:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    flat_load_b32 v2, v[0:1]
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %in, i32* %out) {
 entry:
@@ -7155,10 +7075,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_unordered_store:
@@ -7167,10 +7087,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7271,10 +7191,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_monotonic_store:
@@ -7283,10 +7203,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7391,12 +7311,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_release_store:
@@ -7405,10 +7325,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7513,12 +7433,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_seq_cst_store:
@@ -7527,10 +7447,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32* %out) {
 entry:
@@ -7631,10 +7551,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_monotonic_atomicrmw:
@@ -7643,10 +7563,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7753,8 +7673,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -7767,10 +7686,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -7875,12 +7794,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_release_atomicrmw:
@@ -7889,10 +7808,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8003,8 +7922,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -8019,10 +7937,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8133,8 +8051,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -8149,10 +8066,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8273,14 +8190,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acquire_ret_atomicrmw:
@@ -8289,12 +8206,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8420,8 +8337,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -8430,6 +8346,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acq_rel_ret_atomicrmw:
@@ -8438,12 +8355,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8569,8 +8486,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -8579,6 +8495,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_seq_cst_ret_atomicrmw:
@@ -8587,12 +8504,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in) {
 entry:
@@ -8706,11 +8623,10 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_monotonic_monotonic_cmpxchg:
@@ -8719,11 +8635,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8843,10 +8758,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -8858,11 +8771,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -8980,13 +8892,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_release_monotonic_cmpxchg:
@@ -8995,11 +8906,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9123,10 +9033,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -9140,11 +9048,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9268,10 +9175,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -9285,11 +9190,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9409,10 +9313,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -9424,11 +9326,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9548,10 +9449,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -9563,11 +9462,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9691,10 +9589,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -9708,11 +9604,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9836,10 +9731,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -9853,11 +9746,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -9981,10 +9873,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -9998,11 +9888,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10126,10 +10015,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10143,11 +10030,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10271,10 +10157,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10288,11 +10172,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10416,10 +10299,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10433,11 +10314,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10561,10 +10441,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10578,11 +10456,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10706,10 +10583,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
@@ -10723,11 +10598,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -10865,13 +10739,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_monotonicmonotonic_ret_cmpxchg:
@@ -10880,13 +10753,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11030,15 +10902,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
@@ -11047,13 +10918,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11197,15 +11067,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_release_monotonic_ret_cmpxchg:
@@ -11214,13 +11083,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11368,10 +11236,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -11379,6 +11245,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -11387,13 +11254,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11541,10 +11407,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -11552,6 +11416,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -11560,13 +11425,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11710,15 +11574,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_monotonic_acquire_ret_cmpxchg:
@@ -11727,13 +11590,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -11877,15 +11739,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acquire_acquire_ret_cmpxchg:
@@ -11894,13 +11755,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12048,10 +11908,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12059,6 +11917,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_release_acquire_ret_cmpxchg:
@@ -12067,13 +11926,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12221,10 +12079,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12232,6 +12088,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -12240,13 +12097,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12394,10 +12250,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12405,6 +12259,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -12413,13 +12268,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12567,10 +12421,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12578,6 +12430,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -12586,13 +12439,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12740,10 +12592,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12751,6 +12601,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -12759,13 +12610,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -12913,10 +12763,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -12924,6 +12772,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_release_seq_cst_ret_cmpxchg:
@@ -12932,13 +12781,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13086,10 +12934,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13097,6 +12943,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -13105,13 +12952,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
@@ -13259,10 +13105,8 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
@@ -13270,6 +13114,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: flat_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -13278,13 +13123,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v3, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
-; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v2, s0
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_store_b32 v[0:1], v2
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-agent.ll
@@ -123,6 +123,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_unordered_load:
@@ -133,6 +134,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -253,6 +255,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_monotonic_load:
@@ -263,6 +266,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -395,6 +399,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acquire_load:
@@ -407,6 +412,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -545,6 +551,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_seq_cst_load:
@@ -558,6 +565,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -667,10 +675,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_unordered_store:
@@ -678,10 +686,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -790,10 +798,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_monotonic_store:
@@ -801,10 +809,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -926,12 +934,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_release_store:
@@ -939,12 +947,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1066,12 +1074,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_seq_cst_store:
@@ -1079,12 +1087,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1193,10 +1201,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_monotonic_atomicrmw:
@@ -1204,10 +1212,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1335,9 +1343,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1349,9 +1356,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -1477,12 +1483,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_release_atomicrmw:
@@ -1490,12 +1496,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1636,9 +1642,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1652,9 +1657,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1801,9 +1805,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1817,9 +1820,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1962,14 +1964,14 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acquire_ret_atomicrmw:
@@ -1977,14 +1979,14 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2135,9 +2137,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2145,6 +2146,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acq_rel_ret_atomicrmw:
@@ -2152,9 +2154,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2162,6 +2163,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2312,9 +2314,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2322,6 +2323,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_seq_cst_ret_atomicrmw:
@@ -2329,9 +2331,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2339,6 +2340,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2457,9 +2459,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_monotonic_monotonic_cmpxchg:
@@ -2469,9 +2471,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2609,8 +2611,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -2624,8 +2625,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -2761,11 +2761,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_release_monotonic_cmpxchg:
@@ -2775,11 +2775,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2930,8 +2930,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -2947,8 +2946,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3105,8 +3103,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3122,8 +3119,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3267,8 +3263,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -3282,8 +3277,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -3425,8 +3419,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -3440,8 +3433,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -3596,8 +3588,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3613,8 +3604,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3771,8 +3761,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3788,8 +3777,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3946,8 +3934,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3963,8 +3950,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4121,8 +4107,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4138,8 +4123,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4296,8 +4280,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4313,8 +4296,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4471,8 +4453,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4488,8 +4469,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4646,8 +4626,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4663,8 +4642,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4821,8 +4799,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4838,8 +4815,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4984,11 +4960,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_monotonic_monotonic_ret_cmpxchg:
@@ -4998,11 +4974,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5153,13 +5129,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acquire_monotonic_ret_cmpxchg:
@@ -5169,13 +5145,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5329,13 +5305,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_release_monotonic_ret_cmpxchg:
@@ -5345,13 +5321,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5515,8 +5491,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5524,6 +5499,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acq_rel_monotonic_ret_cmpxchg:
@@ -5533,8 +5509,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5542,6 +5517,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5705,8 +5681,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5714,6 +5689,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_seq_cst_monotonic_ret_cmpxchg:
@@ -5723,8 +5699,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5732,6 +5707,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5882,13 +5858,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_monotonic_acquire_ret_cmpxchg:
@@ -5898,13 +5874,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6055,13 +6031,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acquire_acquire_ret_cmpxchg:
@@ -6071,13 +6047,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6241,8 +6217,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6250,6 +6225,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_release_acquire_ret_cmpxchg:
@@ -6259,8 +6235,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6268,6 +6243,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6431,8 +6407,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6440,6 +6415,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acq_rel_acquire_ret_cmpxchg:
@@ -6449,8 +6425,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6458,6 +6433,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6621,8 +6597,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6630,6 +6605,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_seq_cst_acquire_ret_cmpxchg:
@@ -6639,8 +6615,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6648,6 +6623,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6811,8 +6787,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6820,6 +6795,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_monotonic_seq_cst_ret_cmpxchg:
@@ -6829,8 +6805,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6838,6 +6813,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -7001,8 +6977,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7010,6 +6985,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acquire_seq_cst_ret_cmpxchg:
@@ -7019,8 +6995,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7028,6 +7003,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -7191,8 +7167,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7200,6 +7175,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_release_seq_cst_ret_cmpxchg:
@@ -7209,8 +7185,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7218,6 +7193,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -7381,8 +7357,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7390,6 +7365,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_acq_rel_seq_cst_ret_cmpxchg:
@@ -7399,8 +7375,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7408,6 +7383,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -7571,8 +7547,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7580,6 +7555,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_seq_cst_seq_cst_ret_cmpxchg:
@@ -7589,8 +7565,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -7598,6 +7573,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -7720,6 +7696,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_unordered_load:
@@ -7730,6 +7707,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7850,6 +7828,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_monotonic_load:
@@ -7860,6 +7839,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7992,6 +7972,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acquire_load:
@@ -8004,6 +7985,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -8142,6 +8124,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_seq_cst_load:
@@ -8155,6 +8138,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -8264,10 +8248,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_unordered_store:
@@ -8275,10 +8259,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -8387,10 +8371,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_monotonic_store:
@@ -8398,10 +8382,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -8523,12 +8507,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_release_store:
@@ -8536,12 +8520,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -8663,12 +8647,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_seq_cst_store:
@@ -8676,12 +8660,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -8790,10 +8774,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_monotonic_atomicrmw:
@@ -8801,10 +8785,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8932,9 +8916,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -8946,9 +8929,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -9074,12 +9056,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_release_atomicrmw:
@@ -9087,12 +9069,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9233,9 +9215,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -9249,9 +9230,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -9398,9 +9378,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -9414,9 +9393,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -9559,14 +9537,14 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acquire_ret_atomicrmw:
@@ -9574,14 +9552,14 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9732,9 +9710,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9742,6 +9719,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acq_rel_ret_atomicrmw:
@@ -9749,9 +9727,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9759,6 +9736,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9909,9 +9887,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9919,6 +9896,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_seq_cst_ret_atomicrmw:
@@ -9926,9 +9904,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9936,6 +9913,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -10054,9 +10032,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_monotonic_monotonic_cmpxchg:
@@ -10066,9 +10044,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10206,8 +10184,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10221,8 +10198,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -10358,11 +10334,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_release_monotonic_cmpxchg:
@@ -10372,11 +10348,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10527,8 +10503,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10544,8 +10519,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10702,8 +10676,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10719,8 +10692,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10864,8 +10836,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10879,8 +10850,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -11022,8 +10992,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -11037,8 +11006,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -11193,8 +11161,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11210,8 +11177,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11368,8 +11334,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11385,8 +11350,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11543,8 +11507,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11560,8 +11523,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11718,8 +11680,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11735,8 +11696,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11893,8 +11853,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11910,8 +11869,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -12068,8 +12026,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -12085,8 +12042,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -12243,8 +12199,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -12260,8 +12215,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -12418,8 +12372,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -12435,8 +12388,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -12581,11 +12533,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -12595,11 +12547,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12750,13 +12702,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acquire_monotonic_ret_cmpxchg:
@@ -12766,13 +12718,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12936,8 +12888,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -12945,6 +12896,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -12954,8 +12906,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -12963,6 +12914,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13126,8 +13078,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13135,6 +13086,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -13144,8 +13096,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13153,6 +13104,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13303,13 +13255,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_monotonic_acquire_ret_cmpxchg:
@@ -13319,13 +13271,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13476,13 +13428,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acquire_acquire_ret_cmpxchg:
@@ -13492,13 +13444,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13662,8 +13614,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13671,6 +13622,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_release_acquire_ret_cmpxchg:
@@ -13680,8 +13632,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13689,6 +13640,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13852,8 +13804,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13861,6 +13812,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -13870,8 +13822,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13879,6 +13830,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14042,8 +13994,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14051,6 +14002,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -14060,8 +14012,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14069,6 +14020,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14232,8 +14184,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14241,6 +14192,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -14250,8 +14202,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14259,6 +14210,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14422,8 +14374,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14431,6 +14382,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -14440,8 +14392,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14449,6 +14400,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14612,8 +14564,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14621,6 +14572,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_release_seq_cst_ret_cmpxchg:
@@ -14630,8 +14582,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14639,6 +14590,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14802,8 +14754,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14811,6 +14762,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -14820,8 +14772,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14829,6 +14780,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14992,8 +14944,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -15001,6 +14952,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_agent_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -15010,8 +14962,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -15019,6 +14970,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll
@@ -121,23 +121,23 @@
 ; GFX11-WGP-LABEL: global_nontemporal_load_0:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_nontemporal_load_0:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -265,23 +265,23 @@
 ; GFX11-WGP-LABEL: global_nontemporal_load_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    global_load_b32 v0, v0, s[0:1] slc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_nontemporal_load_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_load_b32 v0, v0, s[0:1] slc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -402,23 +402,23 @@
 ; GFX11-WGP-LABEL: global_nontemporal_store_0:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3] glc slc dlc
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_nontemporal_store_0:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3] glc slc dlc
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -541,23 +541,23 @@
 ; GFX11-WGP-LABEL: global_nontemporal_store_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3] glc slc dlc
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_nontemporal_store_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3] glc slc dlc
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-singlethread.ll
@@ -123,6 +123,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_unordered_load:
@@ -133,6 +134,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -253,6 +255,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_load:
@@ -263,6 +266,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -383,6 +387,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_load:
@@ -393,6 +398,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -513,6 +519,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_load:
@@ -523,6 +530,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -632,10 +640,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_unordered_store:
@@ -643,10 +651,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -755,10 +763,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_store:
@@ -766,10 +774,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -878,10 +886,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_store:
@@ -889,10 +897,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1001,10 +1009,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_store:
@@ -1012,10 +1020,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1124,10 +1132,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_atomicrmw:
@@ -1135,10 +1143,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1247,10 +1255,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_atomicrmw:
@@ -1258,10 +1266,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1370,10 +1378,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_atomicrmw:
@@ -1381,10 +1389,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1493,10 +1501,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_atomicrmw:
@@ -1504,10 +1512,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1616,10 +1624,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_atomicrmw:
@@ -1627,10 +1635,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1757,12 +1765,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_ret_atomicrmw:
@@ -1770,12 +1778,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1903,12 +1911,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_ret_atomicrmw:
@@ -1916,12 +1924,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2049,12 +2057,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_ret_atomicrmw:
@@ -2062,12 +2070,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2186,9 +2194,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_monotonic_cmpxchg:
@@ -2198,9 +2206,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2319,9 +2327,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_monotonic_cmpxchg:
@@ -2331,9 +2339,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2452,9 +2460,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_monotonic_cmpxchg:
@@ -2464,9 +2472,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2585,9 +2593,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_monotonic_cmpxchg:
@@ -2597,9 +2605,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2718,9 +2726,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_monotonic_cmpxchg:
@@ -2730,9 +2738,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2851,9 +2859,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_acquire_cmpxchg:
@@ -2863,9 +2871,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2984,9 +2992,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_acquire_cmpxchg:
@@ -2996,9 +3004,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3117,9 +3125,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_acquire_cmpxchg:
@@ -3129,9 +3137,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3250,9 +3258,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_acquire_cmpxchg:
@@ -3262,9 +3270,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3383,9 +3391,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_acquire_cmpxchg:
@@ -3395,9 +3403,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3516,9 +3524,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_seq_cst_cmpxchg:
@@ -3528,9 +3536,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3649,9 +3657,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_seq_cst_cmpxchg:
@@ -3661,9 +3669,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3782,9 +3790,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_seq_cst_cmpxchg:
@@ -3794,9 +3802,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3915,9 +3923,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_seq_cst_cmpxchg:
@@ -3927,9 +3935,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4048,9 +4056,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_seq_cst_cmpxchg:
@@ -4060,9 +4068,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4201,11 +4209,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_monotonic_ret_cmpxchg:
@@ -4215,11 +4223,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4360,11 +4368,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_monotonic_ret_cmpxchg:
@@ -4374,11 +4382,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4519,11 +4527,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_monotonic_ret_cmpxchg:
@@ -4533,11 +4541,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4678,11 +4686,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_monotonic_ret_cmpxchg:
@@ -4692,11 +4700,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4837,11 +4845,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_monotonic_ret_cmpxchg:
@@ -4851,11 +4859,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4996,11 +5004,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_acquire_ret_cmpxchg:
@@ -5010,11 +5018,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5155,11 +5163,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_acquire_ret_cmpxchg:
@@ -5169,11 +5177,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5314,11 +5322,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_acquire_ret_cmpxchg:
@@ -5328,11 +5336,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5473,11 +5481,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_acquire_ret_cmpxchg:
@@ -5487,11 +5495,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5632,11 +5640,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_acquire_ret_cmpxchg:
@@ -5646,11 +5654,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5791,11 +5799,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_monotonic_seq_cst_ret_cmpxchg:
@@ -5805,11 +5813,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5950,11 +5958,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acquire_seq_cst_ret_cmpxchg:
@@ -5964,11 +5972,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6109,11 +6117,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_release_seq_cst_ret_cmpxchg:
@@ -6123,11 +6131,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6268,11 +6276,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_acq_rel_seq_cst_ret_cmpxchg:
@@ -6282,11 +6290,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6427,11 +6435,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_seq_cst_seq_cst_ret_cmpxchg:
@@ -6441,11 +6449,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6568,6 +6576,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_unordered_load:
@@ -6578,6 +6587,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -6698,6 +6708,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_load:
@@ -6708,6 +6719,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -6828,6 +6840,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_load:
@@ -6838,6 +6851,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -6958,6 +6972,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_load:
@@ -6968,6 +6983,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7077,10 +7093,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_unordered_store:
@@ -7088,10 +7104,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7200,10 +7216,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_store:
@@ -7211,10 +7227,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7323,10 +7339,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_store:
@@ -7334,10 +7350,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7446,10 +7462,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_store:
@@ -7457,10 +7473,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7569,10 +7585,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_atomicrmw:
@@ -7580,10 +7596,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -7692,10 +7708,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_atomicrmw:
@@ -7703,10 +7719,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -7815,10 +7831,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_atomicrmw:
@@ -7826,10 +7842,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -7938,10 +7954,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_atomicrmw:
@@ -7949,10 +7965,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8061,10 +8077,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_atomicrmw:
@@ -8072,10 +8088,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8202,12 +8218,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_ret_atomicrmw:
@@ -8215,12 +8231,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8348,12 +8364,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_ret_atomicrmw:
@@ -8361,12 +8377,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8494,12 +8510,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_ret_atomicrmw:
@@ -8507,12 +8523,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8631,9 +8647,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_cmpxchg:
@@ -8643,9 +8659,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -8764,9 +8780,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_monotonic_cmpxchg:
@@ -8776,9 +8792,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -8897,9 +8913,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_monotonic_cmpxchg:
@@ -8909,9 +8925,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9030,9 +9046,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_cmpxchg:
@@ -9042,9 +9058,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9163,9 +9179,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_cmpxchg:
@@ -9175,9 +9191,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9296,9 +9312,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_acquire_cmpxchg:
@@ -9308,9 +9324,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9429,9 +9445,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_acquire_cmpxchg:
@@ -9441,9 +9457,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9562,9 +9578,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_acquire_cmpxchg:
@@ -9574,9 +9590,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9695,9 +9711,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_cmpxchg:
@@ -9707,9 +9723,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9828,9 +9844,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_cmpxchg:
@@ -9840,9 +9856,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9961,9 +9977,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_cmpxchg:
@@ -9973,9 +9989,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10094,9 +10110,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_cmpxchg:
@@ -10106,9 +10122,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10227,9 +10243,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_seq_cst_cmpxchg:
@@ -10239,9 +10255,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10360,9 +10376,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_cmpxchg:
@@ -10372,9 +10388,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10493,9 +10509,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_cmpxchg:
@@ -10505,9 +10521,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10646,11 +10662,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -10660,11 +10676,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10805,11 +10821,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_monotonic_ret_cmpxchg:
@@ -10819,11 +10835,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10964,11 +10980,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_monotonic_ret_cmpxchg:
@@ -10978,11 +10994,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11123,11 +11139,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -11137,11 +11153,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11282,11 +11298,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -11296,11 +11312,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11441,11 +11457,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_acquire_ret_cmpxchg:
@@ -11455,11 +11471,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11600,11 +11616,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_acquire_ret_cmpxchg:
@@ -11614,11 +11630,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11759,11 +11775,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_acquire_ret_cmpxchg:
@@ -11773,11 +11789,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11918,11 +11934,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -11932,11 +11948,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12077,11 +12093,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -12091,11 +12107,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12236,11 +12252,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -12250,11 +12266,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12395,11 +12411,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -12409,11 +12425,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12554,11 +12570,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_release_seq_cst_ret_cmpxchg:
@@ -12568,11 +12584,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12713,11 +12729,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -12727,11 +12743,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12872,11 +12888,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_singlethread_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -12886,11 +12902,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-system.ll
@@ -123,6 +123,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_unordered_load:
@@ -133,6 +134,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -253,6 +255,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_monotonic_load:
@@ -263,6 +266,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -397,6 +401,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acquire_load:
@@ -409,6 +414,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -549,6 +555,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_seq_cst_load:
@@ -562,6 +569,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -671,10 +679,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_unordered_store:
@@ -682,10 +690,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -794,10 +802,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_monotonic_store:
@@ -805,10 +813,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -932,12 +940,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_release_store:
@@ -945,12 +953,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1074,12 +1082,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_seq_cst_store:
@@ -1087,12 +1095,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1201,10 +1209,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_monotonic_atomicrmw:
@@ -1212,10 +1220,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1345,9 +1353,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1359,9 +1366,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -1489,12 +1495,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_release_atomicrmw:
@@ -1502,12 +1508,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1652,9 +1658,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1668,9 +1673,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1821,9 +1825,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1837,9 +1840,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1984,14 +1986,14 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acquire_ret_atomicrmw:
@@ -1999,14 +2001,14 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2161,9 +2163,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2171,6 +2172,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acq_rel_ret_atomicrmw:
@@ -2178,9 +2180,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2188,6 +2189,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2342,9 +2344,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2352,6 +2353,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_seq_cst_ret_atomicrmw:
@@ -2359,9 +2361,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -2369,6 +2370,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2487,9 +2489,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_monotonic_monotonic_cmpxchg:
@@ -2499,9 +2501,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2641,8 +2643,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -2656,8 +2657,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -2795,11 +2795,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_release_monotonic_cmpxchg:
@@ -2809,11 +2809,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2968,8 +2968,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -2985,8 +2984,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3147,8 +3145,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3164,8 +3161,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3311,8 +3307,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -3326,8 +3321,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -3471,8 +3465,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -3486,8 +3479,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -3646,8 +3638,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3663,8 +3654,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3825,8 +3815,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3842,8 +3831,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4004,8 +3992,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4021,8 +4008,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4183,8 +4169,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4200,8 +4185,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4346,11 +4330,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_monotonic_monotonic_ret_cmpxchg:
@@ -4360,11 +4344,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4517,13 +4501,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acquire_monotonic_ret_cmpxchg:
@@ -4533,13 +4517,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4707,8 +4691,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -4716,6 +4699,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acq_rel_monotonic_ret_cmpxchg:
@@ -4725,8 +4709,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -4734,6 +4717,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4901,8 +4885,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -4910,6 +4893,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_seq_cst_monotonic_ret_cmpxchg:
@@ -4919,8 +4903,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -4928,6 +4911,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5080,13 +5064,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_monotonic_acquire_ret_cmpxchg:
@@ -5096,13 +5080,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5255,13 +5239,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acquire_acquire_ret_cmpxchg:
@@ -5271,13 +5255,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5445,8 +5429,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5454,6 +5437,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_release_acquire_ret_cmpxchg:
@@ -5463,8 +5447,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5472,6 +5455,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5639,8 +5623,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5648,6 +5631,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acq_rel_acquire_ret_cmpxchg:
@@ -5657,8 +5641,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5666,6 +5649,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5833,8 +5817,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5842,6 +5825,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_seq_cst_acquire_ret_cmpxchg:
@@ -5851,8 +5835,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -5860,6 +5843,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6027,8 +6011,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6036,6 +6019,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_monotonic_seq_cst_ret_cmpxchg:
@@ -6045,8 +6029,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6054,6 +6037,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6221,8 +6205,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6230,6 +6213,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acquire_seq_cst_ret_cmpxchg:
@@ -6239,8 +6223,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6248,6 +6231,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6415,8 +6399,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6424,6 +6407,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_relese_seq_cst_ret_cmpxchg:
@@ -6433,8 +6417,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6442,6 +6425,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6609,8 +6593,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6618,6 +6601,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_acq_rel_seq_cst_ret_cmpxchg:
@@ -6627,8 +6611,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6636,6 +6619,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6803,8 +6787,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6812,6 +6795,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_seq_cst_seq_cst_ret_cmpxchg:
@@ -6821,8 +6805,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -6830,6 +6813,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6952,6 +6936,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_unordered_load:
@@ -6962,6 +6947,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7082,6 +7068,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_monotonic_load:
@@ -7092,6 +7079,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7226,6 +7214,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acquire_load:
@@ -7238,6 +7227,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7378,6 +7368,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_seq_cst_load:
@@ -7391,6 +7382,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7500,10 +7492,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_unordered_store:
@@ -7511,10 +7503,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7623,10 +7615,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_monotonic_store:
@@ -7634,10 +7626,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7761,12 +7753,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_release_store:
@@ -7774,12 +7766,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7903,12 +7895,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_seq_cst_store:
@@ -7916,12 +7908,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -8030,10 +8022,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_monotonic_atomicrmw:
@@ -8041,10 +8033,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8174,9 +8166,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -8188,9 +8179,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -8318,12 +8308,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_release_atomicrmw:
@@ -8331,12 +8321,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8481,9 +8471,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -8497,9 +8486,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -8650,9 +8638,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -8666,9 +8653,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -8813,14 +8799,14 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acquire_ret_atomicrmw:
@@ -8828,14 +8814,14 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8990,9 +8976,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9000,6 +8985,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acq_rel_ret_atomicrmw:
@@ -9007,9 +8993,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9017,6 +9002,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9171,9 +9157,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9181,6 +9166,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_seq_cst_ret_atomicrmw:
@@ -9188,9 +9174,8 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
@@ -9198,6 +9183,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9316,9 +9302,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_monotonic_monotonic_cmpxchg:
@@ -9328,9 +9314,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9470,8 +9456,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -9485,8 +9470,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -9624,11 +9608,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_release_monotonic_cmpxchg:
@@ -9638,11 +9622,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9797,8 +9781,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -9814,8 +9797,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -9976,8 +9958,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -9993,8 +9974,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10140,8 +10120,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10155,8 +10134,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -10300,8 +10278,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10315,8 +10292,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -10475,8 +10451,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10492,8 +10467,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10654,8 +10628,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10671,8 +10644,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10833,8 +10805,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10850,8 +10821,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11012,8 +10982,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11029,8 +10998,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11191,8 +11159,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11208,8 +11175,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11370,8 +11336,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11387,8 +11352,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11549,8 +11513,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11566,8 +11529,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11728,8 +11690,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11745,8 +11706,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11891,11 +11851,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -11905,11 +11865,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12062,13 +12022,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acquire_monotonic_ret_cmpxchg:
@@ -12078,13 +12038,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12240,13 +12200,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_release_monotonic_ret_cmpxchg:
@@ -12256,13 +12216,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12430,8 +12390,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -12439,6 +12398,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -12448,8 +12408,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -12457,6 +12416,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12624,8 +12584,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -12633,6 +12592,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -12642,8 +12602,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -12651,6 +12610,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12803,13 +12763,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_monotonic_acquire_ret_cmpxchg:
@@ -12819,13 +12779,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12978,13 +12938,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acquire_acquire_ret_cmpxchg:
@@ -12994,13 +12954,13 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13168,8 +13128,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13177,6 +13136,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_release_acquire_ret_cmpxchg:
@@ -13186,8 +13146,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13195,6 +13154,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13362,8 +13322,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13371,6 +13330,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -13380,8 +13340,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13389,6 +13348,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13556,8 +13516,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13565,6 +13524,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -13574,8 +13534,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13583,6 +13542,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13750,8 +13710,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13759,6 +13718,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -13768,8 +13728,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13777,6 +13736,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13944,8 +13904,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13953,6 +13912,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -13962,8 +13922,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -13971,6 +13930,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14138,8 +14098,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14147,6 +14106,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_release_seq_cst_ret_cmpxchg:
@@ -14156,8 +14116,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14165,6 +14124,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14332,8 +14292,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14341,6 +14300,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -14350,8 +14310,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14359,6 +14318,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -14526,8 +14486,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14535,6 +14494,7 @@
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    buffer_gl1_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_system_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -14544,8 +14504,7 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
@@ -14553,6 +14512,7 @@
 ; GFX11-CU-NEXT:    buffer_gl0_inv
 ; GFX11-CU-NEXT:    buffer_gl1_inv
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-volatile.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-volatile.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-volatile.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-volatile.ll
@@ -79,6 +79,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1] glc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_volatile_load_0:
@@ -89,6 +90,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1] glc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -172,23 +174,23 @@
 ; GFX11-WGP-LABEL: global_volatile_load_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    global_load_b32 v0, v0, s[0:1] glc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_volatile_load_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_load_b32 v0, v0, s[0:1] glc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -270,25 +272,25 @@
 ; GFX11-WGP-LABEL: global_volatile_store_0:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3] dlc
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_volatile_store_0:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3] dlc
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -372,25 +374,25 @@
 ; GFX11-WGP-LABEL: global_volatile_store_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3] dlc
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_volatile_store_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3] dlc
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -475,6 +477,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_volatile_workgroup_acquire_load:
@@ -485,6 +488,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -560,12 +564,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_volatile_workgroup_release_store:
@@ -573,11 +577,11 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32 %in, i32 addrspace(1)* %out) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-wavefront.ll
@@ -123,6 +123,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_unordered_load:
@@ -133,6 +134,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -253,6 +255,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_load:
@@ -263,6 +266,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -383,6 +387,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_load:
@@ -393,6 +398,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -513,6 +519,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_load:
@@ -523,6 +530,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -632,10 +640,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_unordered_store:
@@ -643,10 +651,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -755,10 +763,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_store:
@@ -766,10 +774,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -878,10 +886,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_store:
@@ -889,10 +897,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1001,10 +1009,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_store:
@@ -1012,10 +1020,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1124,10 +1132,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_atomicrmw:
@@ -1135,10 +1143,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1247,10 +1255,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_atomicrmw:
@@ -1258,10 +1266,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1370,10 +1378,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_atomicrmw:
@@ -1381,10 +1389,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1493,10 +1501,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_atomicrmw:
@@ -1504,10 +1512,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1616,10 +1624,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_atomicrmw:
@@ -1627,10 +1635,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1757,12 +1765,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_ret_atomicrmw:
@@ -1770,12 +1778,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1903,12 +1911,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_ret_atomicrmw:
@@ -1916,12 +1924,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2049,12 +2057,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_ret_atomicrmw:
@@ -2062,12 +2070,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2186,9 +2194,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_monotonic_cmpxchg:
@@ -2198,9 +2206,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2319,9 +2327,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_monotonic_cmpxchg:
@@ -2331,9 +2339,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2452,9 +2460,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_monotonic_cmpxchg:
@@ -2464,9 +2472,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2585,9 +2593,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_monotonic_cmpxchg:
@@ -2597,9 +2605,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2718,9 +2726,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_monotonic_cmpxchg:
@@ -2730,9 +2738,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2851,9 +2859,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_acquire_cmpxchg:
@@ -2863,9 +2871,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2984,9 +2992,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_acquire_cmpxchg:
@@ -2996,9 +3004,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3117,9 +3125,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_acquire_cmpxchg:
@@ -3129,9 +3137,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3250,9 +3258,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_acquire_cmpxchg:
@@ -3262,9 +3270,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3383,9 +3391,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_acquire_cmpxchg:
@@ -3395,9 +3403,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3516,9 +3524,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_seq_cst_cmpxchg:
@@ -3528,9 +3536,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3649,9 +3657,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_seq_cst_cmpxchg:
@@ -3661,9 +3669,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3782,9 +3790,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_seq_cst_cmpxchg:
@@ -3794,9 +3802,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3915,9 +3923,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_seq_cst_cmpxchg:
@@ -3927,9 +3935,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4048,9 +4056,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_seq_cst_cmpxchg:
@@ -4060,9 +4068,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4201,11 +4209,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_monotonic_ret_cmpxchg:
@@ -4215,11 +4223,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4360,11 +4368,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_monotonic_ret_cmpxchg:
@@ -4374,11 +4382,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4519,11 +4527,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_monotonic_ret_cmpxchg:
@@ -4533,11 +4541,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
    i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4678,11 +4686,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_monotonic_ret_cmpxchg:
@@ -4692,11 +4700,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4837,11 +4845,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_monotonic_ret_cmpxchg:
@@ -4851,11 +4859,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4996,11 +5004,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_acquire_ret_cmpxchg:
@@ -5010,11 +5018,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5155,11 +5163,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_acquire_ret_cmpxchg:
@@ -5169,11 +5177,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5314,11 +5322,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_acquire_ret_cmpxchg:
@@ -5328,11 +5336,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5473,11 +5481,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_acquire_ret_cmpxchg:
@@ -5487,11 +5495,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5632,11 +5640,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_acquire_ret_cmpxchg:
@@ -5646,11 +5654,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5791,11 +5799,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_monotonic_seq_cst_ret_cmpxchg:
@@ -5805,11 +5813,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5950,11 +5958,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acquire_seq_cst_ret_cmpxchg:
@@ -5964,11 +5972,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6109,11 +6117,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_release_seq_cst_ret_cmpxchg:
@@ -6123,11 +6131,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6268,11 +6276,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_acq_rel_seq_cst_ret_cmpxchg:
@@ -6282,11 +6290,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6427,11 +6435,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_seq_cst_seq_cst_ret_cmpxchg:
@@ -6441,11 +6449,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6568,6 +6576,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_unordered_load:
@@ -6578,6 +6587,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -6698,6 +6708,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_load:
@@ -6708,6 +6719,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -6828,6 +6840,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_load:
@@ -6838,6 +6851,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -6958,6 +6972,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_load:
@@ -6968,6 +6983,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7077,10 +7093,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_unordered_store:
@@ -7088,10 +7104,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7200,10 +7216,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_store:
@@ -7211,10 +7227,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7323,10 +7339,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_store:
@@ -7334,10 +7350,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7446,10 +7462,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_store:
@@ -7457,10 +7473,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7569,10 +7585,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_atomicrmw:
@@ -7580,10 +7596,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -7692,10 +7708,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_atomicrmw:
@@ -7703,10 +7719,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -7815,10 +7831,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_atomicrmw:
@@ -7826,10 +7842,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -7938,10 +7954,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_atomicrmw:
@@ -7949,10 +7965,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8061,10 +8077,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_atomicrmw:
@@ -8072,10 +8088,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8202,12 +8218,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_ret_atomicrmw:
@@ -8215,12 +8231,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8348,12 +8364,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_ret_atomicrmw:
@@ -8361,12 +8377,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8494,12 +8510,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_ret_atomicrmw:
@@ -8507,12 +8523,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8631,9 +8647,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_monotonic_cmpxchg:
@@ -8643,9 +8659,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -8764,9 +8780,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_monotonic_cmpxchg:
@@ -8776,9 +8792,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -8897,9 +8913,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_monotonic_cmpxchg:
@@ -8909,9 +8925,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9030,9 +9046,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_monotonic_cmpxchg:
@@ -9042,9 +9058,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9163,9 +9179,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_monotonic_cmpxchg:
@@ -9175,9 +9191,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9296,9 +9312,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_acquire_cmpxchg:
@@ -9308,9 +9324,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9429,9 +9445,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_acquire_cmpxchg:
@@ -9441,9 +9457,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9562,9 +9578,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_acquire_cmpxchg:
@@ -9574,9 +9590,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9695,9 +9711,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_acquire_cmpxchg:
@@ -9707,9 +9723,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9828,9 +9844,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_acquire_cmpxchg:
@@ -9840,9 +9856,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9961,9 +9977,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_seq_cst_cmpxchg:
@@ -9973,9 +9989,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10094,9 +10110,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_seq_cst_cmpxchg:
@@ -10106,9 +10122,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10227,9 +10243,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_seq_cst_cmpxchg:
@@ -10239,9 +10255,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10360,9 +10376,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_seq_cst_cmpxchg:
@@ -10372,9 +10388,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10493,9 +10509,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_seq_cst_cmpxchg:
@@ -10505,9 +10521,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10646,11 +10662,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -10660,11 +10676,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10805,11 +10821,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_monotonic_ret_cmpxchg:
@@ -10819,11 +10835,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10964,11 +10980,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_monotonic_ret_cmpxchg:
@@ -10978,11 +10994,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11123,11 +11139,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -11137,11 +11153,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11282,11 +11298,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -11296,11 +11312,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11441,11 +11457,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_acquire_ret_cmpxchg:
@@ -11455,11 +11471,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11600,11 +11616,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_acquire_ret_cmpxchg:
@@ -11614,11 +11630,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11759,11 +11775,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_acquire_ret_cmpxchg:
@@ -11773,11 +11789,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11918,11 +11934,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -11932,11 +11948,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12077,11 +12093,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -12091,11 +12107,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12236,11 +12252,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -12250,11 +12266,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12395,11 +12411,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -12409,11 +12425,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12554,11 +12570,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_release_seq_cst_ret_cmpxchg:
@@ -12568,11 +12584,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12713,11 +12729,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -12727,11 +12743,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12872,11 +12888,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_wavefront_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -12886,11 +12902,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-global-workgroup.ll
@@ -123,6 +123,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_unordered_load:
@@ -133,6 +134,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -253,6 +255,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_monotonic_load:
@@ -263,6 +266,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -387,6 +391,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acquire_load:
@@ -397,6 +402,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -526,6 +532,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_seq_cst_load:
@@ -536,6 +543,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -645,10 +653,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_unordered_store:
@@ -656,10 +664,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -768,10 +776,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_monotonic_store:
@@ -779,10 +787,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -901,12 +909,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_release_store:
@@ -914,11 +922,11 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1037,12 +1045,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_seq_cst_store:
@@ -1050,11 +1058,11 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -1163,10 +1171,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_monotonic_atomicrmw:
@@ -1174,10 +1182,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1292,9 +1300,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1305,10 +1312,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1427,12 +1434,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_release_atomicrmw:
@@ -1440,11 +1447,11 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1569,9 +1576,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1584,11 +1590,11 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1713,9 +1719,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -1728,11 +1733,11 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -1862,13 +1867,13 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acquire_ret_atomicrmw:
@@ -1876,12 +1881,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2022,15 +2027,15 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acq_rel_ret_atomicrmw:
@@ -2038,13 +2043,13 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2185,15 +2190,15 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_seq_cst_ret_atomicrmw:
@@ -2201,13 +2206,13 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -2326,9 +2331,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_monotonic_monotonic_cmpxchg:
@@ -2338,9 +2343,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2465,8 +2470,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -2479,9 +2483,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2610,11 +2614,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_release_monotonic_cmpxchg:
@@ -2624,10 +2628,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2762,8 +2766,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -2778,10 +2781,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -2916,8 +2919,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -2932,10 +2934,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3060,8 +3062,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -3074,9 +3075,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3201,8 +3202,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -3215,9 +3215,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3352,8 +3352,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3368,10 +3367,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3506,8 +3505,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3522,10 +3520,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3660,8 +3658,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3676,10 +3673,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3814,8 +3811,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3830,10 +3826,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -3968,8 +3964,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -3984,10 +3979,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4122,8 +4117,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4138,10 +4132,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4276,8 +4270,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4292,10 +4285,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4430,8 +4423,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -4446,10 +4438,10 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4588,11 +4580,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_monotonic_monotonic_ret_cmpxchg:
@@ -4602,11 +4594,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4750,12 +4742,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acquire_monotonic_ret_cmpxchg:
@@ -4765,11 +4757,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -4920,13 +4912,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_release_monotonic_ret_cmpxchg:
@@ -4936,12 +4928,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5095,14 +5087,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acq_rel_monotonic_ret_cmpxchg:
@@ -5112,12 +5104,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5271,14 +5263,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_seq_cst_monotonic_ret_cmpxchg:
@@ -5288,12 +5280,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5437,12 +5429,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_monotonic_acquire_ret_cmpxchg:
@@ -5452,11 +5444,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5600,12 +5592,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acquire_acquire_ret_cmpxchg:
@@ -5615,11 +5607,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5773,14 +5765,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_release_acquire_ret_cmpxchg:
@@ -5790,12 +5782,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -5949,14 +5941,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acq_rel_acquire_ret_cmpxchg:
@@ -5966,12 +5958,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6125,14 +6117,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_seq_cst_acquire_ret_cmpxchg:
@@ -6142,12 +6134,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6301,14 +6293,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_monotonic_seq_cst_ret_cmpxchg:
@@ -6318,12 +6310,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6477,14 +6469,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acquire_seq_cst_ret_cmpxchg:
@@ -6494,12 +6486,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6653,14 +6645,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_release_seq_cst_ret_cmpxchg:
@@ -6670,12 +6662,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -6829,14 +6821,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_acq_rel_seq_cst_ret_cmpxchg:
@@ -6846,12 +6838,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -7005,14 +6997,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_seq_cst_seq_cst_ret_cmpxchg:
@@ -7022,12 +7014,12 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -7150,6 +7142,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_unordered_load:
@@ -7160,6 +7153,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7280,6 +7274,7 @@
 ; GFX11-WGP-NEXT:    global_load_b32 v1, v0, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_monotonic_load:
@@ -7290,6 +7285,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7414,6 +7410,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acquire_load:
@@ -7424,6 +7421,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7550,6 +7548,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_seq_cst_load:
@@ -7560,6 +7559,7 @@
 ; GFX11-CU-NEXT:    global_load_b32 v1, v0, s[0:1]
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[2:3]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -7669,10 +7669,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_unordered_store:
@@ -7680,10 +7680,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7792,10 +7792,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_monotonic_store:
@@ -7803,10 +7803,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -7919,12 +7919,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_release_store:
@@ -7932,10 +7932,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -8048,12 +8048,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_seq_cst_store:
@@ -8061,10 +8061,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(1)* %out) {
 entry:
@@ -8173,10 +8173,10 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_monotonic_atomicrmw:
@@ -8184,10 +8184,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8302,9 +8302,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -8315,10 +8314,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8431,12 +8430,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_release_atomicrmw:
@@ -8444,10 +8443,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8566,9 +8565,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -8581,10 +8579,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8703,9 +8701,8 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
@@ -8718,10 +8715,10 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -8851,13 +8848,13 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acquire_ret_atomicrmw:
@@ -8865,12 +8862,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9005,15 +9002,15 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acq_rel_ret_atomicrmw:
@@ -9021,12 +9018,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9161,15 +9158,15 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_seq_cst_ret_atomicrmw:
@@ -9177,12 +9174,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, 0 :: v_dual_mov_b32 v1, s2
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v0, v1, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in) {
 entry:
@@ -9301,9 +9298,9 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_monotonic_monotonic_cmpxchg:
@@ -9313,9 +9310,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9440,8 +9437,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -9454,9 +9450,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9579,11 +9575,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_release_monotonic_cmpxchg:
@@ -9593,9 +9589,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9724,8 +9720,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -9740,9 +9735,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -9871,8 +9866,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -9887,9 +9881,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10014,8 +10008,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10028,9 +10021,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10155,8 +10148,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -10169,9 +10161,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10300,8 +10292,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10316,9 +10307,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10447,8 +10438,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10463,9 +10453,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10594,8 +10584,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10610,9 +10599,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10741,8 +10730,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10757,9 +10745,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -10888,8 +10876,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -10904,9 +10891,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11035,8 +11022,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11051,9 +11037,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11182,8 +11168,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11198,9 +11183,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11329,8 +11314,7 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
@@ -11345,9 +11329,9 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v2, v[0:1], s[0:1] offset:16
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11486,11 +11470,11 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_monotonic_monotonic_ret_cmpxchg:
@@ -11500,11 +11484,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11648,12 +11632,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acquire_monotonic_ret_cmpxchg:
@@ -11663,11 +11647,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11812,13 +11796,13 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_release_monotonic_ret_cmpxchg:
@@ -11828,11 +11812,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -11980,14 +11964,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acq_rel_monotonic_ret_cmpxchg:
@@ -11997,11 +11981,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12149,14 +12133,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_seq_cst_monotonic_ret_cmpxchg:
@@ -12166,11 +12150,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12314,12 +12298,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_monotonic_acquire_ret_cmpxchg:
@@ -12329,11 +12313,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12477,12 +12461,12 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acquire_acquire_ret_cmpxchg:
@@ -12492,11 +12476,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12644,14 +12628,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_release_acquire_ret_cmpxchg:
@@ -12661,11 +12645,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12813,14 +12797,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acq_rel_acquire_ret_cmpxchg:
@@ -12830,11 +12814,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -12982,14 +12966,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_seq_cst_acquire_ret_cmpxchg:
@@ -12999,11 +12983,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13151,14 +13135,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_monotonic_seq_cst_ret_cmpxchg:
@@ -13168,11 +13152,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13320,14 +13304,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acquire_seq_cst_ret_cmpxchg:
@@ -13337,11 +13321,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13489,14 +13473,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_release_seq_cst_ret_cmpxchg:
@@ -13506,11 +13490,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13658,14 +13642,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_acq_rel_seq_cst_ret_cmpxchg:
@@ -13675,11 +13659,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
@@ -13827,14 +13811,14 @@
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
 ; GFX11-WGP-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: global_workgroup_one_as_seq_cst_seq_cst_ret_cmpxchg:
@@ -13844,11 +13828,11 @@
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s3
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s2 :: v_dual_mov_b32 v1, s3
 ; GFX11-CU-NEXT:    global_atomic_cmpswap_b32 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %out, i32 %in, i32 %old) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-agent.ll
@@ -119,8 +119,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -130,8 +129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -251,8 +249,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -262,8 +259,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -386,8 +382,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -398,8 +393,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -532,8 +526,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
@@ -546,8 +539,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -650,8 +642,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -659,8 +650,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -759,8 +749,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -768,8 +757,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -878,8 +866,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -889,8 +876,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1000,8 +986,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1011,8 +996,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1112,8 +1096,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1121,8 +1104,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1231,8 +1213,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1242,8 +1223,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1353,8 +1333,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1364,8 +1343,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1485,8 +1463,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1498,8 +1475,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1620,8 +1596,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1633,8 +1608,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1756,8 +1730,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1768,8 +1741,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -1902,8 +1874,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
@@ -1916,8 +1887,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2051,8 +2021,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
@@ -2065,8 +2034,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2178,8 +2146,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2188,8 +2155,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2309,8 +2275,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2321,8 +2286,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2443,8 +2407,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2455,8 +2418,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2587,8 +2549,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2601,8 +2562,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2734,8 +2694,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2748,8 +2707,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2871,8 +2829,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2883,8 +2840,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3005,8 +2961,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3017,8 +2972,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3149,8 +3103,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3163,8 +3116,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3296,8 +3248,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3310,8 +3261,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3443,8 +3393,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3457,8 +3406,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3590,8 +3538,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3604,8 +3551,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3737,8 +3683,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3751,8 +3696,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3884,8 +3828,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3898,8 +3841,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4031,8 +3973,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4045,8 +3986,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4178,8 +4118,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4192,8 +4131,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4323,8 +4261,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4335,8 +4272,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4471,8 +4407,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4484,8 +4419,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4627,8 +4561,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4641,8 +4574,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -4788,8 +4720,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4803,8 +4734,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -4950,8 +4880,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4965,8 +4894,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5102,8 +5030,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5115,8 +5042,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5251,8 +5177,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5264,8 +5189,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5410,8 +5334,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5425,8 +5348,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5572,8 +5494,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5587,8 +5508,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5734,8 +5654,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5749,8 +5668,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5896,8 +5814,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5911,8 +5828,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6058,8 +5974,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6073,8 +5988,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6220,8 +6134,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6235,8 +6148,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6382,8 +6294,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6397,8 +6308,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6544,8 +6454,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6559,8 +6468,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6684,8 +6592,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6695,8 +6602,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6816,8 +6722,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6827,8 +6732,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6948,8 +6852,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6959,8 +6862,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -7080,8 +6982,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -7091,8 +6992,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -7194,8 +7094,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7203,8 +7102,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7303,8 +7201,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7312,8 +7209,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7412,8 +7308,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7421,8 +7316,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7521,8 +7415,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7530,8 +7423,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7630,8 +7522,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7639,8 +7530,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7739,8 +7629,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7748,8 +7637,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7848,8 +7736,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7857,8 +7744,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7957,8 +7843,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7966,8 +7851,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -8066,8 +7950,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -8075,8 +7958,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -8193,8 +8075,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8204,8 +8085,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8325,8 +8205,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8336,8 +8215,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8457,8 +8335,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8468,8 +8345,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8580,8 +8456,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8590,8 +8465,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8701,8 +8575,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8711,8 +8584,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8822,8 +8694,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8832,8 +8703,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8943,8 +8813,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8953,8 +8822,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9064,8 +8932,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9074,8 +8941,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9185,8 +9051,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9195,8 +9060,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9306,8 +9170,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9316,8 +9179,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9427,8 +9289,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9437,8 +9298,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9548,8 +9408,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9558,8 +9417,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9669,8 +9527,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9679,8 +9536,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9790,8 +9646,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9800,8 +9655,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9911,8 +9765,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9921,8 +9774,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10032,8 +9884,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10042,8 +9893,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10153,8 +10003,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10163,8 +10012,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10274,8 +10122,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10284,8 +10131,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10413,8 +10259,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10425,8 +10270,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10558,8 +10402,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10570,8 +10413,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10703,8 +10545,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10715,8 +10556,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10848,8 +10688,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10860,8 +10699,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10993,8 +10831,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11005,8 +10842,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11138,8 +10974,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11150,8 +10985,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11283,8 +11117,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11295,8 +11128,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11428,8 +11260,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11440,8 +11271,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11573,8 +11403,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11585,8 +11414,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11718,8 +11546,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11730,8 +11557,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11863,8 +11689,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11875,8 +11700,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12008,8 +11832,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12020,8 +11843,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12153,8 +11975,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12165,8 +11986,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12298,8 +12118,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12310,8 +12129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12443,8 +12261,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12455,8 +12272,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-nontemporal.ll
@@ -133,12 +133,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: local_nontemporal_load_0:
@@ -146,12 +146,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -291,6 +291,7 @@
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: local_nontemporal_load_1:
@@ -304,6 +305,7 @@
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -435,9 +437,8 @@
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s1, s[2:3], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -448,9 +449,8 @@
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s1, s[2:3], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(3)* %out) {
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-singlethread.ll
@@ -119,8 +119,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -130,8 +129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -251,8 +249,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -262,8 +259,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -383,8 +379,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -394,8 +389,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -515,8 +509,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -526,8 +519,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -629,8 +621,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -638,8 +629,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -738,8 +728,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -747,8 +736,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -847,8 +835,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -856,8 +843,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -956,8 +942,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -965,8 +950,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -1065,8 +1049,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1074,8 +1057,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1174,8 +1156,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1183,8 +1164,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1283,8 +1263,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1292,8 +1271,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1392,8 +1370,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1401,8 +1378,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1501,8 +1477,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1510,8 +1485,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1628,8 +1602,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1639,8 +1612,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -1760,8 +1732,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1771,8 +1742,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -1892,8 +1862,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1903,8 +1872,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -2015,8 +1983,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2025,8 +1992,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2136,8 +2102,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2146,8 +2111,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2257,8 +2221,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2267,8 +2230,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2378,8 +2340,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2388,8 +2349,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2499,8 +2459,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2509,8 +2468,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2620,8 +2578,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2630,8 +2587,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2741,8 +2697,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2751,8 +2706,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2862,8 +2816,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2872,8 +2825,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2983,8 +2935,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2993,8 +2944,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3104,8 +3054,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3114,8 +3063,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3225,8 +3173,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3235,8 +3182,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3346,8 +3292,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3356,8 +3301,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3467,8 +3411,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3477,8 +3420,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3588,8 +3530,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3598,8 +3539,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3709,8 +3649,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3719,8 +3658,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3848,8 +3786,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3860,8 +3797,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3993,8 +3929,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4005,8 +3940,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4138,8 +4072,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4150,8 +4083,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4283,8 +4215,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4295,8 +4226,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4428,8 +4358,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4440,8 +4369,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4573,8 +4501,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4585,8 +4512,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4718,8 +4644,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4730,8 +4655,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4863,8 +4787,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4875,8 +4798,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5008,8 +4930,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5020,8 +4941,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5153,8 +5073,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5165,8 +5084,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5298,8 +5216,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5310,8 +5227,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5443,8 +5359,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5455,8 +5370,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5588,8 +5502,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5600,8 +5513,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5733,8 +5645,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5745,8 +5656,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5878,8 +5788,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5890,8 +5799,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -6014,8 +5922,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6025,8 +5932,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6146,8 +6052,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6157,8 +6062,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6278,8 +6182,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6289,8 +6192,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6410,8 +6312,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6421,8 +6322,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6524,8 +6424,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6533,8 +6432,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6633,8 +6531,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6642,8 +6539,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6742,8 +6638,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6751,8 +6646,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6851,8 +6745,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6860,8 +6753,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6960,8 +6852,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6969,8 +6860,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7069,8 +6959,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7078,8 +6967,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7178,8 +7066,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7187,8 +7074,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7287,8 +7173,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7296,8 +7181,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7396,8 +7280,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7405,8 +7288,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7523,8 +7405,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -7534,8 +7415,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -7655,8 +7535,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -7666,8 +7545,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -7787,8 +7665,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -7798,8 +7675,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -7910,8 +7786,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7920,8 +7795,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8031,8 +7905,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8041,8 +7914,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8152,8 +8024,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8162,8 +8033,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8273,8 +8143,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8283,8 +8152,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8394,8 +8262,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8404,8 +8271,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8515,8 +8381,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8525,8 +8390,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8636,8 +8500,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8646,8 +8509,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8757,8 +8619,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8767,8 +8628,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8878,8 +8738,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8888,8 +8747,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8999,8 +8857,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9009,8 +8866,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9120,8 +8976,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9130,8 +8985,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9241,8 +9095,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9251,8 +9104,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9362,8 +9214,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9372,8 +9223,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9483,8 +9333,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9493,8 +9342,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9604,8 +9452,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9614,8 +9461,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9743,8 +9589,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -9755,8 +9600,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -9888,8 +9732,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -9900,8 +9743,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10033,8 +9875,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10045,8 +9886,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10178,8 +10018,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10190,8 +10029,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10323,8 +10161,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10335,8 +10172,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10468,8 +10304,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10480,8 +10315,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10613,8 +10447,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10625,8 +10458,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10758,8 +10590,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10770,8 +10601,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10903,8 +10733,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10915,8 +10744,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11048,8 +10876,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11060,8 +10887,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11193,8 +11019,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11205,8 +11030,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11338,8 +11162,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11350,8 +11173,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11483,8 +11305,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11495,8 +11316,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11628,8 +11448,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11640,8 +11459,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11773,8 +11591,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11785,8 +11602,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-system.ll
@@ -119,8 +119,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -130,8 +129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -251,8 +249,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -262,8 +259,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -386,8 +382,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -398,8 +393,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -532,8 +526,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
@@ -546,8 +539,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -650,8 +642,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -659,8 +650,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -759,8 +749,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -768,8 +757,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -878,8 +866,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -889,8 +876,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1000,8 +986,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1011,8 +996,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1112,8 +1096,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1121,8 +1104,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1231,8 +1213,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1242,8 +1223,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1353,8 +1333,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1364,8 +1343,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1485,8 +1463,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1498,8 +1475,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1620,8 +1596,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1633,8 +1608,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1756,8 +1730,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1768,8 +1741,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -1902,8 +1874,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
@@ -1916,8 +1887,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2051,8 +2021,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
@@ -2065,8 +2034,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2178,8 +2146,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2188,8 +2155,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2309,8 +2275,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2321,8 +2286,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2443,8 +2407,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2455,8 +2418,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2587,8 +2549,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2601,8 +2562,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2734,8 +2694,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2748,8 +2707,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2871,8 +2829,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2883,8 +2840,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3005,8 +2961,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3017,8 +2972,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3149,8 +3103,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3163,8 +3116,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3296,8 +3248,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3310,8 +3261,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3443,8 +3393,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3457,8 +3406,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3590,8 +3538,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3604,8 +3551,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3737,8 +3683,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3751,8 +3696,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3884,8 +3828,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3898,8 +3841,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4031,8 +3973,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4045,8 +3986,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4178,8 +4118,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4192,8 +4131,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4323,8 +4261,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4335,8 +4272,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4471,8 +4407,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4484,8 +4419,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4627,8 +4561,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4641,8 +4574,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -4788,8 +4720,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4803,8 +4734,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -4950,8 +4880,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4965,8 +4894,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5102,8 +5030,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5115,8 +5042,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5251,8 +5177,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5264,8 +5189,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5410,8 +5334,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5425,8 +5348,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5572,8 +5494,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5587,8 +5508,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5734,8 +5654,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5749,8 +5668,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5896,8 +5814,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5911,8 +5828,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6058,8 +5974,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6073,8 +5988,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6220,8 +6134,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6235,8 +6148,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6382,8 +6294,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6397,8 +6308,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6544,8 +6454,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6559,8 +6468,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6684,8 +6592,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6695,8 +6602,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6816,8 +6722,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6827,8 +6732,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6948,8 +6852,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6959,8 +6862,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -7080,8 +6982,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -7091,8 +6992,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -7194,8 +7094,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7203,8 +7102,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7303,8 +7201,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7312,8 +7209,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7412,8 +7308,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7421,8 +7316,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7521,8 +7415,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7530,8 +7423,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7630,8 +7522,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7639,8 +7530,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7739,8 +7629,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7748,8 +7637,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7848,8 +7736,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7857,8 +7744,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7957,8 +7843,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7966,8 +7851,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -8066,8 +7950,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -8075,8 +7958,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -8193,8 +8075,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8204,8 +8085,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8325,8 +8205,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8336,8 +8215,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8457,8 +8335,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8468,8 +8345,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8580,8 +8456,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8590,8 +8465,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8701,8 +8575,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8711,8 +8584,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8822,8 +8694,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8832,8 +8703,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8943,8 +8813,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8953,8 +8822,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9064,8 +8932,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9074,8 +8941,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9185,8 +9051,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9195,8 +9060,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9306,8 +9170,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9316,8 +9179,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9427,8 +9289,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9437,8 +9298,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9548,8 +9408,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9558,8 +9417,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9669,8 +9527,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9679,8 +9536,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9790,8 +9646,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9800,8 +9655,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9911,8 +9765,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9921,8 +9774,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10032,8 +9884,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10042,8 +9893,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10153,8 +10003,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10163,8 +10012,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10274,8 +10122,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10284,8 +10131,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10413,8 +10259,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10425,8 +10270,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10558,8 +10402,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10570,8 +10413,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10703,8 +10545,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10715,8 +10556,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10848,8 +10688,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10860,8 +10699,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10993,8 +10831,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11005,8 +10842,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11138,8 +10974,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11150,8 +10985,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11283,8 +11117,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11295,8 +11128,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11428,8 +11260,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11440,8 +11271,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11573,8 +11403,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11585,8 +11414,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11718,8 +11546,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11730,8 +11557,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11863,8 +11689,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11875,8 +11700,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12008,8 +11832,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12020,8 +11843,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12153,8 +11975,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12165,8 +11986,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12298,8 +12118,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12310,8 +12129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12443,8 +12261,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12455,8 +12272,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-volatile.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-volatile.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-volatile.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-volatile.ll
@@ -81,12 +81,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: local_volatile_load_0:
@@ -94,12 +94,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s2
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_mov_b32 v0, s2
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -191,6 +191,7 @@
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: local_volatile_load_1:
@@ -204,6 +205,7 @@
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -287,9 +289,8 @@
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s1, s[2:3], 0x0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -300,9 +301,8 @@
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s1, s[2:3], 0x0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(3)* %out) {
@@ -480,8 +480,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -492,8 +491,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -565,8 +563,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -576,8 +573,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-wavefront.ll
@@ -119,8 +119,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -130,8 +129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -251,8 +249,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -262,8 +259,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -383,8 +379,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -394,8 +389,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -515,8 +509,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -526,8 +519,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -629,8 +621,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -638,8 +629,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -738,8 +728,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -747,8 +736,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -847,8 +835,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -856,8 +843,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -956,8 +942,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -965,8 +950,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -1065,8 +1049,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1074,8 +1057,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1174,8 +1156,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1183,8 +1164,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1283,8 +1263,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1292,8 +1271,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1392,8 +1370,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1401,8 +1378,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1501,8 +1477,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1510,8 +1485,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1628,8 +1602,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1639,8 +1612,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -1760,8 +1732,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1771,8 +1742,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -1892,8 +1862,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1903,8 +1872,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -2015,8 +1983,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2025,8 +1992,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2136,8 +2102,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2146,8 +2111,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2257,8 +2221,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2267,8 +2230,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2378,8 +2340,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2388,8 +2349,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2499,8 +2459,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2509,8 +2468,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2620,8 +2578,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2630,8 +2587,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2741,8 +2697,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2751,8 +2706,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2862,8 +2816,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2872,8 +2825,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2983,8 +2935,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2993,8 +2944,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3104,8 +3054,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3114,8 +3063,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3225,8 +3173,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3235,8 +3182,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3346,8 +3292,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3356,8 +3301,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3467,8 +3411,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3477,8 +3420,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3588,8 +3530,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3598,8 +3539,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3709,8 +3649,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3719,8 +3658,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3848,8 +3786,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3860,8 +3797,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3993,8 +3929,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4005,8 +3940,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4138,8 +4072,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4150,8 +4083,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4283,8 +4215,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4295,8 +4226,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4428,8 +4358,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4440,8 +4369,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4573,8 +4501,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4585,8 +4512,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4718,8 +4644,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4730,8 +4655,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4863,8 +4787,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4875,8 +4798,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5008,8 +4930,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5020,8 +4941,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5153,8 +5073,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5165,8 +5084,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5298,8 +5216,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5310,8 +5227,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5443,8 +5359,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5455,8 +5370,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5588,8 +5502,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5600,8 +5513,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5733,8 +5645,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5745,8 +5656,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5878,8 +5788,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5890,8 +5799,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -6014,8 +5922,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6025,8 +5932,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6146,8 +6052,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6157,8 +6062,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6278,8 +6182,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6289,8 +6192,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6410,8 +6312,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6421,8 +6322,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6524,8 +6424,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6533,8 +6432,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6633,8 +6531,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6642,8 +6539,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6742,8 +6638,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6751,8 +6646,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6851,8 +6745,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6860,8 +6753,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -6960,8 +6852,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -6969,8 +6860,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7069,8 +6959,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7078,8 +6967,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7178,8 +7066,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7187,8 +7074,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7287,8 +7173,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7296,8 +7181,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7396,8 +7280,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7405,8 +7288,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7523,8 +7405,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -7534,8 +7415,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -7655,8 +7535,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -7666,8 +7545,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -7787,8 +7665,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -7798,8 +7675,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -7910,8 +7786,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7920,8 +7795,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8031,8 +7905,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8041,8 +7914,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8152,8 +8024,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8162,8 +8033,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8273,8 +8143,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8283,8 +8152,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8394,8 +8262,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8404,8 +8271,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8515,8 +8381,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8525,8 +8390,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8636,8 +8500,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8646,8 +8509,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8757,8 +8619,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8767,8 +8628,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8878,8 +8738,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8888,8 +8747,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8999,8 +8857,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9009,8 +8866,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9120,8 +8976,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9130,8 +8985,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9241,8 +9095,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9251,8 +9104,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9362,8 +9214,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9372,8 +9223,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9483,8 +9333,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9493,8 +9342,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9604,8 +9452,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9614,8 +9461,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9743,8 +9589,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -9755,8 +9600,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -9888,8 +9732,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -9900,8 +9743,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10033,8 +9875,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10045,8 +9886,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10178,8 +10018,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10190,8 +10029,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10323,8 +10161,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10335,8 +10172,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10468,8 +10304,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10480,8 +10315,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10613,8 +10447,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10625,8 +10458,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10758,8 +10590,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10770,8 +10601,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10903,8 +10733,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10915,8 +10744,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11048,8 +10876,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11060,8 +10887,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11193,8 +11019,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11205,8 +11030,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11338,8 +11162,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11350,8 +11173,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11483,8 +11305,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11495,8 +11316,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11628,8 +11448,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11640,8 +11459,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11773,8 +11591,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11785,8 +11602,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-local-workgroup.ll
@@ -119,8 +119,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -130,8 +129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -251,8 +249,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -262,8 +259,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -386,8 +382,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -398,8 +393,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -532,8 +526,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
@@ -546,8 +539,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -650,8 +642,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -659,8 +650,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -759,8 +749,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -768,8 +757,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -878,8 +866,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -889,8 +876,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1000,8 +986,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -1011,8 +996,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1112,8 +1096,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -1121,8 +1104,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -1231,8 +1213,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1242,8 +1223,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1353,8 +1333,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1364,8 +1343,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1485,8 +1463,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1498,8 +1475,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1620,8 +1596,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
@@ -1633,8 +1608,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -1756,8 +1730,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1768,8 +1741,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -1902,8 +1874,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
@@ -1916,8 +1887,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2051,8 +2021,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
@@ -2065,8 +2034,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2178,8 +2146,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2188,8 +2155,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2309,8 +2275,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2321,8 +2286,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2443,8 +2407,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2455,8 +2418,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2587,8 +2549,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2601,8 +2562,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2734,8 +2694,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -2748,8 +2707,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -2871,8 +2829,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2883,8 +2840,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3005,8 +2961,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3017,8 +2972,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3149,8 +3103,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3163,8 +3116,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3296,8 +3248,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3310,8 +3261,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3443,8 +3393,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3457,8 +3406,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3590,8 +3538,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3604,8 +3551,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3737,8 +3683,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3751,8 +3696,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -3884,8 +3828,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -3898,8 +3841,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4031,8 +3973,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4045,8 +3986,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4178,8 +4118,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4192,8 +4131,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
@@ -4323,8 +4261,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4335,8 +4272,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4471,8 +4407,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4484,8 +4419,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4627,8 +4561,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4641,8 +4574,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -4788,8 +4720,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4803,8 +4734,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -4950,8 +4880,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -4965,8 +4894,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5102,8 +5030,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5115,8 +5042,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5251,8 +5177,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5264,8 +5189,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -5410,8 +5334,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5425,8 +5348,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5572,8 +5494,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5587,8 +5508,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5734,8 +5654,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5749,8 +5668,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -5896,8 +5814,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -5911,8 +5828,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6058,8 +5974,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6073,8 +5988,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6220,8 +6134,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6235,8 +6148,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6382,8 +6294,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6397,8 +6308,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6544,8 +6454,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
@@ -6559,8 +6468,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
@@ -6684,8 +6592,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6695,8 +6602,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6816,8 +6722,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6827,8 +6732,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -6948,8 +6852,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -6959,8 +6862,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -7080,8 +6982,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_load_b32 v0, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v1, v0
@@ -7091,8 +6992,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_load_b32 v0, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v1, v0
@@ -7194,8 +7094,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7203,8 +7102,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7303,8 +7201,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7312,8 +7209,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7412,8 +7308,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7421,8 +7316,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7521,8 +7415,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7530,8 +7423,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 %in, i32 addrspace(3)* %out) {
@@ -7630,8 +7522,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7639,8 +7530,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7739,8 +7629,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7748,8 +7637,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7848,8 +7736,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7857,8 +7744,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -7957,8 +7843,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -7966,8 +7851,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -8066,8 +7950,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
@@ -8075,8 +7958,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v0, v0, v1
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(3)* %out, i32 %in) {
@@ -8193,8 +8075,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8204,8 +8085,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8325,8 +8205,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8336,8 +8215,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8457,8 +8335,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    ds_store_b32 v0, v1
@@ -8468,8 +8345,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    ds_storexchg_rtn_b32 v1, v0, v1
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    ds_store_b32 v0, v1
@@ -8580,8 +8456,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8590,8 +8465,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8701,8 +8575,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8711,8 +8584,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8822,8 +8694,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8832,8 +8703,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8943,8 +8813,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8953,8 +8822,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9064,8 +8932,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9074,8 +8941,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9185,8 +9051,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9195,8 +9060,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9306,8 +9170,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9316,8 +9179,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9427,8 +9289,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9437,8 +9298,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9548,8 +9408,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9558,8 +9417,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9669,8 +9527,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9679,8 +9536,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9790,8 +9646,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9800,8 +9655,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9911,8 +9765,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9921,8 +9774,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10032,8 +9884,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10042,8 +9893,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10153,8 +10003,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10163,8 +10012,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10274,8 +10122,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10284,8 +10131,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_b32 v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10413,8 +10259,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10425,8 +10270,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10558,8 +10402,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10570,8 +10413,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10703,8 +10545,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10715,8 +10556,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10848,8 +10688,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10860,8 +10699,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -10993,8 +10831,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11005,8 +10842,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11138,8 +10974,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11150,8 +10985,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11283,8 +11117,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11295,8 +11128,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11428,8 +11260,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11440,8 +11271,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11573,8 +11403,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11585,8 +11414,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11718,8 +11546,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11730,8 +11557,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11863,8 +11689,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -11875,8 +11700,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12008,8 +11832,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12020,8 +11843,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12153,8 +11975,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12165,8 +11986,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12298,8 +12118,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12310,8 +12129,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12443,8 +12261,7 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-WGP-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
@@ -12455,8 +12272,7 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v2, s2
 ; GFX11-CU-NEXT:    ds_cmpstore_rtn_b32 v1, v0, v1, v2 offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll
@@ -162,6 +162,7 @@
 ; GFX11-WGP-NEXT:    scratch_load_b32 v0, off, s2 slc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_nontemporal_load_0:
@@ -174,6 +175,7 @@
 ; GFX11-CU-NEXT:    scratch_load_b32 v0, off, s2 slc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -331,25 +333,25 @@
 ; GFX11-WGP-LABEL: private_nontemporal_load_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    scratch_load_b32 v0, v0, s2 slc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_nontemporal_load_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    scratch_load_b32 v0, v0, s2 slc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -509,6 +511,7 @@
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
 ; GFX11-WGP-NEXT:    scratch_store_b32 off, v0, s0 glc slc dlc
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_nontemporal_store_0:
@@ -521,6 +524,7 @@
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
 ; GFX11-CU-NEXT:    scratch_store_b32 off, v0, s0 glc slc dlc
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
 entry:
@@ -678,12 +682,12 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s1, s[2:3], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    scratch_store_b32 v0, v1, s0 glc slc dlc
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_nontemporal_store_1:
@@ -691,12 +695,12 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s1, s[2:3], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    scratch_store_b32 v0, v1, s0 glc slc dlc
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-volatile.ll b/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-volatile.ll
--- a/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-volatile.ll
+++ b/llvm/test/CodeGen/AMDGPU/memory-legalizer-private-volatile.ll
@@ -106,6 +106,7 @@
 ; GFX11-WGP-NEXT:    scratch_load_b32 v0, off, s2 glc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_volatile_load_0:
@@ -118,6 +119,7 @@
 ; GFX11-CU-NEXT:    scratch_load_b32 v0, off, s2 glc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -221,25 +223,25 @@
 ; GFX11-WGP-LABEL: private_volatile_load_1:
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    scratch_load_b32 v0, v0, s2 glc dlc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_volatile_load_1:
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b32 s2, s[0:1], 0x0
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_load_b64 s[0:1], s[0:1], 0x8
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    scratch_load_b32 v0, v0, s2 glc dlc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(5)* %in, i32 addrspace(1)* %out) {
 entry:
@@ -353,6 +355,7 @@
 ; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s1
 ; GFX11-WGP-NEXT:    scratch_store_b32 off, v0, s0 dlc
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_volatile_store_0:
@@ -366,6 +369,7 @@
 ; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s1
 ; GFX11-CU-NEXT:    scratch_store_b32 off, v0, s0 dlc
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
 entry:
@@ -474,13 +478,13 @@
 ; GFX11-WGP-NEXT:    s_clause 0x1
 ; GFX11-WGP-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_load_b32 s0, s[0:1], 0x8
-; GFX11-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-WGP-NEXT:    s_load_b32 s1, s[2:3], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-WGP-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-WGP-NEXT:    scratch_store_b32 v0, v1, s0 dlc
 ; GFX11-WGP-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-WGP-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
 ;
 ; GFX11-CU-LABEL: private_volatile_store_1:
@@ -488,13 +492,13 @@
 ; GFX11-CU-NEXT:    s_clause 0x1
 ; GFX11-CU-NEXT:    s_load_b64 s[2:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_load_b32 s0, s[0:1], 0x8
-; GFX11-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-CU-NEXT:    s_load_b32 s1, s[2:3], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-CU-NEXT:    v_dual_mov_b32 v1, s1 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-CU-NEXT:    scratch_store_b32 v0, v1, s0 dlc
 ; GFX11-CU-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
     i32 addrspace(1)* %in, i32 addrspace(5)* %out) {
 entry:
diff --git a/llvm/test/CodeGen/AMDGPU/verify-vopd.mir b/llvm/test/CodeGen/AMDGPU/verify-vopd.mir
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/verify-vopd.mir
@@ -0,0 +1,11 @@
+# RUN: not --crash llc -march=amdgcn -mcpu=gfx1100 -mattr=+wavefrontsize32,-wavefrontsize64 -run-pass machineverifier -o /dev/null %s 2>&1 | FileCheck -check-prefix=GFX11-ERR %s
+
+# GFX11-ERR: *** Bad machine code: VOP* instruction violates constant bus restriction ***
+# GFX11-ERR: $vgpr2, $vgpr3 = V_DUAL_CNDMASK_B32_e32_X_MUL_F32_e32 $sgpr0, $vgpr0, $sgpr1, $vgpr1, implicit $exec, implicit $mode, implicit $vcc, implicit $vcc_lo
+---
+name: vopd_cndmask_2sgpr
+body:            |
+  bb.0:
+    liveins: $sgpr0, $sgpr1, $vgpr0, $vgpr1
+    $vgpr2, $vgpr3 = V_DUAL_CNDMASK_B32_e32_X_MUL_F32_e32 $sgpr0, $vgpr0, $sgpr1, $vgpr1, implicit $exec, implicit $mode, implicit $vcc, implicit $vcc_lo
+...
diff --git a/llvm/test/CodeGen/AMDGPU/vopd-combine.mir b/llvm/test/CodeGen/AMDGPU/vopd-combine.mir
new file mode 100644
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/vopd-combine.mir
@@ -0,0 +1,543 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -run-pass=postmisched %s -o - | FileCheck -check-prefix=SCHED %s
+# RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs -run-pass=postmisched,gcn-create-vopd %s -o - | FileCheck -check-prefix=PAIR %s
+
+--- |
+  @lds = external addrspace(3) global [8 x i8]
+  define void @vopd_schedule() { ret void }
+  define void @vopd_fmamk() { ret void }
+  define void @vopd_fmamk_fail() { ret void }
+  define void @vopd_cndmask() { ret void }
+  define void @vopd_mov() { ret void }
+  define void @vopd_mov_mov() { ret void }
+  define void @vopd_constants_fail() { ret void }
+  define void @vopd_constants_inlinable() { ret void }
+  define void @vopd_constants_same() { ret void }
+  define void @vopd_mov_fmaak_constants_same() { ret void }
+  define void @vopd_debug() { ret void }
+  define void @vopd_schedule_unconstrained() { ret void }
+  define void @vopd_schedule_unconstrained_2() { ret void }
+  define void @vopd_mov_fixup() { ret void }
+  define void @vopd_mov_fixup_fail() { ret void }
+...
+
+---
+name:            vopd_schedule
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_schedule
+    ; SCHED: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr4 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_schedule
+    ; PAIR: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr3, $vgpr6 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32 $vgpr1, $vgpr1, killed $vgpr0, $vgpr0, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr4 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; can fuse vgpr3 and vgpr6 writing insts only due to reg constraints
+    $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr4 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr6 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_fmamk
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_fmamk
+    ; SCHED: $vgpr2 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr5 = V_FMAMK_F32 killed $vgpr0, 10, killed $vgpr3, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 killed $vgpr1, $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_fmamk
+    ; PAIR: $vgpr2 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr5, $vgpr2 = V_DUAL_FMAMK_F32_X_FMAC_F32_e32 killed $vgpr0, 10, killed $vgpr3, killed $vgpr1, $vgpr1, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = IMPLICIT_DEF
+    $vgpr3 = IMPLICIT_DEF
+    ; should pair
+    $vgpr2 = V_FMAC_F32_e32 $vgpr1, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr5 = V_FMAMK_F32 $vgpr0, 10, $vgpr3, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_fmamk_fail
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_fmamk_fail
+    ; SCHED: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr2 = V_XOR_B32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr4 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr5 = V_FMAMK_F32 killed $vgpr0, 10, killed $vgpr4, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 killed $vgpr1, $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_fmamk_fail
+    ; PAIR: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr2 = V_XOR_B32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr4 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr5 = V_FMAMK_F32 killed $vgpr0, 10, killed $vgpr4, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr2 = V_FMAC_F32_e32 killed $vgpr1, $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = V_XOR_B32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr4 = IMPLICIT_DEF
+    ; should not pair
+    $vgpr2 = V_FMAC_F32_e32 $vgpr1, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr5 = V_FMAMK_F32 $vgpr0, 10, $vgpr4, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_cndmask
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    ; SCHED-LABEL: name: vopd_cndmask
+    ; SCHED: $vgpr2 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; SCHED-NEXT: $sgpr20 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr4 = V_FMAMK_F32 $sgpr20, 12345, $vgpr3, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 $sgpr20, killed $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr5 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr7 = V_CNDMASK_B32_e32 killed $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr6 = V_ADD_F32_e32 $sgpr20, $vgpr3, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr9 = V_CNDMASK_B32_e32 killed $sgpr20, killed $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-LABEL: name: vopd_cndmask
+    ; PAIR: $vgpr2 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; PAIR-NEXT: $sgpr20 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr4 = V_FMAMK_F32 $sgpr20, 12345, $vgpr3, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr2, $vgpr5 = V_DUAL_FMAC_F32_e32_X_CNDMASK_B32_e32 $sgpr20, killed $vgpr1, $vgpr0, $vgpr3, implicit $exec, implicit $vcc, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr7 = V_CNDMASK_B32_e32 killed $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr6 = V_ADD_F32_e32 $sgpr20, $vgpr3, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr9 = V_CNDMASK_B32_e32 killed $sgpr20, killed $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = IMPLICIT_DEF
+    $vgpr3 = IMPLICIT_DEF
+    $sgpr20 = IMPLICIT_DEF
+    ; should pair
+    $vgpr2 = V_FMAC_F32_e32 $sgpr20, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr5 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; should not pair, uses 3 scalars (implicit vcc)
+    $vgpr4 = V_FMAMK_F32 $sgpr20, 12345, $vgpr3, implicit $mode, implicit $exec
+    $vgpr7 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; should not pair, uses 3 scalars (implicit vcc)
+    $vgpr6 = V_ADD_F32_e32 $sgpr20, $vgpr3, implicit $mode, implicit $exec
+    $vgpr9 = V_CNDMASK_B32_e32 $sgpr20, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+
+...
+
+---
+name:            vopd_mov
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_mov
+    ; SCHED: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr2 = V_MOV_B32_e32 killed $vgpr0, implicit $exec
+    ; SCHED-NEXT: $vgpr3 = V_ADD_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_mov
+    ; PAIR: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr2, $vgpr3 = V_DUAL_MOV_B32_e32_X_ADD_F32_e32 killed $vgpr0, killed $vgpr1, $vgpr1, implicit $exec, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = V_MOV_B32_e32 $vgpr0, implicit $exec
+    $vgpr3 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_mov_mov
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_mov_mov
+    ; SCHED: $sgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $sgpr7 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr2 = V_MOV_B32_e32 killed $sgpr0, implicit $exec
+    ; SCHED-NEXT: $vgpr3 = V_MOV_B32_e32 killed $sgpr7, implicit $exec
+    ; PAIR-LABEL: name: vopd_mov_mov
+    ; PAIR: $sgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $sgpr7 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr2, $vgpr3 = V_DUAL_MOV_B32_e32_X_MOV_B32_e32 killed $sgpr0, killed $sgpr7, implicit $exec, implicit $exec, implicit $exec
+    $sgpr0 = IMPLICIT_DEF
+    $sgpr7 = IMPLICIT_DEF
+    $vgpr2 = V_MOV_B32_e32 $sgpr0, implicit $exec
+    $vgpr3 = V_MOV_B32_e32 $sgpr7, implicit $exec
+
+...
+
+
+---
+name:            vopd_constants_fail
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_constants_fail
+    ; SCHED: $vgpr2 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr5 = V_FMAMK_F32 killed $vgpr0, 100, killed $vgpr3, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 99, killed $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_constants_fail
+    ; PAIR: $vgpr2 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr5 = V_FMAMK_F32 killed $vgpr0, 100, killed $vgpr3, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr2 = V_FMAC_F32_e32 99, killed $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = IMPLICIT_DEF
+    $vgpr3 = IMPLICIT_DEF
+    ; should not pair with two different literals
+    $vgpr2 = V_FMAC_F32_e32 99, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr5 = V_FMAMK_F32 $vgpr0, 100, $vgpr3, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_constants_inlinable
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    ; SCHED-LABEL: name: vopd_constants_inlinable
+    ; SCHED: $vgpr2 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr5 = V_FMAMK_F32 killed $vgpr0, 100, killed $vgpr3, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 4, killed $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_constants_inlinable
+    ; PAIR: $vgpr2 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr5, $vgpr2 = V_DUAL_FMAMK_F32_X_FMAC_F32_e32 killed $vgpr0, 100, killed $vgpr3, 4, killed $vgpr1, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = IMPLICIT_DEF
+    $vgpr3 = IMPLICIT_DEF
+    ; can pair since 4 is inlinable
+    $vgpr2 = V_FMAC_F32_e32 4, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr5 = V_FMAMK_F32 $vgpr0, 100, $vgpr3, implicit $mode, implicit $exec
+
+...
+
+
+---
+name:            vopd_constants_same
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_constants_same
+    ; SCHED: $vgpr2 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr5 = V_FMAMK_F32 killed $vgpr0, 100, killed $vgpr3, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 100, killed $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_constants_same
+    ; PAIR: $vgpr2 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr5, $vgpr2 = V_DUAL_FMAMK_F32_X_FMAC_F32_e32 killed $vgpr0, 100, killed $vgpr3, 100, killed $vgpr1, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = IMPLICIT_DEF
+    $vgpr3 = IMPLICIT_DEF
+    ; should be able to pair using 1 deduplicated literal
+    $vgpr2 = V_FMAC_F32_e32 100, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr5 = V_FMAMK_F32 $vgpr0, 100, $vgpr3, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_mov_fmaak_constants_same
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_mov_fmaak_constants_same
+    ; SCHED: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $sgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = V_MOV_B32_e32 981467136, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAAK_F32 killed $sgpr0, killed $vgpr0, 981467136, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_mov_fmaak_constants_same
+    ; PAIR: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $sgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1, $vgpr2 = V_DUAL_MOV_B32_e32_X_FMAAK_F32 981467136, killed $sgpr0, killed $vgpr0, 981467136, implicit $exec, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $sgpr0 = IMPLICIT_DEF
+    ; should be able to pair using 1 deduplicated literal
+    $vgpr1 = V_MOV_B32_e32 981467136, implicit $exec
+    $vgpr2 = V_FMAAK_F32 $sgpr0, $vgpr0, 981467136, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_debug
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_debug
+    ; SCHED: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: DBG_VALUE 0
+    ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_debug
+    ; PAIR: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3, $vgpr6 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32 killed $vgpr1, $vgpr1, killed $vgpr0, $vgpr0, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    ; PAIR-NEXT: DBG_VALUE 0
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    ; TODO Debug values disable VOPD creation
+    $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    DBG_VALUE 0
+    $vgpr6 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_schedule_unconstrained
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_schedule_unconstrained
+    ; SCHED: $vgpr2 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr4 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 10, $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr12 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr19 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr11 = V_CNDMASK_B32_e32 $vgpr0, killed $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr17 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr10 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr15 = V_CNDMASK_B32_e32 $vgpr1, killed $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr16 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr14 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_schedule_unconstrained
+    ; PAIR: $vgpr2 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr4 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr3, $vgpr6 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32 $vgpr1, $vgpr1, $vgpr0, $vgpr0, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr2 = V_FMAC_F32_e32 10, $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr12, $vgpr19 = V_DUAL_ADD_F32_e32_X_CNDMASK_B32_e32 $vgpr1, $vgpr1, $vgpr0, $vgpr3, implicit $exec, implicit $vcc, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr11 = V_CNDMASK_B32_e32 $vgpr0, killed $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr17, $vgpr10 = V_DUAL_MUL_F32_e32_X_CNDMASK_B32_e32 killed $vgpr0, $vgpr0, $vgpr1, $vgpr2, implicit $exec, implicit $vcc, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr15 = V_CNDMASK_B32_e32 $vgpr1, killed $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr16 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr14 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = IMPLICIT_DEF
+    $vgpr3 = IMPLICIT_DEF
+    $vgpr2 = V_FMAC_F32_e32 10, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr10 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr4 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr6 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; $vgpr9 = V_FMAMK_F32 $vgpr0, 10, $vgpr2, implicit $mode, implicit $exec
+    $vgpr11 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr19 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr12 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr17 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; $vgpr18 = V_FMAMK_F32 $vgpr0, 10, $vgpr3, implicit $mode, implicit $exec
+    ; $vgpr11 = V_FMAC_F32_e32 10, $vgpr1, $vgpr11, implicit $mode, implicit $exec
+    $vgpr15 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr16 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr14 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+
+...
+
+---
+name:            vopd_schedule_unconstrained_2
+tracksRegLiveness: true
+body:             |
+  bb.0:
+
+    ; SCHED-LABEL: name: vopd_schedule_unconstrained_2
+    ; SCHED: $vgpr2 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr20 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr16 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr35 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr6 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_FMAC_F32_e32 10, $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr4 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr29 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr19 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr20 = V_FMAC_F32_e32 10, $vgpr1, killed $vgpr20, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr15 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr10 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr17 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr11 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr12 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr37 = V_CNDMASK_B32_e32 $vgpr0, killed $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr14 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr20 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr21 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr24 = V_MUL_F32_e32 killed $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr28 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr22 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr31 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr33 = V_CNDMASK_B32_e32 $vgpr1, killed $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; SCHED-NEXT: $vgpr34 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr32 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-LABEL: name: vopd_schedule_unconstrained_2
+    ; PAIR: $vgpr2 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr3 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr20 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr16, $vgpr35 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32 $vgpr1, $vgpr1, $vgpr0, $vgpr0, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr3, $vgpr6 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32 $vgpr1, $vgpr1, $vgpr0, $vgpr0, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr2 = V_FMAC_F32_e32 10, $vgpr1, killed $vgpr2, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr4, $vgpr29 = V_DUAL_SUB_F32_e32_X_CNDMASK_B32_e32 $vgpr1, $vgpr1, $vgpr0, $vgpr3, implicit $exec, implicit $vcc, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr19, $vgpr20 = V_DUAL_CNDMASK_B32_e32_X_FMAC_F32_e32 $vgpr0, $vgpr3, 10, $vgpr1, implicit $exec, implicit $mode, implicit $vcc, implicit $mode, implicit $exec, implicit $vcc, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr15 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr10, $vgpr17 = V_DUAL_CNDMASK_B32_e32_X_MUL_F32_e32 $vgpr1, $vgpr2, $vgpr0, $vgpr0, implicit $exec, implicit $mode, implicit $vcc, implicit $mode, implicit $exec, implicit $vcc, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr11, $vgpr12 = V_DUAL_CNDMASK_B32_e32_X_ADD_F32_e32 $vgpr0, $vgpr3, $vgpr1, $vgpr1, implicit $exec, implicit $mode, implicit $vcc, implicit $mode, implicit $exec, implicit $vcc, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr37, $vgpr14 = V_DUAL_CNDMASK_B32_e32_X_SUB_F32_e32 $vgpr0, killed $vgpr3, $vgpr1, $vgpr1, implicit $exec, implicit $mode, implicit $vcc, implicit $mode, implicit $exec, implicit $vcc, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr20 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr21, $vgpr24 = V_DUAL_SUB_F32_e32_X_MUL_F32_e32 $vgpr1, $vgpr1, killed $vgpr0, $vgpr0, implicit $exec, implicit $mode, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr28 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr22 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr31 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr33 = V_CNDMASK_B32_e32 $vgpr1, killed $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    ; PAIR-NEXT: $vgpr34 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr32 = V_SUB_F32_e32 killed $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    $vgpr2 = IMPLICIT_DEF
+    $vgpr3 = IMPLICIT_DEF
+    $vgpr2 = V_FMAC_F32_e32 10, $vgpr1, $vgpr2, implicit $mode, implicit $exec
+    $vgpr2 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr3 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr10 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr4 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr6 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; $vgpr9 = V_FMAMK_F32 $vgpr0, 10, $vgpr2, implicit $mode, implicit $exec
+    $vgpr11 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr19 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr12 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr17 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; $vgpr18 = V_FMAMK_F32 $vgpr0, 10, $vgpr3, implicit $mode, implicit $exec
+    ; $vgpr11 = V_FMAC_F32_e32 10, $vgpr1, $vgpr11, implicit $mode, implicit $exec
+    $vgpr15 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr16 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr14 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr20 = IMPLICIT_DEF
+    $vgpr20 = V_FMAC_F32_e32 10, $vgpr1, $vgpr20, implicit $mode, implicit $exec
+    $vgpr20 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr21 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr28 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr22 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr24 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; $vgpr9 = V_FMAMK_F32 $vgpr0, 10, $vgpr2, implicit $mode, implicit $exec
+    $vgpr29 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr37 = V_CNDMASK_B32_e32 $vgpr0, $vgpr3, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr31 = V_ADD_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr35 = V_MUL_F32_e32 $vgpr0, $vgpr0, implicit $mode, implicit $exec
+    ; $vgpr18 = V_FMAMK_F32 $vgpr0, 10, $vgpr3, implicit $mode, implicit $exec
+    ; $vgpr11 = V_FMAC_F32_e32 10, $vgpr1, $vgpr11, implicit $mode, implicit $exec
+    $vgpr33 = V_CNDMASK_B32_e32 $vgpr1, $vgpr2, implicit $mode, implicit $exec, implicit $vcc
+    $vgpr34 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+    $vgpr32 = V_SUB_F32_e32 $vgpr1, $vgpr1, implicit $mode, implicit $exec
+
+...
+
+---
+name: vopd_mov_fixup
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; SCHED-LABEL: name: vopd_mov_fixup
+    ; SCHED: $vgpr0 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; SCHED-NEXT: $vgpr2 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    ; SCHED-NEXT: $vgpr3 = V_ADD_F32_e32 killed $vgpr0, killed $vgpr1, implicit $mode, implicit $exec
+    ; SCHED-NEXT: $vgpr4 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    ; SCHED-NEXT: $vgpr5 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    ; PAIR-LABEL: name: vopd_mov_fixup
+    ; PAIR: $vgpr0 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr1 = IMPLICIT_DEF
+    ; PAIR-NEXT: $vgpr2, $vgpr3 = V_DUAL_MOV_B32_e32_X_ADD_F32_e32 target-flags(amdgpu-abs32-lo) @lds, killed $vgpr0, killed $vgpr1, implicit $exec, implicit $mode, implicit $exec, implicit $mode, implicit $exec
+    ; PAIR-NEXT: $vgpr4, $vgpr5 = V_DUAL_MOV_B32_e32_X_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, target-flags(amdgpu-abs32-lo) @lds, implicit $exec, implicit $exec, implicit $exec
+    $vgpr0 = IMPLICIT_DEF
+    $vgpr1 = IMPLICIT_DEF
+    ; should pair
+    $vgpr2 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    $vgpr3 = V_ADD_F32_e32 $vgpr0, $vgpr1, implicit $mode, implicit $exec
+    ; should pair
+    $vgpr4 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    $vgpr5 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+...
+
+---
+name: vopd_mov_fixup_fail
+tracksRegLiveness: true
+body: |
+  bb.0:
+    ; SCHED-LABEL: name: vopd_mov_fixup_fail
+    ; SCHED: $vgpr0 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    ; SCHED-NEXT: $vgpr1 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds + 4, implicit $exec
+    ; PAIR-LABEL: name: vopd_mov_fixup_fail
+    ; PAIR: $vgpr0 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    ; PAIR-NEXT: $vgpr1 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds + 4, implicit $exec
+    $vgpr0 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds, implicit $exec
+    $vgpr1 = V_MOV_B32_e32 target-flags(amdgpu-abs32-lo) @lds + 4, implicit $exec
+...
diff --git a/llvm/test/CodeGen/AMDGPU/wqm-gfx11.ll b/llvm/test/CodeGen/AMDGPU/wqm-gfx11.ll
--- a/llvm/test/CodeGen/AMDGPU/wqm-gfx11.ll
+++ b/llvm/test/CodeGen/AMDGPU/wqm-gfx11.ll
@@ -12,10 +12,8 @@
 ; CHECK-NEXT:    lds_param_load v4, attr0.y wait_vdst:15
 ; CHECK-NEXT:    lds_param_load v5, attr0.z wait_vdst:15
 ; CHECK-NEXT:    s_mov_b32 exec_lo, s0
-; CHECK-NEXT:    s_waitcnt expcnt(2)
-; CHECK-NEXT:    v_add_f32_e32 v0, v3, v0
 ; CHECK-NEXT:    s_waitcnt expcnt(1)
-; CHECK-NEXT:    v_add_f32_e32 v1, v4, v1
+; CHECK-NEXT:    v_dual_add_f32 v0, v3, v0 :: v_dual_add_f32 v1, v4, v1
 ; CHECK-NEXT:    s_waitcnt expcnt(0)
 ; CHECK-NEXT:    v_add_f32_e32 v2, v5, v2
 ; CHECK-NEXT:    ; return to shader part epilog
@@ -43,10 +41,8 @@
 ; CHECK-NEXT:    s_mov_b32 m0, s2
 ; CHECK-NEXT:    lds_direct_load v5 wait_vdst:15
 ; CHECK-NEXT:    s_mov_b32 exec_lo, s0
-; CHECK-NEXT:    s_waitcnt expcnt(2)
-; CHECK-NEXT:    v_add_f32_e32 v0, v3, v0
 ; CHECK-NEXT:    s_waitcnt expcnt(1)
-; CHECK-NEXT:    v_add_f32_e32 v1, v4, v1
+; CHECK-NEXT:    v_dual_add_f32 v0, v3, v0 :: v_dual_add_f32 v1, v4, v1
 ; CHECK-NEXT:    s_waitcnt expcnt(0)
 ; CHECK-NEXT:    v_add_f32_e32 v2, v5, v2
 ; CHECK-NEXT:    ; return to shader part epilog