Merge memtag instructions with adjacent stack slots.

[eugenis]

Summary

Detect a run of memory tagging instructions for adjacent stack frame slots,
and replace them with a shorter instruction sequence

• replace STG + STG with ST2G
• replace STGloop + STGloop with STGloop

This code needs to run when stack slot offsets are already known, but before
FrameIndex operands in STG instructions are eliminated; that's the
reason for the new hook in PrologueEpilogue.

This change modifies STGloop and STZGloop pseudos to take the size as an
immediate integer operand, and base address as a FI operand when
possible. This is needed to simplify recognizing an STGloop instruction
as operating on a stack slot post-regalloc.

This improves memtag code size by ~0.25%, and it looks like an additional ~0.1%
is possible by rearranging the stack frame such that consecutive STG
instructions reference adjacent slots (patch pending).

[eugenis]

Ping.

Most of the smaller changes in this patch (almost everything outside of AArch64FrameLowering.cpp) are needed to handle FrameIndex operand in STGloop. It is so special because STGloop has a writeback constraint, which regalloc has no way of honoring in the case when the input operand is FrameIndex; therefore eliminateFrameIndex has to be smart about the choice of the scratch register there.

[eugenis]

Cleaned up the code and extended it to allow gaps in a sequence of memory tagging instructions (in which case subsequences before and after the gap would be optimized separatly).

[merge_guards_bot]

Unit tests: pass. 61177 tests passed, 0 failed and 729 were skipped.

clang-tidy: fail. Please fix clang-tidy findings.

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-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

[eugenis]

clang-format

[eugenis]

PTAL

[merge_guards_bot]

Unit tests: pass. 61177 tests passed, 0 failed and 729 were skipped.

clang-tidy: fail. Please fix clang-tidy findings.

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-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

[ostannard]

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

2718

Instructions must be added in the ascending order of Offset, and have to be adjacent.

Could we add an assertion to verify this? Either TagStores must be empty, or this instruction must be adjacent to the last one.

2854

Why do we want to return an empty list if any one input instruction doesn't have a memoperand?

2882

Where does the number 176 come from?

3012

This can be Instrs.back()

llvm/lib/Target/AArch64/AArch64InstrInfo.td

1517

This comment needs updating to mention the size operand.

llvm/lib/Target/AArch64/AArch64RegisterInfo.h

127

Whitespace change.

[eugenis]

Addressed review comments and most of the clang-tidy findings.

[eugenis]

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

2854

Added a comment. No memoperand wins over any memoperand.
See MachineInstr::mayAlias, MachineInstr::cloneMergedMemRefs.

2882

It's approximately the breaking point between loop and unrolled STG sequence in terms of code size. Anything <176 bytes can be tagged with 5 instructions or less; the loop body itself is 3, but it usually needs one or two instructions to prepare the counter and the base pointer.

We can tune it for performance later when the hardware is available.

[merge_guards_bot]

Unit tests: pass. 61260 tests passed, 0 failed and 736 were skipped.

clang-tidy: fail. Please fix clang-tidy findings.

clang-format: pass.

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

[ostannard]

LGTM

[eugenis]

llvm/lib/Target/AArch64/AArch64RegisterInfo.cpp

457

Hmm, there is actually a check against this in the verifier:

• • Bad machine code: Tied use must be a register ***
• function: stg_alloca17
• basic block: %bb.0 entry (0x20076710580)
• instruction: early-clobber %0:gpr64common, early-clobber %1:gpr64sp = STGloop 272, %stack.0.a :: (store 272 into %ir.a, align 16)
• operand 3: %stack.0.a
  
  http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-win/builds/21481/steps/test-check-all/logs/stdio

I've reverted the change for now.

[eugenis]

Fixed the tied use of a non-register problem by introducing _untied versions to STGloop pseudos for use with a FrameIndex operand.

[merge_guards_bot]

Unit tests: pass. 61321 tests passed, 0 failed and 736 were skipped.

clang-tidy: pass.

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-tidy.txt, clang-format.patch, CMakeCache.txt, console-log.txt, test-results.xml

[ostannard]

llvm/lib/Target/AArch64/AArch64RegisterInfo.cpp

460

Based on the name of the function, I wouldn't expect this to modify the instruction passed in. I think it would be better to hoist this code out to the call-site, or rename the function to something like createScratchRegisterForInstruction, and note that it might do this in a comment.

[eugenis]

Updated a comment.
Renamed "old" STGloop instructions to STGloop_wback, and new ones to STGloop.

[merge_guards_bot]

Unit tests: pass. 61728 tests passed, 0 failed and 779 were skipped.

clang-tidy: fail. Please fix clang-tidy findings.

clang-format: pass.

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

[eugenis]

friendly ping

[ostannard]

LGTM

llvm/include/llvm/CodeGen/TargetFrameLowering.h

public:
   /// before the specified function's frame layout (MF.getFrameInfo()) is
   /// finalized.  Once the frame is finalized, MO_FrameIndex operands are
   /// replaced with direct constants.  This method is optional.
   ///
   virtual void processFunctionBeforeFrameFinalized(MachineFunction &MF,
                                              RegScavenger *RS = nullptr) const {
   }
 
+  /// processFunctionBeforeFrameIndicesReplaced - This method is called
+  /// immediately before MO_FrameIndex operands are eliminated, but after the
+  /// frame is finalized. This method is optional.
+  virtual void
+  processFunctionBeforeFrameIndicesReplaced(MachineFunction &MF,
+                                            RegScavenger *RS = nullptr) const {}
+
   virtual unsigned getWinEHParentFrameOffset(const MachineFunction &MF) const {
     report_fatal_error("WinEH not implemented for this target");
   }
 
   /// This method is called during prolog/epilog code insertion to eliminate
   /// call frame setup and destroy pseudo instructions (but only if the Target
   /// is using them).  It is responsible for eliminating these instructions,
   /// replacing them with concrete instructions.  This method need only be

llvm/lib/CodeGen/PrologEpilogInserter.cpp

bool PEI::runOnMachineFunction(MachineFunction &MF) {
   // and MaxCallFrameSize variables.
   if (!F.hasFnAttribute(Attribute::Naked))
     insertPrologEpilogCode(MF);
 
   // Reinsert stashed debug values at the start of the entry blocks.
   for (auto &I : EntryDbgValues)
     I.first->insert(I.first->begin(), I.second.begin(), I.second.end());
 
+  // Allow the target machine to make final modifications to the function
+  // before the frame layout is finalized.
+  TFI->processFunctionBeforeFrameIndicesReplaced(MF, RS);
+
   // Replace all MO_FrameIndex operands with physical register references
   // and actual offsets.
   //
   replaceFrameIndices(MF);
 
   // If register scavenging is needed, as we've enabled doing it as a
   // post-pass, scavenge the virtual registers that frame index elimination
   // inserted.

llvm/lib/Target/AArch64/AArch64ExpandPseudoInsts.cpp

bool AArch64ExpandPseudo::expandCMP_SWAP_128(
   return true;
 }
 
 bool AArch64ExpandPseudo::expandSetTagLoop(
     MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,
     MachineBasicBlock::iterator &NextMBBI) {
   MachineInstr &MI = *MBBI;
   DebugLoc DL = MI.getDebugLoc();
-  Register SizeReg = MI.getOperand(2).getReg();
-  Register AddressReg = MI.getOperand(3).getReg();
+  Register SizeReg = MI.getOperand(0).getReg();
+  Register AddressReg = MI.getOperand(1).getReg();
 
   MachineFunction *MF = MBB.getParent();
 
   bool ZeroData = MI.getOpcode() == AArch64::STZGloop;
-  const unsigned OpCode =
+  const unsigned OpCode1 =
+      ZeroData ? AArch64::STZGPostIndex : AArch64::STGPostIndex;
+  const unsigned OpCode2 =
       ZeroData ? AArch64::STZ2GPostIndex : AArch64::ST2GPostIndex;
 
+  unsigned Size = MI.getOperand(2).getImm();
+  assert(Size > 0 && Size % 16 == 0);
+  if (Size % (16 * 2) != 0) {
+    BuildMI(MBB, MBBI, DL, TII->get(OpCode1), AddressReg)
+        .addReg(AddressReg)
+        .addReg(AddressReg)
+        .addImm(1);
+    Size -= 16;
+  }
+  MachineBasicBlock::iterator I =
+      BuildMI(MBB, MBBI, DL, TII->get(AArch64::MOVi64imm), SizeReg)
+          .addImm(Size);
+  expandMOVImm(MBB, I, 64);
+
   auto LoopBB = MF->CreateMachineBasicBlock(MBB.getBasicBlock());
   auto DoneBB = MF->CreateMachineBasicBlock(MBB.getBasicBlock());
 
   MF->insert(++MBB.getIterator(), LoopBB);
   MF->insert(++LoopBB->getIterator(), DoneBB);
 
-  BuildMI(LoopBB, DL, TII->get(OpCode))
+  BuildMI(LoopBB, DL, TII->get(OpCode2))
       .addDef(AddressReg)
       .addReg(AddressReg)
       .addReg(AddressReg)
       .addImm(2)
       .cloneMemRefs(MI)
       .setMIFlags(MI.getFlags());
   BuildMI(LoopBB, DL, TII->get(AArch64::SUBXri))
       .addDef(SizeReg)

llvm/lib/Target/AArch64/AArch64FrameLowering.h

public:
   }
 
   bool enableStackSlotScavenging(const MachineFunction &MF) const override;
   TargetStackID::Value getStackIDForScalableVectors() const override;
 
   void processFunctionBeforeFrameFinalized(MachineFunction &MF,
                                              RegScavenger *RS) const override;
 
+  void processFunctionBeforeFrameIndicesReplaced(MachineFunction &MF,
+                                                 RegScavenger *RS) const override;
+
   unsigned getWinEHParentFrameOffset(const MachineFunction &MF) const override;
 
   unsigned getWinEHFuncletFrameSize(const MachineFunction &MF) const;
 
   int getFrameIndexReferencePreferSP(const MachineFunction &MF, int FI,
                                      unsigned &FrameReg,
                                      bool IgnoreSPUpdates) const override;
   int getNonLocalFrameIndexReference(const MachineFunction &MF,

 private:
   bool shouldCombineCSRLocalStackBump(MachineFunction &MF,
                                       uint64_t StackBumpBytes) const;
 
   int64_t estimateSVEStackObjectOffsets(MachineFrameInfo &MF) const;
   int64_t assignSVEStackObjectOffsets(MachineFrameInfo &MF,
                                       int &MinCSFrameIndex,
                                       int &MaxCSFrameIndex) const;
+  bool shouldCombineCSRLocalStackBumpInEpilogue(MachineBasicBlock &MBB,
+                                                unsigned StackBumpBytes) const;
 };
 
 } // End llvm namespace
 
 #endif

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

static cl::opt<bool> EnableRedZone("aarch64-redzone",
                                    cl::desc("enable use of redzone on AArch64"),
                                    cl::init(false), cl::Hidden);
 
 static cl::opt<bool>
     ReverseCSRRestoreSeq("reverse-csr-restore-seq",
                          cl::desc("reverse the CSR restore sequence"),
                          cl::init(false), cl::Hidden);
 
+static cl::opt<bool> StackTaggingMergeSetTag("stack-tagging-merge-settag",
+                                   cl::desc("merge settag instruction in function epilog"),
+                                   cl::init(true), cl::Hidden);
+
 STATISTIC(NumRedZoneFunctions, "Number of functions using red zone");
 
 /// This is the biggest offset to the stack pointer we can encode in aarch64
 /// instructions (without using a separate calculation and a temp register).
 /// Note that the exception here are vector stores/loads which cannot encode any
 /// displacements (see estimateRSStackSizeLimit(), isAArch64FrameOffsetLegal()).
 static const unsigned DefaultSafeSPDisplacement = 255;
 
bool AArch64FrameLowering::shouldCombineCSRLocalStackBump(
   // When there is an SVE area on the stack, always allocate the
   // callee-saves and spills/locals separately.
   if (getSVEStackSize(MF))
     return false;
 
   return true;
 }
 
+bool AArch64FrameLowering::shouldCombineCSRLocalStackBumpInEpilogue(
+    MachineBasicBlock &MBB, unsigned StackBumpBytes) const {
+  if (!shouldCombineCSRLocalStackBump(*MBB.getParent(), StackBumpBytes))
+    return false;
+
+  if (MBB.empty())
+    return true;
+
+  // Disable combined SP bump if the last instruction is an MTE tag store. It
+  // is almost always better to merge SP adjustment into those instructions.
+  MachineBasicBlock::iterator LastI = MBB.getFirstTerminator();
+  MachineBasicBlock::iterator Begin = MBB.begin();
+  while (LastI != Begin) {
+    --LastI;
+    if (LastI->isTransient())
+      continue;
+    if (!LastI->getFlag(MachineInstr::FrameDestroy))
+      break;
+  }
+  switch (LastI->getOpcode()) {
+    case AArch64::STGloop:
+    case AArch64::STZGloop:
+    case AArch64::STGOffset:
+    case AArch64::STZGOffset:
+    case AArch64::ST2GOffset:
+    case AArch64::STZ2GOffset:
+      return false;
+    default:
+      return true;
+  }
+  llvm_unreachable("unreachable");
+}
+
 // Given a load or a store instruction, generate an appropriate unwinding SEH
 // code on Windows.
 static MachineBasicBlock::iterator InsertSEH(MachineBasicBlock::iterator MBBI,
                                              const TargetInstrInfo &TII,
                                              MachineInstr::MIFlag Flag) {
   unsigned Opc = MBBI->getOpcode();
   MachineBasicBlock *MBB = MBBI->getParent();
   MachineFunction &MF = *MBB->getParent();
   DebugLoc DL = MBBI->getDebugLoc();
   unsigned ImmIdx = MBBI->getNumOperands() - 1;
   int Imm = MBBI->getOperand(ImmIdx).getImm();
void AArch64FrameLowering::emitEpilogue(MachineFunction &MF,
   uint64_t AfterCSRPopSize = ArgumentPopSize;
   auto PrologueSaveSize = AFI->getCalleeSavedStackSize() + FixedObject;
   // We cannot rely on the local stack size set in emitPrologue if the function
   // has funclets, as funclets have different local stack size requirements, and
   // the current value set in emitPrologue may be that of the containing
   // function.
   if (MF.hasEHFunclets())
     AFI->setLocalStackSize(NumBytes - PrologueSaveSize);
-  bool CombineSPBump = shouldCombineCSRLocalStackBump(MF, NumBytes);
+  bool CombineSPBump = shouldCombineCSRLocalStackBumpInEpilogue(MBB, NumBytes);
   // Assume we can't combine the last pop with the sp restore.
 
   if (!CombineSPBump && PrologueSaveSize != 0) {
     MachineBasicBlock::iterator Pop = std::prev(MBB.getFirstTerminator());
     while (AArch64InstrInfo::isSEHInstruction(*Pop))
       Pop = std::prev(Pop);
     // Converting the last ldp to a post-index ldp is valid only if the last
     // ldp's offset is 0.
void AArch64FrameLowering::processFunctionBeforeFrameFinalized(
   assert(DstReg && "There must be a free register after frame setup");
   BuildMI(MBB, MBBI, DL, TII.get(AArch64::MOVi64imm), DstReg).addImm(-2);
   BuildMI(MBB, MBBI, DL, TII.get(AArch64::STURXi))
       .addReg(DstReg, getKillRegState(true))
       .addFrameIndex(UnwindHelpFI)
       .addImm(0);
 }
 
-/// For Win64 AArch64 EH, the offset to the Unwind object is from the SP before
-/// the update.  This is easily retrieved as it is exactly the offset that is set
-/// in processFunctionBeforeFrameFinalized.
+namespace {
+struct TagStoreInstr {
+  MachineInstr *MI;
+  int64_t Offset, Size;
+  explicit TagStoreInstr(MachineInstr *MI, int64_t Offset, int64_t Size)
+      : MI(MI), Offset(Offset), Size(Size) {}
+};
+
+class TagStoreEdit {
+  MachineFunction *MF;
+  MachineBasicBlock *MBB;
+  MachineRegisterInfo *MRI;
+  // Tag store instructions that are being replaced.
+  SmallVector<TagStoreInstr, 8> TagStores;
+  // Combined memref arguments of the above instructions.
+  SmallVector<MachineMemOperand *, 8> CombinedMemRefs;
+
+  // Replace allocation tags in [FrameReg + FrameRegOffset, FrameReg +
+  // FrameRegOffset + Size) with the address tag of SP.
+  unsigned FrameReg;
+  StackOffset FrameRegOffset;
+  int64_t Size;
+  // If not None, move FrameReg to (FrameReg + FrameRegUpdate) at the end.
+  Optional<int64_t> FrameRegUpdate;
+  // MIFlags for any FrameReg updating instructions.
+  unsigned FrameRegUpdateFlags;
+
+  // Use zeroing instruction variants.
+  bool ZeroData;
+  DebugLoc DL;
+
+  void EmitUnrolled(MachineBasicBlock::iterator InsertI);
+  void EmitLoop(MachineBasicBlock::iterator InsertI);
+
+public:
+  TagStoreEdit(MachineBasicBlock *MBB, bool ZeroData)
+      : MBB(MBB), ZeroData(ZeroData) {
+    MF = MBB->getParent();
+    MRI = &MF->getRegInfo();
+  }
+  // Add an instruction to be replaced. Instructions must be added in the
+  // ascending order of Offset, and have to be adjacent.
+  void AddInstruction(TagStoreInstr I) { TagStores.push_back(I); }
+  void Clear() { TagStores.clear(); }
+  // Emit equivalent code at the given location, and erase the current set of
+  // instructions. May skip if the replacement is not profitable. May invalidate
+  // the input iterator and replace it with a valid one.
+  void EmitCode(MachineBasicBlock::iterator &InsertI,
+                const AArch64FrameLowering *TFI, bool IsLast);
+};
+
+void TagStoreEdit::EmitUnrolled(MachineBasicBlock::iterator InsertI) {
+  const AArch64InstrInfo *TII =
+      MF->getSubtarget<AArch64Subtarget>().getInstrInfo();
+
+  const int64_t kMinOffset = -256 * 16;
+  const int64_t kMaxOffset = 255 * 16;
+
+  unsigned BaseReg = FrameReg;
+  int64_t BaseRegOffsetBytes = FrameRegOffset.getBytes();
+  if (BaseRegOffsetBytes < kMinOffset ||
+      BaseRegOffsetBytes + (Size - Size % 32) > kMaxOffset) {
+    unsigned ScratchReg = MRI->createVirtualRegister(&AArch64::GPR64RegClass);
+    emitFrameOffset(*MBB, InsertI, DL, ScratchReg, BaseReg,
+                    {BaseRegOffsetBytes, MVT::i8}, TII);
+    BaseReg = ScratchReg;
+    BaseRegOffsetBytes = 0;
+  }
+
+  MachineInstr *LastI = nullptr;
+  while (Size) {
+    int64_t InstrSize = (Size > 16) ? 32 : 16;
+    unsigned Opcode =
+        InstrSize == 16
+            ? (ZeroData ? AArch64::STZGOffset : AArch64::STGOffset)
+            : (ZeroData ? AArch64::STZ2GOffset : AArch64::ST2GOffset);
+    MachineInstr *I = BuildMI(*MBB, InsertI, DL, TII->get(Opcode))
+                          .addReg(AArch64::SP)
+                          .addReg(BaseReg)
+                          .addImm(BaseRegOffsetBytes / 16)
+                          .setMemRefs(CombinedMemRefs);
+    // A store to [BaseReg, #0] should go last for an opportunity to fold the
+    // final SP adjustment in the epilogue.
+    if (BaseRegOffsetBytes == 0)
+      LastI = I;
+    BaseRegOffsetBytes += InstrSize;
+    Size -= InstrSize;
+  }
+
+  if (LastI)
+    MBB->splice(InsertI, MBB, LastI);
+}
+
+void TagStoreEdit::EmitLoop(MachineBasicBlock::iterator InsertI) {
+  const AArch64InstrInfo *TII =
+      MF->getSubtarget<AArch64Subtarget>().getInstrInfo();
+
+  unsigned BaseReg =
+      FrameRegUpdate
+          ? BaseReg = FrameReg
+          : BaseReg = MRI->createVirtualRegister(&AArch64::GPR64RegClass);
+  unsigned SizeReg = MRI->createVirtualRegister(&AArch64::GPR64RegClass);
+
+  emitFrameOffset(*MBB, InsertI, DL, BaseReg, FrameReg, FrameRegOffset, TII);
+
+  int64_t LoopSize = Size;
+  // If the loop size is not a multiple of 32, split off one 16-byte store at
+  // the end to fold BaseReg update into.
+  if (FrameRegUpdate && *FrameRegUpdate)
+    LoopSize -= LoopSize % 32;
+  MachineInstr *LoopI =
+      BuildMI(*MBB, InsertI, DL,
+              TII->get(ZeroData ? AArch64::STZGloop : AArch64::STGloop))
+          .addDef(SizeReg)
+          .addDef(BaseReg)
+          .addImm(LoopSize)
+          .addReg(BaseReg)
+          .setMemRefs(CombinedMemRefs);
+  if (FrameRegUpdate)
+    LoopI->setFlags(FrameRegUpdateFlags);
+
+  int64_t ExtraBaseRegUpdate =
+      FrameRegUpdate ? (*FrameRegUpdate - FrameRegOffset.getBytes() - Size) : 0;
+  if (LoopSize < Size) {
+    assert(FrameRegUpdate);
+    assert(Size - LoopSize == 16);
+    // Tag 16 more bytes at BaseReg and update BaseReg.
+    BuildMI(*MBB, InsertI, DL,
+            TII->get(ZeroData ? AArch64::STZGPostIndex : AArch64::STGPostIndex))
+        .addDef(BaseReg)
+        .addReg(BaseReg)
+        .addReg(BaseReg)
+        .addImm(1 + ExtraBaseRegUpdate / 16)
+        .setMemRefs(CombinedMemRefs)
+        .setMIFlags(FrameRegUpdateFlags);
+  } else if (ExtraBaseRegUpdate) {
+    // Update BaseReg.
+    BuildMI(
+        *MBB, InsertI, DL,
+        TII->get(ExtraBaseRegUpdate > 0 ? AArch64::ADDXri : AArch64::SUBXri))
+        .addDef(BaseReg)
+        .addReg(BaseReg)
+        .addImm(std::abs(ExtraBaseRegUpdate))
+        .addImm(0)
+        .setMIFlags(FrameRegUpdateFlags);
+  }
+}
+
+// Check if *II is a register update that can be merged into STGloop that ends
+// at (Reg + Size). RemainingOffset is the required adjustment to Reg after the
+// end of the loop.
+bool canMergeRegUpdate(MachineBasicBlock::iterator II, unsigned Reg,
+                       int64_t Size, int64_t *TotalOffset) {
+  MachineInstr &MI = *II;
+  if ((MI.getOpcode() == AArch64::ADDXri ||
+       MI.getOpcode() == AArch64::SUBXri) &&
+      MI.getOperand(0).getReg() == Reg && MI.getOperand(1).getReg() == Reg) {
+    unsigned Shift = AArch64_AM::getShiftValue(MI.getOperand(3).getImm());
+    int64_t Offset = MI.getOperand(2).getImm() << Shift;
+    if (MI.getOpcode() == AArch64::SUBXri)
+      Offset = -Offset;
+    int64_t AbsPostOffset = std::abs(Offset - Size);
+    const int64_t kMaxOffset =
+        0xFFF; // Max encoding for unshifted ADDXri / SUBXri
+    if (AbsPostOffset <= kMaxOffset && AbsPostOffset % 16 == 0) {
+      *TotalOffset = Offset;
+      return true;
+    }
+  }
+  return false;
+}
+
+void MergeMemRefs(const SmallVectorImpl<TagStoreInstr> &TSE,
+                  SmallVectorImpl<MachineMemOperand *> &MemRefs) {
+  MemRefs.clear();
+  for (auto &TS : TSE) {
+    MachineInstr *MI = TS.MI;
+    if (MI->memoperands_empty()) {
+      MemRefs.clear();
+      return;
+    }
+    MemRefs.append(MI->memoperands_begin(), MI->memoperands_end());
+  }
+}
+
+void TagStoreEdit::EmitCode(MachineBasicBlock::iterator &InsertI,
+                            const AArch64FrameLowering *TFI, bool IsLast) {
+  if (TagStores.empty())
+    return;
+  TagStoreInstr &FirstTagStore = TagStores[0];
+  TagStoreInstr &LastTagStore = TagStores[TagStores.size() - 1];
+  Size = LastTagStore.Offset - FirstTagStore.Offset + LastTagStore.Size;
+  DL = TagStores[0].MI->getDebugLoc();
+
+  FrameRegOffset = TFI->resolveFrameOffsetReference(
+      *MF, FirstTagStore.Offset, false /*isFixed*/, false /*isSVE*/, FrameReg,
+      /*PreferFP=*/false, /*ForSimm=*/true);
+  FrameRegUpdate = None;
+
+  MergeMemRefs(TagStores, CombinedMemRefs);
+
+  LLVM_DEBUG(dbgs() << "Replacing adjacent STG instructions:\n";
+             for (const auto &Instr
+                  : TagStores) { dbgs() << "  " << *Instr.MI; });
+
+  const int kSetTagLoopThreshold = 176;
+  if (Size < kSetTagLoopThreshold) {
+    if (TagStores.size() < 2)
+      return;
+    EmitUnrolled(InsertI);
+  } else {
+    MachineInstr *UpdateInstr = nullptr;
+    int64_t TotalOffset;
+    if (IsLast) {
+      // See if we can merge base register update into the STGloop.
+      // This is done in AArch64LoadStoreOptimizer for "normal" stores,
+      // but STGloop is way too unusual for that, and also it only
+      // realistically happens in function epilogue. Also, STGloop is expanded
+      // before that pass.
+      if (InsertI != MBB->end() &&
+          canMergeRegUpdate(InsertI, FrameReg, FrameRegOffset.getBytes() + Size,
+                            &TotalOffset)) {
+        UpdateInstr = &*InsertI++;
+        LLVM_DEBUG(dbgs() << "Folding SP update into loop:\n  "
+                          << *UpdateInstr);
+      }
+    }
+
+    if (!UpdateInstr && TagStores.size() < 2)
+      return;
+
+    if (UpdateInstr) {
+      FrameRegUpdate = TotalOffset;
+      FrameRegUpdateFlags = UpdateInstr->getFlags();
+    }
+    EmitLoop(InsertI);
+    if (UpdateInstr)
+      UpdateInstr->eraseFromParent();
+  }
+
+  for (auto &TS : TagStores)
+    TS.MI->eraseFromParent();
+}
+
+bool isMergeableStackTaggingInstruction(MachineInstr &MI, int64_t &Offset,
+                                        int64_t &Size, bool &ZeroData) {
+  MachineFunction &MF = *MI.getParent()->getParent();
+  const MachineFrameInfo &MFI = MF.getFrameInfo();
+
+  unsigned Opcode = MI.getOpcode();
+  ZeroData = (Opcode == AArch64::STZGloop || Opcode == AArch64::STZGOffset ||
+              Opcode == AArch64::STZ2GOffset);
+
+  if (Opcode == AArch64::STGloop || Opcode == AArch64::STZGloop) {
+    if (!MI.getOperand(0).isDead() || !MI.getOperand(1).isDead())
+      return false;
+    if (!MI.getOperand(2).isImm() || !MI.getOperand(3).isFI())
+      return false;
+    Offset = MFI.getObjectOffset(MI.getOperand(3).getIndex());
+    Size = MI.getOperand(2).getImm();
+    return true;
+  }
+
+  if (Opcode == AArch64::STGOffset || Opcode == AArch64::STZGOffset)
+    Size = 16;
+  else if (Opcode == AArch64::ST2GOffset || Opcode == AArch64::STZ2GOffset)
+    Size = 32;
+  else
+    return false;
+
+  if (MI.getOperand(0).getReg() != AArch64::SP || !MI.getOperand(1).isFI())
+    return false;
+
+  Offset = MFI.getObjectOffset(MI.getOperand(1).getIndex()) +
+           16 * MI.getOperand(2).getImm();
+  return true;
+}
+
+// Detect a run of memory tagging instructions for adjacent stack frame slots,
+// and replace them with a shorter instruction sequence:
+// * replace STG + STG with ST2G
+// * replace STGloop + STGloop with STGloop
+// This code needs to run when stack slot offsets are already known, but before
+// FrameIndex operands in STG instructions are eliminated.
+MachineBasicBlock::iterator tryMergeAdjacentSTG(MachineBasicBlock::iterator II,
+                                                const AArch64FrameLowering *TFI,
+                                                RegScavenger *RS) {
+  bool FirstZeroData;
+  int64_t Size, Offset;
+  MachineInstr &MI = *II;
+  MachineBasicBlock *MBB = MI.getParent();
+  MachineBasicBlock::iterator NextI = ++II;
+  if (&MI == &MBB->instr_back())
+    return II;
+  if (!isMergeableStackTaggingInstruction(MI, Offset, Size, FirstZeroData))
+    return II;
+
+  SmallVector<TagStoreInstr, 4> Instrs;
+  Instrs.emplace_back(&MI, Offset, Size);
+
+  constexpr int kScanLimit = 10;
+  int Count = 0;
+  for (MachineBasicBlock::iterator E = MBB->end();
+       NextI != E && Count < kScanLimit; ++NextI) {
+    MachineInstr &MI = *NextI;
+    bool ZeroData;
+    int64_t Size, Offset;
+    // Collect instructions that update memory tags with a FrameIndex operand
+    // and (when applicable) constant size, and whose output registers are dead
+    // (the latter is almost always the case in practice). Since these
+    // instructions effectively have no inputs or outputs, we are free to skip
+    // any non-aliasing instructions in between without tracking used registers.
+    if (isMergeableStackTaggingInstruction(MI, Offset, Size, ZeroData)) {
+      if (ZeroData != FirstZeroData)
+        break;
+      Instrs.emplace_back(&MI, Offset, Size);
+      continue;
+    }
+
+    // Only count non-transient, non-tagging instructions toward the scan
+    // limit.
+    if (!MI.isTransient())
+      ++Count;
+
+    // Just in case, stop before the epilogue code starts.
+    if (MI.getFlag(MachineInstr::FrameSetup) ||
+        MI.getFlag(MachineInstr::FrameDestroy))
+      break;
+
+    // Reject anything that may alias the collected instructions.
+    if (MI.mayLoadOrStore() || MI.hasUnmodeledSideEffects())
+      break;
+  }
+
+  // New code will be inserted after the last tagging instruction we've found.
+  MachineBasicBlock::iterator InsertI = Instrs[Instrs.size() - 1].MI;
+  InsertI++;
+
+  llvm::stable_sort(Instrs,
+                    [](const TagStoreInstr &Left, const TagStoreInstr &Right) {
+                      return Left.Offset < Right.Offset;
+                    });
+
+  // Make sure that we don't have any overlapping stores.
+  int64_t CurOffset = Instrs[0].Offset;
+  for (auto &Instr : Instrs) {
+    if (CurOffset > Instr.Offset)
+      return NextI;
+    CurOffset = Instr.Offset + Instr.Size;
+  }
+
+  // Find contiguous runs of tagged memory and emit shorter instruction
+  // sequencies for them when possible.
+  TagStoreEdit TSE(MBB, FirstZeroData);
+  Optional<int64_t> EndOffset;
+  for (auto &Instr : Instrs) {
+    if (EndOffset && *EndOffset != Instr.Offset) {
+      // Found a gap.
+      TSE.EmitCode(InsertI, TFI, /*IsLast = */ false);
+      TSE.Clear();
+    }
+
+    TSE.AddInstruction(Instr);
+    EndOffset = Instr.Offset + Instr.Size;
+  }
+
+  TSE.EmitCode(InsertI, TFI, /*IsLast = */ true);
+
+  return InsertI;
+}
+} // namespace
+
+void AArch64FrameLowering::processFunctionBeforeFrameIndicesReplaced(
+    MachineFunction &MF, RegScavenger *RS = nullptr) const {
+  if (StackTaggingMergeSetTag)
+    for (auto &BB : MF)
+      for (MachineBasicBlock::iterator II = BB.begin(); II != BB.end();)
+        II = tryMergeAdjacentSTG(II, this, RS);
+}
+
+/// For Win64 AArch64 EH, the offset to the Unwind object is from the SP
+/// before the update.  This is easily retrieved as it is exactly the offset
+/// that is set in processFunctionBeforeFrameFinalized.
 int AArch64FrameLowering::getFrameIndexReferencePreferSP(
     const MachineFunction &MF, int FI, unsigned &FrameReg,
     bool IgnoreSPUpdates) const {
   const MachineFrameInfo &MFI = MF.getFrameInfo();
   LLVM_DEBUG(dbgs() << "Offset from the SP for " << FI << " is "
                     << MFI.getObjectOffset(FI) << "\n");
   FrameReg = AArch64::SP;
   return MFI.getObjectOffset(FI);

llvm/lib/Target/AArch64/AArch64InstrInfo.cpp

int llvm::isAArch64FrameOffsetLegal(const MachineInstr &MI,
   case AArch64::ST1Twov2d:
   case AArch64::ST1Threev2d:
   case AArch64::ST1Fourv2d:
   case AArch64::ST1Twov1d:
   case AArch64::ST1Threev1d:
   case AArch64::ST1Fourv1d:
   case AArch64::IRG:
   case AArch64::IRGstack:
+  case AArch64::STGloop:
+  case AArch64::STZGloop:
     return AArch64FrameOffsetCannotUpdate;
   }
 
   // Get the min/max offset and the scale.
   unsigned Scale, Width;
   int64_t MinOff, MaxOff;
   if (!AArch64InstrInfo::getMemOpInfo(MI.getOpcode(), Scale, Width, MinOff,
                                       MaxOff))

llvm/lib/Target/AArch64/AArch64InstrInfo.td

: Pseudo<(outs GPR64sp:$Rd), (ins GPR64sp:$Rn, uimm6s16:$imm6, GPR64sp:$Rm, imm0_15:$imm4), []>,
       Sched<[]>;
 
 // Explicit SP in the first operand prevents ShrinkWrap optimization
 // from leaving this instruction out of the stack frame. When IRGstack
 // is transformed into IRG, this operand is replaced with the actual
 // register / expression for the tagged base pointer of the current function.
 def : Pat<(int_aarch64_irg_sp i64:$Rm), (IRGstack SP, i64:$Rm)>;
 
 // Large STG to be expanded into a loop. $Rm is the size, $Rn is start address.
 // $Rn_wback is one past the end of the range.
 let isCodeGenOnly=1, mayStore=1 in {
 def STGloop
-    : Pseudo<(outs GPR64common:$Rm_wback, GPR64sp:$Rn_wback), (ins GPR64common:$Rm, GPR64sp:$Rn),
-             [], "$Rn = $Rn_wback,@earlyclobber $Rn_wback,$Rm = $Rm_wback,@earlyclobber $Rm_wback" >,
+    : Pseudo<(outs GPR64common:$Rm, GPR64sp:$Rn_wback), (ins i64imm:$sz, GPR64sp:$Rn),
+             [], "$Rn = $Rn_wback,@earlyclobber $Rn_wback,@earlyclobber $Rm" >,
       Sched<[WriteAdr, WriteST]>;
 
 def STZGloop
-    : Pseudo<(outs GPR64common:$Rm_wback, GPR64sp:$Rn_wback), (ins GPR64common:$Rm, GPR64sp:$Rn),
-             [], "$Rn = $Rn_wback,@earlyclobber $Rn_wback,$Rm = $Rm_wback,@earlyclobber $Rm_wback" >,
+    : Pseudo<(outs GPR64common:$Rm, GPR64sp:$Rn_wback), (ins i64imm:$sz, GPR64sp:$Rn),
+             [], "$Rn = $Rn_wback,@earlyclobber $Rn_wback,@earlyclobber $Rm" >,
       Sched<[WriteAdr, WriteST]>;
 }
 
 } // Predicates = [HasMTE]
 
 //===----------------------------------------------------------------------===//
 // Logical instructions.
 //===----------------------------------------------------------------------===//

llvm/lib/Target/AArch64/AArch64RegisterInfo.h

unsigned getRegPressureLimit(const TargetRegisterClass *RC,
                                MachineFunction &MF) const override;
 
   bool trackLivenessAfterRegAlloc(const MachineFunction&) const override {
     return true;
   }
 
   unsigned getLocalAddressRegister(const MachineFunction &MF) const;
 };
-
 } // end namespace llvm
 
 #endif

llvm/lib/Target/AArch64/AArch64RegisterInfo.cpp

bool AArch64RegisterInfo::needsFrameBaseReg(MachineInstr *MI,
   // If we can reference via the stack pointer or base pointer, try that.
   // FIXME: This (and the code that resolves the references) can be improved
   //        to only disallow SP relative references in the live range of
   //        the VLA(s). In practice, it's unclear how much difference that
   //        would make, but it may be worth doing.
   if (isFrameOffsetLegal(MI, AArch64::SP, Offset))
     return false;
 
+  // If even offset 0 is illegal, we don't want a virtual base register.
+  if (!isFrameOffsetLegal(MI, AArch64::SP, 0))
+    return false;
+
   // The offset likely isn't legal; we want to allocate a virtual base register.
   return true;
 }
 
 bool AArch64RegisterInfo::isFrameOffsetLegal(const MachineInstr *MI,
                                              unsigned BaseReg,
                                              int64_t Offset) const {
   assert(MI && "Unable to get the legal offset for nil instruction.");
void AArch64RegisterInfo::resolveFrameIndex(MachineInstr &MI, unsigned BaseReg,
   const MachineFunction *MF = MI.getParent()->getParent();
   const AArch64InstrInfo *TII =
       MF->getSubtarget<AArch64Subtarget>().getInstrInfo();
   bool Done = rewriteAArch64FrameIndex(MI, i, BaseReg, Off, TII);
   assert(Done && "Unable to resolve frame index!");
   (void)Done;
 }
 
+
+static Register getScratchRegisterForInstruction(MachineInstr &MI) {
+  // ST*Gloop can only have #fi in op3, and they have a constraint that op1==op3.
+  // Use op1 as a scratch register.
+  if (MI.getOpcode() == AArch64::STGloop || MI.getOpcode() == AArch64::STZGloop)
+    return MI.getOperand(1).getReg();
+  else
+    return MI.getMF()->getRegInfo().createVirtualRegister(
+        &AArch64::GPR64RegClass);
+}
+
 void AArch64RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
                                               int SPAdj, unsigned FIOperandNum,
                                               RegScavenger *RS) const {
   assert(SPAdj == 0 && "Unexpected");
 
   MachineInstr &MI = *II;
   MachineBasicBlock &MBB = *MI.getParent();
   MachineFunction &MF = *MBB.getParent();
StackOffset SPOffset = {
         MFI.getObjectOffset(FrameIndex) + (int64_t)MFI.getStackSize(), MVT::i8};
     if (MFI.hasVarSizedObjects() ||
         isAArch64FrameOffsetLegal(MI, SPOffset, nullptr, nullptr, nullptr) !=
             (AArch64FrameOffsetCanUpdate | AArch64FrameOffsetIsLegal)) {
       // Can't update to SP + offset in place. Precalculate the tagged pointer
       // in a scratch register.
       Offset = TFI->resolveFrameIndexReference(
           MF, FrameIndex, FrameReg, /*PreferFP=*/false, /*ForSimm=*/true);
-      Register ScratchReg =
-          MF.getRegInfo().createVirtualRegister(&AArch64::GPR64RegClass);
+      Register ScratchReg = getScratchRegisterForInstruction(MI);
       emitFrameOffset(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg, Offset,
                       TII);
       BuildMI(MBB, MI, MI.getDebugLoc(), TII->get(AArch64::LDG), ScratchReg)
           .addReg(ScratchReg)
           .addReg(ScratchReg)
           .addImm(0);
       MI.getOperand(FIOperandNum)
           .ChangeToRegister(ScratchReg, false, false, true);
if (rewriteAArch64FrameIndex(MI, FIOperandNum, FrameReg, Offset, TII))
     return;
 
   assert((!RS || !RS->isScavengingFrameIndex(FrameIndex)) &&
          "Emergency spill slot is out of reach");
 
   // If we get here, the immediate doesn't fit into the instruction.  We folded
   // as much as possible above.  Handle the rest, providing a register that is
   // SP+LargeImm.
-  Register ScratchReg =
-      MF.getRegInfo().createVirtualRegister(&AArch64::GPR64RegClass);
+  Register ScratchReg = getScratchRegisterForInstruction(MI);
   emitFrameOffset(MBB, II, MI.getDebugLoc(), ScratchReg, FrameReg, Offset, TII);
   MI.getOperand(FIOperandNum).ChangeToRegister(ScratchReg, false, false, true);
 }
 
 unsigned AArch64RegisterInfo::getRegPressureLimit(const TargetRegisterClass *RC,
                                                   MachineFunction &MF) const {
   const AArch64FrameLowering *TFI = getFrameLowering(MF);
 

llvm/lib/Target/AArch64/AArch64SelectionDAGInfo.cpp

MachineMemOperand *BaseMemOperand = MF.getMachineMemOperand(
       DstPtrInfo, MachineMemOperand::MOStore, ObjSize, 16);
 
   bool UseSetTagRangeLoop =
       kSetTagLoopThreshold >= 0 && (int)ObjSize >= kSetTagLoopThreshold;
   if (!UseSetTagRangeLoop)
     return EmitUnrolledSetTag(DAG, dl, Chain, Addr, ObjSize, BaseMemOperand,
                               ZeroData);
 
-  if (ObjSize % 32 != 0) {
-    SDNode *St1 = DAG.getMachineNode(
-        ZeroData ? AArch64::STZGPostIndex : AArch64::STGPostIndex, dl,
-        {MVT::i64, MVT::Other},
-        {Addr, Addr, DAG.getTargetConstant(1, dl, MVT::i64), Chain});
-    DAG.setNodeMemRefs(cast<MachineSDNode>(St1), {BaseMemOperand});
-    ObjSize -= 16;
-    Addr = SDValue(St1, 0);
-    Chain = SDValue(St1, 1);
-  }
-
   const EVT ResTys[] = {MVT::i64, MVT::i64, MVT::Other};
-  SDValue Ops[] = {DAG.getConstant(ObjSize, dl, MVT::i64), Addr, Chain};
+
+  if (Addr.getOpcode() == ISD::FrameIndex) {
+    int FI = cast<FrameIndexSDNode>(Addr)->getIndex();
+    Addr = DAG.getTargetFrameIndex(FI, MVT::i64);
+  }
+  SDValue Ops[] = {DAG.getTargetConstant(ObjSize, dl, MVT::i64), Addr, Chain};
   SDNode *St = DAG.getMachineNode(
       ZeroData ? AArch64::STZGloop : AArch64::STGloop, dl, ResTys, Ops);
 
   DAG.setNodeMemRefs(cast<MachineSDNode>(St), {BaseMemOperand});
   return SDValue(St, 2);
 }

llvm/test/CodeGen/AArch64/settag-merge.ll

+; RUN: llc < %s -mtriple=aarch64 -mattr=+mte | FileCheck %s
+
+declare void @use(i8* %p)
+declare void @llvm.aarch64.settag(i8* %p, i64 %a)
+declare void @llvm.aarch64.settag.zero(i8* %p, i64 %a)
+
+define void @stg16_16() {
+entry:
+; CHECK-LABEL: stg16_16:
+; CHECK: st2g sp, [sp], #32
+; CHECK: ret
+  %a = alloca i8, i32 16, align 16
+  %b = alloca i8, i32 16, align 16
+  call void @llvm.aarch64.settag(i8* %a, i64 16)
+  call void @llvm.aarch64.settag(i8* %b, i64 16)
+  ret void
+}
+
+define i32 @stg16_16_16_16_ret() {
+entry:
+; CHECK-LABEL: stg16_16_16_16_ret:
+; CHECK: st2g sp, [sp, #32]
+; CHECK: st2g sp, [sp], #64
+; CHECK: mov  w0, wzr
+; CHECK: ret
+  %a = alloca i8, i32 16, align 16
+  %b = alloca i8, i32 16, align 16
+  %c = alloca i8, i32 16, align 16
+  %d = alloca i8, i32 16, align 16
+  call void @llvm.aarch64.settag(i8* %a, i64 16)
+  call void @llvm.aarch64.settag(i8* %b, i64 16)
+  call void @llvm.aarch64.settag(i8* %c, i64 16)
+  call void @llvm.aarch64.settag(i8* %d, i64 16)
+  ret i32 0
+}
+
+define void @stg16_16_16_16() {
+entry:
+; CHECK-LABEL: stg16_16_16_16:
+; CHECK: st2g sp, [sp, #32]
+; CHECK: st2g sp, [sp], #64
+; CHECK: ret
+  %a = alloca i8, i32 16, align 16
+  %b = alloca i8, i32 16, align 16
+  %c = alloca i8, i32 16, align 16
+  %d = alloca i8, i32 16, align 16
+  call void @llvm.aarch64.settag(i8* %a, i64 16)
+  call void @llvm.aarch64.settag(i8* %b, i64 16)
+  call void @llvm.aarch64.settag(i8* %c, i64 16)
+  call void @llvm.aarch64.settag(i8* %d, i64 16)
+  ret void
+}
+
+define void @stg128_128_128_128() {
+entry:
+; CHECK-LABEL: stg128_128_128_128:
+; CHECK: mov     x8, #512
+; CHECK: st2g    sp, [sp], #32
+; CHECK: sub     x8, x8, #32
+; CHECK: cbnz    x8,
+; CHECK: ret
+  %a = alloca i8, i32 128, align 16
+  %b = alloca i8, i32 128, align 16
+  %c = alloca i8, i32 128, align 16
+  %d = alloca i8, i32 128, align 16
+  call void @llvm.aarch64.settag(i8* %a, i64 128)
+  call void @llvm.aarch64.settag(i8* %b, i64 128)
+  call void @llvm.aarch64.settag(i8* %c, i64 128)
+  call void @llvm.aarch64.settag(i8* %d, i64 128)
+  ret void
+}
+
+define void @stg16_512_16() {
+entry:
+; CHECK-LABEL: stg16_512_16:
+; CHECK: mov     x8, #544
+; CHECK: st2g    sp, [sp], #32
+; CHECK: sub     x8, x8, #32
+; CHECK: cbnz    x8,
+; CHECK: ret
+  %a = alloca i8, i32 16, align 16
+  %b = alloca i8, i32 512, align 16
+  %c = alloca i8, i32 16, align 16
+  call void @llvm.aarch64.settag(i8* %a, i64 16)
+  call void @llvm.aarch64.settag(i8* %b, i64 512)
+  call void @llvm.aarch64.settag(i8* %c, i64 16)
+  ret void
+}
+
+define void @stg512_512_512() {
+entry:
+; CHECK-LABEL: stg512_512_512:
+; CHECK: mov     x8, #1536
+; CHECK: st2g    sp, [sp], #32
+; CHECK: sub     x8, x8, #32
+; CHECK: cbnz    x8,
+; CHECK: ret
+  %a = alloca i8, i32 512, align 16
+  %b = alloca i8, i32 512, align 16
+  %c = alloca i8, i32 512, align 16
+  call void @llvm.aarch64.settag(i8* %a, i64 512)
+  call void @llvm.aarch64.settag(i8* %b, i64 512)
+  call void @llvm.aarch64.settag(i8* %c, i64 512)
+  ret void
+}
+
+define void @early(i1 %flag) {
+entry:
+; CHECK-LABEL: early:
+; CHECK: tbz     w0, #0, [[LABEL:.LBB.*]]
+; CHECK: st2g    sp, [sp, #
+; CHECK: st2g    sp, [sp, #
+; CHECK: st2g    sp, [sp, #
+; CHECK: [[LABEL]]:
+; CHECK: stg     sp, [sp, #
+; CHECK: st2g    sp, [sp], #
+; CHECK: ret
+  %a = alloca i8, i32 48, align 16
+  %b = alloca i8, i32 48, align 16
+  %c = alloca i8, i32 48, align 16
+  br i1 %flag, label %if.then, label %if.end
+
+if.then:
+  call void @llvm.aarch64.settag(i8* %a, i64 48)
+  call void @llvm.aarch64.settag(i8* %b, i64 48)
+  br label %if.end
+
+if.end:
+  call void @llvm.aarch64.settag(i8* %c, i64 48)
+  ret void
+}
+
+define void @early_128_128(i1 %flag) {
+entry:
+; CHECK-LABEL: early_128_128:
+; CHECK: tbz   w0, #0, [[LABEL:.LBB.*]]
+; CHECK: add   x9, sp, #
+; CHECK: mov   x8, #256
+; CHECK: st2g  x9, [x9], #32
+; CHECK: sub   x8, x8, #32
+; CHECK: cbnz  x8,
+; CHECK: [[LABEL]]:
+; CHECK: stg     sp, [sp, #
+; CHECK: st2g    sp, [sp], #
+; CHECK: ret
+  %a = alloca i8, i32 128, align 16
+  %b = alloca i8, i32 128, align 16
+  %c = alloca i8, i32 48, align 16
+  br i1 %flag, label %if.then, label %if.end
+
+if.then:
+  call void @llvm.aarch64.settag(i8* %a, i64 128)
+  call void @llvm.aarch64.settag(i8* %b, i64 128)
+  br label %if.end
+
+if.end:
+  call void @llvm.aarch64.settag(i8* %c, i64 48)
+  ret void
+}
+
+define void @early_512_512(i1 %flag) {
+entry:
+; CHECK-LABEL: early_512_512:
+; CHECK: tbz   w0, #0, [[LABEL:.LBB.*]]
+; CHECK: add   x9, sp, #
+; CHECK: mov   x8, #1024
+; CHECK: st2g  x9, [x9], #32
+; CHECK: sub   x8, x8, #32
+; CHECK: cbnz  x8,
+; CHECK: [[LABEL]]:
+; CHECK: stg     sp, [sp, #
+; CHECK: st2g    sp, [sp], #
+; CHECK: ret
+  %a = alloca i8, i32 512, align 16
+  %b = alloca i8, i32 512, align 16
+  %c = alloca i8, i32 48, align 16
+  br i1 %flag, label %if.then, label %if.end
+
+if.then:
+  call void @llvm.aarch64.settag(i8* %a, i64 512)
+  call void @llvm.aarch64.settag(i8* %b, i64 512)
+  br label %if.end
+
+if.end:
+  call void @llvm.aarch64.settag(i8* %c, i64 48)
+  ret void
+}
+
+; Two loops of size 256; the second loop updates SP.
+define void @stg128_128_gap_128_128() {
+entry:
+; CHECK-LABEL: stg128_128_gap_128_128:
+; CHECK: mov     x9, sp
+; CHECK: mov     x8, #256
+; CHECK: st2g    x9, [x9], #32
+; CHECK: sub     x8, x8, #32
+; CHECK: cbnz    x8,
+; CHECK: mov     x8, #256
+; CHECK: st2g    sp, [sp], #32
+; CHECK: sub     x8, x8, #32
+; CHECK: cbnz    x8,
+; CHECK: ret
+  %a = alloca i8, i32 128, align 16
+  %a2 = alloca i8, i32 128, align 16
+  %b = alloca i8, i32 32, align 16
+  %c = alloca i8, i32 128, align 16
+  %c2 = alloca i8, i32 128, align 16
+  call void @use(i8* %b)
+  call void @llvm.aarch64.settag(i8* %a, i64 128)
+  call void @llvm.aarch64.settag(i8* %a2, i64 128)
+  call void @llvm.aarch64.settag(i8* %c, i64 128)
+  call void @llvm.aarch64.settag(i8* %c2, i64 128)
+  ret void
+}

llvm/test/CodeGen/AArch64/settag-merge.mir

+# RUN: llc -mtriple=aarch64 -mattr=+mte -run-pass=prologepilog %s -o - | FileCheck %s
+
+--- |
+  declare void @llvm.aarch64.settag(i8* nocapture writeonly, i64) argmemonly nounwind writeonly "target-features"="+mte"
+  define i32 @stg16_16_16_16_ret() "target-features"="+mte" {
+  entry:
+    %a = alloca i8, i32 16, align 16
+    %b = alloca i8, i32 16, align 16
+    %c = alloca i8, i32 16, align 16
+    %d = alloca i8, i32 16, align 16
+    call void @llvm.aarch64.settag(i8* %a, i64 16)
+    call void @llvm.aarch64.settag(i8* %b, i64 16)
+    call void @llvm.aarch64.settag(i8* %c, i64 16)
+    call void @llvm.aarch64.settag(i8* %d, i64 16)
+    ret i32 0
+  }
+
+  define void @stg16_store_128() "target-features"="+mte" {
+  entry:
+    %a = alloca i8, i32 16, align 16
+    %b = alloca i8, i32 128, align 16
+    call void @llvm.aarch64.settag(i8* %a, i64 16)
+    store i8 42, i8* %a
+    call void @llvm.aarch64.settag(i8* %b, i64 128)
+    ret void
+  }
+
+...
+---
+# A sequence of STG with a register copy in the middle.
+# Can be merged into ST2G + ST2G.
+# CHECK-LABEL: name:{{.*}}stg16_16_16_16_ret
+# CHECK-DAG: ST2GOffset $sp, $sp, 2
+# CHECK-DAG: ST2GOffset $sp, $sp, 0
+# CHECK-DAG: $w0 = COPY $wzr
+# CHECK-DAG: RET_ReallyLR implicit killed $w0
+
+name:            stg16_16_16_16_ret
+tracksRegLiveness: true
+stack:
+  - { id: 0, name: a, size: 16, alignment: 16 }
+  - { id: 1, name: b, size: 16, alignment: 16 }
+  - { id: 2, name: c, size: 16, alignment: 16 }
+  - { id: 3, name: d, size: 16, alignment: 16 }
+body:             |
+  bb.0.entry:
+    STGOffset $sp, %stack.0.a, 0 :: (store 16 into %ir.a)
+    STGOffset $sp, %stack.1.b, 0 :: (store 16 into %ir.b)
+    STGOffset $sp, %stack.2.c, 0 :: (store 16 into %ir.c)
+    $w0 = COPY $wzr
+    STGOffset $sp, %stack.3.d, 0 :: (store 16 into %ir.d)
+    RET_ReallyLR implicit killed $w0
+
+...
+
+---
+# A store in the middle prevents merging.
+# CHECK-LABEL: name:{{.*}}stg16_store_128
+# CHECK: ST2GOffset $sp, $sp, 2
+# CHECK: ST2GOffset $sp, $sp, 4
+# CHECK: ST2GOffset $sp, $sp, 6
+# CHECK: STGOffset  $sp, $sp, 8
+# CHECK: STRBBui
+# CHECK: ST2GOffset $sp, $sp, 0 
+# CHECK: RET_ReallyLR
+
+name:            stg16_store_128
+tracksRegLiveness: true
+stack:
+  - { id: 0, name: a, size: 16, alignment: 16 }
+  - { id: 1, name: b, size: 128, alignment: 16 }
+body:             |
+  bb.0.entry:
+    STGOffset $sp, %stack.0.a, 0 :: (store 16 into %ir.a)
+    renamable $w8 = MOVi32imm 42
+    ST2GOffset $sp, %stack.1.b, 6 :: (store 32 into %ir.b + 96, align 16)
+    ST2GOffset $sp, %stack.1.b, 4 :: (store 32 into %ir.b + 64, align 16)
+    ST2GOffset $sp, %stack.1.b, 2 :: (store 32 into %ir.b + 32, align 16)
+    STRBBui killed renamable $w8, %stack.0.a, 0 :: (store 1 into %ir.a, align 16)
+    ST2GOffset $sp, %stack.1.b, 0 :: (store 32 into %ir.b, align 16)
+    RET_ReallyLR
+
+...

llvm/test/CodeGen/AArch64/settag.ll

 ; CHECK: ret
   call void @llvm.aarch64.settag(i8* %p, i64 256)
   ret void
 }
 
 define void @stg17(i8* %p) {
 entry:
 ; CHECK-LABEL: stg17:
-; CHECK: mov  {{(w|x)}}[[R:[0-9]+]], #256
 ; CHECK: stg x0, [x0], #16
+; CHECK: mov  {{(w|x)}}[[R:[0-9]+]], #256
 ; CHECK: st2g x0, [x0], #32
 ; CHECK: sub  x[[R]], x[[R]], #32
 ; CHECK: cbnz x[[R]],
 ; CHECK: ret
   call void @llvm.aarch64.settag(i8* %p, i64 272)
   ret void
 }
 
 define void @stzg3(i8* %p) {
 entry:
 ; CHECK-LABEL: stzg3:
 ; CHECK: stzg  x0, [x0, #32]
 ; CHECK: stz2g x0, [x0]
 ; CHECK: ret
   call void @llvm.aarch64.settag.zero(i8* %p, i64 48)
   ret void
 }
 
 define void @stzg17(i8* %p) {
 entry:
 ; CHECK-LABEL: stzg17:
-; CHECK: mov  {{w|x}}[[R:[0-9]+]], #256
 ; CHECK: stzg x0, [x0], #16
+; CHECK: mov  {{w|x}}[[R:[0-9]+]], #256
 ; CHECK: stz2g x0, [x0], #32
 ; CHECK: sub  x[[R]], x[[R]], #32
 ; CHECK: cbnz x[[R]],
 ; CHECK: ret
   call void @llvm.aarch64.settag.zero(i8* %p, i64 272)
   ret void
 }
 
 define void @stg_alloca1() {
 entry:
 ; CHECK-LABEL: stg_alloca1:
 ; CHECK: stg sp, [sp]
 ; CHECK: ret
   %a = alloca i8, i32 16, align 16
   call void @llvm.aarch64.settag(i8* %a, i64 16)
   ret void
 }
 
 define void @stg_alloca5() {
 entry:
 ; CHECK-LABEL: stg_alloca5:
-; CHECK: stg  sp, [sp, #64]
-; CHECK: st2g sp, [sp, #32]
-; CHECK: st2g sp, [sp]
-; CHECK: ret
+; CHECK:         st2g    sp, [sp, #32]
+; CHECK-NEXT:    stg     sp, [sp, #64]
+; CHECK-NEXT:    st2g    sp, [sp], #80
+; CHECK-NEXT:    ret
   %a = alloca i8, i32 80, align 16
   call void @llvm.aarch64.settag(i8* %a, i64 80)
   ret void
 }
 
 define void @stg_alloca17() {
 entry:
 ; CHECK-LABEL: stg_alloca17:
-; CHECK: mov [[P:x[0-9]+]], sp
-; CHECK: stg [[P]], {{\[}}[[P]]{{\]}}, #16
 ; CHECK: mov  {{w|x}}[[R:[0-9]+]], #256
-; CHECK: st2g [[P]], {{\[}}[[P]]{{\]}}, #32
+; CHECK: st2g sp, [sp], #32
 ; CHECK: sub  x[[R]], x[[R]], #32
 ; CHECK: cbnz x[[R]],
+; CHECK: stg sp, [sp], #16
 ; CHECK: ret
   %a = alloca i8, i32 272, align 16
   call void @llvm.aarch64.settag(i8* %a, i64 272)
   ret void
 }
 
 declare void @llvm.aarch64.settag(i8* %p, i64 %a)
 declare void @llvm.aarch64.settag.zero(i8* %p, i64 %a)

llvm/test/CodeGen/AArch64/stack-tagging-unchecked-ld-st.ll

 ; COMMON: bl use8
 
 ; NEVER:  ldrb [[A:w.*]], [x{{.*}}]
 ; NEVER:  ldrb [[B:w.*]], [x{{.*}}]
 
 ; DEFAULT:  ldrb [[A:w.*]], [x{{.*}}]
 ; DEFAULT:  ldrb [[B:w.*]], [x{{.*}}]
 
-; ALWAYS: ldg [[PA:x.*]], [x{{.*}}]
-; ALWAYS: ldrb [[B:w.*]], [sp]
-; ALWAYS: ldrb [[A:w.*]], {{\[}}[[PA]]{{\]}}
+; ALWAYS-DAG: ldg [[PA:x.*]], [x{{.*}}]
+; ALWAYS-DAG: ldrb [[B:w.*]], [sp]
+; ALWAYS-DAG: ldrb [[A:w.*]], {{\[}}[[PA]]{{\]}}
 
-; COMMON: add w0, [[B]], [[A]]
 ; COMMON: ret
 
 ; One of these allocas is closer to FP than to SP, and within 256 bytes
 ; of the former (see hardcoded limit in resolveFrameOffsetReference).
 ; It could be lowered to an FP-relative load, but not when doing an
 ; unchecked access to tagged memory!
 define i8 @FPOffset() "frame-pointer"="all" sanitize_memtag {
   %x = alloca [200 x i8], align 4