[AMDGPU] SIFixSGPRCopies reworking to use one pass over the MIR for analysis and lowering.

[alex-t]

Summary

This change finalizes the series of patches aiming to replace the old strategy of VGPR to SGPR copy lowering.

1. Following the https://reviews.llvm.org/D128252 and https://reviews.llvm.org/D130367 code parts that are no longer used were removed.
2. The first pass over the MachineFunctoin collects all the necessary information.
3. Lowering is done in 3 phases:
   • VGPR to SGPR copies analysis lowering
   • REG_SEQUENCE, PHIs, and SGPR to VGPR copies lowering
   • SCC copies lowering is done in a separate pass over the Machine Function

[rampitec]

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

646

Typo "copiy", "scoreand,hence".

773

Just MI->getParent()

1066

Just !MI->getParent()

1088

Early return here and no else.

1105

New line.

[alex-t]

minor fixes as requested

[rampitec]

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

649

Should not it continue if tryChangeVGPRtoSGPRinCopy returned true? It will skip the next check then, which is always false.

669

Alignment is off.

[alex-t]

'continue' as requested and formatting corrected

[rampitec]

LGTM

[vitalybuka]

This breaks the bot https://lab.llvm.org/buildbot/#/builders/168/builds/8038

[eugenis]

Reverted. Please mind your changes after commit and do not leave LLVM broken for so long.

[alex-t]

In D131246#3713655, @eugenis wrote:

Reverted. Please mind your changes after commit and do not leave LLVM broken for so long.

My apologies for missing the build failure notification. And thanks a lot for cleaning up the mess,

[alex-t]

Fixed address sanitizer error: onDeleteListener interface added as a superclass for any client which needs the capability to handle MachineInstruction deletion in SIInstrInfo::moveToVALU.

[rampitec]

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

1105

New line.

1105

Why not simply clean all of these lists/vectors/maps at the end of the runOnMachineFunction?

[alex-t]

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

1105

The reason for that change is as follows:
We cannot lower SCC copies, REG_SEQUENCE, and PHIs before VGPR to SGPR copies since any lowering could change the MIR and, hence, make the analysis results invalid.
On other hand, in VGPR to SGPR copies lowering "SIInstrInfo::moveToVALU & Co" may be invoked. They remove MIs from the basic block and poison the memory allocated for them. So, all the pointers stored in the lists can potentially point to the poisoned memory and MI->getParent() check becomes invalid. This is the exact reason for the ASAN failure.
The "OnDeleteListener" interface serves as the delegate passed to "SIInstrInfo::moveToVALU" to make the caller informed of the MI->eraseFromParent() calls.

[vpykhtin]

This change isn't looking as refactoring, would be nice to reflect behavior change in the description.

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

572

looks like accidental * removal

[alex-t]

accidental change removed

[arsenm]

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

1105

I don't think this complicated list management is better than a separate loop

[alex-t]

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

1105

Since we have been recently taking efforts to shrink compile time I thought that it is better.

[alex-t]

MachineInstr deletion listener interface removed
SCC copies processing is done by the separate pas over the MIR

[rampitec]

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

647

Accidental brace move.

[alex-t]

formatting

llvm/lib/Target/AMDGPU/SIFixSGPRCopies.cpp

dbgs() << ID << " : " << *Copy << "\n\tS:" << SChain.size()
            << "\n\tSV:" << NumSVCopies << "\n\tSP: " << SiblingPenalty
            << "\nScore: " << Score << "\n";
   }
 #endif
 };
 
 class SIFixSGPRCopies : public MachineFunctionPass {
   MachineDominatorTree *MDT;
+  SmallVector<MachineInstr*, 4> SCCCopies;
+  SmallVector<MachineInstr*, 4> RegSequences;
+  SmallVector<MachineInstr*, 4> PHINodes;
+  SmallVector<MachineInstr*, 4> S2VCopies;
   unsigned NextVGPRToSGPRCopyID;
   DenseMap<unsigned, V2SCopyInfo> V2SCopies;
   DenseMap<MachineInstr *, SetVector<unsigned>> SiblingPenalty;
 
 public:
   static char ID;
 
   MachineRegisterInfo *MRI;
   const SIRegisterInfo *TRI;
   const SIInstrInfo *TII;
 
   SIFixSGPRCopies() : MachineFunctionPass(ID), NextVGPRToSGPRCopyID(0) {}
 
   bool runOnMachineFunction(MachineFunction &MF) override;
+  void fixSCCCopies(bool IsWave32);
+  void prepareRegSequenceAndPHIs(MachineFunction &MF);
   unsigned getNextVGPRToSGPRCopyId() { return ++NextVGPRToSGPRCopyID; }
-  void analyzeVGPRToSGPRCopy(V2SCopyInfo& Info);
+  bool needToBeConvertedToVALU(V2SCopyInfo *I);
+  void analyzeVGPRToSGPRCopy(MachineInstr *MI);
   void lowerVGPR2SGPRCopies(MachineFunction &MF);
   // Handles copies which source register is:
   // 1. Physical register
   // 2. AGPR
   // 3. Defined by the instruction the merely moves the immediate
   bool lowerSpecialCase(MachineInstr &MI);
 
   void processPHINode(MachineInstr &MI);

 char SIFixSGPRCopies::ID = 0;
 
 char &llvm::SIFixSGPRCopiesID = SIFixSGPRCopies::ID;
 
 FunctionPass *llvm::createSIFixSGPRCopiesPass() {
   return new SIFixSGPRCopies();
 }
 
-static bool hasVectorOperands(const MachineInstr &MI,
-                              const SIRegisterInfo *TRI) {
-  const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
-  for (const MachineOperand &MO : MI.operands()) {
-    if (!MO.isReg() || !MO.getReg().isVirtual())
-      continue;
-
-    if (TRI->hasVectorRegisters(MRI.getRegClass(MO.getReg())))
-      return true;
-  }
-  return false;
-}
-
 static std::pair<const TargetRegisterClass *, const TargetRegisterClass *>
 getCopyRegClasses(const MachineInstr &Copy,
                   const SIRegisterInfo &TRI,
                   const MachineRegisterInfo &MRI) {
   Register DstReg = Copy.getOperand(0).getReg();
   Register SrcReg = Copy.getOperand(1).getReg();
 
   const TargetRegisterClass *SrcRC = SrcReg.isVirtual()
while (I != Defs.end()) {
       } else
         ++I;
     }
   }
 
   // Try to schedule SGPR initializations as early as possible in the MBB.
   for (auto &Init : Inits) {
     auto &Defs = Init.second;
     for (auto MI : Defs) {
       auto MBB = MI->getParent();
       MachineInstr &BoundaryMI = *getFirstNonPrologue(MBB, TII);
       MachineBasicBlock::reverse_iterator B(BoundaryMI);
       // Check if B should actually be a boundary. If not set the previous
       // instruction as the boundary instead.
       if (!TII->isBasicBlockPrologue(*B))
         B++;
 
if (MF.getProperties().hasProperty(
     return false;
 
   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
   MRI = &MF.getRegInfo();
   TRI = ST.getRegisterInfo();
   TII = ST.getInstrInfo();
   MDT = &getAnalysis<MachineDominatorTree>();
 
-  // We have to lower VGPR to SGPR copies before the main loop
-  // because the REG_SEQUENCE and PHI lowering in main loop
-  // convert the def-use chains to VALU and close the opportunities
-  // for keeping them scalar.
-  // TODO: REG_SEQENCE and PHIs are semantically copies. The next patch
-  // addresses their lowering and unify the processing in one main loop.
-  lowerVGPR2SGPRCopies(MF);
-
   for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
                                                   BI != BE; ++BI) {
     MachineBasicBlock *MBB = &*BI;
     for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
          ++I) {
       MachineInstr &MI = *I;
 
       switch (MI.getOpcode()) {
       default:
         continue;
       case AMDGPU::COPY:
       case AMDGPU::WQM:
       case AMDGPU::STRICT_WQM:
       case AMDGPU::SOFT_WQM:
       case AMDGPU::STRICT_WWM: {
+        Register SrcReg = MI.getOperand(1).getReg();
         Register DstReg = MI.getOperand(0).getReg();
         const TargetRegisterClass *SrcRC, *DstRC;
         std::tie(SrcRC, DstRC) = getCopyRegClasses(MI, *TRI, *MRI);
 
-        if (MI.isCopy()) {
-          Register SrcReg = MI.getOperand(1).getReg();
-          if (SrcReg == AMDGPU::SCC) {
-            Register SCCCopy = MRI->createVirtualRegister(
-                TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID));
-            I = BuildMI(*MI.getParent(),
-                        std::next(MachineBasicBlock::iterator(MI)),
-                        MI.getDebugLoc(),
-                        TII->get(ST.isWave32() ? AMDGPU::S_CSELECT_B32
-                                               : AMDGPU::S_CSELECT_B64),
-                        SCCCopy)
-                    .addImm(-1)
-                    .addImm(0);
-            I = BuildMI(*MI.getParent(), std::next(I), I->getDebugLoc(),
-                        TII->get(AMDGPU::COPY), DstReg)
-                    .addReg(SCCCopy);
-            MI.eraseFromParent();
-            continue;
-          } else if (DstReg == AMDGPU::SCC) {
-            unsigned Opcode =
-                ST.isWave64() ? AMDGPU::S_AND_B64 : AMDGPU::S_AND_B32;
-            Register Exec = ST.isWave64() ? AMDGPU::EXEC : AMDGPU::EXEC_LO;
-            Register Tmp = MRI->createVirtualRegister(TRI->getBoolRC());
-            I = BuildMI(*MI.getParent(),
-                        std::next(MachineBasicBlock::iterator(MI)),
-                        MI.getDebugLoc(), TII->get(Opcode))
-                    .addReg(Tmp, getDefRegState(true))
-                    .addReg(SrcReg)
-                    .addReg(Exec);
-            MI.eraseFromParent();
-            continue;
-          }
-        }
-
-        if (!DstReg.isVirtual()) {
-          // If the destination register is a physical register there isn't
-          // really much we can do to fix this.
-          // Some special instructions use M0 as an input. Some even only use
-          // the first lane. Insert a readfirstlane and hope for the best.
-          if (DstReg == AMDGPU::M0 && TRI->hasVectorRegisters(SrcRC)) {
-            Register TmpReg
-              = MRI->createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+        if (MI.isCopy() && (SrcReg == AMDGPU::SCC || DstReg == AMDGPU::SCC))
+          SCCCopies.push_back(&MI);
 
-            BuildMI(*MBB, MI, MI.getDebugLoc(),
-                    TII->get(AMDGPU::V_READFIRSTLANE_B32), TmpReg)
-                .add(MI.getOperand(1));
-            MI.getOperand(1).setReg(TmpReg);
-          }
-
+        if (isSGPRToVGPRCopy(SrcRC, DstRC, *TRI)) {
+          // Since VGPR to SGPR copies affect VGPR to SGPR copy
+          // score and, hence the lowering decision, let's try to get rid of
+          // them as early as possible
+          if (tryChangeVGPRtoSGPRinCopy(MI, TRI, TII))
+            continue;
+          // Collect those not changed to try them after VGPR to SGPR copies
+          // lowering as there will be more opportunities.
+          S2VCopies.push_back(&MI);
+        if (!isVGPRToSGPRCopy(SrcRC, DstRC, *TRI))
+          continue;
+        if (lowerSpecialCase(MI))
           continue;
-        }
 
-        if (isSGPRToVGPRCopy(SrcRC, DstRC, *TRI)) {
-          tryChangeVGPRtoSGPRinCopy(MI, TRI, TII);
-        }
+        analyzeVGPRToSGPRCopy(&MI);
 
         break;
       }
-      case AMDGPU::PHI: {
-        processPHINode(MI);
-        break;
-      }
+      case AMDGPU::INSERT_SUBREG:
+      case AMDGPU::PHI:
       case AMDGPU::REG_SEQUENCE: {
-        if (TRI->hasVectorRegisters(TII->getOpRegClass(MI, 0)) ||
-            !hasVectorOperands(MI, TRI)) {
-          foldVGPRCopyIntoRegSequence(MI, TRI, TII, *MRI);
+        if (TRI->isSGPRClass(TII->getOpRegClass(MI, 0))) {
+          for (MachineOperand &MO : MI.operands()) {
+            if (!MO.isReg() || !MO.getReg().isVirtual())
               continue;
+            const TargetRegisterClass *SrcRC = MRI->getRegClass(MO.getReg());
+            if (TRI->hasVectorRegisters(SrcRC)) {
+              const TargetRegisterClass *DestRC =
+                  TRI->getEquivalentSGPRClass(SrcRC);
+              Register NewDst = MRI->createVirtualRegister(DestRC);
+              MachineBasicBlock *BlockToInsertCopy =
+                  MI.isPHI() ? MI.getOperand(MI.getOperandNo(&MO) + 1).getMBB()
+                             : MBB;
+              MachineBasicBlock::iterator PointToInsertCopy =
+                  MI.isPHI() ? BlockToInsertCopy->getFirstInstrTerminator() : I;
+              MachineInstr *NewCopy =
+                  BuildMI(*BlockToInsertCopy, PointToInsertCopy,
+                          PointToInsertCopy->getDebugLoc(),
+                          TII->get(AMDGPU::COPY), NewDst)
+                      .addReg(MO.getReg());
+              MO.setReg(NewDst);
+              analyzeVGPRToSGPRCopy(NewCopy);
             }
-
-        break;
-      }
-      case AMDGPU::INSERT_SUBREG: {
-        const TargetRegisterClass *DstRC, *Src0RC, *Src1RC;
-        DstRC = MRI->getRegClass(MI.getOperand(0).getReg());
-        Src0RC = MRI->getRegClass(MI.getOperand(1).getReg());
-        Src1RC = MRI->getRegClass(MI.getOperand(2).getReg());
-        if (TRI->isSGPRClass(DstRC) &&
-            (TRI->hasVectorRegisters(Src0RC) ||
-             TRI->hasVectorRegisters(Src1RC))) {
-          LLVM_DEBUG(dbgs() << " Fixing INSERT_SUBREG: " << MI);
-          MachineBasicBlock *NewBB = TII->moveToVALU(MI, MDT);
-          if (NewBB && NewBB != MBB) {
-            MBB = NewBB;
-            E = MBB->end();
-            BI = MachineFunction::iterator(MBB);
-            BE = MF.end();
           }
-          assert((!NewBB || NewBB == I->getParent()) &&
-                 "moveToVALU did not return the right basic block");
         }
+
+        if (MI.isPHI())
+          PHINodes.push_back(&MI);
+        else if (MI.isRegSequence())
+          RegSequences.push_back(&MI);
+
         break;
       }
       case AMDGPU::V_WRITELANE_B32: {
         // Some architectures allow more than one constant bus access without
         // SGPR restriction
         if (ST.getConstantBusLimit(MI.getOpcode()) != 1)
           break;
 
for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
           }
         }
         break;
       }
       }
     }
   }
 
+  lowerVGPR2SGPRCopies(MF);
+
+  // Postprocessing
+  fixSCCCopies(ST.isWave32());
+
+  for (auto MI : S2VCopies) {
+    // Check if it is still valid
+    if (MI->getParent() && MI->isCopy()) {
+      const TargetRegisterClass *SrcRC, *DstRC;
+      std::tie(SrcRC, DstRC) = getCopyRegClasses(*MI, *TRI, *MRI);
+      if (isSGPRToVGPRCopy(SrcRC, DstRC, *TRI))
+        tryChangeVGPRtoSGPRinCopy(*MI, TRI, TII);
+    }
+  }
+
+
+  for (auto MI : RegSequences) {
+    // Check if it is still valid
+    if (MI->getParent() && MI->isRegSequence())
+      foldVGPRCopyIntoRegSequence(*MI, TRI, TII, *MRI);
+  }
+
+  for (auto MI : PHINodes) {
+    processPHINode(*MI);
+  }
+
   if (MF.getTarget().getOptLevel() > CodeGenOpt::None && EnableM0Merge)
     hoistAndMergeSGPRInits(AMDGPU::M0, *MRI, TRI, *MDT, TII);
 
   SiblingPenalty.clear();
   V2SCopies.clear();
+  SCCCopies.clear();
+  RegSequences.clear();
+  PHINodes.clear();
+  S2VCopies.clear();
 
   return true;
 }
 
 void SIFixSGPRCopies::processPHINode(MachineInstr &MI) {
   bool AllAGPRUses = true;
   SetVector<const MachineInstr *> worklist;
   SmallSet<const MachineInstr *, 4> Visited;
void SIFixSGPRCopies::processPHINode(MachineInstr &MI) {
 
   // Propagate register class back to PHI operands which are PHI themselves.
   while (!PHIOperands.empty()) {
     processPHINode(*PHIOperands.pop_back_val());
   }
 }
 
 bool SIFixSGPRCopies::lowerSpecialCase(MachineInstr &MI) {
+
+  Register DstReg = MI.getOperand(0).getReg();
   Register SrcReg = MI.getOperand(1).getReg();
+
+  if (!DstReg.isVirtual()) {
+    // If the destination register is a physical register there isn't
+    // really much we can do to fix this.
+    // Some special instructions use M0 as an input. Some even only use
+    // the first lane. Insert a readfirstlane and hope for the best.
+    if (DstReg == AMDGPU::M0 &&
+        TRI->hasVectorRegisters(MRI->getRegClass(SrcReg))) {
+      Register TmpReg =
+          MRI->createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+
+      BuildMI(*MI.getParent(), MI, MI.getDebugLoc(),
+              TII->get(AMDGPU::V_READFIRSTLANE_B32), TmpReg)
+          .add(MI.getOperand(1));
+      MI.getOperand(1).setReg(TmpReg);
+    }
+
+    return true;
+  }
+
   if (!SrcReg.isVirtual() || TRI->isAGPR(*MRI, SrcReg)) {
     TII->moveToVALU(MI, MDT);
     return true;
   }
 
   unsigned SMovOp;
   int64_t Imm;
   // If we are just copying an immediate, we can replace the copy with
   // s_mov_b32.
   if (isSafeToFoldImmIntoCopy(&MI, MRI->getVRegDef(SrcReg), TII, SMovOp, Imm)) {
     MI.getOperand(1).ChangeToImmediate(Imm);
     MI.addImplicitDefUseOperands(*MI.getParent()->getParent());
     MI.setDesc(TII->get(SMovOp));
     return true;
   }
   return false;
 }
 
+void SIFixSGPRCopies::analyzeVGPRToSGPRCopy(MachineInstr* MI) {
+  Register DstReg = MI->getOperand(0).getReg();
+  const TargetRegisterClass *DstRC = MRI->getRegClass(DstReg);
 
+  V2SCopyInfo Info(getNextVGPRToSGPRCopyId(), MI,
+                      TRI->getRegSizeInBits(*DstRC));
 
-void SIFixSGPRCopies::analyzeVGPRToSGPRCopy(V2SCopyInfo& Info) {
   SmallVector<MachineInstr *, 8> AnalysisWorklist;
   // Needed because the SSA is not a tree but a graph and may have
   // forks and joins. We should not then go same way twice.
   DenseSet<MachineInstr *> Visited;
   AnalysisWorklist.push_back(Info.Copy);
   while (!AnalysisWorklist.empty()) {
 
     MachineInstr *Inst = AnalysisWorklist.pop_back_val();
if ((TII->isSALU(*Inst) && Inst->isCompare()) ||
           Users.push_back(&U);
     }
     for (auto U : Users) {
       if (TII->isSALU(*U))
         Info.SChain.insert(U);
       AnalysisWorklist.push_back(U);
     }
   }
+  V2SCopies[Info.ID] = Info;
 }
 
-void SIFixSGPRCopies::lowerVGPR2SGPRCopies(MachineFunction &MF) {
-
-  // The main function that computes the VGPR to SGPR copy score
-  // and determines copy further lowering way: v_readfirstlane_b32 or moveToVALU
-  auto needToBeConvertedToVALU = [&](V2SCopyInfo *I) -> bool {
-    if (I->SChain.empty()) {
-      I->Score = 0;
-      return true;
-    }
-    I->Siblings = SiblingPenalty[*std::max_element(
-        I->SChain.begin(), I->SChain.end(),
-        [&](MachineInstr *A, MachineInstr *B) -> bool {
-          return SiblingPenalty[A].size() < SiblingPenalty[B].size();
-        })];
-    I->Siblings.remove_if([&](unsigned ID) { return ID == I->ID; });
-    // The loop below computes the number of another VGPR to SGPR V2SCopies
-    // which contribute to the current copy SALU chain. We assume that all the
-    // V2SCopies with the same source virtual register will be squashed to one
-    // by regalloc. Also we take care of the V2SCopies of the differnt subregs
-    // of the same register.
-    SmallSet<std::pair<Register, unsigned>, 4> SrcRegs;
-    for (auto J : I->Siblings) {
-      auto InfoIt = V2SCopies.find(J);
-      if (InfoIt != V2SCopies.end()) {
-        MachineInstr *SiblingCopy = InfoIt->getSecond().Copy;
-        if (SiblingCopy->isImplicitDef())
-          // the COPY has already been MoveToVALUed
-          continue;
+bool SIFixSGPRCopies::needToBeConvertedToVALU(V2SCopyInfo *Info) {
+  if (Info->SChain.empty()) {
+    Info->Score = 0;
+  Info->Siblings = SiblingPenalty[*std::max_element(
+      Info->SChain.begin(), Info->SChain.end(),
+  Info->Siblings.remove_if([&](unsigned ID) { return ID == Info->ID; });
+  for (auto J : Info->Siblings) {
 
-        SrcRegs.insert(std::make_pair(SiblingCopy->getOperand(1).getReg(),
-                                      SiblingCopy->getOperand(1).getSubReg()));
-      }
-    }
-    I->SiblingPenalty = SrcRegs.size();
+  Info->SiblingPenalty = SrcRegs.size();
 
-    unsigned Penalty =
-        I->NumSVCopies + I->SiblingPenalty + I->NumReadfirstlanes;
-    unsigned Profit = I->SChain.size();
-    I->Score = Penalty > Profit ? 0 : Profit - Penalty;
-    I->NeedToBeConvertedToVALU = I->Score < 3;
-    return I->NeedToBeConvertedToVALU;
-  };
-
-  auto needProcessing = [](MachineInstr &MI) -> bool {
-    switch (MI.getOpcode()) {
-    case AMDGPU::COPY:
-    case AMDGPU::WQM:
-    case AMDGPU::STRICT_WQM:
-    case AMDGPU::SOFT_WQM:
-    case AMDGPU::STRICT_WWM:
-    case AMDGPU::REG_SEQUENCE:
-    case AMDGPU::PHI:
-      return true;
-    default:
-      return false;
-    }
-  };
-
-  SmallSet<MachineInstr *, 4> OutOfOrderProcessedCopies;
-
-  for (MachineFunction::iterator BI = MF.begin(), BE = MF.end(); BI != BE;
-       ++BI) {
-    MachineBasicBlock *MBB = &*BI;
-    for (MachineBasicBlock::iterator I = MBB->begin(), E = MBB->end(); I != E;
-         ++I) {
-      MachineInstr *MI = &*I;
-      if (!needProcessing(*MI))
-        continue;
-
-      if (MI->isRegSequence() || MI->isPHI()) {
-        MachineBasicBlock::iterator J = I;
-        if (TRI->isSGPRClass(TII->getOpRegClass(*MI, 0))) {
-          for (MachineOperand &MO : MI->operands()) {
-            if (!MO.isReg() || !MO.getReg().isVirtual())
-              continue;
-            const TargetRegisterClass *SrcRC = MRI->getRegClass(MO.getReg());
-            if (TRI->hasVectorRegisters(SrcRC)) {
-              const TargetRegisterClass *DestRC =
-                  TRI->getEquivalentSGPRClass(SrcRC);
-              Register NewDst = MRI->createVirtualRegister(DestRC);
-              MachineBasicBlock *BlockToInsertCopy = MBB;
-              MachineBasicBlock::iterator PointToInsertCopy = I;
-              if (MI->isPHI()) {
-                BlockToInsertCopy =
-                    MI->getOperand(MI->getOperandNo(&MO) + 1).getMBB();
-                PointToInsertCopy =
-                    BlockToInsertCopy->getFirstInstrTerminator();
-              }
-              MachineBasicBlock::iterator NewI =
-                  BuildMI(*BlockToInsertCopy, PointToInsertCopy,
-                          PointToInsertCopy->getDebugLoc(),
-                          TII->get(AMDGPU::COPY), NewDst)
-                      .addReg(MO.getReg());
-              MO.setReg(NewDst);
-              if (!MI->isPHI()) {
-                I = NewI;
-                MI = &*I;
-              } else {
-                // We insert the copy into the basic block that may have been
-                // already processed. Pass it to the analysis explicitly.
-                V2SCopyInfo In(getNextVGPRToSGPRCopyId(), MI,
-                               TRI->getRegSizeInBits(*DestRC));
-                analyzeVGPRToSGPRCopy(In);
-                V2SCopies[In.ID] = In;
-                OutOfOrderProcessedCopies.insert(MI);
-              }
-            }
-          }
-        }
+      Info->NumSVCopies + Info->SiblingPenalty + Info->NumReadfirstlanes;
+  unsigned Profit = Info->SChain.size();
+  Info->Score = Penalty > Profit ? 0 : Profit - Penalty;
+  Info->NeedToBeConvertedToVALU = Info->Score < 3;
+  return Info->NeedToBeConvertedToVALU;
 
-        if (J == I)
-          continue;
-      }
-
-      const TargetRegisterClass *SrcRC, *DstRC;
-      std::tie(SrcRC, DstRC) = getCopyRegClasses(*MI, *TRI, *MRI);
-
-      if (!isVGPRToSGPRCopy(SrcRC, DstRC, *TRI))
-        continue;
-
-      if (lowerSpecialCase(*MI))
-        continue;
-
-      if (OutOfOrderProcessedCopies.contains(MI))
-        continue;
-
-      V2SCopyInfo In(getNextVGPRToSGPRCopyId(), MI,
-                     TRI->getRegSizeInBits(*DstRC));
-
-      analyzeVGPRToSGPRCopy(In);
-
-      V2SCopies[In.ID] = In;
-    }
-  }
+void SIFixSGPRCopies::lowerVGPR2SGPRCopies(MachineFunction &MF) {
 
   SmallVector<unsigned, 8> LoweringWorklist;
   for (auto &C : V2SCopies) {
     if (needToBeConvertedToVALU(&C.second))
       LoweringWorklist.push_back(C.second.ID);
   }
 
   while (!LoweringWorklist.empty()) {
if (SrcSize == 16) {
                 TII->get(AMDGPU::V_READFIRSTLANE_B32), PartialDst)
             .addReg(PartialSrc);
         Result.addReg(PartialDst).addImm(TRI->getSubRegFromChannel(i));
       }
     }
     MI->eraseFromParent();
   }
 }
+
+void SIFixSGPRCopies::fixSCCCopies(bool IsWave32) {
+  for (auto MI : SCCCopies) {
+    // May be lowered out
+    if (!MI->getParent())
+      continue;
+    // May already have been lowered.
+    if (!MI->isCopy())
+      continue;
+    Register SrcReg = MI->getOperand(1).getReg();
+    Register DstReg = MI->getOperand(0).getReg();
+    if (SrcReg == AMDGPU::SCC) {
+      Register SCCCopy = MRI->createVirtualRegister(
+          TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID));
+      MachineBasicBlock::iterator I =
+          BuildMI(*MI->getParent(),
+                  std::next(MachineBasicBlock::iterator(MI)),
+                  MI->getDebugLoc(),
+                  TII->get(IsWave32 ? AMDGPU::S_CSELECT_B32
+                                    : AMDGPU::S_CSELECT_B64),
+                  SCCCopy)
+              .addImm(-1)
+              .addImm(0);
+      BuildMI(*MI->getParent(), std::next(I), I->getDebugLoc(),
+              TII->get(AMDGPU::COPY), DstReg)
+          .addReg(SCCCopy);
+      MI->eraseFromParent();
+      continue;
+    }
+
+    if (DstReg == AMDGPU::SCC) {
+      unsigned Opcode = IsWave32 ? AMDGPU::S_AND_B32 : AMDGPU::S_AND_B64;
+      Register Exec = IsWave32 ? AMDGPU::EXEC_LO : AMDGPU::EXEC;
+      Register Tmp = MRI->createVirtualRegister(TRI->getBoolRC());
+      BuildMI(*MI->getParent(), std::next(MachineBasicBlock::iterator(MI)),
+              MI->getDebugLoc(), TII->get(Opcode))
+          .addReg(Tmp, getDefRegState(true))
+          .addReg(SrcReg)
+          .addReg(Exec);
+      MI->eraseFromParent();
+    }
+  }
+}

llvm/test/CodeGen/AMDGPU/mubuf-legalize-operands.mir

bb.0:
     ; W64-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W64-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W64-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W64-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; W64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
     ; W64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; W64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, [[COPY4]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; W64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, [[COPY3]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; W64-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
     ; W64-NEXT: {{  $}}
     ; W64-NEXT: .1:
     ; W64-NEXT: successors: %bb.2(0x80000000)
     ; W64-NEXT: {{  $}}
     ; W64-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W64-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W64-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
bb.0:
     ; W32-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W32-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W32-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W32-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
     ; W32-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, [[COPY4]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, [[COPY3]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; W32-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: .1:
     ; W32-NEXT: successors: %bb.2(0x80000000)
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W32-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W32-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
bb.0:
     ; W64-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
     ; W64-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W64-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W64-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W64-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
-    ; W64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; W64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
-    ; W64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
-    ; W64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, killed [[COPY8]], %subreg.sub2, [[COPY2]], %subreg.sub3
+    ; W64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
+    ; W64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
+    ; W64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, killed [[COPY6]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; W64-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
     ; W64-NEXT: {{  $}}
     ; W64-NEXT: .1:
     ; W64-NEXT: successors: %bb.2(0x80000000)
     ; W64-NEXT: {{  $}}
     ; W64-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W64-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W64-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
bb.0:
     ; W32-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
     ; W32-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W32-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W32-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W32-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
-    ; W32-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
-    ; W32-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
-    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, killed [[COPY8]], %subreg.sub2, [[COPY2]], %subreg.sub3
+    ; W32-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
+    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W32-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
+    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, killed [[COPY6]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; W32-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: .1:
     ; W32-NEXT: successors: %bb.2(0x80000000)
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W32-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W32-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
bb.0:
     ; W64-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
     ; W64-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W64-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W64-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W64-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
-    ; W64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
-    ; W64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
-    ; W64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; W64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, killed [[COPY6]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; W64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
+    ; W64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
+    ; W64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, killed [[COPY8]], %subreg.sub2, [[COPY2]], %subreg.sub3
     ; W64-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
     ; W64-NEXT: {{  $}}
     ; W64-NEXT: .1:
     ; W64-NEXT: successors: %bb.2(0x80000000)
     ; W64-NEXT: {{  $}}
     ; W64-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W64-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W64-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
bb.0:
     ; W32-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
     ; W32-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W32-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W32-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W32-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
-    ; W32-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
-    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
-    ; W32-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY5]], %subreg.sub0, killed [[COPY6]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; W32-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
+    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
+    ; W32-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, killed [[COPY8]], %subreg.sub2, [[COPY2]], %subreg.sub3
     ; W32-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: .1:
     ; W32-NEXT: successors: %bb.2(0x80000000)
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W32-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W32-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
bb.0:
     ; ADDR64-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; ADDR64-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; ADDR64-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; ADDR64-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; ADDR64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; ADDR64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
     ; ADDR64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; ADDR64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, [[COPY4]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; ADDR64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, [[COPY3]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; ADDR64-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
     ; ADDR64-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
     ; ADDR64-NEXT: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 0
     ; ADDR64-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sgpr_32 = S_MOV_B32 61440
     ; ADDR64-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[S_MOV_B64_]], %subreg.sub0_sub1, [[S_MOV_B32_]], %subreg.sub2, [[S_MOV_B32_1]], %subreg.sub3
     ; ADDR64-NEXT: [[V_ADD_CO_U32_e64_:%[0-9]+]]:vgpr_32, [[V_ADD_CO_U32_e64_1:%[0-9]+]]:sreg_64_xexec = V_ADD_CO_U32_e64 [[COPY9]].sub0, [[COPY1]].sub0, 0, implicit $exec
     ; ADDR64-NEXT: %17:vgpr_32, dead %20:sreg_64_xexec = V_ADDC_U32_e64 [[COPY9]].sub1, [[COPY1]].sub1, killed [[V_ADD_CO_U32_e64_1]], 0, implicit $exec
     ; ADDR64-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_ADD_CO_U32_e64_]], %subreg.sub0, %17, %subreg.sub1
bb.0:
     ; W32-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W32-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W32-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W32-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
     ; W32-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, [[COPY4]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, [[COPY3]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; W32-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
     ; W32-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
     ; W32-NEXT: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 0
     ; W32-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sgpr_32 = S_MOV_B32 822173696
     ; W32-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[S_MOV_B64_]], %subreg.sub0_sub1, [[S_MOV_B32_]], %subreg.sub2, [[S_MOV_B32_1]], %subreg.sub3
     ; W32-NEXT: [[V_ADD_CO_U32_e64_:%[0-9]+]]:vgpr_32, [[V_ADD_CO_U32_e64_1:%[0-9]+]]:sreg_32_xm0_xexec = V_ADD_CO_U32_e64 [[COPY9]].sub0, [[COPY1]].sub0, 0, implicit $exec
     ; W32-NEXT: %17:vgpr_32, dead %20:sreg_32_xm0_xexec = V_ADDC_U32_e64 [[COPY9]].sub1, [[COPY1]].sub1, killed [[V_ADD_CO_U32_e64_1]], 0, implicit $exec
     ; W32-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[V_ADD_CO_U32_e64_]], %subreg.sub0, %17, %subreg.sub1
bb.0:
     ; ADDR64-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; ADDR64-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; ADDR64-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; ADDR64-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; ADDR64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; ADDR64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; ADDR64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
     ; ADDR64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; ADDR64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, [[COPY4]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; ADDR64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, [[COPY3]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; ADDR64-NEXT: [[COPY9:%[0-9]+]]:vreg_64 = COPY [[REG_SEQUENCE]].sub0_sub1
     ; ADDR64-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64 = S_MOV_B64 0
     ; ADDR64-NEXT: [[S_MOV_B32_:%[0-9]+]]:sgpr_32 = S_MOV_B32 0
     ; ADDR64-NEXT: [[S_MOV_B32_1:%[0-9]+]]:sgpr_32 = S_MOV_B32 61440
     ; ADDR64-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[S_MOV_B64_]], %subreg.sub0_sub1, [[S_MOV_B32_]], %subreg.sub2, [[S_MOV_B32_1]], %subreg.sub3
     ; ADDR64-NEXT: [[REG_SEQUENCE2:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY9]].sub0, %subreg.sub0, [[COPY9]].sub1, %subreg.sub1
     ; ADDR64-NEXT: [[BUFFER_LOAD_FORMAT_X_ADDR64_:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_FORMAT_X_ADDR64 [[REG_SEQUENCE2]], [[REG_SEQUENCE1]], 0, 0, 0, 0, 0, implicit $exec
     ; ADDR64-NEXT: $sgpr30_sgpr31 = COPY [[COPY]]
bb.0:
     ; W64-NO-ADDR64-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W64-NO-ADDR64-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W64-NO-ADDR64-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W64-NO-ADDR64-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NO-ADDR64-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NO-ADDR64-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NO-ADDR64-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W64-NO-ADDR64-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; W64-NO-ADDR64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W64-NO-ADDR64-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
     ; W64-NO-ADDR64-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; W64-NO-ADDR64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, [[COPY4]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; W64-NO-ADDR64-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, [[COPY3]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; W64-NO-ADDR64-NEXT: [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
     ; W64-NO-ADDR64-NEXT: {{  $}}
     ; W64-NO-ADDR64-NEXT: .1:
     ; W64-NO-ADDR64-NEXT: successors: %bb.2(0x80000000)
     ; W64-NO-ADDR64-NEXT: {{  $}}
     ; W64-NO-ADDR64-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W64-NO-ADDR64-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W64-NO-ADDR64-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
bb.0:
     ; W32-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
     ; W32-NEXT: [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
     ; W32-NEXT: [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; W32-NEXT: [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
     ; W32-NEXT: [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[COPY5]], implicit $exec
-    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY3]], implicit $exec
+    ; W32-NEXT: [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[COPY4]], implicit $exec
     ; W32-NEXT: [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[COPY2]], implicit $exec
-    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, [[COPY4]], %subreg.sub1, killed [[COPY7]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
+    ; W32-NEXT: [[REG_SEQUENCE:%[0-9]+]]:vreg_128 = REG_SEQUENCE killed [[COPY6]], %subreg.sub0, killed [[COPY7]], %subreg.sub1, [[COPY3]], %subreg.sub2, killed [[COPY8]], %subreg.sub3
     ; W32-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32_xm0_xexec = S_MOV_B32 $exec_lo
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: .1:
     ; W32-NEXT: successors: %bb.2(0x80000000)
     ; W32-NEXT: {{  $}}
     ; W32-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub0, implicit $exec
     ; W32-NEXT: [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE]].sub1, implicit $exec
     ; W32-NEXT: [[REG_SEQUENCE1:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1

llvm/test/CodeGen/AMDGPU/wwm-reserved-spill.ll

 
 define amdgpu_gfx i64 @strict_wwm_called_i64(i64 %a) noinline {
 ; GFX9-O0-LABEL: strict_wwm_called_i64:
 ; GFX9-O0:       ; %bb.0:
 ; GFX9-O0-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v0
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr34
-; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 killed $exec
+; GFX9-O0-NEXT:    ; kill: def $vgpr1 killed $vgpr1 killed $exec
 ; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 def $vgpr2_vgpr3 killed $exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr34_sgpr35
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v2
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v2
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v3
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v3
 ; GFX9-O0-NEXT:    v_add_co_u32_e64 v4, s[34:35], v4, v5

 ; GFX9-O0-NEXT:    v_readlane_b32 s36, v10, 2
 ; GFX9-O0-NEXT:    v_readlane_b32 s37, v10, 3
 ; GFX9-O0-NEXT:    v_readlane_b32 s38, v10, 4
 ; GFX9-O0-NEXT:    v_readlane_b32 s39, v10, 5
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v0
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v3, v1
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr40
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr40
-; GFX9-O0-NEXT:    ; kill: def $vgpr3 killed $vgpr3 killed $exec
+; GFX9-O0-NEXT:    ; kill: def $vgpr2 killed $vgpr2 killed $exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v4, v8
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v9
 ; GFX9-O0-NEXT:    v_add_co_u32_e64 v2, s[40:41], v2, v4
 ; GFX9-O0-NEXT:    v_addc_co_u32_e64 v3, s[40:41], v3, v5, s[40:41]
 ; GFX9-O0-NEXT:    s_mov_b64 exec, s[34:35]
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v0, v2
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v3
 ; GFX9-O0-NEXT:    s_mov_b32 s34, 0

 ; GFX9-O0-NEXT:    s_mov_b32 s34, 0
 ; GFX9-O0-NEXT:    buffer_load_dwordx4 v[10:13], v0, s[36:39], s34 offen
 ; GFX9-O0-NEXT:    buffer_load_dwordx2 v[3:4], v0, s[36:39], s34 offen offset:16
 ; GFX9-O0-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v11
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v10
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
-; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 killed $exec
+; GFX9-O0-NEXT:    ; kill: def $vgpr7 killed $vgpr7 killed $exec
 ; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v7
 ; GFX9-O0-NEXT:    s_mov_b32 s35, 0x7fffffff
 ; GFX9-O0-NEXT:    s_mov_b32 s40, -1
 ; GFX9-O0-NEXT:    ; kill: def $sgpr40 killed $sgpr40 def $sgpr40_sgpr41
 ; GFX9-O0-NEXT:    s_mov_b32 s41, s35
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v5
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v6
 ; GFX9-O0-NEXT:    s_not_b64 exec, exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s40
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s41
 ; GFX9-O0-NEXT:    s_not_b64 exec, exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v1
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v2
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v13
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v12
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
-; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 killed $exec
+; GFX9-O0-NEXT:    ; kill: def $vgpr7 killed $vgpr7 killed $exec
 ; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6 killed $exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v7
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v1, v5
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v2, v6
 ; GFX9-O0-NEXT:    s_not_b64 exec, exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v1, s40
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v2, s41
 ; GFX9-O0-NEXT:    s_not_b64 exec, exec

 ; GFX9-O0-NEXT:    v_mov_b32_e32 v5, v8
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v9, v7
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v10, v6
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
 ; GFX9-O0-NEXT:    ; implicit-def: $sgpr35
 ; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 killed $exec
-; GFX9-O0-NEXT:    ; kill: def $vgpr10 killed $vgpr10 killed $exec
+; GFX9-O0-NEXT:    ; kill: def $vgpr11 killed $vgpr11 killed $exec
 ; GFX9-O0-NEXT:    ; kill: def $vgpr9 killed $vgpr9 killed $exec
 ; GFX9-O0-NEXT:    ; kill: def $vgpr5 killed $vgpr5 def $vgpr5_vgpr6_vgpr7_vgpr8 killed $exec
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v6, v11
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v7, v10
 ; GFX9-O0-NEXT:    v_mov_b32_e32 v8, v9
 ; GFX9-O0-NEXT:    buffer_store_dwordx4 v[5:8], v0, s[36:39], s34 offen
 ; GFX9-O0-NEXT:    buffer_store_dwordx2 v[3:4], v0, s[36:39], s34 offen offset:16
 ; GFX9-O0-NEXT:    s_or_saveexec_b64 s[34:35], -1