[GlobalISel] Implement merging of stores of truncates.

[aemerson]

Summary

This is a port of a combine which matches a pattern where a wide type scalar value is stored by several narrow stores. It folds it into a single store or a BSWAP and a store if the targets supports it.

Assuming little endian target:

 i8 *p = ...
 i32 val = ...
 p[0] = (val >> 0) & 0xFF;
 p[1] = (val >> 8) & 0xFF;
 p[2] = (val >> 16) & 0xFF;
 p[3] = (val >> 24) & 0xFF;
=>
 *((i32)p) = val;

On CTMark AArch64 -Os this results in a good amount of savings:

Program            before        after       diff
             SPASS 412792       412788       -0.0%
                kc 432528       432512       -0.0%
            lencod 430112       430096       -0.0%
  consumer-typeset 419156       419128       -0.0%
            bullet 475840       475752       -0.0%
        tramp3d-v4 367760       367628       -0.0%
          clamscan 383388       383204       -0.0%
    pairlocalalign 249764       249476       -0.1%
    7zip-benchmark 570100       568860       -0.2%
           sqlite3 287628       286920       -0.2%
Geomean difference                           -0.1%

[paquette]

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

83

Looks like this part isn't used in the change?

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

3869

isSimple?

3879

weird that there is no b) case

3886

comment seems incorrect?

3937

maybe not for this patch, but maybe we should have a specific register matcher?

[aemerson]

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

83

Yeah I initially used it but then changed it later. I'll commit this as a separate NFC change.

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

3937

Yeah that would be useful.

[arsenm]

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

76–80

Reverse the order of the fields?

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

3857

This goes up to 8 bytes (also I don't see why bother limiting it)

3923–3924

isLoadFoldBarrier? Also need to watch out for volatile/atomic

[aemerson]

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

76–80

Why?

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

3923–3924

isLoadFoldBarrier() would also trigger for stores so I think its simpler to explicitly check.

[arsenm]

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

76–80

They're currently ordered to maximize padding. I also don't think the NeedBSwap and NeedRotate would be the most important fields (also they should get default initializers)

[aemerson]

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

3923–3924

Actually I could use that and then explicitly check if we have a GStore. That's probably safer.

[paquette]

LGTM

llvm/include/llvm/CodeGen/GlobalISel/CombinerHelper.h

 
 struct ShiftOfShiftedLogic {
   MachineInstr *Logic;
   MachineInstr *Shift2;
   Register LogicNonShiftReg;
   uint64_t ValSum;
 };
 
+struct MergeTruncStoresInfo {
+  SmallVector<GStore *> FoundStores;
+  GStore *LowestIdxStore = nullptr;
+  Register WideSrcVal;
+  bool NeedBSwap = false;
+  bool NeedRotate = false;
+};
+
+
 using OperandBuildSteps =
     SmallVector<std::function<void(MachineInstrBuilder &)>, 4>;
 struct InstructionBuildSteps {
   unsigned Opcode = 0;          /// The opcode for the produced instruction.
   OperandBuildSteps OperandFns; /// Operands to be added to the instruction.
   InstructionBuildSteps() = default;
   InstructionBuildSteps(unsigned Opcode, const OperandBuildSteps &OperandFns)
       : Opcode(Opcode), OperandFns(OperandFns) {}
public:
   ///  sM val = a[0] | (a[1] << N) | (a[2] << 2N) | (a[3] << 3N) ...
   /// \endcode
   ///
   /// And check if the tree can be replaced with a M-bit load + possibly a
   /// bswap.
   bool matchLoadOrCombine(MachineInstr &MI,
                           std::function<void(MachineIRBuilder &)> &MatchInfo);
 
+  bool matchTruncStoreMerge(MachineInstr &MI, MergeTruncStoresInfo &MatchInfo);
+  void applyTruncStoreMerge(MachineInstr &MI, MergeTruncStoresInfo &MatchInfo);
+
   bool matchExtendThroughPhis(MachineInstr &MI, MachineInstr *&ExtMI);
   void applyExtendThroughPhis(MachineInstr &MI, MachineInstr *&ExtMI);
 
   bool matchExtractVecEltBuildVec(MachineInstr &MI, Register &Reg);
   void applyExtractVecEltBuildVec(MachineInstr &MI, Register &Reg);
 
   bool matchExtractAllEltsFromBuildVector(
       MachineInstr &MI,

llvm/include/llvm/Target/GlobalISel/Combine.td

def combine_insert_vec_elts_build_vector : GICombineRule<
   (apply [{ Helper.applyCombineInsertVecElts(*${root}, ${info}); }])>;
 
 def load_or_combine : GICombineRule<
   (defs root:$root, build_fn_matchinfo:$info),
   (match (wip_match_opcode G_OR):$root,
     [{ return Helper.matchLoadOrCombine(*${root}, ${info}); }]),
   (apply [{ Helper.applyBuildFn(*${root}, ${info}); }])>;
 
+
+def truncstore_merge_matcdata : GIDefMatchData<"MergeTruncStoresInfo">;
+def truncstore_merge : GICombineRule<
+  (defs root:$root, truncstore_merge_matcdata:$info),
+  (match (wip_match_opcode G_STORE):$root,
+   [{ return Helper.matchTruncStoreMerge(*${root}, ${info}); }]),
+  (apply [{ Helper.applyTruncStoreMerge(*${root}, ${info}); }])>;
+
 def extend_through_phis_matchdata: GIDefMatchData<"MachineInstr*">;
 def extend_through_phis : GICombineRule<
   (defs root:$root, extend_through_phis_matchdata:$matchinfo),
   (match (wip_match_opcode G_PHI):$root,
     [{ return Helper.matchExtendThroughPhis(*${root}, ${matchinfo}); }]),
   (apply [{ Helper.applyExtendThroughPhis(*${root}, ${matchinfo}); }])>;
 
 // Currently only the one combine above.
def all_combines : GICombineGroup<[trivial_combines, insert_vec_elt_combines,
     shl_ashr_to_sext_inreg, sext_inreg_of_load,
     width_reduction_combines, select_combines,
     known_bits_simplifications, ext_ext_fold,
     not_cmp_fold, opt_brcond_by_inverting_cond,
     unmerge_merge, fabs_fabs_fold, unmerge_cst, unmerge_dead_to_trunc,
     unmerge_zext_to_zext, merge_unmerge, trunc_ext_fold, trunc_shl,
     const_combines, xor_of_and_with_same_reg, ptr_add_with_zero,
     shift_immed_chain, shift_of_shifted_logic_chain, load_or_combine,
-    div_rem_to_divrem, funnel_shift_combines, form_bitfield_extract,
-    constant_fold]>;
+    truncstore_merge, div_rem_to_divrem, funnel_shift_combines,
+    form_bitfield_extract, constant_fold]>;
 
 // A combine group used to for prelegalizer combiners at -O0. The combines in
 // this group have been selected based on experiments to balance code size and
 // compile time performance.
 def optnone_combines : GICombineGroup<[trivial_combines,
     ptr_add_immed_chain, combines_for_extload,
     not_cmp_fold, opt_brcond_by_inverting_cond]>;

llvm/lib/CodeGen/GlobalISel/CombinerHelper.cpp

 #include "llvm/CodeGen/MachineDominators.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstr.h"
 #include "llvm/CodeGen/MachineMemOperand.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/CodeGen/TargetLowering.h"
 #include "llvm/CodeGen/TargetOpcodes.h"
+#include "llvm/IR/DataLayout.h"
+#include "llvm/Support/Casting.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Target/TargetMachine.h"
 #include <tuple>
 
 #define DEBUG_TYPE "gi-combiner"
 
 using namespace llvm;
 using namespace MIPatternMatch;
MatchInfo = [=](MachineIRBuilder &MIB) {
     Register LoadDst = NeedsBSwap ? MRI.cloneVirtualRegister(Dst) : Dst;
     MIB.buildLoad(LoadDst, Ptr, *NewMMO);
     if (NeedsBSwap)
       MIB.buildBSwap(Dst, LoadDst);
   };
   return true;
 }
 
+/// Check if the store \p Store is a truncstore that can be merged. That is,
+/// it's a store of a shifted value of \p SrcVal. If \p SrcVal is an empty
+/// Register then it does not need to match and SrcVal is set to the source
+/// value found.
+/// On match, returns the start byte offset of the \p SrcVal that is being
+/// stored.
+static Optional<int64_t> getTruncStoreByteOffset(GStore &Store, Register &SrcVal,
+                                                 MachineRegisterInfo &MRI) {
+  Register TruncVal;
+  if (!mi_match(Store.getValueReg(), MRI, m_GTrunc(m_Reg(TruncVal))))
+    return None;
+
+  // The shift amount must be a constant multiple of the narrow type.
+  // It is translated to the offset address in the wide source value "y".
+  //
+  // x = G_LSHR y, ShiftAmtC
+  // s8 z = G_TRUNC x
+  // store z, ...
+  Register FoundSrcVal;
+  int64_t ShiftAmt;
+  if (!mi_match(TruncVal, MRI,
+                m_any_of(m_GLShr(m_Reg(FoundSrcVal), m_ICst(ShiftAmt)),
+                         m_GAShr(m_Reg(FoundSrcVal), m_ICst(ShiftAmt))))) {
+    if (!SrcVal.isValid() || TruncVal == SrcVal) {
+      if (!SrcVal.isValid())
+        SrcVal = TruncVal;
+      return 0; // If it's the lowest index store.
+    }
+    return None;
+  }
+
+  unsigned NarrowBits = Store.getMMO().getMemoryType().getScalarSizeInBits();
+  if (ShiftAmt % NarrowBits!= 0)
+    return None;
+  const unsigned Offset = ShiftAmt / NarrowBits;
+
+  if (SrcVal.isValid() && FoundSrcVal != SrcVal)
+    return None;
+
+  if (!SrcVal.isValid())
+    SrcVal = FoundSrcVal;
+  else if (MRI.getType(SrcVal) != MRI.getType(FoundSrcVal))
+    return None;
+  return Offset;
+}
+
+/// Match a pattern where a wide type scalar value is stored by several narrow
+/// stores. Fold it into a single store or a BSWAP and a store if the targets
+/// supports it.
+///
+/// Assuming little endian target:
+///  i8 *p = ...
+///  i32 val = ...
+///  p[0] = (val >> 0) & 0xFF;
+///  p[1] = (val >> 8) & 0xFF;
+///  p[2] = (val >> 16) & 0xFF;
+///  p[3] = (val >> 24) & 0xFF;
+/// =>
+///  *((i32)p) = val;
+///
+///  i8 *p = ...
+///  i32 val = ...
+///  p[0] = (val >> 24) & 0xFF;
+///  p[1] = (val >> 16) & 0xFF;
+///  p[2] = (val >> 8) & 0xFF;
+///  p[3] = (val >> 0) & 0xFF;
+/// =>
+///  *((i32)p) = BSWAP(val);
+bool CombinerHelper::matchTruncStoreMerge(MachineInstr &MI,
+                                          MergeTruncStoresInfo &MatchInfo) {
+  auto &StoreMI = cast<GStore>(MI);
+  LLT MemTy = StoreMI.getMMO().getMemoryType();
+
+  // We only handle merging simple stores of 1-4 bytes.
+  if (!MemTy.isScalar())
+    return false;
+  switch (MemTy.getSizeInBits()) {
+  case 8:
+  case 16:
+  case 32:
+    break;
+  default:
+    return false;
+  }
+  if (!StoreMI.isSimple())
+    return false;
+
+  // We do a simple search for mergeable stores prior to this one.
+  // Any potential alias hazard along the way terminates the search.
+  SmallVector<GStore *> FoundStores;
+
+  // We're looking for:
+  // 1) a (store(trunc(...)))
+  // 2) of an LSHR/ASHR of a single wide value, by the appropriate shift to get
+  //    the partial value stored.
+  // 3) where the offsets form either a little or big-endian sequence.
+
+  auto &LastStore = StoreMI;
+
+  // The single base pointer that all stores must use.
+  Register BaseReg;
+  int64_t LastOffset;
+  if (!mi_match(LastStore.getPointerReg(), MRI,
+                m_GPtrAdd(m_Reg(BaseReg), m_ICst(LastOffset)))) {
+    BaseReg = LastStore.getPointerReg();
+    LastOffset = 0;
+  }
+
+  GStore *LowestIdxStore = &LastStore;
+  int64_t LowestIdxOffset = LastOffset;
+
+  Register WideSrcVal;
+  auto LowestShiftAmt = getTruncStoreByteOffset(LastStore, WideSrcVal, MRI);
+  if (!LowestShiftAmt)
+    return false; // Didn't match a trunc.
+  assert(WideSrcVal.isValid());
+
+  LLT WideStoreTy = MRI.getType(WideSrcVal);
+  const unsigned NumStoresRequired =
+      WideStoreTy.getSizeInBits() / MemTy.getSizeInBits();
+
+  SmallVector<int64_t, 8> OffsetMap(NumStoresRequired, INT64_MAX);
+  OffsetMap[*LowestShiftAmt] = LastOffset;
+  FoundStores.emplace_back(&LastStore);
+
+  // Search the block up for more stores.
+  // We use a search threshold of 10 instructions here because the combiner
+  // works top-down within a block, and we don't want to search an unbounded
+  // number of predecessor instructions trying to find matching stores.
+  // If we moved this optimization into a separate pass then we could probably
+  // use a more efficient search without having a hard-coded threshold.
+  const int MaxInstsToCheck = 10;
+  int NumInstsChecked = 0;
+  for (auto II = ++LastStore.getReverseIterator();
+       II != LastStore.getParent()->rend() && NumInstsChecked < MaxInstsToCheck;
+       ++II) {
+    NumInstsChecked++;
+    GStore *NewStore;
+    if ((NewStore = dyn_cast<GStore>(&*II))) {
+      if (NewStore->getMMO().getMemoryType() != MemTy || !NewStore->isSimple())
+        break;
+    } else if (II->isLoadFoldBarrier() || II->mayLoad()) {
+      break;
+    } else {
+      continue; // This is a safe instruction we can look past.
+    }
+
+    Register NewBaseReg;
+    int64_t MemOffset;
+    // Check we're storing to the same base + some offset.
+    if (!mi_match(NewStore->getPointerReg(), MRI,
+                  m_GPtrAdd(m_Reg(NewBaseReg), m_ICst(MemOffset)))) {
+      NewBaseReg = NewStore->getPointerReg();
+      MemOffset = 0;
+    }
+    if (BaseReg != NewBaseReg)
+      break;
+
+    auto ShiftByteOffset = getTruncStoreByteOffset(*NewStore, WideSrcVal, MRI);
+    if (!ShiftByteOffset)
+      break;
+    if (MemOffset < LowestIdxOffset) {
+      LowestIdxOffset = MemOffset;
+      LowestIdxStore = NewStore;
+    }
+
+    // Map the offset in the store and the offset in the combined value, and
+    // early return if it has been set before.
+    if (*ShiftByteOffset < 0 || *ShiftByteOffset >= NumStoresRequired ||
+        OffsetMap[*ShiftByteOffset] != INT64_MAX)
+      break;
+    OffsetMap[*ShiftByteOffset] = MemOffset;
+
+    FoundStores.emplace_back(NewStore);
+    // Reset counter since we've found a matching inst.
+    NumInstsChecked = 0;
+    if (FoundStores.size() == NumStoresRequired)
+      break;
+  }
+
+  if (FoundStores.size() != NumStoresRequired) {
+    return false;
+  }
+
+  const auto &DL = LastStore.getMF()->getDataLayout();
+  auto &C = LastStore.getMF()->getFunction().getContext();
+  // Check that a store of the wide type is both allowed and fast on the target
+  bool Fast = false;
+  bool Allowed = getTargetLowering().allowsMemoryAccess(
+      C, DL, WideStoreTy, LowestIdxStore->getMMO(), &Fast);
+  if (!Allowed || !Fast)
+    return false;
+
+  // Check if the pieces of the value are going to the expected places in memory
+  // to merge the stores.
+  unsigned NarrowBits = MemTy.getScalarSizeInBits();
+  auto checkOffsets = [&](bool MatchLittleEndian) {
+    if (MatchLittleEndian) {
+      for (unsigned i = 0; i != NumStoresRequired; ++i)
+        if (OffsetMap[i] != i * (NarrowBits / 8) + LowestIdxOffset)
+          return false;
+    } else { // MatchBigEndian by reversing loop counter.
+      for (unsigned i = 0, j = NumStoresRequired - 1; i != NumStoresRequired;
+           ++i, --j)
+        if (OffsetMap[j] != i * (NarrowBits / 8) + LowestIdxOffset)
+          return false;
+    }
+    return true;
+  };
+
+  // Check if the offsets line up for the native data layout of this target.
+  bool NeedBswap = false;
+  bool NeedRotate = false;
+  if (!checkOffsets(DL.isLittleEndian())) {
+    // Special-case: check if byte offsets line up for the opposite endian.
+    if (NarrowBits == 8 && checkOffsets(DL.isBigEndian()))
+      NeedBswap = true;
+    else if (NumStoresRequired == 2 && checkOffsets(DL.isBigEndian()))
+      NeedRotate = true;
+    else
+      return false;
+  }
+
+  if (NeedBswap &&
+      !isLegalOrBeforeLegalizer({TargetOpcode::G_BSWAP, {WideStoreTy}}))
+    return false;
+  if (NeedRotate &&
+      !isLegalOrBeforeLegalizer({TargetOpcode::G_ROTR, {WideStoreTy}}))
+    return false;
+
+  MatchInfo.NeedBSwap = NeedBswap;
+  MatchInfo.NeedRotate = NeedRotate;
+  MatchInfo.LowestIdxStore = LowestIdxStore;
+  MatchInfo.WideSrcVal = WideSrcVal;
+  MatchInfo.FoundStores = std::move(FoundStores);
+  return true;
+}
+
+void CombinerHelper::applyTruncStoreMerge(MachineInstr &MI,
+                                          MergeTruncStoresInfo &MatchInfo) {
+
+  Builder.setInstrAndDebugLoc(MI);
+  Register WideSrcVal = MatchInfo.WideSrcVal;
+  LLT WideStoreTy = MRI.getType(WideSrcVal);
+
+  if (MatchInfo.NeedBSwap) {
+    WideSrcVal = Builder.buildBSwap(WideStoreTy, WideSrcVal).getReg(0);
+  } else if (MatchInfo.NeedRotate) {
+    assert(WideStoreTy.getSizeInBits() % 2 == 0 &&
+           "Unexpected type for rotate");
+    auto RotAmt =
+        Builder.buildConstant(WideStoreTy, WideStoreTy.getSizeInBits() / 2);
+    WideSrcVal =
+        Builder.buildRotateRight(WideStoreTy, WideSrcVal, RotAmt).getReg(0);
+  }
+
+  Builder.buildStore(WideSrcVal, MatchInfo.LowestIdxStore->getPointerReg(),
+                     MatchInfo.LowestIdxStore->getMMO().getPointerInfo(),
+                     MatchInfo.LowestIdxStore->getMMO().getAlign());
+
+  // Erase the old stores.
+  for (auto *ST : MatchInfo.FoundStores)
+    ST->eraseFromParent();
+}
+
 bool CombinerHelper::matchExtendThroughPhis(MachineInstr &MI,
                                             MachineInstr *&ExtMI) {
   assert(MI.getOpcode() == TargetOpcode::G_PHI);
 
   Register DstReg = MI.getOperand(0).getReg();
 
   // TODO: Extending a vector may be expensive, don't do this until heuristics
   // are better.

llvm/test/CodeGen/AArch64/GlobalISel/merge-stores-truncating.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=aarch64-apple-ios -global-isel -global-isel-abort=1 | FileCheck %s
+
+define dso_local void @trunc_i16_to_i8(i16 %x, i8* %p) {
+; CHECK-LABEL: trunc_i16_to_i8:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    strh w0, [x1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i16 %x to i8
+  %sh = lshr i16 %x, 8
+  %t2 = trunc i16 %sh to i8
+  store i8 %t1, i8* %p, align 1
+  %p1 = getelementptr inbounds i8, i8* %p, i64 1
+  store i8 %t2, i8* %p1, align 1
+  ret void
+}
+
+define dso_local void @trunc_i32_to_i8(i32 %x, i8* %p) {
+; CHECK-LABEL: trunc_i32_to_i8:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    str w0, [x1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i32 %x to i8
+  %sh1 = lshr i32 %x, 8
+  %t2 = trunc i32 %sh1 to i8
+  %sh2 = lshr i32 %x, 16
+  %t3 = trunc i32 %sh2 to i8
+  %sh3 = lshr i32 %x, 24
+  %t4 = trunc i32 %sh3 to i8
+  store i8 %t1, i8* %p, align 1
+  %p1 = getelementptr inbounds i8, i8* %p, i64 1
+  store i8 %t2, i8* %p1, align 1
+  %p2 = getelementptr inbounds i8, i8* %p, i64 2
+  store i8 %t3, i8* %p2, align 1
+  %p3 = getelementptr inbounds i8, i8* %p, i64 3
+  store i8 %t4, i8* %p3, align 1
+  ret void
+}
+
+define dso_local void @trunc_i32_to_i16(i32 %x, i16* %p) {
+; CHECK-LABEL: trunc_i32_to_i16:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    str w0, [x1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i32 %x to i16
+  %sh = lshr i32 %x, 16
+  %t2 = trunc i32 %sh to i16
+  store i16 %t1, i16* %p, align 2
+  %p1 = getelementptr inbounds i16, i16* %p, i64 1
+  store i16 %t2, i16* %p1, align 2
+  ret void
+}
+
+define dso_local void @be_i32_to_i16(i32 %x, i16* %p0) {
+; CHECK-LABEL: be_i32_to_i16:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    ror w8, w0, #16
+; CHECK-NEXT:    str w8, [x1]
+; CHECK-NEXT:    ret
+  %sh1 = lshr i32 %x, 16
+  %t0 = trunc i32 %x to i16
+  %t1 = trunc i32 %sh1 to i16
+  %p1 = getelementptr inbounds i16, i16* %p0, i64 1
+  store i16 %t0, i16* %p1, align 2
+  store i16 %t1, i16* %p0, align 2
+  ret void
+}
+
+define dso_local void @be_i32_to_i16_order(i32 %x, i16* %p0) {
+; CHECK-LABEL: be_i32_to_i16_order:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    ror w8, w0, #16
+; CHECK-NEXT:    str w8, [x1]
+; CHECK-NEXT:    ret
+  %sh1 = lshr i32 %x, 16
+  %t0 = trunc i32 %x to i16
+  %t1 = trunc i32 %sh1 to i16
+  %p1 = getelementptr inbounds i16, i16* %p0, i64 1
+  store i16 %t1, i16* %p0, align 2
+  store i16 %t0, i16* %p1, align 2
+  ret void
+}
+
+define dso_local void @trunc_i64_to_i8(i64 %x, i8* %p) {
+; CHECK-LABEL: trunc_i64_to_i8:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    str x0, [x1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i64 %x to i8
+  %sh1 = lshr i64 %x, 8
+  %t2 = trunc i64 %sh1 to i8
+  %sh2 = lshr i64 %x, 16
+  %t3 = trunc i64 %sh2 to i8
+  %sh3 = lshr i64 %x, 24
+  %t4 = trunc i64 %sh3 to i8
+  %sh4 = lshr i64 %x, 32
+  %t5 = trunc i64 %sh4 to i8
+  %sh5 = lshr i64 %x, 40
+  %t6 = trunc i64 %sh5 to i8
+  %sh6 = lshr i64 %x, 48
+  %t7 = trunc i64 %sh6 to i8
+  %sh7 = lshr i64 %x, 56
+  %t8 = trunc i64 %sh7 to i8
+  store i8 %t1, i8* %p, align 1
+  %p1 = getelementptr inbounds i8, i8* %p, i64 1
+  store i8 %t2, i8* %p1, align 1
+  %p2 = getelementptr inbounds i8, i8* %p, i64 2
+  store i8 %t3, i8* %p2, align 1
+  %p3 = getelementptr inbounds i8, i8* %p, i64 3
+  store i8 %t4, i8* %p3, align 1
+  %p4 = getelementptr inbounds i8, i8* %p, i64 4
+  store i8 %t5, i8* %p4, align 1
+  %p5 = getelementptr inbounds i8, i8* %p, i64 5
+  store i8 %t6, i8* %p5, align 1
+  %p6 = getelementptr inbounds i8, i8* %p, i64 6
+  store i8 %t7, i8* %p6, align 1
+  %p7 = getelementptr inbounds i8, i8* %p, i64 7
+  store i8 %t8, i8* %p7, align 1
+  ret void
+}
+
+define dso_local void @trunc_i64_to_i16(i64 %x, i16* %p) {
+; CHECK-LABEL: trunc_i64_to_i16:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    str x0, [x1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i64 %x to i16
+  %sh1 = lshr i64 %x, 16
+  %t2 = trunc i64 %sh1 to i16
+  %sh2 = lshr i64 %x, 32
+  %t3 = trunc i64 %sh2 to i16
+  %sh3 = lshr i64 %x, 48
+  %t4 = trunc i64 %sh3 to i16
+  store i16 %t1, i16* %p, align 2
+  %p1 = getelementptr inbounds i16, i16* %p, i64 1
+  store i16 %t2, i16* %p1, align 2
+  %p2 = getelementptr inbounds i16, i16* %p, i64 2
+  store i16 %t3, i16* %p2, align 2
+  %p3 = getelementptr inbounds i16, i16* %p, i64 3
+  store i16 %t4, i16* %p3, align 2
+  ret void
+}
+
+define dso_local void @trunc_i64_to_i32(i64 %x, i32* %p) {
+; CHECK-LABEL: trunc_i64_to_i32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    str x0, [x1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i64 %x to i32
+  %sh = lshr i64 %x, 32
+  %t2 = trunc i64 %sh to i32
+  store i32 %t1, i32* %p, align 4
+  %p1 = getelementptr inbounds i32, i32* %p, i64 1
+  store i32 %t2, i32* %p1, align 4
+  ret void
+}
+define dso_local void @be_i64_to_i32(i64 %x, i32* %p0) {
+; CHECK-LABEL: be_i64_to_i32:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    ror x8, x0, #32
+; CHECK-NEXT:    str x8, [x1]
+; CHECK-NEXT:    ret
+  %sh1 = lshr i64 %x, 32
+  %t0 = trunc i64 %x to i32
+  %t1 = trunc i64 %sh1 to i32
+  %p1 = getelementptr inbounds i32, i32* %p0, i64 1
+  store i32 %t0, i32* %p1, align 4
+  store i32 %t1, i32* %p0, align 4
+  ret void
+}
+
+define dso_local void @be_i64_to_i32_order(i64 %x, i32* %p0) {
+; CHECK-LABEL: be_i64_to_i32_order:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    ror x8, x0, #32
+; CHECK-NEXT:    str x8, [x1]
+; CHECK-NEXT:    ret
+  %sh1 = lshr i64 %x, 32
+  %t0 = trunc i64 %x to i32
+  %t1 = trunc i64 %sh1 to i32
+  %p1 = getelementptr inbounds i32, i32* %p0, i64 1
+  store i32 %t1, i32* %p0, align 4
+  store i32 %t0, i32* %p1, align 4
+  ret void
+}
+
+; Negative tests.
+
+define void @merge_hole(i32 %x, i8* %p) {
+; CHECK-LABEL: merge_hole:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    lsr w8, w0, #16
+; CHECK-NEXT:    strb w0, [x1]
+; CHECK-NEXT:    strh w8, [x1, #2]
+; CHECK-NEXT:    ret
+  %pcast = bitcast i8* %p to i16*
+  %p2 = getelementptr inbounds i16, i16* %pcast, i64 1
+  %x3 = trunc i32 %x to i8
+  store i8 %x3, i8* %p, align 1
+  %sh = lshr i32 %x, 16
+  %x01 = trunc i32 %sh to i16
+  store i16 %x01, i16* %p2, align 1
+  ret void
+}
+
+define void @merge_hole2(i32 %x, i8* %p) {
+; CHECK-LABEL: merge_hole2:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    lsr w8, w0, #16
+; CHECK-NEXT:    strh w8, [x1, #2]
+; CHECK-NEXT:    strb w0, [x1]
+; CHECK-NEXT:    ret
+  %pcast = bitcast i8* %p to i16*
+  %p2 = getelementptr inbounds i16, i16* %pcast, i64 1
+  %sh = lshr i32 %x, 16
+  %x01 = trunc i32 %sh to i16
+  store i16 %x01, i16* %p2, align 1
+  %x3 = trunc i32 %x to i8
+  store i8 %x3, i8* %p, align 1
+  ret void
+}
+
+define void @merge_hole3(i32 %x, i8* %p) {
+; CHECK-LABEL: merge_hole3:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    lsr w8, w0, #16
+; CHECK-NEXT:    strb w0, [x1, #1]
+; CHECK-NEXT:    strh w8, [x1, #2]
+; CHECK-NEXT:    ret
+  %p1 = getelementptr inbounds i8, i8* %p, i64 1
+  %pcast = bitcast i8* %p to i16*
+  %p2 = getelementptr inbounds i16, i16* %pcast, i64 1
+  %x3 = trunc i32 %x to i8
+  store i8 %x3, i8* %p1, align 1
+  %sh = lshr i32 %x, 16
+  %x01 = trunc i32 %sh to i16
+  store i16 %x01, i16* %p2, align 1
+  ret void
+}
+
+define void @merge_hole4(i32 %x, i8* %p) {
+; CHECK-LABEL: merge_hole4:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    lsr w8, w0, #16
+; CHECK-NEXT:    strb w0, [x1, #2]
+; CHECK-NEXT:    strh w8, [x1]
+; CHECK-NEXT:    ret
+  %pcast = bitcast i8* %p to i16*
+  %p2 = getelementptr inbounds i8, i8* %p, i64 2
+  %x3 = trunc i32 %x to i8
+  store i8 %x3, i8* %p2, align 1
+  %sh = lshr i32 %x, 16
+  %x01 = trunc i32 %sh to i16
+  store i16 %x01, i16* %pcast, align 1
+  ret void
+}
+
+define dso_local i32 @load_between_stores(i32 %x, i16* %p, i32 *%ptr) {
+; CHECK-LABEL: load_between_stores:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    strh w0, [x1]
+; CHECK-NEXT:    ldr w8, [x2]
+; CHECK-NEXT:    lsr w9, w0, #16
+; CHECK-NEXT:    strh w9, [x1, #2]
+; CHECK-NEXT:    mov w0, w8
+; CHECK-NEXT:    ret
+  %t1 = trunc i32 %x to i16
+  %sh = lshr i32 %x, 16
+  %t2 = trunc i32 %sh to i16
+  store i16 %t1, i16* %p, align 2
+  %ld = load i32, i32 *%ptr
+  %p1 = getelementptr inbounds i16, i16* %p, i64 1
+  store i16 %t2, i16* %p1, align 2
+  ret i32 %ld
+}
+
+define dso_local void @invalid_shift(i16 %x, i8* %p) {
+; CHECK-LABEL: invalid_shift:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    and w8, w0, #0xffff
+; CHECK-NEXT:    lsr w8, w8, #4
+; CHECK-NEXT:    strb w0, [x1]
+; CHECK-NEXT:    strb w8, [x1, #1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i16 %x to i8
+  %sh = lshr i16 %x, 4
+  %t2 = trunc i16 %sh to i8
+  store i8 %t1, i8* %p, align 1
+  %p1 = getelementptr inbounds i8, i8* %p, i64 1
+  store i8 %t2, i8* %p1, align 1
+  ret void
+}
+
+define dso_local void @missing_store(i32 %x, i8* %p) {
+; CHECK-LABEL: missing_store:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    lsr w8, w0, #8
+; CHECK-NEXT:    lsr w9, w0, #24
+; CHECK-NEXT:    strb w0, [x1]
+; CHECK-NEXT:    strb w8, [x1, #1]
+; CHECK-NEXT:    strb w9, [x1, #3]
+; CHECK-NEXT:    ret
+  %t1 = trunc i32 %x to i8
+  %sh1 = lshr i32 %x, 8
+  %t2 = trunc i32 %sh1 to i8
+  %sh3 = lshr i32 %x, 24
+  %t4 = trunc i32 %sh3 to i8
+  store i8 %t1, i8* %p, align 1
+  %p1 = getelementptr inbounds i8, i8* %p, i64 1
+  store i8 %t2, i8* %p1, align 1
+  %p3 = getelementptr inbounds i8, i8* %p, i64 3
+  store i8 %t4, i8* %p3, align 1
+  ret void
+}
+
+define dso_local void @different_base_reg(i16 %x, i8* %p, i8 *%p2) {
+; CHECK-LABEL: different_base_reg:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    and w8, w0, #0xffff
+; CHECK-NEXT:    lsr w8, w8, #8
+; CHECK-NEXT:    strb w0, [x1]
+; CHECK-NEXT:    strb w8, [x2, #1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i16 %x to i8
+  %sh = lshr i16 %x, 8
+  %t2 = trunc i16 %sh to i8
+  store i8 %t1, i8* %p, align 1
+  %p1 = getelementptr inbounds i8, i8* %p2, i64 1
+  store i8 %t2, i8* %p1, align 1
+  ret void
+}
+
+define dso_local void @second_store_is_volatile(i16 %x, i8* %p) {
+; CHECK-LABEL: second_store_is_volatile:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    and w8, w0, #0xffff
+; CHECK-NEXT:    lsr w8, w8, #8
+; CHECK-NEXT:    strb w0, [x1]
+; CHECK-NEXT:    strb w8, [x1, #1]
+; CHECK-NEXT:    ret
+  %t1 = trunc i16 %x to i8
+  %sh = lshr i16 %x, 8
+  %t2 = trunc i16 %sh to i8
+  store volatile i8 %t1, i8* %p, align 1
+  %p1 = getelementptr inbounds i8, i8* %p, i64 1
+  store i8 %t2, i8* %p1, align 1
+  ret void
+}

llvm/test/CodeGen/AArch64/GlobalISel/merge-stores-truncating.mir

+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple aarch64 -run-pass=aarch64-prelegalizer-combiner -verify-machineinstrs %s -o - | FileCheck %s
+---
+name:            trunc_i16_to_i8
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: trunc_i16_to_i8
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: G_STORE [[TRUNC]](s16), [[COPY1]](p0) :: (store (s16), align 1)
+    ; CHECK: RET_ReallyLR
+    %2:_(s32) = COPY $w0
+    %0:_(s16) = G_TRUNC %2(s32)
+    %1:_(p0) = COPY $x1
+    %4:_(s16) = G_CONSTANT i16 8
+    %3:_(s8) = G_TRUNC %0(s16)
+    %5:_(s16) = G_LSHR %0, %4(s16)
+    %6:_(s8) = G_TRUNC %5(s16)
+    G_STORE %3(s8), %1(p0) :: (store (s8))
+    %7:_(s64) = G_CONSTANT i64 1
+    %8:_(p0) = G_PTR_ADD %1, %7(s64)
+    G_STORE %6(s8), %8(p0) :: (store (s8))
+    RET_ReallyLR
+
+...
+---
+name:            trunc_i32_to_i8
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: trunc_i32_to_i8
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: G_STORE [[COPY]](s32), [[COPY1]](p0) :: (store (s32), align 1)
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %3:_(s32) = G_CONSTANT i32 8
+    %6:_(s32) = G_CONSTANT i32 16
+    %9:_(s32) = G_CONSTANT i32 24
+    %2:_(s8) = G_TRUNC %0(s32)
+    %4:_(s32) = G_LSHR %0, %3(s32)
+    %5:_(s8) = G_TRUNC %4(s32)
+    %7:_(s32) = G_LSHR %0, %6(s32)
+    %8:_(s8) = G_TRUNC %7(s32)
+    %10:_(s32) = G_LSHR %0, %9(s32)
+    %11:_(s8) = G_TRUNC %10(s32)
+    G_STORE %2(s8), %1(p0) :: (store (s8))
+    %12:_(s64) = G_CONSTANT i64 1
+    %13:_(p0) = G_PTR_ADD %1, %12(s64)
+    G_STORE %5(s8), %13(p0) :: (store (s8))
+    %14:_(s64) = G_CONSTANT i64 2
+    %15:_(p0) = G_PTR_ADD %1, %14(s64)
+    G_STORE %8(s8), %15(p0) :: (store (s8))
+    %16:_(s64) = G_CONSTANT i64 3
+    %17:_(p0) = G_PTR_ADD %1, %16(s64)
+    G_STORE %11(s8), %17(p0) :: (store (s8))
+    RET_ReallyLR
+
+...
+---
+name:            trunc_i32_to_i16
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: trunc_i32_to_i16
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: G_STORE [[COPY]](s32), [[COPY1]](p0) :: (store (s32), align 2)
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %3:_(s32) = G_CONSTANT i32 16
+    %2:_(s16) = G_TRUNC %0(s32)
+    %4:_(s32) = G_LSHR %0, %3(s32)
+    %5:_(s16) = G_TRUNC %4(s32)
+    G_STORE %2(s16), %1(p0) :: (store (s16))
+    %6:_(s64) = G_CONSTANT i64 2
+    %7:_(p0) = G_PTR_ADD %1, %6(s64)
+    G_STORE %5(s16), %7(p0) :: (store (s16))
+    RET_ReallyLR
+
+...
+---
+name:            be_i32_to_i16
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: be_i32_to_i16
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[ROTR:%[0-9]+]]:_(s32) = G_ROTR [[COPY]], [[C]](s32)
+    ; CHECK: G_STORE [[ROTR]](s32), [[COPY1]](p0) :: (store (s32), align 2)
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %2:_(s32) = G_CONSTANT i32 16
+    %3:_(s32) = G_LSHR %0, %2(s32)
+    %4:_(s16) = G_TRUNC %0(s32)
+    %5:_(s16) = G_TRUNC %3(s32)
+    %6:_(s64) = G_CONSTANT i64 2
+    %7:_(p0) = G_PTR_ADD %1, %6(s64)
+    G_STORE %4(s16), %7(p0) :: (store (s16))
+    G_STORE %5(s16), %1(p0) :: (store (s16))
+    RET_ReallyLR
+
+...
+---
+name:            be_i32_to_i16_order
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: be_i32_to_i16_order
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[ROTR:%[0-9]+]]:_(s32) = G_ROTR [[COPY]], [[C]](s32)
+    ; CHECK: G_STORE [[ROTR]](s32), [[COPY1]](p0) :: (store (s32), align 2)
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %2:_(s32) = G_CONSTANT i32 16
+    %3:_(s32) = G_LSHR %0, %2(s32)
+    %4:_(s16) = G_TRUNC %0(s32)
+    %5:_(s16) = G_TRUNC %3(s32)
+    %6:_(s64) = G_CONSTANT i64 2
+    %7:_(p0) = G_PTR_ADD %1, %6(s64)
+    G_STORE %5(s16), %1(p0) :: (store (s16))
+    G_STORE %4(s16), %7(p0) :: (store (s16))
+    RET_ReallyLR
+
+...
+---
+name:            trunc_i64_to_i8
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: trunc_i64_to_i8
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: G_STORE [[COPY]](s64), [[COPY1]](p0) :: (store (s64), align 1)
+    ; CHECK: RET_ReallyLR
+    %0:_(s64) = COPY $x0
+    %1:_(p0) = COPY $x1
+    %3:_(s64) = G_CONSTANT i64 8
+    %6:_(s64) = G_CONSTANT i64 16
+    %9:_(s64) = G_CONSTANT i64 24
+    %12:_(s64) = G_CONSTANT i64 32
+    %15:_(s64) = G_CONSTANT i64 40
+    %18:_(s64) = G_CONSTANT i64 48
+    %21:_(s64) = G_CONSTANT i64 56
+    %2:_(s8) = G_TRUNC %0(s64)
+    %4:_(s64) = G_LSHR %0, %3(s64)
+    %5:_(s8) = G_TRUNC %4(s64)
+    %7:_(s64) = G_LSHR %0, %6(s64)
+    %8:_(s8) = G_TRUNC %7(s64)
+    %10:_(s64) = G_LSHR %0, %9(s64)
+    %11:_(s8) = G_TRUNC %10(s64)
+    %13:_(s64) = G_LSHR %0, %12(s64)
+    %14:_(s8) = G_TRUNC %13(s64)
+    %16:_(s64) = G_LSHR %0, %15(s64)
+    %17:_(s8) = G_TRUNC %16(s64)
+    %19:_(s64) = G_LSHR %0, %18(s64)
+    %20:_(s8) = G_TRUNC %19(s64)
+    %22:_(s64) = G_LSHR %0, %21(s64)
+    %23:_(s8) = G_TRUNC %22(s64)
+    G_STORE %2(s8), %1(p0) :: (store (s8))
+    %24:_(s64) = G_CONSTANT i64 1
+    %25:_(p0) = G_PTR_ADD %1, %24(s64)
+    G_STORE %5(s8), %25(p0) :: (store (s8))
+    %26:_(s64) = G_CONSTANT i64 2
+    %27:_(p0) = G_PTR_ADD %1, %26(s64)
+    G_STORE %8(s8), %27(p0) :: (store (s8))
+    %28:_(s64) = G_CONSTANT i64 3
+    %29:_(p0) = G_PTR_ADD %1, %28(s64)
+    G_STORE %11(s8), %29(p0) :: (store (s8))
+    %30:_(s64) = G_CONSTANT i64 4
+    %31:_(p0) = G_PTR_ADD %1, %30(s64)
+    G_STORE %14(s8), %31(p0) :: (store (s8))
+    %32:_(s64) = G_CONSTANT i64 5
+    %33:_(p0) = G_PTR_ADD %1, %32(s64)
+    G_STORE %17(s8), %33(p0) :: (store (s8))
+    %34:_(s64) = G_CONSTANT i64 6
+    %35:_(p0) = G_PTR_ADD %1, %34(s64)
+    G_STORE %20(s8), %35(p0) :: (store (s8))
+    %36:_(s64) = G_CONSTANT i64 7
+    %37:_(p0) = G_PTR_ADD %1, %36(s64)
+    G_STORE %23(s8), %37(p0) :: (store (s8))
+    RET_ReallyLR
+
+...
+---
+name:            trunc_i64_to_i16
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: trunc_i64_to_i16
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: G_STORE [[COPY]](s64), [[COPY1]](p0) :: (store (s64), align 2)
+    ; CHECK: RET_ReallyLR
+    %0:_(s64) = COPY $x0
+    %1:_(p0) = COPY $x1
+    %3:_(s64) = G_CONSTANT i64 16
+    %6:_(s64) = G_CONSTANT i64 32
+    %9:_(s64) = G_CONSTANT i64 48
+    %2:_(s16) = G_TRUNC %0(s64)
+    %4:_(s64) = G_LSHR %0, %3(s64)
+    %5:_(s16) = G_TRUNC %4(s64)
+    %7:_(s64) = G_LSHR %0, %6(s64)
+    %8:_(s16) = G_TRUNC %7(s64)
+    %10:_(s64) = G_LSHR %0, %9(s64)
+    %11:_(s16) = G_TRUNC %10(s64)
+    G_STORE %2(s16), %1(p0) :: (store (s16))
+    %12:_(s64) = G_CONSTANT i64 2
+    %13:_(p0) = G_PTR_ADD %1, %12(s64)
+    G_STORE %5(s16), %13(p0) :: (store (s16))
+    %14:_(s64) = G_CONSTANT i64 4
+    %15:_(p0) = G_PTR_ADD %1, %14(s64)
+    G_STORE %8(s16), %15(p0) :: (store (s16))
+    %16:_(s64) = G_CONSTANT i64 6
+    %17:_(p0) = G_PTR_ADD %1, %16(s64)
+    G_STORE %11(s16), %17(p0) :: (store (s16))
+    RET_ReallyLR
+
+...
+---
+name:            trunc_i64_to_i32
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: trunc_i64_to_i32
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: G_STORE [[COPY]](s64), [[COPY1]](p0) :: (store (s64), align 4)
+    ; CHECK: RET_ReallyLR
+    %0:_(s64) = COPY $x0
+    %1:_(p0) = COPY $x1
+    %3:_(s64) = G_CONSTANT i64 32
+    %2:_(s32) = G_TRUNC %0(s64)
+    %4:_(s64) = G_LSHR %0, %3(s64)
+    %5:_(s32) = G_TRUNC %4(s64)
+    G_STORE %2(s32), %1(p0) :: (store (s32))
+    %6:_(s64) = G_CONSTANT i64 4
+    %7:_(p0) = G_PTR_ADD %1, %6(s64)
+    G_STORE %5(s32), %7(p0) :: (store (s32))
+    RET_ReallyLR
+
+...
+---
+name:            be_i64_to_i32
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: be_i64_to_i32
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; CHECK: [[ROTR:%[0-9]+]]:_(s64) = G_ROTR [[COPY]], [[C]](s64)
+    ; CHECK: G_STORE [[ROTR]](s64), [[COPY1]](p0) :: (store (s64), align 4)
+    ; CHECK: RET_ReallyLR
+    %0:_(s64) = COPY $x0
+    %1:_(p0) = COPY $x1
+    %2:_(s64) = G_CONSTANT i64 32
+    %3:_(s64) = G_LSHR %0, %2(s64)
+    %4:_(s32) = G_TRUNC %0(s64)
+    %5:_(s32) = G_TRUNC %3(s64)
+    %6:_(s64) = G_CONSTANT i64 4
+    %7:_(p0) = G_PTR_ADD %1, %6(s64)
+    G_STORE %4(s32), %7(p0) :: (store (s32))
+    G_STORE %5(s32), %1(p0) :: (store (s32))
+    RET_ReallyLR
+
+...
+---
+name:            be_i64_to_i32_order
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$x0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $x0, $x1
+
+    ; CHECK-LABEL: name: be_i64_to_i32_order
+    ; CHECK: liveins: $x0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY $x0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 32
+    ; CHECK: [[ROTR:%[0-9]+]]:_(s64) = G_ROTR [[COPY]], [[C]](s64)
+    ; CHECK: G_STORE [[ROTR]](s64), [[COPY1]](p0) :: (store (s64), align 4)
+    ; CHECK: RET_ReallyLR
+    %0:_(s64) = COPY $x0
+    %1:_(p0) = COPY $x1
+    %2:_(s64) = G_CONSTANT i64 32
+    %3:_(s64) = G_LSHR %0, %2(s64)
+    %4:_(s32) = G_TRUNC %0(s64)
+    %5:_(s32) = G_TRUNC %3(s64)
+    %6:_(s64) = G_CONSTANT i64 4
+    %7:_(p0) = G_PTR_ADD %1, %6(s64)
+    G_STORE %5(s32), %1(p0) :: (store (s32))
+    G_STORE %4(s32), %7(p0) :: (store (s32))
+    RET_ReallyLR
+
+...
+---
+name:            merge_hole
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: merge_hole
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C1]](s64)
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
+    ; CHECK: G_STORE [[TRUNC]](s8), [[COPY1]](p0) :: (store (s8))
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C]](s32)
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; CHECK: G_STORE [[TRUNC1]](s16), [[PTR_ADD]](p0) :: (store (s16), align 1)
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %5:_(s32) = G_CONSTANT i32 16
+    %2:_(s64) = G_CONSTANT i64 2
+    %3:_(p0) = G_PTR_ADD %1, %2(s64)
+    %4:_(s8) = G_TRUNC %0(s32)
+    G_STORE %4(s8), %1(p0) :: (store (s8))
+    %6:_(s32) = G_LSHR %0, %5(s32)
+    %7:_(s16) = G_TRUNC %6(s32)
+    G_STORE %7(s16), %3(p0) :: (store (s16), align 1)
+    RET_ReallyLR
+
+...
+---
+name:            merge_hole2
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: merge_hole2
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C1]](s64)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C]](s32)
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; CHECK: G_STORE [[TRUNC]](s16), [[PTR_ADD]](p0) :: (store (s16), align 1)
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
+    ; CHECK: G_STORE [[TRUNC1]](s8), [[COPY1]](p0) :: (store (s8))
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %4:_(s32) = G_CONSTANT i32 16
+    %2:_(s64) = G_CONSTANT i64 2
+    %3:_(p0) = G_PTR_ADD %1, %2(s64)
+    %5:_(s32) = G_LSHR %0, %4(s32)
+    %6:_(s16) = G_TRUNC %5(s32)
+    G_STORE %6(s16), %3(p0) :: (store (s16), align 1)
+    %7:_(s8) = G_TRUNC %0(s32)
+    G_STORE %7(s8), %1(p0) :: (store (s8))
+    RET_ReallyLR
+
+...
+---
+name:            merge_hole3
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: merge_hole3
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C1]](s64)
+    ; CHECK: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CHECK: [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C2]](s64)
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
+    ; CHECK: G_STORE [[TRUNC]](s8), [[PTR_ADD]](p0) :: (store (s8))
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C]](s32)
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; CHECK: G_STORE [[TRUNC1]](s16), [[PTR_ADD1]](p0) :: (store (s16), align 1)
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %7:_(s32) = G_CONSTANT i32 16
+    %2:_(s64) = G_CONSTANT i64 1
+    %3:_(p0) = G_PTR_ADD %1, %2(s64)
+    %4:_(s64) = G_CONSTANT i64 2
+    %5:_(p0) = G_PTR_ADD %1, %4(s64)
+    %6:_(s8) = G_TRUNC %0(s32)
+    G_STORE %6(s8), %3(p0) :: (store (s8))
+    %8:_(s32) = G_LSHR %0, %7(s32)
+    %9:_(s16) = G_TRUNC %8(s32)
+    G_STORE %9(s16), %5(p0) :: (store (s16), align 1)
+    RET_ReallyLR
+
+...
+---
+name:            merge_hole4
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: merge_hole4
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C1]](s64)
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
+    ; CHECK: G_STORE [[TRUNC]](s8), [[PTR_ADD]](p0) :: (store (s8))
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C]](s32)
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; CHECK: G_STORE [[TRUNC1]](s16), [[COPY1]](p0) :: (store (s16), align 1)
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %5:_(s32) = G_CONSTANT i32 16
+    %2:_(s64) = G_CONSTANT i64 2
+    %3:_(p0) = G_PTR_ADD %1, %2(s64)
+    %4:_(s8) = G_TRUNC %0(s32)
+    G_STORE %4(s8), %3(p0) :: (store (s8))
+    %6:_(s32) = G_LSHR %0, %5(s32)
+    %7:_(s16) = G_TRUNC %6(s32)
+    G_STORE %7(s16), %1(p0) :: (store (s16), align 1)
+    RET_ReallyLR
+
+...
+---
+name:            load_between_stores
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+  - { reg: '$x2' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1, $x2
+
+    ; CHECK-LABEL: name: load_between_stores
+    ; CHECK: liveins: $w0, $x1, $x2
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[COPY2:%[0-9]+]]:_(p0) = COPY $x2
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C]](s32)
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; CHECK: G_STORE [[TRUNC]](s16), [[COPY1]](p0) :: (store (s16))
+    ; CHECK: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY2]](p0) :: (load (s32))
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C1]](s64)
+    ; CHECK: G_STORE [[TRUNC1]](s16), [[PTR_ADD]](p0) :: (store (s16))
+    ; CHECK: $w0 = COPY [[LOAD]](s32)
+    ; CHECK: RET_ReallyLR implicit $w0
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %2:_(p0) = COPY $x2
+    %4:_(s32) = G_CONSTANT i32 16
+    %3:_(s16) = G_TRUNC %0(s32)
+    %5:_(s32) = G_LSHR %0, %4(s32)
+    %6:_(s16) = G_TRUNC %5(s32)
+    G_STORE %3(s16), %1(p0) :: (store (s16))
+    %7:_(s32) = G_LOAD %2(p0) :: (load (s32))
+    %8:_(s64) = G_CONSTANT i64 2
+    %9:_(p0) = G_PTR_ADD %1, %8(s64)
+    G_STORE %6(s16), %9(p0) :: (store (s16))
+    $w0 = COPY %7(s32)
+    RET_ReallyLR implicit $w0
+
+...
+---
+name:            invalid_shift
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: invalid_shift
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 4
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[TRUNC]](s16)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[C]](s16)
+    ; CHECK: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s16)
+    ; CHECK: G_STORE [[TRUNC1]](s8), [[COPY1]](p0) :: (store (s8))
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C1]](s64)
+    ; CHECK: G_STORE [[TRUNC2]](s8), [[PTR_ADD]](p0) :: (store (s8))
+    ; CHECK: RET_ReallyLR
+    %2:_(s32) = COPY $w0
+    %0:_(s16) = G_TRUNC %2(s32)
+    %1:_(p0) = COPY $x1
+    %4:_(s16) = G_CONSTANT i16 4
+    %3:_(s8) = G_TRUNC %0(s16)
+    %5:_(s16) = G_LSHR %0, %4(s16)
+    %6:_(s8) = G_TRUNC %5(s16)
+    G_STORE %3(s8), %1(p0) :: (store (s8))
+    %7:_(s64) = G_CONSTANT i64 1
+    %8:_(p0) = G_PTR_ADD %1, %7(s64)
+    G_STORE %6(s8), %8(p0) :: (store (s8))
+    RET_ReallyLR
+
+...
+---
+name:            missing_store
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: missing_store
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CHECK: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s8) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C]](s32)
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s32)
+    ; CHECK: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[COPY]], [[C1]](s32)
+    ; CHECK: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR1]](s32)
+    ; CHECK: G_STORE [[TRUNC]](s8), [[COPY1]](p0) :: (store (s8))
+    ; CHECK: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C2]](s64)
+    ; CHECK: G_STORE [[TRUNC1]](s8), [[PTR_ADD]](p0) :: (store (s8))
+    ; CHECK: [[C3:%[0-9]+]]:_(s64) = G_CONSTANT i64 3
+    ; CHECK: [[PTR_ADD1:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C3]](s64)
+    ; CHECK: G_STORE [[TRUNC2]](s8), [[PTR_ADD1]](p0) :: (store (s8))
+    ; CHECK: RET_ReallyLR
+    %0:_(s32) = COPY $w0
+    %1:_(p0) = COPY $x1
+    %3:_(s32) = G_CONSTANT i32 8
+    %6:_(s32) = G_CONSTANT i32 24
+    %2:_(s8) = G_TRUNC %0(s32)
+    %4:_(s32) = G_LSHR %0, %3(s32)
+    %5:_(s8) = G_TRUNC %4(s32)
+    %7:_(s32) = G_LSHR %0, %6(s32)
+    %8:_(s8) = G_TRUNC %7(s32)
+    G_STORE %2(s8), %1(p0) :: (store (s8))
+    %9:_(s64) = G_CONSTANT i64 1
+    %10:_(p0) = G_PTR_ADD %1, %9(s64)
+    G_STORE %5(s8), %10(p0) :: (store (s8))
+    %11:_(s64) = G_CONSTANT i64 3
+    %12:_(p0) = G_PTR_ADD %1, %11(s64)
+    G_STORE %8(s8), %12(p0) :: (store (s8))
+    RET_ReallyLR
+
+...
+---
+name:            different_base_reg
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+  - { reg: '$x2' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1, $x2
+
+    ; CHECK-LABEL: name: different_base_reg
+    ; CHECK: liveins: $w0, $x1, $x2
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[COPY2:%[0-9]+]]:_(p0) = COPY $x2
+    ; CHECK: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[TRUNC]](s16)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[C]](s16)
+    ; CHECK: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s16)
+    ; CHECK: G_STORE [[TRUNC1]](s8), [[COPY1]](p0) :: (store (s8))
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY2]], [[C1]](s64)
+    ; CHECK: G_STORE [[TRUNC2]](s8), [[PTR_ADD]](p0) :: (store (s8))
+    ; CHECK: RET_ReallyLR
+    %3:_(s32) = COPY $w0
+    %0:_(s16) = G_TRUNC %3(s32)
+    %1:_(p0) = COPY $x1
+    %2:_(p0) = COPY $x2
+    %5:_(s16) = G_CONSTANT i16 8
+    %4:_(s8) = G_TRUNC %0(s16)
+    %6:_(s16) = G_LSHR %0, %5(s16)
+    %7:_(s8) = G_TRUNC %6(s16)
+    G_STORE %4(s8), %1(p0) :: (store (s8))
+    %8:_(s64) = G_CONSTANT i64 1
+    %9:_(p0) = G_PTR_ADD %2, %8(s64)
+    G_STORE %7(s8), %9(p0) :: (store (s8))
+    RET_ReallyLR
+
+...
+---
+name:            second_store_is_volatile
+alignment:       4
+tracksRegLiveness: true
+liveins:
+  - { reg: '$w0' }
+  - { reg: '$x1' }
+body:             |
+  bb.1:
+    liveins: $w0, $x1
+
+    ; CHECK-LABEL: name: second_store_is_volatile
+    ; CHECK: liveins: $w0, $x1
+    ; CHECK: [[COPY:%[0-9]+]]:_(s32) = COPY $w0
+    ; CHECK: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32)
+    ; CHECK: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
+    ; CHECK: [[C:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; CHECK: [[TRUNC1:%[0-9]+]]:_(s8) = G_TRUNC [[TRUNC]](s16)
+    ; CHECK: [[LSHR:%[0-9]+]]:_(s16) = G_LSHR [[TRUNC]], [[C]](s16)
+    ; CHECK: [[TRUNC2:%[0-9]+]]:_(s8) = G_TRUNC [[LSHR]](s16)
+    ; CHECK: G_STORE [[TRUNC1]](s8), [[COPY1]](p0) :: (volatile store (s8))
+    ; CHECK: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CHECK: [[PTR_ADD:%[0-9]+]]:_(p0) = G_PTR_ADD [[COPY1]], [[C1]](s64)
+    ; CHECK: G_STORE [[TRUNC2]](s8), [[PTR_ADD]](p0) :: (store (s8))
+    ; CHECK: RET_ReallyLR
+    %2:_(s32) = COPY $w0
+    %0:_(s16) = G_TRUNC %2(s32)
+    %1:_(p0) = COPY $x1
+    %4:_(s16) = G_CONSTANT i16 8
+    %3:_(s8) = G_TRUNC %0(s16)
+    %5:_(s16) = G_LSHR %0, %4(s16)
+    %6:_(s8) = G_TRUNC %5(s16)
+    G_STORE %3(s8), %1(p0) :: (volatile store (s8))
+    %7:_(s64) = G_CONSTANT i64 1
+    %8:_(p0) = G_PTR_ADD %1, %7(s64)
+    G_STORE %6(s8), %8(p0) :: (store (s8))
+    RET_ReallyLR
+
+...