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);
@@ -920,6 +929,8 @@
     DAG->addMutation(ST.createFillMFMAShadowMutation(DAG->TII));
     DAG->addMutation(createIGroupLPDAGMutation());
     DAG->addMutation(createSchedBarrierDAGMutation());
+    if (isPassEnabled(EnableVOPD, CodeGenOpt::Less))
+      DAG->addMutation(createVOPDPairingMutation());
     return DAG;
   }
 
@@ -1399,6 +1410,8 @@
 }
 
 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 +1,181 @@
+//===- 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 << " from\tX: "
+                      << *Pair.first << "\tY: " << *Pair.second << "\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();
+        llvm::AMDGPU::CanBeVOPD FirstCanBeVOPD = AMDGPU::getCanBeVOPD(Opc);
+        llvm::AMDGPU::CanBeVOPD 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
+                    << "\n\tY: " << SecondMI << "\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
@@ -3982,6 +3982,14 @@
   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);
@@ -4255,9 +4263,9 @@
     // 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
@@ -2950,6 +2950,27 @@
   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
@@ -470,6 +470,14 @@
 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,
@@ -482,6 +490,12 @@
 LLVM_READONLY
 int getMCOpcode(uint16_t Opcode, unsigned Gen);
 
+LLVM_READONLY
+unsigned getVOPDOpcode(unsigned Opc);
+
+LLVM_READONLY
+int getVOPDFull(unsigned OpX, unsigned OpY);
+
 LLVM_READONLY
 unsigned mapWMMA2AddrTo3AddrOpcode(unsigned Opc);
 
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
 #define GET_WMMAOpcode2AddrMappingTable_DECL
 #define GET_WMMAOpcode2AddrMappingTable_IMPL
 #define GET_WMMAOpcode3AddrMappingTable_DECL
@@ -398,6 +414,19 @@
   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;
+}
+
 unsigned mapWMMA2AddrTo3AddrOpcode(unsigned Opc) {
   const WMMAOpcodeMappingInfo *Info = getWMMAMappingInfoFrom2AddrOpcode(Opc);
   return Info ? Info->Opcode3Addr : ~0u;
@@ -415,6 +444,11 @@
   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/GlobalISel/saddsat.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/saddsat.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/saddsat.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/saddsat.ll
@@ -4277,8 +4277,7 @@
 ; GFX11-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[4:5], v[0:1]
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v6, s1
 ; GFX11-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v4, v6, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v4, v6 :: v_dual_cndmask_b32 v1, v5, v1
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %result = call i64 @llvm.sadd.sat.i64(i64 %lhs, i64 %rhs)
   ret i64 %result
@@ -4354,24 +4353,42 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s1, v1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_saddsat_i64:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_add_u32 s4, s0, s2
-; GFX10PLUS-NEXT:    s_addc_u32 s5, s1, s3
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s2, s[2:3], 0
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s1, s[4:5], s[0:1]
-; GFX10PLUS-NEXT:    s_mov_b32 s3, 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s5, 31
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, s4
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, s5
-; GFX10PLUS-NEXT:    s_xor_b32 s2, s2, s1
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s2
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0x80000000
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_saddsat_i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_add_u32 s4, s0, s2
+; GFX10-NEXT:    s_addc_u32 s5, s1, s3
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s2, s[2:3], 0
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s1, s[4:5], s[0:1]
+; GFX10-NEXT:    s_mov_b32 s3, 0
+; GFX10-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX10-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-NEXT:    s_xor_b32 s2, s2, s1
+; GFX10-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s2
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s2
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_saddsat_i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_add_u32 s4, s0, s2
+; GFX11-NEXT:    s_addc_u32 s5, s1, s3
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s2, s[2:3], 0
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s1, s[4:5], s[0:1]
+; GFX11-NEXT:    s_mov_b32 s3, 0
+; GFX11-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX11-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-NEXT:    s_xor_b32 s2, s2, s1
+; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s2
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s2
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call i64 @llvm.sadd.sat.i64(i64 %lhs, i64 %rhs)
   ret i64 %result
 }
@@ -4425,19 +4442,32 @@
 ; GFX9-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: saddsat_i64_sv:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    v_add_co_u32 v2, vcc_lo, s0, v0
-; GFX10PLUS-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
-; GFX10PLUS-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e32 vcc_lo, s[0:1], v[2:3]
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s0, 0, v[0:1]
-; GFX10PLUS-NEXT:    s_mov_b32 s1, 0
-; GFX10PLUS-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v4, s1
-; GFX10PLUS-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: saddsat_i64_sv:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_add_co_u32 v2, vcc_lo, s0, v0
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX10-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc_lo, s[0:1], v[2:3]
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s0, 0, v[0:1]
+; GFX10-NEXT:    s_mov_b32 s1, 0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v4, s1
+; GFX10-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: saddsat_i64_sv:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, s0, v0
+; GFX11-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX11-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc_lo, s[0:1], v[2:3]
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s0, 0, v[0:1]
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v4, s1
+; GFX11-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v4 :: v_dual_cndmask_b32 v1, v3, v1
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call i64 @llvm.sadd.sat.i64(i64 %lhs, i64 %rhs)
   %cast = bitcast i64 %result to <2 x float>
   ret <2 x float> %cast
@@ -4492,19 +4522,32 @@
 ; GFX9-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: saddsat_i64_vs:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    v_add_co_u32 v2, vcc_lo, v0, s0
-; GFX10PLUS-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s1, s[0:1], 0
-; GFX10PLUS-NEXT:    s_mov_b32 s0, 0
-; GFX10PLUS-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[2:3], v[0:1]
-; GFX10PLUS-NEXT:    v_add_co_ci_u32_e64 v1, s0, 0x80000000, v4, s0
-; GFX10PLUS-NEXT:    s_xor_b32 vcc_lo, s1, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: saddsat_i64_vs:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_add_co_u32 v2, vcc_lo, v0, s0
+; GFX10-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s1, s[0:1], 0
+; GFX10-NEXT:    s_mov_b32 s0, 0
+; GFX10-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX10-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[2:3], v[0:1]
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v1, s0, 0x80000000, v4, s0
+; GFX10-NEXT:    s_xor_b32 vcc_lo, s1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: saddsat_i64_vs:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_add_co_u32 v2, vcc_lo, v0, s0
+; GFX11-NEXT:    v_add_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s1, s[0:1], 0
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX11-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[2:3], v[0:1]
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, s0, 0x80000000, v4, s0
+; GFX11-NEXT:    s_xor_b32 vcc_lo, s1, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v4 :: v_dual_cndmask_b32 v1, v3, v1
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call i64 @llvm.sadd.sat.i64(i64 %lhs, i64 %rhs)
   %cast = bitcast i64 %result to <2 x float>
   ret <2 x float> %cast
@@ -4629,11 +4672,9 @@
 ; GFX11-NEXT:    v_cmp_lt_i64_e64 s2, v[10:11], v[2:3]
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, s1, 0x80000000, v4, s1
 ; GFX11-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v8, v0, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v8, v0 :: v_dual_cndmask_b32 v1, v9, v1
 ; GFX11-NEXT:    s_xor_b32 vcc_lo, s3, s2
-; GFX11-NEXT:    v_cndmask_b32_e32 v2, v10, v4, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v3, v11, v3, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v10, v4 :: v_dual_cndmask_b32 v3, v11, v3
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %result = call <2 x i64> @llvm.sadd.sat.v2i64(<2 x i64> %lhs, <2 x i64> %rhs)
   ret <2 x i64> %result
@@ -4766,38 +4807,69 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s3, v1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_saddsat_v2i64:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_add_u32 s8, s0, s4
-; GFX10PLUS-NEXT:    s_addc_u32 s9, s1, s5
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s4, s[4:5], 0
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s1, s[8:9], s[0:1]
-; GFX10PLUS-NEXT:    s_mov_b32 s10, 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s9, 31
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, s8
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, s9
-; GFX10PLUS-NEXT:    s_xor_b32 s8, s4, s1
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s10, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s8
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0x80000000
-; GFX10PLUS-NEXT:    s_add_u32 s4, s2, s6
-; GFX10PLUS-NEXT:    s_addc_u32 s5, s3, s7
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, s4
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s2, s[4:5], s[2:3]
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s3, s[6:7], 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s5, 31
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, s5
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s8
-; GFX10PLUS-NEXT:    s_xor_b32 s2, s3, s2
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s10, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, v2, s0, s2
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0x80000000
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, v3, s1, s2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s2, v2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s3, v3
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_saddsat_v2i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_add_u32 s8, s0, s4
+; GFX10-NEXT:    s_addc_u32 s9, s1, s5
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s4, s[4:5], 0
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s1, s[8:9], s[0:1]
+; GFX10-NEXT:    s_mov_b32 s10, 0
+; GFX10-NEXT:    s_ashr_i32 s0, s9, 31
+; GFX10-NEXT:    v_mov_b32_e32 v0, s8
+; GFX10-NEXT:    v_mov_b32_e32 v1, s9
+; GFX10-NEXT:    s_xor_b32 s8, s4, s1
+; GFX10-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s8
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX10-NEXT:    s_add_u32 s4, s2, s6
+; GFX10-NEXT:    s_addc_u32 s5, s3, s7
+; GFX10-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s2, s[4:5], s[2:3]
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s3, s[6:7], 0
+; GFX10-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX10-NEXT:    v_mov_b32_e32 v3, s5
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s8
+; GFX10-NEXT:    s_xor_b32 s2, s3, s2
+; GFX10-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, s0, s2
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, s1, s2
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX10-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX10-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_saddsat_v2i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_add_u32 s8, s0, s4
+; GFX11-NEXT:    s_addc_u32 s9, s1, s5
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s4, s[4:5], 0
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s1, s[8:9], s[0:1]
+; GFX11-NEXT:    s_mov_b32 s10, 0
+; GFX11-NEXT:    s_ashr_i32 s0, s9, 31
+; GFX11-NEXT:    v_dual_mov_b32 v0, s8 :: v_dual_mov_b32 v1, s9
+; GFX11-NEXT:    s_xor_b32 s8, s4, s1
+; GFX11-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s8
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX11-NEXT:    s_add_u32 s4, s2, s6
+; GFX11-NEXT:    s_addc_u32 s5, s3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s2, s[4:5], s[2:3]
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s3, s[6:7], 0
+; GFX11-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s8
+; GFX11-NEXT:    s_xor_b32 s2, s3, s2
+; GFX11-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, v2, s0, s2
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, v3, s1, s2
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call <2 x i64> @llvm.sadd.sat.v2i64(<2 x i64> %lhs, <2 x i64> %rhs)
   ret <2 x i64> %result
 }
@@ -5871,91 +5943,176 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s7, v3
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_saddsat_v2i128:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_add_u32 s8, s0, s8
-; GFX10PLUS-NEXT:    s_addc_u32 s9, s1, s9
-; GFX10PLUS-NEXT:    s_addc_u32 s16, s2, s10
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s0, s[8:9], s[0:1]
-; GFX10PLUS-NEXT:    s_addc_u32 s17, s3, s11
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[16:17], s[2:3]
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v5, s17
-; GFX10PLUS-NEXT:    s_cselect_b32 s18, 1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s0, s[16:17], s[2:3]
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s2, s[10:11], 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
-; GFX10PLUS-NEXT:    s_and_b32 s0, 1, s18
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[10:11], 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
-; GFX10PLUS-NEXT:    s_cselect_b32 s1, 1, 0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
-; GFX10PLUS-NEXT:    s_and_b32 s1, 1, s1
-; GFX10PLUS-NEXT:    s_ashr_i32 s2, s17, 31
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 s0, 0, s1
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v2, 0, s0
-; GFX10PLUS-NEXT:    s_mov_b32 s0, 0
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s0, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s2, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s10, s2, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s3, s2, 0x80000000
-; GFX10PLUS-NEXT:    s_add_u32 s12, s4, s12
-; GFX10PLUS-NEXT:    s_addc_u32 s13, s5, s13
-; GFX10PLUS-NEXT:    s_addc_u32 s18, s6, s14
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s4, s[12:13], s[4:5]
-; GFX10PLUS-NEXT:    s_addc_u32 s19, s7, s15
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s5, s[14:15], 0
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[18:19], s[6:7]
-; GFX10PLUS-NEXT:    v_xor_b32_e32 v0, v1, v0
-; GFX10PLUS-NEXT:    s_cselect_b32 s0, 1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s4
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s4, s[18:19], s[6:7]
-; GFX10PLUS-NEXT:    s_and_b32 s0, 1, s0
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[14:15], 0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 s0, 0, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s5
-; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s4
-; GFX10PLUS-NEXT:    s_cselect_b32 s4, 1, 0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, s8
-; GFX10PLUS-NEXT:    s_and_b32 s4, 1, s4
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, v3, v2, s0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 s0, 0, s4
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, s9
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v6, s13
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v1, s2, vcc_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v7, s19
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, v4, 0, s0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v4, s16
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, v0, s1, vcc_lo
-; GFX10PLUS-NEXT:    s_mov_b32 s1, 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s19, 31
-; GFX10PLUS-NEXT:    v_xor_b32_e32 v2, v3, v2
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, v4, s10, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v4, v5, s3, vcc_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v5, s12
-; GFX10PLUS-NEXT:    v_and_b32_e32 v2, 1, v2
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s2, s0, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s3, s0, 0x80000000
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, s18
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v5, v5, s0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v6, v6, s1, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, v2, s2, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v7, v7, s3, vcc_lo
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v1
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s2, v3
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s3, v4
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s5, v6
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s6, v2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s7, v7
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_saddsat_v2i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_add_u32 s8, s0, s8
+; GFX10-NEXT:    s_addc_u32 s9, s1, s9
+; GFX10-NEXT:    s_addc_u32 s16, s2, s10
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s0, s[8:9], s[0:1]
+; GFX10-NEXT:    s_addc_u32 s17, s3, s11
+; GFX10-NEXT:    s_cmp_eq_u64 s[16:17], s[2:3]
+; GFX10-NEXT:    v_mov_b32_e32 v5, s17
+; GFX10-NEXT:    s_cselect_b32 s18, 1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s0, s[16:17], s[2:3]
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s2, s[10:11], 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX10-NEXT:    s_and_b32 s0, 1, s18
+; GFX10-NEXT:    s_cmp_eq_u64 s[10:11], 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
+; GFX10-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX10-NEXT:    s_and_b32 s1, 1, s1
+; GFX10-NEXT:    s_ashr_i32 s2, s17, 31
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, s1
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v2, 0, s0
+; GFX10-NEXT:    s_mov_b32 s0, 0
+; GFX10-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX10-NEXT:    s_addc_u32 s1, s2, 0
+; GFX10-NEXT:    s_addc_u32 s10, s2, 0
+; GFX10-NEXT:    s_addc_u32 s3, s2, 0x80000000
+; GFX10-NEXT:    s_add_u32 s12, s4, s12
+; GFX10-NEXT:    s_addc_u32 s13, s5, s13
+; GFX10-NEXT:    s_addc_u32 s18, s6, s14
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s4, s[12:13], s[4:5]
+; GFX10-NEXT:    s_addc_u32 s19, s7, s15
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s5, s[14:15], 0
+; GFX10-NEXT:    s_cmp_eq_u64 s[18:19], s[6:7]
+; GFX10-NEXT:    v_xor_b32_e32 v0, v1, v0
+; GFX10-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s4
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s4, s[18:19], s[6:7]
+; GFX10-NEXT:    s_and_b32 s0, 1, s0
+; GFX10-NEXT:    s_cmp_eq_u64 s[14:15], 0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s5
+; GFX10-NEXT:    v_and_b32_e32 v0, 1, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s4
+; GFX10-NEXT:    s_cselect_b32 s4, 1, 0
+; GFX10-NEXT:    v_mov_b32_e32 v1, s8
+; GFX10-NEXT:    s_and_b32 s4, 1, s4
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v3, v2, s0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, s4
+; GFX10-NEXT:    v_mov_b32_e32 v0, s9
+; GFX10-NEXT:    v_mov_b32_e32 v6, s13
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, s2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v7, s19
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v4, 0, s0
+; GFX10-NEXT:    v_mov_b32_e32 v4, s16
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, s1, vcc_lo
+; GFX10-NEXT:    s_mov_b32 s1, 0
+; GFX10-NEXT:    s_ashr_i32 s0, s19, 31
+; GFX10-NEXT:    v_xor_b32_e32 v2, v3, v2
+; GFX10-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v4, s10, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, v5, s3, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v5, s12
+; GFX10-NEXT:    v_and_b32_e32 v2, 1, v2
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0
+; GFX10-NEXT:    s_addc_u32 s2, s0, 0
+; GFX10-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX10-NEXT:    v_mov_b32_e32 v2, s18
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, v5, s0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v6, s1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, s2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v7, s3, vcc_lo
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v0
+; GFX10-NEXT:    v_readfirstlane_b32 s2, v3
+; GFX10-NEXT:    v_readfirstlane_b32 s3, v4
+; GFX10-NEXT:    v_readfirstlane_b32 s4, v5
+; GFX10-NEXT:    v_readfirstlane_b32 s5, v6
+; GFX10-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX10-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_saddsat_v2i128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_add_u32 s8, s0, s8
+; GFX11-NEXT:    s_addc_u32 s9, s1, s9
+; GFX11-NEXT:    s_addc_u32 s16, s2, s10
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s0, s[8:9], s[0:1]
+; GFX11-NEXT:    s_addc_u32 s17, s3, s11
+; GFX11-NEXT:    s_cmp_eq_u64 s[16:17], s[2:3]
+; GFX11-NEXT:    v_mov_b32_e32 v5, s17
+; GFX11-NEXT:    s_cselect_b32 s18, 1, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s0, s[16:17], s[2:3]
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s2, s[10:11], 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX11-NEXT:    s_and_b32 s0, 1, s18
+; GFX11-NEXT:    s_cmp_eq_u64 s[10:11], 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
+; GFX11-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX11-NEXT:    s_and_b32 s1, 1, s1
+; GFX11-NEXT:    s_ashr_i32 s2, s17, 31
+; GFX11-NEXT:    v_cmp_ne_u32_e64 s0, 0, s1
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v2, 0, s0
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s0, 0
+; GFX11-NEXT:    s_addc_u32 s1, s2, 0
+; GFX11-NEXT:    s_addc_u32 s10, s2, 0
+; GFX11-NEXT:    s_addc_u32 s3, s2, 0x80000000
+; GFX11-NEXT:    s_add_u32 s12, s4, s12
+; GFX11-NEXT:    s_addc_u32 s13, s5, s13
+; GFX11-NEXT:    s_addc_u32 s18, s6, s14
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s4, s[12:13], s[4:5]
+; GFX11-NEXT:    s_addc_u32 s19, s7, s15
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s5, s[14:15], 0
+; GFX11-NEXT:    s_cmp_eq_u64 s[18:19], s[6:7]
+; GFX11-NEXT:    v_xor_b32_e32 v0, v1, v0
+; GFX11-NEXT:    v_mov_b32_e32 v1, s8
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s4
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s4, s[18:19], s[6:7]
+; GFX11-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s5
+; GFX11-NEXT:    s_and_b32 s0, 1, s0
+; GFX11-NEXT:    s_cmp_eq_u64 s[14:15], 0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 s0, 0, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s4
+; GFX11-NEXT:    s_cselect_b32 s4, 1, 0
+; GFX11-NEXT:    v_and_b32_e32 v0, 1, v0
+; GFX11-NEXT:    s_and_b32 s4, 1, s4
+; GFX11-NEXT:    v_mov_b32_e32 v6, s13
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, v3, v2, s0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 s0, 0, s4
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    v_dual_mov_b32 v0, s9 :: v_dual_mov_b32 v7, s19
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, v4, 0, s0
+; GFX11-NEXT:    v_mov_b32_e32 v4, s16
+; GFX11-NEXT:    s_ashr_i32 s0, s19, 31
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, s2, vcc_lo
+; GFX11-NEXT:    v_xor_b32_e32 v2, v3, v2
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, v4, s10, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v4, v5, s3, vcc_lo
+; GFX11-NEXT:    v_mov_b32_e32 v5, s12
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, s1, vcc_lo
+; GFX11-NEXT:    v_and_b32_e32 v2, 1, v2
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v2
+; GFX11-NEXT:    v_mov_b32_e32 v2, s18
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0
+; GFX11-NEXT:    s_addc_u32 s2, s0, 0
+; GFX11-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX11-NEXT:    v_cndmask_b32_e64 v5, v5, s0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v6, v6, s1, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, v2, s2, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v7, v7, s3, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call <2 x i128> @llvm.sadd.sat.v2i128(<2 x i128> %lhs, <2 x i128> %rhs)
   ret <2 x i128> %result
 }
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/ssubsat.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/ssubsat.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/ssubsat.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/ssubsat.ll
@@ -4263,8 +4263,7 @@
 ; GFX11-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[4:5], v[0:1]
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v6, s1
 ; GFX11-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v4, v6, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v1, v5, v1, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v4, v6 :: v_dual_cndmask_b32 v1, v5, v1
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %result = call i64 @llvm.ssub.sat.i64(i64 %lhs, i64 %rhs)
   ret i64 %result
@@ -4340,24 +4339,42 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s1, v1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_ssubsat_i64:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_sub_u32 s4, s0, s2
-; GFX10PLUS-NEXT:    s_subb_u32 s5, s1, s3
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s2, s[2:3], 0
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s1, s[4:5], s[0:1]
-; GFX10PLUS-NEXT:    s_mov_b32 s3, 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s5, 31
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, s4
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, s5
-; GFX10PLUS-NEXT:    s_xor_b32 s2, s2, s1
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s3, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s2
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0x80000000
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_ssubsat_i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_sub_u32 s4, s0, s2
+; GFX10-NEXT:    s_subb_u32 s5, s1, s3
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s2, s[2:3], 0
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s1, s[4:5], s[0:1]
+; GFX10-NEXT:    s_mov_b32 s3, 0
+; GFX10-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX10-NEXT:    v_mov_b32_e32 v0, s4
+; GFX10-NEXT:    v_mov_b32_e32 v1, s5
+; GFX10-NEXT:    s_xor_b32 s2, s2, s1
+; GFX10-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s2
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s2
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_ssubsat_i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_sub_u32 s4, s0, s2
+; GFX11-NEXT:    s_subb_u32 s5, s1, s3
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s2, s[2:3], 0
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s1, s[4:5], s[0:1]
+; GFX11-NEXT:    s_mov_b32 s3, 0
+; GFX11-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX11-NEXT:    v_dual_mov_b32 v0, s4 :: v_dual_mov_b32 v1, s5
+; GFX11-NEXT:    s_xor_b32 s2, s2, s1
+; GFX11-NEXT:    s_cmp_lg_u32 s3, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s2
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s2
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call i64 @llvm.ssub.sat.i64(i64 %lhs, i64 %rhs)
   ret i64 %result
 }
@@ -4411,19 +4428,32 @@
 ; GFX9-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: ssubsat_i64_sv:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    v_sub_co_u32 v2, vcc_lo, s0, v0
-; GFX10PLUS-NEXT:    v_sub_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
-; GFX10PLUS-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e32 vcc_lo, s[0:1], v[2:3]
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s0, 0, v[0:1]
-; GFX10PLUS-NEXT:    s_mov_b32 s1, 0
-; GFX10PLUS-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v4, s1
-; GFX10PLUS-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: ssubsat_i64_sv:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_sub_co_u32 v2, vcc_lo, s0, v0
+; GFX10-NEXT:    v_sub_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX10-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX10-NEXT:    v_cmp_gt_i64_e32 vcc_lo, s[0:1], v[2:3]
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s0, 0, v[0:1]
+; GFX10-NEXT:    s_mov_b32 s1, 0
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v4, s1
+; GFX10-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: ssubsat_i64_sv:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_sub_co_u32 v2, vcc_lo, s0, v0
+; GFX11-NEXT:    v_sub_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX11-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX11-NEXT:    v_cmp_gt_i64_e32 vcc_lo, s[0:1], v[2:3]
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s0, 0, v[0:1]
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, s1, 0x80000000, v4, s1
+; GFX11-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v4 :: v_dual_cndmask_b32 v1, v3, v1
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call i64 @llvm.ssub.sat.i64(i64 %lhs, i64 %rhs)
   %cast = bitcast i64 %result to <2 x float>
   ret <2 x float> %cast
@@ -4478,19 +4508,32 @@
 ; GFX9-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: ssubsat_i64_vs:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    v_sub_co_u32 v2, vcc_lo, v0, s0
-; GFX10PLUS-NEXT:    v_subrev_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s1, s[0:1], 0
-; GFX10PLUS-NEXT:    s_mov_b32 s0, 0
-; GFX10PLUS-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[2:3], v[0:1]
-; GFX10PLUS-NEXT:    v_add_co_ci_u32_e64 v1, s0, 0x80000000, v4, s0
-; GFX10PLUS-NEXT:    s_xor_b32 vcc_lo, s1, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: ssubsat_i64_vs:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    v_sub_co_u32 v2, vcc_lo, v0, s0
+; GFX10-NEXT:    v_subrev_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s1, s[0:1], 0
+; GFX10-NEXT:    s_mov_b32 s0, 0
+; GFX10-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX10-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[2:3], v[0:1]
+; GFX10-NEXT:    v_add_co_ci_u32_e64 v1, s0, 0x80000000, v4, s0
+; GFX10-NEXT:    s_xor_b32 vcc_lo, s1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v2, v4, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v1, vcc_lo
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: ssubsat_i64_vs:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    v_sub_co_u32 v2, vcc_lo, v0, s0
+; GFX11-NEXT:    v_subrev_co_ci_u32_e32 v3, vcc_lo, s1, v1, vcc_lo
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s1, s[0:1], 0
+; GFX11-NEXT:    s_mov_b32 s0, 0
+; GFX11-NEXT:    v_ashrrev_i32_e32 v4, 31, v3
+; GFX11-NEXT:    v_cmp_lt_i64_e32 vcc_lo, v[2:3], v[0:1]
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, s0, 0x80000000, v4, s0
+; GFX11-NEXT:    s_xor_b32 vcc_lo, s1, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v2, v4 :: v_dual_cndmask_b32 v1, v3, v1
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call i64 @llvm.ssub.sat.i64(i64 %lhs, i64 %rhs)
   %cast = bitcast i64 %result to <2 x float>
   ret <2 x float> %cast
@@ -4615,11 +4658,9 @@
 ; GFX11-NEXT:    v_cmp_lt_i64_e64 s2, v[10:11], v[2:3]
 ; GFX11-NEXT:    v_add_co_ci_u32_e64 v3, s1, 0x80000000, v4, s1
 ; GFX11-NEXT:    s_xor_b32 vcc_lo, s0, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v0, v8, v0, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v1, v9, v1, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v0, v8, v0 :: v_dual_cndmask_b32 v1, v9, v1
 ; GFX11-NEXT:    s_xor_b32 vcc_lo, s3, s2
-; GFX11-NEXT:    v_cndmask_b32_e32 v2, v10, v4, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e32 v3, v11, v3, vcc_lo
+; GFX11-NEXT:    v_dual_cndmask_b32 v2, v10, v4 :: v_dual_cndmask_b32 v3, v11, v3
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %result = call <2 x i64> @llvm.ssub.sat.v2i64(<2 x i64> %lhs, <2 x i64> %rhs)
   ret <2 x i64> %result
@@ -4752,38 +4793,69 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s3, v1
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_ssubsat_v2i64:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_sub_u32 s8, s0, s4
-; GFX10PLUS-NEXT:    s_subb_u32 s9, s1, s5
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s4, s[4:5], 0
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s1, s[8:9], s[0:1]
-; GFX10PLUS-NEXT:    s_mov_b32 s10, 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s9, 31
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, s8
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, s9
-; GFX10PLUS-NEXT:    s_xor_b32 s8, s4, s1
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s10, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s8
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0x80000000
-; GFX10PLUS-NEXT:    s_sub_u32 s4, s2, s6
-; GFX10PLUS-NEXT:    s_subb_u32 s5, s3, s7
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, s4
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s2, s[4:5], s[2:3]
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s3, s[6:7], 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s5, 31
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, s5
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s8
-; GFX10PLUS-NEXT:    s_xor_b32 s2, s3, s2
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s10, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, v2, s0, s2
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0x80000000
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, v3, s1, s2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s2, v2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s3, v3
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_ssubsat_v2i64:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_sub_u32 s8, s0, s4
+; GFX10-NEXT:    s_subb_u32 s9, s1, s5
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s4, s[4:5], 0
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s1, s[8:9], s[0:1]
+; GFX10-NEXT:    s_mov_b32 s10, 0
+; GFX10-NEXT:    s_ashr_i32 s0, s9, 31
+; GFX10-NEXT:    v_mov_b32_e32 v0, s8
+; GFX10-NEXT:    v_mov_b32_e32 v1, s9
+; GFX10-NEXT:    s_xor_b32 s8, s4, s1
+; GFX10-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s8
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX10-NEXT:    s_sub_u32 s4, s2, s6
+; GFX10-NEXT:    s_subb_u32 s5, s3, s7
+; GFX10-NEXT:    v_mov_b32_e32 v2, s4
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s2, s[4:5], s[2:3]
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s3, s[6:7], 0
+; GFX10-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX10-NEXT:    v_mov_b32_e32 v3, s5
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s8
+; GFX10-NEXT:    s_xor_b32 s2, s3, s2
+; GFX10-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, s0, s2
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, s1, s2
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX10-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX10-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_ssubsat_v2i64:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_sub_u32 s8, s0, s4
+; GFX11-NEXT:    s_subb_u32 s9, s1, s5
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s4, s[4:5], 0
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s1, s[8:9], s[0:1]
+; GFX11-NEXT:    s_mov_b32 s10, 0
+; GFX11-NEXT:    s_ashr_i32 s0, s9, 31
+; GFX11-NEXT:    v_dual_mov_b32 v0, s8 :: v_dual_mov_b32 v1, s9
+; GFX11-NEXT:    s_xor_b32 s8, s4, s1
+; GFX11-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, s0, s8
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX11-NEXT:    s_sub_u32 s4, s2, s6
+; GFX11-NEXT:    s_subb_u32 s5, s3, s7
+; GFX11-NEXT:    v_dual_mov_b32 v2, s4 :: v_dual_mov_b32 v3, s5
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s2, s[4:5], s[2:3]
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s3, s[6:7], 0
+; GFX11-NEXT:    s_ashr_i32 s0, s5, 31
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, s1, s8
+; GFX11-NEXT:    s_xor_b32 s2, s3, s2
+; GFX11-NEXT:    s_cmp_lg_u32 s10, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, v2, s0, s2
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0x80000000
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, v3, s1, s2
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call <2 x i64> @llvm.ssub.sat.v2i64(<2 x i64> %lhs, <2 x i64> %rhs)
   ret <2 x i64> %result
 }
@@ -4948,52 +5020,98 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s3, v3
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_ssubsat_i128:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_sub_u32 s8, s0, s4
-; GFX10PLUS-NEXT:    s_subb_u32 s9, s1, s5
-; GFX10PLUS-NEXT:    s_subb_u32 s10, s2, s6
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s0, s[8:9], s[0:1]
-; GFX10PLUS-NEXT:    s_subb_u32 s11, s3, s7
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[10:11], s[2:3]
-; GFX10PLUS-NEXT:    s_cselect_b32 s12, 1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s0, s[10:11], s[2:3]
-; GFX10PLUS-NEXT:    v_cmp_gt_u64_e64 s2, s[4:5], 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
-; GFX10PLUS-NEXT:    s_and_b32 s0, 1, s12
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[6:7], 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s2, s[6:7], 0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
-; GFX10PLUS-NEXT:    s_cselect_b32 s1, 1, 0
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s11, 31
-; GFX10PLUS-NEXT:    s_and_b32 s1, 1, s1
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s2
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s1
-; GFX10PLUS-NEXT:    s_mov_b32 s1, 0
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v3, v2, vcc_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, s9
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, s11
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s2, s0, 0
-; GFX10PLUS-NEXT:    v_xor_b32_e32 v0, v1, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, s8
-; GFX10PLUS-NEXT:    s_addc_u32 s3, s0, 0x80000000
-; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, s10
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v1, s0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, v2, s1, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, v3, s3, vcc_lo
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v1
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s2, v0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s3, v3
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_ssubsat_i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_sub_u32 s8, s0, s4
+; GFX10-NEXT:    s_subb_u32 s9, s1, s5
+; GFX10-NEXT:    s_subb_u32 s10, s2, s6
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s0, s[8:9], s[0:1]
+; GFX10-NEXT:    s_subb_u32 s11, s3, s7
+; GFX10-NEXT:    s_cmp_eq_u64 s[10:11], s[2:3]
+; GFX10-NEXT:    s_cselect_b32 s12, 1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s0, s[10:11], s[2:3]
+; GFX10-NEXT:    v_cmp_gt_u64_e64 s2, s[4:5], 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX10-NEXT:    s_and_b32 s0, 1, s12
+; GFX10-NEXT:    s_cmp_eq_u64 s[6:7], 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s2, s[6:7], 0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX10-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX10-NEXT:    s_ashr_i32 s0, s11, 31
+; GFX10-NEXT:    s_and_b32 s1, 1, s1
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s2
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s1
+; GFX10-NEXT:    s_mov_b32 s1, 0
+; GFX10-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v2, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v2, s9
+; GFX10-NEXT:    v_mov_b32_e32 v3, s11
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0
+; GFX10-NEXT:    s_addc_u32 s2, s0, 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, v1, v0
+; GFX10-NEXT:    v_mov_b32_e32 v1, s8
+; GFX10-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX10-NEXT:    v_and_b32_e32 v0, 1, v0
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT:    v_mov_b32_e32 v0, s10
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, s0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, s1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, s3, vcc_lo
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v2
+; GFX10-NEXT:    v_readfirstlane_b32 s2, v0
+; GFX10-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_ssubsat_i128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_sub_u32 s8, s0, s4
+; GFX11-NEXT:    s_subb_u32 s9, s1, s5
+; GFX11-NEXT:    s_subb_u32 s10, s2, s6
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s0, s[8:9], s[0:1]
+; GFX11-NEXT:    s_subb_u32 s11, s3, s7
+; GFX11-NEXT:    s_cmp_eq_u64 s[10:11], s[2:3]
+; GFX11-NEXT:    s_cselect_b32 s12, 1, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s0, s[10:11], s[2:3]
+; GFX11-NEXT:    v_cmp_gt_u64_e64 s2, s[4:5], 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX11-NEXT:    s_and_b32 s0, 1, s12
+; GFX11-NEXT:    s_cmp_eq_u64 s[6:7], 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s2, s[6:7], 0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX11-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX11-NEXT:    s_ashr_i32 s0, s11, 31
+; GFX11-NEXT:    s_and_b32 s1, 1, s1
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s2
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s1
+; GFX11-NEXT:    s_mov_b32 s1, 0
+; GFX11-NEXT:    s_cmp_lg_u32 s1, 0
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v3, v2 :: v_dual_mov_b32 v2, s9
+; GFX11-NEXT:    v_mov_b32_e32 v3, s11
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0
+; GFX11-NEXT:    s_addc_u32 s2, s0, 0
+; GFX11-NEXT:    v_xor_b32_e32 v0, v1, v0
+; GFX11-NEXT:    v_mov_b32_e32 v1, s8
+; GFX11-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX11-NEXT:    v_and_b32_e32 v0, 1, v0
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    v_mov_b32_e32 v0, s10
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, s0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, v2, s1, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, v3, s3, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call i128 @llvm.ssub.sat.i128(i128 %lhs, i128 %rhs)
   ret i128 %result
 }
@@ -5913,94 +6031,181 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s7, v3
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_ssubsat_v2i128:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_sub_u32 s16, s0, s8
-; GFX10PLUS-NEXT:    s_subb_u32 s17, s1, s9
-; GFX10PLUS-NEXT:    s_subb_u32 s18, s2, s10
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s0, s[16:17], s[0:1]
-; GFX10PLUS-NEXT:    s_subb_u32 s19, s3, s11
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[18:19], s[2:3]
-; GFX10PLUS-NEXT:    s_cselect_b32 s20, 1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s0, s[18:19], s[2:3]
-; GFX10PLUS-NEXT:    v_cmp_gt_u64_e64 s2, s[8:9], 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
-; GFX10PLUS-NEXT:    s_and_b32 s0, 1, s20
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[10:11], 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
-; GFX10PLUS-NEXT:    s_cselect_b32 s1, 1, 0
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s2, s[10:11], 0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
-; GFX10PLUS-NEXT:    s_mov_b32 s20, 0
-; GFX10PLUS-NEXT:    s_and_b32 s1, 1, s1
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s19, 31
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s1
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s2
-; GFX10PLUS-NEXT:    s_addc_u32 s2, s0, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s3, s0, 0x80000000
-; GFX10PLUS-NEXT:    s_sub_u32 s8, s4, s12
-; GFX10PLUS-NEXT:    s_subb_u32 s9, s5, s13
-; GFX10PLUS-NEXT:    s_subb_u32 s10, s6, s14
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s4, s[8:9], s[4:5]
-; GFX10PLUS-NEXT:    s_subb_u32 s11, s7, s15
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v3, v2, vcc_lo
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[10:11], s[6:7]
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v2, s17
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v7, s11
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s4
-; GFX10PLUS-NEXT:    v_cmp_lt_i64_e64 s4, s[10:11], s[6:7]
-; GFX10PLUS-NEXT:    v_cmp_gt_u64_e64 s6, s[12:13], 0
-; GFX10PLUS-NEXT:    v_xor_b32_e32 v0, v1, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v1, s16
-; GFX10PLUS-NEXT:    s_cselect_b32 s16, 1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s4
-; GFX10PLUS-NEXT:    s_and_b32 s4, 1, s16
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[14:15], 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s6
-; GFX10PLUS-NEXT:    v_cmp_gt_i64_e64 s6, s[14:15], 0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s4
-; GFX10PLUS-NEXT:    s_cselect_b32 s5, 1, 0
-; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX10PLUS-NEXT:    s_and_b32 s5, 1, s5
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s6
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v3, v4, v3, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s5
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v4, v6, v5, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v0, s18
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v5, s19
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v6, s9
-; GFX10PLUS-NEXT:    v_xor_b32_e32 v3, v4, v3
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, v1, s0, vcc_lo
-; GFX10PLUS-NEXT:    s_ashr_i32 s0, s11, 31
-; GFX10PLUS-NEXT:    s_cmp_lg_u32 s20, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, v2, s1, vcc_lo
-; GFX10PLUS-NEXT:    v_and_b32_e32 v3, 1, v3
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v4, v5, s3, vcc_lo
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v5, s8
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s0, 0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v3
-; GFX10PLUS-NEXT:    v_mov_b32_e32 v3, s10
-; GFX10PLUS-NEXT:    s_addc_u32 s2, s0, 0
-; GFX10PLUS-NEXT:    s_addc_u32 s3, s0, 0x80000000
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v5, v5, s0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v6, v6, s1, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, v3, s2, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v7, v7, s3, vcc_lo
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v1
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s2, v0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s3, v4
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s4, v5
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s5, v6
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s6, v3
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s7, v7
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_ssubsat_v2i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_sub_u32 s16, s0, s8
+; GFX10-NEXT:    s_subb_u32 s17, s1, s9
+; GFX10-NEXT:    s_subb_u32 s18, s2, s10
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s0, s[16:17], s[0:1]
+; GFX10-NEXT:    s_subb_u32 s19, s3, s11
+; GFX10-NEXT:    s_cmp_eq_u64 s[18:19], s[2:3]
+; GFX10-NEXT:    s_cselect_b32 s20, 1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s0, s[18:19], s[2:3]
+; GFX10-NEXT:    v_cmp_gt_u64_e64 s2, s[8:9], 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX10-NEXT:    s_and_b32 s0, 1, s20
+; GFX10-NEXT:    s_cmp_eq_u64 s[10:11], 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
+; GFX10-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s2, s[10:11], 0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX10-NEXT:    s_mov_b32 s20, 0
+; GFX10-NEXT:    s_and_b32 s1, 1, s1
+; GFX10-NEXT:    s_ashr_i32 s0, s19, 31
+; GFX10-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s1
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s2
+; GFX10-NEXT:    s_addc_u32 s2, s0, 0
+; GFX10-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX10-NEXT:    s_sub_u32 s8, s4, s12
+; GFX10-NEXT:    s_subb_u32 s9, s5, s13
+; GFX10-NEXT:    s_subb_u32 s10, s6, s14
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s4, s[8:9], s[4:5]
+; GFX10-NEXT:    s_subb_u32 s11, s7, s15
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v3, v2, vcc_lo
+; GFX10-NEXT:    s_cmp_eq_u64 s[10:11], s[6:7]
+; GFX10-NEXT:    v_mov_b32_e32 v2, s17
+; GFX10-NEXT:    v_mov_b32_e32 v7, s11
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s4
+; GFX10-NEXT:    v_cmp_lt_i64_e64 s4, s[10:11], s[6:7]
+; GFX10-NEXT:    v_cmp_gt_u64_e64 s6, s[12:13], 0
+; GFX10-NEXT:    v_xor_b32_e32 v0, v1, v0
+; GFX10-NEXT:    v_mov_b32_e32 v1, s16
+; GFX10-NEXT:    s_cselect_b32 s16, 1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s4
+; GFX10-NEXT:    s_and_b32 s4, 1, s16
+; GFX10-NEXT:    s_cmp_eq_u64 s[14:15], 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s6
+; GFX10-NEXT:    v_cmp_gt_i64_e64 s6, s[14:15], 0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s4
+; GFX10-NEXT:    s_cselect_b32 s5, 1, 0
+; GFX10-NEXT:    v_and_b32_e32 v0, 1, v0
+; GFX10-NEXT:    s_and_b32 s5, 1, s5
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s6
+; GFX10-NEXT:    v_cndmask_b32_e32 v3, v4, v3, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s5
+; GFX10-NEXT:    v_cndmask_b32_e32 v4, v6, v5, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT:    v_mov_b32_e32 v0, s18
+; GFX10-NEXT:    v_mov_b32_e32 v5, s19
+; GFX10-NEXT:    v_mov_b32_e32 v6, s9
+; GFX10-NEXT:    v_xor_b32_e32 v3, v4, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, v1, s0, vcc_lo
+; GFX10-NEXT:    s_ashr_i32 s0, s11, 31
+; GFX10-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, v2, s1, vcc_lo
+; GFX10-NEXT:    v_and_b32_e32 v3, 1, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, v5, s3, vcc_lo
+; GFX10-NEXT:    v_mov_b32_e32 v5, s8
+; GFX10-NEXT:    s_addc_u32 s1, s0, 0
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v3
+; GFX10-NEXT:    v_mov_b32_e32 v3, s10
+; GFX10-NEXT:    s_addc_u32 s2, s0, 0
+; GFX10-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, v5, s0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, v6, s1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, v3, s2, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, v7, s3, vcc_lo
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v2
+; GFX10-NEXT:    v_readfirstlane_b32 s2, v0
+; GFX10-NEXT:    v_readfirstlane_b32 s3, v4
+; GFX10-NEXT:    v_readfirstlane_b32 s4, v5
+; GFX10-NEXT:    v_readfirstlane_b32 s5, v6
+; GFX10-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX10-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_ssubsat_v2i128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_sub_u32 s16, s0, s8
+; GFX11-NEXT:    s_subb_u32 s17, s1, s9
+; GFX11-NEXT:    s_subb_u32 s18, s2, s10
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s0, s[16:17], s[0:1]
+; GFX11-NEXT:    s_subb_u32 s19, s3, s11
+; GFX11-NEXT:    s_cmp_eq_u64 s[18:19], s[2:3]
+; GFX11-NEXT:    s_cselect_b32 s20, 1, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s0, s[18:19], s[2:3]
+; GFX11-NEXT:    v_cmp_gt_u64_e64 s2, s[8:9], 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX11-NEXT:    s_and_b32 s0, 1, s20
+; GFX11-NEXT:    s_cmp_eq_u64 s[10:11], 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s2
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s2, s[10:11], 0
+; GFX11-NEXT:    s_cselect_b32 s1, 1, 0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX11-NEXT:    s_mov_b32 s20, 0
+; GFX11-NEXT:    s_and_b32 s1, 1, s1
+; GFX11-NEXT:    s_ashr_i32 s0, s19, 31
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s2
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s1
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0
+; GFX11-NEXT:    s_addc_u32 s2, s0, 0
+; GFX11-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX11-NEXT:    s_sub_u32 s8, s4, s12
+; GFX11-NEXT:    v_cndmask_b32_e32 v1, v3, v2, vcc_lo
+; GFX11-NEXT:    s_subb_u32 s9, s5, s13
+; GFX11-NEXT:    s_subb_u32 s10, s6, s14
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s4, s[8:9], s[4:5]
+; GFX11-NEXT:    s_subb_u32 s11, s7, s15
+; GFX11-NEXT:    v_xor_b32_e32 v0, v1, v0
+; GFX11-NEXT:    s_cmp_eq_u64 s[10:11], s[6:7]
+; GFX11-NEXT:    v_mov_b32_e32 v1, s16
+; GFX11-NEXT:    s_cselect_b32 s16, 1, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, 0, 1, s4
+; GFX11-NEXT:    v_cmp_lt_i64_e64 s4, s[10:11], s[6:7]
+; GFX11-NEXT:    v_cmp_gt_u64_e64 s6, s[12:13], 0
+; GFX11-NEXT:    v_dual_mov_b32 v7, s11 :: v_dual_and_b32 v0, 1, v0
+; GFX11-NEXT:    v_cndmask_b32_e64 v4, 0, 1, s4
+; GFX11-NEXT:    s_and_b32 s4, 1, s16
+; GFX11-NEXT:    s_cmp_eq_u64 s[14:15], 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v5, 0, 1, s6
+; GFX11-NEXT:    v_cmp_gt_i64_e64 s6, s[14:15], 0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s4
+; GFX11-NEXT:    s_cselect_b32 s5, 1, 0
+; GFX11-NEXT:    v_mov_b32_e32 v2, s17
+; GFX11-NEXT:    s_and_b32 s5, 1, s5
+; GFX11-NEXT:    v_cndmask_b32_e32 v3, v4, v3, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v6, 0, 1, s6
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s5
+; GFX11-NEXT:    v_dual_cndmask_b32 v4, v6, v5 :: v_dual_mov_b32 v5, s19
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    v_mov_b32_e32 v6, s9
+; GFX11-NEXT:    v_xor_b32_e32 v3, v4, v3
+; GFX11-NEXT:    v_mov_b32_e32 v0, s18
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, s0, vcc_lo
+; GFX11-NEXT:    s_ashr_i32 s0, s11, 31
+; GFX11-NEXT:    s_cmp_lg_u32 s20, 0
+; GFX11-NEXT:    v_and_b32_e32 v3, 1, v3
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, v2, s1, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, s2, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v4, v5, s3, vcc_lo
+; GFX11-NEXT:    v_mov_b32_e32 v5, s8
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v3
+; GFX11-NEXT:    v_mov_b32_e32 v3, s10
+; GFX11-NEXT:    s_addc_u32 s1, s0, 0
+; GFX11-NEXT:    s_addc_u32 s2, s0, 0
+; GFX11-NEXT:    s_addc_u32 s3, s0, 0x80000000
+; GFX11-NEXT:    v_cndmask_b32_e64 v5, v5, s0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v6, v6, s1, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, v3, s2, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v7, v7, s3, vcc_lo
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call <2 x i128> @llvm.ssub.sat.v2i128(<2 x i128> %lhs, <2 x i128> %rhs)
   ret <2 x i128> %result
 }
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/uaddsat.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/uaddsat.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/uaddsat.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/uaddsat.ll
@@ -3549,16 +3549,16 @@
 ; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[2:3], v[10:11]
 ; GFX11-NEXT:    v_cndmask_b32_e32 v8, v9, v8, vcc_lo
 ; GFX11-NEXT:    v_cmp_eq_u64_e32 vcc_lo, v[6:7], v[14:15]
-; GFX11-NEXT:    v_and_b32_e32 v8, 1, v8
 ; GFX11-NEXT:    v_cndmask_b32_e32 v9, v13, v12, vcc_lo
-; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
 ; GFX11-NEXT:    v_and_b32_e32 v9, 1, v9
+; GFX11-NEXT:    v_and_b32_e32 v8, 1, v8
+; GFX11-NEXT:    v_cmp_ne_u32_e64 s0, 0, v9
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v8
+; GFX11-NEXT:    v_cndmask_b32_e64 v4, v4, -1, s0
 ; GFX11-NEXT:    v_cndmask_b32_e64 v0, v0, -1, vcc_lo
 ; GFX11-NEXT:    v_cndmask_b32_e64 v1, v1, -1, vcc_lo
-; GFX11-NEXT:    v_cmp_ne_u32_e64 s0, 0, v9
 ; GFX11-NEXT:    v_cndmask_b32_e64 v2, v2, -1, vcc_lo
 ; GFX11-NEXT:    v_cndmask_b32_e64 v3, v3, -1, vcc_lo
-; GFX11-NEXT:    v_cndmask_b32_e64 v4, v4, -1, s0
 ; GFX11-NEXT:    v_cndmask_b32_e64 v5, v5, -1, s0
 ; GFX11-NEXT:    v_cndmask_b32_e64 v6, v6, -1, s0
 ; GFX11-NEXT:    v_cndmask_b32_e64 v7, v7, -1, s0
@@ -3760,55 +3760,104 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s7, v3
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_uaddsat_v2i128:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_add_u32 s0, s0, s8
-; GFX10PLUS-NEXT:    s_addc_u32 s1, s1, s9
-; GFX10PLUS-NEXT:    s_addc_u32 s2, s2, s10
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s8, s[0:1], s[8:9]
-; GFX10PLUS-NEXT:    s_addc_u32 s3, s3, s11
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[2:3], s[10:11]
-; GFX10PLUS-NEXT:    s_cselect_b32 s16, 1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s8
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s8, s[2:3], s[10:11]
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s8
-; GFX10PLUS-NEXT:    s_and_b32 s8, 1, s16
-; GFX10PLUS-NEXT:    s_add_u32 s4, s4, s12
-; GFX10PLUS-NEXT:    s_addc_u32 s5, s5, s13
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s8
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s9, s[4:5], s[12:13]
-; GFX10PLUS-NEXT:    s_addc_u32 s6, s6, s14
-; GFX10PLUS-NEXT:    s_addc_u32 s7, s7, s15
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[6:7], s[14:15]
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s9
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s9, s[6:7], s[14:15]
-; GFX10PLUS-NEXT:    s_cselect_b32 s8, 1, 0
-; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX10PLUS-NEXT:    s_and_b32 s8, 1, s8
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s8
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s9
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v2, v1, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX10PLUS-NEXT:    v_and_b32_e32 v1, 1, v1
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, s0, -1, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, s2, -1, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, s3, -1, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 s0, 0, v1
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, s1, -1, vcc_lo
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s2, v2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s3, v3
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v4, s4, -1, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v5, s5, -1, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v6, s6, -1, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v7, s7, -1, s0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s4, v4
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s5, v5
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s6, v6
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s7, v7
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_uaddsat_v2i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_add_u32 s0, s0, s8
+; GFX10-NEXT:    s_addc_u32 s1, s1, s9
+; GFX10-NEXT:    s_addc_u32 s2, s2, s10
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s8, s[0:1], s[8:9]
+; GFX10-NEXT:    s_addc_u32 s3, s3, s11
+; GFX10-NEXT:    s_cmp_eq_u64 s[2:3], s[10:11]
+; GFX10-NEXT:    s_cselect_b32 s16, 1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s8
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s8, s[2:3], s[10:11]
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s8
+; GFX10-NEXT:    s_and_b32 s8, 1, s16
+; GFX10-NEXT:    s_add_u32 s4, s4, s12
+; GFX10-NEXT:    s_addc_u32 s5, s5, s13
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s8
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s9, s[4:5], s[12:13]
+; GFX10-NEXT:    s_addc_u32 s6, s6, s14
+; GFX10-NEXT:    s_addc_u32 s7, s7, s15
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX10-NEXT:    s_cmp_eq_u64 s[6:7], s[14:15]
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s9
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s9, s[6:7], s[14:15]
+; GFX10-NEXT:    s_cselect_b32 s8, 1, 0
+; GFX10-NEXT:    v_and_b32_e32 v0, 1, v0
+; GFX10-NEXT:    s_and_b32 s8, 1, s8
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s8
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s9
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v2, v1, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT:    v_and_b32_e32 v1, 1, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, s0, -1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, s2, -1, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, s3, -1, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, s1, -1, vcc_lo
+; GFX10-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX10-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, s4, -1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, s5, -1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, s6, -1, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, s7, -1, s0
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX10-NEXT:    v_readfirstlane_b32 s4, v4
+; GFX10-NEXT:    v_readfirstlane_b32 s5, v5
+; GFX10-NEXT:    v_readfirstlane_b32 s6, v6
+; GFX10-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_uaddsat_v2i128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_add_u32 s0, s0, s8
+; GFX11-NEXT:    s_addc_u32 s1, s1, s9
+; GFX11-NEXT:    s_addc_u32 s2, s2, s10
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s8, s[0:1], s[8:9]
+; GFX11-NEXT:    s_addc_u32 s3, s3, s11
+; GFX11-NEXT:    s_cmp_eq_u64 s[2:3], s[10:11]
+; GFX11-NEXT:    s_cselect_b32 s16, 1, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s8
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s8, s[2:3], s[10:11]
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s8
+; GFX11-NEXT:    s_and_b32 s8, 1, s16
+; GFX11-NEXT:    s_add_u32 s4, s4, s12
+; GFX11-NEXT:    s_addc_u32 s5, s5, s13
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s8
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s9, s[4:5], s[12:13]
+; GFX11-NEXT:    s_addc_u32 s6, s6, s14
+; GFX11-NEXT:    s_addc_u32 s7, s7, s15
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX11-NEXT:    s_cmp_eq_u64 s[6:7], s[14:15]
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s9
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s9, s[6:7], s[14:15]
+; GFX11-NEXT:    s_cselect_b32 s8, 1, 0
+; GFX11-NEXT:    s_and_b32 s8, 1, s8
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s8
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s9
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v2, v1 :: v_dual_and_b32 v0, 1, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 1, v1
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, s0, -1, vcc_lo
+; GFX11-NEXT:    v_cmp_ne_u32_e64 s0, 0, v1
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, s1, -1, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, s2, -1, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, s3, -1, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v4, s4, -1, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v5, s5, -1, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v6, s6, -1, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v7, s7, -1, s0
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call <2 x i128> @llvm.uadd.sat.v2i128(<2 x i128> %lhs, <2 x i128> %rhs)
   ret <2 x i128> %result
 }
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/usubsat.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/usubsat.ll
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/usubsat.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/usubsat.ll
@@ -3630,55 +3630,104 @@
 ; GFX9-NEXT:    v_readfirstlane_b32 s7, v3
 ; GFX9-NEXT:    ; return to shader part epilog
 ;
-; GFX10PLUS-LABEL: s_usubsat_v2i128:
-; GFX10PLUS:       ; %bb.0:
-; GFX10PLUS-NEXT:    s_sub_u32 s16, s0, s8
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s0, s[0:1], s[8:9]
-; GFX10PLUS-NEXT:    s_subb_u32 s17, s1, s9
-; GFX10PLUS-NEXT:    s_subb_u32 s18, s2, s10
-; GFX10PLUS-NEXT:    s_subb_u32 s19, s3, s11
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[2:3], s[10:11]
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s0, s[2:3], s[10:11]
-; GFX10PLUS-NEXT:    s_cselect_b32 s20, 1, 0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
-; GFX10PLUS-NEXT:    s_and_b32 s0, 1, s20
-; GFX10PLUS-NEXT:    s_sub_u32 s2, s4, s12
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s4, s[4:5], s[12:13]
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
-; GFX10PLUS-NEXT:    s_subb_u32 s1, s5, s13
-; GFX10PLUS-NEXT:    s_subb_u32 s8, s6, s14
-; GFX10PLUS-NEXT:    s_subb_u32 s3, s7, s15
-; GFX10PLUS-NEXT:    s_cmp_eq_u64 s[6:7], s[14:15]
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s4
-; GFX10PLUS-NEXT:    v_cmp_lt_u64_e64 s4, s[6:7], s[14:15]
-; GFX10PLUS-NEXT:    s_cselect_b32 s0, 1, 0
-; GFX10PLUS-NEXT:    s_and_b32 s0, 1, s0
-; GFX10PLUS-NEXT:    v_and_b32_e32 v0, 1, v0
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s4
-; GFX10PLUS-NEXT:    v_cndmask_b32_e32 v1, v2, v1, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
-; GFX10PLUS-NEXT:    v_and_b32_e32 v1, 1, v1
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v0, s16, 0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v2, s18, 0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v3, s19, 0, vcc_lo
-; GFX10PLUS-NEXT:    v_cmp_ne_u32_e64 s0, 0, v1
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v1, s17, 0, vcc_lo
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v4, s2, 0, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v5, s1, 0, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v6, s8, 0, s0
-; GFX10PLUS-NEXT:    v_cndmask_b32_e64 v7, s3, 0, s0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s1, v1
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s2, v2
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s3, v3
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s4, v4
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s5, v5
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s6, v6
-; GFX10PLUS-NEXT:    v_readfirstlane_b32 s7, v7
-; GFX10PLUS-NEXT:    ; return to shader part epilog
+; GFX10-LABEL: s_usubsat_v2i128:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_sub_u32 s16, s0, s8
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s0, s[0:1], s[8:9]
+; GFX10-NEXT:    s_subb_u32 s17, s1, s9
+; GFX10-NEXT:    s_subb_u32 s18, s2, s10
+; GFX10-NEXT:    s_subb_u32 s19, s3, s11
+; GFX10-NEXT:    s_cmp_eq_u64 s[2:3], s[10:11]
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s0, s[2:3], s[10:11]
+; GFX10-NEXT:    s_cselect_b32 s20, 1, 0
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX10-NEXT:    s_and_b32 s0, 1, s20
+; GFX10-NEXT:    s_sub_u32 s2, s4, s12
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s4, s[4:5], s[12:13]
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX10-NEXT:    s_subb_u32 s1, s5, s13
+; GFX10-NEXT:    s_subb_u32 s8, s6, s14
+; GFX10-NEXT:    s_subb_u32 s3, s7, s15
+; GFX10-NEXT:    s_cmp_eq_u64 s[6:7], s[14:15]
+; GFX10-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s4
+; GFX10-NEXT:    v_cmp_lt_u64_e64 s4, s[6:7], s[14:15]
+; GFX10-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX10-NEXT:    s_and_b32 s0, 1, s0
+; GFX10-NEXT:    v_and_b32_e32 v0, 1, v0
+; GFX10-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s4
+; GFX10-NEXT:    v_cndmask_b32_e32 v1, v2, v1, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX10-NEXT:    v_and_b32_e32 v1, 1, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v0, s16, 0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v2, s18, 0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v3, s19, 0, vcc_lo
+; GFX10-NEXT:    v_cmp_ne_u32_e64 s0, 0, v1
+; GFX10-NEXT:    v_cndmask_b32_e64 v1, s17, 0, vcc_lo
+; GFX10-NEXT:    v_cndmask_b32_e64 v4, s2, 0, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v5, s1, 0, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v6, s8, 0, s0
+; GFX10-NEXT:    v_cndmask_b32_e64 v7, s3, 0, s0
+; GFX10-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX10-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX10-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX10-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX10-NEXT:    v_readfirstlane_b32 s4, v4
+; GFX10-NEXT:    v_readfirstlane_b32 s5, v5
+; GFX10-NEXT:    v_readfirstlane_b32 s6, v6
+; GFX10-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX10-NEXT:    ; return to shader part epilog
+;
+; GFX11-LABEL: s_usubsat_v2i128:
+; GFX11:       ; %bb.0:
+; GFX11-NEXT:    s_sub_u32 s16, s0, s8
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s0, s[0:1], s[8:9]
+; GFX11-NEXT:    s_subb_u32 s17, s1, s9
+; GFX11-NEXT:    s_subb_u32 s18, s2, s10
+; GFX11-NEXT:    s_subb_u32 s19, s3, s11
+; GFX11-NEXT:    s_cmp_eq_u64 s[2:3], s[10:11]
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s0, s[2:3], s[10:11]
+; GFX11-NEXT:    s_cselect_b32 s20, 1, 0
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s0
+; GFX11-NEXT:    s_and_b32 s0, 1, s20
+; GFX11-NEXT:    s_sub_u32 s2, s4, s12
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s4, s[4:5], s[12:13]
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX11-NEXT:    s_subb_u32 s1, s5, s13
+; GFX11-NEXT:    s_subb_u32 s8, s6, s14
+; GFX11-NEXT:    s_subb_u32 s3, s7, s15
+; GFX11-NEXT:    s_cmp_eq_u64 s[6:7], s[14:15]
+; GFX11-NEXT:    v_cndmask_b32_e32 v0, v1, v0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, 0, 1, s4
+; GFX11-NEXT:    v_cmp_lt_u64_e64 s4, s[6:7], s[14:15]
+; GFX11-NEXT:    s_cselect_b32 s0, 1, 0
+; GFX11-NEXT:    s_and_b32 s0, 1, s0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 vcc_lo, 0, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, 0, 1, s4
+; GFX11-NEXT:    v_dual_cndmask_b32 v1, v2, v1 :: v_dual_and_b32 v0, 1, v0
+; GFX11-NEXT:    v_and_b32_e32 v1, 1, v1
+; GFX11-NEXT:    v_cmp_ne_u32_e32 vcc_lo, 0, v0
+; GFX11-NEXT:    v_cmp_ne_u32_e64 s0, 0, v1
+; GFX11-NEXT:    v_cndmask_b32_e64 v0, s16, 0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v1, s17, 0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v2, s18, 0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v3, s19, 0, vcc_lo
+; GFX11-NEXT:    v_cndmask_b32_e64 v4, s2, 0, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v5, s1, 0, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v6, s8, 0, s0
+; GFX11-NEXT:    v_cndmask_b32_e64 v7, s3, 0, s0
+; GFX11-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX11-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX11-NEXT:    v_readfirstlane_b32 s2, v2
+; GFX11-NEXT:    v_readfirstlane_b32 s3, v3
+; GFX11-NEXT:    v_readfirstlane_b32 s4, v4
+; GFX11-NEXT:    v_readfirstlane_b32 s5, v5
+; GFX11-NEXT:    v_readfirstlane_b32 s6, v6
+; GFX11-NEXT:    v_readfirstlane_b32 s7, v7
+; GFX11-NEXT:    ; return to shader part epilog
   %result = call <2 x i128> @llvm.usub.sat.v2i128(<2 x i128> %lhs, <2 x i128> %rhs)
   ret <2 x i128> %result
 }
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/build_vector.ll b/llvm/test/CodeGen/AMDGPU/build_vector.ll
--- a/llvm/test/CodeGen/AMDGPU/build_vector.ll
+++ b/llvm/test/CodeGen/AMDGPU/build_vector.ll
@@ -2,7 +2,7 @@
 ; RUN: llc < %s -march=amdgcn -verify-machineinstrs | FileCheck %s --check-prefixes=GFX6,GFX678,ALL
 ; RUN: llc < %s -march=amdgcn -mcpu=tonga -mattr=-flat-for-global -verify-machineinstrs | FileCheck %s --check-prefixes=GFX8,GFX678,ALL
 ; RUN: llc < %s -mtriple=amdgcn-amd-amdpal -mcpu=gfx1030 -verify-machineinstrs | FileCheck %s --check-prefixes=GFX10,GFX1011,ALL
-; RUN: llc < %s -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -verify-machineinstrs | FileCheck %s --check-prefixes=GFX11,GFX1011,ALL
+; RUN: llc < %s -mtriple=amdgcn-amd-amdpal -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs | FileCheck %s --check-prefixes=GFX11,GFX1011,ALL
 
 ; ALL-LABEL: {{^}}build_vector2:
 ; R600: MOV
diff --git a/llvm/test/CodeGen/AMDGPU/clamp.ll b/llvm/test/CodeGen/AMDGPU/clamp.ll
--- a/llvm/test/CodeGen/AMDGPU/clamp.ll
+++ b/llvm/test/CodeGen/AMDGPU/clamp.ll
@@ -1358,8 +1358,7 @@
 ; GFX11-LABEL: v_clamp_constants_to_one_f32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 1.0
+; GFX11-NEXT:    v_dual_mov_b32 v1, 1.0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1407,8 +1406,7 @@
 ; GFX11-LABEL: v_clamp_constants_to_zero_f32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1457,8 +1455,7 @@
 ; GFX11-LABEL: v_clamp_constant_preserve_f32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0.5
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0.5 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1507,8 +1504,7 @@
 ; GFX11-LABEL: v_clamp_constant_preserve_denorm_f32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0x7fffff
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0x7fffff :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1556,8 +1552,7 @@
 ; GFX11-LABEL: v_clamp_constant_qnan_f32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1605,8 +1600,7 @@
 ; GFX11-LABEL: v_clamp_constant_snan_f32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -2335,8 +2329,7 @@
 ; GFX11-LABEL: v_clamp_constant_qnan_f32_no_dx10_clamp:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0x7fc00000
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0x7fc00000 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -2385,8 +2378,7 @@
 ; GFX11-LABEL: v_clamp_constant_snan_f32_no_dx10_clamp:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x24
-; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX11-NEXT:    v_mov_b32_e32 v1, 0x7f800001
+; GFX11-NEXT:    v_dual_mov_b32 v1, 0x7f800001 :: v_dual_lshlrev_b32 v0, 2, v0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b32 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
diff --git a/llvm/test/CodeGen/AMDGPU/cluster_stores.ll b/llvm/test/CodeGen/AMDGPU/cluster_stores.ll
--- a/llvm/test/CodeGen/AMDGPU/cluster_stores.ll
+++ b/llvm/test/CodeGen/AMDGPU/cluster_stores.ll
@@ -106,15 +106,13 @@
 ; GFX11-NEXT:    s_load_b64 s[2:3], s[0:1], 0x24
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x2c
 ; 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:    s_clause 0x3
 ; GFX11-NEXT:    flat_load_b32 v2, v[0:1]
 ; GFX11-NEXT:    flat_load_b32 v3, v[0:1] offset:8
 ; GFX11-NEXT:    flat_load_b32 v4, v[0:1] offset:16
 ; GFX11-NEXT:    flat_load_b32 v5, v[0:1] offset:24
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-NEXT:    s_waitcnt vmcnt(3) lgkmcnt(3)
 ; GFX11-NEXT:    flat_store_b32 v[0:1], v2
 ; GFX11-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(3)
@@ -246,15 +244,13 @@
 ; GFX11-NEXT:    s_load_b64 s[2:3], s[0:1], 0x24
 ; GFX11-NEXT:    s_load_b64 s[0:1], s[0:1], 0x2c
 ; 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:    s_clause 0x3
 ; GFX11-NEXT:    flat_load_b32 v2, v[0:1] offset:8
 ; GFX11-NEXT:    flat_load_b32 v3, v[0:1]
 ; GFX11-NEXT:    flat_load_b32 v4, v[0:1] offset:16
 ; GFX11-NEXT:    flat_load_b32 v5, v[0:1] offset:24
-; GFX11-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-NEXT:    v_mov_b32_e32 v1, s1
+; GFX11-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
 ; GFX11-NEXT:    s_waitcnt vmcnt(3) lgkmcnt(3)
 ; GFX11-NEXT:    v_add_nc_u32_e32 v2, 1, v2
 ; GFX11-NEXT:    s_waitcnt vmcnt(2) lgkmcnt(2)
@@ -341,10 +337,8 @@
 ; GFX11-NEXT:    image_load v[2:5], v[2:3], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
 ; GFX11-NEXT:    image_load v[6:9], v[6:7], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_add_f32_e32 v5, v5, v9
-; GFX11-NEXT:    v_add_f32_e32 v4, v4, v8
-; GFX11-NEXT:    v_add_f32_e32 v3, v3, v7
-; GFX11-NEXT:    v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT:    v_dual_add_f32 v2, v2, v6 :: v_dual_add_f32 v5, v5, v9
+; GFX11-NEXT:    v_dual_add_f32 v4, v4, v8 :: v_dual_add_f32 v3, v3, v7
 ; GFX11-NEXT:    image_store v[2:5], v[0:1], s[8:15] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
@@ -398,10 +392,8 @@
 ; GFX11-NEXT:    image_load_mip v[2:5], [v0, v1, v6], s[0:7] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
 ; GFX11-NEXT:    image_load_mip v[6:9], [v0, v1, v6], s[8:15] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_add_f32_e32 v5, v5, v9
-; GFX11-NEXT:    v_add_f32_e32 v4, v4, v8
-; GFX11-NEXT:    v_add_f32_e32 v3, v3, v7
-; GFX11-NEXT:    v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT:    v_dual_add_f32 v5, v5, v9 :: v_dual_add_f32 v4, v4, v8
+; GFX11-NEXT:    v_dual_add_f32 v3, v3, v7 :: v_dual_add_f32 v2, v2, v6
 ; GFX11-NEXT:    image_store v[2:5], v[0:1], s[16:23] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
@@ -476,30 +468,23 @@
 ;
 ; GFX11-LABEL: cluster_image_sample:
 ; GFX11:       ; %bb.0: ; %entry
-; GFX11-NEXT:    v_cvt_f32_i32_e32 v8, v0
 ; GFX11-NEXT:    v_cvt_f32_i32_e32 v9, v1
-; GFX11-NEXT:    v_mov_b32_e32 v4, 0
-; GFX11-NEXT:    v_mov_b32_e32 v10, 1.0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
-; GFX11-NEXT:    v_add_f32_e32 v2, 1.0, v8
-; GFX11-NEXT:    v_add_f32_e32 v3, 1.0, v9
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT:    v_mov_b32_e32 v5, v4
-; GFX11-NEXT:    v_mov_b32_e32 v6, v4
-; GFX11-NEXT:    v_mov_b32_e32 v7, v4
-; GFX11-NEXT:    v_add_f32_e32 v8, 2.0, v8
-; GFX11-NEXT:    v_add_f32_e32 v9, 2.0, v9
-; GFX11-NEXT:    v_mov_b32_e32 v11, v10
+; GFX11-NEXT:    v_cvt_f32_i32_e32 v8, v0
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_dual_mov_b32 v4, 0 :: v_dual_add_f32 v3, 1.0, v9
+; GFX11-NEXT:    v_dual_mov_b32 v10, 1.0 :: v_dual_mov_b32 v7, v4
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT:    v_dual_add_f32 v2, 1.0, v8 :: v_dual_mov_b32 v5, v4
+; GFX11-NEXT:    v_dual_mov_b32 v6, v4 :: v_dual_add_f32 v9, 2.0, v9
+; GFX11-NEXT:    v_dual_add_f32 v8, 2.0, v8 :: v_dual_mov_b32 v11, v10
 ; GFX11-NEXT:    v_mov_b32_e32 v12, v10
 ; GFX11-NEXT:    v_mov_b32_e32 v13, v10
 ; GFX11-NEXT:    s_clause 0x1
 ; GFX11-NEXT:    image_sample_d v[2:5], v[2:7], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D
 ; GFX11-NEXT:    image_sample_d v[6:9], v[8:13], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_add_f32_e32 v5, v5, v9
-; GFX11-NEXT:    v_add_f32_e32 v4, v4, v8
-; GFX11-NEXT:    v_add_f32_e32 v3, v3, v7
-; GFX11-NEXT:    v_add_f32_e32 v2, v2, v6
+; GFX11-NEXT:    v_dual_add_f32 v4, v4, v8 :: v_dual_add_f32 v5, v5, v9
+; GFX11-NEXT:    v_dual_add_f32 v3, v3, v7 :: v_dual_add_f32 v2, v2, v6
 ; GFX11-NEXT:    image_store v[2:5], v[0:1], s[12:19] dmask:0xf dim:SQ_RSRC_IMG_2D unorm
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
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/gfx11-user-sgpr-init16-bug.ll b/llvm/test/CodeGen/AMDGPU/gfx11-user-sgpr-init16-bug.ll
--- a/llvm/test/CodeGen/AMDGPU/gfx11-user-sgpr-init16-bug.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx11-user-sgpr-init16-bug.ll
@@ -1,7 +1,7 @@
-; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1102 < %s | FileCheck -check-prefix=GCN %s
-; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1102 < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -amdgpu-enable-vopd=0 < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -amdgpu-enable-vopd=0 < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -global-isel=0 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1102 -amdgpu-enable-vopd=0 < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -global-isel=1 -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1102 -amdgpu-enable-vopd=0 < %s | FileCheck -check-prefix=GCN %s
 
 ; There aren't any stack objects, but we still enable the
 ; private_segment_wavefront_offset to get to 16, and the workgroup ID
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
@@ -378,6 +378,7 @@
 ; 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:        MachineDominator Tree Construction
 ; GCN-O1-NEXT:        Machine Natural Loop Construction
@@ -667,6 +668,7 @@
 ; 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:        MachineDominator Tree Construction
 ; GCN-O1-OPTS-NEXT:        Machine Natural Loop Construction
@@ -958,6 +960,7 @@
 ; 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:        MachineDominator Tree Construction
 ; GCN-O2-NEXT:        Machine Natural Loop Construction
@@ -1262,6 +1265,7 @@
 ; 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:        MachineDominator Tree Construction
 ; GCN-O3-NEXT:        Machine Natural Loop Construction
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll
@@ -1,5 +1,5 @@
-; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
-; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
+; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
 
 ; FUNC-LABEL: {{^}}ds_ordered_add:
 ; GCN-DAG: v_mov_b32_e32 v[[INCR:[0-9]+]], 31
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.exp.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.exp.ll
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.exp.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.exp.ll
@@ -1,6 +1,6 @@
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -strict-whitespace -check-prefixes=GCN,PREGFX11 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck -strict-whitespace -check-prefixes=GCN,GFX10,PREGFX11 %s
-; RUN: llc -march=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -strict-whitespace -check-prefixes=GCN,GFX11 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -strict-whitespace -check-prefixes=GCN,GFX11 %s
 
 declare void @llvm.amdgcn.exp.f32(i32, i32, float, float, float, float, i1, i1) #1
 declare void @llvm.amdgcn.exp.i32(i32, i32, i32, i32, i32, i32, i1, i1) #1
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
@@ -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=GCN,GFX9 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx1030 --verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10 %s
-; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 --verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX11 %s
+; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -amdgpu-enable-delay-alu=0 --verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX11 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx1030 -mattr=+wavefrontsize64 --verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX10 %s
 ; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-delay-alu=0 -mattr=+wavefrontsize64 --verify-machineinstrs < %s | FileCheck --check-prefixes=GCN,GFX11 %s
 
@@ -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
@@ -168,13 +147,6 @@
 ; 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
@@ -119,11 +119,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -133,11 +131,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -258,11 +254,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -272,11 +266,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -406,14 +398,12 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -422,14 +412,12 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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
@@ -568,16 +556,14 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -586,16 +572,14 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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
@@ -699,8 +683,7 @@
 ; 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)
@@ -712,8 +695,7 @@
 ; 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)
@@ -817,8 +799,7 @@
 ; 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)
@@ -830,8 +811,7 @@
 ; 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)
@@ -947,8 +927,7 @@
 ; 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
@@ -962,8 +941,7 @@
 ; 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
@@ -1081,8 +1059,7 @@
 ; 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
@@ -1096,8 +1073,7 @@
 ; 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
@@ -1203,8 +1179,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1216,8 +1191,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1340,8 +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)
@@ -1356,8 +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)
@@ -1476,8 +1448,7 @@
 ; 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
@@ -1491,8 +1462,7 @@
 ; 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
@@ -1629,8 +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
@@ -1647,8 +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
@@ -1788,8 +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
@@ -1806,8 +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
@@ -1941,8 +1907,7 @@
 ; 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) lgkmcnt(0)
@@ -1958,8 +1923,7 @@
 ; 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) lgkmcnt(0)
@@ -2105,8 +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
@@ -2124,8 +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
@@ -2273,8 +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
@@ -2292,8 +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
@@ -2416,10 +2376,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2430,10 +2388,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2568,10 +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
@@ -2585,10 +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
@@ -2719,10 +2671,8 @@
 ; 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
@@ -2735,10 +2685,8 @@
 ; 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
@@ -2887,10 +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
@@ -2906,10 +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
@@ -3061,10 +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
@@ -3080,10 +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
@@ -3223,10 +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
@@ -3240,10 +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
@@ -3381,10 +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
@@ -3398,10 +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
@@ -3551,10 +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
@@ -3570,10 +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
@@ -3725,10 +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
@@ -3744,10 +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
@@ -3899,10 +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
@@ -3918,10 +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
@@ -4073,10 +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
@@ -4092,10 +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
@@ -4247,10 +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
@@ -4266,10 +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
@@ -4421,10 +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
@@ -4440,10 +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
@@ -4595,10 +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
@@ -4614,10 +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
@@ -4769,10 +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
@@ -4788,10 +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
@@ -4936,10 +4836,8 @@
 ; 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 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
@@ -4952,10 +4850,8 @@
 ; 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 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
@@ -5108,10 +5004,8 @@
 ; 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 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
@@ -5126,10 +5020,8 @@
 ; 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 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
@@ -5287,10 +5179,8 @@
 ; 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
@@ -5305,10 +5195,8 @@
 ; 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
@@ -5475,10 +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
@@ -5495,10 +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
@@ -5667,10 +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
@@ -5687,10 +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
@@ -5847,10 +5727,8 @@
 ; 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 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
@@ -5865,10 +5743,8 @@
 ; 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 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
@@ -6023,10 +5899,8 @@
 ; 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 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
@@ -6041,10 +5915,8 @@
 ; 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 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
@@ -6211,10 +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
@@ -6231,10 +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
@@ -6403,10 +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
@@ -6423,10 +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
@@ -6595,10 +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
@@ -6615,10 +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
@@ -6787,10 +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
@@ -6807,10 +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
@@ -6979,10 +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
@@ -6999,10 +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
@@ -7171,10 +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
@@ -7191,10 +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
@@ -7363,10 +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
@@ -7383,10 +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
@@ -7555,10 +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
@@ -7575,10 +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
@@ -7706,11 +7546,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -7720,11 +7558,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7845,11 +7681,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -7859,11 +7693,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7999,14 +7831,12 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -8016,14 +7846,12 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -8169,16 +7997,14 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -8188,16 +8014,14 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -8302,8 +8126,7 @@
 ; 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)
@@ -8315,8 +8138,7 @@
 ; 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)
@@ -8420,8 +8242,7 @@
 ; 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)
@@ -8433,8 +8254,7 @@
 ; 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)
@@ -8550,8 +8370,7 @@
 ; 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
@@ -8565,8 +8384,7 @@
 ; 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
@@ -8684,8 +8502,7 @@
 ; 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
@@ -8699,8 +8516,7 @@
 ; 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
@@ -8806,8 +8622,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8819,8 +8634,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8941,8 +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
@@ -8956,8 +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
@@ -9075,8 +8887,7 @@
 ; 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
@@ -9090,8 +8901,7 @@
 ; 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
@@ -9226,8 +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
@@ -9243,8 +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
@@ -9381,8 +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
@@ -9398,8 +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
@@ -9537,8 +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)
@@ -9555,8 +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)
@@ -9708,8 +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
@@ -9728,8 +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
@@ -9883,8 +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
@@ -9903,8 +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
@@ -10028,10 +9828,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10042,10 +9840,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10178,10 +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
@@ -10194,10 +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
@@ -10327,10 +10119,8 @@
 ; 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
@@ -10343,10 +10133,8 @@
 ; 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
@@ -10493,10 +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
@@ -10511,10 +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
@@ -10663,10 +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
@@ -10681,10 +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
@@ -10821,10 +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
@@ -10837,10 +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
@@ -10975,10 +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
@@ -10991,10 +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
@@ -11141,10 +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
@@ -11159,10 +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
@@ -11311,10 +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
@@ -11329,10 +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
@@ -11481,10 +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
@@ -11499,10 +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
@@ -11651,10 +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
@@ -11669,10 +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
@@ -11821,10 +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
@@ -11839,10 +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
@@ -11991,10 +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
@@ -12009,10 +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
@@ -12161,10 +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
@@ -12179,10 +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
@@ -12331,10 +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
@@ -12349,10 +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
@@ -12496,10 +12236,8 @@
 ; 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 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
@@ -12512,10 +12250,8 @@
 ; 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 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
@@ -12674,10 +12410,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12693,10 +12427,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -12855,10 +12587,8 @@
 ; 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
@@ -12873,10 +12603,8 @@
 ; 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
@@ -13049,10 +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
@@ -13070,10 +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
@@ -13249,10 +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
@@ -13270,10 +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
@@ -13437,10 +13157,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -13456,10 +13174,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13621,10 +13337,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -13640,10 +13354,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13817,10 +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
@@ -13838,10 +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
@@ -14017,10 +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
@@ -14038,10 +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
@@ -14217,10 +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
@@ -14238,10 +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
@@ -14417,10 +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
@@ -14438,10 +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
@@ -14617,10 +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
@@ -14638,10 +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
@@ -14817,10 +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
@@ -14838,10 +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
@@ -15017,10 +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
@@ -15038,10 +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
@@ -15217,10 +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
@@ -15238,10 +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
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
@@ -119,11 +119,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -133,11 +131,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -275,8 +271,7 @@
 ; 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)
@@ -291,8 +286,7 @@
 ; 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)
@@ -415,11 +409,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -429,11 +421,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -565,11 +555,10 @@
 ; 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
@@ -581,11 +570,10 @@
 ; 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
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
@@ -119,11 +119,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -133,11 +131,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -258,11 +254,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -272,11 +266,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -397,11 +389,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -411,11 +401,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -536,11 +524,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -550,11 +536,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -659,8 +643,7 @@
 ; 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)
@@ -672,8 +655,7 @@
 ; 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)
@@ -777,8 +759,7 @@
 ; 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)
@@ -790,8 +771,7 @@
 ; 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)
@@ -895,8 +875,7 @@
 ; 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)
@@ -908,8 +887,7 @@
 ; 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)
@@ -1013,8 +991,7 @@
 ; 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)
@@ -1026,8 +1003,7 @@
 ; 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)
@@ -1131,8 +1107,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1144,8 +1119,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1249,8 +1223,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1262,8 +1235,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1367,8 +1339,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1380,8 +1351,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1485,8 +1455,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1498,8 +1467,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1603,8 +1571,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1616,8 +1583,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1737,8 +1703,7 @@
 ; 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) lgkmcnt(0)
@@ -1752,8 +1717,7 @@
 ; 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) lgkmcnt(0)
@@ -1876,8 +1840,7 @@
 ; 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) lgkmcnt(0)
@@ -1891,8 +1854,7 @@
 ; 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) lgkmcnt(0)
@@ -2015,8 +1977,7 @@
 ; 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) lgkmcnt(0)
@@ -2030,8 +1991,7 @@
 ; 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) lgkmcnt(0)
@@ -2150,10 +2110,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2164,10 +2122,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2283,10 +2239,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2297,10 +2251,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2416,10 +2368,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2430,10 +2380,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2549,10 +2497,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2563,10 +2509,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2682,10 +2626,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2696,10 +2638,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2815,10 +2755,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2829,10 +2767,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2948,10 +2884,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2962,10 +2896,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3081,10 +3013,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3095,10 +3025,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3214,10 +3142,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3228,10 +3154,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3347,10 +3271,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3361,10 +3283,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3480,10 +3400,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3494,10 +3412,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3613,10 +3529,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3627,10 +3541,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3746,10 +3658,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3760,10 +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:    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
@@ -3879,10 +3787,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3893,10 +3799,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4012,10 +3916,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -4026,10 +3928,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4169,10 +4069,8 @@
 ; 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 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
@@ -4185,10 +4083,8 @@
 ; 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 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
@@ -4332,10 +4228,8 @@
 ; 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 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
@@ -4348,10 +4242,8 @@
 ; 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 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
@@ -4495,10 +4387,8 @@
 ; 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 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
@@ -4511,10 +4401,8 @@
 ; 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 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
@@ -4658,10 +4546,8 @@
 ; 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 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
@@ -4674,10 +4560,8 @@
 ; 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 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
@@ -4821,10 +4705,8 @@
 ; 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 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
@@ -4837,10 +4719,8 @@
 ; 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 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
@@ -4984,10 +4864,8 @@
 ; 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 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
@@ -5000,10 +4878,8 @@
 ; 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 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
@@ -5147,10 +5023,8 @@
 ; 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 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
@@ -5163,10 +5037,8 @@
 ; 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 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
@@ -5310,10 +5182,8 @@
 ; 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 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
@@ -5326,10 +5196,8 @@
 ; 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 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
@@ -5473,10 +5341,8 @@
 ; 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 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
@@ -5489,10 +5355,8 @@
 ; 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 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
@@ -5636,10 +5500,8 @@
 ; 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 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
@@ -5652,10 +5514,8 @@
 ; 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 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
@@ -5799,10 +5659,8 @@
 ; 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 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
@@ -5815,10 +5673,8 @@
 ; 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 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
@@ -5962,10 +5818,8 @@
 ; 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 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
@@ -5978,10 +5832,8 @@
 ; 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 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
@@ -6125,10 +5977,8 @@
 ; 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 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
@@ -6141,10 +5991,8 @@
 ; 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 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
@@ -6288,10 +6136,8 @@
 ; 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 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
@@ -6304,10 +6150,8 @@
 ; 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 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
@@ -6451,10 +6295,8 @@
 ; 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 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
@@ -6467,10 +6309,8 @@
 ; 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 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
@@ -6594,11 +6434,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6608,11 +6446,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -6733,11 +6569,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6747,11 +6581,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -6872,11 +6704,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6886,11 +6716,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7011,11 +6839,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -7025,11 +6851,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7134,8 +6958,7 @@
 ; 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)
@@ -7147,8 +6970,7 @@
 ; 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)
@@ -7252,8 +7074,7 @@
 ; 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)
@@ -7265,8 +7086,7 @@
 ; 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)
@@ -7370,8 +7190,7 @@
 ; 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)
@@ -7383,8 +7202,7 @@
 ; 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)
@@ -7488,8 +7306,7 @@
 ; 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)
@@ -7501,8 +7318,7 @@
 ; 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)
@@ -7606,8 +7422,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7619,8 +7434,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7724,8 +7538,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7737,8 +7550,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7842,8 +7654,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7855,8 +7666,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7960,8 +7770,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7973,8 +7782,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8078,8 +7886,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8091,8 +7898,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8212,8 +8018,7 @@
 ; 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) lgkmcnt(0)
@@ -8227,8 +8032,7 @@
 ; 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) lgkmcnt(0)
@@ -8351,8 +8155,7 @@
 ; 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) lgkmcnt(0)
@@ -8366,8 +8169,7 @@
 ; 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) lgkmcnt(0)
@@ -8490,8 +8292,7 @@
 ; 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) lgkmcnt(0)
@@ -8505,8 +8306,7 @@
 ; 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) lgkmcnt(0)
@@ -8625,10 +8425,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8639,10 +8437,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8758,10 +8554,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8772,10 +8566,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8891,10 +8683,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8905,10 +8695,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9024,10 +8812,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9038,10 +8824,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9157,10 +8941,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9171,10 +8953,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9290,10 +9070,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9304,10 +9082,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9423,10 +9199,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9437,10 +9211,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9556,10 +9328,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9570,10 +9340,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9689,10 +9457,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9703,10 +9469,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9822,10 +9586,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9836,10 +9598,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9955,10 +9715,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9969,10 +9727,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10088,10 +9844,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10102,10 +9856,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10221,10 +9973,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10235,10 +9985,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10354,10 +10102,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10368,10 +10114,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10487,10 +10231,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10501,10 +10243,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10644,10 +10384,8 @@
 ; 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 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
@@ -10660,10 +10398,8 @@
 ; 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 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
@@ -10807,10 +10543,8 @@
 ; 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 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
@@ -10823,10 +10557,8 @@
 ; 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 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
@@ -10970,10 +10702,8 @@
 ; 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 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
@@ -10986,10 +10716,8 @@
 ; 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 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
@@ -11133,10 +10861,8 @@
 ; 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 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
@@ -11149,10 +10875,8 @@
 ; 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 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
@@ -11296,10 +11020,8 @@
 ; 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 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
@@ -11312,10 +11034,8 @@
 ; 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 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
@@ -11459,10 +11179,8 @@
 ; 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 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
@@ -11475,10 +11193,8 @@
 ; 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 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
@@ -11622,10 +11338,8 @@
 ; 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 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
@@ -11638,10 +11352,8 @@
 ; 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 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
@@ -11785,10 +11497,8 @@
 ; 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 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
@@ -11801,10 +11511,8 @@
 ; 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 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
@@ -11948,10 +11656,8 @@
 ; 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 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
@@ -11964,10 +11670,8 @@
 ; 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 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
@@ -12111,10 +11815,8 @@
 ; 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 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
@@ -12127,10 +11829,8 @@
 ; 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 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
@@ -12274,10 +11974,8 @@
 ; 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 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
@@ -12290,10 +11988,8 @@
 ; 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 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
@@ -12437,10 +12133,8 @@
 ; 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 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
@@ -12453,10 +12147,8 @@
 ; 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 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
@@ -12600,10 +12292,8 @@
 ; 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 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
@@ -12616,10 +12306,8 @@
 ; 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 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
@@ -12763,10 +12451,8 @@
 ; 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 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
@@ -12779,10 +12465,8 @@
 ; 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 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
@@ -12926,10 +12610,8 @@
 ; 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 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
@@ -12942,10 +12624,8 @@
 ; 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 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
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
@@ -119,11 +119,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -133,11 +131,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -258,11 +254,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -272,11 +266,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -408,14 +400,12 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -424,14 +414,12 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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
@@ -572,16 +560,14 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -590,16 +576,14 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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
@@ -703,8 +687,7 @@
 ; 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)
@@ -716,8 +699,7 @@
 ; 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)
@@ -821,8 +803,7 @@
 ; 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)
@@ -834,8 +815,7 @@
 ; 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)
@@ -953,8 +933,7 @@
 ; 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
@@ -968,8 +947,7 @@
 ; 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
@@ -1089,8 +1067,7 @@
 ; 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
@@ -1104,8 +1081,7 @@
 ; 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
@@ -1211,8 +1187,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1224,8 +1199,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1350,8 +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)
@@ -1366,8 +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)
@@ -1488,8 +1460,7 @@
 ; 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
@@ -1503,8 +1474,7 @@
 ; 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
@@ -1645,8 +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
@@ -1663,8 +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
@@ -1808,8 +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
@@ -1826,8 +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
@@ -1963,8 +1929,7 @@
 ; 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) lgkmcnt(0)
@@ -1980,8 +1945,7 @@
 ; 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) lgkmcnt(0)
@@ -2131,8 +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
@@ -2150,8 +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
@@ -2303,8 +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
@@ -2322,8 +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
@@ -2446,10 +2406,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2460,10 +2418,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2600,10 +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
@@ -2617,10 +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
@@ -2753,10 +2705,8 @@
 ; 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
@@ -2769,10 +2719,8 @@
 ; 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
@@ -2925,10 +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
@@ -2944,10 +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
@@ -3103,10 +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
@@ -3122,10 +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
@@ -3267,10 +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
@@ -3284,10 +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
@@ -3427,10 +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
@@ -3444,10 +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
@@ -3601,10 +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
@@ -3620,10 +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
@@ -3779,10 +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
@@ -3798,10 +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
@@ -3957,10 +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
@@ -3976,10 +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
@@ -4135,10 +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
@@ -4154,10 +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
@@ -4313,10 +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
@@ -4332,10 +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
@@ -4491,10 +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
@@ -4510,10 +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
@@ -4669,10 +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
@@ -4688,10 +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
@@ -4847,10 +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
@@ -4866,10 +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
@@ -5014,10 +4914,8 @@
 ; 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 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
@@ -5030,10 +4928,8 @@
 ; 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 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
@@ -5188,10 +5084,8 @@
 ; 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 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
@@ -5206,10 +5100,8 @@
 ; 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 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
@@ -5369,10 +5261,8 @@
 ; 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
@@ -5387,10 +5277,8 @@
 ; 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
@@ -5561,10 +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
@@ -5581,10 +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
@@ -5757,10 +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
@@ -5777,10 +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
@@ -5939,10 +5819,8 @@
 ; 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 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
@@ -5957,10 +5835,8 @@
 ; 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 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
@@ -6117,10 +5993,8 @@
 ; 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 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
@@ -6135,10 +6009,8 @@
 ; 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 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
@@ -6309,10 +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
@@ -6329,10 +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
@@ -6505,10 +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
@@ -6525,10 +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
@@ -6701,10 +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
@@ -6721,10 +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
@@ -6897,10 +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
@@ -6917,10 +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
@@ -7093,10 +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
@@ -7113,10 +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
@@ -7289,10 +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
@@ -7309,10 +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
@@ -7485,10 +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
@@ -7505,10 +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
@@ -7681,10 +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
@@ -7701,10 +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
@@ -7832,11 +7672,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -7846,11 +7684,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7971,11 +7807,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -7985,11 +7819,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -8127,14 +7959,12 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -8144,14 +7974,12 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -8299,16 +8127,14 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -8318,16 +8144,14 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -8432,8 +8256,7 @@
 ; 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)
@@ -8445,8 +8268,7 @@
 ; 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)
@@ -8550,8 +8372,7 @@
 ; 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)
@@ -8563,8 +8384,7 @@
 ; 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)
@@ -8682,8 +8502,7 @@
 ; 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
@@ -8697,8 +8516,7 @@
 ; 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
@@ -8818,8 +8636,7 @@
 ; 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
@@ -8833,8 +8650,7 @@
 ; 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
@@ -8940,8 +8756,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8953,8 +8768,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -9077,8 +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
@@ -9092,8 +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
@@ -9213,8 +9025,7 @@
 ; 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
@@ -9228,8 +9039,7 @@
 ; 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
@@ -9368,8 +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
@@ -9385,8 +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
@@ -9527,8 +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
@@ -9544,8 +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
@@ -9685,8 +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)
@@ -9703,8 +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)
@@ -9860,8 +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
@@ -9880,8 +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
@@ -10039,8 +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
@@ -10059,8 +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
@@ -10184,10 +9984,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10198,10 +9996,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10336,10 +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
@@ -10352,10 +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
@@ -10487,10 +10279,8 @@
 ; 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
@@ -10503,10 +10293,8 @@
 ; 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
@@ -10657,10 +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
@@ -10675,10 +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
@@ -10831,10 +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
@@ -10849,10 +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
@@ -10991,10 +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
@@ -11007,10 +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
@@ -11147,10 +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
@@ -11163,10 +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
@@ -11317,10 +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
@@ -11335,10 +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
@@ -11491,10 +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
@@ -11509,10 +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
@@ -11665,10 +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
@@ -11683,10 +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
@@ -11839,10 +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
@@ -11857,10 +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
@@ -12013,10 +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
@@ -12031,10 +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
@@ -12187,10 +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
@@ -12205,10 +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
@@ -12361,10 +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
@@ -12379,10 +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
@@ -12535,10 +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
@@ -12553,10 +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
@@ -12700,10 +12440,8 @@
 ; 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 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
@@ -12716,10 +12454,8 @@
 ; 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 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
@@ -12880,10 +12616,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12899,10 +12633,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13063,10 +12795,8 @@
 ; 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
@@ -13081,10 +12811,8 @@
 ; 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
@@ -13261,10 +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
@@ -13282,10 +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
@@ -13465,10 +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
@@ -13486,10 +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
@@ -13655,10 +13375,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -13674,10 +13392,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13841,10 +13557,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -13860,10 +13574,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -14041,10 +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
@@ -14062,10 +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
@@ -14245,10 +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
@@ -14266,10 +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
@@ -14449,10 +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
@@ -14470,10 +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
@@ -14653,10 +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
@@ -14674,10 +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
@@ -14857,10 +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
@@ -14878,10 +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
@@ -15061,10 +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
@@ -15082,10 +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
@@ -15265,10 +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
@@ -15286,10 +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
@@ -15469,10 +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
@@ -15490,10 +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
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
@@ -67,12 +67,10 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -82,12 +80,10 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -172,8 +168,7 @@
 ; 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)
@@ -189,8 +184,7 @@
 ; 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)
@@ -265,11 +259,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -280,11 +272,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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
@@ -363,11 +353,10 @@
 ; 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
@@ -380,11 +369,10 @@
 ; 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
@@ -463,13 +451,11 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -478,12 +464,10 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -553,8 +537,7 @@
 ; 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
@@ -568,8 +551,7 @@
 ; 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
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
@@ -119,11 +119,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -133,11 +131,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -258,11 +254,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -272,11 +266,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -397,11 +389,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -411,11 +401,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -536,11 +524,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -550,11 +536,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -659,8 +643,7 @@
 ; 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)
@@ -672,8 +655,7 @@
 ; 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)
@@ -777,8 +759,7 @@
 ; 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)
@@ -790,8 +771,7 @@
 ; 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)
@@ -895,8 +875,7 @@
 ; 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)
@@ -908,8 +887,7 @@
 ; 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)
@@ -1013,8 +991,7 @@
 ; 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)
@@ -1026,8 +1003,7 @@
 ; 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)
@@ -1131,8 +1107,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1144,8 +1119,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1249,8 +1223,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1262,8 +1235,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1367,8 +1339,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1380,8 +1351,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1485,8 +1455,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1498,8 +1467,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1603,8 +1571,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1616,8 +1583,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1737,8 +1703,7 @@
 ; 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) lgkmcnt(0)
@@ -1752,8 +1717,7 @@
 ; 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) lgkmcnt(0)
@@ -1876,8 +1840,7 @@
 ; 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) lgkmcnt(0)
@@ -1891,8 +1854,7 @@
 ; 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) lgkmcnt(0)
@@ -2015,8 +1977,7 @@
 ; 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) lgkmcnt(0)
@@ -2030,8 +1991,7 @@
 ; 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) lgkmcnt(0)
@@ -2150,10 +2110,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2164,10 +2122,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2283,10 +2239,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2297,10 +2251,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2416,10 +2368,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2430,10 +2380,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2549,10 +2497,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2563,10 +2509,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2682,10 +2626,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2696,10 +2638,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2815,10 +2755,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2829,10 +2767,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2948,10 +2884,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2962,10 +2896,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3081,10 +3013,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3095,10 +3025,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3214,10 +3142,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3228,10 +3154,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3347,10 +3271,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3361,10 +3283,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3480,10 +3400,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3494,10 +3412,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3613,10 +3529,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3627,10 +3541,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3746,10 +3658,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3760,10 +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:    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
@@ -3879,10 +3787,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3893,10 +3799,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4012,10 +3916,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -4026,10 +3928,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4169,10 +4069,8 @@
 ; 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 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
@@ -4185,10 +4083,8 @@
 ; 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 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
@@ -4332,10 +4228,8 @@
 ; 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 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
@@ -4348,10 +4242,8 @@
 ; 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 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
@@ -4495,10 +4387,8 @@
 ; 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 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
@@ -4511,10 +4401,8 @@
 ; 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 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
@@ -4658,10 +4546,8 @@
 ; 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 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
@@ -4674,10 +4560,8 @@
 ; 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 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
@@ -4821,10 +4705,8 @@
 ; 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 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
@@ -4837,10 +4719,8 @@
 ; 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 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
@@ -4984,10 +4864,8 @@
 ; 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 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
@@ -5000,10 +4878,8 @@
 ; 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 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
@@ -5147,10 +5023,8 @@
 ; 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 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
@@ -5163,10 +5037,8 @@
 ; 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 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
@@ -5310,10 +5182,8 @@
 ; 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 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
@@ -5326,10 +5196,8 @@
 ; 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 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
@@ -5473,10 +5341,8 @@
 ; 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 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
@@ -5489,10 +5355,8 @@
 ; 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 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
@@ -5636,10 +5500,8 @@
 ; 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 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
@@ -5652,10 +5514,8 @@
 ; 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 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
@@ -5799,10 +5659,8 @@
 ; 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 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
@@ -5815,10 +5673,8 @@
 ; 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 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
@@ -5962,10 +5818,8 @@
 ; 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 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
@@ -5978,10 +5832,8 @@
 ; 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 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
@@ -6125,10 +5977,8 @@
 ; 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 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
@@ -6141,10 +5991,8 @@
 ; 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 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
@@ -6288,10 +6136,8 @@
 ; 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 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
@@ -6304,10 +6150,8 @@
 ; 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 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
@@ -6451,10 +6295,8 @@
 ; 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 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
@@ -6467,10 +6309,8 @@
 ; 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 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
@@ -6594,11 +6434,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6608,11 +6446,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -6733,11 +6569,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6747,11 +6581,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -6872,11 +6704,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6886,11 +6716,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7011,11 +6839,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -7025,11 +6851,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7134,8 +6958,7 @@
 ; 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)
@@ -7147,8 +6970,7 @@
 ; 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)
@@ -7252,8 +7074,7 @@
 ; 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)
@@ -7265,8 +7086,7 @@
 ; 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)
@@ -7370,8 +7190,7 @@
 ; 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)
@@ -7383,8 +7202,7 @@
 ; 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)
@@ -7488,8 +7306,7 @@
 ; 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)
@@ -7501,8 +7318,7 @@
 ; 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)
@@ -7606,8 +7422,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7619,8 +7434,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7724,8 +7538,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7737,8 +7550,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7842,8 +7654,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7855,8 +7666,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7960,8 +7770,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7973,8 +7782,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8078,8 +7886,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8091,8 +7898,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8212,8 +8018,7 @@
 ; 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) lgkmcnt(0)
@@ -8227,8 +8032,7 @@
 ; 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) lgkmcnt(0)
@@ -8351,8 +8155,7 @@
 ; 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) lgkmcnt(0)
@@ -8366,8 +8169,7 @@
 ; 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) lgkmcnt(0)
@@ -8490,8 +8292,7 @@
 ; 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) lgkmcnt(0)
@@ -8505,8 +8306,7 @@
 ; 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) lgkmcnt(0)
@@ -8625,10 +8425,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8639,10 +8437,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8758,10 +8554,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8772,10 +8566,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8891,10 +8683,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8905,10 +8695,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9024,10 +8812,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9038,10 +8824,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9157,10 +8941,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9171,10 +8953,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9290,10 +9070,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9304,10 +9082,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9423,10 +9199,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9437,10 +9211,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9556,10 +9328,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9570,10 +9340,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9689,10 +9457,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9703,10 +9469,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9822,10 +9586,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9836,10 +9598,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9955,10 +9715,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9969,10 +9727,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10088,10 +9844,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10102,10 +9856,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10221,10 +9973,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10235,10 +9985,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10354,10 +10102,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10368,10 +10114,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10487,10 +10231,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10501,10 +10243,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10644,10 +10384,8 @@
 ; 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 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
@@ -10660,10 +10398,8 @@
 ; 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 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
@@ -10807,10 +10543,8 @@
 ; 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 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
@@ -10823,10 +10557,8 @@
 ; 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 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
@@ -10970,10 +10702,8 @@
 ; 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 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
@@ -10986,10 +10716,8 @@
 ; 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 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
@@ -11133,10 +10861,8 @@
 ; 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 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
@@ -11149,10 +10875,8 @@
 ; 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 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
@@ -11296,10 +11020,8 @@
 ; 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 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
@@ -11312,10 +11034,8 @@
 ; 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 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
@@ -11459,10 +11179,8 @@
 ; 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 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
@@ -11475,10 +11193,8 @@
 ; 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 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
@@ -11622,10 +11338,8 @@
 ; 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 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
@@ -11638,10 +11352,8 @@
 ; 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 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
@@ -11785,10 +11497,8 @@
 ; 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 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
@@ -11801,10 +11511,8 @@
 ; 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 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
@@ -11948,10 +11656,8 @@
 ; 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 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
@@ -11964,10 +11670,8 @@
 ; 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 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
@@ -12111,10 +11815,8 @@
 ; 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 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
@@ -12127,10 +11829,8 @@
 ; 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 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
@@ -12274,10 +11974,8 @@
 ; 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 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
@@ -12290,10 +11988,8 @@
 ; 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 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
@@ -12437,10 +12133,8 @@
 ; 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 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
@@ -12453,10 +12147,8 @@
 ; 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 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
@@ -12600,10 +12292,8 @@
 ; 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 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
@@ -12616,10 +12306,8 @@
 ; 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 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
@@ -12763,10 +12451,8 @@
 ; 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 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
@@ -12779,10 +12465,8 @@
 ; 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 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
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
@@ -119,11 +119,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -133,11 +131,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -258,11 +254,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -272,11 +266,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -405,13 +397,11 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -420,12 +410,10 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -563,15 +551,13 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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
@@ -580,13 +566,11 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -691,8 +675,7 @@
 ; 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)
@@ -704,8 +687,7 @@
 ; 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)
@@ -809,8 +791,7 @@
 ; 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)
@@ -822,8 +803,7 @@
 ; 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)
@@ -936,8 +916,7 @@
 ; 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
@@ -951,8 +930,7 @@
 ; 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
@@ -1066,8 +1044,7 @@
 ; 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
@@ -1081,8 +1058,7 @@
 ; 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
@@ -1187,8 +1163,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1200,8 +1175,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1317,8 +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)
@@ -1332,8 +1305,7 @@
 ; 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)
@@ -1447,8 +1419,7 @@
 ; 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
@@ -1462,8 +1433,7 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
@@ -1589,8 +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
@@ -1606,8 +1575,7 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
@@ -1734,8 +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
@@ -1751,8 +1718,7 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v[0:1], v2
@@ -1877,8 +1843,7 @@
 ; 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) lgkmcnt(0)
@@ -1893,8 +1858,7 @@
 ; 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) lgkmcnt(0)
@@ -2029,8 +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
@@ -2047,8 +2010,7 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
@@ -2184,8 +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
@@ -2202,8 +2163,7 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_swap_b32 v2, v[0:1], v2 glc
@@ -2323,10 +2283,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2337,10 +2295,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2468,10 +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
@@ -2484,10 +2438,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -2613,10 +2565,8 @@
 ; 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
@@ -2629,10 +2579,8 @@
 ; 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 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)
@@ -2770,10 +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
@@ -2788,10 +2734,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -2930,10 +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
@@ -2948,10 +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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3081,10 +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
@@ -3097,10 +3035,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -3229,10 +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
@@ -3245,10 +3179,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -3386,10 +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
@@ -3404,10 +3334,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3546,10 +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
@@ -3564,10 +3490,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3706,10 +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
@@ -3724,10 +3646,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -3866,10 +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
@@ -3884,10 +3802,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    flat_atomic_cmpswap_b32 v[0:1], v[2:3] offset:16
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
@@ -4029,10 +3945,8 @@
 ; 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 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
@@ -4045,10 +3959,8 @@
 ; 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 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
@@ -4198,10 +4110,8 @@
 ; 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 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
@@ -4215,10 +4125,8 @@
 ; 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 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
@@ -4371,10 +4279,8 @@
 ; 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
@@ -4389,10 +4295,8 @@
 ; 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 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)
@@ -4552,10 +4456,8 @@
 ; 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
@@ -4571,10 +4473,8 @@
 ; 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 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)
@@ -4734,10 +4634,8 @@
 ; 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
@@ -4753,10 +4651,8 @@
 ; 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 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)
@@ -4907,10 +4803,8 @@
 ; 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 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
@@ -4924,10 +4818,8 @@
 ; 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 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
@@ -5077,10 +4969,8 @@
 ; 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 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
@@ -5094,10 +4984,8 @@
 ; 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 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
@@ -5256,10 +5144,8 @@
 ; 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
@@ -5275,10 +5161,8 @@
 ; 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 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)
@@ -5438,10 +5322,8 @@
 ; 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
@@ -5457,10 +5339,8 @@
 ; 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 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)
@@ -5620,10 +5500,8 @@
 ; 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
@@ -5639,10 +5517,8 @@
 ; 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 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)
@@ -5802,10 +5678,8 @@
 ; 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
@@ -5821,10 +5695,8 @@
 ; 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 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)
@@ -5984,10 +5856,8 @@
 ; 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
@@ -6003,10 +5873,8 @@
 ; 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 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)
@@ -6166,10 +6034,8 @@
 ; 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
@@ -6185,10 +6051,8 @@
 ; 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 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)
@@ -6348,10 +6212,8 @@
 ; 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
@@ -6367,10 +6229,8 @@
 ; 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 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)
@@ -6530,10 +6390,8 @@
 ; 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
@@ -6549,10 +6407,8 @@
 ; 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 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)
@@ -6677,11 +6533,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6691,11 +6545,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -6816,11 +6668,9 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6830,11 +6680,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -6959,13 +6807,11 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -6975,11 +6821,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7108,15 +6952,13 @@
 ; GFX11-WGP:       ; %bb.0: ; %entry
 ; GFX11-WGP-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-WGP-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-WGP-NEXT:    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)
@@ -7126,11 +6968,9 @@
 ; GFX11-CU:       ; %bb.0: ; %entry
 ; GFX11-CU-NEXT:    s_load_b128 s[0:3], s[0:1], 0x0
 ; GFX11-CU-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX11-CU-NEXT:    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)
@@ -7235,8 +7075,7 @@
 ; 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)
@@ -7248,8 +7087,7 @@
 ; 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)
@@ -7353,8 +7191,7 @@
 ; 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)
@@ -7366,8 +7203,7 @@
 ; 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)
@@ -7475,8 +7311,7 @@
 ; 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
@@ -7490,8 +7325,7 @@
 ; 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)
@@ -7599,8 +7433,7 @@
 ; 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
@@ -7614,8 +7447,7 @@
 ; 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)
@@ -7719,8 +7551,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7732,8 +7563,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7843,8 +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
@@ -7857,8 +7686,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -7966,8 +7794,7 @@
 ; 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
@@ -7981,8 +7808,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8096,8 +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
@@ -8112,8 +7937,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8227,8 +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
@@ -8243,8 +8066,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -8368,8 +8190,7 @@
 ; 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)
@@ -8385,8 +8206,7 @@
 ; 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) lgkmcnt(0)
@@ -8517,8 +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
@@ -8536,8 +8355,7 @@
 ; 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) lgkmcnt(0)
@@ -8668,8 +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
@@ -8687,8 +8504,7 @@
 ; 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) lgkmcnt(0)
@@ -8807,10 +8623,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8821,10 +8635,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8946,10 +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
@@ -8961,10 +8771,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9084,10 +8892,8 @@
 ; 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
@@ -9100,10 +8906,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9229,10 +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
@@ -9246,10 +9048,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9375,10 +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
@@ -9392,10 +9190,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9517,10 +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
@@ -9532,10 +9326,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9657,10 +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
@@ -9672,10 +9462,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9801,10 +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
@@ -9818,10 +9604,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9947,10 +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
@@ -9964,10 +9746,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10093,10 +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
@@ -10110,10 +9888,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10239,10 +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
@@ -10256,10 +10030,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10385,10 +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
@@ -10402,10 +10172,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10531,10 +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
@@ -10548,10 +10314,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10677,10 +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
@@ -10694,10 +10456,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10823,10 +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
@@ -10840,10 +10598,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10983,10 +10739,8 @@
 ; 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 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
@@ -10999,10 +10753,8 @@
 ; 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 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
@@ -11150,10 +10902,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -11168,10 +10918,8 @@
 ; 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 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
@@ -11319,10 +11067,8 @@
 ; 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
@@ -11337,10 +11083,8 @@
 ; 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 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
@@ -11492,10 +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
@@ -11512,10 +11254,8 @@
 ; 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 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
@@ -11667,10 +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
@@ -11687,10 +11425,8 @@
 ; 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 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
@@ -11838,10 +11574,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -11856,10 +11590,8 @@
 ; 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 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
@@ -12007,10 +11739,8 @@
 ; 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 v2, v[0:1], v[2:3] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12025,10 +11755,8 @@
 ; 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 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
@@ -12180,10 +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
@@ -12200,10 +11926,8 @@
 ; 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 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
@@ -12355,10 +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
@@ -12375,10 +12097,8 @@
 ; 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 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
@@ -12530,10 +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
@@ -12550,10 +12268,8 @@
 ; 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 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
@@ -12705,10 +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
@@ -12725,10 +12439,8 @@
 ; 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 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
@@ -12880,10 +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
@@ -12900,10 +12610,8 @@
 ; 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 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
@@ -13055,10 +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
@@ -13075,10 +12781,8 @@
 ; 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 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
@@ -13230,10 +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
@@ -13250,10 +12952,8 @@
 ; 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 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
@@ -13405,10 +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
@@ -13425,10 +13123,8 @@
 ; 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 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
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
@@ -675,9 +675,8 @@
 ; 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
@@ -687,9 +686,8 @@
 ; 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
@@ -800,9 +798,8 @@
 ; 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
@@ -812,9 +809,8 @@
 ; 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
@@ -938,9 +934,8 @@
 ; 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]
@@ -952,9 +947,8 @@
 ; 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]
@@ -1080,9 +1074,8 @@
 ; 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]
@@ -1094,9 +1087,8 @@
 ; 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]
@@ -1209,9 +1201,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1221,9 +1212,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1353,9 +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
@@ -1367,9 +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
@@ -1495,9 +1483,8 @@
 ; 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]
@@ -1509,9 +1496,8 @@
 ; 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]
@@ -1656,9 +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]
@@ -1672,9 +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]
@@ -1821,9 +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]
@@ -1837,9 +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]
@@ -1982,9 +1964,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1998,9 +1979,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -2157,9 +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
@@ -2175,9 +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
@@ -2336,9 +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
@@ -2354,9 +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
@@ -2483,8 +2459,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2496,8 +2471,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2637,8 +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
@@ -2652,8 +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
@@ -2789,8 +2761,7 @@
 ; 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
@@ -2804,8 +2775,7 @@
 ; 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
@@ -2960,8 +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
@@ -2977,8 +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
@@ -3135,8 +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
@@ -3152,8 +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
@@ -3297,8 +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
@@ -3312,8 +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
@@ -3455,8 +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
@@ -3470,8 +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
@@ -3626,8 +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
@@ -3643,8 +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
@@ -3801,8 +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
@@ -3818,8 +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
@@ -3976,8 +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
@@ -3993,8 +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
@@ -4151,8 +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
@@ -4168,8 +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
@@ -4326,8 +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
@@ -4343,8 +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
@@ -4501,8 +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
@@ -4518,8 +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
@@ -4676,8 +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
@@ -4693,8 +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
@@ -4851,8 +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
@@ -4868,8 +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
@@ -5014,8 +4960,7 @@
 ; 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 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]
@@ -5029,8 +4974,7 @@
 ; 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 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]
@@ -5185,8 +5129,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -5202,8 +5145,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -5363,8 +5305,7 @@
 ; 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
@@ -5380,8 +5321,7 @@
 ; 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
@@ -5551,8 +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
@@ -5570,8 +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
@@ -5743,8 +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
@@ -5762,8 +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
@@ -5922,8 +5858,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -5939,8 +5874,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -6097,8 +6031,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -6114,8 +6047,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -6285,8 +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
@@ -6304,8 +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
@@ -6477,8 +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
@@ -6496,8 +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
@@ -6669,8 +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
@@ -6688,8 +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
@@ -6861,8 +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
@@ -6880,8 +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
@@ -7053,8 +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
@@ -7072,8 +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
@@ -7245,8 +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
@@ -7264,8 +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
@@ -7437,8 +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
@@ -7456,8 +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
@@ -7629,8 +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
@@ -7648,8 +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
@@ -8332,9 +8248,8 @@
 ; 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
@@ -8344,9 +8259,8 @@
 ; 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
@@ -8457,9 +8371,8 @@
 ; 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
@@ -8469,9 +8382,8 @@
 ; 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
@@ -8595,9 +8507,8 @@
 ; 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]
@@ -8609,9 +8520,8 @@
 ; 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]
@@ -8737,9 +8647,8 @@
 ; 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]
@@ -8751,9 +8660,8 @@
 ; 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]
@@ -8866,9 +8774,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8878,9 +8785,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9010,9 +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
@@ -9024,9 +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
@@ -9152,9 +9056,8 @@
 ; 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]
@@ -9166,9 +9069,8 @@
 ; 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]
@@ -9313,9 +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]
@@ -9329,9 +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]
@@ -9478,9 +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]
@@ -9494,9 +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]
@@ -9639,9 +9537,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -9655,9 +9552,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -9814,9 +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
@@ -9832,9 +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
@@ -9993,9 +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
@@ -10011,9 +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
@@ -10140,8 +10032,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10153,8 +10044,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10294,8 +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
@@ -10309,8 +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
@@ -10446,8 +10334,7 @@
 ; 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
@@ -10461,8 +10348,7 @@
 ; 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
@@ -10617,8 +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
@@ -10634,8 +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
@@ -10792,8 +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
@@ -10809,8 +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
@@ -10954,8 +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
@@ -10969,8 +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
@@ -11112,8 +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
@@ -11127,8 +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
@@ -11283,8 +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
@@ -11300,8 +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
@@ -11458,8 +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
@@ -11475,8 +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
@@ -11633,8 +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
@@ -11650,8 +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
@@ -11808,8 +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
@@ -11825,8 +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
@@ -11983,8 +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
@@ -12000,8 +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
@@ -12158,8 +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
@@ -12175,8 +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
@@ -12333,8 +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
@@ -12350,8 +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
@@ -12508,8 +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
@@ -12525,8 +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
@@ -12671,8 +12533,7 @@
 ; 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 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]
@@ -12686,8 +12547,7 @@
 ; 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 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]
@@ -12842,8 +12702,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12859,8 +12718,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13030,8 +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
@@ -13049,8 +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
@@ -13222,8 +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
@@ -13241,8 +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
@@ -13401,8 +13255,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -13418,8 +13271,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13576,8 +13428,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -13593,8 +13444,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13764,8 +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
@@ -13783,8 +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
@@ -13956,8 +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
@@ -13975,8 +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
@@ -14148,8 +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
@@ -14167,8 +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
@@ -14340,8 +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
@@ -14359,8 +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
@@ -14532,8 +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
@@ -14551,8 +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
@@ -14724,8 +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
@@ -14743,8 +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
@@ -14916,8 +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
@@ -14935,8 +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
@@ -15108,8 +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
@@ -15127,8 +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
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
@@ -121,11 +121,10 @@
 ; 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
@@ -133,11 +132,10 @@
 ; 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
@@ -267,8 +265,7 @@
 ; 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)
@@ -279,8 +276,7 @@
 ; 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)
@@ -406,11 +402,10 @@
 ; 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
@@ -418,11 +413,10 @@
 ; 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
@@ -547,11 +541,10 @@
 ; 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
@@ -559,11 +552,10 @@
 ; 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
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
@@ -640,9 +640,8 @@
 ; 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
@@ -652,9 +651,8 @@
 ; 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
@@ -765,9 +763,8 @@
 ; 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
@@ -777,9 +774,8 @@
 ; 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
@@ -890,9 +886,8 @@
 ; 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
@@ -902,9 +897,8 @@
 ; 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
@@ -1015,9 +1009,8 @@
 ; 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
@@ -1027,9 +1020,8 @@
 ; 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
@@ -1140,9 +1132,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1152,9 +1143,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1265,9 +1255,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1277,9 +1266,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1390,9 +1378,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1402,9 +1389,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1515,9 +1501,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1527,9 +1512,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1640,9 +1624,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1652,9 +1635,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1783,9 +1765,8 @@
 ; 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 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]
@@ -1797,9 +1778,8 @@
 ; 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 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]
@@ -1931,9 +1911,8 @@
 ; 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 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]
@@ -1945,9 +1924,8 @@
 ; 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 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]
@@ -2079,9 +2057,8 @@
 ; 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 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]
@@ -2093,9 +2070,8 @@
 ; 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 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]
@@ -2218,8 +2194,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2231,8 +2206,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2353,8 +2327,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2366,8 +2339,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2488,8 +2460,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2501,8 +2472,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2623,8 +2593,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2636,8 +2605,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2758,8 +2726,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2771,8 +2738,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2893,8 +2859,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2906,8 +2871,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3028,8 +2992,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3041,8 +3004,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3163,8 +3125,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3176,8 +3137,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3298,8 +3258,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3311,8 +3270,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3433,8 +3391,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3446,8 +3403,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3568,8 +3524,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3581,8 +3536,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3703,8 +3657,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3716,8 +3669,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3838,8 +3790,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3851,8 +3802,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3973,8 +3923,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3986,8 +3935,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4108,8 +4056,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -4121,8 +4068,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4263,8 +4209,7 @@
 ; 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 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]
@@ -4278,8 +4223,7 @@
 ; 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 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]
@@ -4424,8 +4368,7 @@
 ; 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 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]
@@ -4439,8 +4382,7 @@
 ; 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 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]
@@ -4585,8 +4527,7 @@
 ; 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 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]
@@ -4600,8 +4541,7 @@
 ; 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 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]
@@ -4746,8 +4686,7 @@
 ; 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 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]
@@ -4761,8 +4700,7 @@
 ; 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 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]
@@ -4907,8 +4845,7 @@
 ; 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 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]
@@ -4922,8 +4859,7 @@
 ; 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 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]
@@ -5068,8 +5004,7 @@
 ; 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 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]
@@ -5083,8 +5018,7 @@
 ; 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 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]
@@ -5229,8 +5163,7 @@
 ; 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 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]
@@ -5244,8 +5177,7 @@
 ; 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 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]
@@ -5390,8 +5322,7 @@
 ; 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 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]
@@ -5405,8 +5336,7 @@
 ; 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 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]
@@ -5551,8 +5481,7 @@
 ; 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 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]
@@ -5566,8 +5495,7 @@
 ; 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 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]
@@ -5712,8 +5640,7 @@
 ; 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 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]
@@ -5727,8 +5654,7 @@
 ; 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 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]
@@ -5873,8 +5799,7 @@
 ; 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 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]
@@ -5888,8 +5813,7 @@
 ; 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 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]
@@ -6034,8 +5958,7 @@
 ; 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 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]
@@ -6049,8 +5972,7 @@
 ; 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 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]
@@ -6195,8 +6117,7 @@
 ; 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 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]
@@ -6210,8 +6131,7 @@
 ; 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 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]
@@ -6356,8 +6276,7 @@
 ; 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 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]
@@ -6371,8 +6290,7 @@
 ; 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 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]
@@ -6517,8 +6435,7 @@
 ; 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 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]
@@ -6532,8 +6449,7 @@
 ; 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 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]
@@ -7177,9 +7093,8 @@
 ; 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
@@ -7189,9 +7104,8 @@
 ; 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
@@ -7302,9 +7216,8 @@
 ; 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
@@ -7314,9 +7227,8 @@
 ; 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
@@ -7427,9 +7339,8 @@
 ; 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
@@ -7439,9 +7350,8 @@
 ; 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
@@ -7552,9 +7462,8 @@
 ; 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
@@ -7564,9 +7473,8 @@
 ; 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
@@ -7677,9 +7585,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7689,9 +7596,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -7802,9 +7708,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7814,9 +7719,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -7927,9 +7831,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7939,9 +7842,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8052,9 +7954,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8064,9 +7965,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8177,9 +8077,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8189,9 +8088,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8320,9 +8218,8 @@
 ; 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 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]
@@ -8334,9 +8231,8 @@
 ; 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 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]
@@ -8468,9 +8364,8 @@
 ; 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 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]
@@ -8482,9 +8377,8 @@
 ; 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 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]
@@ -8616,9 +8510,8 @@
 ; 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 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]
@@ -8630,9 +8523,8 @@
 ; 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 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]
@@ -8755,8 +8647,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8768,8 +8659,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8890,8 +8780,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8903,8 +8792,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9025,8 +8913,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9038,8 +8925,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9160,8 +9046,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9173,8 +9058,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9295,8 +9179,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9308,8 +9191,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9430,8 +9312,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9443,8 +9324,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9565,8 +9445,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9578,8 +9457,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9700,8 +9578,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9713,8 +9590,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9835,8 +9711,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9848,8 +9723,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9970,8 +9844,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9983,8 +9856,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10105,8 +9977,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10118,8 +9989,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10240,8 +10110,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10253,8 +10122,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10375,8 +10243,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10388,8 +10255,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10510,8 +10376,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10523,8 +10388,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10645,8 +10509,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10658,8 +10521,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10800,8 +10662,7 @@
 ; 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 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]
@@ -10815,8 +10676,7 @@
 ; 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 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]
@@ -10961,8 +10821,7 @@
 ; 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 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]
@@ -10976,8 +10835,7 @@
 ; 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 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]
@@ -11122,8 +10980,7 @@
 ; 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 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]
@@ -11137,8 +10994,7 @@
 ; 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 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]
@@ -11283,8 +11139,7 @@
 ; 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 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]
@@ -11298,8 +11153,7 @@
 ; 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 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]
@@ -11444,8 +11298,7 @@
 ; 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 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]
@@ -11459,8 +11312,7 @@
 ; 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 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]
@@ -11605,8 +11457,7 @@
 ; 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 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]
@@ -11620,8 +11471,7 @@
 ; 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 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]
@@ -11766,8 +11616,7 @@
 ; 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 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]
@@ -11781,8 +11630,7 @@
 ; 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 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]
@@ -11927,8 +11775,7 @@
 ; 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 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]
@@ -11942,8 +11789,7 @@
 ; 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 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]
@@ -12088,8 +11934,7 @@
 ; 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 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]
@@ -12103,8 +11948,7 @@
 ; 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 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]
@@ -12249,8 +12093,7 @@
 ; 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 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]
@@ -12264,8 +12107,7 @@
 ; 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 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]
@@ -12410,8 +12252,7 @@
 ; 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 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]
@@ -12425,8 +12266,7 @@
 ; 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 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]
@@ -12571,8 +12411,7 @@
 ; 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 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]
@@ -12586,8 +12425,7 @@
 ; 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 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]
@@ -12732,8 +12570,7 @@
 ; 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 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]
@@ -12747,8 +12584,7 @@
 ; 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 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]
@@ -12893,8 +12729,7 @@
 ; 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 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]
@@ -12908,8 +12743,7 @@
 ; 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 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]
@@ -13054,8 +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:    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]
@@ -13069,8 +12902,7 @@
 ; 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 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]
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
@@ -679,9 +679,8 @@
 ; 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
@@ -691,9 +690,8 @@
 ; 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
@@ -804,9 +802,8 @@
 ; 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
@@ -816,9 +813,8 @@
 ; 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
@@ -944,9 +940,8 @@
 ; 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]
@@ -958,9 +953,8 @@
 ; 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]
@@ -1088,9 +1082,8 @@
 ; 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]
@@ -1102,9 +1095,8 @@
 ; 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]
@@ -1217,9 +1209,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1229,9 +1220,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1363,9 +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
@@ -1377,9 +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
@@ -1507,9 +1495,8 @@
 ; 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]
@@ -1521,9 +1508,8 @@
 ; 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]
@@ -1672,9 +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]
@@ -1688,9 +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]
@@ -1841,9 +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]
@@ -1857,9 +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]
@@ -2004,9 +1986,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -2020,9 +2001,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -2183,9 +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
@@ -2201,9 +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
@@ -2366,9 +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
@@ -2384,9 +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
@@ -2513,8 +2489,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2526,8 +2501,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2669,8 +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
@@ -2684,8 +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
@@ -2823,8 +2795,7 @@
 ; 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
@@ -2838,8 +2809,7 @@
 ; 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
@@ -2998,8 +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
@@ -3015,8 +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
@@ -3177,8 +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
@@ -3194,8 +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
@@ -3341,8 +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
@@ -3356,8 +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
@@ -3501,8 +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
@@ -3516,8 +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
@@ -3676,8 +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
@@ -3693,8 +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
@@ -3855,8 +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
@@ -3872,8 +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
@@ -4034,8 +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
@@ -4051,8 +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
@@ -4213,8 +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
@@ -4230,8 +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
@@ -4376,8 +4330,7 @@
 ; 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 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]
@@ -4391,8 +4344,7 @@
 ; 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 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]
@@ -4549,8 +4501,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -4566,8 +4517,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -4741,8 +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
@@ -4760,8 +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
@@ -4937,8 +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
@@ -4956,8 +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
@@ -5118,8 +5064,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -5135,8 +5080,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -5295,8 +5239,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -5312,8 +5255,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -5487,8 +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
@@ -5506,8 +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
@@ -5683,8 +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
@@ -5702,8 +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
@@ -5879,8 +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
@@ -5898,8 +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
@@ -6075,8 +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
@@ -6094,8 +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
@@ -6271,8 +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
@@ -6290,8 +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
@@ -6467,8 +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
@@ -6486,8 +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
@@ -6663,8 +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
@@ -6682,8 +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
@@ -6859,8 +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
@@ -6878,8 +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
@@ -7566,9 +7492,8 @@
 ; 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
@@ -7578,9 +7503,8 @@
 ; 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
@@ -7691,9 +7615,8 @@
 ; 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
@@ -7703,9 +7626,8 @@
 ; 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
@@ -7831,9 +7753,8 @@
 ; 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]
@@ -7845,9 +7766,8 @@
 ; 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]
@@ -7975,9 +7895,8 @@
 ; 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]
@@ -7989,9 +7908,8 @@
 ; 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]
@@ -8104,9 +8022,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8116,9 +8033,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8250,9 +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
@@ -8264,9 +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
@@ -8394,9 +8308,8 @@
 ; 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]
@@ -8408,9 +8321,8 @@
 ; 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]
@@ -8559,9 +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]
@@ -8575,9 +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]
@@ -8728,9 +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]
@@ -8744,9 +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]
@@ -8891,9 +8799,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -8907,9 +8814,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -9070,9 +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
@@ -9088,9 +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
@@ -9253,9 +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
@@ -9271,9 +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
@@ -9400,8 +9302,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9413,8 +9314,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9556,8 +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
@@ -9571,8 +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
@@ -9710,8 +9608,7 @@
 ; 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
@@ -9725,8 +9622,7 @@
 ; 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
@@ -9885,8 +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
@@ -9902,8 +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
@@ -10064,8 +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
@@ -10081,8 +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
@@ -10228,8 +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
@@ -10243,8 +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
@@ -10388,8 +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
@@ -10403,8 +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
@@ -10563,8 +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
@@ -10580,8 +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
@@ -10742,8 +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
@@ -10759,8 +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
@@ -10921,8 +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
@@ -10938,8 +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
@@ -11100,8 +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
@@ -11117,8 +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
@@ -11279,8 +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
@@ -11296,8 +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
@@ -11458,8 +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
@@ -11475,8 +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
@@ -11637,8 +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
@@ -11654,8 +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
@@ -11816,8 +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
@@ -11833,8 +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
@@ -11979,8 +11851,7 @@
 ; 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 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]
@@ -11994,8 +11865,7 @@
 ; 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 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]
@@ -12152,8 +12022,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12169,8 +12038,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -12332,8 +12200,7 @@
 ; 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
@@ -12349,8 +12216,7 @@
 ; 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
@@ -12524,8 +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
@@ -12543,8 +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
@@ -12720,8 +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
@@ -12739,8 +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
@@ -12901,8 +12763,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12918,8 +12779,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13078,8 +12938,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -13095,8 +12954,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-CU-NEXT:    buffer_gl0_inv
@@ -13270,8 +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
@@ -13289,8 +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
@@ -13466,8 +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
@@ -13485,8 +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
@@ -13662,8 +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
@@ -13681,8 +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
@@ -13858,8 +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
@@ -13877,8 +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
@@ -14054,8 +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
@@ -14073,8 +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
@@ -14250,8 +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
@@ -14269,8 +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
@@ -14446,8 +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
@@ -14465,8 +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
@@ -14642,8 +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
@@ -14661,8 +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
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
@@ -174,8 +174,7 @@
 ; 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)
@@ -186,8 +185,7 @@
 ; 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)
@@ -274,11 +272,10 @@
 ; 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)
@@ -287,11 +284,10 @@
 ; 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)
@@ -378,11 +374,10 @@
 ; 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)
@@ -391,11 +386,10 @@
 ; 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)
@@ -570,9 +564,8 @@
 ; 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]
@@ -584,9 +577,8 @@
 ; 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)
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
@@ -640,9 +640,8 @@
 ; 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
@@ -652,9 +651,8 @@
 ; 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
@@ -765,9 +763,8 @@
 ; 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
@@ -777,9 +774,8 @@
 ; 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
@@ -890,9 +886,8 @@
 ; 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
@@ -902,9 +897,8 @@
 ; 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
@@ -1015,9 +1009,8 @@
 ; 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
@@ -1027,9 +1020,8 @@
 ; 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
@@ -1140,9 +1132,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1152,9 +1143,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1265,9 +1255,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1277,9 +1266,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1390,9 +1378,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1402,9 +1389,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1515,9 +1501,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1527,9 +1512,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1640,9 +1624,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1652,9 +1635,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1783,9 +1765,8 @@
 ; 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 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]
@@ -1797,9 +1778,8 @@
 ; 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 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]
@@ -1931,9 +1911,8 @@
 ; 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 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]
@@ -1945,9 +1924,8 @@
 ; 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 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]
@@ -2079,9 +2057,8 @@
 ; 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 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]
@@ -2093,9 +2070,8 @@
 ; 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 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]
@@ -2218,8 +2194,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2231,8 +2206,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2353,8 +2327,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2366,8 +2339,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2488,8 +2460,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2501,8 +2472,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2623,8 +2593,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2636,8 +2605,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2758,8 +2726,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2771,8 +2738,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2893,8 +2859,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2906,8 +2871,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3028,8 +2992,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3041,8 +3004,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3163,8 +3125,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3176,8 +3137,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3298,8 +3258,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3311,8 +3270,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3433,8 +3391,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3446,8 +3403,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3568,8 +3524,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3581,8 +3536,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3703,8 +3657,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3716,8 +3669,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3838,8 +3790,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3851,8 +3802,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3973,8 +3923,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -3986,8 +3935,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4108,8 +4056,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -4121,8 +4068,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -4263,8 +4209,7 @@
 ; 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 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]
@@ -4278,8 +4223,7 @@
 ; 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 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]
@@ -4424,8 +4368,7 @@
 ; 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 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]
@@ -4439,8 +4382,7 @@
 ; 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 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]
@@ -4585,8 +4527,7 @@
 ; 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 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]
@@ -4600,8 +4541,7 @@
 ; 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 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]
@@ -4746,8 +4686,7 @@
 ; 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 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]
@@ -4761,8 +4700,7 @@
 ; 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 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]
@@ -4907,8 +4845,7 @@
 ; 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 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]
@@ -4922,8 +4859,7 @@
 ; 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 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]
@@ -5068,8 +5004,7 @@
 ; 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 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]
@@ -5083,8 +5018,7 @@
 ; 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 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]
@@ -5229,8 +5163,7 @@
 ; 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 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]
@@ -5244,8 +5177,7 @@
 ; 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 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]
@@ -5390,8 +5322,7 @@
 ; 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 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]
@@ -5405,8 +5336,7 @@
 ; 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 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]
@@ -5551,8 +5481,7 @@
 ; 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 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]
@@ -5566,8 +5495,7 @@
 ; 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 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]
@@ -5712,8 +5640,7 @@
 ; 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 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]
@@ -5727,8 +5654,7 @@
 ; 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 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]
@@ -5873,8 +5799,7 @@
 ; 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 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]
@@ -5888,8 +5813,7 @@
 ; 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 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]
@@ -6034,8 +5958,7 @@
 ; 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 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]
@@ -6049,8 +5972,7 @@
 ; 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 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]
@@ -6195,8 +6117,7 @@
 ; 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 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]
@@ -6210,8 +6131,7 @@
 ; 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 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]
@@ -6356,8 +6276,7 @@
 ; 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 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]
@@ -6371,8 +6290,7 @@
 ; 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 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]
@@ -6517,8 +6435,7 @@
 ; 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 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]
@@ -6532,8 +6449,7 @@
 ; 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 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]
@@ -7177,9 +7093,8 @@
 ; 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
@@ -7189,9 +7104,8 @@
 ; 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
@@ -7302,9 +7216,8 @@
 ; 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
@@ -7314,9 +7227,8 @@
 ; 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
@@ -7427,9 +7339,8 @@
 ; 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
@@ -7439,9 +7350,8 @@
 ; 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
@@ -7552,9 +7462,8 @@
 ; 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
@@ -7564,9 +7473,8 @@
 ; 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
@@ -7677,9 +7585,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7689,9 +7596,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -7802,9 +7708,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7814,9 +7719,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -7927,9 +7831,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -7939,9 +7842,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8052,9 +7954,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8064,9 +7965,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8177,9 +8077,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8189,9 +8088,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8320,9 +8218,8 @@
 ; 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 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]
@@ -8334,9 +8231,8 @@
 ; 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 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]
@@ -8468,9 +8364,8 @@
 ; 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 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]
@@ -8482,9 +8377,8 @@
 ; 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 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]
@@ -8616,9 +8510,8 @@
 ; 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 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]
@@ -8630,9 +8523,8 @@
 ; 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 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]
@@ -8755,8 +8647,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8768,8 +8659,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8890,8 +8780,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8903,8 +8792,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9025,8 +8913,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9038,8 +8925,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9160,8 +9046,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9173,8 +9058,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9295,8 +9179,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9308,8 +9191,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9430,8 +9312,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9443,8 +9324,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9565,8 +9445,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9578,8 +9457,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9700,8 +9578,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9713,8 +9590,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9835,8 +9711,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9848,8 +9723,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9970,8 +9844,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9983,8 +9856,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10105,8 +9977,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10118,8 +9989,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10240,8 +10110,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10253,8 +10122,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10375,8 +10243,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10388,8 +10255,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10510,8 +10376,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10523,8 +10388,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10645,8 +10509,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -10658,8 +10521,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10800,8 +10662,7 @@
 ; 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 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]
@@ -10815,8 +10676,7 @@
 ; 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 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]
@@ -10961,8 +10821,7 @@
 ; 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 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]
@@ -10976,8 +10835,7 @@
 ; 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 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]
@@ -11122,8 +10980,7 @@
 ; 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 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]
@@ -11137,8 +10994,7 @@
 ; 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 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]
@@ -11283,8 +11139,7 @@
 ; 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 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]
@@ -11298,8 +11153,7 @@
 ; 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 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]
@@ -11444,8 +11298,7 @@
 ; 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 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]
@@ -11459,8 +11312,7 @@
 ; 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 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]
@@ -11605,8 +11457,7 @@
 ; 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 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]
@@ -11620,8 +11471,7 @@
 ; 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 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]
@@ -11766,8 +11616,7 @@
 ; 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 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]
@@ -11781,8 +11630,7 @@
 ; 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 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]
@@ -11927,8 +11775,7 @@
 ; 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 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]
@@ -11942,8 +11789,7 @@
 ; 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 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]
@@ -12088,8 +11934,7 @@
 ; 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 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]
@@ -12103,8 +11948,7 @@
 ; 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 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]
@@ -12249,8 +12093,7 @@
 ; 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 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]
@@ -12264,8 +12107,7 @@
 ; 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 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]
@@ -12410,8 +12252,7 @@
 ; 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 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]
@@ -12425,8 +12266,7 @@
 ; 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 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]
@@ -12571,8 +12411,7 @@
 ; 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 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]
@@ -12586,8 +12425,7 @@
 ; 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 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]
@@ -12732,8 +12570,7 @@
 ; 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 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]
@@ -12747,8 +12584,7 @@
 ; 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 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]
@@ -12893,8 +12729,7 @@
 ; 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 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]
@@ -12908,8 +12743,7 @@
 ; 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 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]
@@ -13054,8 +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:    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]
@@ -13069,8 +12902,7 @@
 ; 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 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]
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
@@ -653,9 +653,8 @@
 ; 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
@@ -665,9 +664,8 @@
 ; 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
@@ -778,9 +776,8 @@
 ; 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
@@ -790,9 +787,8 @@
 ; 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
@@ -913,9 +909,8 @@
 ; 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]
@@ -927,9 +922,8 @@
 ; 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)
@@ -1051,9 +1045,8 @@
 ; 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]
@@ -1065,9 +1058,8 @@
 ; 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)
@@ -1179,9 +1171,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -1191,9 +1182,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1310,9 +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
@@ -1323,9 +1312,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -1446,9 +1434,8 @@
 ; 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]
@@ -1460,9 +1447,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1590,9 +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]
@@ -1605,9 +1590,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1735,9 +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]
@@ -1750,9 +1733,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v0, v1, s[0:1]
 ; GFX11-CU-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
@@ -1885,9 +1867,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -1900,9 +1881,8 @@
 ; 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 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]
@@ -2047,9 +2027,8 @@
 ; 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
@@ -2064,9 +2043,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
@@ -2212,9 +2190,8 @@
 ; 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
@@ -2229,9 +2206,8 @@
 ; 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 lgkmcnt(0)
 ; GFX11-CU-NEXT:    global_atomic_swap_b32 v1, v0, v1, s[0:1] glc
 ; GFX11-CU-NEXT:    s_waitcnt vmcnt(0)
@@ -2355,8 +2331,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -2368,8 +2343,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2496,8 +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
@@ -2510,8 +2483,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -2642,8 +2614,7 @@
 ; 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
@@ -2657,8 +2628,7 @@
 ; 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 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)
@@ -2796,8 +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
@@ -2812,8 +2781,7 @@
 ; 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 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)
@@ -2951,8 +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
@@ -2967,8 +2934,7 @@
 ; 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 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)
@@ -3096,8 +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
@@ -3110,8 +3075,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3238,8 +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
@@ -3252,8 +3215,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -3390,8 +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
@@ -3406,8 +3367,7 @@
 ; 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 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)
@@ -3545,8 +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
@@ -3561,8 +3520,7 @@
 ; 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 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)
@@ -3700,8 +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
@@ -3716,8 +3673,7 @@
 ; 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 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)
@@ -3855,8 +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
@@ -3871,8 +3826,7 @@
 ; 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 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)
@@ -4010,8 +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
@@ -4026,8 +3979,7 @@
 ; 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 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)
@@ -4165,8 +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
@@ -4181,8 +4132,7 @@
 ; 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 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)
@@ -4320,8 +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
@@ -4336,8 +4285,7 @@
 ; 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 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)
@@ -4475,8 +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
@@ -4491,8 +4438,7 @@
 ; 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 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)
@@ -4634,8 +4580,7 @@
 ; 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 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]
@@ -4649,8 +4594,7 @@
 ; 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 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]
@@ -4798,8 +4742,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -4814,8 +4757,7 @@
 ; 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 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]
@@ -4970,8 +4912,7 @@
 ; 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
@@ -4987,8 +4928,7 @@
 ; 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 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)
@@ -5147,8 +5087,7 @@
 ; 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
@@ -5165,8 +5104,7 @@
 ; 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 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)
@@ -5325,8 +5263,7 @@
 ; 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
@@ -5343,8 +5280,7 @@
 ; 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 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)
@@ -5493,8 +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:    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
@@ -5509,8 +5444,7 @@
 ; 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 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]
@@ -5658,8 +5592,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -5674,8 +5607,7 @@
 ; 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 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]
@@ -5833,8 +5765,7 @@
 ; 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
@@ -5851,8 +5782,7 @@
 ; 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 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)
@@ -6011,8 +5941,7 @@
 ; 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
@@ -6029,8 +5958,7 @@
 ; 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 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)
@@ -6189,8 +6117,7 @@
 ; 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
@@ -6207,8 +6134,7 @@
 ; 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 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)
@@ -6367,8 +6293,7 @@
 ; 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
@@ -6385,8 +6310,7 @@
 ; 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 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)
@@ -6545,8 +6469,7 @@
 ; 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
@@ -6563,8 +6486,7 @@
 ; 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 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)
@@ -6723,8 +6645,7 @@
 ; 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
@@ -6741,8 +6662,7 @@
 ; 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 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)
@@ -6901,8 +6821,7 @@
 ; 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
@@ -6919,8 +6838,7 @@
 ; 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 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)
@@ -7079,8 +6997,7 @@
 ; 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
@@ -7097,8 +7014,7 @@
 ; 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 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)
@@ -7753,9 +7669,8 @@
 ; 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
@@ -7765,9 +7680,8 @@
 ; 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
@@ -7878,9 +7792,8 @@
 ; 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
@@ -7890,9 +7803,8 @@
 ; 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
@@ -8007,9 +7919,8 @@
 ; 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]
@@ -8021,9 +7932,8 @@
 ; 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
@@ -8138,9 +8048,8 @@
 ; 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]
@@ -8152,9 +8061,8 @@
 ; 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
@@ -8265,9 +8173,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -8277,9 +8184,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8396,9 +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
@@ -8409,9 +8314,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8526,9 +8430,8 @@
 ; 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]
@@ -8540,9 +8443,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8663,9 +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]
@@ -8678,9 +8579,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8801,9 +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]
@@ -8816,9 +8715,8 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -8950,9 +8848,8 @@
 ; 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 v1, v0, v1, s[0:1] glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -8965,9 +8862,8 @@
 ; 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 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]
@@ -9106,9 +9002,8 @@
 ; 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
@@ -9123,9 +9018,8 @@
 ; 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 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]
@@ -9264,9 +9158,8 @@
 ; 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
@@ -9281,9 +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:    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]
@@ -9406,8 +9298,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-WGP-NEXT:    s_endpgm
@@ -9419,8 +9310,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9547,8 +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
@@ -9561,8 +9450,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9687,8 +9575,7 @@
 ; 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
@@ -9702,8 +9589,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9834,8 +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
@@ -9850,8 +9735,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -9982,8 +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
@@ -9998,8 +9881,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10126,8 +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
@@ -10140,8 +10021,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10268,8 +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
@@ -10282,8 +10161,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10414,8 +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
@@ -10430,8 +10307,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10562,8 +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
@@ -10578,8 +10453,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10710,8 +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
@@ -10726,8 +10599,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -10858,8 +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
@@ -10874,8 +10745,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -11006,8 +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
@@ -11022,8 +10891,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -11154,8 +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
@@ -11170,8 +11037,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -11302,8 +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
@@ -11318,8 +11183,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -11450,8 +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
@@ -11466,8 +11329,7 @@
 ; 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_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-CU-NEXT:    s_endpgm
@@ -11608,8 +11470,7 @@
 ; 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 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]
@@ -11623,8 +11484,7 @@
 ; 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 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]
@@ -11772,8 +11632,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -11788,8 +11647,7 @@
 ; 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 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]
@@ -11938,8 +11796,7 @@
 ; 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
@@ -11955,8 +11812,7 @@
 ; 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 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]
@@ -12108,8 +11964,7 @@
 ; 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
@@ -12126,8 +11981,7 @@
 ; 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 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]
@@ -12279,8 +12133,7 @@
 ; 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
@@ -12297,8 +12150,7 @@
 ; 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 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]
@@ -12446,8 +12298,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12462,8 +12313,7 @@
 ; 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 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]
@@ -12611,8 +12461,7 @@
 ; 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 v0, v2, v[0:1], s[0:1] offset:16 glc
 ; GFX11-WGP-NEXT:    s_waitcnt vmcnt(0)
 ; GFX11-WGP-NEXT:    buffer_gl0_inv
@@ -12627,8 +12476,7 @@
 ; 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 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]
@@ -12780,8 +12628,7 @@
 ; 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
@@ -12798,8 +12645,7 @@
 ; 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 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]
@@ -12951,8 +12797,7 @@
 ; 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
@@ -12969,8 +12814,7 @@
 ; 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 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]
@@ -13122,8 +12966,7 @@
 ; 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
@@ -13140,8 +12983,7 @@
 ; 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 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]
@@ -13293,8 +13135,7 @@
 ; 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
@@ -13311,8 +13152,7 @@
 ; 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 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]
@@ -13464,8 +13304,7 @@
 ; 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
@@ -13482,8 +13321,7 @@
 ; 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 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]
@@ -13635,8 +13473,7 @@
 ; 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
@@ -13653,8 +13490,7 @@
 ; 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 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]
@@ -13806,8 +13642,7 @@
 ; 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
@@ -13824,8 +13659,7 @@
 ; 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 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]
@@ -13977,8 +13811,7 @@
 ; 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
@@ -13995,8 +13828,7 @@
 ; 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 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]
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
@@ -133,9 +133,8 @@
 ; 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]
@@ -147,9 +146,8 @@
 ; 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]
@@ -439,9 +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
 ;
@@ -452,9 +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
@@ -81,9 +81,8 @@
 ; 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]
@@ -95,9 +94,8 @@
 ; 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]
@@ -291,9 +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
 ;
@@ -304,9 +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) {
@@ -484,8 +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
@@ -496,8 +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
@@ -569,8 +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
@@ -580,8 +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
@@ -333,9 +333,8 @@
 ; 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)
@@ -346,9 +345,8 @@
 ; 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)
@@ -684,11 +682,10 @@
 ; 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
@@ -698,11 +695,10 @@
 ; 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
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
@@ -223,9 +223,8 @@
 ; 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)
@@ -236,9 +235,8 @@
 ; 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)
@@ -480,11 +478,10 @@
 ; 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)
@@ -495,11 +492,10 @@
 ; 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)
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 $vgpr0, 0, 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 $vgpr0, 0, 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 $vgpr0, 0, 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