AMDGPU/SILoadStoreOptimizer: Improve merging of out of order offsets

[tstellar]

Summary

This improves merging of sequences like:

store a, ptr + 4
store b, ptr + 8
store c, ptr + 12
store d, ptr + 16
store e, ptr + 20
store f, ptr

Prior to this patch the basic block was scanned in order to find instructions
to merge and the above sequence would be transformed to:

store4 <a, b, c, d>, ptr + 4
store e, ptr + 20
store r, ptr

With this change, we now sort all the candidate merge instructions by their offset,
so instructions are visited in offset order rather than in the order they appear
in the basic block. We now transform this sequnce into:

store4 <f, a, b, c>, ptr
store2 <d, e>, ptr + 16

Another benefit of this change is that since we have sorted the mergeable lists
by offset, we can easily check if an instruction is mergeable by checking the
offset of the instruction that becomes before or after it in the sorted list.
Once we determine an instruction is not mergeable we can remove it from the list
and avoid having to do the more expensive mergeablilty checks.

[arsenm]

llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp

228

Why std::list?

278–279

Weird format

1936–1940

Don't you know which is first from which was encountered first?

[nhaehnle]

I think the code would benefit from the refactoring I've mentioned on the other patch, where the lists only hold a structure with information on a single instruction. Maybe call it CandidateInfo (information of one instruction, persistent in lists) vs. CombineInfo (information on a pair, only temporary).

llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp

567–572

Can that really happen? Only instructions with the same InstClass should be added to the same list.

1936

Wasn't there some talk about ordered basic blocks? They exist for IR apparently, but not for MIR unless we're tracking live ranges, which we don't do here, so...

This pass could perhaps number the CombineInfo instructions in order as they're collected at the start? It'd have to be kept uptodate as instructions are merged.

[tstellar]

Rebase on master and update for a recent refactoring.

[merge_guards_bot]

Build result: FAILURE - Could not check out parent git hash "8d7b8123bba35d4092d66750a83f9d0980bd169c". It was not found in the repository. Did you configure the "Parent Revision" in Phabricator properly? Trying to apply the patch to the master branch instead...

ERROR: arc patch failed with error code 1. Check build log for details.
Log files: console-log.txt, CMakeCache.txt

[tstellar]

In D65966#1701129, @nhaehnle wrote:

I think the code would benefit from the refactoring I've mentioned on the other patch, where the lists only hold a structure with information on a single instruction. Maybe call it CandidateInfo (information of one instruction, persistent in lists) vs. CombineInfo (information on a pair, only temporary).

See D71045 .

llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp

1936

I added an Order field to keep track of the ordering.

[nhaehnle]

Thanks. This basically looks good to me, some minor nitpicks.

llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp

786

Should be "Handle DS instructions."

849–850

This method should probably be renamed now since its meaning has shifted by quite a lot. Not sure what name is best, maybe checkAndPrepareMerge? The "prepare" part refers to collecting the InstsToMove.

Come to think of it, maybe the InstsToMove could be moved out of the CombineInfo struct? Then the CombineInfo arguments here could almost be passed as const, except for the fixup by the final offsetsCanBeCombined.

2010–2011

Maybe move this check into optimizeInstsWithSameBaseAddr? That function is already responsible for handling the case where the size is 1...

[tstellar]

+ Remove InstsToMove from CombineInfo struct
+ Address other review comments.

[arsenm]

LGTM

[merge_guards_bot]

Unit tests: pass. 61849 tests passed, 0 failed and 781 were skipped.

clang-tidy: unknown.

clang-format: fail. Please format your changes with clang-format by running git-clang-format HEAD^ or applying this patch.

Build artifacts: diff.json, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

llvm/lib/Target/AMDGPU/SILoadStoreOptimizer.cpp

struct CombineInfo {
     unsigned Format;
     unsigned BaseOff;
     unsigned DMask;
     InstClassEnum InstClass;
     bool GLC;
     bool SLC;
     bool DLC;
     bool UseST64;
-    SmallVector<MachineInstr *, 8> InstsToMove;
     int AddrIdx[5];
     const MachineOperand *AddrReg[5];
     unsigned NumAddresses;
+    unsigned Order;
 
     bool hasSameBaseAddress(const MachineInstr &MI) {
       for (unsigned i = 0; i < NumAddresses; i++) {
         const MachineOperand &AddrRegNext = MI.getOperand(AddrIdx[i]);
 
         if (AddrReg[i]->isImm() || AddrRegNext.isImm()) {
           if (AddrReg[i]->isImm() != AddrRegNext.isImm() ||
               AddrReg[i]->getImm() != AddrRegNext.getImm()) {
private:
   MachineRegisterInfo *MRI = nullptr;
   AliasAnalysis *AA = nullptr;
   bool OptimizeAgain;
 
   static bool dmasksCanBeCombined(const CombineInfo &CI,
                                   const SIInstrInfo &TII,
                                   const CombineInfo &Paired);
   static bool offsetsCanBeCombined(CombineInfo &CI, const MCSubtargetInfo &STI,
-                                   CombineInfo &Paired);
+                                   CombineInfo &Paired, bool Modify = false);
   static bool widthsFit(const GCNSubtarget &STM, const CombineInfo &CI,
                         const CombineInfo &Paired);
   static unsigned getNewOpcode(const CombineInfo &CI, const CombineInfo &Paired);
   static std::pair<unsigned, unsigned> getSubRegIdxs(const CombineInfo &CI,
                                                      const CombineInfo &Paired);
   const TargetRegisterClass *getTargetRegisterClass(const CombineInfo &CI,
                                                     const CombineInfo &Paired);
 
-  bool findMatchingInst(CombineInfo &CI, CombineInfo &Paired);
+  bool checkAndPrepareMerge(CombineInfo &CI, CombineInfo &Paired,
+                            SmallVectorImpl<MachineInstr *> &InstsToMove);
 
   unsigned read2Opcode(unsigned EltSize) const;
   unsigned read2ST64Opcode(unsigned EltSize) const;
-  MachineBasicBlock::iterator mergeRead2Pair(CombineInfo &CI, CombineInfo &Paired);
+  MachineBasicBlock::iterator mergeRead2Pair(CombineInfo &CI,
+                                             CombineInfo &Paired,
+                  const SmallVectorImpl<MachineInstr *> &InstsToMove);
 
   unsigned write2Opcode(unsigned EltSize) const;
   unsigned write2ST64Opcode(unsigned EltSize) const;
-  MachineBasicBlock::iterator mergeWrite2Pair(CombineInfo &CI, CombineInfo &Paired);
-  MachineBasicBlock::iterator mergeImagePair(CombineInfo &CI, CombineInfo &Paired);
-  MachineBasicBlock::iterator mergeSBufferLoadImmPair(CombineInfo &CI, CombineInfo &Paired);
-  MachineBasicBlock::iterator mergeBufferLoadPair(CombineInfo &CI, CombineInfo &Paired);
-  MachineBasicBlock::iterator mergeBufferStorePair(CombineInfo &CI, CombineInfo &Paired);
-  MachineBasicBlock::iterator mergeTBufferLoadPair(CombineInfo &CI, CombineInfo &Paired);
-  MachineBasicBlock::iterator mergeTBufferStorePair(CombineInfo &CI, CombineInfo &Paired);
+  MachineBasicBlock::iterator
+  mergeWrite2Pair(CombineInfo &CI, CombineInfo &Paired,
+                  const SmallVectorImpl<MachineInstr *> &InstsToMove);
+  MachineBasicBlock::iterator
+  mergeImagePair(CombineInfo &CI, CombineInfo &Paired,
+                 const SmallVectorImpl<MachineInstr *> &InstsToMove);
+  MachineBasicBlock::iterator
+  mergeSBufferLoadImmPair(CombineInfo &CI, CombineInfo &Paired,
+                          const SmallVectorImpl<MachineInstr *> &InstsToMove);
+  MachineBasicBlock::iterator
+  mergeBufferLoadPair(CombineInfo &CI, CombineInfo &Paired,
+                      const SmallVectorImpl<MachineInstr *> &InstsToMove);
+  MachineBasicBlock::iterator
+  mergeBufferStorePair(CombineInfo &CI, CombineInfo &Paired,
+                       const SmallVectorImpl<MachineInstr *> &InstsToMove);
+  MachineBasicBlock::iterator
+  mergeTBufferLoadPair(CombineInfo &CI, CombineInfo &Paired,
+                       const SmallVectorImpl<MachineInstr *> &InstsToMove);
+  MachineBasicBlock::iterator
+  mergeTBufferStorePair(CombineInfo &CI, CombineInfo &Paired,
+                        const SmallVectorImpl<MachineInstr *> &InstsToMove);
 
   void updateBaseAndOffset(MachineInstr &I, unsigned NewBase,
                            int32_t NewOffset) const;
   unsigned computeBase(MachineInstr &MI, const MemAddress &Addr) const;
   MachineOperand createRegOrImm(int32_t Val, MachineInstr &MI) const;
   Optional<int32_t> extractConstOffset(const MachineOperand &Op) const;
   void processBaseWithConstOffset(const MachineOperand &Base, MemAddress &Addr) const;
   /// Promotes constant offset to the immediate by adjusting the base. It
   /// tries to use a base from the nearby instructions that allows it to have
   /// a 13bit constant offset which gets promoted to the immediate.
   bool promoteConstantOffsetToImm(MachineInstr &CI,
                                   MemInfoMap &Visited,
                                   SmallPtrSet<MachineInstr *, 4> &Promoted) const;
   void addInstToMergeableList(const CombineInfo &CI,
                   std::list<std::list<CombineInfo> > &MergeableInsts) const;
   bool collectMergeableInsts(MachineBasicBlock &MBB,
                   std::list<std::list<CombineInfo> > &MergeableInsts) const;
 
 public:
   static char ID;
 
   SILoadStoreOptimizer() : MachineFunctionPass(ID) {
     initializeSILoadStoreOptimizerPass(*PassRegistry::getPassRegistry());
   }
 
-  void removeCombinedInst(std::list<CombineInfo> &MergeList,
-                                         const MachineInstr &MI);
   bool optimizeInstsWithSameBaseAddr(std::list<CombineInfo> &MergeList,
                                      bool &OptimizeListAgain);
   bool optimizeBlock(std::list<std::list<CombineInfo> > &MergeableInsts);
 
   bool runOnMachineFunction(MachineFunction &MF) override;
 
   StringRef getPassName() const override { return "SI Load Store Optimizer"; }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
void SILoadStoreOptimizer::CombineInfo::setMI(MachineBasicBlock::iterator MI,
   if (Regs & SSAMP) {
     AddrOpName[NumAddresses++] = AMDGPU::OpName::ssamp;
   }
 
   for (unsigned i = 0; i < NumAddresses; i++) {
     AddrIdx[i] = AMDGPU::getNamedOperandIdx(I->getOpcode(), AddrOpName[i]);
     AddrReg[i] = &I->getOperand(AddrIdx[i]);
   }
-
-  InstsToMove.clear();
 }
 
 } // end anonymous namespace.
 
 INITIALIZE_PASS_BEGIN(SILoadStoreOptimizer, DEBUG_TYPE,
                       "SI Load Store Optimizer", false, false)
 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
 INITIALIZE_PASS_END(SILoadStoreOptimizer, DEBUG_TYPE, "SI Load Store Optimizer",
                     false, false)
 
 char SILoadStoreOptimizer::ID = 0;
 
 char &llvm::SILoadStoreOptimizerID = SILoadStoreOptimizer::ID;
 
 FunctionPass *llvm::createSILoadStoreOptimizerPass() {
   return new SILoadStoreOptimizer();
static unsigned getBufferFormatWithCompCount(unsigned OldFormat,
   assert(NewFormatInfo->NumFormat == OldFormatInfo->NumFormat &&
          NewFormatInfo->BitsPerComp == OldFormatInfo->BitsPerComp);
 
   return NewFormatInfo->Format;
 }
 
 bool SILoadStoreOptimizer::offsetsCanBeCombined(CombineInfo &CI,
                                                 const MCSubtargetInfo &STI,
-                                                CombineInfo &Paired) {
+                                                CombineInfo &Paired,
+                                                bool Modify) {
   assert(CI.InstClass != MIMG);
 
   // XXX - Would the same offset be OK? Is there any reason this would happen or
   // be useful?
   if (CI.Offset == Paired.Offset)
     return false;
 
   // This won't be valid if the offset isn't aligned.
if (getBufferFormatWithCompCount(CI.Format, CI.Width + Paired.Width, STI) == 0)
       return false;
   }
 
   unsigned EltOffset0 = CI.Offset / CI.EltSize;
   unsigned EltOffset1 = Paired.Offset / CI.EltSize;
   CI.UseST64 = false;
   CI.BaseOff = 0;
 
-  // Handle SMEM and VMEM instructions.
+  // Handle DS instructions.
   if ((CI.InstClass != DS_READ) && (CI.InstClass != DS_WRITE)) {
     return (EltOffset0 + CI.Width == EltOffset1 ||
             EltOffset1 + Paired.Width == EltOffset0) &&
            CI.GLC == Paired.GLC && CI.DLC == Paired.DLC &&
            (CI.InstClass == S_BUFFER_LOAD_IMM || CI.SLC == Paired.SLC);
   }
 
+  // Handle SMEM and VMEM instructions.
   // If the offset in elements doesn't fit in 8-bits, we might be able to use
   // the stride 64 versions.
   if ((EltOffset0 % 64 == 0) && (EltOffset1 % 64) == 0 &&
       isUInt<8>(EltOffset0 / 64) && isUInt<8>(EltOffset1 / 64)) {
+    if (Modify) {
       CI.Offset = EltOffset0 / 64;
       Paired.Offset = EltOffset1 / 64;
       CI.UseST64 = true;
+    }
     return true;
   }
 
   // Check if the new offsets fit in the reduced 8-bit range.
   if (isUInt<8>(EltOffset0) && isUInt<8>(EltOffset1)) {
+    if (Modify) {
       CI.Offset = EltOffset0;
       Paired.Offset = EltOffset1;
+    }
     return true;
   }
 
   // Try to shift base address to decrease offsets.
   unsigned OffsetDiff = std::abs((int)EltOffset1 - (int)EltOffset0);
   CI.BaseOff = std::min(CI.Offset, Paired.Offset);
 
   if ((OffsetDiff % 64 == 0) && isUInt<8>(OffsetDiff / 64)) {
+    if (Modify) {
       CI.Offset = (EltOffset0 - CI.BaseOff / CI.EltSize) / 64;
       Paired.Offset = (EltOffset1 - CI.BaseOff / CI.EltSize) / 64;
       CI.UseST64 = true;
+    }
     return true;
   }
 
   if (isUInt<8>(OffsetDiff)) {
+    if (Modify) {
       CI.Offset = EltOffset0 - CI.BaseOff / CI.EltSize;
       Paired.Offset = EltOffset1 - CI.BaseOff / CI.EltSize;
+    }
     return true;
   }
 
   return false;
 }
 
 bool SILoadStoreOptimizer::widthsFit(const GCNSubtarget &STM,
                                      const CombineInfo &CI,
                                      const CombineInfo &Paired) {
   const unsigned Width = (CI.Width + Paired.Width);
   switch (CI.InstClass) {
   default:
     return (Width <= 4) && (STM.hasDwordx3LoadStores() || (Width != 3));
   case S_BUFFER_LOAD_IMM:
     switch (Width) {
     default:
       return false;
     case 2:
     case 4:
       return true;
     }
   }
 }
 
-bool SILoadStoreOptimizer::findMatchingInst(CombineInfo &CI,
-                                            CombineInfo &Paired) {
-  MachineBasicBlock *MBB = CI.I->getParent();
-  MachineBasicBlock::iterator E = MBB->end();
-  MachineBasicBlock::iterator MBBI = CI.I;
+/// This function assumes that CI comes before Paired in a basic block.
+bool SILoadStoreOptimizer::checkAndPrepareMerge(
+    CombineInfo &CI, CombineInfo &Paired,
+    SmallVectorImpl<MachineInstr *> &InstsToMove) {
+
+  // Check both offsets (or masks for MIMG) can be combined and fit in the
+  // reduced range.
+  if (CI.InstClass == MIMG && !dmasksCanBeCombined(CI, *TII, Paired))
+    return false;
+
+  if (CI.InstClass != MIMG &&
+      (!widthsFit(*STM, CI, Paired) || !offsetsCanBeCombined(CI, *STI, Paired)))
+    return false;
 
   const unsigned Opc = CI.I->getOpcode();
   const InstClassEnum InstClass = getInstClass(Opc, *TII);
 
   if (InstClass == UNKNOWN) {
     return false;
   }
   const unsigned InstSubclass = getInstSubclass(Opc, *TII);
 
   // Do not merge VMEM buffer instructions with "swizzled" bit set.
   int Swizzled =
       AMDGPU::getNamedOperandIdx(CI.I->getOpcode(), AMDGPU::OpName::swz);
   if (Swizzled != -1 && CI.I->getOperand(Swizzled).getImm())
     return false;
 
-  ++MBBI;
-
   DenseSet<unsigned> RegDefsToMove;
   DenseSet<unsigned> PhysRegUsesToMove;
   addDefsUsesToList(*CI.I, RegDefsToMove, PhysRegUsesToMove);
 
+  MachineBasicBlock::iterator E = std::next(Paired.I);
+  MachineBasicBlock::iterator MBBI = std::next(CI.I);
   for (; MBBI != E; ++MBBI) {
 
     if ((getInstClass(MBBI->getOpcode(), *TII) != InstClass) ||
         (getInstSubclass(MBBI->getOpcode(), *TII) != InstSubclass)) {
       // This is not a matching instruction, but we can keep looking as
       // long as one of these conditions are met:
       // 1. It is safe to move I down past MBBI.
       // 2. It is safe to move MBBI down past the instruction that I will
       //    be merged into.
 
       if (MBBI->hasUnmodeledSideEffects()) {
         // We can't re-order this instruction with respect to other memory
         // operations, so we fail both conditions mentioned above.
         return false;
       }
 
       if (MBBI->mayLoadOrStore() &&
           (!memAccessesCanBeReordered(*CI.I, *MBBI, AA) ||
-           !canMoveInstsAcrossMemOp(*MBBI, CI.InstsToMove, AA))) {
+           !canMoveInstsAcrossMemOp(*MBBI, InstsToMove, AA))) {
         // We fail condition #1, but we may still be able to satisfy condition
         // #2.  Add this instruction to the move list and then we will check
         // if condition #2 holds once we have selected the matching instruction.
-        CI.InstsToMove.push_back(&*MBBI);
+        InstsToMove.push_back(&*MBBI);
         addDefsUsesToList(*MBBI, RegDefsToMove, PhysRegUsesToMove);
         continue;
       }
 
       // When we match I with another DS instruction we will be moving I down
       // to the location of the matched instruction any uses of I will need to
       // be moved down as well.
       addToListsIfDependent(*MBBI, RegDefsToMove, PhysRegUsesToMove,
-                            CI.InstsToMove);
+                            InstsToMove);
       continue;
     }
 
     // Don't merge volatiles.
     if (MBBI->hasOrderedMemoryRef())
       return false;
 
     int Swizzled =
         AMDGPU::getNamedOperandIdx(MBBI->getOpcode(), AMDGPU::OpName::swz);
     if (Swizzled != -1 && MBBI->getOperand(Swizzled).getImm())
       return false;
 
     // Handle a case like
     //   DS_WRITE_B32 addr, v, idx0
     //   w = DS_READ_B32 addr, idx0
     //   DS_WRITE_B32 addr, f(w), idx1
     // where the DS_READ_B32 ends up in InstsToMove and therefore prevents
     // merging of the two writes.
     if (addToListsIfDependent(*MBBI, RegDefsToMove, PhysRegUsesToMove,
-                              CI.InstsToMove))
+                              InstsToMove))
       continue;
 
-    bool Match = CI.hasSameBaseAddress(*MBBI);
-
-    if (Match) {
-      Paired.setMI(MBBI, *TII, *STM);
-
-      // Check both offsets (or masks for MIMG) can be combined and fit in the
-      // reduced range.
-      bool canBeCombined =
-          CI.InstClass == MIMG
-              ? dmasksCanBeCombined(CI, *TII, Paired)
-              : widthsFit(*STM, CI, Paired) && offsetsCanBeCombined(CI, *STI, Paired);
-
-      // We also need to go through the list of instructions that we plan to
+    if (&*MBBI == &*Paired.I) {
+      // We need to go through the list of instructions that we plan to
       // move and make sure they are all safe to move down past the merged
       // instruction.
-      if (canBeCombined && canMoveInstsAcrossMemOp(*MBBI, CI.InstsToMove, AA))
+      if (canMoveInstsAcrossMemOp(*MBBI, InstsToMove, AA)) {
+
+        // Call offsetsCanBeCombined with modify = true so that the offsets are
+        // correct for the new instruction.  This should return true, because
+        // this function should only be called on CombineInfo objects that
+        // have already been confirmed to be mergeable.
+        if (CI.InstClass != MIMG)
+          offsetsCanBeCombined(CI, *STI, Paired, true);
         return true;
       }
+      return false;
+    }
 
     // We've found a load/store that we couldn't merge for some reason.
     // We could potentially keep looking, but we'd need to make sure that
     // it was safe to move I and also all the instruction in InstsToMove
     // down past this instruction.
     // check if we can move I across MBBI and if we can move all I's users
     if (!memAccessesCanBeReordered(*CI.I, *MBBI, AA) ||
-        !canMoveInstsAcrossMemOp(*MBBI, CI.InstsToMove, AA))
+        !canMoveInstsAcrossMemOp(*MBBI, InstsToMove, AA))
       break;
   }
   return false;
 }
 
 unsigned SILoadStoreOptimizer::read2Opcode(unsigned EltSize) const {
   if (STM->ldsRequiresM0Init())
     return (EltSize == 4) ? AMDGPU::DS_READ2_B32 : AMDGPU::DS_READ2_B64;
   return (EltSize == 4) ? AMDGPU::DS_READ2_B32_gfx9 : AMDGPU::DS_READ2_B64_gfx9;
 }
 
 unsigned SILoadStoreOptimizer::read2ST64Opcode(unsigned EltSize) const {
   if (STM->ldsRequiresM0Init())
     return (EltSize == 4) ? AMDGPU::DS_READ2ST64_B32 : AMDGPU::DS_READ2ST64_B64;
 
   return (EltSize == 4) ? AMDGPU::DS_READ2ST64_B32_gfx9
                         : AMDGPU::DS_READ2ST64_B64_gfx9;
 }
 
 MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeRead2Pair(CombineInfo &CI, CombineInfo &Paired) {
+SILoadStoreOptimizer::mergeRead2Pair(CombineInfo &CI, CombineInfo &Paired,
+    const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
 
   // Be careful, since the addresses could be subregisters themselves in weird
   // cases, like vectors of pointers.
   const auto *AddrReg = TII->getNamedOperand(*CI.I, AMDGPU::OpName::addr);
 
   const auto *Dest0 = TII->getNamedOperand(*CI.I, AMDGPU::OpName::vdst);
   const auto *Dest1 = TII->getNamedOperand(*Paired.I, AMDGPU::OpName::vdst);
SILoadStoreOptimizer::mergeRead2Pair(CombineInfo &CI, CombineInfo &Paired,
   // Copy to the old destination registers.
   BuildMI(*MBB, Paired.I, DL, CopyDesc)
       .add(*Dest0) // Copy to same destination including flags and sub reg.
       .addReg(DestReg, 0, SubRegIdx0);
   MachineInstr *Copy1 = BuildMI(*MBB, Paired.I, DL, CopyDesc)
                             .add(*Dest1)
                             .addReg(DestReg, RegState::Kill, SubRegIdx1);
 
-  moveInstsAfter(Copy1, CI.InstsToMove);
+  moveInstsAfter(Copy1, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
 
   LLVM_DEBUG(dbgs() << "Inserted read2: " << *Read2 << '\n');
   return Read2;
 }
 
 unsigned SILoadStoreOptimizer::write2Opcode(unsigned EltSize) const {
   if (STM->ldsRequiresM0Init())
     return (EltSize == 4) ? AMDGPU::DS_WRITE2_B32 : AMDGPU::DS_WRITE2_B64;
   return (EltSize == 4) ? AMDGPU::DS_WRITE2_B32_gfx9
                         : AMDGPU::DS_WRITE2_B64_gfx9;
 }
 
 unsigned SILoadStoreOptimizer::write2ST64Opcode(unsigned EltSize) const {
   if (STM->ldsRequiresM0Init())
     return (EltSize == 4) ? AMDGPU::DS_WRITE2ST64_B32
                           : AMDGPU::DS_WRITE2ST64_B64;
 
   return (EltSize == 4) ? AMDGPU::DS_WRITE2ST64_B32_gfx9
                         : AMDGPU::DS_WRITE2ST64_B64_gfx9;
 }
 
 MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeWrite2Pair(CombineInfo &CI, CombineInfo &Paired) {
+SILoadStoreOptimizer::mergeWrite2Pair(CombineInfo &CI, CombineInfo &Paired,
+                                      const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
 
   // Be sure to use .addOperand(), and not .addReg() with these. We want to be
   // sure we preserve the subregister index and any register flags set on them.
   const MachineOperand *AddrReg =
       TII->getNamedOperand(*CI.I, AMDGPU::OpName::addr);
   const MachineOperand *Data0 =
       TII->getNamedOperand(*CI.I, AMDGPU::OpName::data0);
MachineInstrBuilder Write2 =
           .addReg(BaseReg, BaseRegFlags, BaseSubReg) // addr
           .add(*Data0)                               // data0
           .add(*Data1)                               // data1
           .addImm(NewOffset0)                        // offset0
           .addImm(NewOffset1)                        // offset1
           .addImm(0)                                 // gds
           .cloneMergedMemRefs({&*CI.I, &*Paired.I});
 
-  moveInstsAfter(Write2, CI.InstsToMove);
+  moveInstsAfter(Write2, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
 
   LLVM_DEBUG(dbgs() << "Inserted write2 inst: " << *Write2 << '\n');
   return Write2;
 }
 
 MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeImagePair(CombineInfo &CI, CombineInfo &Paired) {
+SILoadStoreOptimizer::mergeImagePair(CombineInfo &CI, CombineInfo &Paired,
+                           const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
   DebugLoc DL = CI.I->getDebugLoc();
   const unsigned Opcode = getNewOpcode(CI, Paired);
 
   const TargetRegisterClass *SuperRC = getTargetRegisterClass(CI, Paired);
 
   Register DestReg = MRI->createVirtualRegister(SuperRC);
   unsigned MergedDMask = CI.DMask | Paired.DMask;
SILoadStoreOptimizer::mergeImagePair(CombineInfo &CI, CombineInfo &Paired,
 
   BuildMI(*MBB, Paired.I, DL, CopyDesc)
       .add(*Dest0) // Copy to same destination including flags and sub reg.
       .addReg(DestReg, 0, SubRegIdx0);
   MachineInstr *Copy1 = BuildMI(*MBB, Paired.I, DL, CopyDesc)
                             .add(*Dest1)
                             .addReg(DestReg, RegState::Kill, SubRegIdx1);
 
-  moveInstsAfter(Copy1, CI.InstsToMove);
+  moveInstsAfter(Copy1, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
   return New;
 }
 
-MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeSBufferLoadImmPair(CombineInfo &CI, CombineInfo &Paired) {
+MachineBasicBlock::iterator SILoadStoreOptimizer::mergeSBufferLoadImmPair(
+    CombineInfo &CI, CombineInfo &Paired,
+    const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
   DebugLoc DL = CI.I->getDebugLoc();
   const unsigned Opcode = getNewOpcode(CI, Paired);
 
   const TargetRegisterClass *SuperRC = getTargetRegisterClass(CI, Paired);
 
   Register DestReg = MRI->createVirtualRegister(SuperRC);
   unsigned MergedOffset = std::min(CI.Offset, Paired.Offset);
MachineBasicBlock::iterator SILoadStoreOptimizer::mergeSBufferLoadImmPair(
 
   BuildMI(*MBB, Paired.I, DL, CopyDesc)
       .add(*Dest0) // Copy to same destination including flags and sub reg.
       .addReg(DestReg, 0, SubRegIdx0);
   MachineInstr *Copy1 = BuildMI(*MBB, Paired.I, DL, CopyDesc)
                             .add(*Dest1)
                             .addReg(DestReg, RegState::Kill, SubRegIdx1);
 
-  moveInstsAfter(Copy1, CI.InstsToMove);
+  moveInstsAfter(Copy1, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
   return New;
 }
 
-MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeBufferLoadPair(CombineInfo &CI, CombineInfo &Paired) {
+MachineBasicBlock::iterator SILoadStoreOptimizer::mergeBufferLoadPair(
+    CombineInfo &CI, CombineInfo &Paired,
+    const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
   DebugLoc DL = CI.I->getDebugLoc();
 
   const unsigned Opcode = getNewOpcode(CI, Paired);
 
   const TargetRegisterClass *SuperRC = getTargetRegisterClass(CI, Paired);
 
   // Copy to the new source register.
MachineBasicBlock::iterator SILoadStoreOptimizer::mergeBufferLoadPair(
 
   BuildMI(*MBB, Paired.I, DL, CopyDesc)
       .add(*Dest0) // Copy to same destination including flags and sub reg.
       .addReg(DestReg, 0, SubRegIdx0);
   MachineInstr *Copy1 = BuildMI(*MBB, Paired.I, DL, CopyDesc)
                             .add(*Dest1)
                             .addReg(DestReg, RegState::Kill, SubRegIdx1);
 
-  moveInstsAfter(Copy1, CI.InstsToMove);
+  moveInstsAfter(Copy1, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
   return New;
 }
 
-MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeTBufferLoadPair(CombineInfo &CI, CombineInfo &Paired) {
+MachineBasicBlock::iterator SILoadStoreOptimizer::mergeTBufferLoadPair(
+    CombineInfo &CI, CombineInfo &Paired,
+    const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
   DebugLoc DL = CI.I->getDebugLoc();
 
   const unsigned Opcode = getNewOpcode(CI, Paired);
 
   const TargetRegisterClass *SuperRC = getTargetRegisterClass(CI, Paired);
 
   // Copy to the new source register.
MachineBasicBlock::iterator SILoadStoreOptimizer::mergeTBufferLoadPair(
 
   BuildMI(*MBB, Paired.I, DL, CopyDesc)
       .add(*Dest0) // Copy to same destination including flags and sub reg.
       .addReg(DestReg, 0, SubRegIdx0);
   MachineInstr *Copy1 = BuildMI(*MBB, Paired.I, DL, CopyDesc)
                             .add(*Dest1)
                             .addReg(DestReg, RegState::Kill, SubRegIdx1);
 
-  moveInstsAfter(Copy1, CI.InstsToMove);
+  moveInstsAfter(Copy1, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
   return New;
 }
 
-MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeTBufferStorePair(CombineInfo &CI, CombineInfo &Paired) {
+MachineBasicBlock::iterator SILoadStoreOptimizer::mergeTBufferStorePair(
+    CombineInfo &CI, CombineInfo &Paired,
+    const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
   DebugLoc DL = CI.I->getDebugLoc();
 
   const unsigned Opcode = getNewOpcode(CI, Paired);
 
   std::pair<unsigned, unsigned> SubRegIdx = getSubRegIdxs(CI, Paired);
   const unsigned SubRegIdx0 = std::get<0>(SubRegIdx);
   const unsigned SubRegIdx1 = std::get<1>(SubRegIdx);
MachineInstr *New =
           .addImm(CI.GLC)                          // glc
           .addImm(CI.SLC)                          // slc
           .addImm(0)                                // tfe
           .addImm(CI.DLC)                          // dlc
           .addImm(0)                                // swz
           .addMemOperand(
               combineKnownAdjacentMMOs(*MBB->getParent(), MMOa, MMOb));
 
-  moveInstsAfter(MIB, CI.InstsToMove);
+  moveInstsAfter(MIB, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
   return New;
 }
 
 unsigned SILoadStoreOptimizer::getNewOpcode(const CombineInfo &CI,
                                             const CombineInfo &Paired) {
if (CI.InstClass == S_BUFFER_LOAD_IMM) {
     case 3:
       return &AMDGPU::VReg_96RegClass;
     case 4:
       return &AMDGPU::VReg_128RegClass;
     }
   }
 }
 
-MachineBasicBlock::iterator
-SILoadStoreOptimizer::mergeBufferStorePair(CombineInfo &CI, CombineInfo &Paired) {
+MachineBasicBlock::iterator SILoadStoreOptimizer::mergeBufferStorePair(
+    CombineInfo &CI, CombineInfo &Paired,
+    const SmallVectorImpl<MachineInstr *> &InstsToMove) {
   MachineBasicBlock *MBB = CI.I->getParent();
   DebugLoc DL = CI.I->getDebugLoc();
 
   const unsigned Opcode = getNewOpcode(CI, Paired);
 
   std::pair<unsigned, unsigned> SubRegIdx = getSubRegIdxs(CI, Paired);
   const unsigned SubRegIdx0 = std::get<0>(SubRegIdx);
   const unsigned SubRegIdx1 = std::get<1>(SubRegIdx);
MIB.add(*TII->getNamedOperand(*CI.I, AMDGPU::OpName::srsrc))
         .addImm(std::min(CI.Offset, Paired.Offset)) // offset
         .addImm(CI.GLC)      // glc
         .addImm(CI.SLC)      // slc
         .addImm(0)            // tfe
         .addImm(CI.DLC)      // dlc
         .addImm(0)            // swz
         .addMemOperand(combineKnownAdjacentMMOs(*MBB->getParent(), MMOa, MMOb));
 
-  moveInstsAfter(MIB, CI.InstsToMove);
+  moveInstsAfter(MIB, InstsToMove);
 
   CI.I->eraseFromParent();
   Paired.I->eraseFromParent();
   return New;
 }
 
 MachineOperand
 SILoadStoreOptimizer::createRegOrImm(int32_t Val, MachineInstr &MI) const {
void SILoadStoreOptimizer::addInstToMergeableList(const CombineInfo &CI,
   }
 
   // Base address not found, so add a new list.
   MergeableInsts.emplace_back(1, CI);
 }
 
 bool SILoadStoreOptimizer::collectMergeableInsts(MachineBasicBlock &MBB,
                  std::list<std::list<CombineInfo> > &MergeableInsts) const {
   bool Modified = false;
   // Contain the list
   MemInfoMap Visited;
   // Contains the list of instructions for which constant offsets are being
   // promoted to the IMM.
   SmallPtrSet<MachineInstr *, 4> AnchorList;
 
   // Sort potential mergeable instructions into lists.  One list per base address.
+  unsigned Order = 0;
   for (MachineInstr &MI : MBB.instrs()) {
     // We run this before checking if an address is mergeable, because it can produce
     // better code even if the instructions aren't mergeable.
     if (promoteConstantOffsetToImm(MI, Visited, AnchorList))
       Modified = true;
 
     const InstClassEnum InstClass = getInstClass(MI.getOpcode(), *TII);
     if (InstClass == UNKNOWN)
       continue;
 
     // Don't combine if volatile.
     if (MI.hasOrderedMemoryRef())
       continue;
 
     CombineInfo CI;
     CI.setMI(MI, *TII, *STM);
+    CI.Order = Order++;
 
     if (!CI.hasMergeableAddress(*MRI))
       continue;
 
     addInstToMergeableList(CI, MergeableInsts);
   }
+
+  // At this point we have lists of Mergeable instructions.
+  //
+  // Part 2: Sort lists by offset and then for each CombineInfo object in the
+  // list try to find an instruction that can be merged with I.  If an instruction
+  // is found, it is stored in the Paired field.  If no instructions are found, then
+  // the CombineInfo object is deleted from the list.
+  
+  for (std::list<std::list<CombineInfo>>::iterator I = MergeableInsts.begin(),
+                                                   E = MergeableInsts.end(); I != E;) {
+
+    std::list<CombineInfo> &MergeList = *I;
+    if (MergeList.size() <= 1) {
+      // This means we have found only one instruction with a given address
+      // that can be merged, and we need at least 2 instructions to do a merge,
+      // so this list can be discarded.
+      I = MergeableInsts.erase(I);
+      continue;
+    }
+
+    // Sort the lists by offsets, this way mergeable instructions will be
+    // adjacent to each other in the list, which will make it easier to find
+    // matches.
+    MergeList.sort(
+        [] (const CombineInfo &A, CombineInfo &B) {
+          return A.Offset < B.Offset;
+        });
+    ++I;
+  }
+
   return Modified;
 }
 
 // Scan through looking for adjacent LDS operations with constant offsets from
 // the same base register. We rely on the scheduler to do the hard work of
 // clustering nearby loads, and assume these are all adjacent.
 bool SILoadStoreOptimizer::optimizeBlock(
                        std::list<std::list<CombineInfo> > &MergeableInsts) {
   bool Modified = false;
 
-  for (std::list<CombineInfo> &MergeList : MergeableInsts) {
-    if (MergeList.size() < 2)
-      continue;
+  for (std::list<std::list<CombineInfo>>::iterator I = MergeableInsts.begin(),
+                                                   E = MergeableInsts.end(); I != E;) {
+    std::list<CombineInfo> &MergeList = *I;
 
     bool OptimizeListAgain = false;
     if (!optimizeInstsWithSameBaseAddr(MergeList, OptimizeListAgain)) {
-      // We weren't able to make any changes, so clear the list so we don't
+      // We weren't able to make any changes, so delete the list so we don't
       // process the same instructions the next time we try to optimize this
       // block.
-      MergeList.clear();
+      I = MergeableInsts.erase(I);
       continue;
     }
 
-    // We made changes, but also determined that there were no more optimization
-    // opportunities, so we don't need to reprocess the list
-    if (!OptimizeListAgain)
-      MergeList.clear();
-
-    OptimizeAgain |= OptimizeListAgain;
     Modified = true;
-  }
-  return Modified;
-}
 
-void
-SILoadStoreOptimizer::removeCombinedInst(std::list<CombineInfo> &MergeList,
-                                         const MachineInstr &MI) {
-
-  for (auto CI = MergeList.begin(), E = MergeList.end(); CI != E; ++CI) {
-    if (&*CI->I == &MI) {
-      MergeList.erase(CI);
-      return;
+    // We made changes, but also determined that there were no more optimization
+    // opportunities, so we don't need to reprocess the list
+    if (!OptimizeListAgain) {
+      I = MergeableInsts.erase(I);
+      continue;
     }
+    OptimizeAgain = true;
   }
+  return Modified;
 }
 
 bool
 SILoadStoreOptimizer::optimizeInstsWithSameBaseAddr(
                                           std::list<CombineInfo> &MergeList,
                                           bool &OptimizeListAgain) {
+  if (MergeList.empty())
+    return false;
+
   bool Modified = false;
-  for (auto I = MergeList.begin(); I != MergeList.end(); ++I) {
-    CombineInfo &CI = *I;
-    CombineInfo Paired;
 
-    if (CI.InstClass == UNKNOWN)
-      continue;
+  for (auto I = MergeList.begin(), Next = std::next(I); Next != MergeList.end();
+       Next = std::next(I)) {
+
+    auto First = I;
+    auto Second = Next;
+
+    if ((*First).Order > (*Second).Order)
+      std::swap(First, Second);
+    CombineInfo &CI = *First;
+    CombineInfo &Paired = *Second;
 
-    if (!findMatchingInst(CI, Paired))
-      goto done;
+    SmallVector<MachineInstr *, 8> InstsToMove;
+    if (!checkAndPrepareMerge(CI, Paired, InstsToMove)) {
+      ++I;
+      continue;
+    }
 
     Modified = true;
-    removeCombinedInst(MergeList, *Paired.I);
 
     switch (CI.InstClass) {
     default:
       llvm_unreachable("unknown InstClass");
       break;
     case DS_READ: {
-      MachineBasicBlock::iterator NewMI = mergeRead2Pair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeRead2Pair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       break;
     }
     case DS_WRITE: {
-      MachineBasicBlock::iterator NewMI = mergeWrite2Pair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeWrite2Pair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       break;
     }
     case S_BUFFER_LOAD_IMM: {
-      MachineBasicBlock::iterator NewMI = mergeSBufferLoadImmPair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeSBufferLoadImmPair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       OptimizeListAgain |= (CI.Width + Paired.Width) < 16;
       break;
     }
     case BUFFER_LOAD: {
-      MachineBasicBlock::iterator NewMI = mergeBufferLoadPair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeBufferLoadPair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       OptimizeListAgain |= (CI.Width + Paired.Width) < 4;
       break;
     }
     case BUFFER_STORE: {
-      MachineBasicBlock::iterator NewMI = mergeBufferStorePair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeBufferStorePair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       OptimizeListAgain |= (CI.Width + Paired.Width) < 4;
       break;
     }
     case MIMG: {
-      MachineBasicBlock::iterator NewMI = mergeImagePair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeImagePair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       OptimizeListAgain |= (CI.Width + Paired.Width) < 4;
       break;
     }
     case TBUFFER_LOAD: {
-      MachineBasicBlock::iterator NewMI = mergeTBufferLoadPair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeTBufferLoadPair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       OptimizeListAgain |= (CI.Width + Paired.Width) < 4;
       break;
     }
     case TBUFFER_STORE: {
-      MachineBasicBlock::iterator NewMI = mergeTBufferStorePair(CI, Paired);
+      MachineBasicBlock::iterator NewMI =
+          mergeTBufferStorePair(CI, Paired, InstsToMove);
       CI.setMI(NewMI, *TII, *STM);
       OptimizeListAgain |= (CI.Width + Paired.Width) < 4;
       break;
     }
     }
+    CI.Order = Paired.Order;
+    if (I == Second)
+      I = Next;
 
-done:
-    // Clear the InstsToMove after we have finished searching so we don't have
-    // stale values left over if we search for this CI again in another pass
-    // over the block.
-    CI.InstsToMove.clear();
+    MergeList.erase(Second);
   }
 
   return Modified;
 }
 
 bool SILoadStoreOptimizer::runOnMachineFunction(MachineFunction &MF) {
   if (skipFunction(MF.getFunction()))
     return false;

llvm/test/CodeGen/AMDGPU/ds_read2_offset_order.ll

 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs < %s | FileCheck -strict-whitespace -check-prefix=SI %s
 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -strict-whitespace -check-prefix=SI %s
 
 @lds = addrspace(3) global [512 x float] undef, align 4
 
 ; offset0 is larger than offset1
 
 ; SI-LABEL: {{^}}offset_order:
-; SI-DAG: ds_read2st64_b32 v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} offset1:4{{$}}
+; SI-DAG: ds_read2_b32 v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} offset1:14{{$}}
 ; SI-DAG: ds_read2_b32 v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} offset0:2 offset1:3
-; SI-DAG: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:56
+; SI-DAG: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:1024
 ; SI-DAG: ds_read2_b32 v[{{[0-9]+}}:{{[0-9]+}}], v{{[0-9]+}} offset0:11 offset1:12
 define amdgpu_kernel void @offset_order(float addrspace(1)* %out) {
 entry:
   %ptr0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 0
   %val0 = load float, float addrspace(3)* %ptr0
 
   %ptr1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 256
   %val1 = load float, float addrspace(3)* %ptr1

llvm/test/CodeGen/AMDGPU/merge-load-store.mir

--- |
 
   define amdgpu_kernel void @move_waw_hazards() #0 {
     ret void
   }
 
   attributes #0 = { convergent nounwind }
 
   define amdgpu_kernel void @merge_mmos(i32 addrspace(1)* %ptr_addr1) { ret void }
+  define amdgpu_kernel void @reorder_offsets(i32 addrspace(1)* %reorder_addr1) { ret void }
 
 ...
 ---
 name:            mem_dependency
 alignment:       1
 exposesReturnsTwice: false
 legalized:       false
 regBankSelected: false
bb.0:
     %5:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET %0, 0, 64, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from %ir.ptr_addr1 + 64)
     %6:vgpr_32 = BUFFER_LOAD_DWORD_OFFSET %0, 0, 68, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 4 from %ir.ptr_addr1 + 68)
     BUFFER_STORE_DWORD_OFFSET_exact %5, %0, 0, 64, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.ptr_addr1 + 64)
     BUFFER_STORE_DWORD_OFFSET_exact %6, %0, 0, 68, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.ptr_addr1 + 68)
 
     S_ENDPGM 0
 
 ...
+---
+# CHECK-LABEL: reorder_offsets
+# CHECK-DAG: BUFFER_STORE_DWORDX2_OFFSET_exact killed %{{[0-9]+}}, %0, 0, 16, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8 into %ir.reorder_addr1 + 16, align 4, addrspace 1)
+# CHECK-DAG: BUFFER_STORE_DWORDX4_OFFSET_exact killed %{{[0-9]+}}, %0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16 into %ir.reorder_addr1, align 4, addrspace 1)
+
+name: reorder_offsets
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3
+
+    %0:sgpr_128 = COPY $sgpr0_sgpr1_sgpr2_sgpr3
+    %1:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    BUFFER_STORE_DWORD_OFFSET_exact %1, %0, 0, 4, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.reorder_addr1 + 4)
+    BUFFER_STORE_DWORD_OFFSET_exact %1, %0, 0, 8, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.reorder_addr1 + 8)
+    BUFFER_STORE_DWORD_OFFSET_exact %1, %0, 0, 12, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.reorder_addr1 + 12)
+    BUFFER_STORE_DWORD_OFFSET_exact %1, %0, 0, 16, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.reorder_addr1 + 16)
+    BUFFER_STORE_DWORD_OFFSET_exact %1, %0, 0, 20, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.reorder_addr1 + 20)
+    BUFFER_STORE_DWORD_OFFSET_exact %1, %0, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 4 into %ir.reorder_addr1)
+    S_ENDPGM 0
+
+
+...

llvm/test/CodeGen/AMDGPU/merge-tbuffer.mir

 ---
 
 
 # GFX9-LABEL: name: gfx9_tbuffer_load_float_32
 # GFX9: %{{[0-9]+}}:vreg_64 = TBUFFER_LOAD_FORMAT_XY_OFFSET %4, 0, 4, 123, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY %14.sub0
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %14.sub1
 # GFX9: %{{[0-9]+}}:vreg_128 = TBUFFER_LOAD_FORMAT_XYZW_OFFSET %4, 0, 16, 126, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 1, addrspace 4)
-# GFX9: %{{[0-9]+}}:vreg_64 = COPY %18.sub0_sub1
-# GFX9: %{{[0-9]+}}:vreg_64 = COPY killed %18.sub2_sub3
+# GFX9: %{{[0-9]+}}:vreg_96 = COPY %17.sub0_sub1_sub2
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub3
+# GFX9: %{{[0-9]+}}:vreg_64 = COPY %16.sub0_sub1
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub2
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY %15.sub0
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %15.sub1
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %16.sub0
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub1
 # GFX9: %{{[0-9]+}}:vreg_96 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET %4, 0, 36, 125, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 12, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = COPY %19.sub0_sub1
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %19.sub2
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %17.sub0
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub1
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %18.sub0
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %18.sub1
 
 name: gfx9_tbuffer_load_float_32
 body:             |
   bb.0.entry:
     %0:sgpr_32 = COPY $sgpr0
     %1:sgpr_32 = COPY $sgpr1
     %2:sgpr_32 = COPY $sgpr2
     %3:sgpr_32 = COPY $sgpr3

 ...
 ---
 
 # GFX9-LABEL: name: gfx9_tbuffer_load_sint_32
 # GFX9: %{{[0-9]+}}:vreg_64 = TBUFFER_LOAD_FORMAT_XY_OFFSET %4, 0, 4, 91, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY %14.sub0
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %14.sub1
 # GFX9: %{{[0-9]+}}:vreg_128 = TBUFFER_LOAD_FORMAT_XYZW_OFFSET %4, 0, 16, 94, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 1, addrspace 4)
-# GFX9: %{{[0-9]+}}:vreg_64 = COPY %18.sub0_sub1
-# GFX9: %{{[0-9]+}}:vreg_64 = COPY killed %18.sub2_sub3
+# GFX9: %{{[0-9]+}}:vreg_96 = COPY %17.sub0_sub1_sub2
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub3
+# GFX9: %{{[0-9]+}}:vreg_64 = COPY %16.sub0_sub1
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub2
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY %15.sub0
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %15.sub1
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %16.sub0
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub1
 # GFX9: %{{[0-9]+}}:vreg_96 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET %4, 0, 36, 93, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 12, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = COPY %19.sub0_sub1
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %19.sub2
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %17.sub0
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub1
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %18.sub0
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %18.sub1
 
 name: gfx9_tbuffer_load_sint_32
 body:             |
   bb.0.entry:
     %0:sgpr_32 = COPY $sgpr0
     %1:sgpr_32 = COPY $sgpr1
     %2:sgpr_32 = COPY $sgpr2
     %3:sgpr_32 = COPY $sgpr3

 ...
 ---
 
 # GFX9-LABEL: name: gfx9_tbuffer_load_uint_32
 # GFX9: %{{[0-9]+}}:vreg_64 = TBUFFER_LOAD_FORMAT_XY_OFFSET %4, 0, 4, 75, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY %14.sub0
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %14.sub1
 # GFX9: %{{[0-9]+}}:vreg_128 = TBUFFER_LOAD_FORMAT_XYZW_OFFSET %4, 0, 16, 78, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 1, addrspace 4)
-# GFX9: %{{[0-9]+}}:vreg_64 = COPY %18.sub0_sub1
-# GFX9: %{{[0-9]+}}:vreg_64 = COPY killed %18.sub2_sub3
+# GFX9: %{{[0-9]+}}:vreg_96 = COPY %17.sub0_sub1_sub2
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub3
+# GFX9: %{{[0-9]+}}:vreg_64 = COPY %16.sub0_sub1
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub2
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY %15.sub0
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %15.sub1
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %16.sub0
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub1
 # GFX9: %{{[0-9]+}}:vreg_96 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET %4, 0, 36, 77, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 12, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = COPY %19.sub0_sub1
 # GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %19.sub2
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %17.sub0
-# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub1
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY %18.sub0
+# GFX9: %{{[0-9]+}}:vgpr_32 = COPY killed %18.sub1
 
 name: gfx9_tbuffer_load_uint_32
 body:             |
   bb.0.entry:
     %0:sgpr_32 = COPY $sgpr0
     %1:sgpr_32 = COPY $sgpr1
     %2:sgpr_32 = COPY $sgpr2
     %3:sgpr_32 = COPY $sgpr3

 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr3
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr2
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr1
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr0
 # GFX9: %{{[0-9]+}}:sgpr_128 = REG_SEQUENCE %12, %subreg.sub0, %11, %subreg.sub1, %10, %subreg.sub2, %9, %subreg.sub3
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %8, %subreg.sub0, %7, %subreg.sub1
 # GFX9: TBUFFER_STORE_FORMAT_XY_OFFSET_exact killed %14, %13, 0, 4, 123, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %6, %subreg.sub0, %5, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %4, %subreg.sub0, %3, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, killed %16, %subreg.sub2_sub3
-# GFX9: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %18, %13, 0, 16, 126, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
+# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, %4, %subreg.sub2
+# GFX9: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %16, %subreg.sub0_sub1_sub2, %3, %subreg.sub3
+# GFX9: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %17, %13, 0, 16, 126, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, %1, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %17, %subreg.sub0_sub1, %0, %subreg.sub2
+# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %18, %subreg.sub0_sub1, %0, %subreg.sub2
 # GFX9: TBUFFER_STORE_FORMAT_XYZ_OFFSET_exact killed %19, %13, 0, 36, 125, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 12, align 1, addrspace 4)
 name: gfx9_tbuffer_store_float32
 body:             |
   bb.0.entry:
     liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
     %12:vgpr_32 = COPY $vgpr8
     %11:vgpr_32 = COPY $vgpr7
     %10:vgpr_32 = COPY $vgpr6

 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr3
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr2
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr1
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr0
 # GFX9: %{{[0-9]+}}:sgpr_128 = REG_SEQUENCE %12, %subreg.sub0, %11, %subreg.sub1, %10, %subreg.sub2, %9, %subreg.sub3
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %8, %subreg.sub0, %7, %subreg.sub1
 # GFX9: TBUFFER_STORE_FORMAT_XY_OFFSET_exact killed %14, %13, 0, 4, 91, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %6, %subreg.sub0, %5, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %4, %subreg.sub0, %3, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, killed %16, %subreg.sub2_sub3
-# GFX9: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %18, %13, 0, 16, 94, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
+# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, %4, %subreg.sub2
+# GFX9: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %16, %subreg.sub0_sub1_sub2, %3, %subreg.sub3
+# GFX9: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %17, %13, 0, 16, 94, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, %1, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %17, %subreg.sub0_sub1, %0, %subreg.sub2
+# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %18, %subreg.sub0_sub1, %0, %subreg.sub2
 # GFX9: TBUFFER_STORE_FORMAT_XYZ_OFFSET_exact killed %19, %13, 0, 36, 93, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 12, align 1, addrspace 4)
 name: gfx9_tbuffer_store_sint32
 body:             |
   bb.0.entry:
     liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
     %12:vgpr_32 = COPY $vgpr8
     %11:vgpr_32 = COPY $vgpr7
     %10:vgpr_32 = COPY $vgpr6

 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr3
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr2
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr1
 # GFX9: %{{[0-9]+}}:sgpr_32 = COPY $sgpr0
 # GFX9: %{{[0-9]+}}:sgpr_128 = REG_SEQUENCE %12, %subreg.sub0, %11, %subreg.sub1, %10, %subreg.sub2, %9, %subreg.sub3
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %8, %subreg.sub0, %7, %subreg.sub1
 # GFX9: TBUFFER_STORE_FORMAT_XY_OFFSET_exact killed %14, %13, 0, 4, 75, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %6, %subreg.sub0, %5, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %4, %subreg.sub0, %3, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, killed %16, %subreg.sub2_sub3
-# GFX9: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %18, %13, 0, 16, 78, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
+# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, %4, %subreg.sub2
+# GFX9: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %16, %subreg.sub0_sub1_sub2, %3, %subreg.sub3
+# GFX9: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %17, %13, 0, 16, 78, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
 # GFX9: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, %1, %subreg.sub1
-# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %17, %subreg.sub0_sub1, %0, %subreg.sub2
+# GFX9: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %18, %subreg.sub0_sub1, %0, %subreg.sub2
 # GFX9: TBUFFER_STORE_FORMAT_XYZ_OFFSET_exact killed %19, %13, 0, 36, 77, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 12, align 1, addrspace 4)
 name: gfx9_tbuffer_store_uint32
 body:             |
   bb.0.entry:
     liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
     %12:vgpr_32 = COPY $vgpr8
     %11:vgpr_32 = COPY $vgpr7
     %10:vgpr_32 = COPY $vgpr6

 ---
 
 
 # GFX10-LABEL: name: gfx10_tbuffer_load_float_32
 # GFX10: %{{[0-9]+}}:vreg_64 = TBUFFER_LOAD_FORMAT_XY_OFFSET %4, 0, 4, 64, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY %14.sub0
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %14.sub1
 # GFX10: %{{[0-9]+}}:vreg_128 = TBUFFER_LOAD_FORMAT_XYZW_OFFSET %4, 0, 16, 77, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 1, addrspace 4)
-# GFX10: %{{[0-9]+}}:vreg_64 = COPY %18.sub0_sub1
-# GFX10: %{{[0-9]+}}:vreg_64 = COPY killed %18.sub2_sub3
+# GFX10: %{{[0-9]+}}:vreg_96 = COPY %17.sub0_sub1_sub2
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub3
+# GFX10: %{{[0-9]+}}:vreg_64 = COPY %16.sub0_sub1
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub2
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY %15.sub0
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %15.sub1
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %16.sub0
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub1
 # GFX10: %{{[0-9]+}}:vreg_96 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET %4, 0, 36, 74, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 12, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = COPY %19.sub0_sub1
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %19.sub2
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %17.sub0
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub1
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %18.sub0
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %18.sub1
 
 name: gfx10_tbuffer_load_float_32
 body:             |
   bb.0.entry:
     %0:sgpr_32 = COPY $sgpr0
     %1:sgpr_32 = COPY $sgpr1
     %2:sgpr_32 = COPY $sgpr2
     %3:sgpr_32 = COPY $sgpr3

 ...
 ---
 
 # GFX10-LABEL: name: gfx10_tbuffer_load_sint_32
 # GFX10: %{{[0-9]+}}:vreg_64 = TBUFFER_LOAD_FORMAT_XY_OFFSET %4, 0, 4, 63, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY %14.sub0
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %14.sub1
 # GFX10: %{{[0-9]+}}:vreg_128 = TBUFFER_LOAD_FORMAT_XYZW_OFFSET %4, 0, 16, 76, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 1, addrspace 4)
-# GFX10: %{{[0-9]+}}:vreg_64 = COPY %18.sub0_sub1
-# GFX10: %{{[0-9]+}}:vreg_64 = COPY killed %18.sub2_sub3
+# GFX10: %{{[0-9]+}}:vreg_96 = COPY %17.sub0_sub1_sub2
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub3
+# GFX10: %{{[0-9]+}}:vreg_64 = COPY %16.sub0_sub1
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub2
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY %15.sub0
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %15.sub1
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %16.sub0
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub1
 # GFX10: %{{[0-9]+}}:vreg_96 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET %4, 0, 36, 73, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 12, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = COPY %19.sub0_sub1
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %19.sub2
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %17.sub0
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub1
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %18.sub0
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %18.sub1
 
 name: gfx10_tbuffer_load_sint_32
 body:             |
   bb.0.entry:
     %0:sgpr_32 = COPY $sgpr0
     %1:sgpr_32 = COPY $sgpr1
     %2:sgpr_32 = COPY $sgpr2
     %3:sgpr_32 = COPY $sgpr3

 ...
 ---
 
 # GFX10-LABEL: name: gfx10_tbuffer_load_uint_32
 # GFX10: %{{[0-9]+}}:vreg_64 = TBUFFER_LOAD_FORMAT_XY_OFFSET %4, 0, 4, 62, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 8, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY %14.sub0
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %14.sub1
 # GFX10: %{{[0-9]+}}:vreg_128 = TBUFFER_LOAD_FORMAT_XYZW_OFFSET %4, 0, 16, 75, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16, align 1, addrspace 4)
-# GFX10: %{{[0-9]+}}:vreg_64 = COPY %18.sub0_sub1
-# GFX10: %{{[0-9]+}}:vreg_64 = COPY killed %18.sub2_sub3
+# GFX10: %{{[0-9]+}}:vreg_96 = COPY %17.sub0_sub1_sub2
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub3
+# GFX10: %{{[0-9]+}}:vreg_64 = COPY %16.sub0_sub1
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub2
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY %15.sub0
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %15.sub1
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %16.sub0
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %16.sub1
 # GFX10: %{{[0-9]+}}:vreg_96 = TBUFFER_LOAD_FORMAT_XYZ_OFFSET %4, 0, 36, 72, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 12, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = COPY %19.sub0_sub1
 # GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %19.sub2
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %17.sub0
-# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %17.sub1
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY %18.sub0
+# GFX10: %{{[0-9]+}}:vgpr_32 = COPY killed %18.sub1
 
 name: gfx10_tbuffer_load_uint_32
 body:             |
   bb.0.entry:
     %0:sgpr_32 = COPY $sgpr0
     %1:sgpr_32 = COPY $sgpr1
     %2:sgpr_32 = COPY $sgpr2
     %3:sgpr_32 = COPY $sgpr3

 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr3
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr2
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr1
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr0
 # GFX10: %{{[0-9]+}}:sgpr_128 = REG_SEQUENCE %12, %subreg.sub0, %11, %subreg.sub1, %10, %subreg.sub2, %9, %subreg.sub3
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %8, %subreg.sub0, %7, %subreg.sub1
 # GFX10: TBUFFER_STORE_FORMAT_XY_OFFSET_exact killed %14, %13, 0, 4, 64, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %6, %subreg.sub0, %5, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %4, %subreg.sub0, %3, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, killed %16, %subreg.sub2_sub3
-# GFX10: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %18, %13, 0, 16, 77, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
+# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, %4, %subreg.sub2
+# GFX10: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %16, %subreg.sub0_sub1_sub2, %3, %subreg.sub3
+# GFX10: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %17, %13, 0, 16, 77, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, %1, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %17, %subreg.sub0_sub1, %0, %subreg.sub2
+# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %18, %subreg.sub0_sub1, %0, %subreg.sub2
 # GFX10: TBUFFER_STORE_FORMAT_XYZ_OFFSET_exact killed %19, %13, 0, 36, 74, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 12, align 1, addrspace 4)
 name: gfx10_tbuffer_store_float32
 body:             |
   bb.0.entry:
     liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
     %12:vgpr_32 = COPY $vgpr8
     %11:vgpr_32 = COPY $vgpr7
     %10:vgpr_32 = COPY $vgpr6

 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr3
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr2
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr1
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr0
 # GFX10: %{{[0-9]+}}:sgpr_128 = REG_SEQUENCE %12, %subreg.sub0, %11, %subreg.sub1, %10, %subreg.sub2, %9, %subreg.sub3
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %8, %subreg.sub0, %7, %subreg.sub1
 # GFX10: TBUFFER_STORE_FORMAT_XY_OFFSET_exact killed %14, %13, 0, 4, 63, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %6, %subreg.sub0, %5, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %4, %subreg.sub0, %3, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, killed %16, %subreg.sub2_sub3
-# GFX10: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %18, %13, 0, 16, 76, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
+# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, %4, %subreg.sub2
+# GFX10: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %16, %subreg.sub0_sub1_sub2, %3, %subreg.sub3
+# GFX10: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %17, %13, 0, 16, 76, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, %1, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %17, %subreg.sub0_sub1, %0, %subreg.sub2
+# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %18, %subreg.sub0_sub1, %0, %subreg.sub2
 # GFX10: TBUFFER_STORE_FORMAT_XYZ_OFFSET_exact killed %19, %13, 0, 36, 73, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 12, align 1, addrspace 4)
 name: gfx10_tbuffer_store_sint32
 body:             |
   bb.0.entry:
     liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
     %12:vgpr_32 = COPY $vgpr8
     %11:vgpr_32 = COPY $vgpr7
     %10:vgpr_32 = COPY $vgpr6

 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr3
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr2
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr1
 # GFX10: %{{[0-9]+}}:sgpr_32 = COPY $sgpr0
 # GFX10: %{{[0-9]+}}:sgpr_128 = REG_SEQUENCE %12, %subreg.sub0, %11, %subreg.sub1, %10, %subreg.sub2, %9, %subreg.sub3
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %8, %subreg.sub0, %7, %subreg.sub1
 # GFX10: TBUFFER_STORE_FORMAT_XY_OFFSET_exact killed %14, %13, 0, 4, 62, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 8, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %6, %subreg.sub0, %5, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %4, %subreg.sub0, %3, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, killed %16, %subreg.sub2_sub3
-# GFX10: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %18, %13, 0, 16, 75, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
+# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %15, %subreg.sub0_sub1, %4, %subreg.sub2
+# GFX10: %{{[0-9]+}}:vreg_128 = REG_SEQUENCE killed %16, %subreg.sub0_sub1_sub2, %3, %subreg.sub3
+# GFX10: TBUFFER_STORE_FORMAT_XYZW_OFFSET_exact killed %17, %13, 0, 16, 75, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 16, align 1, addrspace 4)
 # GFX10: %{{[0-9]+}}:vreg_64 = REG_SEQUENCE %2, %subreg.sub0, %1, %subreg.sub1
-# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %17, %subreg.sub0_sub1, %0, %subreg.sub2
+# GFX10: %{{[0-9]+}}:vreg_96 = REG_SEQUENCE killed %18, %subreg.sub0_sub1, %0, %subreg.sub2
 # GFX10: TBUFFER_STORE_FORMAT_XYZ_OFFSET_exact killed %19, %13, 0, 36, 72, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable store 12, align 1, addrspace 4)
 name: gfx10_tbuffer_store_uint32
 body:             |
   bb.0.entry:
     liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
     %12:vgpr_32 = COPY $vgpr8
     %11:vgpr_32 = COPY $vgpr7
     %10:vgpr_32 = COPY $vgpr6