LegalizerHelper.cpp

Go to the documentation of this file.

//===-- llvm/CodeGen/GlobalISel/LegalizerHelper.cpp -----------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
/// \file This file implements the LegalizerHelper class to legalize
/// individual instructions and the LegalizeMachineIR wrapper pass for the
/// primary legalization.
//
//===----------------------------------------------------------------------===//
 
#include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
#include "llvm/CodeGen/GlobalISel/CallLowering.h"
#include "llvm/CodeGen/GlobalISel/GISelChangeObserver.h"
#include "llvm/CodeGen/GlobalISel/GISelKnownBits.h"
#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
#include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
#include "llvm/CodeGen/GlobalISel/LostDebugLocObserver.h"
#include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
#include "llvm/CodeGen/GlobalISel/Utils.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetFrameLowering.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/CodeGen/TargetSubtargetInfo.h"
#include "llvm/IR/Instructions.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include <numeric>
#include <optional>
 
#define DEBUG_TYPE "legalizer"
 
using namespace llvm;
using namespace LegalizeActions;
using namespace MIPatternMatch;
 
/// Try to break down \p OrigTy into \p NarrowTy sized pieces.
///
/// Returns the number of \p NarrowTy elements needed to reconstruct \p OrigTy,
/// with any leftover piece as type \p LeftoverTy
///
/// Returns -1 in the first element of the pair if the breakdown is not
/// satisfiable.
static std::pair<int, int>
getNarrowTypeBreakDown(LLT OrigTy, LLT NarrowTy, LLT &LeftoverTy) {
  assert(!LeftoverTy.isValid() && "this is an out argument");
 
  unsigned Size = OrigTy.getSizeInBits();
  unsigned NarrowSize = NarrowTy.getSizeInBits();
  unsigned NumParts = Size / NarrowSize;
  unsigned LeftoverSize = Size - NumParts * NarrowSize;
  assert(Size > NarrowSize);
 
  if (LeftoverSize == 0)
    return {NumParts, 0};
 
  if (NarrowTy.isVector()) {
    unsigned EltSize = OrigTy.getScalarSizeInBits();
    if (LeftoverSize % EltSize != 0)
      return {-1, -1};
    LeftoverTy = LLT::scalarOrVector(
        ElementCount::getFixed(LeftoverSize / EltSize), EltSize);
  } else {
    LeftoverTy = LLT::scalar(LeftoverSize);
  }
 
  int NumLeftover = LeftoverSize / LeftoverTy.getSizeInBits();
  return std::make_pair(NumParts, NumLeftover);
}
 
static Type *getFloatTypeForLLT(LLVMContext &Ctx, LLT Ty) {
 
  if (!Ty.isScalar())
    return nullptr;
 
  switch (Ty.getSizeInBits()) {
  case 16:
    return Type::getHalfTy(Ctx);
  case 32:
    return Type::getFloatTy(Ctx);
  case 64:
    return Type::getDoubleTy(Ctx);
  case 80:
    return Type::getX86_FP80Ty(Ctx);
  case 128:
    return Type::getFP128Ty(Ctx);
  default:
    return nullptr;
  }
}
 
LegalizerHelper::LegalizerHelper(MachineFunction &MF,
                                 GISelChangeObserver &Observer,
                                 MachineIRBuilder &Builder)
    : MIRBuilder(Builder), Observer(Observer), MRI(MF.getRegInfo()),
      LI(*MF.getSubtarget().getLegalizerInfo()),
      TLI(*MF.getSubtarget().getTargetLowering()), KB(nullptr) {}
 
LegalizerHelper::LegalizerHelper(MachineFunction &MF, const LegalizerInfo &LI,
                                 GISelChangeObserver &Observer,
                                 MachineIRBuilder &B, GISelKnownBits *KB)
    : MIRBuilder(B), Observer(Observer), MRI(MF.getRegInfo()), LI(LI),
      TLI(*MF.getSubtarget().getTargetLowering()), KB(KB) {}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::legalizeInstrStep(MachineInstr &MI,
                                   LostDebugLocObserver &LocObserver) {
  LLVM_DEBUG(dbgs() << "Legalizing: " << MI);
 
  MIRBuilder.setInstrAndDebugLoc(MI);
 
  if (isa<GIntrinsic>(MI))
    return LI.legalizeIntrinsic(*this, MI) ? Legalized : UnableToLegalize;
  auto Step = LI.getAction(MI, MRI);
  switch (Step.Action) {
  case Legal:
    LLVM_DEBUG(dbgs() << ".. Already legal\n");
    return AlreadyLegal;
  case Libcall:
    LLVM_DEBUG(dbgs() << ".. Convert to libcall\n");
    return libcall(MI, LocObserver);
  case NarrowScalar:
    LLVM_DEBUG(dbgs() << ".. Narrow scalar\n");
    return narrowScalar(MI, Step.TypeIdx, Step.NewType);
  case WidenScalar:
    LLVM_DEBUG(dbgs() << ".. Widen scalar\n");
    return widenScalar(MI, Step.TypeIdx, Step.NewType);
  case Bitcast:
    LLVM_DEBUG(dbgs() << ".. Bitcast type\n");
    return bitcast(MI, Step.TypeIdx, Step.NewType);
  case Lower:
    LLVM_DEBUG(dbgs() << ".. Lower\n");
    return lower(MI, Step.TypeIdx, Step.NewType);
  case FewerElements:
    LLVM_DEBUG(dbgs() << ".. Reduce number of elements\n");
    return fewerElementsVector(MI, Step.TypeIdx, Step.NewType);
  case MoreElements:
    LLVM_DEBUG(dbgs() << ".. Increase number of elements\n");
    return moreElementsVector(MI, Step.TypeIdx, Step.NewType);
  case Custom:
    LLVM_DEBUG(dbgs() << ".. Custom legalization\n");
    return LI.legalizeCustom(*this, MI) ? Legalized : UnableToLegalize;
  default:
    LLVM_DEBUG(dbgs() << ".. Unable to legalize\n");
    return UnableToLegalize;
  }
}
 
void LegalizerHelper::extractParts(Register Reg, LLT Ty, int NumParts,
                                   SmallVectorImpl<Register> &VRegs) {
  for (int i = 0; i < NumParts; ++i)
    VRegs.push_back(MRI.createGenericVirtualRegister(Ty));
  MIRBuilder.buildUnmerge(VRegs, Reg);
}
 
bool LegalizerHelper::extractParts(Register Reg, LLT RegTy,
                                   LLT MainTy, LLT &LeftoverTy,
                                   SmallVectorImpl<Register> &VRegs,
                                   SmallVectorImpl<Register> &LeftoverRegs) {
  assert(!LeftoverTy.isValid() && "this is an out argument");
 
  unsigned RegSize = RegTy.getSizeInBits();
  unsigned MainSize = MainTy.getSizeInBits();
  unsigned NumParts = RegSize / MainSize;
  unsigned LeftoverSize = RegSize - NumParts * MainSize;
 
  // Use an unmerge when possible.
  if (LeftoverSize == 0) {
    for (unsigned I = 0; I < NumParts; ++I)
      VRegs.push_back(MRI.createGenericVirtualRegister(MainTy));
    MIRBuilder.buildUnmerge(VRegs, Reg);
    return true;
  }
 
  // Perform irregular split. Leftover is last element of RegPieces.
  if (MainTy.isVector()) {
    SmallVector<Register, 8> RegPieces;
    extractVectorParts(Reg, MainTy.getNumElements(), RegPieces);
    for (unsigned i = 0; i < RegPieces.size() - 1; ++i)
      VRegs.push_back(RegPieces[i]);
    LeftoverRegs.push_back(RegPieces[RegPieces.size() - 1]);
    LeftoverTy = MRI.getType(LeftoverRegs[0]);
    return true;
  }
 
  LeftoverTy = LLT::scalar(LeftoverSize);
  // For irregular sizes, extract the individual parts.
  for (unsigned I = 0; I != NumParts; ++I) {
    Register NewReg = MRI.createGenericVirtualRegister(MainTy);
    VRegs.push_back(NewReg);
    MIRBuilder.buildExtract(NewReg, Reg, MainSize * I);
  }
 
  for (unsigned Offset = MainSize * NumParts; Offset < RegSize;
       Offset += LeftoverSize) {
    Register NewReg = MRI.createGenericVirtualRegister(LeftoverTy);
    LeftoverRegs.push_back(NewReg);
    MIRBuilder.buildExtract(NewReg, Reg, Offset);
  }
 
  return true;
}
 
void LegalizerHelper::extractVectorParts(Register Reg, unsigned NumElts,
                                         SmallVectorImpl<Register> &VRegs) {
  LLT RegTy = MRI.getType(Reg);
  assert(RegTy.isVector() && "Expected a vector type");
 
  LLT EltTy = RegTy.getElementType();
  LLT NarrowTy = (NumElts == 1) ? EltTy : LLT::fixed_vector(NumElts, EltTy);
  unsigned RegNumElts = RegTy.getNumElements();
  unsigned LeftoverNumElts = RegNumElts % NumElts;
  unsigned NumNarrowTyPieces = RegNumElts / NumElts;
 
  // Perfect split without leftover
  if (LeftoverNumElts == 0)
    return extractParts(Reg, NarrowTy, NumNarrowTyPieces, VRegs);
 
  // Irregular split. Provide direct access to all elements for artifact
  // combiner using unmerge to elements. Then build vectors with NumElts
  // elements. Remaining element(s) will be (used to build vector) Leftover.
  SmallVector<Register, 8> Elts;
  extractParts(Reg, EltTy, RegNumElts, Elts);
 
  unsigned Offset = 0;
  // Requested sub-vectors of NarrowTy.
  for (unsigned i = 0; i < NumNarrowTyPieces; ++i, Offset += NumElts) {
    ArrayRef<Register> Pieces(&Elts[Offset], NumElts);
    VRegs.push_back(MIRBuilder.buildMergeLikeInstr(NarrowTy, Pieces).getReg(0));
  }
 
  // Leftover element(s).
  if (LeftoverNumElts == 1) {
    VRegs.push_back(Elts[Offset]);
  } else {
    LLT LeftoverTy = LLT::fixed_vector(LeftoverNumElts, EltTy);
    ArrayRef<Register> Pieces(&Elts[Offset], LeftoverNumElts);
    VRegs.push_back(
        MIRBuilder.buildMergeLikeInstr(LeftoverTy, Pieces).getReg(0));
  }
}
 
void LegalizerHelper::insertParts(Register DstReg,
                                  LLT ResultTy, LLT PartTy,
                                  ArrayRef<Register> PartRegs,
                                  LLT LeftoverTy,
                                  ArrayRef<Register> LeftoverRegs) {
  if (!LeftoverTy.isValid()) {
    assert(LeftoverRegs.empty());
 
    if (!ResultTy.isVector()) {
      MIRBuilder.buildMergeLikeInstr(DstReg, PartRegs);
      return;
    }
 
    if (PartTy.isVector())
      MIRBuilder.buildConcatVectors(DstReg, PartRegs);
    else
      MIRBuilder.buildBuildVector(DstReg, PartRegs);
    return;
  }
 
  // Merge sub-vectors with different number of elements and insert into DstReg.
  if (ResultTy.isVector()) {
    assert(LeftoverRegs.size() == 1 && "Expected one leftover register");
    SmallVector<Register, 8> AllRegs;
    for (auto Reg : concat<const Register>(PartRegs, LeftoverRegs))
      AllRegs.push_back(Reg);
    return mergeMixedSubvectors(DstReg, AllRegs);
  }
 
  SmallVector<Register> GCDRegs;
  LLT GCDTy = getGCDType(getGCDType(ResultTy, LeftoverTy), PartTy);
  for (auto PartReg : concat<const Register>(PartRegs, LeftoverRegs))
    extractGCDType(GCDRegs, GCDTy, PartReg);
  LLT ResultLCMTy = buildLCMMergePieces(ResultTy, LeftoverTy, GCDTy, GCDRegs);
  buildWidenedRemergeToDst(DstReg, ResultLCMTy, GCDRegs);
}
 
void LegalizerHelper::appendVectorElts(SmallVectorImpl<Register> &Elts,
                                       Register Reg) {
  LLT Ty = MRI.getType(Reg);
  SmallVector<Register, 8> RegElts;
  extractParts(Reg, Ty.getScalarType(), Ty.getNumElements(), RegElts);
  Elts.append(RegElts);
}
 
/// Merge \p PartRegs with different types into \p DstReg.
void LegalizerHelper::mergeMixedSubvectors(Register DstReg,
                                           ArrayRef<Register> PartRegs) {
  SmallVector<Register, 8> AllElts;
  for (unsigned i = 0; i < PartRegs.size() - 1; ++i)
    appendVectorElts(AllElts, PartRegs[i]);
 
  Register Leftover = PartRegs[PartRegs.size() - 1];
  if (MRI.getType(Leftover).isScalar())
    AllElts.push_back(Leftover);
  else
    appendVectorElts(AllElts, Leftover);
 
  MIRBuilder.buildMergeLikeInstr(DstReg, AllElts);
}
 
/// Append the result registers of G_UNMERGE_VALUES \p MI to \p Regs.
static void getUnmergeResults(SmallVectorImpl<Register> &Regs,
                              const MachineInstr &MI) {
  assert(MI.getOpcode() == TargetOpcode::G_UNMERGE_VALUES);
 
  const int StartIdx = Regs.size();
  const int NumResults = MI.getNumOperands() - 1;
  Regs.resize(Regs.size() + NumResults);
  for (int I = 0; I != NumResults; ++I)
    Regs[StartIdx + I] = MI.getOperand(I).getReg();
}
 
void LegalizerHelper::extractGCDType(SmallVectorImpl<Register> &Parts,
                                     LLT GCDTy, Register SrcReg) {
  LLT SrcTy = MRI.getType(SrcReg);
  if (SrcTy == GCDTy) {
    // If the source already evenly divides the result type, we don't need to do
    // anything.
    Parts.push_back(SrcReg);
  } else {
    // Need to split into common type sized pieces.
    auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg);
    getUnmergeResults(Parts, *Unmerge);
  }
}
 
LLT LegalizerHelper::extractGCDType(SmallVectorImpl<Register> &Parts, LLT DstTy,
                                    LLT NarrowTy, Register SrcReg) {
  LLT SrcTy = MRI.getType(SrcReg);
  LLT GCDTy = getGCDType(getGCDType(SrcTy, NarrowTy), DstTy);
  extractGCDType(Parts, GCDTy, SrcReg);
  return GCDTy;
}
 
LLT LegalizerHelper::buildLCMMergePieces(LLT DstTy, LLT NarrowTy, LLT GCDTy,
                                         SmallVectorImpl<Register> &VRegs,
                                         unsigned PadStrategy) {
  LLT LCMTy = getLCMType(DstTy, NarrowTy);
 
  int NumParts = LCMTy.getSizeInBits() / NarrowTy.getSizeInBits();
  int NumSubParts = NarrowTy.getSizeInBits() / GCDTy.getSizeInBits();
  int NumOrigSrc = VRegs.size();
 
  Register PadReg;
 
  // Get a value we can use to pad the source value if the sources won't evenly
  // cover the result type.
  if (NumOrigSrc < NumParts * NumSubParts) {
    if (PadStrategy == TargetOpcode::G_ZEXT)
      PadReg = MIRBuilder.buildConstant(GCDTy, 0).getReg(0);
    else if (PadStrategy == TargetOpcode::G_ANYEXT)
      PadReg = MIRBuilder.buildUndef(GCDTy).getReg(0);
    else {
      assert(PadStrategy == TargetOpcode::G_SEXT);
 
      // Shift the sign bit of the low register through the high register.
      auto ShiftAmt =
        MIRBuilder.buildConstant(LLT::scalar(64), GCDTy.getSizeInBits() - 1);
      PadReg = MIRBuilder.buildAShr(GCDTy, VRegs.back(), ShiftAmt).getReg(0);
    }
  }
 
  // Registers for the final merge to be produced.
  SmallVector<Register, 4> Remerge(NumParts);
 
  // Registers needed for intermediate merges, which will be merged into a
  // source for Remerge.
  SmallVector<Register, 4> SubMerge(NumSubParts);
 
  // Once we've fully read off the end of the original source bits, we can reuse
  // the same high bits for remaining padding elements.
  Register AllPadReg;
 
  // Build merges to the LCM type to cover the original result type.
  for (int I = 0; I != NumParts; ++I) {
    bool AllMergePartsArePadding = true;
 
    // Build the requested merges to the requested type.
    for (int J = 0; J != NumSubParts; ++J) {
      int Idx = I * NumSubParts + J;
      if (Idx >= NumOrigSrc) {
        SubMerge[J] = PadReg;
        continue;
      }
 
      SubMerge[J] = VRegs[Idx];
 
      // There are meaningful bits here we can't reuse later.
      AllMergePartsArePadding = false;
    }
 
    // If we've filled up a complete piece with padding bits, we can directly
    // emit the natural sized constant if applicable, rather than a merge of
    // smaller constants.
    if (AllMergePartsArePadding && !AllPadReg) {
      if (PadStrategy == TargetOpcode::G_ANYEXT)
        AllPadReg = MIRBuilder.buildUndef(NarrowTy).getReg(0);
      else if (PadStrategy == TargetOpcode::G_ZEXT)
        AllPadReg = MIRBuilder.buildConstant(NarrowTy, 0).getReg(0);
 
      // If this is a sign extension, we can't materialize a trivial constant
      // with the right type and have to produce a merge.
    }
 
    if (AllPadReg) {
      // Avoid creating additional instructions if we're just adding additional
      // copies of padding bits.
      Remerge[I] = AllPadReg;
      continue;
    }
 
    if (NumSubParts == 1)
      Remerge[I] = SubMerge[0];
    else
      Remerge[I] = MIRBuilder.buildMergeLikeInstr(NarrowTy, SubMerge).getReg(0);
 
    // In the sign extend padding case, re-use the first all-signbit merge.
    if (AllMergePartsArePadding && !AllPadReg)
      AllPadReg = Remerge[I];
  }
 
  VRegs = std::move(Remerge);
  return LCMTy;
}
 
void LegalizerHelper::buildWidenedRemergeToDst(Register DstReg, LLT LCMTy,
                                               ArrayRef<Register> RemergeRegs) {
  LLT DstTy = MRI.getType(DstReg);
 
  // Create the merge to the widened source, and extract the relevant bits into
  // the result.
 
  if (DstTy == LCMTy) {
    MIRBuilder.buildMergeLikeInstr(DstReg, RemergeRegs);
    return;
  }
 
  auto Remerge = MIRBuilder.buildMergeLikeInstr(LCMTy, RemergeRegs);
  if (DstTy.isScalar() && LCMTy.isScalar()) {
    MIRBuilder.buildTrunc(DstReg, Remerge);
    return;
  }
 
  if (LCMTy.isVector()) {
    unsigned NumDefs = LCMTy.getSizeInBits() / DstTy.getSizeInBits();
    SmallVector<Register, 8> UnmergeDefs(NumDefs);
    UnmergeDefs[0] = DstReg;
    for (unsigned I = 1; I != NumDefs; ++I)
      UnmergeDefs[I] = MRI.createGenericVirtualRegister(DstTy);
 
    MIRBuilder.buildUnmerge(UnmergeDefs,
                            MIRBuilder.buildMergeLikeInstr(LCMTy, RemergeRegs));
    return;
  }
 
  llvm_unreachable("unhandled case");
}
 
static RTLIB::Libcall getRTLibDesc(unsigned Opcode, unsigned Size) {
#define RTLIBCASE_INT(LibcallPrefix)                                           \
  do {                                                                         \
    switch (Size) {                                                            \
    case 32:                                                                   \
      return RTLIB::LibcallPrefix##32;                                         \
    case 64:                                                                   \
      return RTLIB::LibcallPrefix##64;                                         \
    case 128:                                                                  \
      return RTLIB::LibcallPrefix##128;                                        \
    default:                                                                   \
      llvm_unreachable("unexpected size");                                     \
    }                                                                          \
  } while (0)
 
#define RTLIBCASE(LibcallPrefix)                                               \
  do {                                                                         \
    switch (Size) {                                                            \
    case 32:                                                                   \
      return RTLIB::LibcallPrefix##32;                                         \
    case 64:                                                                   \
      return RTLIB::LibcallPrefix##64;                                         \
    case 80:                                                                   \
      return RTLIB::LibcallPrefix##80;                                         \
    case 128:                                                                  \
      return RTLIB::LibcallPrefix##128;                                        \
    default:                                                                   \
      llvm_unreachable("unexpected size");                                     \
    }                                                                          \
  } while (0)
 
  switch (Opcode) {
  case TargetOpcode::G_MUL:
    RTLIBCASE_INT(MUL_I);
  case TargetOpcode::G_SDIV:
    RTLIBCASE_INT(SDIV_I);
  case TargetOpcode::G_UDIV:
    RTLIBCASE_INT(UDIV_I);
  case TargetOpcode::G_SREM:
    RTLIBCASE_INT(SREM_I);
  case TargetOpcode::G_UREM:
    RTLIBCASE_INT(UREM_I);
  case TargetOpcode::G_CTLZ_ZERO_UNDEF:
    RTLIBCASE_INT(CTLZ_I);
  case TargetOpcode::G_FADD:
    RTLIBCASE(ADD_F);
  case TargetOpcode::G_FSUB:
    RTLIBCASE(SUB_F);
  case TargetOpcode::G_FMUL:
    RTLIBCASE(MUL_F);
  case TargetOpcode::G_FDIV:
    RTLIBCASE(DIV_F);
  case TargetOpcode::G_FEXP:
    RTLIBCASE(EXP_F);
  case TargetOpcode::G_FEXP2:
    RTLIBCASE(EXP2_F);
  case TargetOpcode::G_FEXP10:
    RTLIBCASE(EXP10_F);
  case TargetOpcode::G_FREM:
    RTLIBCASE(REM_F);
  case TargetOpcode::G_FPOW:
    RTLIBCASE(POW_F);
  case TargetOpcode::G_FMA:
    RTLIBCASE(FMA_F);
  case TargetOpcode::G_FSIN:
    RTLIBCASE(SIN_F);
  case TargetOpcode::G_FCOS:
    RTLIBCASE(COS_F);
  case TargetOpcode::G_FLOG10:
    RTLIBCASE(LOG10_F);
  case TargetOpcode::G_FLOG:
    RTLIBCASE(LOG_F);
  case TargetOpcode::G_FLOG2:
    RTLIBCASE(LOG2_F);
  case TargetOpcode::G_FLDEXP:
    RTLIBCASE(LDEXP_F);
  case TargetOpcode::G_FCEIL:
    RTLIBCASE(CEIL_F);
  case TargetOpcode::G_FFLOOR:
    RTLIBCASE(FLOOR_F);
  case TargetOpcode::G_FMINNUM:
    RTLIBCASE(FMIN_F);
  case TargetOpcode::G_FMAXNUM:
    RTLIBCASE(FMAX_F);
  case TargetOpcode::G_FSQRT:
    RTLIBCASE(SQRT_F);
  case TargetOpcode::G_FRINT:
    RTLIBCASE(RINT_F);
  case TargetOpcode::G_FNEARBYINT:
    RTLIBCASE(NEARBYINT_F);
  case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
    RTLIBCASE(ROUNDEVEN_F);
  }
  llvm_unreachable("Unknown libcall function");
}
 
/// True if an instruction is in tail position in its caller. Intended for
/// legalizing libcalls as tail calls when possible.
static bool isLibCallInTailPosition(MachineInstr &MI,
                                    const TargetInstrInfo &TII,
                                    MachineRegisterInfo &MRI) {
  MachineBasicBlock &MBB = *MI.getParent();
  const Function &F = MBB.getParent()->getFunction();
 
  // Conservatively require the attributes of the call to match those of
  // the return. Ignore NoAlias and NonNull because they don't affect the
  // call sequence.
  AttributeList CallerAttrs = F.getAttributes();
  if (AttrBuilder(F.getContext(), CallerAttrs.getRetAttrs())
          .removeAttribute(Attribute::NoAlias)
          .removeAttribute(Attribute::NonNull)
          .hasAttributes())
    return false;
 
  // It's not safe to eliminate the sign / zero extension of the return value.
  if (CallerAttrs.hasRetAttr(Attribute::ZExt) ||
      CallerAttrs.hasRetAttr(Attribute::SExt))
    return false;
 
  // Only tail call if the following instruction is a standard return or if we
  // have a `thisreturn` callee, and a sequence like:
  //
  //   G_MEMCPY %0, %1, %2
  //   $x0 = COPY %0
  //   RET_ReallyLR implicit $x0
  auto Next = next_nodbg(MI.getIterator(), MBB.instr_end());
  if (Next != MBB.instr_end() && Next->isCopy()) {
    switch (MI.getOpcode()) {
    default:
      llvm_unreachable("unsupported opcode");
    case TargetOpcode::G_BZERO:
      return false;
    case TargetOpcode::G_MEMCPY:
    case TargetOpcode::G_MEMMOVE:
    case TargetOpcode::G_MEMSET:
      break;
    }
 
    Register VReg = MI.getOperand(0).getReg();
    if (!VReg.isVirtual() || VReg != Next->getOperand(1).getReg())
      return false;
 
    Register PReg = Next->getOperand(0).getReg();
    if (!PReg.isPhysical())
      return false;
 
    auto Ret = next_nodbg(Next, MBB.instr_end());
    if (Ret == MBB.instr_end() || !Ret->isReturn())
      return false;
 
    if (Ret->getNumImplicitOperands() != 1)
      return false;
 
    if (PReg != Ret->getOperand(0).getReg())
      return false;
 
    // Skip over the COPY that we just validated.
    Next = Ret;
  }
 
  if (Next == MBB.instr_end() || TII.isTailCall(*Next) || !Next->isReturn())
    return false;
 
  return true;
}
 
LegalizerHelper::LegalizeResult
llvm::createLibcall(MachineIRBuilder &MIRBuilder, const char *Name,
                    const CallLowering::ArgInfo &Result,
                    ArrayRef<CallLowering::ArgInfo> Args,
                    const CallingConv::ID CC) {
  auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering();
 
  CallLowering::CallLoweringInfo Info;
  Info.CallConv = CC;
  Info.Callee = MachineOperand::CreateES(Name);
  Info.OrigRet = Result;
  std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs));
  if (!CLI.lowerCall(MIRBuilder, Info))
    return LegalizerHelper::UnableToLegalize;
 
  return LegalizerHelper::Legalized;
}
 
LegalizerHelper::LegalizeResult
llvm::createLibcall(MachineIRBuilder &MIRBuilder, RTLIB::Libcall Libcall,
                    const CallLowering::ArgInfo &Result,
                    ArrayRef<CallLowering::ArgInfo> Args) {
  auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering();
  const char *Name = TLI.getLibcallName(Libcall);
  const CallingConv::ID CC = TLI.getLibcallCallingConv(Libcall);
  return createLibcall(MIRBuilder, Name, Result, Args, CC);
}
 
// Useful for libcalls where all operands have the same type.
static LegalizerHelper::LegalizeResult
simpleLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, unsigned Size,
              Type *OpType) {
  auto Libcall = getRTLibDesc(MI.getOpcode(), Size);
 
  // FIXME: What does the original arg index mean here?
  SmallVector<CallLowering::ArgInfo, 3> Args;
  for (const MachineOperand &MO : llvm::drop_begin(MI.operands()))
    Args.push_back({MO.getReg(), OpType, 0});
  return createLibcall(MIRBuilder, Libcall,
                       {MI.getOperand(0).getReg(), OpType, 0}, Args);
}
 
LegalizerHelper::LegalizeResult
llvm::createMemLibcall(MachineIRBuilder &MIRBuilder, MachineRegisterInfo &MRI,
                       MachineInstr &MI, LostDebugLocObserver &LocObserver) {
  auto &Ctx = MIRBuilder.getMF().getFunction().getContext();
 
  SmallVector<CallLowering::ArgInfo, 3> Args;
  // Add all the args, except for the last which is an imm denoting 'tail'.
  for (unsigned i = 0; i < MI.getNumOperands() - 1; ++i) {
    Register Reg = MI.getOperand(i).getReg();
 
    // Need derive an IR type for call lowering.
    LLT OpLLT = MRI.getType(Reg);
    Type *OpTy = nullptr;
    if (OpLLT.isPointer())
      OpTy = PointerType::get(Ctx, OpLLT.getAddressSpace());
    else
      OpTy = IntegerType::get(Ctx, OpLLT.getSizeInBits());
    Args.push_back({Reg, OpTy, 0});
  }
 
  auto &CLI = *MIRBuilder.getMF().getSubtarget().getCallLowering();
  auto &TLI = *MIRBuilder.getMF().getSubtarget().getTargetLowering();
  RTLIB::Libcall RTLibcall;
  unsigned Opc = MI.getOpcode();
  switch (Opc) {
  case TargetOpcode::G_BZERO:
    RTLibcall = RTLIB::BZERO;
    break;
  case TargetOpcode::G_MEMCPY:
    RTLibcall = RTLIB::MEMCPY;
    Args[0].Flags[0].setReturned();
    break;
  case TargetOpcode::G_MEMMOVE:
    RTLibcall = RTLIB::MEMMOVE;
    Args[0].Flags[0].setReturned();
    break;
  case TargetOpcode::G_MEMSET:
    RTLibcall = RTLIB::MEMSET;
    Args[0].Flags[0].setReturned();
    break;
  default:
    llvm_unreachable("unsupported opcode");
  }
  const char *Name = TLI.getLibcallName(RTLibcall);
 
  // Unsupported libcall on the target.
  if (!Name) {
    LLVM_DEBUG(dbgs() << ".. .. Could not find libcall name for "
                      << MIRBuilder.getTII().getName(Opc) << "\n");
    return LegalizerHelper::UnableToLegalize;
  }
 
  CallLowering::CallLoweringInfo Info;
  Info.CallConv = TLI.getLibcallCallingConv(RTLibcall);
  Info.Callee = MachineOperand::CreateES(Name);
  Info.OrigRet = CallLowering::ArgInfo({0}, Type::getVoidTy(Ctx), 0);
  Info.IsTailCall = MI.getOperand(MI.getNumOperands() - 1).getImm() &&
                    isLibCallInTailPosition(MI, MIRBuilder.getTII(), MRI);
 
  std::copy(Args.begin(), Args.end(), std::back_inserter(Info.OrigArgs));
  if (!CLI.lowerCall(MIRBuilder, Info))
    return LegalizerHelper::UnableToLegalize;
 
  if (Info.LoweredTailCall) {
    assert(Info.IsTailCall && "Lowered tail call when it wasn't a tail call?");
 
    // Check debug locations before removing the return.
    LocObserver.checkpoint(true);
 
    // We must have a return following the call (or debug insts) to get past
    // isLibCallInTailPosition.
    do {
      MachineInstr *Next = MI.getNextNode();
      assert(Next &&
             (Next->isCopy() || Next->isReturn() || Next->isDebugInstr()) &&
             "Expected instr following MI to be return or debug inst?");
      // We lowered a tail call, so the call is now the return from the block.
      // Delete the old return.
      Next->eraseFromParent();
    } while (MI.getNextNode());
 
    // We expect to lose the debug location from the return.
    LocObserver.checkpoint(false);
  }
 
  return LegalizerHelper::Legalized;
}
 
static RTLIB::Libcall getConvRTLibDesc(unsigned Opcode, Type *ToType,
                                       Type *FromType) {
  auto ToMVT = MVT::getVT(ToType);
  auto FromMVT = MVT::getVT(FromType);
 
  switch (Opcode) {
  case TargetOpcode::G_FPEXT:
    return RTLIB::getFPEXT(FromMVT, ToMVT);
  case TargetOpcode::G_FPTRUNC:
    return RTLIB::getFPROUND(FromMVT, ToMVT);
  case TargetOpcode::G_FPTOSI:
    return RTLIB::getFPTOSINT(FromMVT, ToMVT);
  case TargetOpcode::G_FPTOUI:
    return RTLIB::getFPTOUINT(FromMVT, ToMVT);
  case TargetOpcode::G_SITOFP:
    return RTLIB::getSINTTOFP(FromMVT, ToMVT);
  case TargetOpcode::G_UITOFP:
    return RTLIB::getUINTTOFP(FromMVT, ToMVT);
  }
  llvm_unreachable("Unsupported libcall function");
}
 
static LegalizerHelper::LegalizeResult
conversionLibcall(MachineInstr &MI, MachineIRBuilder &MIRBuilder, Type *ToType,
                  Type *FromType) {
  RTLIB::Libcall Libcall = getConvRTLibDesc(MI.getOpcode(), ToType, FromType);
  return createLibcall(MIRBuilder, Libcall,
                       {MI.getOperand(0).getReg(), ToType, 0},
                       {{MI.getOperand(1).getReg(), FromType, 0}});
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::libcall(MachineInstr &MI, LostDebugLocObserver &LocObserver) {
  LLT LLTy = MRI.getType(MI.getOperand(0).getReg());
  unsigned Size = LLTy.getSizeInBits();
  auto &Ctx = MIRBuilder.getMF().getFunction().getContext();
 
  switch (MI.getOpcode()) {
  default:
    return UnableToLegalize;
  case TargetOpcode::G_MUL:
  case TargetOpcode::G_SDIV:
  case TargetOpcode::G_UDIV:
  case TargetOpcode::G_SREM:
  case TargetOpcode::G_UREM:
  case TargetOpcode::G_CTLZ_ZERO_UNDEF: {
    Type *HLTy = IntegerType::get(Ctx, Size);
    auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
    if (Status != Legalized)
      return Status;
    break;
  }
  case TargetOpcode::G_FADD:
  case TargetOpcode::G_FSUB:
  case TargetOpcode::G_FMUL:
  case TargetOpcode::G_FDIV:
  case TargetOpcode::G_FMA:
  case TargetOpcode::G_FPOW:
  case TargetOpcode::G_FREM:
  case TargetOpcode::G_FCOS:
  case TargetOpcode::G_FSIN:
  case TargetOpcode::G_FLOG10:
  case TargetOpcode::G_FLOG:
  case TargetOpcode::G_FLOG2:
  case TargetOpcode::G_FLDEXP:
  case TargetOpcode::G_FEXP:
  case TargetOpcode::G_FEXP2:
  case TargetOpcode::G_FEXP10:
  case TargetOpcode::G_FCEIL:
  case TargetOpcode::G_FFLOOR:
  case TargetOpcode::G_FMINNUM:
  case TargetOpcode::G_FMAXNUM:
  case TargetOpcode::G_FSQRT:
  case TargetOpcode::G_FRINT:
  case TargetOpcode::G_FNEARBYINT:
  case TargetOpcode::G_INTRINSIC_ROUNDEVEN: {
    Type *HLTy = getFloatTypeForLLT(Ctx, LLTy);
    if (!HLTy || (Size != 32 && Size != 64 && Size != 80 && Size != 128)) {
      LLVM_DEBUG(dbgs() << "No libcall available for type " << LLTy << ".\n");
      return UnableToLegalize;
    }
    auto Status = simpleLibcall(MI, MIRBuilder, Size, HLTy);
    if (Status != Legalized)
      return Status;
    break;
  }
  case TargetOpcode::G_FPEXT:
  case TargetOpcode::G_FPTRUNC: {
    Type *FromTy = getFloatTypeForLLT(Ctx,  MRI.getType(MI.getOperand(1).getReg()));
    Type *ToTy = getFloatTypeForLLT(Ctx, MRI.getType(MI.getOperand(0).getReg()));
    if (!FromTy || !ToTy)
      return UnableToLegalize;
    LegalizeResult Status = conversionLibcall(MI, MIRBuilder, ToTy, FromTy );
    if (Status != Legalized)
      return Status;
    break;
  }
  case TargetOpcode::G_FPTOSI:
  case TargetOpcode::G_FPTOUI: {
    // FIXME: Support other types
    unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
    unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
    if ((ToSize != 32 && ToSize != 64) || (FromSize != 32 && FromSize != 64))
      return UnableToLegalize;
    LegalizeResult Status = conversionLibcall(
        MI, MIRBuilder,
        ToSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx),
        FromSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx));
    if (Status != Legalized)
      return Status;
    break;
  }
  case TargetOpcode::G_SITOFP:
  case TargetOpcode::G_UITOFP: {
    // FIXME: Support other types
    unsigned FromSize = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
    unsigned ToSize = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
    if ((FromSize != 32 && FromSize != 64) || (ToSize != 32 && ToSize != 64))
      return UnableToLegalize;
    LegalizeResult Status = conversionLibcall(
        MI, MIRBuilder,
        ToSize == 64 ? Type::getDoubleTy(Ctx) : Type::getFloatTy(Ctx),
        FromSize == 32 ? Type::getInt32Ty(Ctx) : Type::getInt64Ty(Ctx));
    if (Status != Legalized)
      return Status;
    break;
  }
  case TargetOpcode::G_BZERO:
  case TargetOpcode::G_MEMCPY:
  case TargetOpcode::G_MEMMOVE:
  case TargetOpcode::G_MEMSET: {
    LegalizeResult Result =
        createMemLibcall(MIRBuilder, *MIRBuilder.getMRI(), MI, LocObserver);
    if (Result != Legalized)
      return Result;
    MI.eraseFromParent();
    return Result;
  }
  }
 
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult LegalizerHelper::narrowScalar(MachineInstr &MI,
                                                              unsigned TypeIdx,
                                                              LLT NarrowTy) {
  uint64_t SizeOp0 = MRI.getType(MI.getOperand(0).getReg()).getSizeInBits();
  uint64_t NarrowSize = NarrowTy.getSizeInBits();
 
  switch (MI.getOpcode()) {
  default:
    return UnableToLegalize;
  case TargetOpcode::G_IMPLICIT_DEF: {
    Register DstReg = MI.getOperand(0).getReg();
    LLT DstTy = MRI.getType(DstReg);
 
    // If SizeOp0 is not an exact multiple of NarrowSize, emit
    // G_ANYEXT(G_IMPLICIT_DEF). Cast result to vector if needed.
    // FIXME: Although this would also be legal for the general case, it causes
    //  a lot of regressions in the emitted code (superfluous COPYs, artifact
    //  combines not being hit). This seems to be a problem related to the
    //  artifact combiner.
    if (SizeOp0 % NarrowSize != 0) {
      LLT ImplicitTy = NarrowTy;
      if (DstTy.isVector())
        ImplicitTy = LLT::vector(DstTy.getElementCount(), ImplicitTy);
 
      Register ImplicitReg = MIRBuilder.buildUndef(ImplicitTy).getReg(0);
      MIRBuilder.buildAnyExt(DstReg, ImplicitReg);
 
      MI.eraseFromParent();
      return Legalized;
    }
 
    int NumParts = SizeOp0 / NarrowSize;
 
    SmallVector<Register, 2> DstRegs;
    for (int i = 0; i < NumParts; ++i)
      DstRegs.push_back(MIRBuilder.buildUndef(NarrowTy).getReg(0));
 
    if (DstTy.isVector())
      MIRBuilder.buildBuildVector(DstReg, DstRegs);
    else
      MIRBuilder.buildMergeLikeInstr(DstReg, DstRegs);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_CONSTANT: {
    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
    const APInt &Val = MI.getOperand(1).getCImm()->getValue();
    unsigned TotalSize = Ty.getSizeInBits();
    unsigned NarrowSize = NarrowTy.getSizeInBits();
    int NumParts = TotalSize / NarrowSize;
 
    SmallVector<Register, 4> PartRegs;
    for (int I = 0; I != NumParts; ++I) {
      unsigned Offset = I * NarrowSize;
      auto K = MIRBuilder.buildConstant(NarrowTy,
                                        Val.lshr(Offset).trunc(NarrowSize));
      PartRegs.push_back(K.getReg(0));
    }
 
    LLT LeftoverTy;
    unsigned LeftoverBits = TotalSize - NumParts * NarrowSize;
    SmallVector<Register, 1> LeftoverRegs;
    if (LeftoverBits != 0) {
      LeftoverTy = LLT::scalar(LeftoverBits);
      auto K = MIRBuilder.buildConstant(
        LeftoverTy,
        Val.lshr(NumParts * NarrowSize).trunc(LeftoverBits));
      LeftoverRegs.push_back(K.getReg(0));
    }
 
    insertParts(MI.getOperand(0).getReg(),
                Ty, NarrowTy, PartRegs, LeftoverTy, LeftoverRegs);
 
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_SEXT:
  case TargetOpcode::G_ZEXT:
  case TargetOpcode::G_ANYEXT:
    return narrowScalarExt(MI, TypeIdx, NarrowTy);
  case TargetOpcode::G_TRUNC: {
    if (TypeIdx != 1)
      return UnableToLegalize;
 
    uint64_t SizeOp1 = MRI.getType(MI.getOperand(1).getReg()).getSizeInBits();
    if (NarrowTy.getSizeInBits() * 2 != SizeOp1) {
      LLVM_DEBUG(dbgs() << "Can't narrow trunc to type " << NarrowTy << "\n");
      return UnableToLegalize;
    }
 
    auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1));
    MIRBuilder.buildCopy(MI.getOperand(0), Unmerge.getReg(0));
    MI.eraseFromParent();
    return Legalized;
  }
 
  case TargetOpcode::G_FREEZE: {
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
    // Should widen scalar first
    if (Ty.getSizeInBits() % NarrowTy.getSizeInBits() != 0)
      return UnableToLegalize;
 
    auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, MI.getOperand(1).getReg());
    SmallVector<Register, 8> Parts;
    for (unsigned i = 0; i < Unmerge->getNumDefs(); ++i) {
      Parts.push_back(
          MIRBuilder.buildFreeze(NarrowTy, Unmerge.getReg(i)).getReg(0));
    }
 
    MIRBuilder.buildMergeLikeInstr(MI.getOperand(0).getReg(), Parts);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_ADD:
  case TargetOpcode::G_SUB:
  case TargetOpcode::G_SADDO:
  case TargetOpcode::G_SSUBO:
  case TargetOpcode::G_SADDE:
  case TargetOpcode::G_SSUBE:
  case TargetOpcode::G_UADDO:
  case TargetOpcode::G_USUBO:
  case TargetOpcode::G_UADDE:
  case TargetOpcode::G_USUBE:
    return narrowScalarAddSub(MI, TypeIdx, NarrowTy);
  case TargetOpcode::G_MUL:
  case TargetOpcode::G_UMULH:
    return narrowScalarMul(MI, NarrowTy);
  case TargetOpcode::G_EXTRACT:
    return narrowScalarExtract(MI, TypeIdx, NarrowTy);
  case TargetOpcode::G_INSERT:
    return narrowScalarInsert(MI, TypeIdx, NarrowTy);
  case TargetOpcode::G_LOAD: {
    auto &LoadMI = cast<GLoad>(MI);
    Register DstReg = LoadMI.getDstReg();
    LLT DstTy = MRI.getType(DstReg);
    if (DstTy.isVector())
      return UnableToLegalize;
 
    if (8 * LoadMI.getMemSize() != DstTy.getSizeInBits()) {
      Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
      MIRBuilder.buildLoad(TmpReg, LoadMI.getPointerReg(), LoadMI.getMMO());
      MIRBuilder.buildAnyExt(DstReg, TmpReg);
      LoadMI.eraseFromParent();
      return Legalized;
    }
 
    return reduceLoadStoreWidth(LoadMI, TypeIdx, NarrowTy);
  }
  case TargetOpcode::G_ZEXTLOAD:
  case TargetOpcode::G_SEXTLOAD: {
    auto &LoadMI = cast<GExtLoad>(MI);
    Register DstReg = LoadMI.getDstReg();
    Register PtrReg = LoadMI.getPointerReg();
 
    Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
    auto &MMO = LoadMI.getMMO();
    unsigned MemSize = MMO.getSizeInBits();
 
    if (MemSize == NarrowSize) {
      MIRBuilder.buildLoad(TmpReg, PtrReg, MMO);
    } else if (MemSize < NarrowSize) {
      MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), TmpReg, PtrReg, MMO);
    } else if (MemSize > NarrowSize) {
      // FIXME: Need to split the load.
      return UnableToLegalize;
    }
 
    if (isa<GZExtLoad>(LoadMI))
      MIRBuilder.buildZExt(DstReg, TmpReg);
    else
      MIRBuilder.buildSExt(DstReg, TmpReg);
 
    LoadMI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_STORE: {
    auto &StoreMI = cast<GStore>(MI);
 
    Register SrcReg = StoreMI.getValueReg();
    LLT SrcTy = MRI.getType(SrcReg);
    if (SrcTy.isVector())
      return UnableToLegalize;
 
    int NumParts = SizeOp0 / NarrowSize;
    unsigned HandledSize = NumParts * NarrowTy.getSizeInBits();
    unsigned LeftoverBits = SrcTy.getSizeInBits() - HandledSize;
    if (SrcTy.isVector() && LeftoverBits != 0)
      return UnableToLegalize;
 
    if (8 * StoreMI.getMemSize() != SrcTy.getSizeInBits()) {
      Register TmpReg = MRI.createGenericVirtualRegister(NarrowTy);
      MIRBuilder.buildTrunc(TmpReg, SrcReg);
      MIRBuilder.buildStore(TmpReg, StoreMI.getPointerReg(), StoreMI.getMMO());
      StoreMI.eraseFromParent();
      return Legalized;
    }
 
    return reduceLoadStoreWidth(StoreMI, 0, NarrowTy);
  }
  case TargetOpcode::G_SELECT:
    return narrowScalarSelect(MI, TypeIdx, NarrowTy);
  case TargetOpcode::G_AND:
  case TargetOpcode::G_OR:
  case TargetOpcode::G_XOR: {
    // Legalize bitwise operation:
    // A = BinOp<Ty> B, C
    // into:
    // B1, ..., BN = G_UNMERGE_VALUES B
    // C1, ..., CN = G_UNMERGE_VALUES C
    // A1 = BinOp<Ty/N> B1, C2
    // ...
    // AN = BinOp<Ty/N> BN, CN
    // A = G_MERGE_VALUES A1, ..., AN
    return narrowScalarBasic(MI, TypeIdx, NarrowTy);
  }
  case TargetOpcode::G_SHL:
  case TargetOpcode::G_LSHR:
  case TargetOpcode::G_ASHR:
    return narrowScalarShift(MI, TypeIdx, NarrowTy);
  case TargetOpcode::G_CTLZ:
  case TargetOpcode::G_CTLZ_ZERO_UNDEF:
  case TargetOpcode::G_CTTZ:
  case TargetOpcode::G_CTTZ_ZERO_UNDEF:
  case TargetOpcode::G_CTPOP:
    if (TypeIdx == 1)
      switch (MI.getOpcode()) {
      case TargetOpcode::G_CTLZ:
      case TargetOpcode::G_CTLZ_ZERO_UNDEF:
        return narrowScalarCTLZ(MI, TypeIdx, NarrowTy);
      case TargetOpcode::G_CTTZ:
      case TargetOpcode::G_CTTZ_ZERO_UNDEF:
        return narrowScalarCTTZ(MI, TypeIdx, NarrowTy);
      case TargetOpcode::G_CTPOP:
        return narrowScalarCTPOP(MI, TypeIdx, NarrowTy);
      default:
        return UnableToLegalize;
      }
 
    Observer.changingInstr(MI);
    narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_INTTOPTR:
    if (TypeIdx != 1)
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    narrowScalarSrc(MI, NarrowTy, 1);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_PTRTOINT:
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_ZEXT);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_PHI: {
    // FIXME: add support for when SizeOp0 isn't an exact multiple of
    // NarrowSize.
    if (SizeOp0 % NarrowSize != 0)
      return UnableToLegalize;
 
    unsigned NumParts = SizeOp0 / NarrowSize;
    SmallVector<Register, 2> DstRegs(NumParts);
    SmallVector<SmallVector<Register, 2>, 2> SrcRegs(MI.getNumOperands() / 2);
    Observer.changingInstr(MI);
    for (unsigned i = 1; i < MI.getNumOperands(); i += 2) {
      MachineBasicBlock &OpMBB = *MI.getOperand(i + 1).getMBB();
      MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminatorForward());
      extractParts(MI.getOperand(i).getReg(), NarrowTy, NumParts,
                   SrcRegs[i / 2]);
    }
    MachineBasicBlock &MBB = *MI.getParent();
    MIRBuilder.setInsertPt(MBB, MI);
    for (unsigned i = 0; i < NumParts; ++i) {
      DstRegs[i] = MRI.createGenericVirtualRegister(NarrowTy);
      MachineInstrBuilder MIB =
          MIRBuilder.buildInstr(TargetOpcode::G_PHI).addDef(DstRegs[i]);
      for (unsigned j = 1; j < MI.getNumOperands(); j += 2)
        MIB.addUse(SrcRegs[j / 2][i]).add(MI.getOperand(j + 1));
    }
    MIRBuilder.setInsertPt(MBB, MBB.getFirstNonPHI());
    MIRBuilder.buildMergeLikeInstr(MI.getOperand(0), DstRegs);
    Observer.changedInstr(MI);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_EXTRACT_VECTOR_ELT:
  case TargetOpcode::G_INSERT_VECTOR_ELT: {
    if (TypeIdx != 2)
      return UnableToLegalize;
 
    int OpIdx = MI.getOpcode() == TargetOpcode::G_EXTRACT_VECTOR_ELT ? 2 : 3;
    Observer.changingInstr(MI);
    narrowScalarSrc(MI, NarrowTy, OpIdx);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_ICMP: {
    Register LHS = MI.getOperand(2).getReg();
    LLT SrcTy = MRI.getType(LHS);
    uint64_t SrcSize = SrcTy.getSizeInBits();
    CmpInst::Predicate Pred =
        static_cast<CmpInst::Predicate>(MI.getOperand(1).getPredicate());
 
    // TODO: Handle the non-equality case for weird sizes.
    if (NarrowSize * 2 != SrcSize && !ICmpInst::isEquality(Pred))
      return UnableToLegalize;
 
    LLT LeftoverTy; // Example: s88 -> s64 (NarrowTy) + s24 (leftover)
    SmallVector<Register, 4> LHSPartRegs, LHSLeftoverRegs;
    if (!extractParts(LHS, SrcTy, NarrowTy, LeftoverTy, LHSPartRegs,
                      LHSLeftoverRegs))
      return UnableToLegalize;
 
    LLT Unused; // Matches LeftoverTy; G_ICMP LHS and RHS are the same type.
    SmallVector<Register, 4> RHSPartRegs, RHSLeftoverRegs;
    if (!extractParts(MI.getOperand(3).getReg(), SrcTy, NarrowTy, Unused,
                      RHSPartRegs, RHSLeftoverRegs))
      return UnableToLegalize;
 
    // We now have the LHS and RHS of the compare split into narrow-type
    // registers, plus potentially some leftover type.
    Register Dst = MI.getOperand(0).getReg();
    LLT ResTy = MRI.getType(Dst);
    if (ICmpInst::isEquality(Pred)) {
      // For each part on the LHS and RHS, keep track of the result of XOR-ing
      // them together. For each equal part, the result should be all 0s. For
      // each non-equal part, we'll get at least one 1.
      auto Zero = MIRBuilder.buildConstant(NarrowTy, 0);
      SmallVector<Register, 4> Xors;
      for (auto LHSAndRHS : zip(LHSPartRegs, RHSPartRegs)) {
        auto LHS = std::get<0>(LHSAndRHS);
        auto RHS = std::get<1>(LHSAndRHS);
        auto Xor = MIRBuilder.buildXor(NarrowTy, LHS, RHS).getReg(0);
        Xors.push_back(Xor);
      }
 
      // Build a G_XOR for each leftover register. Each G_XOR must be widened
      // to the desired narrow type so that we can OR them together later.
      SmallVector<Register, 4> WidenedXors;
      for (auto LHSAndRHS : zip(LHSLeftoverRegs, RHSLeftoverRegs)) {
        auto LHS = std::get<0>(LHSAndRHS);
        auto RHS = std::get<1>(LHSAndRHS);
        auto Xor = MIRBuilder.buildXor(LeftoverTy, LHS, RHS).getReg(0);
        LLT GCDTy = extractGCDType(WidenedXors, NarrowTy, LeftoverTy, Xor);
        buildLCMMergePieces(LeftoverTy, NarrowTy, GCDTy, WidenedXors,
                            /* PadStrategy = */ TargetOpcode::G_ZEXT);
        Xors.insert(Xors.end(), WidenedXors.begin(), WidenedXors.end());
      }
 
      // Now, for each part we broke up, we know if they are equal/not equal
      // based off the G_XOR. We can OR these all together and compare against
      // 0 to get the result.
      assert(Xors.size() >= 2 && "Should have gotten at least two Xors?");
      auto Or = MIRBuilder.buildOr(NarrowTy, Xors[0], Xors[1]);
      for (unsigned I = 2, E = Xors.size(); I < E; ++I)
        Or = MIRBuilder.buildOr(NarrowTy, Or, Xors[I]);
      MIRBuilder.buildICmp(Pred, Dst, Or, Zero);
    } else {
      // TODO: Handle non-power-of-two types.
      assert(LHSPartRegs.size() == 2 && "Expected exactly 2 LHS part regs?");
      assert(RHSPartRegs.size() == 2 && "Expected exactly 2 RHS part regs?");
      Register LHSL = LHSPartRegs[0];
      Register LHSH = LHSPartRegs[1];
      Register RHSL = RHSPartRegs[0];
      Register RHSH = RHSPartRegs[1];
      MachineInstrBuilder CmpH = MIRBuilder.buildICmp(Pred, ResTy, LHSH, RHSH);
      MachineInstrBuilder CmpHEQ =
          MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_EQ, ResTy, LHSH, RHSH);
      MachineInstrBuilder CmpLU = MIRBuilder.buildICmp(
          ICmpInst::getUnsignedPredicate(Pred), ResTy, LHSL, RHSL);
      MIRBuilder.buildSelect(Dst, CmpHEQ, CmpLU, CmpH);
    }
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_SEXT_INREG: {
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    int64_t SizeInBits = MI.getOperand(2).getImm();
 
    // So long as the new type has more bits than the bits we're extending we
    // don't need to break it apart.
    if (NarrowTy.getScalarSizeInBits() >= SizeInBits) {
      Observer.changingInstr(MI);
      // We don't lose any non-extension bits by truncating the src and
      // sign-extending the dst.
      MachineOperand &MO1 = MI.getOperand(1);
      auto TruncMIB = MIRBuilder.buildTrunc(NarrowTy, MO1);
      MO1.setReg(TruncMIB.getReg(0));
 
      MachineOperand &MO2 = MI.getOperand(0);
      Register DstExt = MRI.createGenericVirtualRegister(NarrowTy);
      MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
      MIRBuilder.buildSExt(MO2, DstExt);
      MO2.setReg(DstExt);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    // Break it apart. Components below the extension point are unmodified. The
    // component containing the extension point becomes a narrower SEXT_INREG.
    // Components above it are ashr'd from the component containing the
    // extension point.
    if (SizeOp0 % NarrowSize != 0)
      return UnableToLegalize;
    int NumParts = SizeOp0 / NarrowSize;
 
    // List the registers where the destination will be scattered.
    SmallVector<Register, 2> DstRegs;
    // List the registers where the source will be split.
    SmallVector<Register, 2> SrcRegs;
 
    // Create all the temporary registers.
    for (int i = 0; i < NumParts; ++i) {
      Register SrcReg = MRI.createGenericVirtualRegister(NarrowTy);
 
      SrcRegs.push_back(SrcReg);
    }
 
    // Explode the big arguments into smaller chunks.
    MIRBuilder.buildUnmerge(SrcRegs, MI.getOperand(1));
 
    Register AshrCstReg =
        MIRBuilder.buildConstant(NarrowTy, NarrowTy.getScalarSizeInBits() - 1)
            .getReg(0);
    Register FullExtensionReg = 0;
    Register PartialExtensionReg = 0;
 
    // Do the operation on each small part.
    for (int i = 0; i < NumParts; ++i) {
      if ((i + 1) * NarrowTy.getScalarSizeInBits() < SizeInBits)
        DstRegs.push_back(SrcRegs[i]);
      else if (i * NarrowTy.getScalarSizeInBits() > SizeInBits) {
        assert(PartialExtensionReg &&
               "Expected to visit partial extension before full");
        if (FullExtensionReg) {
          DstRegs.push_back(FullExtensionReg);
          continue;
        }
        DstRegs.push_back(
            MIRBuilder.buildAShr(NarrowTy, PartialExtensionReg, AshrCstReg)
                .getReg(0));
        FullExtensionReg = DstRegs.back();
      } else {
        DstRegs.push_back(
            MIRBuilder
                .buildInstr(
                    TargetOpcode::G_SEXT_INREG, {NarrowTy},
                    {SrcRegs[i], SizeInBits % NarrowTy.getScalarSizeInBits()})
                .getReg(0));
        PartialExtensionReg = DstRegs.back();
      }
    }
 
    // Gather the destination registers into the final destination.
    Register DstReg = MI.getOperand(0).getReg();
    MIRBuilder.buildMergeLikeInstr(DstReg, DstRegs);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_BSWAP:
  case TargetOpcode::G_BITREVERSE: {
    if (SizeOp0 % NarrowSize != 0)
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    SmallVector<Register, 2> SrcRegs, DstRegs;
    unsigned NumParts = SizeOp0 / NarrowSize;
    extractParts(MI.getOperand(1).getReg(), NarrowTy, NumParts, SrcRegs);
 
    for (unsigned i = 0; i < NumParts; ++i) {
      auto DstPart = MIRBuilder.buildInstr(MI.getOpcode(), {NarrowTy},
                                           {SrcRegs[NumParts - 1 - i]});
      DstRegs.push_back(DstPart.getReg(0));
    }
 
    MIRBuilder.buildMergeLikeInstr(MI.getOperand(0), DstRegs);
 
    Observer.changedInstr(MI);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_PTR_ADD:
  case TargetOpcode::G_PTRMASK: {
    if (TypeIdx != 1)
      return UnableToLegalize;
    Observer.changingInstr(MI);
    narrowScalarSrc(MI, NarrowTy, 2);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_FPTOUI:
  case TargetOpcode::G_FPTOSI:
    return narrowScalarFPTOI(MI, TypeIdx, NarrowTy);
  case TargetOpcode::G_FPEXT:
    if (TypeIdx != 0)
      return UnableToLegalize;
    Observer.changingInstr(MI);
    narrowScalarDst(MI, NarrowTy, 0, TargetOpcode::G_FPEXT);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_FLDEXP:
  case TargetOpcode::G_STRICT_FLDEXP:
    return narrowScalarFLDEXP(MI, TypeIdx, NarrowTy);
  }
}
 
Register LegalizerHelper::coerceToScalar(Register Val) {
  LLT Ty = MRI.getType(Val);
  if (Ty.isScalar())
    return Val;
 
  const DataLayout &DL = MIRBuilder.getDataLayout();
  LLT NewTy = LLT::scalar(Ty.getSizeInBits());
  if (Ty.isPointer()) {
    if (DL.isNonIntegralAddressSpace(Ty.getAddressSpace()))
      return Register();
    return MIRBuilder.buildPtrToInt(NewTy, Val).getReg(0);
  }
 
  Register NewVal = Val;
 
  assert(Ty.isVector());
  LLT EltTy = Ty.getElementType();
  if (EltTy.isPointer())
    NewVal = MIRBuilder.buildPtrToInt(NewTy, NewVal).getReg(0);
  return MIRBuilder.buildBitcast(NewTy, NewVal).getReg(0);
}
 
void LegalizerHelper::widenScalarSrc(MachineInstr &MI, LLT WideTy,
                                     unsigned OpIdx, unsigned ExtOpcode) {
  MachineOperand &MO = MI.getOperand(OpIdx);
  auto ExtB = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MO});
  MO.setReg(ExtB.getReg(0));
}
 
void LegalizerHelper::narrowScalarSrc(MachineInstr &MI, LLT NarrowTy,
                                      unsigned OpIdx) {
  MachineOperand &MO = MI.getOperand(OpIdx);
  auto ExtB = MIRBuilder.buildTrunc(NarrowTy, MO);
  MO.setReg(ExtB.getReg(0));
}
 
void LegalizerHelper::widenScalarDst(MachineInstr &MI, LLT WideTy,
                                     unsigned OpIdx, unsigned TruncOpcode) {
  MachineOperand &MO = MI.getOperand(OpIdx);
  Register DstExt = MRI.createGenericVirtualRegister(WideTy);
  MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
  MIRBuilder.buildInstr(TruncOpcode, {MO}, {DstExt});
  MO.setReg(DstExt);
}
 
void LegalizerHelper::narrowScalarDst(MachineInstr &MI, LLT NarrowTy,
                                      unsigned OpIdx, unsigned ExtOpcode) {
  MachineOperand &MO = MI.getOperand(OpIdx);
  Register DstTrunc = MRI.createGenericVirtualRegister(NarrowTy);
  MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
  MIRBuilder.buildInstr(ExtOpcode, {MO}, {DstTrunc});
  MO.setReg(DstTrunc);
}
 
void LegalizerHelper::moreElementsVectorDst(MachineInstr &MI, LLT WideTy,
                                            unsigned OpIdx) {
  MachineOperand &MO = MI.getOperand(OpIdx);
  MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
  Register Dst = MO.getReg();
  Register DstExt = MRI.createGenericVirtualRegister(WideTy);
  MO.setReg(DstExt);
  MIRBuilder.buildDeleteTrailingVectorElements(Dst, DstExt);
}
 
void LegalizerHelper::moreElementsVectorSrc(MachineInstr &MI, LLT MoreTy,
                                            unsigned OpIdx) {
  MachineOperand &MO = MI.getOperand(OpIdx);
  SmallVector<Register, 8> Regs;
  MO.setReg(MIRBuilder.buildPadVectorWithUndefElements(MoreTy, MO).getReg(0));
}
 
void LegalizerHelper::bitcastSrc(MachineInstr &MI, LLT CastTy, unsigned OpIdx) {
  MachineOperand &Op = MI.getOperand(OpIdx);
  Op.setReg(MIRBuilder.buildBitcast(CastTy, Op).getReg(0));
}
 
void LegalizerHelper::bitcastDst(MachineInstr &MI, LLT CastTy, unsigned OpIdx) {
  MachineOperand &MO = MI.getOperand(OpIdx);
  Register CastDst = MRI.createGenericVirtualRegister(CastTy);
  MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
  MIRBuilder.buildBitcast(MO, CastDst);
  MO.setReg(CastDst);
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarMergeValues(MachineInstr &MI, unsigned TypeIdx,
                                        LLT WideTy) {
  if (TypeIdx != 1)
    return UnableToLegalize;
 
  auto [DstReg, DstTy, Src1Reg, Src1Ty] = MI.getFirst2RegLLTs();
  if (DstTy.isVector())
    return UnableToLegalize;
 
  LLT SrcTy = MRI.getType(Src1Reg);
  const int DstSize = DstTy.getSizeInBits();
  const int SrcSize = SrcTy.getSizeInBits();
  const int WideSize = WideTy.getSizeInBits();
  const int NumMerge = (DstSize + WideSize - 1) / WideSize;
 
  unsigned NumOps = MI.getNumOperands();
  unsigned NumSrc = MI.getNumOperands() - 1;
  unsigned PartSize = DstTy.getSizeInBits() / NumSrc;
 
  if (WideSize >= DstSize) {
    // Directly pack the bits in the target type.
    Register ResultReg = MIRBuilder.buildZExt(WideTy, Src1Reg).getReg(0);
 
    for (unsigned I = 2; I != NumOps; ++I) {
      const unsigned Offset = (I - 1) * PartSize;
 
      Register SrcReg = MI.getOperand(I).getReg();
      assert(MRI.getType(SrcReg) == LLT::scalar(PartSize));
 
      auto ZextInput = MIRBuilder.buildZExt(WideTy, SrcReg);
 
      Register NextResult = I + 1 == NumOps && WideTy == DstTy ? DstReg :
        MRI.createGenericVirtualRegister(WideTy);
 
      auto ShiftAmt = MIRBuilder.buildConstant(WideTy, Offset);
      auto Shl = MIRBuilder.buildShl(WideTy, ZextInput, ShiftAmt);
      MIRBuilder.buildOr(NextResult, ResultReg, Shl);
      ResultReg = NextResult;
    }
 
    if (WideSize > DstSize)
      MIRBuilder.buildTrunc(DstReg, ResultReg);
    else if (DstTy.isPointer())
      MIRBuilder.buildIntToPtr(DstReg, ResultReg);
 
    MI.eraseFromParent();
    return Legalized;
  }
 
  // Unmerge the original values to the GCD type, and recombine to the next
  // multiple greater than the original type.
  //
  // %3:_(s12) = G_MERGE_VALUES %0:_(s4), %1:_(s4), %2:_(s4) -> s6
  // %4:_(s2), %5:_(s2) = G_UNMERGE_VALUES %0
  // %6:_(s2), %7:_(s2) = G_UNMERGE_VALUES %1
  // %8:_(s2), %9:_(s2) = G_UNMERGE_VALUES %2
  // %10:_(s6) = G_MERGE_VALUES %4, %5, %6
  // %11:_(s6) = G_MERGE_VALUES %7, %8, %9
  // %12:_(s12) = G_MERGE_VALUES %10, %11
  //
  // Padding with undef if necessary:
  //
  // %2:_(s8) = G_MERGE_VALUES %0:_(s4), %1:_(s4) -> s6
  // %3:_(s2), %4:_(s2) = G_UNMERGE_VALUES %0
  // %5:_(s2), %6:_(s2) = G_UNMERGE_VALUES %1
  // %7:_(s2) = G_IMPLICIT_DEF
  // %8:_(s6) = G_MERGE_VALUES %3, %4, %5
  // %9:_(s6) = G_MERGE_VALUES %6, %7, %7
  // %10:_(s12) = G_MERGE_VALUES %8, %9
 
  const int GCD = std::gcd(SrcSize, WideSize);
  LLT GCDTy = LLT::scalar(GCD);
 
  SmallVector<Register, 8> Parts;
  SmallVector<Register, 8> NewMergeRegs;
  SmallVector<Register, 8> Unmerges;
  LLT WideDstTy = LLT::scalar(NumMerge * WideSize);
 
  // Decompose the original operands if they don't evenly divide.
  for (const MachineOperand &MO : llvm::drop_begin(MI.operands())) {
    Register SrcReg = MO.getReg();
    if (GCD == SrcSize) {
      Unmerges.push_back(SrcReg);
    } else {
      auto Unmerge = MIRBuilder.buildUnmerge(GCDTy, SrcReg);
      for (int J = 0, JE = Unmerge->getNumOperands() - 1; J != JE; ++J)
        Unmerges.push_back(Unmerge.getReg(J));
    }
  }
 
  // Pad with undef to the next size that is a multiple of the requested size.
  if (static_cast<int>(Unmerges.size()) != NumMerge * WideSize) {
    Register UndefReg = MIRBuilder.buildUndef(GCDTy).getReg(0);
    for (int I = Unmerges.size(); I != NumMerge * WideSize; ++I)
      Unmerges.push_back(UndefReg);
  }
 
  const int PartsPerGCD = WideSize / GCD;
 
  // Build merges of each piece.
  ArrayRef<Register> Slicer(Unmerges);
  for (int I = 0; I != NumMerge; ++I, Slicer = Slicer.drop_front(PartsPerGCD)) {
    auto Merge =
        MIRBuilder.buildMergeLikeInstr(WideTy, Slicer.take_front(PartsPerGCD));
    NewMergeRegs.push_back(Merge.getReg(0));
  }
 
  // A truncate may be necessary if the requested type doesn't evenly divide the
  // original result type.
  if (DstTy.getSizeInBits() == WideDstTy.getSizeInBits()) {
    MIRBuilder.buildMergeLikeInstr(DstReg, NewMergeRegs);
  } else {
    auto FinalMerge = MIRBuilder.buildMergeLikeInstr(WideDstTy, NewMergeRegs);
    MIRBuilder.buildTrunc(DstReg, FinalMerge.getReg(0));
  }
 
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarUnmergeValues(MachineInstr &MI, unsigned TypeIdx,
                                          LLT WideTy) {
  if (TypeIdx != 0)
    return UnableToLegalize;
 
  int NumDst = MI.getNumOperands() - 1;
  Register SrcReg = MI.getOperand(NumDst).getReg();
  LLT SrcTy = MRI.getType(SrcReg);
  if (SrcTy.isVector())
    return UnableToLegalize;
 
  Register Dst0Reg = MI.getOperand(0).getReg();
  LLT DstTy = MRI.getType(Dst0Reg);
  if (!DstTy.isScalar())
    return UnableToLegalize;
 
  if (WideTy.getSizeInBits() >= SrcTy.getSizeInBits()) {
    if (SrcTy.isPointer()) {
      const DataLayout &DL = MIRBuilder.getDataLayout();
      if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace())) {
        LLVM_DEBUG(
            dbgs() << "Not casting non-integral address space integer\n");
        return UnableToLegalize;
      }
 
      SrcTy = LLT::scalar(SrcTy.getSizeInBits());
      SrcReg = MIRBuilder.buildPtrToInt(SrcTy, SrcReg).getReg(0);
    }
 
    // Widen SrcTy to WideTy. This does not affect the result, but since the
    // user requested this size, it is probably better handled than SrcTy and
    // should reduce the total number of legalization artifacts.
    if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) {
      SrcTy = WideTy;
      SrcReg = MIRBuilder.buildAnyExt(WideTy, SrcReg).getReg(0);
    }
 
    // Theres no unmerge type to target. Directly extract the bits from the
    // source type
    unsigned DstSize = DstTy.getSizeInBits();
 
    MIRBuilder.buildTrunc(Dst0Reg, SrcReg);
    for (int I = 1; I != NumDst; ++I) {
      auto ShiftAmt = MIRBuilder.buildConstant(SrcTy, DstSize * I);
      auto Shr = MIRBuilder.buildLShr(SrcTy, SrcReg, ShiftAmt);
      MIRBuilder.buildTrunc(MI.getOperand(I), Shr);
    }
 
    MI.eraseFromParent();
    return Legalized;
  }
 
  // Extend the source to a wider type.
  LLT LCMTy = getLCMType(SrcTy, WideTy);
 
  Register WideSrc = SrcReg;
  if (LCMTy.getSizeInBits() != SrcTy.getSizeInBits()) {
    // TODO: If this is an integral address space, cast to integer and anyext.
    if (SrcTy.isPointer()) {
      LLVM_DEBUG(dbgs() << "Widening pointer source types not implemented\n");
      return UnableToLegalize;
    }
 
    WideSrc = MIRBuilder.buildAnyExt(LCMTy, WideSrc).getReg(0);
  }
 
  auto Unmerge = MIRBuilder.buildUnmerge(WideTy, WideSrc);
 
  // Create a sequence of unmerges and merges to the original results. Since we
  // may have widened the source, we will need to pad the results with dead defs
  // to cover the source register.
  // e.g. widen s48 to s64:
  // %1:_(s48), %2:_(s48) = G_UNMERGE_VALUES %0:_(s96)
  //
  // =>
  //  %4:_(s192) = G_ANYEXT %0:_(s96)
  //  %5:_(s64), %6, %7 = G_UNMERGE_VALUES %4 ; Requested unmerge
  //  ; unpack to GCD type, with extra dead defs
  //  %8:_(s16), %9, %10, %11 = G_UNMERGE_VALUES %5:_(s64)
  //  %12:_(s16), %13, dead %14, dead %15 = G_UNMERGE_VALUES %6:_(s64)
  //  dead %16:_(s16), dead %17, dead %18, dead %18 = G_UNMERGE_VALUES %7:_(s64)
  //  %1:_(s48) = G_MERGE_VALUES %8:_(s16), %9, %10   ; Remerge to destination
  //  %2:_(s48) = G_MERGE_VALUES %11:_(s16), %12, %13 ; Remerge to destination
  const LLT GCDTy = getGCDType(WideTy, DstTy);
  const int NumUnmerge = Unmerge->getNumOperands() - 1;
  const int PartsPerRemerge = DstTy.getSizeInBits() / GCDTy.getSizeInBits();
 
  // Directly unmerge to the destination without going through a GCD type
  // if possible
  if (PartsPerRemerge == 1) {
    const int PartsPerUnmerge = WideTy.getSizeInBits() / DstTy.getSizeInBits();
 
    for (int I = 0; I != NumUnmerge; ++I) {
      auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES);
 
      for (int J = 0; J != PartsPerUnmerge; ++J) {
        int Idx = I * PartsPerUnmerge + J;
        if (Idx < NumDst)
          MIB.addDef(MI.getOperand(Idx).getReg());
        else {
          // Create dead def for excess components.
          MIB.addDef(MRI.createGenericVirtualRegister(DstTy));
        }
      }
 
      MIB.addUse(Unmerge.getReg(I));
    }
  } else {
    SmallVector<Register, 16> Parts;
    for (int J = 0; J != NumUnmerge; ++J)
      extractGCDType(Parts, GCDTy, Unmerge.getReg(J));
 
    SmallVector<Register, 8> RemergeParts;
    for (int I = 0; I != NumDst; ++I) {
      for (int J = 0; J < PartsPerRemerge; ++J) {
        const int Idx = I * PartsPerRemerge + J;
        RemergeParts.emplace_back(Parts[Idx]);
      }
 
      MIRBuilder.buildMergeLikeInstr(MI.getOperand(I).getReg(), RemergeParts);
      RemergeParts.clear();
    }
  }
 
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarExtract(MachineInstr &MI, unsigned TypeIdx,
                                    LLT WideTy) {
  auto [DstReg, DstTy, SrcReg, SrcTy] = MI.getFirst2RegLLTs();
  unsigned Offset = MI.getOperand(2).getImm();
 
  if (TypeIdx == 0) {
    if (SrcTy.isVector() || DstTy.isVector())
      return UnableToLegalize;
 
    SrcOp Src(SrcReg);
    if (SrcTy.isPointer()) {
      // Extracts from pointers can be handled only if they are really just
      // simple integers.
      const DataLayout &DL = MIRBuilder.getDataLayout();
      if (DL.isNonIntegralAddressSpace(SrcTy.getAddressSpace()))
        return UnableToLegalize;
 
      LLT SrcAsIntTy = LLT::scalar(SrcTy.getSizeInBits());
      Src = MIRBuilder.buildPtrToInt(SrcAsIntTy, Src);
      SrcTy = SrcAsIntTy;
    }
 
    if (DstTy.isPointer())
      return UnableToLegalize;
 
    if (Offset == 0) {
      // Avoid a shift in the degenerate case.
      MIRBuilder.buildTrunc(DstReg,
                            MIRBuilder.buildAnyExtOrTrunc(WideTy, Src));
      MI.eraseFromParent();
      return Legalized;
    }
 
    // Do a shift in the source type.
    LLT ShiftTy = SrcTy;
    if (WideTy.getSizeInBits() > SrcTy.getSizeInBits()) {
      Src = MIRBuilder.buildAnyExt(WideTy, Src);
      ShiftTy = WideTy;
    }
 
    auto LShr = MIRBuilder.buildLShr(
      ShiftTy, Src, MIRBuilder.buildConstant(ShiftTy, Offset));
    MIRBuilder.buildTrunc(DstReg, LShr);
    MI.eraseFromParent();
    return Legalized;
  }
 
  if (SrcTy.isScalar()) {
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
    Observer.changedInstr(MI);
    return Legalized;
  }
 
  if (!SrcTy.isVector())
    return UnableToLegalize;
 
  if (DstTy != SrcTy.getElementType())
    return UnableToLegalize;
 
  if (Offset % SrcTy.getScalarSizeInBits() != 0)
    return UnableToLegalize;
 
  Observer.changingInstr(MI);
  widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
 
  MI.getOperand(2).setImm((WideTy.getSizeInBits() / SrcTy.getSizeInBits()) *
                          Offset);
  widenScalarDst(MI, WideTy.getScalarType(), 0);
  Observer.changedInstr(MI);
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarInsert(MachineInstr &MI, unsigned TypeIdx,
                                   LLT WideTy) {
  if (TypeIdx != 0 || WideTy.isVector())
    return UnableToLegalize;
  Observer.changingInstr(MI);
  widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
  widenScalarDst(MI, WideTy);
  Observer.changedInstr(MI);
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarAddSubOverflow(MachineInstr &MI, unsigned TypeIdx,
                                           LLT WideTy) {
  unsigned Opcode;
  unsigned ExtOpcode;
  std::optional<Register> CarryIn;
  switch (MI.getOpcode()) {
  default:
    llvm_unreachable("Unexpected opcode!");
  case TargetOpcode::G_SADDO:
    Opcode = TargetOpcode::G_ADD;
    ExtOpcode = TargetOpcode::G_SEXT;
    break;
  case TargetOpcode::G_SSUBO:
    Opcode = TargetOpcode::G_SUB;
    ExtOpcode = TargetOpcode::G_SEXT;
    break;
  case TargetOpcode::G_UADDO:
    Opcode = TargetOpcode::G_ADD;
    ExtOpcode = TargetOpcode::G_ZEXT;
    break;
  case TargetOpcode::G_USUBO:
    Opcode = TargetOpcode::G_SUB;
    ExtOpcode = TargetOpcode::G_ZEXT;
    break;
  case TargetOpcode::G_SADDE:
    Opcode = TargetOpcode::G_UADDE;
    ExtOpcode = TargetOpcode::G_SEXT;
    CarryIn = MI.getOperand(4).getReg();
    break;
  case TargetOpcode::G_SSUBE:
    Opcode = TargetOpcode::G_USUBE;
    ExtOpcode = TargetOpcode::G_SEXT;
    CarryIn = MI.getOperand(4).getReg();
    break;
  case TargetOpcode::G_UADDE:
    Opcode = TargetOpcode::G_UADDE;
    ExtOpcode = TargetOpcode::G_ZEXT;
    CarryIn = MI.getOperand(4).getReg();
    break;
  case TargetOpcode::G_USUBE:
    Opcode = TargetOpcode::G_USUBE;
    ExtOpcode = TargetOpcode::G_ZEXT;
    CarryIn = MI.getOperand(4).getReg();
    break;
  }
 
  if (TypeIdx == 1) {
    unsigned BoolExtOp = MIRBuilder.getBoolExtOp(WideTy.isVector(), false);
 
    Observer.changingInstr(MI);
    if (CarryIn)
      widenScalarSrc(MI, WideTy, 4, BoolExtOp);
    widenScalarDst(MI, WideTy, 1);
 
    Observer.changedInstr(MI);
    return Legalized;
  }
 
  auto LHSExt = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MI.getOperand(2)});
  auto RHSExt = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {MI.getOperand(3)});
  // Do the arithmetic in the larger type.
  Register NewOp;
  if (CarryIn) {
    LLT CarryOutTy = MRI.getType(MI.getOperand(1).getReg());
    NewOp = MIRBuilder
                .buildInstr(Opcode, {WideTy, CarryOutTy},
                            {LHSExt, RHSExt, *CarryIn})
                .getReg(0);
  } else {
    NewOp = MIRBuilder.buildInstr(Opcode, {WideTy}, {LHSExt, RHSExt}).getReg(0);
  }
  LLT OrigTy = MRI.getType(MI.getOperand(0).getReg());
  auto TruncOp = MIRBuilder.buildTrunc(OrigTy, NewOp);
  auto ExtOp = MIRBuilder.buildInstr(ExtOpcode, {WideTy}, {TruncOp});
  // There is no overflow if the ExtOp is the same as NewOp.
  MIRBuilder.buildICmp(CmpInst::ICMP_NE, MI.getOperand(1), NewOp, ExtOp);
  // Now trunc the NewOp to the original result.
  MIRBuilder.buildTrunc(MI.getOperand(0), NewOp);
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarAddSubShlSat(MachineInstr &MI, unsigned TypeIdx,
                                         LLT WideTy) {
  bool IsSigned = MI.getOpcode() == TargetOpcode::G_SADDSAT ||
                  MI.getOpcode() == TargetOpcode::G_SSUBSAT ||
                  MI.getOpcode() == TargetOpcode::G_SSHLSAT;
  bool IsShift = MI.getOpcode() == TargetOpcode::G_SSHLSAT ||
                 MI.getOpcode() == TargetOpcode::G_USHLSAT;
  // We can convert this to:
  //   1. Any extend iN to iM
  //   2. SHL by M-N
  //   3. [US][ADD|SUB|SHL]SAT
  //   4. L/ASHR by M-N
  //
  // It may be more efficient to lower this to a min and a max operation in
  // the higher precision arithmetic if the promoted operation isn't legal,
  // but this decision is up to the target's lowering request.
  Register DstReg = MI.getOperand(0).getReg();
 
  unsigned NewBits = WideTy.getScalarSizeInBits();
  unsigned SHLAmount = NewBits - MRI.getType(DstReg).getScalarSizeInBits();
 
  // Shifts must zero-extend the RHS to preserve the unsigned quantity, and
  // must not left shift the RHS to preserve the shift amount.
  auto LHS = MIRBuilder.buildAnyExt(WideTy, MI.getOperand(1));
  auto RHS = IsShift ? MIRBuilder.buildZExt(WideTy, MI.getOperand(2))
                     : MIRBuilder.buildAnyExt(WideTy, MI.getOperand(2));
  auto ShiftK = MIRBuilder.buildConstant(WideTy, SHLAmount);
  auto ShiftL = MIRBuilder.buildShl(WideTy, LHS, ShiftK);
  auto ShiftR = IsShift ? RHS : MIRBuilder.buildShl(WideTy, RHS, ShiftK);
 
  auto WideInst = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy},
                                        {ShiftL, ShiftR}, MI.getFlags());
 
  // Use a shift that will preserve the number of sign bits when the trunc is
  // folded away.
  auto Result = IsSigned ? MIRBuilder.buildAShr(WideTy, WideInst, ShiftK)
                         : MIRBuilder.buildLShr(WideTy, WideInst, ShiftK);
 
  MIRBuilder.buildTrunc(DstReg, Result);
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalarMulo(MachineInstr &MI, unsigned TypeIdx,
                                 LLT WideTy) {
  if (TypeIdx == 1) {
    Observer.changingInstr(MI);
    widenScalarDst(MI, WideTy, 1);
    Observer.changedInstr(MI);
    return Legalized;
  }
 
  bool IsSigned = MI.getOpcode() == TargetOpcode::G_SMULO;
  auto [Result, OriginalOverflow, LHS, RHS] = MI.getFirst4Regs();
  LLT SrcTy = MRI.getType(LHS);
  LLT OverflowTy = MRI.getType(OriginalOverflow);
  unsigned SrcBitWidth = SrcTy.getScalarSizeInBits();
 
  // To determine if the result overflowed in the larger type, we extend the
  // input to the larger type, do the multiply (checking if it overflows),
  // then also check the high bits of the result to see if overflow happened
  // there.
  unsigned ExtOp = IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT;
  auto LeftOperand = MIRBuilder.buildInstr(ExtOp, {WideTy}, {LHS});
  auto RightOperand = MIRBuilder.buildInstr(ExtOp, {WideTy}, {RHS});
 
  auto Mulo = MIRBuilder.buildInstr(MI.getOpcode(), {WideTy, OverflowTy},
                                    {LeftOperand, RightOperand});
  auto Mul = Mulo->getOperand(0);
  MIRBuilder.buildTrunc(Result, Mul);
 
  MachineInstrBuilder ExtResult;
  // Overflow occurred if it occurred in the larger type, or if the high part
  // of the result does not zero/sign-extend the low part.  Check this second
  // possibility first.
  if (IsSigned) {
    // For signed, overflow occurred when the high part does not sign-extend
    // the low part.
    ExtResult = MIRBuilder.buildSExtInReg(WideTy, Mul, SrcBitWidth);
  } else {
    // Unsigned overflow occurred when the high part does not zero-extend the
    // low part.
    ExtResult = MIRBuilder.buildZExtInReg(WideTy, Mul, SrcBitWidth);
  }
 
  // Multiplication cannot overflow if the WideTy is >= 2 * original width,
  // so we don't need to check the overflow result of larger type Mulo.
  if (WideTy.getScalarSizeInBits() < 2 * SrcBitWidth) {
    auto Overflow =
        MIRBuilder.buildICmp(CmpInst::ICMP_NE, OverflowTy, Mul, ExtResult);
    // Finally check if the multiplication in the larger type itself overflowed.
    MIRBuilder.buildOr(OriginalOverflow, Mulo->getOperand(1), Overflow);
  } else {
    MIRBuilder.buildICmp(CmpInst::ICMP_NE, OriginalOverflow, Mul, ExtResult);
  }
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::widenScalar(MachineInstr &MI, unsigned TypeIdx, LLT WideTy) {
  switch (MI.getOpcode()) {
  default:
    return UnableToLegalize;
  case TargetOpcode::G_ATOMICRMW_XCHG:
  case TargetOpcode::G_ATOMICRMW_ADD:
  case TargetOpcode::G_ATOMICRMW_SUB:
  case TargetOpcode::G_ATOMICRMW_AND:
  case TargetOpcode::G_ATOMICRMW_OR:
  case TargetOpcode::G_ATOMICRMW_XOR:
  case TargetOpcode::G_ATOMICRMW_MIN:
  case TargetOpcode::G_ATOMICRMW_MAX:
  case TargetOpcode::G_ATOMICRMW_UMIN:
  case TargetOpcode::G_ATOMICRMW_UMAX:
    assert(TypeIdx == 0 && "atomicrmw with second scalar type");
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
    widenScalarDst(MI, WideTy, 0);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_ATOMIC_CMPXCHG:
    assert(TypeIdx == 0 && "G_ATOMIC_CMPXCHG with second scalar type");
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
    widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT);
    widenScalarDst(MI, WideTy, 0);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS:
    if (TypeIdx == 0) {
      Observer.changingInstr(MI);
      widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT);
      widenScalarSrc(MI, WideTy, 4, TargetOpcode::G_ANYEXT);
      widenScalarDst(MI, WideTy, 0);
      Observer.changedInstr(MI);
      return Legalized;
    }
    assert(TypeIdx == 1 &&
           "G_ATOMIC_CMPXCHG_WITH_SUCCESS with third scalar type");
    Observer.changingInstr(MI);
    widenScalarDst(MI, WideTy, 1);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_EXTRACT:
    return widenScalarExtract(MI, TypeIdx, WideTy);
  case TargetOpcode::G_INSERT:
    return widenScalarInsert(MI, TypeIdx, WideTy);
  case TargetOpcode::G_MERGE_VALUES:
    return widenScalarMergeValues(MI, TypeIdx, WideTy);
  case TargetOpcode::G_UNMERGE_VALUES:
    return widenScalarUnmergeValues(MI, TypeIdx, WideTy);
  case TargetOpcode::G_SADDO:
  case TargetOpcode::G_SSUBO:
  case TargetOpcode::G_UADDO:
  case TargetOpcode::G_USUBO:
  case TargetOpcode::G_SADDE:
  case TargetOpcode::G_SSUBE:
  case TargetOpcode::G_UADDE:
  case TargetOpcode::G_USUBE:
    return widenScalarAddSubOverflow(MI, TypeIdx, WideTy);
  case TargetOpcode::G_UMULO:
  case TargetOpcode::G_SMULO:
    return widenScalarMulo(MI, TypeIdx, WideTy);
  case TargetOpcode::G_SADDSAT:
  case TargetOpcode::G_SSUBSAT:
  case TargetOpcode::G_SSHLSAT:
  case TargetOpcode::G_UADDSAT:
  case TargetOpcode::G_USUBSAT:
  case TargetOpcode::G_USHLSAT:
    return widenScalarAddSubShlSat(MI, TypeIdx, WideTy);
  case TargetOpcode::G_CTTZ:
  case TargetOpcode::G_CTTZ_ZERO_UNDEF:
  case TargetOpcode::G_CTLZ:
  case TargetOpcode::G_CTLZ_ZERO_UNDEF:
  case TargetOpcode::G_CTPOP: {
    if (TypeIdx == 0) {
      Observer.changingInstr(MI);
      widenScalarDst(MI, WideTy, 0);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    Register SrcReg = MI.getOperand(1).getReg();
 
    // First extend the input.
    unsigned ExtOpc = MI.getOpcode() == TargetOpcode::G_CTTZ ||
                              MI.getOpcode() == TargetOpcode::G_CTTZ_ZERO_UNDEF
                          ? TargetOpcode::G_ANYEXT
                          : TargetOpcode::G_ZEXT;
    auto MIBSrc = MIRBuilder.buildInstr(ExtOpc, {WideTy}, {SrcReg});
    LLT CurTy = MRI.getType(SrcReg);
    unsigned NewOpc = MI.getOpcode();
    if (NewOpc == TargetOpcode::G_CTTZ) {
      // The count is the same in the larger type except if the original
      // value was zero.  This can be handled by setting the bit just off
      // the top of the original type.
      auto TopBit =
          APInt::getOneBitSet(WideTy.getSizeInBits(), CurTy.getSizeInBits());
      MIBSrc = MIRBuilder.buildOr(
        WideTy, MIBSrc, MIRBuilder.buildConstant(WideTy, TopBit));
      // Now we know the operand is non-zero, use the more relaxed opcode.
      NewOpc = TargetOpcode::G_CTTZ_ZERO_UNDEF;
    }
 
    // Perform the operation at the larger size.
    auto MIBNewOp = MIRBuilder.buildInstr(NewOpc, {WideTy}, {MIBSrc});
    // This is already the correct result for CTPOP and CTTZs
    if (MI.getOpcode() == TargetOpcode::G_CTLZ ||
        MI.getOpcode() == TargetOpcode::G_CTLZ_ZERO_UNDEF) {
      // The correct result is NewOp - (Difference in widety and current ty).
      unsigned SizeDiff = WideTy.getSizeInBits() - CurTy.getSizeInBits();
      MIBNewOp = MIRBuilder.buildSub(
          WideTy, MIBNewOp, MIRBuilder.buildConstant(WideTy, SizeDiff));
    }
 
    MIRBuilder.buildZExtOrTrunc(MI.getOperand(0), MIBNewOp);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_BSWAP: {
    Observer.changingInstr(MI);
    Register DstReg = MI.getOperand(0).getReg();
 
    Register ShrReg = MRI.createGenericVirtualRegister(WideTy);
    Register DstExt = MRI.createGenericVirtualRegister(WideTy);
    Register ShiftAmtReg = MRI.createGenericVirtualRegister(WideTy);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
 
    MI.getOperand(0).setReg(DstExt);
 
    MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
 
    LLT Ty = MRI.getType(DstReg);
    unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits();
    MIRBuilder.buildConstant(ShiftAmtReg, DiffBits);
    MIRBuilder.buildLShr(ShrReg, DstExt, ShiftAmtReg);
 
    MIRBuilder.buildTrunc(DstReg, ShrReg);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_BITREVERSE: {
    Observer.changingInstr(MI);
 
    Register DstReg = MI.getOperand(0).getReg();
    LLT Ty = MRI.getType(DstReg);
    unsigned DiffBits = WideTy.getScalarSizeInBits() - Ty.getScalarSizeInBits();
 
    Register DstExt = MRI.createGenericVirtualRegister(WideTy);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
    MI.getOperand(0).setReg(DstExt);
    MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
 
    auto ShiftAmt = MIRBuilder.buildConstant(WideTy, DiffBits);
    auto Shift = MIRBuilder.buildLShr(WideTy, DstExt, ShiftAmt);
    MIRBuilder.buildTrunc(DstReg, Shift);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_FREEZE:
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_ABS:
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_ADD:
  case TargetOpcode::G_AND:
  case TargetOpcode::G_MUL:
  case TargetOpcode::G_OR:
  case TargetOpcode::G_XOR:
  case TargetOpcode::G_SUB:
    // Perform operation at larger width (any extension is fines here, high bits
    // don't affect the result) and then truncate the result back to the
    // original type.
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_SBFX:
  case TargetOpcode::G_UBFX:
    Observer.changingInstr(MI);
 
    if (TypeIdx == 0) {
      widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
      widenScalarDst(MI, WideTy);
    } else {
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
      widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT);
    }
 
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_SHL:
    Observer.changingInstr(MI);
 
    if (TypeIdx == 0) {
      widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
      widenScalarDst(MI, WideTy);
    } else {
      assert(TypeIdx == 1);
      // The "number of bits to shift" operand must preserve its value as an
      // unsigned integer:
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
    }
 
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_SDIV:
  case TargetOpcode::G_SREM:
  case TargetOpcode::G_SMIN:
  case TargetOpcode::G_SMAX:
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_SDIVREM:
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
    widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT);
    widenScalarDst(MI, WideTy);
    widenScalarDst(MI, WideTy, 1);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_ASHR:
  case TargetOpcode::G_LSHR:
    Observer.changingInstr(MI);
 
    if (TypeIdx == 0) {
      unsigned CvtOp = MI.getOpcode() == TargetOpcode::G_ASHR ?
        TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT;
 
      widenScalarSrc(MI, WideTy, 1, CvtOp);
      widenScalarDst(MI, WideTy);
    } else {
      assert(TypeIdx == 1);
      // The "number of bits to shift" operand must preserve its value as an
      // unsigned integer:
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
    }
 
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_UDIV:
  case TargetOpcode::G_UREM:
  case TargetOpcode::G_UMIN:
  case TargetOpcode::G_UMAX:
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_UDIVREM:
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
    widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ZEXT);
    widenScalarDst(MI, WideTy);
    widenScalarDst(MI, WideTy, 1);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_SELECT:
    Observer.changingInstr(MI);
    if (TypeIdx == 0) {
      // Perform operation at larger width (any extension is fine here, high
      // bits don't affect the result) and then truncate the result back to the
      // original type.
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
      widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_ANYEXT);
      widenScalarDst(MI, WideTy);
    } else {
      bool IsVec = MRI.getType(MI.getOperand(1).getReg()).isVector();
      // Explicit extension is required here since high bits affect the result.
      widenScalarSrc(MI, WideTy, 1, MIRBuilder.getBoolExtOp(IsVec, false));
    }
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_FPTOSI:
  case TargetOpcode::G_FPTOUI:
    Observer.changingInstr(MI);
 
    if (TypeIdx == 0)
      widenScalarDst(MI, WideTy);
    else
      widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT);
 
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_SITOFP:
    Observer.changingInstr(MI);
 
    if (TypeIdx == 0)
      widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
    else
      widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_SEXT);
 
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_UITOFP:
    Observer.changingInstr(MI);
 
    if (TypeIdx == 0)
      widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
    else
      widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
 
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_LOAD:
  case TargetOpcode::G_SEXTLOAD:
  case TargetOpcode::G_ZEXTLOAD:
    Observer.changingInstr(MI);
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_STORE: {
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
    if (!Ty.isScalar())
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
 
    unsigned ExtType = Ty.getScalarSizeInBits() == 1 ?
      TargetOpcode::G_ZEXT : TargetOpcode::G_ANYEXT;
    widenScalarSrc(MI, WideTy, 0, ExtType);
 
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_CONSTANT: {
    MachineOperand &SrcMO = MI.getOperand(1);
    LLVMContext &Ctx = MIRBuilder.getMF().getFunction().getContext();
    unsigned ExtOpc = LI.getExtOpcodeForWideningConstant(
        MRI.getType(MI.getOperand(0).getReg()));
    assert((ExtOpc == TargetOpcode::G_ZEXT || ExtOpc == TargetOpcode::G_SEXT ||
            ExtOpc == TargetOpcode::G_ANYEXT) &&
           "Illegal Extend");
    const APInt &SrcVal = SrcMO.getCImm()->getValue();
    const APInt &Val = (ExtOpc == TargetOpcode::G_SEXT)
                           ? SrcVal.sext(WideTy.getSizeInBits())
                           : SrcVal.zext(WideTy.getSizeInBits());
    Observer.changingInstr(MI);
    SrcMO.setCImm(ConstantInt::get(Ctx, Val));
 
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_FCONSTANT: {
    // To avoid changing the bits of the constant due to extension to a larger
    // type and then using G_FPTRUNC, we simply convert to a G_CONSTANT.
    MachineOperand &SrcMO = MI.getOperand(1);
    APInt Val = SrcMO.getFPImm()->getValueAPF().bitcastToAPInt();
    MIRBuilder.setInstrAndDebugLoc(MI);
    auto IntCst = MIRBuilder.buildConstant(MI.getOperand(0).getReg(), Val);
    widenScalarDst(*IntCst, WideTy, 0, TargetOpcode::G_TRUNC);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_IMPLICIT_DEF: {
    Observer.changingInstr(MI);
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_BRCOND:
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 0, MIRBuilder.getBoolExtOp(false, false));
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_FCMP:
    Observer.changingInstr(MI);
    if (TypeIdx == 0)
      widenScalarDst(MI, WideTy);
    else {
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT);
      widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_FPEXT);
    }
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_ICMP:
    Observer.changingInstr(MI);
    if (TypeIdx == 0)
      widenScalarDst(MI, WideTy);
    else {
      unsigned ExtOpcode = CmpInst::isSigned(static_cast<CmpInst::Predicate>(
                               MI.getOperand(1).getPredicate()))
                               ? TargetOpcode::G_SEXT
                               : TargetOpcode::G_ZEXT;
      widenScalarSrc(MI, WideTy, 2, ExtOpcode);
      widenScalarSrc(MI, WideTy, 3, ExtOpcode);
    }
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_PTR_ADD:
    assert(TypeIdx == 1 && "unable to legalize pointer of G_PTR_ADD");
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
    Observer.changedInstr(MI);
    return Legalized;
 
  case TargetOpcode::G_PHI: {
    assert(TypeIdx == 0 && "Expecting only Idx 0");
 
    Observer.changingInstr(MI);
    for (unsigned I = 1; I < MI.getNumOperands(); I += 2) {
      MachineBasicBlock &OpMBB = *MI.getOperand(I + 1).getMBB();
      MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminatorForward());
      widenScalarSrc(MI, WideTy, I, TargetOpcode::G_ANYEXT);
    }
 
    MachineBasicBlock &MBB = *MI.getParent();
    MIRBuilder.setInsertPt(MBB, --MBB.getFirstNonPHI());
    widenScalarDst(MI, WideTy);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_EXTRACT_VECTOR_ELT: {
    if (TypeIdx == 0) {
      Register VecReg = MI.getOperand(1).getReg();
      LLT VecTy = MRI.getType(VecReg);
      Observer.changingInstr(MI);
 
      widenScalarSrc(
          MI, LLT::vector(VecTy.getElementCount(), WideTy.getSizeInBits()), 1,
          TargetOpcode::G_ANYEXT);
 
      widenScalarDst(MI, WideTy, 0);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    if (TypeIdx != 2)
      return UnableToLegalize;
    Observer.changingInstr(MI);
    // TODO: Probably should be zext
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_INSERT_VECTOR_ELT: {
    if (TypeIdx == 0) {
      Observer.changingInstr(MI);
      const LLT WideEltTy = WideTy.getElementType();
 
      widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
      widenScalarSrc(MI, WideEltTy, 2, TargetOpcode::G_ANYEXT);
      widenScalarDst(MI, WideTy, 0);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    if (TypeIdx == 1) {
      Observer.changingInstr(MI);
 
      Register VecReg = MI.getOperand(1).getReg();
      LLT VecTy = MRI.getType(VecReg);
      LLT WideVecTy = LLT::vector(VecTy.getElementCount(), WideTy);
 
      widenScalarSrc(MI, WideVecTy, 1, TargetOpcode::G_ANYEXT);
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ANYEXT);
      widenScalarDst(MI, WideVecTy, 0);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    if (TypeIdx == 2) {
      Observer.changingInstr(MI);
      // TODO: Probably should be zext
      widenScalarSrc(MI, WideTy, 3, TargetOpcode::G_SEXT);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    return UnableToLegalize;
  }
  case TargetOpcode::G_FADD:
  case TargetOpcode::G_FMUL:
  case TargetOpcode::G_FSUB:
  case TargetOpcode::G_FMA:
  case TargetOpcode::G_FMAD:
  case TargetOpcode::G_FNEG:
  case TargetOpcode::G_FABS:
  case TargetOpcode::G_FCANONICALIZE:
  case TargetOpcode::G_FMINNUM:
  case TargetOpcode::G_FMAXNUM:
  case TargetOpcode::G_FMINNUM_IEEE:
  case TargetOpcode::G_FMAXNUM_IEEE:
  case TargetOpcode::G_FMINIMUM:
  case TargetOpcode::G_FMAXIMUM:
  case TargetOpcode::G_FDIV:
  case TargetOpcode::G_FREM:
  case TargetOpcode::G_FCEIL:
  case TargetOpcode::G_FFLOOR:
  case TargetOpcode::G_FCOS:
  case TargetOpcode::G_FSIN:
  case TargetOpcode::G_FLOG10:
  case TargetOpcode::G_FLOG:
  case TargetOpcode::G_FLOG2:
  case TargetOpcode::G_FRINT:
  case TargetOpcode::G_FNEARBYINT:
  case TargetOpcode::G_FSQRT:
  case TargetOpcode::G_FEXP:
  case TargetOpcode::G_FEXP2:
  case TargetOpcode::G_FEXP10:
  case TargetOpcode::G_FPOW:
  case TargetOpcode::G_INTRINSIC_TRUNC:
  case TargetOpcode::G_INTRINSIC_ROUND:
  case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
    assert(TypeIdx == 0);
    Observer.changingInstr(MI);
 
    for (unsigned I = 1, E = MI.getNumOperands(); I != E; ++I)
      widenScalarSrc(MI, WideTy, I, TargetOpcode::G_FPEXT);
 
    widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_FPOWI:
  case TargetOpcode::G_FLDEXP:
  case TargetOpcode::G_STRICT_FLDEXP: {
    if (TypeIdx == 0) {
      if (MI.getOpcode() == TargetOpcode::G_STRICT_FLDEXP)
        return UnableToLegalize;
 
      Observer.changingInstr(MI);
      widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_FPEXT);
      widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    if (TypeIdx == 1) {
      // For some reason SelectionDAG tries to promote to a libcall without
      // actually changing the integer type for promotion.
      Observer.changingInstr(MI);
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_SEXT);
      Observer.changedInstr(MI);
      return Legalized;
    }
 
    return UnableToLegalize;
  }
  case TargetOpcode::G_FFREXP: {
    Observer.changingInstr(MI);
 
    if (TypeIdx == 0) {
      widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_FPEXT);
      widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
    } else {
      widenScalarDst(MI, WideTy, 1);
    }
 
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_INTTOPTR:
    if (TypeIdx != 1)
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ZEXT);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_PTRTOINT:
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    widenScalarDst(MI, WideTy, 0);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_BUILD_VECTOR: {
    Observer.changingInstr(MI);
 
    const LLT WideEltTy = TypeIdx == 1 ? WideTy : WideTy.getElementType();
    for (int I = 1, E = MI.getNumOperands(); I != E; ++I)
      widenScalarSrc(MI, WideEltTy, I, TargetOpcode::G_ANYEXT);
 
    // Avoid changing the result vector type if the source element type was
    // requested.
    if (TypeIdx == 1) {
      MI.setDesc(MIRBuilder.getTII().get(TargetOpcode::G_BUILD_VECTOR_TRUNC));
    } else {
      widenScalarDst(MI, WideTy, 0);
    }
 
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_SEXT_INREG:
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 1, TargetOpcode::G_ANYEXT);
    widenScalarDst(MI, WideTy, 0, TargetOpcode::G_TRUNC);
    Observer.changedInstr(MI);
    return Legalized;
  case TargetOpcode::G_PTRMASK: {
    if (TypeIdx != 1)
      return UnableToLegalize;
    Observer.changingInstr(MI);
    widenScalarSrc(MI, WideTy, 2, TargetOpcode::G_ZEXT);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_VECREDUCE_FMIN:
  case TargetOpcode::G_VECREDUCE_FMAX:
  case TargetOpcode::G_VECREDUCE_FMINIMUM:
  case TargetOpcode::G_VECREDUCE_FMAXIMUM:
    if (TypeIdx != 0)
      return UnableToLegalize;
    Observer.changingInstr(MI);
    Register VecReg = MI.getOperand(1).getReg();
    LLT VecTy = MRI.getType(VecReg);
    LLT WideVecTy = VecTy.isVector()
                        ? LLT::vector(VecTy.getElementCount(), WideTy)
                        : WideTy;
    widenScalarSrc(MI, WideVecTy, 1, TargetOpcode::G_FPEXT);
    widenScalarDst(MI, WideTy, 0, TargetOpcode::G_FPTRUNC);
    Observer.changedInstr(MI);
    return Legalized;
  }
}
 
static void getUnmergePieces(SmallVectorImpl<Register> &Pieces,
                             MachineIRBuilder &B, Register Src, LLT Ty) {
  auto Unmerge = B.buildUnmerge(Ty, Src);
  for (int I = 0, E = Unmerge->getNumOperands() - 1; I != E; ++I)
    Pieces.push_back(Unmerge.getReg(I));
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::lowerFConstant(MachineInstr &MI) {
  Register Dst = MI.getOperand(0).getReg();
 
  MachineFunction &MF = MIRBuilder.getMF();
  const DataLayout &DL = MIRBuilder.getDataLayout();
 
  unsigned AddrSpace = DL.getDefaultGlobalsAddressSpace();
  LLT AddrPtrTy = LLT::pointer(AddrSpace, DL.getPointerSizeInBits(AddrSpace));
  Align Alignment = Align(DL.getABITypeAlign(
      getFloatTypeForLLT(MF.getFunction().getContext(), MRI.getType(Dst))));
 
  auto Addr = MIRBuilder.buildConstantPool(
      AddrPtrTy, MF.getConstantPool()->getConstantPoolIndex(
                     MI.getOperand(1).getFPImm(), Alignment));
 
  MachineMemOperand *MMO = MF.getMachineMemOperand(
      MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad,
      MRI.getType(Dst), Alignment);
 
  MIRBuilder.buildLoadInstr(TargetOpcode::G_LOAD, Dst, Addr, *MMO);
  MI.eraseFromParent();
 
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::lowerBitcast(MachineInstr &MI) {
  auto [Dst, DstTy, Src, SrcTy] = MI.getFirst2RegLLTs();
  if (SrcTy.isVector()) {
    LLT SrcEltTy = SrcTy.getElementType();
    SmallVector<Register, 8> SrcRegs;
 
    if (DstTy.isVector()) {
      int NumDstElt = DstTy.getNumElements();
      int NumSrcElt = SrcTy.getNumElements();
 
      LLT DstEltTy = DstTy.getElementType();
      LLT DstCastTy = DstEltTy; // Intermediate bitcast result type
      LLT SrcPartTy = SrcEltTy; // Original unmerge result type.
 
      // If there's an element size mismatch, insert intermediate casts to match
      // the result element type.
      if (NumSrcElt < NumDstElt) { // Source element type is larger.
        // %1:_(<4 x s8>) = G_BITCAST %0:_(<2 x s16>)
        //
        // =>
        //
        // %2:_(s16), %3:_(s16) = G_UNMERGE_VALUES %0
        // %3:_(<2 x s8>) = G_BITCAST %2
        // %4:_(<2 x s8>) = G_BITCAST %3
        // %1:_(<4 x s16>) = G_CONCAT_VECTORS %3, %4
        DstCastTy = LLT::fixed_vector(NumDstElt / NumSrcElt, DstEltTy);
        SrcPartTy = SrcEltTy;
      } else if (NumSrcElt > NumDstElt) { // Source element type is smaller.
        //
        // %1:_(<2 x s16>) = G_BITCAST %0:_(<4 x s8>)
        //
        // =>
        //
        // %2:_(<2 x s8>), %3:_(<2 x s8>) = G_UNMERGE_VALUES %0
        // %3:_(s16) = G_BITCAST %2
        // %4:_(s16) = G_BITCAST %3
        // %1:_(<2 x s16>) = G_BUILD_VECTOR %3, %4
        SrcPartTy = LLT::fixed_vector(NumSrcElt / NumDstElt, SrcEltTy);
        DstCastTy = DstEltTy;
      }
 
      getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcPartTy);
      for (Register &SrcReg : SrcRegs)
        SrcReg = MIRBuilder.buildBitcast(DstCastTy, SrcReg).getReg(0);
    } else
      getUnmergePieces(SrcRegs, MIRBuilder, Src, SrcEltTy);
 
    MIRBuilder.buildMergeLikeInstr(Dst, SrcRegs);
    MI.eraseFromParent();
    return Legalized;
  }
 
  if (DstTy.isVector()) {
    SmallVector<Register, 8> SrcRegs;
    getUnmergePieces(SrcRegs, MIRBuilder, Src, DstTy.getElementType());
    MIRBuilder.buildMergeLikeInstr(Dst, SrcRegs);
    MI.eraseFromParent();
    return Legalized;
  }
 
  return UnableToLegalize;
}
 
/// Figure out the bit offset into a register when coercing a vector index for
/// the wide element type. This is only for the case when promoting vector to
/// one with larger elements.
//
///
/// %offset_idx = G_AND %idx, ~(-1 << Log2(DstEltSize / SrcEltSize))
/// %offset_bits = G_SHL %offset_idx, Log2(SrcEltSize)
static Register getBitcastWiderVectorElementOffset(MachineIRBuilder &B,
                                                   Register Idx,
                                                   unsigned NewEltSize,
                                                   unsigned OldEltSize) {
  const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize);
  LLT IdxTy = B.getMRI()->getType(Idx);
 
  // Now figure out the amount we need to shift to get the target bits.
  auto OffsetMask = B.buildConstant(
      IdxTy, ~(APInt::getAllOnes(IdxTy.getSizeInBits()) << Log2EltRatio));
  auto OffsetIdx = B.buildAnd(IdxTy, Idx, OffsetMask);
  return B.buildShl(IdxTy, OffsetIdx,
                    B.buildConstant(IdxTy, Log2_32(OldEltSize))).getReg(0);
}
 
/// Perform a G_EXTRACT_VECTOR_ELT in a different sized vector element. If this
/// is casting to a vector with a smaller element size, perform multiple element
/// extracts and merge the results. If this is coercing to a vector with larger
/// elements, index the bitcasted vector and extract the target element with bit
/// operations. This is intended to force the indexing in the native register
/// size for architectures that can dynamically index the register file.
LegalizerHelper::LegalizeResult
LegalizerHelper::bitcastExtractVectorElt(MachineInstr &MI, unsigned TypeIdx,
                                         LLT CastTy) {
  if (TypeIdx != 1)
    return UnableToLegalize;
 
  auto [Dst, DstTy, SrcVec, SrcVecTy, Idx, IdxTy] = MI.getFirst3RegLLTs();
 
  LLT SrcEltTy = SrcVecTy.getElementType();
  unsigned NewNumElts = CastTy.isVector() ? CastTy.getNumElements() : 1;
  unsigned OldNumElts = SrcVecTy.getNumElements();
 
  LLT NewEltTy = CastTy.isVector() ? CastTy.getElementType() : CastTy;
  Register CastVec = MIRBuilder.buildBitcast(CastTy, SrcVec).getReg(0);
 
  const unsigned NewEltSize = NewEltTy.getSizeInBits();
  const unsigned OldEltSize = SrcEltTy.getSizeInBits();
  if (NewNumElts > OldNumElts) {
    // Decreasing the vector element size
    //
    // e.g. i64 = extract_vector_elt x:v2i64, y:i32
    //  =>
    //  v4i32:castx = bitcast x:v2i64
    //
    // i64 = bitcast
    //   (v2i32 build_vector (i32 (extract_vector_elt castx, (2 * y))),
    //                       (i32 (extract_vector_elt castx, (2 * y + 1)))
    //
    if (NewNumElts % OldNumElts != 0)
      return UnableToLegalize;
 
    // Type of the intermediate result vector.
    const unsigned NewEltsPerOldElt = NewNumElts / OldNumElts;
    LLT MidTy =
        LLT::scalarOrVector(ElementCount::getFixed(NewEltsPerOldElt), NewEltTy);
 
    auto NewEltsPerOldEltK = MIRBuilder.buildConstant(IdxTy, NewEltsPerOldElt);
 
    SmallVector<Register, 8> NewOps(NewEltsPerOldElt);
    auto NewBaseIdx = MIRBuilder.buildMul(IdxTy, Idx, NewEltsPerOldEltK);
 
    for (unsigned I = 0; I < NewEltsPerOldElt; ++I) {
      auto IdxOffset = MIRBuilder.buildConstant(IdxTy, I);
      auto TmpIdx = MIRBuilder.buildAdd(IdxTy, NewBaseIdx, IdxOffset);
      auto Elt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec, TmpIdx);
      NewOps[I] = Elt.getReg(0);
    }
 
    auto NewVec = MIRBuilder.buildBuildVector(MidTy, NewOps);
    MIRBuilder.buildBitcast(Dst, NewVec);
    MI.eraseFromParent();
    return Legalized;
  }
 
  if (NewNumElts < OldNumElts) {
    if (NewEltSize % OldEltSize != 0)
      return UnableToLegalize;
 
    // This only depends on powers of 2 because we use bit tricks to figure out
    // the bit offset we need to shift to get the target element. A general
    // expansion could emit division/multiply.
    if (!isPowerOf2_32(NewEltSize / OldEltSize))
      return UnableToLegalize;
 
    // Increasing the vector element size.
    // %elt:_(small_elt) = G_EXTRACT_VECTOR_ELT %vec:_(<N x small_elt>), %idx
    //
    //   =>
    //
    // %cast = G_BITCAST %vec
    // %scaled_idx = G_LSHR %idx, Log2(DstEltSize / SrcEltSize)
    // %wide_elt  = G_EXTRACT_VECTOR_ELT %cast, %scaled_idx
    // %offset_idx = G_AND %idx, ~(-1 << Log2(DstEltSize / SrcEltSize))
    // %offset_bits = G_SHL %offset_idx, Log2(SrcEltSize)
    // %elt_bits = G_LSHR %wide_elt, %offset_bits
    // %elt = G_TRUNC %elt_bits
 
    const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize);
    auto Log2Ratio = MIRBuilder.buildConstant(IdxTy, Log2EltRatio);
 
    // Divide to get the index in the wider element type.
    auto ScaledIdx = MIRBuilder.buildLShr(IdxTy, Idx, Log2Ratio);
 
    Register WideElt = CastVec;
    if (CastTy.isVector()) {
      WideElt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec,
                                                     ScaledIdx).getReg(0);
    }
 
    // Compute the bit offset into the register of the target element.
    Register OffsetBits = getBitcastWiderVectorElementOffset(
      MIRBuilder, Idx, NewEltSize, OldEltSize);
 
    // Shift the wide element to get the target element.
    auto ExtractedBits = MIRBuilder.buildLShr(NewEltTy, WideElt, OffsetBits);
    MIRBuilder.buildTrunc(Dst, ExtractedBits);
    MI.eraseFromParent();
    return Legalized;
  }
 
  return UnableToLegalize;
}
 
/// Emit code to insert \p InsertReg into \p TargetRet at \p OffsetBits in \p
/// TargetReg, while preserving other bits in \p TargetReg.
///
/// (InsertReg << Offset) | (TargetReg & ~(-1 >> InsertReg.size()) << Offset)
static Register buildBitFieldInsert(MachineIRBuilder &B,
                                    Register TargetReg, Register InsertReg,
                                    Register OffsetBits) {
  LLT TargetTy = B.getMRI()->getType(TargetReg);
  LLT InsertTy = B.getMRI()->getType(InsertReg);
  auto ZextVal = B.buildZExt(TargetTy, InsertReg);
  auto ShiftedInsertVal = B.buildShl(TargetTy, ZextVal, OffsetBits);
 
  // Produce a bitmask of the value to insert
  auto EltMask = B.buildConstant(
    TargetTy, APInt::getLowBitsSet(TargetTy.getSizeInBits(),
                                   InsertTy.getSizeInBits()));
  // Shift it into position
  auto ShiftedMask = B.buildShl(TargetTy, EltMask, OffsetBits);
  auto InvShiftedMask = B.buildNot(TargetTy, ShiftedMask);
 
  // Clear out the bits in the wide element
  auto MaskedOldElt = B.buildAnd(TargetTy, TargetReg, InvShiftedMask);
 
  // The value to insert has all zeros already, so stick it into the masked
  // wide element.
  return B.buildOr(TargetTy, MaskedOldElt, ShiftedInsertVal).getReg(0);
}
 
/// Perform a G_INSERT_VECTOR_ELT in a different sized vector element. If this
/// is increasing the element size, perform the indexing in the target element
/// type, and use bit operations to insert at the element position. This is
/// intended for architectures that can dynamically index the register file and
/// want to force indexing in the native register size.
LegalizerHelper::LegalizeResult
LegalizerHelper::bitcastInsertVectorElt(MachineInstr &MI, unsigned TypeIdx,
                                        LLT CastTy) {
  if (TypeIdx != 0)
    return UnableToLegalize;
 
  auto [Dst, DstTy, SrcVec, SrcVecTy, Val, ValTy, Idx, IdxTy] =
      MI.getFirst4RegLLTs();
  LLT VecTy = DstTy;
 
  LLT VecEltTy = VecTy.getElementType();
  LLT NewEltTy = CastTy.isVector() ? CastTy.getElementType() : CastTy;
  const unsigned NewEltSize = NewEltTy.getSizeInBits();
  const unsigned OldEltSize = VecEltTy.getSizeInBits();
 
  unsigned NewNumElts = CastTy.isVector() ? CastTy.getNumElements() : 1;
  unsigned OldNumElts = VecTy.getNumElements();
 
  Register CastVec = MIRBuilder.buildBitcast(CastTy, SrcVec).getReg(0);
  if (NewNumElts < OldNumElts) {
    if (NewEltSize % OldEltSize != 0)
      return UnableToLegalize;
 
    // This only depends on powers of 2 because we use bit tricks to figure out
    // the bit offset we need to shift to get the target element. A general
    // expansion could emit division/multiply.
    if (!isPowerOf2_32(NewEltSize / OldEltSize))
      return UnableToLegalize;
 
    const unsigned Log2EltRatio = Log2_32(NewEltSize / OldEltSize);
    auto Log2Ratio = MIRBuilder.buildConstant(IdxTy, Log2EltRatio);
 
    // Divide to get the index in the wider element type.
    auto ScaledIdx = MIRBuilder.buildLShr(IdxTy, Idx, Log2Ratio);
 
    Register ExtractedElt = CastVec;
    if (CastTy.isVector()) {
      ExtractedElt = MIRBuilder.buildExtractVectorElement(NewEltTy, CastVec,
                                                          ScaledIdx).getReg(0);
    }
 
    // Compute the bit offset into the register of the target element.
    Register OffsetBits = getBitcastWiderVectorElementOffset(
      MIRBuilder, Idx, NewEltSize, OldEltSize);
 
    Register InsertedElt = buildBitFieldInsert(MIRBuilder, ExtractedElt,
                                               Val, OffsetBits);
    if (CastTy.isVector()) {
      InsertedElt = MIRBuilder.buildInsertVectorElement(
        CastTy, CastVec, InsertedElt, ScaledIdx).getReg(0);
    }
 
    MIRBuilder.buildBitcast(Dst, InsertedElt);
    MI.eraseFromParent();
    return Legalized;
  }
 
  return UnableToLegalize;
}
 
LegalizerHelper::LegalizeResult LegalizerHelper::lowerLoad(GAnyLoad &LoadMI) {
  // Lower to a memory-width G_LOAD and a G_SEXT/G_ZEXT/G_ANYEXT
  Register DstReg = LoadMI.getDstReg();
  Register PtrReg = LoadMI.getPointerReg();
  LLT DstTy = MRI.getType(DstReg);
  MachineMemOperand &MMO = LoadMI.getMMO();
  LLT MemTy = MMO.getMemoryType();
  MachineFunction &MF = MIRBuilder.getMF();
 
  unsigned MemSizeInBits = MemTy.getSizeInBits();
  unsigned MemStoreSizeInBits = 8 * MemTy.getSizeInBytes();
 
  if (MemSizeInBits != MemStoreSizeInBits) {
    if (MemTy.isVector())
      return UnableToLegalize;
 
    // Promote to a byte-sized load if not loading an integral number of
    // bytes.  For example, promote EXTLOAD:i20 -> EXTLOAD:i24.
    LLT WideMemTy = LLT::scalar(MemStoreSizeInBits);
    MachineMemOperand *NewMMO =
        MF.getMachineMemOperand(&MMO, MMO.getPointerInfo(), WideMemTy);
 
    Register LoadReg = DstReg;
    LLT LoadTy = DstTy;
 
    // If this wasn't already an extending load, we need to widen the result
    // register to avoid creating a load with a narrower result than the source.
    if (MemStoreSizeInBits > DstTy.getSizeInBits()) {
      LoadTy = WideMemTy;
      LoadReg = MRI.createGenericVirtualRegister(WideMemTy);
    }
 
    if (isa<GSExtLoad>(LoadMI)) {
      auto NewLoad = MIRBuilder.buildLoad(LoadTy, PtrReg, *NewMMO);
      MIRBuilder.buildSExtInReg(LoadReg, NewLoad, MemSizeInBits);
    } else if (isa<GZExtLoad>(LoadMI) || WideMemTy == LoadTy) {
      auto NewLoad = MIRBuilder.buildLoad(LoadTy, PtrReg, *NewMMO);
      // The extra bits are guaranteed to be zero, since we stored them that
      // way.  A zext load from Wide thus automatically gives zext from MemVT.
      MIRBuilder.buildAssertZExt(LoadReg, NewLoad, MemSizeInBits);
    } else {
      MIRBuilder.buildLoad(LoadReg, PtrReg, *NewMMO);
    }
 
    if (DstTy != LoadTy)
      MIRBuilder.buildTrunc(DstReg, LoadReg);
 
    LoadMI.eraseFromParent();
    return Legalized;
  }
 
  // Big endian lowering not implemented.
  if (MIRBuilder.getDataLayout().isBigEndian())
    return UnableToLegalize;
 
  // This load needs splitting into power of 2 sized loads.
  //
  // Our strategy here is to generate anyextending loads for the smaller
  // types up to next power-2 result type, and then combine the two larger
  // result values together, before truncating back down to the non-pow-2
  // type.
  // E.g. v1 = i24 load =>
  // v2 = i32 zextload (2 byte)
  // v3 = i32 load (1 byte)
  // v4 = i32 shl v3, 16
  // v5 = i32 or v4, v2
  // v1 = i24 trunc v5
  // By doing this we generate the correct truncate which should get
  // combined away as an artifact with a matching extend.
 
  uint64_t LargeSplitSize, SmallSplitSize;
 
  if (!isPowerOf2_32(MemSizeInBits)) {
    // This load needs splitting into power of 2 sized loads.
    LargeSplitSize = llvm::bit_floor(MemSizeInBits);
    SmallSplitSize = MemSizeInBits - LargeSplitSize;
  } else {
    // This is already a power of 2, but we still need to split this in half.
    //
    // Assume we're being asked to decompose an unaligned load.
    // TODO: If this requires multiple splits, handle them all at once.
    auto &Ctx = MF.getFunction().getContext();
    if (TLI.allowsMemoryAccess(Ctx, MIRBuilder.getDataLayout(), MemTy, MMO))
      return UnableToLegalize;
 
    SmallSplitSize = LargeSplitSize = MemSizeInBits / 2;
  }
 
  if (MemTy.isVector()) {
    // TODO: Handle vector extloads
    if (MemTy != DstTy)
      return UnableToLegalize;
 
    // TODO: We can do better than scalarizing the vector and at least split it
    // in half.
    return reduceLoadStoreWidth(LoadMI, 0, DstTy.getElementType());
  }
 
  MachineMemOperand *LargeMMO =
      MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8);
  MachineMemOperand *SmallMMO =
      MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8);
 
  LLT PtrTy = MRI.getType(PtrReg);
  unsigned AnyExtSize = PowerOf2Ceil(DstTy.getSizeInBits());
  LLT AnyExtTy = LLT::scalar(AnyExtSize);
  auto LargeLoad = MIRBuilder.buildLoadInstr(TargetOpcode::G_ZEXTLOAD, AnyExtTy,
                                             PtrReg, *LargeMMO);
 
  auto OffsetCst = MIRBuilder.buildConstant(LLT::scalar(PtrTy.getSizeInBits()),
                                            LargeSplitSize / 8);
  Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy);
  auto SmallPtr = MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst);
  auto SmallLoad = MIRBuilder.buildLoadInstr(LoadMI.getOpcode(), AnyExtTy,
                                             SmallPtr, *SmallMMO);
 
  auto ShiftAmt = MIRBuilder.buildConstant(AnyExtTy, LargeSplitSize);
  auto Shift = MIRBuilder.buildShl(AnyExtTy, SmallLoad, ShiftAmt);
 
  if (AnyExtTy == DstTy)
    MIRBuilder.buildOr(DstReg, Shift, LargeLoad);
  else if (AnyExtTy.getSizeInBits() != DstTy.getSizeInBits()) {
    auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad);
    MIRBuilder.buildTrunc(DstReg, {Or});
  } else {
    assert(DstTy.isPointer() && "expected pointer");
    auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad);
 
    // FIXME: We currently consider this to be illegal for non-integral address
    // spaces, but we need still need a way to reinterpret the bits.
    MIRBuilder.buildIntToPtr(DstReg, Or);
  }
 
  LoadMI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult LegalizerHelper::lowerStore(GStore &StoreMI) {
  // Lower a non-power of 2 store into multiple pow-2 stores.
  // E.g. split an i24 store into an i16 store + i8 store.
  // We do this by first extending the stored value to the next largest power
  // of 2 type, and then using truncating stores to store the components.
  // By doing this, likewise with G_LOAD, generate an extend that can be
  // artifact-combined away instead of leaving behind extracts.
  Register SrcReg = StoreMI.getValueReg();
  Register PtrReg = StoreMI.getPointerReg();
  LLT SrcTy = MRI.getType(SrcReg);
  MachineFunction &MF = MIRBuilder.getMF();
  MachineMemOperand &MMO = **StoreMI.memoperands_begin();
  LLT MemTy = MMO.getMemoryType();
 
  unsigned StoreWidth = MemTy.getSizeInBits();
  unsigned StoreSizeInBits = 8 * MemTy.getSizeInBytes();
 
  if (StoreWidth != StoreSizeInBits) {
    if (SrcTy.isVector())
      return UnableToLegalize;
 
    // Promote to a byte-sized store with upper bits zero if not
    // storing an integral number of bytes.  For example, promote
    // TRUNCSTORE:i1 X -> TRUNCSTORE:i8 (and X, 1)
    LLT WideTy = LLT::scalar(StoreSizeInBits);
 
    if (StoreSizeInBits > SrcTy.getSizeInBits()) {
      // Avoid creating a store with a narrower source than result.
      SrcReg = MIRBuilder.buildAnyExt(WideTy, SrcReg).getReg(0);
      SrcTy = WideTy;
    }
 
    auto ZextInReg = MIRBuilder.buildZExtInReg(SrcTy, SrcReg, StoreWidth);
 
    MachineMemOperand *NewMMO =
        MF.getMachineMemOperand(&MMO, MMO.getPointerInfo(), WideTy);
    MIRBuilder.buildStore(ZextInReg, PtrReg, *NewMMO);
    StoreMI.eraseFromParent();
    return Legalized;
  }
 
  if (MemTy.isVector()) {
    // TODO: Handle vector trunc stores
    if (MemTy != SrcTy)
      return UnableToLegalize;
 
    // TODO: We can do better than scalarizing the vector and at least split it
    // in half.
    return reduceLoadStoreWidth(StoreMI, 0, SrcTy.getElementType());
  }
 
  unsigned MemSizeInBits = MemTy.getSizeInBits();
  uint64_t LargeSplitSize, SmallSplitSize;
 
  if (!isPowerOf2_32(MemSizeInBits)) {
    LargeSplitSize = llvm::bit_floor<uint64_t>(MemTy.getSizeInBits());
    SmallSplitSize = MemTy.getSizeInBits() - LargeSplitSize;
  } else {
    auto &Ctx = MF.getFunction().getContext();
    if (TLI.allowsMemoryAccess(Ctx, MIRBuilder.getDataLayout(), MemTy, MMO))
      return UnableToLegalize; // Don't know what we're being asked to do.
 
    SmallSplitSize = LargeSplitSize = MemSizeInBits / 2;
  }
 
  // Extend to the next pow-2. If this store was itself the result of lowering,
  // e.g. an s56 store being broken into s32 + s24, we might have a stored type
  // that's wider than the stored size.
  unsigned AnyExtSize = PowerOf2Ceil(MemTy.getSizeInBits());
  const LLT NewSrcTy = LLT::scalar(AnyExtSize);
 
  if (SrcTy.isPointer()) {
    const LLT IntPtrTy = LLT::scalar(SrcTy.getSizeInBits());
    SrcReg = MIRBuilder.buildPtrToInt(IntPtrTy, SrcReg).getReg(0);
  }
 
  auto ExtVal = MIRBuilder.buildAnyExtOrTrunc(NewSrcTy, SrcReg);
 
  // Obtain the smaller value by shifting away the larger value.
  auto ShiftAmt = MIRBuilder.buildConstant(NewSrcTy, LargeSplitSize);
  auto SmallVal = MIRBuilder.buildLShr(NewSrcTy, ExtVal, ShiftAmt);
 
  // Generate the PtrAdd and truncating stores.
  LLT PtrTy = MRI.getType(PtrReg);
  auto OffsetCst = MIRBuilder.buildConstant(
    LLT::scalar(PtrTy.getSizeInBits()), LargeSplitSize / 8);
  auto SmallPtr =
    MIRBuilder.buildPtrAdd(PtrTy, PtrReg, OffsetCst);
 
  MachineMemOperand *LargeMMO =
    MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8);
  MachineMemOperand *SmallMMO =
    MF.getMachineMemOperand(&MMO, LargeSplitSize / 8, SmallSplitSize / 8);
  MIRBuilder.buildStore(ExtVal, PtrReg, *LargeMMO);
  MIRBuilder.buildStore(SmallVal, SmallPtr, *SmallMMO);
  StoreMI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::bitcast(MachineInstr &MI, unsigned TypeIdx, LLT CastTy) {
  switch (MI.getOpcode()) {
  case TargetOpcode::G_LOAD: {
    if (TypeIdx != 0)
      return UnableToLegalize;
    MachineMemOperand &MMO = **MI.memoperands_begin();
 
    // Not sure how to interpret a bitcast of an extending load.
    if (MMO.getMemoryType().getSizeInBits() != CastTy.getSizeInBits())
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    bitcastDst(MI, CastTy, 0);
    MMO.setType(CastTy);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_STORE: {
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    MachineMemOperand &MMO = **MI.memoperands_begin();
 
    // Not sure how to interpret a bitcast of a truncating store.
    if (MMO.getMemoryType().getSizeInBits() != CastTy.getSizeInBits())
      return UnableToLegalize;
 
    Observer.changingInstr(MI);
    bitcastSrc(MI, CastTy, 0);
    MMO.setType(CastTy);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_SELECT: {
    if (TypeIdx != 0)
      return UnableToLegalize;
 
    if (MRI.getType(MI.getOperand(1).getReg()).isVector()) {
      LLVM_DEBUG(
          dbgs() << "bitcast action not implemented for vector select\n");
      return UnableToLegalize;
    }
 
    Observer.changingInstr(MI);
    bitcastSrc(MI, CastTy, 2);
    bitcastSrc(MI, CastTy, 3);
    bitcastDst(MI, CastTy, 0);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_AND:
  case TargetOpcode::G_OR:
  case TargetOpcode::G_XOR: {
    Observer.changingInstr(MI);
    bitcastSrc(MI, CastTy, 1);
    bitcastSrc(MI, CastTy, 2);
    bitcastDst(MI, CastTy, 0);
    Observer.changedInstr(MI);
    return Legalized;
  }
  case TargetOpcode::G_EXTRACT_VECTOR_ELT:
    return bitcastExtractVectorElt(MI, TypeIdx, CastTy);
  case TargetOpcode::G_INSERT_VECTOR_ELT:
    return bitcastInsertVectorElt(MI, TypeIdx, CastTy);
  default:
    return UnableToLegalize;
  }
}
 
// Legalize an instruction by changing the opcode in place.
void LegalizerHelper::changeOpcode(MachineInstr &MI, unsigned NewOpcode) {
    Observer.changingInstr(MI);
    MI.setDesc(MIRBuilder.getTII().get(NewOpcode));
    Observer.changedInstr(MI);
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::lower(MachineInstr &MI, unsigned TypeIdx, LLT LowerHintTy) {
  using namespace TargetOpcode;
 
  switch(MI.getOpcode()) {
  default:
    return UnableToLegalize;
  case TargetOpcode::G_FCONSTANT:
    return lowerFConstant(MI);
  case TargetOpcode::G_BITCAST:
    return lowerBitcast(MI);
  case TargetOpcode::G_SREM:
  case TargetOpcode::G_UREM: {
    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
    auto Quot =
        MIRBuilder.buildInstr(MI.getOpcode() == G_SREM ? G_SDIV : G_UDIV, {Ty},
                              {MI.getOperand(1), MI.getOperand(2)});
 
    auto Prod = MIRBuilder.buildMul(Ty, Quot, MI.getOperand(2));
    MIRBuilder.buildSub(MI.getOperand(0), MI.getOperand(1), Prod);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_SADDO:
  case TargetOpcode::G_SSUBO:
    return lowerSADDO_SSUBO(MI);
  case TargetOpcode::G_UMULH:
  case TargetOpcode::G_SMULH:
    return lowerSMULH_UMULH(MI);
  case TargetOpcode::G_SMULO:
  case TargetOpcode::G_UMULO: {
    // Generate G_UMULH/G_SMULH to check for overflow and a normal G_MUL for the
    // result.
    auto [Res, Overflow, LHS, RHS] = MI.getFirst4Regs();
    LLT Ty = MRI.getType(Res);
 
    unsigned Opcode = MI.getOpcode() == TargetOpcode::G_SMULO
                          ? TargetOpcode::G_SMULH
                          : TargetOpcode::G_UMULH;
 
    Observer.changingInstr(MI);
    const auto &TII = MIRBuilder.getTII();
    MI.setDesc(TII.get(TargetOpcode::G_MUL));
    MI.removeOperand(1);
    Observer.changedInstr(MI);
 
    auto HiPart = MIRBuilder.buildInstr(Opcode, {Ty}, {LHS, RHS});
    auto Zero = MIRBuilder.buildConstant(Ty, 0);
 
    // Move insert point forward so we can use the Res register if needed.
    MIRBuilder.setInsertPt(MIRBuilder.getMBB(), ++MIRBuilder.getInsertPt());
 
    // For *signed* multiply, overflow is detected by checking:
    // (hi != (lo >> bitwidth-1))
    if (Opcode == TargetOpcode::G_SMULH) {
      auto ShiftAmt = MIRBuilder.buildConstant(Ty, Ty.getSizeInBits() - 1);
      auto Shifted = MIRBuilder.buildAShr(Ty, Res, ShiftAmt);
      MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Shifted);
    } else {
      MIRBuilder.buildICmp(CmpInst::ICMP_NE, Overflow, HiPart, Zero);
    }
    return Legalized;
  }
  case TargetOpcode::G_FNEG: {
    auto [Res, SubByReg] = MI.getFirst2Regs();
    LLT Ty = MRI.getType(Res);
 
    // TODO: Handle vector types once we are able to
    // represent them.
    if (Ty.isVector())
      return UnableToLegalize;
    auto SignMask =
        MIRBuilder.buildConstant(Ty, APInt::getSignMask(Ty.getSizeInBits()));
    MIRBuilder.buildXor(Res, SubByReg, SignMask);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_FSUB:
  case TargetOpcode::G_STRICT_FSUB: {
    auto [Res, LHS, RHS] = MI.getFirst3Regs();
    LLT Ty = MRI.getType(Res);
 
    // Lower (G_FSUB LHS, RHS) to (G_FADD LHS, (G_FNEG RHS)).
    auto Neg = MIRBuilder.buildFNeg(Ty, RHS);
 
    if (MI.getOpcode() == TargetOpcode::G_STRICT_FSUB)
      MIRBuilder.buildStrictFAdd(Res, LHS, Neg, MI.getFlags());
    else
      MIRBuilder.buildFAdd(Res, LHS, Neg, MI.getFlags());
 
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_FMAD:
    return lowerFMad(MI);
  case TargetOpcode::G_FFLOOR:
    return lowerFFloor(MI);
  case TargetOpcode::G_INTRINSIC_ROUND:
    return lowerIntrinsicRound(MI);
  case TargetOpcode::G_INTRINSIC_ROUNDEVEN: {
    // Since round even is the assumed rounding mode for unconstrained FP
    // operations, rint and roundeven are the same operation.
    changeOpcode(MI, TargetOpcode::G_FRINT);
    return Legalized;
  }
  case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: {
    auto [OldValRes, SuccessRes, Addr, CmpVal, NewVal] = MI.getFirst5Regs();
    MIRBuilder.buildAtomicCmpXchg(OldValRes, Addr, CmpVal, NewVal,
                                  **MI.memoperands_begin());
    MIRBuilder.buildICmp(CmpInst::ICMP_EQ, SuccessRes, OldValRes, CmpVal);
    MI.eraseFromParent();
    return Legalized;
  }
  case TargetOpcode::G_LOAD:
  case TargetOpcode::G_SEXTLOAD:
  case TargetOpcode::G_ZEXTLOAD:
    return lowerLoad(cast<GAnyLoad>(MI));
  case TargetOpcode::G_STORE:
    return lowerStore(cast<GStore>(MI));
  case TargetOpcode::G_CTLZ_ZERO_UNDEF:
  case TargetOpcode::G_CTTZ_ZERO_UNDEF:
  case TargetOpcode::G_CTLZ:
  case TargetOpcode::G_CTTZ:
  case TargetOpcode::G_CTPOP:
    return lowerBitCount(MI);
  case G_UADDO: {
    auto [Res, CarryOut, LHS, RHS] = MI.getFirst4Regs();
 
    MIRBuilder.buildAdd(Res, LHS, RHS);
    MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CarryOut, Res, RHS);
 
    MI.eraseFromParent();
    return Legalized;
  }
  case G_UADDE: {
    auto [Res, CarryOut, LHS, RHS, CarryIn] = MI.getFirst5Regs();
    const LLT CondTy = MRI.getType(CarryOut);
    const LLT Ty = MRI.getType(Res);
 
    // Initial add of the two operands.
    auto TmpRes = MIRBuilder.buildAdd(Ty, LHS, RHS);
 
    // Initial check for carry.
    auto Carry = MIRBuilder.buildICmp(CmpInst::ICMP_ULT, CondTy, TmpRes, LHS);
 
    // Add the sum and the carry.
    auto ZExtCarryIn = MIRBuilder.buildZExt(Ty, CarryIn);
    MIRBuilder.buildAdd(Res, TmpRes, ZExtCarryIn);
 
    // Second check for carry. We can only carry if the initial sum is all 1s
    // and the carry is set, resulting in a new sum of 0.
    auto Zero = MIRBuilder.buildConstant(Ty, 0);
    auto ResEqZero = MIRBuilder.buildICmp(CmpInst::ICMP_EQ, CondTy, Res, Zero);
    auto Carry2 = MIRBuilder.buildAnd(CondTy, ResEqZero, CarryIn);
    MIRBuilder.buildOr(CarryOut, Carry, Carry2);
 
    MI.eraseFromParent();
    return Legalized;
  }
  case G_USUBO: {
    auto [Res, BorrowOut, LHS, RHS] = MI.getFirst4Regs();
 
    MIRBuilder.buildSub(Res, LHS, RHS);
    MIRBuilder.buildICmp(CmpInst::ICMP_ULT, BorrowOut, LHS, RHS);
 
    MI.eraseFromParent();
    return Legalized;
  }
  case G_USUBE: {
    auto [Res, BorrowOut, LHS, RHS, BorrowIn] = MI.getFirst5Regs();
    const LLT CondTy = MRI.getType(BorrowOut);
    const LLT Ty = MRI.getType(Res);
 
    // Initial subtract of the two operands.
    auto TmpRes = MIRBuilder.buildSub(Ty, LHS, RHS);
 
    // Initial check for borrow.
    auto Borrow = MIRBuilder.buildICmp(CmpInst::ICMP_UGT, CondTy, TmpRes, LHS);
 
    // Subtract the borrow from the first subtract.
    auto ZExtBorrowIn = MIRBuilder.buildZExt(Ty, BorrowIn);
    MIRBuilder.buildSub(Res, TmpRes, ZExtBorrowIn);
 
    // Second check for borrow. We can only borrow if the initial difference is
    // 0 and the borrow is set, resulting in a new difference of all 1s.
    auto Zero = MIRBuilder.buildConstant(Ty, 0);
    auto TmpResEqZero =
        MIRBuilder.buildICmp(CmpInst::ICMP_EQ, CondTy, TmpRes, Zero);
    auto Borrow2 = MIRBuilder.buildAnd(CondTy, TmpResEqZero, BorrowIn);
    MIRBuilder.buildOr(BorrowOut, Borrow, Borrow2);
 
    MI.eraseFromParent();
    return Legalized;
  }
  case G_UITOFP:
    return lowerUITOFP(MI);
  case G_SITOFP:
    return lowerSITOFP(MI);
  case G_FPTOUI:
    return lowerFPTOUI(MI);
  case G_FPTOSI:
    return lowerFPTOSI(MI);
  case G_FPTRUNC:
    return lowerFPTRUNC(MI);
  case G_FPOWI:
    return lowerFPOWI(MI);
  case G_SMIN:
  case G_SMAX:
  case G_UMIN:
  case G_UMAX:
    return lowerMinMax(MI);
  case G_FCOPYSIGN:
    return lowerFCopySign(MI);
  case G_FMINNUM:
  case G_FMAXNUM:
    return lowerFMinNumMaxNum(MI);
  case G_MERGE_VALUES:
    return lowerMergeValues(MI);
  case G_UNMERGE_VALUES:
    return lowerUnmergeValues(MI);
  case TargetOpcode::G_SEXT_INREG: {
    assert(MI.getOperand(2).isImm() && "Expected immediate");
    int64_t SizeInBits = MI.getOperand(2).getImm();
 
    auto [DstReg, SrcReg] = MI.getFirst2Regs();
    LLT DstTy = MRI.getType(DstReg);
    Register TmpRes = MRI.createGenericVirtualRegister(DstTy);
 
    auto MIBSz = MIRBuilder.buildConstant(DstTy, DstTy.getScalarSizeInBits() - SizeInBits);
    MIRBuilder.buildShl(TmpRes, SrcReg, MIBSz->getOperand(0));
    MIRBuilder.buildAShr(DstReg, TmpRes, MIBSz->getOperand(0));
    MI.eraseFromParent();
    return Legalized;
  }
  case G_EXTRACT_VECTOR_ELT:
  case G_INSERT_VECTOR_ELT:
    return lowerExtractInsertVectorElt(MI);
  case G_SHUFFLE_VECTOR:
    return lowerShuffleVector(MI);
  case G_DYN_STACKALLOC:
    return lowerDynStackAlloc(MI);
  case G_STACKSAVE:
    return lowerStackSave(MI);
  case G_STACKRESTORE:
    return lowerStackRestore(MI);
  case G_EXTRACT:
    return lowerExtract(MI);
  case G_INSERT:
    return lowerInsert(MI);
  case G_BSWAP:
    return lowerBswap(MI);
  case G_BITREVERSE:
    return lowerBitreverse(MI);
  case G_READ_REGISTER:
  case G_WRITE_REGISTER:
    return lowerReadWriteRegister(MI);
  case G_UADDSAT:
  case G_USUBSAT: {
    // Try to make a reasonable guess about which lowering strategy to use. The
    // target can override this with custom lowering and calling the
    // implementation functions.
    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
    if (LI.isLegalOrCustom({G_UMIN, Ty}))
      return lowerAddSubSatToMinMax(MI);
    return lowerAddSubSatToAddoSubo(MI);
  }
  case G_SADDSAT:
  case G_SSUBSAT: {
    LLT Ty = MRI.getType(MI.getOperand(0).getReg());
 
    // FIXME: It would probably make more sense to see if G_SADDO is preferred,
    // since it's a shorter expansion. However, we would need to figure out the
    // preferred boolean type for the carry out for the query.
    if (LI.isLegalOrCustom({G_SMIN, Ty}) && LI.isLegalOrCustom({G_SMAX, Ty}))
      return lowerAddSubSatToMinMax(MI);
    return lowerAddSubSatToAddoSubo(MI);
  }
  case G_SSHLSAT:
  case G_USHLSAT:
    return lowerShlSat(MI);
  case G_ABS:
    return lowerAbsToAddXor(MI);
  case G_SELECT:
    return lowerSelect(MI);
  case G_IS_FPCLASS:
    return lowerISFPCLASS(MI);
  case G_SDIVREM:
  case G_UDIVREM:
    return lowerDIVREM(MI);
  case G_FSHL:
  case G_FSHR:
    return lowerFunnelShift(MI);
  case G_ROTL:
  case G_ROTR:
    return lowerRotate(MI);
  case G_MEMSET:
  case G_MEMCPY:
  case G_MEMMOVE:
    return lowerMemCpyFamily(MI);
  case G_MEMCPY_INLINE:
    return lowerMemcpyInline(MI);
  case G_ZEXT:
  case G_SEXT:
  case G_ANYEXT:
    return lowerEXT(MI);
  GISEL_VECREDUCE_CASES_NONSEQ
    return lowerVectorReduction(MI);
  }
}
 
Align LegalizerHelper::getStackTemporaryAlignment(LLT Ty,
                                                  Align MinAlign) const {
  // FIXME: We're missing a way to go back from LLT to llvm::Type to query the
  // datalayout for the preferred alignment. Also there should be a target hook
  // for this to allow targets to reduce the alignment and ignore the
  // datalayout. e.g. AMDGPU should always use a 4-byte alignment, regardless of
  // the type.
  return std::max(Align(PowerOf2Ceil(Ty.getSizeInBytes())), MinAlign);
}
 
MachineInstrBuilder
LegalizerHelper::createStackTemporary(TypeSize Bytes, Align Alignment,
                                      MachinePointerInfo &PtrInfo) {
  MachineFunction &MF = MIRBuilder.getMF();
  const DataLayout &DL = MIRBuilder.getDataLayout();
  int FrameIdx = MF.getFrameInfo().CreateStackObject(Bytes, Alignment, false);
 
  unsigned AddrSpace = DL.getAllocaAddrSpace();
  LLT FramePtrTy = LLT::pointer(AddrSpace, DL.getPointerSizeInBits(AddrSpace));
 
  PtrInfo = MachinePointerInfo::getFixedStack(MF, FrameIdx);
  return MIRBuilder.buildFrameIndex(FramePtrTy, FrameIdx);
}
 
static Register clampDynamicVectorIndex(MachineIRBuilder &B, Register IdxReg,
                                        LLT VecTy) {
  int64_t IdxVal;
  if (mi_match(IdxReg, *B.getMRI(), m_ICst(IdxVal)))
    return IdxReg;
 
  LLT IdxTy = B.getMRI()->getType(IdxReg);
  unsigned NElts = VecTy.getNumElements();
  if (isPowerOf2_32(NElts)) {
    APInt Imm = APInt::getLowBitsSet(IdxTy.getSizeInBits(), Log2_32(NElts));
    return B.buildAnd(IdxTy, IdxReg, B.buildConstant(IdxTy, Imm)).getReg(0);
  }
 
  return B.buildUMin(IdxTy, IdxReg, B.buildConstant(IdxTy, NElts - 1))
      .getReg(0);
}
 
Register LegalizerHelper::getVectorElementPointer(Register VecPtr, LLT VecTy,
                                                  Register Index) {
  LLT EltTy = VecTy.getElementType();
 
  // Calculate the element offset and add it to the pointer.
  unsigned EltSize = EltTy.getSizeInBits() / 8; // FIXME: should be ABI size.
  assert(EltSize * 8 == EltTy.getSizeInBits() &&
         "Converting bits to bytes lost precision");
 
  Index = clampDynamicVectorIndex(MIRBuilder, Index, VecTy);
 
  LLT IdxTy = MRI.getType(Index);
  auto Mul = MIRBuilder.buildMul(IdxTy, Index,
                                 MIRBuilder.buildConstant(IdxTy, EltSize));
 
  LLT PtrTy = MRI.getType(VecPtr);
  return MIRBuilder.buildPtrAdd(PtrTy, VecPtr, Mul).getReg(0);
}
 
#ifndef NDEBUG
/// Check that all vector operands have same number of elements. Other operands
/// should be listed in NonVecOp.
static bool hasSameNumEltsOnAllVectorOperands(
    GenericMachineInstr &MI, MachineRegisterInfo &MRI,
    std::initializer_list<unsigned> NonVecOpIndices) {
  if (MI.getNumMemOperands() != 0)
    return false;
 
  LLT VecTy = MRI.getType(MI.getReg(0));
  if (!VecTy.isVector())
    return false;
  unsigned NumElts = VecTy.getNumElements();
 
  for (unsigned OpIdx = 1; OpIdx < MI.getNumOperands(); ++OpIdx) {
    MachineOperand &Op = MI.getOperand(OpIdx);
    if (!Op.isReg()) {
      if (!is_contained(NonVecOpIndices, OpIdx))
        return false;
      continue;
    }
 
    LLT Ty = MRI.getType(Op.getReg());
    if (!Ty.isVector()) {
      if (!is_contained(NonVecOpIndices, OpIdx))
        return false;
      continue;
    }
 
    if (Ty.getNumElements() != NumElts)
      return false;
  }
 
  return true;
}
#endif
 
/// Fill \p DstOps with DstOps that have same number of elements combined as
/// the Ty. These DstOps have either scalar type when \p NumElts = 1 or are
/// vectors with \p NumElts elements. When Ty.getNumElements() is not multiple
/// of \p NumElts last DstOp (leftover) has fewer then \p NumElts elements.
static void makeDstOps(SmallVectorImpl<DstOp> &DstOps, LLT Ty,
                       unsigned NumElts) {
  LLT LeftoverTy;
  assert(Ty.isVector() && "Expected vector type");
  LLT EltTy = Ty.getElementType();
  LLT NarrowTy = (NumElts == 1) ? EltTy : LLT::fixed_vector(NumElts, EltTy);
  int NumParts, NumLeftover;
  std::tie(NumParts, NumLeftover) =
      getNarrowTypeBreakDown(Ty, NarrowTy, LeftoverTy);
 
  assert(NumParts > 0 && "Error in getNarrowTypeBreakDown");
  for (int i = 0; i < NumParts; ++i) {
    DstOps.push_back(NarrowTy);
  }
 
  if (LeftoverTy.isValid()) {
    assert(NumLeftover == 1 && "expected exactly one leftover");
    DstOps.push_back(LeftoverTy);
  }
}
 
/// Operand \p Op is used on \p N sub-instructions. Fill \p Ops with \p N SrcOps
/// made from \p Op depending on operand type.
static void broadcastSrcOp(SmallVectorImpl<SrcOp> &Ops, unsigned N,
                           MachineOperand &Op) {
  for (unsigned i = 0; i < N; ++i) {
    if (Op.isReg())
      Ops.push_back(Op.getReg());
    else if (Op.isImm())
      Ops.push_back(Op.getImm());
    else if (Op.isPredicate())
      Ops.push_back(static_cast<CmpInst::Predicate>(Op.getPredicate()));
    else
      llvm_unreachable("Unsupported type");
  }
}
 
// Handle splitting vector operations which need to have the same number of
// elements in each type index, but each type index may have a different element
// type.
//
// e.g.  <4 x s64> = G_SHL <4 x s64>, <4 x s32> ->
//       <2 x s64> = G_SHL <2 x s64>, <2 x s32>
//       <2 x s64> = G_SHL <2 x s64>, <2 x s32>
//
// Also handles some irregular breakdown cases, e.g.
// e.g.  <3 x s64> = G_SHL <3 x s64>, <3 x s32> ->
//       <2 x s64> = G_SHL <2 x s64>, <2 x s32>
//             s64 = G_SHL s64, s32
LegalizerHelper::LegalizeResult
LegalizerHelper::fewerElementsVectorMultiEltType(
    GenericMachineInstr &MI, unsigned NumElts,
    std::initializer_list<unsigned> NonVecOpIndices) {
  assert(hasSameNumEltsOnAllVectorOperands(MI, MRI, NonVecOpIndices) &&
         "Non-compatible opcode or not specified non-vector operands");
  unsigned OrigNumElts = MRI.getType(MI.getReg(0)).getNumElements();
 
  unsigned NumInputs = MI.getNumOperands() - MI.getNumDefs();
  unsigned NumDefs = MI.getNumDefs();
 
  // Create DstOps (sub-vectors with NumElts elts + Leftover) for each output.
  // Build instructions with DstOps to use instruction found by CSE directly.
  // CSE copies found instruction into given vreg when building with vreg dest.
  SmallVector<SmallVector<DstOp, 8>, 2> OutputOpsPieces(NumDefs);
  // Output registers will be taken from created instructions.
  SmallVector<SmallVector<Register, 8>, 2> OutputRegs(NumDefs);
  for (unsigned i = 0; i < NumDefs; ++i) {
    makeDstOps(OutputOpsPieces[i], MRI.getType(MI.getReg(i)), NumElts);
  }
 
  // Split vector input operands into sub-vectors with NumElts elts + Leftover.
  // Operands listed in NonVecOpIndices will be used as is without splitting;
  // examples: compare predicate in icmp and fcmp (op 1), vector select with i1
  // scalar condition (op 1), immediate in sext_inreg (op 2).
  SmallVector<SmallVector<SrcOp, 8>, 3> InputOpsPieces(NumInputs);
  for (unsigned UseIdx = NumDefs, UseNo = 0; UseIdx < MI.getNumOperands();
       ++UseIdx, ++UseNo) {
    if (is_contained(NonVecOpIndices, UseIdx)) {
      broadcastSrcOp(InputOpsPieces[UseNo], OutputOpsPieces[0].size(),
                     MI.getOperand(UseIdx));
    } else {
      SmallVector<Register, 8> SplitPieces;
      extractVectorParts(MI.getReg(UseIdx), NumElts, SplitPieces);
      for (auto Reg : SplitPieces)
        InputOpsPieces[UseNo].push_back(Reg);
    }
  }
 
  unsigned NumLeftovers = OrigNumElts % NumElts ? 1 : 0;
 
  // Take i-th piece of each input operand split and build sub-vector/scalar
  // instruction. Set i-th DstOp(s) from OutputOpsPieces as destination(s).
  for (unsigned i = 0; i < OrigNumElts / NumElts + NumLeftovers; ++i) {
    SmallVector<DstOp, 2> Defs;
    for (unsigned DstNo = 0; DstNo < NumDefs; ++DstNo)
      Defs.push_back(OutputOpsPieces[DstNo][i]);
 
    SmallVector<SrcOp, 3> Uses;
    for (unsigned InputNo = 0; InputNo < NumInputs; ++InputNo)
      Uses.push_back(InputOpsPieces[InputNo][i]);
 
    auto I = MIRBuilder.buildInstr(MI.getOpcode(), Defs, Uses, MI.getFlags());
    for (unsigned DstNo = 0; DstNo < NumDefs; ++DstNo)
      OutputRegs[DstNo].push_back(I.getReg(DstNo));
  }
 
  // Merge small outputs into MI's output for each def operand.
  if (NumLeftovers) {
    for (unsigned i = 0; i < NumDefs; ++i)
      mergeMixedSubvectors(MI.getReg(i), OutputRegs[i]);
  } else {
    for (unsigned i = 0; i < NumDefs; ++i)
      MIRBuilder.buildMergeLikeInstr(MI.getReg(i), OutputRegs[i]);
  }
 
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::fewerElementsVectorPhi(GenericMachineInstr &MI,
                                        unsigned NumElts) {
  unsigned OrigNumElts = MRI.getType(MI.getReg(0)).getNumElements();
 
  unsigned NumInputs = MI.getNumOperands() - MI.getNumDefs();
  unsigned NumDefs = MI.getNumDefs();
 
  SmallVector<DstOp, 8> OutputOpsPieces;
  SmallVector<Register, 8> OutputRegs;
  makeDstOps(OutputOpsPieces, MRI.getType(MI.getReg(0)), NumElts);
 
  // Instructions that perform register split will be inserted in basic block
  // where register is defined (basic block is in the next operand).
  SmallVector<SmallVector<Register, 8>, 3> InputOpsPieces(NumInputs / 2);
  for (unsigned UseIdx = NumDefs, UseNo = 0; UseIdx < MI.getNumOperands();
       UseIdx += 2, ++UseNo) {
    MachineBasicBlock &OpMBB = *MI.getOperand(UseIdx + 1).getMBB();
    MIRBuilder.setInsertPt(OpMBB, OpMBB.getFirstTerminatorForward());
    extractVectorParts(MI.getReg(UseIdx), NumElts, InputOpsPieces[UseNo]);
  }
 
  // Build PHIs with fewer elements.
  unsigned NumLeftovers = OrigNumElts % NumElts ? 1 : 0;
  MIRBuilder.setInsertPt(*MI.getParent(), MI);
  for (unsigned i = 0; i < OrigNumElts / NumElts + NumLeftovers; ++i) {
    auto Phi = MIRBuilder.buildInstr(TargetOpcode::G_PHI);
    Phi.addDef(
        MRI.createGenericVirtualRegister(OutputOpsPieces[i].getLLTTy(MRI)));
    OutputRegs.push_back(Phi.getReg(0));
 
    for (unsigned j = 0; j < NumInputs / 2; ++j) {
      Phi.addUse(InputOpsPieces[j][i]);
      Phi.add(MI.getOperand(1 + j * 2 + 1));
    }
  }
 
  // Merge small outputs into MI's def.
  if (NumLeftovers) {
    mergeMixedSubvectors(MI.getReg(0), OutputRegs);
  } else {
    MIRBuilder.buildMergeLikeInstr(MI.getReg(0), OutputRegs);
  }
 
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::fewerElementsVectorUnmergeValues(MachineInstr &MI,
                                                  unsigned TypeIdx,
                                                  LLT NarrowTy) {
  const int NumDst = MI.getNumOperands() - 1;
  const Register SrcReg = MI.getOperand(NumDst).getReg();
  LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
  LLT SrcTy = MRI.getType(SrcReg);
 
  if (TypeIdx != 1 || NarrowTy == DstTy)
    return UnableToLegalize;
 
  // Requires compatible types. Otherwise SrcReg should have been defined by
  // merge-like instruction that would get artifact combined. Most likely
  // instruction that defines SrcReg has to perform more/fewer elements
  // legalization compatible with NarrowTy.
  assert(SrcTy.isVector() && NarrowTy.isVector() && "Expected vector types");
  assert((SrcTy.getScalarType() == NarrowTy.getScalarType()) && "bad type");
 
  if ((SrcTy.getSizeInBits() % NarrowTy.getSizeInBits() != 0) ||
      (NarrowTy.getSizeInBits() % DstTy.getSizeInBits() != 0))
    return UnableToLegalize;
 
  // This is most likely DstTy (smaller then register size) packed in SrcTy
  // (larger then register size) and since unmerge was not combined it will be
  // lowered to bit sequence extracts from register. Unpack SrcTy to NarrowTy
  // (register size) pieces first. Then unpack each of NarrowTy pieces to DstTy.
 
  // %1:_(DstTy), %2, %3, %4 = G_UNMERGE_VALUES %0:_(SrcTy)
  //
  // %5:_(NarrowTy), %6 = G_UNMERGE_VALUES %0:_(SrcTy) - reg sequence
  // %1:_(DstTy), %2 = G_UNMERGE_VALUES %5:_(NarrowTy) - sequence of bits in reg
  // %3:_(DstTy), %4 = G_UNMERGE_VALUES %6:_(NarrowTy)
  auto Unmerge = MIRBuilder.buildUnmerge(NarrowTy, SrcReg);
  const int NumUnmerge = Unmerge->getNumOperands() - 1;
  const int PartsPerUnmerge = NumDst / NumUnmerge;
 
  for (int I = 0; I != NumUnmerge; ++I) {
    auto MIB = MIRBuilder.buildInstr(TargetOpcode::G_UNMERGE_VALUES);
 
    for (int J = 0; J != PartsPerUnmerge; ++J)
      MIB.addDef(MI.getOperand(I * PartsPerUnmerge + J).getReg());
    MIB.addUse(Unmerge.getReg(I));
  }
 
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::fewerElementsVectorMerge(MachineInstr &MI, unsigned TypeIdx,
                                          LLT NarrowTy) {
  auto [DstReg, DstTy, SrcReg, SrcTy] = MI.getFirst2RegLLTs();
  // Requires compatible types. Otherwise user of DstReg did not perform unmerge
  // that should have been artifact combined. Most likely instruction that uses
  // DstReg has to do more/fewer elements legalization compatible with NarrowTy.
  assert(DstTy.isVector() && NarrowTy.isVector() && "Expected vector types");
  assert((DstTy.getScalarType() == NarrowTy.getScalarType()) && "bad type");
  if (NarrowTy == SrcTy)
    return UnableToLegalize;
 
  // This attempts to lower part of LCMTy merge/unmerge sequence. Intended use
  // is for old mir tests. Since the changes to more/fewer elements it should no
  // longer be possible to generate MIR like this when starting from llvm-ir
  // because LCMTy approach was replaced with merge/unmerge to vector elements.
  if (TypeIdx == 1) {
    assert(SrcTy.isVector() && "Expected vector types");
    assert((SrcTy.getScalarType() == NarrowTy.getScalarType()) && "bad type");
    if ((DstTy.getSizeInBits() % NarrowTy.getSizeInBits() != 0) ||
        (NarrowTy.getNumElements() >= SrcTy.getNumElements()))
      return UnableToLegalize;
    // %2:_(DstTy) = G_CONCAT_VECTORS %0:_(SrcTy), %1:_(SrcTy)
    //
    // %3:_(EltTy), %4, %5 = G_UNMERGE_VALUES %0:_(SrcTy)
    // %6:_(EltTy), %7, %8 = G_UNMERGE_VALUES %1:_(SrcTy)
    // %9:_(NarrowTy) = G_BUILD_VECTOR %3:_(EltTy), %4
    // %10:_(NarrowTy) = G_BUILD_VECTOR %5:_(EltTy), %6
    // %11:_(NarrowTy) = G_BUILD_VECTOR %7:_(EltTy), %8
    // %2:_(DstTy) = G_CONCAT_VECTORS %9:_(NarrowTy), %10, %11
 
    SmallVector<Register, 8> Elts;
    LLT EltTy = MRI.getType(MI.getOperand(1).getReg()).getScalarType();
    for (unsigned i = 1; i < MI.getNumOperands(); ++i) {
      auto Unmerge = MIRBuilder.buildUnmerge(EltTy, MI.getOperand(i).getReg());
      for (unsigned j = 0; j < Unmerge->getNumDefs(); ++j)
        Elts.push_back(Unmerge.getReg(j));
    }
 
    SmallVector<Register, 8> NarrowTyElts;
    unsigned NumNarrowTyElts = NarrowTy.getNumElements();
    unsigned NumNarrowTyPieces = DstTy.getNumElements() / NumNarrowTyElts;
    for (unsigned i = 0, Offset = 0; i < NumNarrowTyPieces;
         ++i, Offset += NumNarrowTyElts) {
      ArrayRef<Register> Pieces(&Elts[Offset], NumNarrowTyElts);
      NarrowTyElts.push_back(
          MIRBuilder.buildMergeLikeInstr(NarrowTy, Pieces).getReg(0));
    }
 
    MIRBuilder.buildMergeLikeInstr(DstReg, NarrowTyElts);
    MI.eraseFromParent();
    return Legalized;
  }
 
  assert(TypeIdx == 0 && "Bad type index");
  if ((NarrowTy.getSizeInBits() % SrcTy.getSizeInBits() != 0) ||
      (DstTy.getSizeInBits() % NarrowTy.getSizeInBits() != 0))
    return UnableToLegalize;
 
  // This is most likely SrcTy (smaller then register size) packed in DstTy
  // (larger then register size) and since merge was not combined it will be
  // lowered to bit sequence packing into register. Merge SrcTy to NarrowTy
  // (register size) pieces first. Then merge each of NarrowTy pieces to DstTy.
 
  // %0:_(DstTy) = G_MERGE_VALUES %1:_(SrcTy), %2, %3, %4
  //
  // %5:_(NarrowTy) = G_MERGE_VALUES %1:_(SrcTy), %2 - sequence of bits in reg
  // %6:_(NarrowTy) = G_MERGE_VALUES %3:_(SrcTy), %4
  // %0:_(DstTy)  = G_MERGE_VALUES %5:_(NarrowTy), %6 - reg sequence
  SmallVector<Register, 8> NarrowTyElts;
  unsigned NumParts = DstTy.getNumElements() / NarrowTy.getNumElements();
  unsigned NumSrcElts = SrcTy.isVector() ? SrcTy.getNumElements() : 1;
  unsigned NumElts = NarrowTy.getNumElements() / NumSrcElts;
  for (unsigned i = 0; i < NumParts; ++i) {
    SmallVector<Register, 8> Sources;
    for (unsigned j = 0; j < NumElts; ++j)
      Sources.push_back(MI.getOperand(1 + i * NumElts + j).getReg());
    NarrowTyElts.push_back(
        MIRBuilder.buildMergeLikeInstr(NarrowTy, Sources).getReg(0));
  }
 
  MIRBuilder.buildMergeLikeInstr(DstReg, NarrowTyElts);
  MI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::fewerElementsVectorExtractInsertVectorElt(MachineInstr &MI,
                                                           unsigned TypeIdx,
                                                           LLT NarrowVecTy) {
  auto [DstReg, SrcVec] = MI.getFirst2Regs();
  Register InsertVal;
  bool IsInsert = MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT;
 
  assert((IsInsert ? TypeIdx == 0 : TypeIdx == 1) && "not a vector type index");
  if (IsInsert)
    InsertVal = MI.getOperand(2).getReg();
 
  Register Idx = MI.getOperand(MI.getNumOperands() - 1).getReg();
 
  // TODO: Handle total scalarization case.
  if (!NarrowVecTy.isVector())
    return UnableToLegalize;
 
  LLT VecTy = MRI.getType(SrcVec);
 
  // If the index is a constant, we can really break this down as you would
  // expect, and index into the target size pieces.
  int64_t IdxVal;
  auto MaybeCst = getIConstantVRegValWithLookThrough(Idx, MRI);
  if (MaybeCst) {
    IdxVal = MaybeCst->Value.getSExtValue();
    // Avoid out of bounds indexing the pieces.
    if (IdxVal >= VecTy.getNumElements()) {
      MIRBuilder.buildUndef(DstReg);
      MI.eraseFromParent();
      return Legalized;
    }
 
    SmallVector<Register, 8> VecParts;
    LLT GCDTy = extractGCDType(VecParts, VecTy, NarrowVecTy, SrcVec);
 
    // Build a sequence of NarrowTy pieces in VecParts for this operand.
    LLT LCMTy = buildLCMMergePieces(VecTy, NarrowVecTy, GCDTy, VecParts,
                                    TargetOpcode::G_ANYEXT);
 
    unsigned NewNumElts = NarrowVecTy.getNumElements();
 
    LLT IdxTy = MRI.getType(Idx);
    int64_t PartIdx = IdxVal / NewNumElts;
    auto NewIdx =
        MIRBuilder.buildConstant(IdxTy, IdxVal - NewNumElts * PartIdx);
 
    if (IsInsert) {
      LLT PartTy = MRI.getType(VecParts[PartIdx]);
 
      // Use the adjusted index to insert into one of the subvectors.
      auto InsertPart = MIRBuilder.buildInsertVectorElement(
          PartTy, VecParts[PartIdx], InsertVal, NewIdx);
      VecParts[PartIdx] = InsertPart.getReg(0);
 
      // Recombine the inserted subvector with the others to reform the result
      // vector.
      buildWidenedRemergeToDst(DstReg, LCMTy, VecParts);
    } else {
      MIRBuilder.buildExtractVectorElement(DstReg, VecParts[PartIdx], NewIdx);
    }
 
    MI.eraseFromParent();
    return Legalized;
  }
 
  // With a variable index, we can't perform the operation in a smaller type, so
  // we're forced to expand this.
  //
  // TODO: We could emit a chain of compare/select to figure out which piece to
  // index.
  return lowerExtractInsertVectorElt(MI);
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::reduceLoadStoreWidth(GLoadStore &LdStMI, unsigned TypeIdx,
                                      LLT NarrowTy) {
  // FIXME: Don't know how to handle secondary types yet.
  if (TypeIdx != 0)
    return UnableToLegalize;
 
  // This implementation doesn't work for atomics. Give up instead of doing
  // something invalid.
  if (LdStMI.isAtomic())
    return UnableToLegalize;
 
  bool IsLoad = isa<GLoad>(LdStMI);
  Register ValReg = LdStMI.getReg(0);
  Register AddrReg = LdStMI.getPointerReg();
  LLT ValTy = MRI.getType(ValReg);
 
  // FIXME: Do we need a distinct NarrowMemory legalize action?
  if (ValTy.getSizeInBits() != 8 * LdStMI.getMemSize()) {
    LLVM_DEBUG(dbgs() << "Can't narrow extload/truncstore\n");
    return UnableToLegalize;
  }
 
  int NumParts = -1;
  int NumLeftover = -1;
  LLT LeftoverTy;
  SmallVector<Register, 8> NarrowRegs, NarrowLeftoverRegs;
  if (IsLoad) {
    std::tie(NumParts, NumLeftover) = getNarrowTypeBreakDown(ValTy, NarrowTy, LeftoverTy);
  } else {
    if (extractParts(ValReg, ValTy, NarrowTy, LeftoverTy, NarrowRegs,
                     NarrowLeftoverRegs)) {
      NumParts = NarrowRegs.size();
      NumLeftover = NarrowLeftoverRegs.size();
    }
  }
 
  if (NumParts == -1)
    return UnableToLegalize;
 
  LLT PtrTy = MRI.getType(AddrReg);
  const LLT OffsetTy = LLT::scalar(PtrTy.getSizeInBits());
 
  unsigned TotalSize = ValTy.getSizeInBits();
 
  // Split the load/store into PartTy sized pieces starting at Offset. If this
  // is a load, return the new registers in ValRegs. For a store, each elements
  // of ValRegs should be PartTy. Returns the next offset that needs to be
  // handled.
  bool isBigEndian = MIRBuilder.getDataLayout().isBigEndian();
  auto MMO = LdStMI.getMMO();
  auto splitTypePieces = [=](LLT PartTy, SmallVectorImpl<Register> &ValRegs,
                             unsigned NumParts, unsigned Offset) -> unsigned {
    MachineFunction &MF = MIRBuilder.getMF();
    unsigned PartSize = PartTy.getSizeInBits();
    for (unsigned Idx = 0, E = NumParts; Idx != E && Offset < TotalSize;
         ++Idx) {
      unsigned ByteOffset = Offset / 8;
      Register NewAddrReg;
 
      MIRBuilder.materializePtrAdd(NewAddrReg, AddrReg, OffsetTy, ByteOffset);
 
      MachineMemOperand *NewMMO =
          MF.getMachineMemOperand(&MMO, ByteOffset, PartTy);
 
      if (IsLoad) {
        Register Dst = MRI.createGenericVirtualRegister(PartTy);
        ValRegs.push_back(Dst);
        MIRBuilder.buildLoad(Dst, NewAddrReg, *NewMMO);
      } else {
        MIRBuilder.buildStore(ValRegs[Idx], NewAddrReg, *NewMMO);
      }
      Offset = isBigEndian ? Offset - PartSize : Offset + PartSize;
    }
 
    return Offset;
  };
 
  unsigned Offset = isBigEndian ? TotalSize - NarrowTy.getSizeInBits() : 0;
  unsigned HandledOffset =
      splitTypePieces(NarrowTy, NarrowRegs, NumParts, Offset);
 
  // Handle the rest of the register if this isn't an even type breakdown.
  if (LeftoverTy.isValid())
    splitTypePieces(LeftoverTy, NarrowLeftoverRegs, NumLeftover, HandledOffset);
 
  if (IsLoad) {
    insertParts(ValReg, ValTy, NarrowTy, NarrowRegs,
                LeftoverTy, NarrowLeftoverRegs);
  }
 
  LdStMI.eraseFromParent();
  return Legalized;
}
 
LegalizerHelper::LegalizeResult
LegalizerHelper::fewerElementsVector(MachineInstr &MI, unsigned TypeIdx,
                                     LLT NarrowTy) {
  using namespace TargetOpcode;
  GenericMachineInstr &GMI = cast<GenericMachineInstr>(MI);
  unsigned NumElts = NarrowTy.isVector() ? NarrowTy.getNumElements() : 1;
 
  switch (MI.getOpcode()) {
  case G_IMPLICIT_DEF:
  case G_TRUNC:
  case G_AND:
  case G_OR:
  case G_XOR:
  case G_ADD:
  case G_SUB:
  case G_MUL:
  case G_PTR_ADD:
  case G_SMULH:
  case G_UMULH:
  case G_FADD:
  case G_FMUL:
  case G_FSUB:
  case G_FNEG:
  case G_FABS:
  case G_FCANONICALIZE:
  case G_FDIV:
  case G_FREM:
  case G_FMA:
  case G_FMAD:
  case G_FPOW:
  case G_FEXP:
  case G_FEXP2:
  case G_FEXP10:
  case G_FLOG:
  case G_FLOG2:
  case G_FLOG10:
  case G_FLDEXP:
  case G_FNEARBYINT:
  case G_FCEIL:
  case G_FFLOOR:
  case G_FRINT:
  case G_INTRINSIC_ROUND:
  case G_INTRINSIC_ROUNDEVEN:
  case G_INTRINSIC_TRUNC:
  case G_FCOS:
  case G_FSIN:
  case G_FSQRT:
  case G_BSWAP:
  case G_BITREVERSE:
  case G_SDIV:
  case G_UDIV:
  case G_SREM:
  case G_UREM:
  case G_SDIVREM:
  case G_UDIVREM:
  case G_SMIN:
  case G_SMAX:
  case G_UMIN:
  case G_UMAX:
  case G_ABS:
  case G_FMINNUM:
  case G_FMAXNUM:
  case G_FMINNUM_IEEE:
  case G_FMAXNUM_IEEE:
  case G_FMINIMUM:
  case G_FMAXIMUM:
  case G_FSHL:
  case G_FSHR:
  case G_ROTL:
  case G_ROTR:
  case G_FREEZE:
  case G_SADDSAT:
  case G_SSUBSAT:
  case G_UADDSAT:
  case G_USUBSAT:
  case G_UMULO:
  case G_SMULO:
  case G_SHL:
  case G_LSHR:
  case G_ASHR:
  case G_SSHLSAT:
  case G_USHLSAT:
  case G_CTLZ:
  case G_CTLZ_ZERO_UNDEF:
  case G_CTTZ:
  case G_CTTZ_ZERO_UNDEF:
  case G_CTPOP:
  case G_FCOPYSIGN:
  case G_ZEXT:
  case G_SEXT:
  case G_ANYEXT:
  case G_FPEXT:
  case G_FPTRUNC:
  case G_SITOFP:
  case G_UITOFP:
  case G_FPTOSI:
  case G_FPTOUI:
  case G_INTTOPTR:
  case G_PTRTOINT:
  case G_ADDRSPACE_CAST:
  case G_UADDO:
  case G_USUBO:
  case G_UADDE:
  case G_USUBE:
  case G_SADDO:
  case G_SSUBO:
  case G_SADDE:
  case G_SSUBE:
  case G_STRICT_FADD:
  case G_STRICT_FSUB:
  case G_STRICT_FMUL:
  case G_STRICT_FMA:
  case G_STRICT_FLDEXP:
  case G_FFREXP:
    return fewerElementsVectorMultiEltType(GMI, NumElts);
  case G_ICMP:
  case G_FCMP:
    return fewerElementsVectorMultiEltType(GMI, NumElts, {1 /*cpm predicate*/});
  case G_IS_FPCLASS:
    return fewerElementsVectorMultiEltType(GMI, NumElts, {2, 3 /*mask,fpsem*/});
  case G_SELECT:
    if (MRI.getType(MI.getOperand(1).getReg()).isVector())
      return fewerElementsVectorMultiEltType(GMI, NumElts);
    return fewerElementsVectorMultiEltType(GMI, NumElts, {1 /*scalar cond*/});
  case G_PHI:
    return fewerElementsVectorPhi(GMI, NumElts);
  case G_UNMERGE_VALUES:
    return fewerElementsVectorUnmergeValues(MI, TypeIdx, NarrowTy);
  case G_BUILD_VECTOR:
    assert(TypeIdx == 0 && "not a vector type index");
    return fewerElementsVectorMerge(MI, TypeIdx, NarrowTy);
  case G_CONCAT_VECTORS:
    if (TypeIdx != 1) // TODO: This probably does work as expected already.
      return UnableToLegalize;
    return fewerElementsVectorMerge(MI, TypeIdx, NarrowTy);
  case G_EXTRACT_VECTOR_ELT:
  case G_INSERT_VECTOR_ELT:
    return fewerElementsVectorExtractInsertVectorElt(MI, TypeIdx, NarrowTy);
  case G_LOAD:
  case G_STORE:
    return reduceLoadStoreWidth(cast<GLoadStore>(MI), TypeIdx, NarrowTy);
  case G_SEXT_INREG:
    return fewerElementsVectorMultiEltType(GMI, NumElts, {2 /*imm*/});
  GISEL_VECREDUCE_CASES_NONSEQ
    return fewerElementsVectorReductions(MI, TypeIdx, NarrowTy);
  case G_SHUFFLE_VECTOR:
    return fewerElementsVectorShuffle(MI, TypeIdx, NarrowTy);
  default:
    return UnableToLegalize;
  }
}
 
LegalizerHelper::LegalizeResult LegalizerHelper::fewerElementsVectorShuffle(
    MachineInstr &MI, unsigned int TypeIdx, LLT NarrowTy) {
  assert(MI.getOpcode() == TargetOpcode::G_SHUFFLE_VECTOR);
  if (TypeIdx != 0)
    return UnableToLegalize;
 
  auto [DstReg, DstTy, Src1Reg, Src1Ty, Src2Reg, Src2Ty] =
      MI.getFirst3RegLLTs();
  ArrayRef<int> Mask = MI.getOperand(3).getShuffleMask();
  // The shuffle should be canonicalized by now.
  if (DstTy != Src1Ty)
    return UnableToLegalize;
  if (DstTy != Src2Ty)
    return UnableToLegalize;
 
  if (!isPowerOf2_32(DstTy.getNumElements()))
    return UnableToLegalize;
 
  // We only support splitting a shuffle into 2, so adjust NarrowTy accordingly.
  // Further legalization attempts will be needed to do split further.
  NarrowTy =
      DstTy.changeElementCount(DstTy.getElementCount().divideCoefficientBy(2));
  unsigned NewElts = NarrowTy.getNumElements();
 
  SmallVector<Register> SplitSrc1Regs, SplitSrc2Regs;
  extractParts(Src1Reg, NarrowTy, 2, SplitSrc1Regs);
  extractParts(Src2Reg, NarrowTy, 2, SplitSrc2Regs);
  Register Inputs[4] = {SplitSrc1Regs[0], SplitSrc1Regs[1], SplitSrc2Regs[0],
                        SplitSrc2Regs[1]};
 
  Register Hi, Lo;
 
  // If Lo or Hi uses elements from at most two of the four input vectors, then
  // express it as a vector shuffle of those two inputs.  Otherwise extract the
  // input elements by hand and construct the Lo/Hi output using a BUILD_VECTOR.
  SmallVector<int, 16> Ops;
  for (unsigned High = 0; High < 2; ++High) {
    Register &Output = High ? Hi : Lo;
 
    // Build a shuffle mask for the output, discovering on the fly which
    // input vectors to use as shuffle operands (recorded in InputUsed).
    // If building a suitable shuffle vector proves too hard, then bail
    // out with useBuildVector set.
    unsigned InputUsed[2] = {-1U, -1U}; // Not yet discovered.
    unsigned FirstMaskIdx = High * NewElts;
    bool UseBuildVector = false;
    for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) {
      // The mask element.  This indexes into the input.
      int Idx = Mask[FirstMaskIdx + MaskOffset];
 
      // The input vector this mask element indexes into.
      unsigned Input = (unsigned)Idx / NewElts;
 
      if (Input >= std::size(Inputs)) {
        // The mask element does not index into any input vector.
        Ops.push_back(-1);
        continue;
      }
 
      // Turn the index into an offset from the start of the input vector.
      Idx -= Input * NewElts;
 
      // Find or create a shuffle vector operand to hold this input.
      unsigned OpNo;
      for (OpNo = 0; OpNo < std::size(InputUsed); ++OpNo) {
        if (InputUsed[OpNo] == Input) {
          // This input vector is already an operand.
          break;
        } else if (InputUsed[OpNo] == -1U) {
          // Create a new operand for this input vector.
          InputUsed[OpNo] = Input;
          break;
        }
      }
 
      if (OpNo >= std::size(InputUsed)) {
        // More than two input vectors used!  Give up on trying to create a
        // shuffle vector.  Insert all elements into a BUILD_VECTOR instead.
        UseBuildVector = true;
        break;
      }
 
      // Add the mask index for the new shuffle vector.
      Ops.push_back(Idx + OpNo * NewElts);
    }
 
    if (UseBuildVector) {
      LLT EltTy = NarrowTy.getElementType();
      SmallVector<Register, 16> SVOps;
 
      // Extract the input elements by hand.
      for (unsigned MaskOffset = 0; MaskOffset < NewElts; ++MaskOffset) {
        // The mask element.  This indexes into the input.
        int Idx = Mask[FirstMaskIdx + MaskOffset];
 
        // The input vector this mask element indexes into.
        unsigned Input = (unsigned)Idx / NewElts;
 
        if (Input >= std::size(Inputs)) {
          // The mask element is "undef" or indexes off the end of the input.
          SVOps.push_back(MIRBuilder.buildUndef(EltTy).getReg(0));
          continue;
        }
 
        // Turn the index into an offset from the start of the input vector.
        Idx -= Input * NewElts;
 
        // Extract the vector element by hand.
        SVOps.push_back(MIRBuilder
                            .buildExtractVectorElement(
                                EltTy, Inputs[Input],
                                MIRBuilder.buildConstant(LLT::scalar(32), Idx))
                            .getReg(0));
      }
 
      // Construct the Lo/Hi output using a G_BUILD_VECTOR.
      Output = MIRBuilder.buildBuildVector(NarrowTy, SVOps).getReg(0);
    } else if (InputUsed[0] == -1U) {
      // No input vectors were used! The result is undefined.
      Output = MIRBuilder.buildUndef(NarrowTy).getReg(0);
    } else {
      Register Op0 = Inputs[InputUsed[0]];
      // If only one input was used, use an undefined vector for the other.
      Register Op1 = InputUsed[1] == -1U
                         ? MIRBuilder.buildUndef(NarrowTy).getReg(0)
                         : Inputs[InputUsed[1]];
      // At least one input vector was used. Create a new shuffle vector.
      Output = MIRBuilder.buildShuffleVector(NarrowTy, Op0, Op1, Ops).getReg(0);
    }
 
    Ops.clear();
  }
 
  MIRBuilder.buildConcatVectors(DstReg, {Lo, Hi});
  MI.eraseFromParent();
  return