[AMDGPU] Construct memory clauses before RA

[rampitec]

Summary

Memory clauses are formed into bundles in presence of xnack.
Their source operands are marked as early-clobber.
This allows to allocate distinct source and destination registers within a clause and prevent breaking the clause with s_nop in the hazard recognizer.
Clauses are undone before post-RA scheduler to allow some rescheduling, which will not break the clause since artificial edges are created in the dag to keep memory operations together. Yet this allows a better ILP in some cases.

[rampitec]

Stripped patch off the parent revision D47509 changes.

[mgrang]

lib/Target/AMDGPU/SIFormMemoryClauses.cpp

171 ↗

(On Diff #149023)

Please use llvm::sort instead of std::sort. See https://llvm.org/docs/CodingStandards.html#beware-of-non-deterministic-sorting-order-of-equal-elements.

[rampitec]

Replaced std::sort with llvm::sort.

[arsenm]

Needs some tests with frame indexes, and with the d16 lo/hi operations with tied operands

lib/Target/AMDGPU/SIFormMemoryClauses.cpp

32–36 ↗

(On Diff #149040)

I thought this was a subtarget specific limit.

I apparently never committed D40578 which made the hazard recognizer aware of the subtarget specific sizes

41 ↗

(On Diff #149040)

Should this be using the BitVector of RegUnits instead?

312 ↗

(On Diff #149040)

SReg_32_XM0_XEXEC?

363 ↗

(On Diff #149040)

Why && (and for the rest)?

lib/Target/AMDGPU/SIInstructions.td

548–550 ↗

(On Diff #149040)

Why do you use this instead of using the generic TargetOpcode::BUNDLE?
I think trying to use a custom bundle instruction is going to cause issues. For example instruction size computation for this won't work properly, and any other code that expects to see BUNDLE

[rampitec]

In D47511#1116028, @arsenm wrote:

Needs some tests with frame indexes, and with the d16 lo/hi operations with tied operands

There is a test with frame indexes in mir (function name "stack"). Pass skips all instructions with FI operands, so the only test needed is to check there are no clauses formed.

lib/Target/AMDGPU/SIFormMemoryClauses.cpp

32–36 ↗

(On Diff #149040)

4 bits is a smallest counter for all ASICs (excluding exp, but we are not using exp here). On GFX9 there are two additional bits for vmcnt, but it does not make too much sense to create too big clauses, so to fit into 4 bit seems reasonable at least as a starting point.

Take in mind we cannot be precise here because if there are outstanding operations counter may be higher than zero when we come to a clause. In this case HW will break it anyway when it is saturated. Thus the value is not calculated from available counter bits or hardcoded, but rather a tunable option.

41 ↗

(On Diff #149040)

RegUnits are for Phys. Anyway, this carries lane bitmask *and* register state, so I need two values.

312 ↗

(On Diff #149040)

When SIRegisterInfo::getReservedRegs() works it reserves SGPR_32 for amdgpu-num-sgpr attribute. This limit accounts primarily for that attribute use (and debug reserved registers too). I also do not believe I want to use vcc or anything exotic here, I'd rather split the clause if we are running that low of registers.

363 ↗

(On Diff #149040)

I could use const auto &, but forwarding reference is less verbose (and usually recommended in fact).

lib/Target/AMDGPU/SIInstructions.td

548–550 ↗

(On Diff #149040)

Primarily to know what to remove post-RA. I do not think it creates any problems though, for instance getFunctionCodeSize() is called when memory clauses are already unbundled.

[rampitec]

Added tests for d16 and d16_hi loads. Since operands are tied we skip creating a clause for them.

[rampitec]

Replaced custom instruction with TargetOpcode::BUNDLE.

lib/Target/AMDGPU/SIInstructions.td

548–550 ↗

(On Diff #149040)

Actually I can do it with BUNDLE too. Resulting bundle returns mayLoad() and can be identified.

[vpykhtin]

lib/Target/AMDGPU/SIFormMemoryClauses.cpp

242 ↗

(On Diff #149225)

moveMaxPressure actually clears accumulated MaxPressure - is this intended?

[rampitec]

lib/Target/AMDGPU/SIFormMemoryClauses.cpp

242 ↗

(On Diff #149225)

That is the only interface we made here ;) It does not really matter, as the next step will recreate MaxPressure. Take in mind that pressure cannot go down within the clause, only after. Therefor CurPressure is always a MaxPressure too while inside the clause.

[vpykhtin]

lib/Target/AMDGPU/SIFormMemoryClauses.cpp

242 ↗

(On Diff #149225)

ok, then. I will add something like getMaxPressure next time.

llvm/trunk/lib/Target/AMDGPU/AMDGPU.h

 FunctionPass *createSIWholeQuadModePass();
 FunctionPass *createSIFixControlFlowLiveIntervalsPass();
 FunctionPass *createSIOptimizeExecMaskingPreRAPass();
 FunctionPass *createSIFixSGPRCopiesPass();
 FunctionPass *createSIMemoryLegalizerPass();
 FunctionPass *createSIDebuggerInsertNopsPass();
 FunctionPass *createSIInsertWaitcntsPass();
 FunctionPass *createSIFixWWMLivenessPass();
+FunctionPass *createSIFormMemoryClausesPass();
 FunctionPass *createAMDGPUSimplifyLibCallsPass(const TargetOptions &);
 FunctionPass *createAMDGPUUseNativeCallsPass();
 FunctionPass *createAMDGPUCodeGenPreparePass();
 FunctionPass *createAMDGPUMachineCFGStructurizerPass();
 FunctionPass *createAMDGPURewriteOutArgumentsPass();
 
 void initializeAMDGPUDAGToDAGISelPass(PassRegistry&);
 

 extern char &SIMemoryLegalizerID;
 
 void initializeSIDebuggerInsertNopsPass(PassRegistry&);
 extern char &SIDebuggerInsertNopsID;
 
 void initializeSIInsertWaitcntsPass(PassRegistry&);
 extern char &SIInsertWaitcntsID;
 
+void initializeSIFormMemoryClausesPass(PassRegistry&);
+extern char &SIFormMemoryClausesID;
+
 void initializeAMDGPUUnifyDivergentExitNodesPass(PassRegistry&);
 extern char &AMDGPUUnifyDivergentExitNodesID;
 
 ImmutablePass *createAMDGPUAAWrapperPass();
 void initializeAMDGPUAAWrapperPassPass(PassRegistry&);
 
 void initializeAMDGPUArgumentUsageInfoPass(PassRegistry &);
 

llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

extern "C" void LLVMInitializeAMDGPUTarget() {
   initializeSIInsertWaitcntsPass(*PR);
   initializeSIWholeQuadModePass(*PR);
   initializeSILowerControlFlowPass(*PR);
   initializeSIInsertSkipsPass(*PR);
   initializeSIMemoryLegalizerPass(*PR);
   initializeSIDebuggerInsertNopsPass(*PR);
   initializeSIOptimizeExecMaskingPass(*PR);
   initializeSIFixWWMLivenessPass(*PR);
+  initializeSIFormMemoryClausesPass(*PR);
   initializeAMDGPUUnifyDivergentExitNodesPass(*PR);
   initializeAMDGPUAAWrapperPassPass(*PR);
   initializeAMDGPUUseNativeCallsPass(*PR);
   initializeAMDGPUSimplifyLibCallsPass(*PR);
   initializeAMDGPUInlinerPass(*PR);
 }
 
 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
void GCNPassConfig::addFastRegAlloc(FunctionPass *RegAllocPass) {
   insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);
 
   TargetPassConfig::addFastRegAlloc(RegAllocPass);
 }
 
 void GCNPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
   insertPass(&MachineSchedulerID, &SIOptimizeExecMaskingPreRAID);
 
+  insertPass(&SIOptimizeExecMaskingPreRAID, &SIFormMemoryClausesID);
+
   // This must be run immediately after phi elimination and before
   // TwoAddressInstructions, otherwise the processing of the tied operand of
   // SI_ELSE will introduce a copy of the tied operand source after the else.
   insertPass(&PHIEliminationID, &SILowerControlFlowID, false);
 
   // This must be run after SILowerControlFlow, since it needs to use the
   // machine-level CFG, but before register allocation.
   insertPass(&SILowerControlFlowID, &SIFixWWMLivenessID, false);

llvm/trunk/lib/Target/AMDGPU/CMakeLists.txt

add_llvm_target(AMDGPUCodeGen
   R600Packetizer.cpp
   R600RegisterInfo.cpp
   SIAnnotateControlFlow.cpp
   SIDebuggerInsertNops.cpp
   SIFixSGPRCopies.cpp
   SIFixVGPRCopies.cpp
   SIFixWWMLiveness.cpp
   SIFoldOperands.cpp
+  SIFormMemoryClauses.cpp
   SIFrameLowering.cpp
   SIInsertSkips.cpp
   SIInsertWaitcnts.cpp
   SIInstrInfo.cpp
   SIISelLowering.cpp
   SILoadStoreOptimizer.cpp
   SILowerControlFlow.cpp
   SILowerI1Copies.cpp

llvm/trunk/lib/Target/AMDGPU/SIFormMemoryClauses.cpp

+//===-- SIFormMemoryClauses.cpp -------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+/// \file
+/// This pass creates bundles of SMEM and VMEM instructions forming memory
+/// clauses if XNACK is enabled. Def operands of clauses are marked as early
+/// clobber to make sure we will not override any source within a clause.
+///
+//===----------------------------------------------------------------------===//
+
+#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
+#include "GCNRegPressure.h"
+#include "SIInstrInfo.h"
+#include "SIMachineFunctionInfo.h"
+#include "SIRegisterInfo.h"
+#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/CodeGen/LiveIntervals.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "si-form-memory-clauses"
+
+// Clauses longer then 15 instructions would overflow one of the counters
+// and stall. They can stall even earlier if there are outstanding counters.
+static cl::opt<unsigned>
+MaxClause("amdgpu-max-memory-clause", cl::Hidden, cl::init(15),
+          cl::desc("Maximum length of a memory clause, instructions"));
+
+namespace {
+
+class SIFormMemoryClauses : public MachineFunctionPass {
+  typedef DenseMap<unsigned, std::pair<unsigned, LaneBitmask>> RegUse;
+
+public:
+  static char ID;
+
+public:
+  SIFormMemoryClauses() : MachineFunctionPass(ID) {
+    initializeSIFormMemoryClausesPass(*PassRegistry::getPassRegistry());
+  }
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+  StringRef getPassName() const override {
+    return "SI Form memory clauses";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<LiveIntervals>();
+    AU.setPreservesAll();
+    MachineFunctionPass::getAnalysisUsage(AU);
+  }
+
+private:
+  template <typename Callable>
+  void forAllLanes(unsigned Reg, LaneBitmask LaneMask, Callable Func) const;
+
+  bool canBundle(const MachineInstr &MI, RegUse &Defs, RegUse &Uses) const;
+  bool checkPressure(const MachineInstr &MI, GCNDownwardRPTracker &RPT);
+  void collectRegUses(const MachineInstr &MI, RegUse &Defs, RegUse &Uses) const;
+  bool processRegUses(const MachineInstr &MI, RegUse &Defs, RegUse &Uses,
+                      GCNDownwardRPTracker &RPT);
+
+  const SISubtarget *ST;
+  const SIRegisterInfo *TRI;
+  const MachineRegisterInfo *MRI;
+  SIMachineFunctionInfo *MFI;
+
+  unsigned LastRecordedOccupancy;
+  unsigned MaxVGPRs;
+  unsigned MaxSGPRs;
+};
+
+} // End anonymous namespace.
+
+INITIALIZE_PASS_BEGIN(SIFormMemoryClauses, DEBUG_TYPE,
+                      "SI Form memory clauses", false, false)
+INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
+INITIALIZE_PASS_END(SIFormMemoryClauses, DEBUG_TYPE,
+                    "SI Form memory clauses", false, false)
+
+
+char SIFormMemoryClauses::ID = 0;
+
+char &llvm::SIFormMemoryClausesID = SIFormMemoryClauses::ID;
+
+FunctionPass *llvm::createSIFormMemoryClausesPass() {
+  return new SIFormMemoryClauses();
+}
+
+static bool isVMEMClauseInst(const MachineInstr &MI) {
+  return SIInstrInfo::isFLAT(MI) || SIInstrInfo::isVMEM(MI);
+}
+
+static bool isSMEMClauseInst(const MachineInstr &MI) {
+  return SIInstrInfo::isSMRD(MI);
+}
+
+// There no sense to create store clauses, they do not define anything,
+// thus there is nothing to set early-clobber.
+static bool isValidClauseInst(const MachineInstr &MI, bool IsVMEMClause) {
+  if (MI.isDebugValue() || MI.isBundled())
+    return false;
+  if (!MI.mayLoad() || MI.mayStore())
+    return false;
+  if (AMDGPU::getAtomicNoRetOp(MI.getOpcode()) != -1 ||
+      AMDGPU::getAtomicRetOp(MI.getOpcode()) != -1)
+    return false;
+  if (IsVMEMClause && !isVMEMClauseInst(MI))
+    return false;
+  if (!IsVMEMClause && !isSMEMClauseInst(MI))
+    return false;
+  return true;
+}
+
+static unsigned getMopState(const MachineOperand &MO) {
+  unsigned S = 0;
+  if (MO.isImplicit())
+    S |= RegState::Implicit;
+  if (MO.isDead())
+    S |= RegState::Dead;
+  if (MO.isUndef())
+    S |= RegState::Undef;
+  if (MO.isKill())
+    S |= RegState::Kill;
+  if (MO.isEarlyClobber())
+    S |= RegState::EarlyClobber;
+  if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()) && MO.isRenamable())
+    S |= RegState::Renamable;
+  return S;
+}
+
+template <typename Callable>
+void SIFormMemoryClauses::forAllLanes(unsigned Reg, LaneBitmask LaneMask,
+                                      Callable Func) const {
+  if (LaneMask.all() || TargetRegisterInfo::isPhysicalRegister(Reg) ||
+      LaneMask == MRI->getMaxLaneMaskForVReg(Reg)) {
+    Func(0);
+    return;
+  }
+
+  const TargetRegisterClass *RC = MRI->getRegClass(Reg);
+  unsigned E = TRI->getNumSubRegIndices();
+  SmallVector<unsigned, AMDGPU::NUM_TARGET_SUBREGS> CoveringSubregs;
+  for (unsigned Idx = 1; Idx < E; ++Idx) {
+    // Is this index even compatible with the given class?
+    if (TRI->getSubClassWithSubReg(RC, Idx) != RC)
+      continue;
+    LaneBitmask SubRegMask = TRI->getSubRegIndexLaneMask(Idx);
+    // Early exit if we found a perfect match.
+    if (SubRegMask == LaneMask) {
+      Func(Idx);
+      return;
+    }
+
+    if ((SubRegMask & ~LaneMask).any() || (SubRegMask & LaneMask).none())
+      continue;
+
+    CoveringSubregs.push_back(Idx);
+  }
+
+  llvm::sort(CoveringSubregs.begin(), CoveringSubregs.end(),
+             [this](unsigned A, unsigned B) {
+               LaneBitmask MaskA = TRI->getSubRegIndexLaneMask(A);
+               LaneBitmask MaskB = TRI->getSubRegIndexLaneMask(B);
+               unsigned NA = MaskA.getNumLanes();
+               unsigned NB = MaskB.getNumLanes();
+               if (NA != NB)
+                 return NA > NB;
+               return MaskA.getHighestLane() > MaskB.getHighestLane();
+             });
+
+  for (unsigned Idx : CoveringSubregs) {
+    LaneBitmask SubRegMask = TRI->getSubRegIndexLaneMask(Idx);
+    if ((SubRegMask & ~LaneMask).any() || (SubRegMask & LaneMask).none())
+      continue;
+
+    Func(Idx);
+    LaneMask &= ~SubRegMask;
+    if (LaneMask.none())
+      return;
+  }
+
+  llvm_unreachable("Failed to find all subregs to cover lane mask");
+}
+
+// Returns false if there is a use of a def already in the map.
+// In this case we must break the clause.
+bool SIFormMemoryClauses::canBundle(const MachineInstr &MI,
+                                    RegUse &Defs, RegUse &Uses) const {
+  // Check interference with defs.
+  for (const MachineOperand &MO : MI.operands()) {
+    // TODO: Prologue/Epilogue Insertion pass does not process bundled
+    //       instructions.
+    if (MO.isFI())
+      return false;
+
+    if (!MO.isReg())
+      continue;
+
+    unsigned Reg = MO.getReg();
+
+    // If it is tied we will need to write same register as we read.
+    if (MO.isTied())
+      return false;
+
+    RegUse &Map = MO.isDef() ? Uses : Defs;
+    auto Conflict = Map.find(Reg);
+    if (Conflict == Map.end())
+      continue;
+
+    if (TargetRegisterInfo::isPhysicalRegister(Reg))
+      return false;
+
+    LaneBitmask Mask = TRI->getSubRegIndexLaneMask(MO.getSubReg());
+    if ((Conflict->second.second & Mask).any())
+      return false;
+  }
+
+  return true;
+}
+
+// Since all defs in the clause are early clobber we can run out of registers.
+// Function returns false if pressure would hit the limit if instruction is
+// bundled into a memory clause.
+bool SIFormMemoryClauses::checkPressure(const MachineInstr &MI,
+                                        GCNDownwardRPTracker &RPT) {
+  // NB: skip advanceBeforeNext() call. Since all defs will be marked
+  // early-clobber they will all stay alive at least to the end of the
+  // clause. Therefor we should not decrease pressure even if load
+  // pointer becomes dead and could otherwise be reused for destination.
+  RPT.advanceToNext();
+  GCNRegPressure MaxPressure = RPT.moveMaxPressure();
+  unsigned Occupancy = MaxPressure.getOccupancy(*ST);
+  if (Occupancy >= MFI->getMinAllowedOccupancy() &&
+      MaxPressure.getVGPRNum() <= MaxVGPRs &&
+      MaxPressure.getSGPRNum() <= MaxSGPRs) {
+    LastRecordedOccupancy = Occupancy;
+    return true;
+  }
+  return false;
+}
+
+// Collect register defs and uses along with their lane masks and states.
+void SIFormMemoryClauses::collectRegUses(const MachineInstr &MI,
+                                         RegUse &Defs, RegUse &Uses) const {
+  for (const MachineOperand &MO : MI.operands()) {
+    if (!MO.isReg())
+      continue;
+    unsigned Reg = MO.getReg();
+    if (!Reg)
+      continue;
+
+    LaneBitmask Mask = TargetRegisterInfo::isVirtualRegister(Reg) ?
+                         TRI->getSubRegIndexLaneMask(MO.getSubReg()) :
+                         LaneBitmask::getAll();
+    RegUse &Map = MO.isDef() ? Defs : Uses;
+
+    auto Loc = Map.find(Reg);
+    unsigned State = getMopState(MO);
+    if (Loc == Map.end()) {
+      Map[Reg] = std::make_pair(State, Mask);
+    } else {
+      Loc->second.first |= State;
+      Loc->second.second |= Mask;
+    }
+  }
+}
+
+// Check register def/use conflicts, occupancy limits and collect def/use maps.
+// Return true if instruction can be bundled with previous. It it cannot
+// def/use maps are not updated.
+bool SIFormMemoryClauses::processRegUses(const MachineInstr &MI,
+                                         RegUse &Defs, RegUse &Uses,
+                                         GCNDownwardRPTracker &RPT) {
+  if (!canBundle(MI, Defs, Uses))
+    return false;
+
+  if (!checkPressure(MI, RPT))
+    return false;
+
+  collectRegUses(MI, Defs, Uses);
+  return true;
+}
+
+bool SIFormMemoryClauses::runOnMachineFunction(MachineFunction &MF) {
+  if (skipFunction(MF.getFunction()))
+    return false;
+
+  ST = &MF.getSubtarget<SISubtarget>();
+  if (!ST->isXNACKEnabled())
+    return false;
+
+  const SIInstrInfo *TII = ST->getInstrInfo();
+  TRI = ST->getRegisterInfo();
+  MRI = &MF.getRegInfo();
+  MFI = MF.getInfo<SIMachineFunctionInfo>();
+  LiveIntervals *LIS = &getAnalysis<LiveIntervals>();
+  SlotIndexes *Ind = LIS->getSlotIndexes();
+  bool Changed = false;
+
+  MaxVGPRs = TRI->getAllocatableSet(MF, &AMDGPU::VGPR_32RegClass).count();
+  MaxSGPRs = TRI->getAllocatableSet(MF, &AMDGPU::SGPR_32RegClass).count();
+
+  for (MachineBasicBlock &MBB : MF) {
+    MachineBasicBlock::instr_iterator Next;
+    for (auto I = MBB.instr_begin(), E = MBB.instr_end(); I != E; I = Next) {
+      MachineInstr &MI = *I;
+      Next = std::next(I);
+
+      bool IsVMEM = isVMEMClauseInst(MI);
+
+      if (!isValidClauseInst(MI, IsVMEM))
+        continue;
+
+      RegUse Defs, Uses;
+      GCNDownwardRPTracker RPT(*LIS);
+      RPT.reset(MI);
+
+      if (!processRegUses(MI, Defs, Uses, RPT))
+        continue;
+
+      unsigned Length = 1;
+      for ( ; Next != E && Length < MaxClause; ++Next) {
+        if (!isValidClauseInst(*Next, IsVMEM))
+          break;
+
+        // A load from pointer which was loaded inside the same bundle is an
+        // impossible clause because we will need to write and read the same
+        // register inside. In this case processRegUses will return false.
+        if (!processRegUses(*Next, Defs, Uses, RPT))
+          break;
+
+        ++Length;
+      }
+      if (Length < 2)
+        continue;
+
+      Changed = true;
+      MFI->limitOccupancy(LastRecordedOccupancy);
+
+      auto B = BuildMI(MBB, I, DebugLoc(), TII->get(TargetOpcode::BUNDLE));
+      Ind->insertMachineInstrInMaps(*B);
+
+      for (auto BI = I; BI != Next; ++BI) {
+        BI->bundleWithPred();
+        Ind->removeSingleMachineInstrFromMaps(*BI);
+
+        for (MachineOperand &MO : BI->defs())
+          if (MO.readsReg())
+            MO.setIsInternalRead(true);
+      }
+
+      for (auto &&R : Defs) {
+        forAllLanes(R.first, R.second.second, [&R, &B](unsigned SubReg) {
+          unsigned S = R.second.first | RegState::EarlyClobber;
+          if (!SubReg)
+            S &= ~(RegState::Undef | RegState::Dead);
+          B.addDef(R.first, S, SubReg);
+        });
+      }
+
+      for (auto &&R : Uses) {
+        forAllLanes(R.first, R.second.second, [&R, &B](unsigned SubReg) {
+          B.addUse(R.first, R.second.first & ~RegState::Kill, SubReg);
+        });
+      }
+
+      for (auto &&R : Defs) {
+        unsigned Reg = R.first;
+        Uses.erase(Reg);
+        if (TargetRegisterInfo::isPhysicalRegister(Reg))
+          continue;
+        LIS->removeInterval(Reg);
+        LIS->createAndComputeVirtRegInterval(Reg);
+      }
+
+      for (auto &&R : Uses) {
+        unsigned Reg = R.first;
+        if (TargetRegisterInfo::isPhysicalRegister(Reg))
+          continue;
+        LIS->removeInterval(Reg);
+        LIS->createAndComputeVirtRegInterval(Reg);
+      }
+    }
+  }
+
+  return Changed;
+}

llvm/trunk/lib/Target/AMDGPU/SIInstrInfo.cpp

case AMDGPU::SI_PC_ADD_REL_OFFSET: {
     break;
   }
   case AMDGPU::EXIT_WWM: {
     // This only gets its own opcode so that SIFixWWMLiveness can tell when WWM
     // is exited.
     MI.setDesc(get(AMDGPU::S_MOV_B64));
     break;
   }
+  case TargetOpcode::BUNDLE: {
+    if (!MI.mayLoad())
+      return false;
+
+    // If it is a load it must be a memory clause
+    for (MachineBasicBlock::instr_iterator I = MI.getIterator();
+         I->isBundledWithSucc(); ++I) {
+      I->unbundleFromSucc();
+      for (MachineOperand &MO : I->operands())
+        if (MO.isReg())
+          MO.setIsInternalRead(false);
+    }
+
+    MI.eraseFromParent();
+    break;
+  }
   }
   return true;
 }
 
 bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
                                       MachineOperand &Src0,
                                       unsigned Src0OpName,
                                       MachineOperand &Src1,

llvm/trunk/test/CodeGen/AMDGPU/memory_clause.ll

+; RUN: llc -march=amdgcn -mcpu=gfx902 -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+; GCN-LABEL: {{^}}vector_clause:
+; GCN:      global_load_dwordx4
+; GCN-NEXT: global_load_dwordx4
+; GCN-NEXT: global_load_dwordx4
+; GCN-NEXT: global_load_dwordx4
+; GCN-NEXT: s_nop
+define amdgpu_kernel void @vector_clause(<4 x i32> addrspace(1)* noalias nocapture readonly %arg, <4 x i32> addrspace(1)* noalias nocapture %arg1) {
+bb:
+  %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %tmp2 = zext i32 %tmp to i64
+  %tmp3 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg, i64 %tmp2
+  %tmp4 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp3, align 16
+  %tmp5 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg1, i64 %tmp2
+  %tmp6 = add nuw nsw i64 %tmp2, 1
+  %tmp7 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg, i64 %tmp6
+  %tmp8 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp7, align 16
+  %tmp9 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg1, i64 %tmp6
+  %tmp10 = add nuw nsw i64 %tmp2, 2
+  %tmp11 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg, i64 %tmp10
+  %tmp12 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp11, align 16
+  %tmp13 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg1, i64 %tmp10
+  %tmp14 = add nuw nsw i64 %tmp2, 3
+  %tmp15 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg, i64 %tmp14
+  %tmp16 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp15, align 16
+  %tmp17 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg1, i64 %tmp14
+  store <4 x i32> %tmp4, <4 x i32> addrspace(1)* %tmp5, align 16
+  store <4 x i32> %tmp8, <4 x i32> addrspace(1)* %tmp9, align 16
+  store <4 x i32> %tmp12, <4 x i32> addrspace(1)* %tmp13, align 16
+  store <4 x i32> %tmp16, <4 x i32> addrspace(1)* %tmp17, align 16
+  ret void
+}
+
+; GCN-LABEL: {{^}}scalar_clause:
+; GCN:      s_load_dwordx2
+; GCN-NEXT: s_load_dwordx2
+; GCN-NEXT: s_nop
+; GCN-NEXT: s_waitcnt lgkmcnt(0)
+; GCN-NEXT: s_load_dwordx4
+; GCN-NEXT: s_load_dwordx4
+; GCN-NEXT: s_load_dwordx4
+; GCN-NEXT: s_load_dwordx4
+define amdgpu_kernel void @scalar_clause(<4 x i32> addrspace(1)* noalias nocapture readonly %arg, <4 x i32> addrspace(1)* noalias nocapture %arg1) {
+bb:
+  %tmp = load <4 x i32>, <4 x i32> addrspace(1)* %arg, align 16
+  %tmp2 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg, i64 1
+  %tmp3 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp2, align 16
+  %tmp4 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg1, i64 1
+  %tmp5 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg, i64 2
+  %tmp6 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp5, align 16
+  %tmp7 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg1, i64 2
+  %tmp8 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg, i64 3
+  %tmp9 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp8, align 16
+  %tmp10 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg1, i64 3
+  store <4 x i32> %tmp, <4 x i32> addrspace(1)* %arg1, align 16
+  store <4 x i32> %tmp3, <4 x i32> addrspace(1)* %tmp4, align 16
+  store <4 x i32> %tmp6, <4 x i32> addrspace(1)* %tmp7, align 16
+  store <4 x i32> %tmp9, <4 x i32> addrspace(1)* %tmp10, align 16
+  ret void
+}
+
+; GCN-LABEL: {{^}}mubuf_clause:
+; GCN:      buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: buffer_load_dword
+; GCN-NEXT: s_nop
+; GCN-NEXT: buffer_load_dword
+define void @mubuf_clause(<4 x i32> addrspace(5)* noalias nocapture readonly %arg, <4 x i32> addrspace(5)* noalias nocapture %arg1) {
+bb:
+  %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %tmp2 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg, i32 %tmp
+  %tmp3 = load <4 x i32>, <4 x i32> addrspace(5)* %tmp2, align 16
+  %tmp4 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg1, i32 %tmp
+  %tmp5 = add nuw nsw i32 %tmp, 1
+  %tmp6 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg, i32 %tmp5
+  %tmp7 = load <4 x i32>, <4 x i32> addrspace(5)* %tmp6, align 16
+  %tmp8 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg1, i32 %tmp5
+  %tmp9 = add nuw nsw i32 %tmp, 2
+  %tmp10 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg, i32 %tmp9
+  %tmp11 = load <4 x i32>, <4 x i32> addrspace(5)* %tmp10, align 16
+  %tmp12 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg1, i32 %tmp9
+  %tmp13 = add nuw nsw i32 %tmp, 3
+  %tmp14 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg, i32 %tmp13
+  %tmp15 = load <4 x i32>, <4 x i32> addrspace(5)* %tmp14, align 16
+  %tmp16 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(5)* %arg1, i32 %tmp13
+  store <4 x i32> %tmp3, <4 x i32> addrspace(5)* %tmp4, align 16
+  store <4 x i32> %tmp7, <4 x i32> addrspace(5)* %tmp8, align 16
+  store <4 x i32> %tmp11, <4 x i32> addrspace(5)* %tmp12, align 16
+  store <4 x i32> %tmp15, <4 x i32> addrspace(5)* %tmp16, align 16
+  ret void
+}
+
+; GCN-LABEL: {{^}}vector_clause_indirect:
+; GCN: global_load_dwordx2 [[ADDR:v\[[0-9:]+\]]], v[{{[0-9:]+}}], off
+; GCN-NEXT: s_nop
+; GCN-NEXT: s_waitcnt vmcnt(0)
+; GCN-NEXT: global_load_dwordx4 v[{{[0-9:]+}}], [[ADDR]], off
+; GCN-NEXT: global_load_dwordx4 v[{{[0-9:]+}}], [[ADDR]], off offset:16
+define amdgpu_kernel void @vector_clause_indirect(i64 addrspace(1)* noalias nocapture readonly %arg, <4 x i32> addrspace(1)* noalias nocapture readnone %arg1, <4 x i32> addrspace(1)* noalias nocapture %arg2) {
+bb:
+  %tmp = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %tmp3 = zext i32 %tmp to i64
+  %tmp4 = getelementptr inbounds i64, i64 addrspace(1)* %arg, i64 %tmp3
+  %tmp5 = bitcast i64 addrspace(1)* %tmp4 to <4 x i32> addrspace(1)* addrspace(1)*
+  %tmp6 = load <4 x i32> addrspace(1)*, <4 x i32> addrspace(1)* addrspace(1)* %tmp5, align 8
+  %tmp7 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp6, align 16
+  %tmp8 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %tmp6, i64 1
+  %tmp9 = load <4 x i32>, <4 x i32> addrspace(1)* %tmp8, align 16
+  store <4 x i32> %tmp7, <4 x i32> addrspace(1)* %arg2, align 16
+  %tmp10 = getelementptr inbounds <4 x i32>, <4 x i32> addrspace(1)* %arg2, i64 1
+  store <4 x i32> %tmp9, <4 x i32> addrspace(1)* %tmp10, align 16
+  ret void
+}
+
+; GCN-LABEL: {{^}}load_global_d16_hi:
+; GCN:      global_load_short_d16_hi v
+; GCN-NEXT: s_nop
+; GCN-NEXT: global_load_short_d16_hi v
+define void @load_global_d16_hi(i16 addrspace(1)* %in, i16 %reg, <2 x i16> addrspace(1)* %out) {
+entry:
+  %gep = getelementptr inbounds i16, i16 addrspace(1)* %in, i64 32
+  %load1 = load i16, i16 addrspace(1)* %in
+  %load2 = load i16, i16 addrspace(1)* %gep
+  %build0 = insertelement <2 x i16> undef, i16 %reg, i32 0
+  %build1 = insertelement <2 x i16> %build0, i16 %load1, i32 1
+  store <2 x i16> %build1, <2 x i16> addrspace(1)* %out
+  %build2 = insertelement <2 x i16> undef, i16 %reg, i32 0
+  %build3 = insertelement <2 x i16> %build2, i16 %load2, i32 1
+  %gep2 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 32
+  store <2 x i16> %build3, <2 x i16> addrspace(1)* %gep2
+  ret void
+}
+
+; GCN-LABEL: {{^}}load_global_d16_lo:
+; GCN:      global_load_short_d16 v
+; GCN-NEXT: s_nop
+; GCN-NEXT: global_load_short_d16 v
+define void @load_global_d16_lo(i16 addrspace(1)* %in, i32 %reg, <2 x i16> addrspace(1)* %out) {
+entry:
+  %gep = getelementptr inbounds i16, i16 addrspace(1)* %in, i64 32
+  %reg.bc1 = bitcast i32 %reg to <2 x i16>
+  %reg.bc2 = bitcast i32 %reg to <2 x i16>
+  %load1 = load i16, i16 addrspace(1)* %in
+  %load2 = load i16, i16 addrspace(1)* %gep
+  %build1 = insertelement <2 x i16> %reg.bc1, i16 %load1, i32 0
+  %build2 = insertelement <2 x i16> %reg.bc2, i16 %load2, i32 0
+  %gep2 = getelementptr inbounds <2 x i16>, <2 x i16> addrspace(1)* %out, i64 32
+  store <2 x i16> %build1, <2 x i16> addrspace(1)* %out
+  store <2 x i16> %build2, <2 x i16> addrspace(1)* %gep2
+  ret void
+}
+
+declare i32 @llvm.amdgcn.workitem.id.x()

llvm/trunk/test/CodeGen/AMDGPU/memory_clause.mir

+# RUN: llc -march=amdgcn -mcpu=gfx902 -verify-machineinstrs -run-pass=phi-node-elimination,si-form-memory-clauses %s -o - | FileCheck -check-prefix=GCN %s
+
+# GCN-LABEL: {{^}}name: vector_clause{{$}}
+# GCN:         early-clobber %2:vreg_128, early-clobber %4:vreg_128, early-clobber %1:vreg_128, early-clobber %3:vreg_128 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:    %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+# GCN-NEXT:    %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 16, 0, 0, implicit $exec
+# GCN-NEXT:    %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 32, 0, 0, implicit $exec
+# GCN-NEXT:    %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 48, 0, 0, implicit $exec
+# GCN-NEXT:  }
+# GCN-NEXT:  GLOBAL_STORE_DWORDX4 %0, %1, 0, 0, 0, implicit $exec
+
+---
+name:            vector_clause
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_128 }
+  - { id: 2, class: vreg_128 }
+  - { id: 3, class: vreg_128 }
+  - { id: 4, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+    %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 16, 0, 0, implicit $exec
+    %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 32, 0, 0, implicit $exec
+    %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 48, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %1, 0, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %2, 16, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %3, 32, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %4, 48, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: subreg_full{{$}}
+# GCN:      early-clobber %1:vreg_128 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:   undef %1.sub0:vreg_128 = GLOBAL_LOAD_DWORD %0.sub0_sub1, 0, 0, 0, implicit $exec
+# GCN-NEXT:   internal %1.sub1:vreg_128 = GLOBAL_LOAD_DWORD %0.sub1_sub2, 16, 0, 0, implicit $exec
+# GCN-NEXT:   internal %1.sub2:vreg_128 = GLOBAL_LOAD_DWORD %0.sub2_sub3, 32, 0, 0, implicit $exec
+# GCN-NEXT:   internal %1.sub3:vreg_128 = GLOBAL_LOAD_DWORD %0.sub2_sub3, 32, 0, 0, implicit $exec
+# GCN-NEXT: }
+# GCN-NEXT: GLOBAL_STORE_DWORDX4 %0.sub0_sub1, %1, 0, 0, 0, implicit $exec
+
+---
+name:            subreg_full
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_128 }
+  - { id: 1, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    undef %1.sub0:vreg_128 = GLOBAL_LOAD_DWORD %0.sub0_sub1, 0, 0, 0, implicit $exec
+    %1.sub1:vreg_128 = GLOBAL_LOAD_DWORD %0.sub1_sub2, 16, 0, 0, implicit $exec
+    %1.sub2:vreg_128 = GLOBAL_LOAD_DWORD %0.sub2_sub3, 32, 0, 0, implicit $exec
+    %1.sub3:vreg_128 = GLOBAL_LOAD_DWORD %0.sub2_sub3, 32, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0.sub0_sub1, %1, 0, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: subreg_part{{$}}
+# GCN:      undef early-clobber %1.sub0_sub1:vreg_128, undef early-clobber %1.sub3:vreg_128 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:   undef %1.sub0:vreg_128 = GLOBAL_LOAD_DWORD %0.sub0_sub1, 0, 0, 0, implicit $exec
+# GCN-NEXT:   internal %1.sub1:vreg_128 = GLOBAL_LOAD_DWORD %0.sub1_sub2, 16, 0, 0, implicit $exec
+# GCN-NEXT:   internal %1.sub3:vreg_128 = GLOBAL_LOAD_DWORD %0.sub2_sub3, 32, 0, 0, implicit $exec
+# GCN-NEXT: }
+# GCN-NEXT: GLOBAL_STORE_DWORDX4 %0.sub0_sub1, %1, 0, 0, 0, implicit $exec
+
+---
+name:            subreg_part
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_128 }
+  - { id: 1, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    undef %1.sub0:vreg_128 = GLOBAL_LOAD_DWORD %0.sub0_sub1, 0, 0, 0, implicit $exec
+    %1.sub1:vreg_128 = GLOBAL_LOAD_DWORD %0.sub1_sub2, 16, 0, 0, implicit $exec
+    %1.sub3:vreg_128 = GLOBAL_LOAD_DWORD %0.sub2_sub3, 32, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0.sub0_sub1, %1, 0, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: dead{{$}}
+# GCN:      dead early-clobber %2:vreg_128, dead early-clobber %4:vreg_128, dead early-clobber %1:vreg_128, dead early-clobber %3:vreg_128 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:   dead %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 16, 0, 0, implicit $exec
+# GCN-NEXT:   dead %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 32, 0, 0, implicit $exec
+# GCN-NEXT:   dead %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 48, 0, 0, implicit $exec
+# GCN-NEXT: }
+
+---
+name:            dead
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_128 }
+  - { id: 2, class: vreg_128 }
+  - { id: 3, class: vreg_128 }
+  - { id: 4, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    dead %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+    dead %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 16, 0, 0, implicit $exec
+    dead %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 32, 0, 0, implicit $exec
+    dead %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 48, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: subreg_dead{{$}}
+# GCN:      early-clobber %1:vreg_64 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:    %1.sub0:vreg_64 = GLOBAL_LOAD_DWORD %0, 16, 0, 0, implicit $exec
+# GCN-NEXT:    dead %1.sub1:vreg_64 = GLOBAL_LOAD_DWORD %0, 32, 0, 0, implicit $exec
+# GCN-NEXT: }
+# GCN-NEXT: GLOBAL_STORE_DWORD %0, %1.sub0, 0, 0, 0, implicit $exec
+
+---
+name:            subreg_dead
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_64 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    undef %1.sub0:vreg_64 = GLOBAL_LOAD_DWORD %0, 16, 0, 0, implicit $exec
+    dead %1.sub1:vreg_64 = GLOBAL_LOAD_DWORD %0, 32, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORD %0, %1.sub0, 0, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: kill{{$}}
+# GCN:      early-clobber %2:vreg_128, early-clobber %3:vreg_128 = BUNDLE %0, %1, implicit $exec {
+# GCN-NEXT:   %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %1, 16, 0, 0, implicit $exec
+# GCN-NEXT: }
+
+---
+name:            kill
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_64 }
+  - { id: 2, class: vreg_128 }
+  - { id: 3, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1 = IMPLICIT_DEF
+    %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+    %3:vreg_128 = GLOBAL_LOAD_DWORDX4 killed %1, 16, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %3, 16, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: indirect{{$}}
+# GCN:      %1:vreg_64 = GLOBAL_LOAD_DWORDX2 %0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   early-clobber %2:vreg_128, early-clobber %3:vreg_128 = BUNDLE %1, implicit $exec {
+# GCN-NEXT:   %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %1, 0, 0, 0, implicit $exec
+# GCN-NEXT:   %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %1, 16, 0, 0, implicit $exec
+# GCN-NEXT: }
+
+---
+name:            indirect
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_64 }
+  - { id: 2, class: vreg_128 }
+  - { id: 3, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1:vreg_64 = GLOBAL_LOAD_DWORDX2 %0, 0, 0, 0, implicit $exec
+    %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %1, 0, 0, 0, implicit $exec
+    %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %1, 16, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %2, 0, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %3, 16, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: stack{{$}}
+# GCN:      %0:vreg_64 = IMPLICIT_DEF
+# GCN-NEXT: %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %stack.0, 0, 0, 0, implicit $exec
+# GCN-NEXT: %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %stack.0, 16, 0, 0, implicit $exec
+# GCN-NEXT: GLOBAL_STORE_DWORDX4 %0, %1, 0, 0, 0, implicit $exec
+
+---
+name:            stack
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_128 }
+  - { id: 2, class: vreg_128 }
+stack:
+  - { id: 0, type: default, offset: 0, size: 64, alignment: 8 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %stack.0, 0, 0, 0, implicit $exec
+    %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %stack.0, 16, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %1, 0, 0, 0, implicit $exec
+    GLOBAL_STORE_DWORDX4 %0, %2, 16, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: overflow_counter{{$}}
+# GCN:      dead early-clobber %7:vgpr_32, dead early-clobber %14:vgpr_32, dead early-clobber %2:vgpr_32, dead early-clobber %9:vgpr_32, dead early-clobber %4:vgpr_32, dead early-clobber %11:vgpr_32, dead early-clobber %6:vgpr_32, dead early-clobber %13:vgpr_32, dead early-clobber %1:vgpr_32, dead early-clobber %8:vgpr_32, dead early-clobber %15:vgpr_32, dead early-clobber %3:vgpr_32, dead early-clobber %10:vgpr_32, dead early-clobber %5:vgpr_32, dead early-clobber %12:vgpr_32 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:   dead %1:vgpr_32 = GLOBAL_LOAD_DWORD %0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   dead %2:vgpr_32 = GLOBAL_LOAD_DWORD %0, 4, 0, 0, implicit $exec
+# GCN-NEXT:   dead %3:vgpr_32 = GLOBAL_LOAD_DWORD %0, 8, 0, 0, implicit $exec
+# GCN-NEXT:   dead %4:vgpr_32 = GLOBAL_LOAD_DWORD %0, 12, 0, 0, implicit $exec
+# GCN-NEXT:   dead %5:vgpr_32 = GLOBAL_LOAD_DWORD %0, 16, 0, 0, implicit $exec
+# GCN-NEXT:   dead %6:vgpr_32 = GLOBAL_LOAD_DWORD %0, 20, 0, 0, implicit $exec
+# GCN-NEXT:   dead %7:vgpr_32 = GLOBAL_LOAD_DWORD %0, 24, 0, 0, implicit $exec
+# GCN-NEXT:   dead %8:vgpr_32 = GLOBAL_LOAD_DWORD %0, 28, 0, 0, implicit $exec
+# GCN-NEXT:   dead %9:vgpr_32 = GLOBAL_LOAD_DWORD %0, 32, 0, 0, implicit $exec
+# GCN-NEXT:   dead %10:vgpr_32 = GLOBAL_LOAD_DWORD %0, 36, 0, 0, implicit $exec
+# GCN-NEXT:   dead %11:vgpr_32 = GLOBAL_LOAD_DWORD %0, 40, 0, 0, implicit $exec
+# GCN-NEXT:   dead %12:vgpr_32 = GLOBAL_LOAD_DWORD %0, 44, 0, 0, implicit $exec
+# GCN-NEXT:   dead %13:vgpr_32 = GLOBAL_LOAD_DWORD %0, 48, 0, 0, implicit $exec
+# GCN-NEXT:   dead %14:vgpr_32 = GLOBAL_LOAD_DWORD %0, 52, 0, 0, implicit $exec
+# GCN-NEXT:   dead %15:vgpr_32 = GLOBAL_LOAD_DWORD %0, 56, 0, 0, implicit $exec
+# GCN-NEXT: }
+# GCN-NEXT: dead early-clobber %16:vgpr_32, dead early-clobber %17:vgpr_32 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:   dead %16:vgpr_32 = GLOBAL_LOAD_DWORD %0, 60, 0, 0, implicit $exec
+# GCN-NEXT:   dead %17:vgpr_32 = GLOBAL_LOAD_DWORD %0, 64, 0, 0, implicit $exec
+# GCN-NEXT: }
+
+---
+name:            overflow_counter
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: vgpr_32 }
+  - { id: 4, class: vgpr_32 }
+  - { id: 5, class: vgpr_32 }
+  - { id: 6, class: vgpr_32 }
+  - { id: 7, class: vgpr_32 }
+  - { id: 8, class: vgpr_32 }
+  - { id: 9, class: vgpr_32 }
+  - { id: 10, class: vgpr_32 }
+  - { id: 11, class: vgpr_32 }
+  - { id: 12, class: vgpr_32 }
+  - { id: 13, class: vgpr_32 }
+  - { id: 14, class: vgpr_32 }
+  - { id: 15, class: vgpr_32 }
+  - { id: 16, class: vgpr_32 }
+  - { id: 17, class: vgpr_32 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1:vgpr_32 = GLOBAL_LOAD_DWORD %0, 0, 0, 0, implicit $exec
+    %2:vgpr_32 = GLOBAL_LOAD_DWORD %0, 4, 0, 0, implicit $exec
+    %3:vgpr_32 = GLOBAL_LOAD_DWORD %0, 8, 0, 0, implicit $exec
+    %4:vgpr_32 = GLOBAL_LOAD_DWORD %0, 12, 0, 0, implicit $exec
+    %5:vgpr_32 = GLOBAL_LOAD_DWORD %0, 16, 0, 0, implicit $exec
+    %6:vgpr_32 = GLOBAL_LOAD_DWORD %0, 20, 0, 0, implicit $exec
+    %7:vgpr_32 = GLOBAL_LOAD_DWORD %0, 24, 0, 0, implicit $exec
+    %8:vgpr_32 = GLOBAL_LOAD_DWORD %0, 28, 0, 0, implicit $exec
+    %9:vgpr_32 = GLOBAL_LOAD_DWORD %0, 32, 0, 0, implicit $exec
+    %10:vgpr_32 = GLOBAL_LOAD_DWORD %0, 36, 0, 0, implicit $exec
+    %11:vgpr_32 = GLOBAL_LOAD_DWORD %0, 40, 0, 0, implicit $exec
+    %12:vgpr_32 = GLOBAL_LOAD_DWORD %0, 44, 0, 0, implicit $exec
+    %13:vgpr_32 = GLOBAL_LOAD_DWORD %0, 48, 0, 0, implicit $exec
+    %14:vgpr_32 = GLOBAL_LOAD_DWORD %0, 52, 0, 0, implicit $exec
+    %15:vgpr_32 = GLOBAL_LOAD_DWORD %0, 56, 0, 0, implicit $exec
+    %16:vgpr_32 = GLOBAL_LOAD_DWORD %0, 60, 0, 0, implicit $exec
+    %17:vgpr_32 = GLOBAL_LOAD_DWORD %0, 64, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: reg_pressure{{$}}
+# GCN:      dead early-clobber %2:vreg_128, dead early-clobber %4:vreg_128, dead early-clobber %1:vreg_128, dead early-clobber %3:vreg_128, dead early-clobber %5:vreg_128 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:   dead %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   dead %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 16, 0, 0, implicit $exec
+# GCN-NEXT:   dead %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 32, 0, 0, implicit $exec
+# GCN-NEXT:   dead %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 48, 0, 0, implicit $exec
+# GCN-NEXT:   dead %5:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 64, 0, 0, implicit $exec
+# GCN-NEXT: }
+# GCN-NEXT: dead early-clobber %7:vreg_128, dead early-clobber %6:vreg_128 = BUNDLE %0, implicit $exec {
+# GCN-NEXT:   dead %6:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 80, 0, 0, implicit $exec
+# GCN-NEXT:   dead %7:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 96, 0, 0, implicit $exec
+# GCN-NEXT: }
+
+---
+name:            reg_pressure
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vreg_128 }
+  - { id: 2, class: vreg_128 }
+  - { id: 3, class: vreg_128 }
+  - { id: 4, class: vreg_128 }
+  - { id: 5, class: vreg_128 }
+  - { id: 6, class: vreg_128 }
+  - { id: 7, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+    %2:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 16, 0, 0, implicit $exec
+    %3:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 32, 0, 0, implicit $exec
+    %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 48, 0, 0, implicit $exec
+    %5:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 64, 0, 0, implicit $exec
+    %6:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 80, 0, 0, implicit $exec
+    %7:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 96, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: image_clause{{$}}
+# GCN:      early-clobber %4:vreg_128, early-clobber %3:vreg_128, early-clobber %5:vreg_128 = BUNDLE %0, undef %2:sreg_128, %1, implicit $exec {
+# GCN-NEXT:   %3:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, undef %2:sreg_128, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   %4:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, undef %2:sreg_128, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   %5:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, undef %2:sreg_128, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+# GCN-NEXT: }
+# GCN-NEXT: IMAGE_STORE_V4_V2 %3, %0, %1, 15, -1, 0, 0, 0, 0, 0, 0, implicit $exec
+
+---
+name:            image_clause
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: sreg_256 }
+  - { id: 2, class: sreg_128 }
+  - { id: 3, class: vreg_128 }
+  - { id: 4, class: vreg_128 }
+  - { id: 5, class: vreg_128 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1 = IMPLICIT_DEF
+    %3:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, undef %2:sreg_128, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    %4:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, undef %2:sreg_128, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    %5:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, undef %2:sreg_128, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    IMAGE_STORE_V4_V2 %3, %0, %1, 15, -1, 0, 0, 0, 0, 0, 0, implicit $exec
+    IMAGE_STORE_V4_V2 %4, %0, %1, 15, -1, 0, 0, 0, 0, 0, 0, implicit $exec
+    IMAGE_STORE_V4_V2 %5, %0, %1, 15, -1, 0, 0, 0, 0, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: mixed_clause{{$}}
+# GCN:      dead early-clobber %4:vreg_128, dead early-clobber %3:vreg_128, dead early-clobber %5:vgpr_32 = BUNDLE %0, %2, %1, implicit $exec {
+# GCN-NEXT:   dead %3:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, %2, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   dead %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+# GCN-NEXT:   dead %5:vgpr_32 = BUFFER_LOAD_DWORD_ADDR64 %0, %2, 0, 0, 0, 0, 0, implicit $exec
+# GCN-NEXT: }
+
+---
+name:            mixed_clause
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: sreg_256 }
+  - { id: 2, class: sreg_128 }
+  - { id: 3, class: vreg_128 }
+  - { id: 4, class: vreg_128 }
+  - { id: 5, class: vgpr_32 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1 = IMPLICIT_DEF
+    %2 = IMPLICIT_DEF
+    %3:vreg_128 = IMAGE_SAMPLE_LZ_V4_V2 %0, %1, %2, 15, 0, 0, 0, 0, 0, 0, 0, implicit $exec
+    %4:vreg_128 = GLOBAL_LOAD_DWORDX4 %0, 0, 0, 0, implicit $exec
+    %5:vgpr_32 = BUFFER_LOAD_DWORD_ADDR64 %0, %2, 0, 0, 0, 0, 0, implicit $exec
+...
+
+# GCN-LABEL: {{^}}name: atomic{{$}}
+# GCN:      %1:vgpr_32 = IMPLICIT_DEF
+# GCN-NEXT: dead %2:vgpr_32 = FLAT_ATOMIC_ADD_RTN %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+# GCN-NEXT: dead %3:vgpr_32 = FLAT_ATOMIC_ADD_RTN %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+# GCN-NEXT: FLAT_ATOMIC_ADD %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+# GCN-NEXT: FLAT_ATOMIC_ADD %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+# GCN-NEXT: S_ENDPGM
+
+---
+name:            atomic
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: vreg_64 }
+  - { id: 1, class: vgpr_32 }
+  - { id: 2, class: vgpr_32 }
+  - { id: 3, class: vgpr_32 }
+body:             |
+  bb.0:
+    %0 = IMPLICIT_DEF
+    %1 = IMPLICIT_DEF
+    %2:vgpr_32 = FLAT_ATOMIC_ADD_RTN %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+    %3:vgpr_32 = FLAT_ATOMIC_ADD_RTN %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+    FLAT_ATOMIC_ADD %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+    FLAT_ATOMIC_ADD %0, %1, 0, 0, implicit $exec, implicit $flat_scr
+    S_ENDPGM
+...