llvm-project/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

//===- AMDGPULegalizerInfo.cpp -----------------------------------*- C++ -*-==//
//
// 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 targeting of the Machinelegalizer class for
/// AMDGPU.
/// \todo This should be generated by TableGen.
//===----------------------------------------------------------------------===//

#if defined(_MSC_VER) || defined(__MINGW32__)
// According to Microsoft, one must set _USE_MATH_DEFINES in order to get M_PI
// from the Visual C++ cmath / math.h headers:
// https://docs.microsoft.com/en-us/cpp/c-runtime-library/math-constants?view=vs-2019
#define _USE_MATH_DEFINES
#endif

#include "AMDGPULegalizerInfo.h"

#include "AMDGPU.h"
#include "AMDGPUGlobalISelUtils.h"
#include "AMDGPUTargetMachine.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/ADT/ScopeExit.h"
#include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
#include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/Debug.h"

#define DEBUG_TYPE "amdgpu-legalinfo"

using namespace llvm;
using namespace LegalizeActions;
using namespace LegalizeMutations;
using namespace LegalityPredicates;
using namespace MIPatternMatch;

// Round the number of elements to the next power of two elements
static LLT getPow2VectorType(LLT Ty) {
  unsigned NElts = Ty.getNumElements();
  unsigned Pow2NElts = 1 <<  Log2_32_Ceil(NElts);
  return Ty.changeNumElements(Pow2NElts);
}

// Round the number of bits to the next power of two bits
static LLT getPow2ScalarType(LLT Ty) {
  unsigned Bits = Ty.getSizeInBits();
  unsigned Pow2Bits = 1 <<  Log2_32_Ceil(Bits);
  return LLT::scalar(Pow2Bits);
}

static LegalityPredicate isMultiple32(unsigned TypeIdx,
                                      unsigned MaxSize = 1024) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    const LLT EltTy = Ty.getScalarType();
    return Ty.getSizeInBits() <= MaxSize && EltTy.getSizeInBits() % 32 == 0;
  };
}

static LegalityPredicate sizeIs(unsigned TypeIdx, unsigned Size) {
  return [=](const LegalityQuery &Query) {
    return Query.Types[TypeIdx].getSizeInBits() == Size;
  };
}

static LegalityPredicate isSmallOddVector(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    return Ty.isVector() &&
           Ty.getNumElements() % 2 != 0 &&
           Ty.getElementType().getSizeInBits() < 32 &&
           Ty.getSizeInBits() % 32 != 0;
  };
}

static LegalityPredicate isWideVec16(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    const LLT EltTy = Ty.getScalarType();
    return EltTy.getSizeInBits() == 16 && Ty.getNumElements() > 2;
  };
}

static LegalizeMutation oneMoreElement(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    const LLT EltTy = Ty.getElementType();
    return std::make_pair(TypeIdx, LLT::vector(Ty.getNumElements() + 1, EltTy));
  };
}

static LegalizeMutation fewerEltsToSize64Vector(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    const LLT EltTy = Ty.getElementType();
    unsigned Size = Ty.getSizeInBits();
    unsigned Pieces = (Size + 63) / 64;
    unsigned NewNumElts = (Ty.getNumElements() + 1) / Pieces;
    return std::make_pair(TypeIdx, LLT::scalarOrVector(NewNumElts, EltTy));
  };
}

// Increase the number of vector elements to reach the next multiple of 32-bit
// type.
static LegalizeMutation moreEltsToNext32Bit(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];

    const LLT EltTy = Ty.getElementType();
    const int Size = Ty.getSizeInBits();
    const int EltSize = EltTy.getSizeInBits();
    const int NextMul32 = (Size + 31) / 32;

    assert(EltSize < 32);

    const int NewNumElts = (32 * NextMul32 + EltSize - 1) / EltSize;
    return std::make_pair(TypeIdx, LLT::vector(NewNumElts, EltTy));
  };
}

static LegalityPredicate vectorSmallerThan(unsigned TypeIdx, unsigned Size) {
  return [=](const LegalityQuery &Query) {
    const LLT QueryTy = Query.Types[TypeIdx];
    return QueryTy.isVector() && QueryTy.getSizeInBits() < Size;
  };
}

static LegalityPredicate vectorWiderThan(unsigned TypeIdx, unsigned Size) {
  return [=](const LegalityQuery &Query) {
    const LLT QueryTy = Query.Types[TypeIdx];
    return QueryTy.isVector() && QueryTy.getSizeInBits() > Size;
  };
}

static LegalityPredicate numElementsNotEven(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT QueryTy = Query.Types[TypeIdx];
    return QueryTy.isVector() && QueryTy.getNumElements() % 2 != 0;
  };
}

// Any combination of 32 or 64-bit elements up to 1024 bits, and multiples of
// v2s16.
static LegalityPredicate isRegisterType(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    if (Ty.isVector()) {
      const int EltSize = Ty.getElementType().getSizeInBits();
      return EltSize == 32 || EltSize == 64 ||
            (EltSize == 16 && Ty.getNumElements() % 2 == 0) ||
             EltSize == 128 || EltSize == 256;
    }

    return Ty.getSizeInBits() % 32 == 0 && Ty.getSizeInBits() <= 1024;
  };
}

static LegalityPredicate elementTypeIs(unsigned TypeIdx, LLT Type) {
  return [=](const LegalityQuery &Query) {
    const LLT QueryTy = Query.Types[TypeIdx];
    return QueryTy.isVector() && QueryTy.getElementType() == Type;
  };
}

static LegalityPredicate elementTypeIsLegal(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT QueryTy = Query.Types[TypeIdx];
    if (!QueryTy.isVector())
      return false;
    const LLT EltTy = QueryTy.getElementType();
    return EltTy == LLT::scalar(16) || EltTy.getSizeInBits() >= 32;
  };
}

static LegalityPredicate isWideScalarTruncStore(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    return !Ty.isVector() && Ty.getSizeInBits() > 32 &&
           Query.MMODescrs[0].SizeInBits < Ty.getSizeInBits();
  };
}

static LegalityPredicate smallerThan(unsigned TypeIdx0, unsigned TypeIdx1) {
  return [=](const LegalityQuery &Query) {
    return Query.Types[TypeIdx0].getSizeInBits() <
           Query.Types[TypeIdx1].getSizeInBits();
  };
}

static LegalityPredicate greaterThan(unsigned TypeIdx0, unsigned TypeIdx1) {
  return [=](const LegalityQuery &Query) {
    return Query.Types[TypeIdx0].getSizeInBits() >
           Query.Types[TypeIdx1].getSizeInBits();
  };
}

AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
                                         const GCNTargetMachine &TM)
  :  ST(ST_) {
  using namespace TargetOpcode;

  auto GetAddrSpacePtr = [&TM](unsigned AS) {
    return LLT::pointer(AS, TM.getPointerSizeInBits(AS));
  };

  const LLT S1 = LLT::scalar(1);
  const LLT S16 = LLT::scalar(16);
  const LLT S32 = LLT::scalar(32);
  const LLT S64 = LLT::scalar(64);
  const LLT S128 = LLT::scalar(128);
  const LLT S256 = LLT::scalar(256);
  const LLT S512 = LLT::scalar(512);
  const LLT S1024 = LLT::scalar(1024);

  const LLT V2S16 = LLT::vector(2, 16);
  const LLT V4S16 = LLT::vector(4, 16);

  const LLT V2S32 = LLT::vector(2, 32);
  const LLT V3S32 = LLT::vector(3, 32);
  const LLT V4S32 = LLT::vector(4, 32);
  const LLT V5S32 = LLT::vector(5, 32);
  const LLT V6S32 = LLT::vector(6, 32);
  const LLT V7S32 = LLT::vector(7, 32);
  const LLT V8S32 = LLT::vector(8, 32);
  const LLT V9S32 = LLT::vector(9, 32);
  const LLT V10S32 = LLT::vector(10, 32);
  const LLT V11S32 = LLT::vector(11, 32);
  const LLT V12S32 = LLT::vector(12, 32);
  const LLT V13S32 = LLT::vector(13, 32);
  const LLT V14S32 = LLT::vector(14, 32);
  const LLT V15S32 = LLT::vector(15, 32);
  const LLT V16S32 = LLT::vector(16, 32);
  const LLT V32S32 = LLT::vector(32, 32);

  const LLT V2S64 = LLT::vector(2, 64);
  const LLT V3S64 = LLT::vector(3, 64);
  const LLT V4S64 = LLT::vector(4, 64);
  const LLT V5S64 = LLT::vector(5, 64);
  const LLT V6S64 = LLT::vector(6, 64);
  const LLT V7S64 = LLT::vector(7, 64);
  const LLT V8S64 = LLT::vector(8, 64);
  const LLT V16S64 = LLT::vector(16, 64);

  std::initializer_list<LLT> AllS32Vectors =
    {V2S32, V3S32, V4S32, V5S32, V6S32, V7S32, V8S32,
     V9S32, V10S32, V11S32, V12S32, V13S32, V14S32, V15S32, V16S32, V32S32};
  std::initializer_list<LLT> AllS64Vectors =
    {V2S64, V3S64, V4S64, V5S64, V6S64, V7S64, V8S64, V16S64};

  const LLT GlobalPtr = GetAddrSpacePtr(AMDGPUAS::GLOBAL_ADDRESS);
  const LLT ConstantPtr = GetAddrSpacePtr(AMDGPUAS::CONSTANT_ADDRESS);
  const LLT Constant32Ptr = GetAddrSpacePtr(AMDGPUAS::CONSTANT_ADDRESS_32BIT);
  const LLT LocalPtr = GetAddrSpacePtr(AMDGPUAS::LOCAL_ADDRESS);
  const LLT RegionPtr = GetAddrSpacePtr(AMDGPUAS::REGION_ADDRESS);
  const LLT FlatPtr = GetAddrSpacePtr(AMDGPUAS::FLAT_ADDRESS);
  const LLT PrivatePtr = GetAddrSpacePtr(AMDGPUAS::PRIVATE_ADDRESS);

  const LLT CodePtr = FlatPtr;

  const std::initializer_list<LLT> AddrSpaces64 = {
    GlobalPtr, ConstantPtr, FlatPtr
  };

  const std::initializer_list<LLT> AddrSpaces32 = {
    LocalPtr, PrivatePtr, Constant32Ptr, RegionPtr
  };

  const std::initializer_list<LLT> FPTypesBase = {
    S32, S64
  };

  const std::initializer_list<LLT> FPTypes16 = {
    S32, S64, S16
  };

  const std::initializer_list<LLT> FPTypesPK16 = {
    S32, S64, S16, V2S16
  };

  const LLT MinScalarFPTy = ST.has16BitInsts() ? S16 : S32;

  setAction({G_BRCOND, S1}, Legal); // VCC branches
  setAction({G_BRCOND, S32}, Legal); // SCC branches

  // TODO: All multiples of 32, vectors of pointers, all v2s16 pairs, more
  // elements for v3s16
  getActionDefinitionsBuilder(G_PHI)
    .legalFor({S32, S64, V2S16, V4S16, S1, S128, S256})
    .legalFor(AllS32Vectors)
    .legalFor(AllS64Vectors)
    .legalFor(AddrSpaces64)
    .legalFor(AddrSpaces32)
    .clampScalar(0, S32, S256)
    .widenScalarToNextPow2(0, 32)
    .clampMaxNumElements(0, S32, 16)
    .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
    .legalIf(isPointer(0));

  if (ST.hasVOP3PInsts()) {
    getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL})
      .legalFor({S32, S16, V2S16})
      .clampScalar(0, S16, S32)
      .clampMaxNumElements(0, S16, 2)
      .scalarize(0)
      .widenScalarToNextPow2(0, 32);
  } else if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL})
      .legalFor({S32, S16})
      .clampScalar(0, S16, S32)
      .scalarize(0)
      .widenScalarToNextPow2(0, 32);
  } else {
    getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL})
      .legalFor({S32})
      .clampScalar(0, S32, S32)
      .scalarize(0);
  }

  // FIXME: Not really legal. Placeholder for custom lowering.
  getActionDefinitionsBuilder({G_SDIV, G_UDIV, G_SREM, G_UREM})
    .customFor({S32, S64})
    .clampScalar(0, S32, S64)
    .widenScalarToNextPow2(0, 32)
    .scalarize(0);

  getActionDefinitionsBuilder({G_UMULH, G_SMULH})
    .legalFor({S32})
    .clampScalar(0, S32, S32)
    .scalarize(0);

  // Report legal for any types we can handle anywhere. For the cases only legal
  // on the SALU, RegBankSelect will be able to re-legalize.
  getActionDefinitionsBuilder({G_AND, G_OR, G_XOR})
    .legalFor({S32, S1, S64, V2S32, S16, V2S16, V4S16})
    .clampScalar(0, S32, S64)
    .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
    .fewerElementsIf(vectorWiderThan(0, 64), fewerEltsToSize64Vector(0))
    .widenScalarToNextPow2(0)
    .scalarize(0);

  getActionDefinitionsBuilder({G_UADDO, G_USUBO,
                               G_UADDE, G_SADDE, G_USUBE, G_SSUBE})
    .legalFor({{S32, S1}, {S32, S32}})
    .minScalar(0, S32)
    // TODO: .scalarize(0)
    .lower();

  getActionDefinitionsBuilder(G_BITCAST)
    // Don't worry about the size constraint.
    .legalIf(all(isRegisterType(0), isRegisterType(1)))
    .lower();


  getActionDefinitionsBuilder(G_CONSTANT)
    .legalFor({S1, S32, S64, S16, GlobalPtr,
               LocalPtr, ConstantPtr, PrivatePtr, FlatPtr })
    .clampScalar(0, S32, S64)
    .widenScalarToNextPow2(0)
    .legalIf(isPointer(0));

  getActionDefinitionsBuilder(G_FCONSTANT)
    .legalFor({S32, S64, S16})
    .clampScalar(0, S16, S64);

  getActionDefinitionsBuilder({G_IMPLICIT_DEF, G_FREEZE})
      .legalFor({S1, S32, S64, S16, V2S32, V4S32, V2S16, V4S16, GlobalPtr,
                 ConstantPtr, LocalPtr, FlatPtr, PrivatePtr})
      .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
      .clampScalarOrElt(0, S32, S1024)
      .legalIf(isMultiple32(0))
      .widenScalarToNextPow2(0, 32)
      .clampMaxNumElements(0, S32, 16);

  setAction({G_FRAME_INDEX, PrivatePtr}, Legal);
  getActionDefinitionsBuilder(G_GLOBAL_VALUE)
    .unsupportedFor({PrivatePtr})
    .custom();
  setAction({G_BLOCK_ADDR, CodePtr}, Legal);

  auto &FPOpActions = getActionDefinitionsBuilder(
    { G_FADD, G_FMUL, G_FMA, G_FCANONICALIZE})
    .legalFor({S32, S64});
  auto &TrigActions = getActionDefinitionsBuilder({G_FSIN, G_FCOS})
    .customFor({S32, S64});
  auto &FDIVActions = getActionDefinitionsBuilder(G_FDIV)
    .customFor({S32, S64});

  if (ST.has16BitInsts()) {
    if (ST.hasVOP3PInsts())
      FPOpActions.legalFor({S16, V2S16});
    else
      FPOpActions.legalFor({S16});

    TrigActions.customFor({S16});
    FDIVActions.customFor({S16});
  }

  auto &MinNumMaxNum = getActionDefinitionsBuilder({
      G_FMINNUM, G_FMAXNUM, G_FMINNUM_IEEE, G_FMAXNUM_IEEE});

  if (ST.hasVOP3PInsts()) {
    MinNumMaxNum.customFor(FPTypesPK16)
      .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
      .clampMaxNumElements(0, S16, 2)
      .clampScalar(0, S16, S64)
      .scalarize(0);
  } else if (ST.has16BitInsts()) {
    MinNumMaxNum.customFor(FPTypes16)
      .clampScalar(0, S16, S64)
      .scalarize(0);
  } else {
    MinNumMaxNum.customFor(FPTypesBase)
      .clampScalar(0, S32, S64)
      .scalarize(0);
  }

  if (ST.hasVOP3PInsts())
    FPOpActions.clampMaxNumElements(0, S16, 2);

  FPOpActions
    .scalarize(0)
    .clampScalar(0, ST.has16BitInsts() ? S16 : S32, S64);

  TrigActions
    .scalarize(0)
    .clampScalar(0, ST.has16BitInsts() ? S16 : S32, S64);

  FDIVActions
    .scalarize(0)
    .clampScalar(0, ST.has16BitInsts() ? S16 : S32, S64);

  getActionDefinitionsBuilder({G_FNEG, G_FABS})
    .legalFor(FPTypesPK16)
    .clampMaxNumElements(0, S16, 2)
    .scalarize(0)
    .clampScalar(0, S16, S64);

  if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder({G_FSQRT, G_FFLOOR})
      .legalFor({S32, S64, S16})
      .scalarize(0)
      .clampScalar(0, S16, S64);
  } else {
    getActionDefinitionsBuilder(G_FSQRT)
      .legalFor({S32, S64})
      .scalarize(0)
      .clampScalar(0, S32, S64);

    if (ST.hasFractBug()) {
      getActionDefinitionsBuilder(G_FFLOOR)
        .customFor({S64})
        .legalFor({S32, S64})
        .scalarize(0)
        .clampScalar(0, S32, S64);
    } else {
      getActionDefinitionsBuilder(G_FFLOOR)
        .legalFor({S32, S64})
        .scalarize(0)
        .clampScalar(0, S32, S64);
    }
  }

  getActionDefinitionsBuilder(G_FPTRUNC)
    .legalFor({{S32, S64}, {S16, S32}})
    .scalarize(0)
    .lower();

  getActionDefinitionsBuilder(G_FPEXT)
    .legalFor({{S64, S32}, {S32, S16}})
    .lowerFor({{S64, S16}}) // FIXME: Implement
    .scalarize(0);

  getActionDefinitionsBuilder(G_FSUB)
      // Use actual fsub instruction
      .legalFor({S32})
      // Must use fadd + fneg
      .lowerFor({S64, S16, V2S16})
      .scalarize(0)
      .clampScalar(0, S32, S64);

  // Whether this is legal depends on the floating point mode for the function.
  auto &FMad = getActionDefinitionsBuilder(G_FMAD);
  if (ST.hasMadF16())
    FMad.customFor({S32, S16});
  else
    FMad.customFor({S32});
  FMad.scalarize(0)
      .lower();

  // TODO: Do we need to clamp maximum bitwidth?
  getActionDefinitionsBuilder(G_TRUNC)
    .legalIf(isScalar(0))
    .legalFor({{V2S16, V2S32}})
    .clampMaxNumElements(0, S16, 2)
    // Avoid scalarizing in cases that should be truly illegal. In unresolvable
    // situations (like an invalid implicit use), we don't want to infinite loop
    // in the legalizer.
    .fewerElementsIf(elementTypeIsLegal(0), LegalizeMutations::scalarize(0))
    .alwaysLegal();

  getActionDefinitionsBuilder({G_SEXT, G_ZEXT, G_ANYEXT})
    .legalFor({{S64, S32}, {S32, S16}, {S64, S16},
               {S32, S1}, {S64, S1}, {S16, S1}})
    .scalarize(0)
    .clampScalar(0, S32, S64)
    .widenScalarToNextPow2(1, 32);

  // TODO: Split s1->s64 during regbankselect for VALU.
  auto &IToFP = getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
    .legalFor({{S32, S32}, {S64, S32}, {S16, S32}})
    .lowerFor({{S32, S64}})
    .lowerIf(typeIs(1, S1))
    .customFor({{S64, S64}});
  if (ST.has16BitInsts())
    IToFP.legalFor({{S16, S16}});
  IToFP.clampScalar(1, S32, S64)
       .scalarize(0)
       .widenScalarToNextPow2(1);

  auto &FPToI = getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
    .legalFor({{S32, S32}, {S32, S64}, {S32, S16}})
    .customFor({{S64, S64}});
  if (ST.has16BitInsts())
    FPToI.legalFor({{S16, S16}});
  else
    FPToI.minScalar(1, S32);

  FPToI.minScalar(0, S32)
       .scalarize(0)
       .lower();

  getActionDefinitionsBuilder(G_INTRINSIC_ROUND)
    .scalarize(0)
    .lower();

  if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
      .legalFor({S16, S32, S64})
      .clampScalar(0, S16, S64)
      .scalarize(0);
  } else if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
    getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
      .legalFor({S32, S64})
      .clampScalar(0, S32, S64)
      .scalarize(0);
  } else {
    getActionDefinitionsBuilder({G_INTRINSIC_TRUNC, G_FCEIL, G_FRINT})
      .legalFor({S32})
      .customFor({S64})
      .clampScalar(0, S32, S64)
      .scalarize(0);
  }

  getActionDefinitionsBuilder(G_PTR_ADD)
    .scalarize(0)
    .alwaysLegal();

  // TODO: Clamp mask to pointer sizes
  getActionDefinitionsBuilder(G_PTRMASK)
    .scalarize(0)
    .alwaysLegal();

  auto &CmpBuilder =
    getActionDefinitionsBuilder(G_ICMP)
    // The compare output type differs based on the register bank of the output,
    // so make both s1 and s32 legal.
    //
    // Scalar compares producing output in scc will be promoted to s32, as that
    // is the allocatable register type that will be needed for the copy from
    // scc. This will be promoted during RegBankSelect, and we assume something
    // before that won't try to use s32 result types.
    //
    // Vector compares producing an output in vcc/SGPR will use s1 in VCC reg
    // bank.
    .legalForCartesianProduct(
      {S1}, {S32, S64, GlobalPtr, LocalPtr, ConstantPtr, PrivatePtr, FlatPtr})
    .legalForCartesianProduct(
      {S32}, {S32, S64, GlobalPtr, LocalPtr, ConstantPtr, PrivatePtr, FlatPtr});
  if (ST.has16BitInsts()) {
    CmpBuilder.legalFor({{S1, S16}});
  }

  CmpBuilder
    .widenScalarToNextPow2(1)
    .clampScalar(1, S32, S64)
    .scalarize(0)
    .legalIf(all(typeInSet(0, {S1, S32}), isPointer(1)));

  getActionDefinitionsBuilder(G_FCMP)
    .legalForCartesianProduct({S1}, ST.has16BitInsts() ? FPTypes16 : FPTypesBase)
    .widenScalarToNextPow2(1)
    .clampScalar(1, S32, S64)
    .scalarize(0);

  // FIXME: fpow has a selection pattern that should move to custom lowering.
  auto &Exp2Ops = getActionDefinitionsBuilder({G_FEXP2, G_FLOG2});
  if (ST.has16BitInsts())
    Exp2Ops.legalFor({S32, S16});
  else
    Exp2Ops.legalFor({S32});
  Exp2Ops.clampScalar(0, MinScalarFPTy, S32);
  Exp2Ops.scalarize(0);

  auto &ExpOps = getActionDefinitionsBuilder({G_FEXP, G_FLOG, G_FLOG10, G_FPOW});
  if (ST.has16BitInsts())
    ExpOps.customFor({{S32}, {S16}});
  else
    ExpOps.customFor({S32});
  ExpOps.clampScalar(0, MinScalarFPTy, S32)
        .scalarize(0);

  // The 64-bit versions produce 32-bit results, but only on the SALU.
  getActionDefinitionsBuilder(G_CTPOP)
    .legalFor({{S32, S32}, {S32, S64}})
    .clampScalar(0, S32, S32)
    .clampScalar(1, S32, S64)
    .scalarize(0)
    .widenScalarToNextPow2(0, 32)
    .widenScalarToNextPow2(1, 32);

  // The hardware instructions return a different result on 0 than the generic
  // instructions expect. The hardware produces -1, but these produce the
  // bitwidth.
  getActionDefinitionsBuilder({G_CTLZ, G_CTTZ})
    .scalarize(0)
    .clampScalar(0, S32, S32)
    .clampScalar(1, S32, S64)
    .widenScalarToNextPow2(0, 32)
    .widenScalarToNextPow2(1, 32)
    .lower();

  // The 64-bit versions produce 32-bit results, but only on the SALU.
  getActionDefinitionsBuilder({G_CTLZ_ZERO_UNDEF, G_CTTZ_ZERO_UNDEF})
    .legalFor({{S32, S32}, {S32, S64}})
    .clampScalar(0, S32, S32)
    .clampScalar(1, S32, S64)
    .scalarize(0)
    .widenScalarToNextPow2(0, 32)
    .widenScalarToNextPow2(1, 32);

  getActionDefinitionsBuilder(G_BITREVERSE)
    .legalFor({S32})
    .clampScalar(0, S32, S32)
    .scalarize(0);

  if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder(G_BSWAP)
      .legalFor({S16, S32, V2S16})
      .clampMaxNumElements(0, S16, 2)
      // FIXME: Fixing non-power-of-2 before clamp is workaround for
      // narrowScalar limitation.
      .widenScalarToNextPow2(0)
      .clampScalar(0, S16, S32)
      .scalarize(0);

    if (ST.hasVOP3PInsts()) {
      getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
        .legalFor({S32, S16, V2S16})
        .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
        .clampMaxNumElements(0, S16, 2)
        .minScalar(0, S16)
        .widenScalarToNextPow2(0)
        .scalarize(0)
        .lower();
    } else {
      getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
        .legalFor({S32, S16})
        .widenScalarToNextPow2(0)
        .minScalar(0, S16)
        .scalarize(0)
        .lower();
    }
  } else {
    // TODO: Should have same legality without v_perm_b32
    getActionDefinitionsBuilder(G_BSWAP)
      .legalFor({S32})
      .lowerIf(narrowerThan(0, 32))
      // FIXME: Fixing non-power-of-2 before clamp is workaround for
      // narrowScalar limitation.
      .widenScalarToNextPow2(0)
      .maxScalar(0, S32)
      .scalarize(0)
      .lower();

    getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
      .legalFor({S32})
      .minScalar(0, S32)
      .widenScalarToNextPow2(0)
      .scalarize(0)
      .lower();
  }

  getActionDefinitionsBuilder(G_INTTOPTR)
    // List the common cases
    .legalForCartesianProduct(AddrSpaces64, {S64})
    .legalForCartesianProduct(AddrSpaces32, {S32})
    .scalarize(0)
    // Accept any address space as long as the size matches
    .legalIf(sameSize(0, 1))
    .widenScalarIf(smallerThan(1, 0),
      [](const LegalityQuery &Query) {
        return std::make_pair(1, LLT::scalar(Query.Types[0].getSizeInBits()));
      })
    .narrowScalarIf(greaterThan(1, 0),
      [](const LegalityQuery &Query) {
        return std::make_pair(1, LLT::scalar(Query.Types[0].getSizeInBits()));
      });

  getActionDefinitionsBuilder(G_PTRTOINT)
    // List the common cases
    .legalForCartesianProduct(AddrSpaces64, {S64})
    .legalForCartesianProduct(AddrSpaces32, {S32})
    .scalarize(0)
    // Accept any address space as long as the size matches
    .legalIf(sameSize(0, 1))
    .widenScalarIf(smallerThan(0, 1),
      [](const LegalityQuery &Query) {
        return std::make_pair(0, LLT::scalar(Query.Types[1].getSizeInBits()));
      })
    .narrowScalarIf(
      greaterThan(0, 1),
      [](const LegalityQuery &Query) {
        return std::make_pair(0, LLT::scalar(Query.Types[1].getSizeInBits()));
      });

  getActionDefinitionsBuilder(G_ADDRSPACE_CAST)
    .scalarize(0)
    .custom();

  // TODO: Should load to s16 be legal? Most loads extend to 32-bits, but we
  // handle some operations by just promoting the register during
  // selection. There are also d16 loads on GFX9+ which preserve the high bits.
  auto maxSizeForAddrSpace = [this](unsigned AS, bool IsLoad) -> unsigned {
    switch (AS) {
    // FIXME: Private element size.
    case AMDGPUAS::PRIVATE_ADDRESS:
      return 32;
    // FIXME: Check subtarget
    case AMDGPUAS::LOCAL_ADDRESS:
      return ST.useDS128() ? 128 : 64;

    // Treat constant and global as identical. SMRD loads are sometimes usable
    // for global loads (ideally constant address space should be eliminated)
    // depending on the context. Legality cannot be context dependent, but
    // RegBankSelect can split the load as necessary depending on the pointer
    // register bank/uniformity and if the memory is invariant or not written in
    // a kernel.
    case AMDGPUAS::CONSTANT_ADDRESS:
    case AMDGPUAS::GLOBAL_ADDRESS:
      return IsLoad ? 512 : 128;
    default:
      return 128;
    }
  };

  const auto needToSplitMemOp = [=](const LegalityQuery &Query,
                                    bool IsLoad) -> bool {
    const LLT DstTy = Query.Types[0];

    // Split vector extloads.
    unsigned MemSize = Query.MMODescrs[0].SizeInBits;
    unsigned Align = Query.MMODescrs[0].AlignInBits;

    if (MemSize < DstTy.getSizeInBits())
      MemSize = std::max(MemSize, Align);

    if (DstTy.isVector() && DstTy.getSizeInBits() > MemSize)
      return true;

    const LLT PtrTy = Query.Types[1];
    unsigned AS = PtrTy.getAddressSpace();
    if (MemSize > maxSizeForAddrSpace(AS, IsLoad))
      return true;

    // Catch weird sized loads that don't evenly divide into the access sizes
    // TODO: May be able to widen depending on alignment etc.
    unsigned NumRegs = (MemSize + 31) / 32;
    if (NumRegs == 3) {
      if (!ST.hasDwordx3LoadStores())
        return true;
    } else {
      // If the alignment allows, these should have been widened.
      if (!isPowerOf2_32(NumRegs))
        return true;
    }

    if (Align < MemSize) {
      const SITargetLowering *TLI = ST.getTargetLowering();
      return !TLI->allowsMisalignedMemoryAccessesImpl(MemSize, AS, Align / 8);
    }

    return false;
  };

  const auto shouldWidenLoadResult = [=](const LegalityQuery &Query) -> bool {
    unsigned Size = Query.Types[0].getSizeInBits();
    if (isPowerOf2_32(Size))
      return false;

    if (Size == 96 && ST.hasDwordx3LoadStores())
      return false;

    unsigned AddrSpace = Query.Types[1].getAddressSpace();
    if (Size >= maxSizeForAddrSpace(AddrSpace, true))
      return false;

    unsigned Align = Query.MMODescrs[0].AlignInBits;
    unsigned RoundedSize = NextPowerOf2(Size);
    return (Align >= RoundedSize);
  };

  unsigned GlobalAlign32 = ST.hasUnalignedBufferAccess() ? 0 : 32;
  unsigned GlobalAlign16 = ST.hasUnalignedBufferAccess() ? 0 : 16;
  unsigned GlobalAlign8 = ST.hasUnalignedBufferAccess() ? 0 : 8;

  // TODO: Refine based on subtargets which support unaligned access or 128-bit
  // LDS
  // TODO: Unsupported flat for SI.

  for (unsigned Op : {G_LOAD, G_STORE}) {
    const bool IsStore = Op == G_STORE;

    auto &Actions = getActionDefinitionsBuilder(Op);
    // Whitelist the common cases.
    // TODO: Loads to s16 on gfx9
    Actions.legalForTypesWithMemDesc({{S32, GlobalPtr, 32, GlobalAlign32},
                                      {V2S32, GlobalPtr, 64, GlobalAlign32},
                                      {V4S32, GlobalPtr, 128, GlobalAlign32},
                                      {S128, GlobalPtr, 128, GlobalAlign32},
                                      {S64, GlobalPtr, 64, GlobalAlign32},
                                      {V2S64, GlobalPtr, 128, GlobalAlign32},
                                      {V2S16, GlobalPtr, 32, GlobalAlign32},
                                      {S32, GlobalPtr, 8, GlobalAlign8},
                                      {S32, GlobalPtr, 16, GlobalAlign16},

                                      {S32, LocalPtr, 32, 32},
                                      {S64, LocalPtr, 64, 32},
                                      {V2S32, LocalPtr, 64, 32},
                                      {S32, LocalPtr, 8, 8},
                                      {S32, LocalPtr, 16, 16},
                                      {V2S16, LocalPtr, 32, 32},

                                      {S32, PrivatePtr, 32, 32},
                                      {S32, PrivatePtr, 8, 8},
                                      {S32, PrivatePtr, 16, 16},
                                      {V2S16, PrivatePtr, 32, 32},

                                      {S32, FlatPtr, 32, GlobalAlign32},
                                      {S32, FlatPtr, 16, GlobalAlign16},
                                      {S32, FlatPtr, 8, GlobalAlign8},
                                      {V2S16, FlatPtr, 32, GlobalAlign32},

                                      {S32, ConstantPtr, 32, GlobalAlign32},
                                      {V2S32, ConstantPtr, 64, GlobalAlign32},
                                      {V4S32, ConstantPtr, 128, GlobalAlign32},
                                      {S64, ConstantPtr, 64, GlobalAlign32},
                                      {S128, ConstantPtr, 128, GlobalAlign32},
                                      {V2S32, ConstantPtr, 32, GlobalAlign32}});
    Actions
        .customIf(typeIs(1, Constant32Ptr))
        // Widen suitably aligned loads by loading extra elements.
        .moreElementsIf([=](const LegalityQuery &Query) {
            const LLT Ty = Query.Types[0];
            return Op == G_LOAD && Ty.isVector() &&
                   shouldWidenLoadResult(Query);
          }, moreElementsToNextPow2(0))
        .widenScalarIf([=](const LegalityQuery &Query) {
            const LLT Ty = Query.Types[0];
            return Op == G_LOAD && !Ty.isVector() &&
                   shouldWidenLoadResult(Query);
          }, widenScalarOrEltToNextPow2(0))
        .narrowScalarIf(
            [=](const LegalityQuery &Query) -> bool {
              return !Query.Types[0].isVector() &&
                     needToSplitMemOp(Query, Op == G_LOAD);
            },
            [=](const LegalityQuery &Query) -> std::pair<unsigned, LLT> {
              const LLT DstTy = Query.Types[0];
              const LLT PtrTy = Query.Types[1];

              const unsigned DstSize = DstTy.getSizeInBits();
              unsigned MemSize = Query.MMODescrs[0].SizeInBits;

              // Split extloads.
              if (DstSize > MemSize)
                return std::make_pair(0, LLT::scalar(MemSize));

              if (!isPowerOf2_32(DstSize)) {
                // We're probably decomposing an odd sized store. Try to split
                // to the widest type. TODO: Account for alignment. As-is it
                // should be OK, since the new parts will be further legalized.
                unsigned FloorSize = PowerOf2Floor(DstSize);
                return std::make_pair(0, LLT::scalar(FloorSize));
              }

              if (DstSize > 32 && (DstSize % 32 != 0)) {
                // FIXME: Need a way to specify non-extload of larger size if
                // suitably aligned.
                return std::make_pair(0, LLT::scalar(32 * (DstSize / 32)));
              }

              unsigned MaxSize = maxSizeForAddrSpace(PtrTy.getAddressSpace(),
                                                     Op == G_LOAD);
              if (MemSize > MaxSize)
                return std::make_pair(0, LLT::scalar(MaxSize));

              unsigned Align = Query.MMODescrs[0].AlignInBits;
              return std::make_pair(0, LLT::scalar(Align));
            })
        .fewerElementsIf(
            [=](const LegalityQuery &Query) -> bool {
              return Query.Types[0].isVector() &&
                     needToSplitMemOp(Query, Op == G_LOAD);
            },
            [=](const LegalityQuery &Query) -> std::pair<unsigned, LLT> {
              const LLT DstTy = Query.Types[0];
              const LLT PtrTy = Query.Types[1];

              LLT EltTy = DstTy.getElementType();
              unsigned MaxSize = maxSizeForAddrSpace(PtrTy.getAddressSpace(),
                                                     Op == G_LOAD);

              // FIXME: Handle widened to power of 2 results better. This ends
              // up scalarizing.
              // FIXME: 3 element stores scalarized on SI

              // Split if it's too large for the address space.
              if (Query.MMODescrs[0].SizeInBits > MaxSize) {
                unsigned NumElts = DstTy.getNumElements();
                unsigned EltSize = EltTy.getSizeInBits();

                if (MaxSize % EltSize == 0) {
                  return std::make_pair(
                    0, LLT::scalarOrVector(MaxSize / EltSize, EltTy));
                }

                unsigned NumPieces = Query.MMODescrs[0].SizeInBits / MaxSize;

                // FIXME: Refine when odd breakdowns handled
                // The scalars will need to be re-legalized.
                if (NumPieces == 1 || NumPieces >= NumElts ||
                    NumElts % NumPieces != 0)
                  return std::make_pair(0, EltTy);

                return std::make_pair(0,
                                      LLT::vector(NumElts / NumPieces, EltTy));
              }

              // FIXME: We could probably handle weird extending loads better.
              unsigned MemSize = Query.MMODescrs[0].SizeInBits;
              if (DstTy.getSizeInBits() > MemSize)
                return std::make_pair(0, EltTy);

              unsigned EltSize = EltTy.getSizeInBits();
              unsigned DstSize = DstTy.getSizeInBits();
              if (!isPowerOf2_32(DstSize)) {
                // We're probably decomposing an odd sized store. Try to split
                // to the widest type. TODO: Account for alignment. As-is it
                // should be OK, since the new parts will be further legalized.
                unsigned FloorSize = PowerOf2Floor(DstSize);
                return std::make_pair(
                  0, LLT::scalarOrVector(FloorSize / EltSize, EltTy));
              }

              // Need to split because of alignment.
              unsigned Align = Query.MMODescrs[0].AlignInBits;
              if (EltSize > Align &&
                  (EltSize / Align < DstTy.getNumElements())) {
                return std::make_pair(0, LLT::vector(EltSize / Align, EltTy));
              }

              // May need relegalization for the scalars.
              return std::make_pair(0, EltTy);
            })
        .minScalar(0, S32);

    if (IsStore)
      Actions.narrowScalarIf(isWideScalarTruncStore(0), changeTo(0, S32));

    // TODO: Need a bitcast lower option?
    Actions
        .legalIf([=](const LegalityQuery &Query) {
          const LLT Ty0 = Query.Types[0];
          unsigned Size = Ty0.getSizeInBits();
          unsigned MemSize = Query.MMODescrs[0].SizeInBits;
          unsigned Align = Query.MMODescrs[0].AlignInBits;

          // FIXME: Widening store from alignment not valid.
          if (MemSize < Size)
            MemSize = std::max(MemSize, Align);

          // No extending vector loads.
          if (Size > MemSize && Ty0.isVector())
            return false;

          switch (MemSize) {
          case 8:
          case 16:
            return Size == 32;
          case 32:
          case 64:
          case 128:
            return true;
          case 96:
            return ST.hasDwordx3LoadStores();
          case 256:
          case 512:
            return true;
          default:
            return false;
          }
        })
        .widenScalarToNextPow2(0)
        .moreElementsIf(vectorSmallerThan(0, 32), moreEltsToNext32Bit(0));
  }

  auto &ExtLoads = getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
                       .legalForTypesWithMemDesc({{S32, GlobalPtr, 8, 8},
                                                  {S32, GlobalPtr, 16, 2 * 8},
                                                  {S32, LocalPtr, 8, 8},
                                                  {S32, LocalPtr, 16, 16},
                                                  {S32, PrivatePtr, 8, 8},
                                                  {S32, PrivatePtr, 16, 16},
                                                  {S32, ConstantPtr, 8, 8},
                                                  {S32, ConstantPtr, 16, 2 * 8}});
  if (ST.hasFlatAddressSpace()) {
    ExtLoads.legalForTypesWithMemDesc(
        {{S32, FlatPtr, 8, 8}, {S32, FlatPtr, 16, 16}});
  }

  ExtLoads.clampScalar(0, S32, S32)
          .widenScalarToNextPow2(0)
          .unsupportedIfMemSizeNotPow2()
          .lower();

  auto &Atomics = getActionDefinitionsBuilder(
    {G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD, G_ATOMICRMW_SUB,
     G_ATOMICRMW_AND, G_ATOMICRMW_OR, G_ATOMICRMW_XOR,
     G_ATOMICRMW_MAX, G_ATOMICRMW_MIN, G_ATOMICRMW_UMAX,
     G_ATOMICRMW_UMIN})
    .legalFor({{S32, GlobalPtr}, {S32, LocalPtr},
               {S64, GlobalPtr}, {S64, LocalPtr}});
  if (ST.hasFlatAddressSpace()) {
    Atomics.legalFor({{S32, FlatPtr}, {S64, FlatPtr}});
  }

  getActionDefinitionsBuilder(G_ATOMICRMW_FADD)
    .legalFor({{S32, LocalPtr}});

  // BUFFER/FLAT_ATOMIC_CMP_SWAP on GCN GPUs needs input marshalling, and output
  // demarshalling
  getActionDefinitionsBuilder(G_ATOMIC_CMPXCHG)
    .customFor({{S32, GlobalPtr}, {S64, GlobalPtr},
                {S32, FlatPtr}, {S64, FlatPtr}})
    .legalFor({{S32, LocalPtr}, {S64, LocalPtr},
               {S32, RegionPtr}, {S64, RegionPtr}});
  // TODO: Pointer types, any 32-bit or 64-bit vector

  // Condition should be s32 for scalar, s1 for vector.
  getActionDefinitionsBuilder(G_SELECT)
    .legalForCartesianProduct({S32, S64, S16, V2S32, V2S16, V4S16,
          GlobalPtr, LocalPtr, FlatPtr, PrivatePtr,
          LLT::vector(2, LocalPtr), LLT::vector(2, PrivatePtr)}, {S1, S32})
    .clampScalar(0, S16, S64)
    .scalarize(1)
    .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
    .fewerElementsIf(numElementsNotEven(0), scalarize(0))
    .clampMaxNumElements(0, S32, 2)
    .clampMaxNumElements(0, LocalPtr, 2)
    .clampMaxNumElements(0, PrivatePtr, 2)
    .scalarize(0)
    .widenScalarToNextPow2(0)
    .legalIf(all(isPointer(0), typeInSet(1, {S1, S32})));

  // TODO: Only the low 4/5/6 bits of the shift amount are observed, so we can
  // be more flexible with the shift amount type.
  auto &Shifts = getActionDefinitionsBuilder({G_SHL, G_LSHR, G_ASHR})
    .legalFor({{S32, S32}, {S64, S32}});
  if (ST.has16BitInsts()) {
    if (ST.hasVOP3PInsts()) {
      Shifts.legalFor({{S16, S16}, {V2S16, V2S16}})
            .clampMaxNumElements(0, S16, 2);
    } else
      Shifts.legalFor({{S16, S16}});

    // TODO: Support 16-bit shift amounts for all types
    Shifts.widenScalarIf(
      [=](const LegalityQuery &Query) {
        // Use 16-bit shift amounts for any 16-bit shift. Otherwise we want a
        // 32-bit amount.
        const LLT ValTy = Query.Types[0];
        const LLT AmountTy = Query.Types[1];
        return ValTy.getSizeInBits() <= 16 &&
               AmountTy.getSizeInBits() < 16;
      }, changeTo(1, S16));
    Shifts.maxScalarIf(typeIs(0, S16), 1, S16);
    Shifts.clampScalar(1, S32, S32);
    Shifts.clampScalar(0, S16, S64);
    Shifts.widenScalarToNextPow2(0, 16);
  } else {
    // Make sure we legalize the shift amount type first, as the general
    // expansion for the shifted type will produce much worse code if it hasn't
    // been truncated already.
    Shifts.clampScalar(1, S32, S32);
    Shifts.clampScalar(0, S32, S64);
    Shifts.widenScalarToNextPow2(0, 32);
  }
  Shifts.scalarize(0);

  for (unsigned Op : {G_EXTRACT_VECTOR_ELT, G_INSERT_VECTOR_ELT}) {
    unsigned VecTypeIdx = Op == G_EXTRACT_VECTOR_ELT ? 1 : 0;
    unsigned EltTypeIdx = Op == G_EXTRACT_VECTOR_ELT ? 0 : 1;
    unsigned IdxTypeIdx = 2;

    getActionDefinitionsBuilder(Op)
      .customIf([=](const LegalityQuery &Query) {
          const LLT EltTy = Query.Types[EltTypeIdx];
          const LLT VecTy = Query.Types[VecTypeIdx];
          const LLT IdxTy = Query.Types[IdxTypeIdx];
          return (EltTy.getSizeInBits() == 16 ||
                  EltTy.getSizeInBits() % 32 == 0) &&
                 VecTy.getSizeInBits() % 32 == 0 &&
                 VecTy.getSizeInBits() <= 1024 &&
                 IdxTy.getSizeInBits() == 32;
        })
      .clampScalar(EltTypeIdx, S32, S64)
      .clampScalar(VecTypeIdx, S32, S64)
      .clampScalar(IdxTypeIdx, S32, S32);
  }

  getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
    .unsupportedIf([=](const LegalityQuery &Query) {
        const LLT &EltTy = Query.Types[1].getElementType();
        return Query.Types[0] != EltTy;
      });

  for (unsigned Op : {G_EXTRACT, G_INSERT}) {
    unsigned BigTyIdx = Op == G_EXTRACT ? 1 : 0;
    unsigned LitTyIdx = Op == G_EXTRACT ? 0 : 1;

    // FIXME: Doesn't handle extract of illegal sizes.
    getActionDefinitionsBuilder(Op)
      .lowerIf(all(typeIs(LitTyIdx, S16), sizeIs(BigTyIdx, 32)))
      // FIXME: Multiples of 16 should not be legal.
      .legalIf([=](const LegalityQuery &Query) {
          const LLT BigTy = Query.Types[BigTyIdx];
          const LLT LitTy = Query.Types[LitTyIdx];
          return (BigTy.getSizeInBits() % 32 == 0) &&
                 (LitTy.getSizeInBits() % 16 == 0);
        })
      .widenScalarIf(
        [=](const LegalityQuery &Query) {
          const LLT BigTy = Query.Types[BigTyIdx];
          return (BigTy.getScalarSizeInBits() < 16);
        },
        LegalizeMutations::widenScalarOrEltToNextPow2(BigTyIdx, 16))
      .widenScalarIf(
        [=](const LegalityQuery &Query) {
          const LLT LitTy = Query.Types[LitTyIdx];
          return (LitTy.getScalarSizeInBits() < 16);
        },
        LegalizeMutations::widenScalarOrEltToNextPow2(LitTyIdx, 16))
      .moreElementsIf(isSmallOddVector(BigTyIdx), oneMoreElement(BigTyIdx))
      .widenScalarToNextPow2(BigTyIdx, 32);

  }

  auto &BuildVector = getActionDefinitionsBuilder(G_BUILD_VECTOR)
    .legalForCartesianProduct(AllS32Vectors, {S32})
    .legalForCartesianProduct(AllS64Vectors, {S64})
    .clampNumElements(0, V16S32, V32S32)
    .clampNumElements(0, V2S64, V16S64)
    .fewerElementsIf(isWideVec16(0), changeTo(0, V2S16));

  if (ST.hasScalarPackInsts()) {
    BuildVector
      // FIXME: Should probably widen s1 vectors straight to s32
      .minScalarOrElt(0, S16)
      // Widen source elements and produce a G_BUILD_VECTOR_TRUNC
      .minScalar(1, S32);

    getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC)
      .legalFor({V2S16, S32})
      .lower();
    BuildVector.minScalarOrElt(0, S32);
  } else {
    BuildVector.customFor({V2S16, S16});
    BuildVector.minScalarOrElt(0, S32);

    getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC)
      .customFor({V2S16, S32})
      .lower();
  }

  BuildVector.legalIf(isRegisterType(0));

  // FIXME: Clamp maximum size
  getActionDefinitionsBuilder(G_CONCAT_VECTORS)
    .legalIf(isRegisterType(0));

  // TODO: Don't fully scalarize v2s16 pieces? Or combine out thosse
  // pre-legalize.
  if (ST.hasVOP3PInsts()) {
    getActionDefinitionsBuilder(G_SHUFFLE_VECTOR)
      .customFor({V2S16, V2S16})
      .lower();
  } else
    getActionDefinitionsBuilder(G_SHUFFLE_VECTOR).lower();

  // Merge/Unmerge
  for (unsigned Op : {G_MERGE_VALUES, G_UNMERGE_VALUES}) {
    unsigned BigTyIdx = Op == G_MERGE_VALUES ? 0 : 1;
    unsigned LitTyIdx = Op == G_MERGE_VALUES ? 1 : 0;

    auto notValidElt = [=](const LegalityQuery &Query, unsigned TypeIdx) {
      const LLT Ty = Query.Types[TypeIdx];
      if (Ty.isVector()) {
        const LLT &EltTy = Ty.getElementType();
        if (EltTy.getSizeInBits() < 8 || EltTy.getSizeInBits() > 512)
          return true;
        if (!isPowerOf2_32(EltTy.getSizeInBits()))
          return true;
      }
      return false;
    };

    auto &Builder = getActionDefinitionsBuilder(Op)
      .lowerFor({{S16, V2S16}})
      .lowerIf([=](const LegalityQuery &Query) {
          const LLT BigTy = Query.Types[BigTyIdx];
          return BigTy.getSizeInBits() == 32;
        })
      // Try to widen to s16 first for small types.
      // TODO: Only do this on targets with legal s16 shifts
      .minScalarOrEltIf(narrowerThan(LitTyIdx, 16), LitTyIdx, S16)
      .widenScalarToNextPow2(LitTyIdx, /*Min*/ 16)
      .moreElementsIf(isSmallOddVector(BigTyIdx), oneMoreElement(BigTyIdx))
      .fewerElementsIf(all(typeIs(0, S16), vectorWiderThan(1, 32),
                           elementTypeIs(1, S16)),
                       changeTo(1, V2S16))
      // Clamp the little scalar to s8-s256 and make it a power of 2. It's not
      // worth considering the multiples of 64 since 2*192 and 2*384 are not
      // valid.
      .clampScalar(LitTyIdx, S32, S512)
      .widenScalarToNextPow2(LitTyIdx, /*Min*/ 32)
      // Break up vectors with weird elements into scalars
      .fewerElementsIf(
        [=](const LegalityQuery &Query) { return notValidElt(Query, LitTyIdx); },
        scalarize(0))
      .fewerElementsIf(
        [=](const LegalityQuery &Query) { return notValidElt(Query, BigTyIdx); },
        scalarize(1))
      .clampScalar(BigTyIdx, S32, S1024);

    if (Op == G_MERGE_VALUES) {
      Builder.widenScalarIf(
        // TODO: Use 16-bit shifts if legal for 8-bit values?
        [=](const LegalityQuery &Query) {
          const LLT Ty = Query.Types[LitTyIdx];
          return Ty.getSizeInBits() < 32;
        },
        changeTo(LitTyIdx, S32));
    }

    Builder.widenScalarIf(
      [=](const LegalityQuery &Query) {
        const LLT Ty = Query.Types[BigTyIdx];
        return !isPowerOf2_32(Ty.getSizeInBits()) &&
          Ty.getSizeInBits() % 16 != 0;
      },
      [=](const LegalityQuery &Query) {
        // Pick the next power of 2, or a multiple of 64 over 128.
        // Whichever is smaller.
        const LLT &Ty = Query.Types[BigTyIdx];
        unsigned NewSizeInBits = 1 << Log2_32_Ceil(Ty.getSizeInBits() + 1);
        if (NewSizeInBits >= 256) {
          unsigned RoundedTo = alignTo<64>(Ty.getSizeInBits() + 1);
          if (RoundedTo < NewSizeInBits)
            NewSizeInBits = RoundedTo;
        }
        return std::make_pair(BigTyIdx, LLT::scalar(NewSizeInBits));
      })
      .legalIf([=](const LegalityQuery &Query) {
          const LLT &BigTy = Query.Types[BigTyIdx];
          const LLT &LitTy = Query.Types[LitTyIdx];

          if (BigTy.isVector() && BigTy.getSizeInBits() < 32)
            return false;
          if (LitTy.isVector() && LitTy.getSizeInBits() < 32)
            return false;

          return BigTy.getSizeInBits() % 16 == 0 &&
                 LitTy.getSizeInBits() % 16 == 0 &&
                 BigTy.getSizeInBits() <= 1024;
        })
      // Any vectors left are the wrong size. Scalarize them.
      .scalarize(0)
      .scalarize(1);
  }

  // S64 is only legal on SALU, and needs to be broken into 32-bit elements in
  // RegBankSelect.
  auto &SextInReg = getActionDefinitionsBuilder(G_SEXT_INREG)
    .legalFor({{S32}, {S64}});

  if (ST.hasVOP3PInsts()) {
    SextInReg.lowerFor({{V2S16}})
      // Prefer to reduce vector widths for 16-bit vectors before lowering, to
      // get more vector shift opportunities, since we'll get those when
      // expanded.
      .fewerElementsIf(elementTypeIs(0, S16), changeTo(0, V2S16));
  } else if (ST.has16BitInsts()) {
    SextInReg.lowerFor({{S32}, {S64}, {S16}});
  } else {
    // Prefer to promote to s32 before lowering if we don't have 16-bit
    // shifts. This avoid a lot of intermediate truncate and extend operations.
    SextInReg.lowerFor({{S32}, {S64}});
  }

  SextInReg
    .scalarize(0)
    .clampScalar(0, S32, S64)
    .lower();

  getActionDefinitionsBuilder(G_FSHR)
    .legalFor({{S32, S32}})
    .scalarize(0)
    .lower();

  getActionDefinitionsBuilder(G_READCYCLECOUNTER)
    .legalFor({S64});

  getActionDefinitionsBuilder({
      // TODO: Verify V_BFI_B32 is generated from expanded bit ops
      G_FCOPYSIGN,

      G_ATOMIC_CMPXCHG_WITH_SUCCESS,
      G_READ_REGISTER,
      G_WRITE_REGISTER,

      G_SADDO, G_SSUBO,

       // TODO: Implement
      G_FMINIMUM, G_FMAXIMUM,
      G_FSHL
    }).lower();

  getActionDefinitionsBuilder({G_VASTART, G_VAARG, G_BRJT, G_JUMP_TABLE,
        G_DYN_STACKALLOC, G_INDEXED_LOAD, G_INDEXED_SEXTLOAD,
        G_INDEXED_ZEXTLOAD, G_INDEXED_STORE})
    .unsupported();

  computeTables();
  verify(*ST.getInstrInfo());
}

bool AMDGPULegalizerInfo::legalizeCustom(MachineInstr &MI,
                                         MachineRegisterInfo &MRI,
                                         MachineIRBuilder &B,
                                         GISelChangeObserver &Observer) const {
  switch (MI.getOpcode()) {
  case TargetOpcode::G_ADDRSPACE_CAST:
    return legalizeAddrSpaceCast(MI, MRI, B);
  case TargetOpcode::G_FRINT:
    return legalizeFrint(MI, MRI, B);
  case TargetOpcode::G_FCEIL:
    return legalizeFceil(MI, MRI, B);
  case TargetOpcode::G_INTRINSIC_TRUNC:
    return legalizeIntrinsicTrunc(MI, MRI, B);
  case TargetOpcode::G_SITOFP:
    return legalizeITOFP(MI, MRI, B, true);
  case TargetOpcode::G_UITOFP:
    return legalizeITOFP(MI, MRI, B, false);
  case TargetOpcode::G_FPTOSI:
    return legalizeFPTOI(MI, MRI, B, true);
  case TargetOpcode::G_FPTOUI:
    return legalizeFPTOI(MI, MRI, B, false);
  case TargetOpcode::G_FMINNUM:
  case TargetOpcode::G_FMAXNUM:
  case TargetOpcode::G_FMINNUM_IEEE:
  case TargetOpcode::G_FMAXNUM_IEEE:
    return legalizeMinNumMaxNum(MI, MRI, B);
  case TargetOpcode::G_EXTRACT_VECTOR_ELT:
    return legalizeExtractVectorElt(MI, MRI, B);
  case TargetOpcode::G_INSERT_VECTOR_ELT:
    return legalizeInsertVectorElt(MI, MRI, B);
  case TargetOpcode::G_SHUFFLE_VECTOR:
    return legalizeShuffleVector(MI, MRI, B);
  case TargetOpcode::G_FSIN:
  case TargetOpcode::G_FCOS:
    return legalizeSinCos(MI, MRI, B);
  case TargetOpcode::G_GLOBAL_VALUE:
    return legalizeGlobalValue(MI, MRI, B);
  case TargetOpcode::G_LOAD:
    return legalizeLoad(MI, MRI, B, Observer);
  case TargetOpcode::G_FMAD:
    return legalizeFMad(MI, MRI, B);
  case TargetOpcode::G_FDIV:
    return legalizeFDIV(MI, MRI, B);
  case TargetOpcode::G_UDIV:
  case TargetOpcode::G_UREM:
    return legalizeUDIV_UREM(MI, MRI, B);
  case TargetOpcode::G_SDIV:
  case TargetOpcode::G_SREM:
    return legalizeSDIV_SREM(MI, MRI, B);
  case TargetOpcode::G_ATOMIC_CMPXCHG:
    return legalizeAtomicCmpXChg(MI, MRI, B);
  case TargetOpcode::G_FLOG:
    return legalizeFlog(MI, B, 1.0f / numbers::log2ef);
  case TargetOpcode::G_FLOG10:
    return legalizeFlog(MI, B, numbers::ln2f / numbers::ln10f);
  case TargetOpcode::G_FEXP:
    return legalizeFExp(MI, B);
  case TargetOpcode::G_FPOW:
    return legalizeFPow(MI, B);
  case TargetOpcode::G_FFLOOR:
    return legalizeFFloor(MI, MRI, B);
  case TargetOpcode::G_BUILD_VECTOR:
    return legalizeBuildVector(MI, MRI, B);
  default:
    return false;
  }

  llvm_unreachable("expected switch to return");
}

Register AMDGPULegalizerInfo::getSegmentAperture(
  unsigned AS,
  MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  MachineFunction &MF = B.getMF();
  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
  const LLT S32 = LLT::scalar(32);

  assert(AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::PRIVATE_ADDRESS);

  if (ST.hasApertureRegs()) {
    // FIXME: Use inline constants (src_{shared, private}_base) instead of
    // getreg.
    unsigned Offset = AS == AMDGPUAS::LOCAL_ADDRESS ?
        AMDGPU::Hwreg::OFFSET_SRC_SHARED_BASE :
        AMDGPU::Hwreg::OFFSET_SRC_PRIVATE_BASE;
    unsigned WidthM1 = AS == AMDGPUAS::LOCAL_ADDRESS ?
        AMDGPU::Hwreg::WIDTH_M1_SRC_SHARED_BASE :
        AMDGPU::Hwreg::WIDTH_M1_SRC_PRIVATE_BASE;
    unsigned Encoding =
        AMDGPU::Hwreg::ID_MEM_BASES << AMDGPU::Hwreg::ID_SHIFT_ |
        Offset << AMDGPU::Hwreg::OFFSET_SHIFT_ |
        WidthM1 << AMDGPU::Hwreg::WIDTH_M1_SHIFT_;

    Register GetReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);

    B.buildInstr(AMDGPU::S_GETREG_B32)
      .addDef(GetReg)
      .addImm(Encoding);
    MRI.setType(GetReg, S32);

    auto ShiftAmt = B.buildConstant(S32, WidthM1 + 1);
    return B.buildShl(S32, GetReg, ShiftAmt).getReg(0);
  }

  Register QueuePtr = MRI.createGenericVirtualRegister(
    LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64));

  const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
  if (!loadInputValue(QueuePtr, B, &MFI->getArgInfo().QueuePtr))
    return Register();

  // Offset into amd_queue_t for group_segment_aperture_base_hi /
  // private_segment_aperture_base_hi.
  uint32_t StructOffset = (AS == AMDGPUAS::LOCAL_ADDRESS) ? 0x40 : 0x44;

  // TODO: can we be smarter about machine pointer info?
  MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS);
  MachineMemOperand *MMO = MF.getMachineMemOperand(
      PtrInfo,
      MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
          MachineMemOperand::MOInvariant,
      4, commonAlignment(Align(64), StructOffset));

  Register LoadAddr;

  B.materializePtrAdd(LoadAddr, QueuePtr, LLT::scalar(64), StructOffset);
  return B.buildLoad(S32, LoadAddr, *MMO).getReg(0);
}

bool AMDGPULegalizerInfo::legalizeAddrSpaceCast(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  MachineFunction &MF = B.getMF();

  B.setInstr(MI);

  const LLT S32 = LLT::scalar(32);
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();

  LLT DstTy = MRI.getType(Dst);
  LLT SrcTy = MRI.getType(Src);
  unsigned DestAS = DstTy.getAddressSpace();
  unsigned SrcAS = SrcTy.getAddressSpace();

  // TODO: Avoid reloading from the queue ptr for each cast, or at least each
  // vector element.
  assert(!DstTy.isVector());

  const AMDGPUTargetMachine &TM
    = static_cast<const AMDGPUTargetMachine &>(MF.getTarget());

  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
  if (ST.getTargetLowering()->isNoopAddrSpaceCast(SrcAS, DestAS)) {
    MI.setDesc(B.getTII().get(TargetOpcode::G_BITCAST));
    return true;
  }

  if (DestAS == AMDGPUAS::CONSTANT_ADDRESS_32BIT) {
    // Truncate.
    B.buildExtract(Dst, Src, 0);
    MI.eraseFromParent();
    return true;
  }

  if (SrcAS == AMDGPUAS::CONSTANT_ADDRESS_32BIT) {
    const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
    uint32_t AddrHiVal = Info->get32BitAddressHighBits();

    // FIXME: This is a bit ugly due to creating a merge of 2 pointers to
    // another. Merge operands are required to be the same type, but creating an
    // extra ptrtoint would be kind of pointless.
    auto HighAddr = B.buildConstant(
      LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS_32BIT, 32), AddrHiVal);
    B.buildMerge(Dst, {Src, HighAddr});
    MI.eraseFromParent();
    return true;
  }

  if (SrcAS == AMDGPUAS::FLAT_ADDRESS) {
    assert(DestAS == AMDGPUAS::LOCAL_ADDRESS ||
           DestAS == AMDGPUAS::PRIVATE_ADDRESS);
    unsigned NullVal = TM.getNullPointerValue(DestAS);

    auto SegmentNull = B.buildConstant(DstTy, NullVal);
    auto FlatNull = B.buildConstant(SrcTy, 0);

    // Extract low 32-bits of the pointer.
    auto PtrLo32 = B.buildExtract(DstTy, Src, 0);

    auto CmpRes =
        B.buildICmp(CmpInst::ICMP_NE, LLT::scalar(1), Src, FlatNull.getReg(0));
    B.buildSelect(Dst, CmpRes, PtrLo32, SegmentNull.getReg(0));

    MI.eraseFromParent();
    return true;
  }

  if (SrcAS != AMDGPUAS::LOCAL_ADDRESS && SrcAS != AMDGPUAS::PRIVATE_ADDRESS)
    return false;

  if (!ST.hasFlatAddressSpace())
    return false;

  auto SegmentNull =
      B.buildConstant(SrcTy, TM.getNullPointerValue(SrcAS));
  auto FlatNull =
      B.buildConstant(DstTy, TM.getNullPointerValue(DestAS));

  Register ApertureReg = getSegmentAperture(SrcAS, MRI, B);
  if (!ApertureReg.isValid())
    return false;

  auto CmpRes =
      B.buildICmp(CmpInst::ICMP_NE, LLT::scalar(1), Src, SegmentNull.getReg(0));

  // Coerce the type of the low half of the result so we can use merge_values.
  Register SrcAsInt = B.buildPtrToInt(S32, Src).getReg(0);

  // TODO: Should we allow mismatched types but matching sizes in merges to
  // avoid the ptrtoint?
  auto BuildPtr = B.buildMerge(DstTy, {SrcAsInt, ApertureReg});
  B.buildSelect(Dst, CmpRes, BuildPtr, FlatNull);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeFrint(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  B.setInstr(MI);

  Register Src = MI.getOperand(1).getReg();
  LLT Ty = MRI.getType(Src);
  assert(Ty.isScalar() && Ty.getSizeInBits() == 64);

  APFloat C1Val(APFloat::IEEEdouble(), "0x1.0p+52");
  APFloat C2Val(APFloat::IEEEdouble(), "0x1.fffffffffffffp+51");

  auto C1 = B.buildFConstant(Ty, C1Val);
  auto CopySign = B.buildFCopysign(Ty, C1, Src);

  // TODO: Should this propagate fast-math-flags?
  auto Tmp1 = B.buildFAdd(Ty, Src, CopySign);
  auto Tmp2 = B.buildFSub(Ty, Tmp1, CopySign);

  auto C2 = B.buildFConstant(Ty, C2Val);
  auto Fabs = B.buildFAbs(Ty, Src);

  auto Cond = B.buildFCmp(CmpInst::FCMP_OGT, LLT::scalar(1), Fabs, C2);
  B.buildSelect(MI.getOperand(0).getReg(), Cond, Src, Tmp2);
  return true;
}

bool AMDGPULegalizerInfo::legalizeFceil(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  B.setInstr(MI);

  const LLT S1 = LLT::scalar(1);
  const LLT S64 = LLT::scalar(64);

  Register Src = MI.getOperand(1).getReg();
  assert(MRI.getType(Src) == S64);

  // result = trunc(src)
  // if (src > 0.0 && src != result)
  //   result += 1.0

  auto Trunc = B.buildIntrinsicTrunc(S64, Src);

  const auto Zero = B.buildFConstant(S64, 0.0);
  const auto One = B.buildFConstant(S64, 1.0);
  auto Lt0 = B.buildFCmp(CmpInst::FCMP_OGT, S1, Src, Zero);
  auto NeTrunc = B.buildFCmp(CmpInst::FCMP_ONE, S1, Src, Trunc);
  auto And = B.buildAnd(S1, Lt0, NeTrunc);
  auto Add = B.buildSelect(S64, And, One, Zero);

  // TODO: Should this propagate fast-math-flags?
  B.buildFAdd(MI.getOperand(0).getReg(), Trunc, Add);
  return true;
}

static MachineInstrBuilder extractF64Exponent(unsigned Hi,
                                              MachineIRBuilder &B) {
  const unsigned FractBits = 52;
  const unsigned ExpBits = 11;
  LLT S32 = LLT::scalar(32);

  auto Const0 = B.buildConstant(S32, FractBits - 32);
  auto Const1 = B.buildConstant(S32, ExpBits);

  auto ExpPart = B.buildIntrinsic(Intrinsic::amdgcn_ubfe, {S32}, false)
    .addUse(Const0.getReg(0))
    .addUse(Const1.getReg(0));

  return B.buildSub(S32, ExpPart, B.buildConstant(S32, 1023));
}

bool AMDGPULegalizerInfo::legalizeIntrinsicTrunc(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  B.setInstr(MI);

  const LLT S1 = LLT::scalar(1);
  const LLT S32 = LLT::scalar(32);
  const LLT S64 = LLT::scalar(64);

  Register Src = MI.getOperand(1).getReg();
  assert(MRI.getType(Src) == S64);

  // TODO: Should this use extract since the low half is unused?
  auto Unmerge = B.buildUnmerge({S32, S32}, Src);
  Register Hi = Unmerge.getReg(1);

  // Extract the upper half, since this is where we will find the sign and
  // exponent.
  auto Exp = extractF64Exponent(Hi, B);

  const unsigned FractBits = 52;

  // Extract the sign bit.
  const auto SignBitMask = B.buildConstant(S32, UINT32_C(1) << 31);
  auto SignBit = B.buildAnd(S32, Hi, SignBitMask);

  const auto FractMask = B.buildConstant(S64, (UINT64_C(1) << FractBits) - 1);

  const auto Zero32 = B.buildConstant(S32, 0);

  // Extend back to 64-bits.
  auto SignBit64 = B.buildMerge(S64, {Zero32, SignBit});

  auto Shr = B.buildAShr(S64, FractMask, Exp);
  auto Not = B.buildNot(S64, Shr);
  auto Tmp0 = B.buildAnd(S64, Src, Not);
  auto FiftyOne = B.buildConstant(S32, FractBits - 1);

  auto ExpLt0 = B.buildICmp(CmpInst::ICMP_SLT, S1, Exp, Zero32);
  auto ExpGt51 = B.buildICmp(CmpInst::ICMP_SGT, S1, Exp, FiftyOne);

  auto Tmp1 = B.buildSelect(S64, ExpLt0, SignBit64, Tmp0);
  B.buildSelect(MI.getOperand(0).getReg(), ExpGt51, Src, Tmp1);
  return true;
}

bool AMDGPULegalizerInfo::legalizeITOFP(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B, bool Signed) const {
  B.setInstr(MI);

  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();

  const LLT S64 = LLT::scalar(64);
  const LLT S32 = LLT::scalar(32);

  assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S64);

  auto Unmerge = B.buildUnmerge({S32, S32}, Src);

  auto CvtHi = Signed ?
    B.buildSITOFP(S64, Unmerge.getReg(1)) :
    B.buildUITOFP(S64, Unmerge.getReg(1));

  auto CvtLo = B.buildUITOFP(S64, Unmerge.getReg(0));

  auto ThirtyTwo = B.buildConstant(S32, 32);
  auto LdExp = B.buildIntrinsic(Intrinsic::amdgcn_ldexp, {S64}, false)
    .addUse(CvtHi.getReg(0))
    .addUse(ThirtyTwo.getReg(0));

  // TODO: Should this propagate fast-math-flags?
  B.buildFAdd(Dst, LdExp, CvtLo);
  MI.eraseFromParent();
  return true;
}

// TODO: Copied from DAG implementation. Verify logic and document how this
// actually works.
bool AMDGPULegalizerInfo::legalizeFPTOI(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B, bool Signed) const {
  B.setInstr(MI);

  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();

  const LLT S64 = LLT::scalar(64);
  const LLT S32 = LLT::scalar(32);

  assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S64);

  unsigned Flags = MI.getFlags();

  auto Trunc = B.buildIntrinsicTrunc(S64, Src, Flags);
  auto K0 = B.buildFConstant(S64, BitsToDouble(UINT64_C(0x3df0000000000000)));
  auto K1 = B.buildFConstant(S64, BitsToDouble(UINT64_C(0xc1f0000000000000)));

  auto Mul = B.buildFMul(S64, Trunc, K0, Flags);
  auto FloorMul = B.buildFFloor(S64, Mul, Flags);
  auto Fma = B.buildFMA(S64, FloorMul, K1, Trunc, Flags);

  auto Hi = Signed ?
    B.buildFPTOSI(S32, FloorMul) :
    B.buildFPTOUI(S32, FloorMul);
  auto Lo = B.buildFPTOUI(S32, Fma);

  B.buildMerge(Dst, { Lo, Hi });
  MI.eraseFromParent();

  return true;
}

bool AMDGPULegalizerInfo::legalizeMinNumMaxNum(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  MachineFunction &MF = B.getMF();
  const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

  const bool IsIEEEOp = MI.getOpcode() == AMDGPU::G_FMINNUM_IEEE ||
                        MI.getOpcode() == AMDGPU::G_FMAXNUM_IEEE;

  // With ieee_mode disabled, the instructions have the correct behavior
  // already for G_FMINNUM/G_FMAXNUM
  if (!MFI->getMode().IEEE)
    return !IsIEEEOp;

  if (IsIEEEOp)
    return true;

  MachineIRBuilder HelperBuilder(MI);
  GISelObserverWrapper DummyObserver;
  LegalizerHelper Helper(MF, DummyObserver, HelperBuilder);
  HelperBuilder.setInstr(MI);
  return Helper.lowerFMinNumMaxNum(MI) == LegalizerHelper::Legalized;
}

bool AMDGPULegalizerInfo::legalizeExtractVectorElt(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  // TODO: Should move some of this into LegalizerHelper.

  // TODO: Promote dynamic indexing of s16 to s32

  // FIXME: Artifact combiner probably should have replaced the truncated
  // constant before this, so we shouldn't need
  // getConstantVRegValWithLookThrough.
  Optional<ValueAndVReg> IdxVal = getConstantVRegValWithLookThrough(
    MI.getOperand(2).getReg(), MRI);
  if (!IdxVal) // Dynamic case will be selected to register indexing.
    return true;

  Register Dst = MI.getOperand(0).getReg();
  Register Vec = MI.getOperand(1).getReg();

  LLT VecTy = MRI.getType(Vec);
  LLT EltTy = VecTy.getElementType();
  assert(EltTy == MRI.getType(Dst));

  B.setInstr(MI);

  if (IdxVal->Value < VecTy.getNumElements())
    B.buildExtract(Dst, Vec, IdxVal->Value * EltTy.getSizeInBits());
  else
    B.buildUndef(Dst);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeInsertVectorElt(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  // TODO: Should move some of this into LegalizerHelper.

  // TODO: Promote dynamic indexing of s16 to s32

  // FIXME: Artifact combiner probably should have replaced the truncated
  // constant before this, so we shouldn't need
  // getConstantVRegValWithLookThrough.
  Optional<ValueAndVReg> IdxVal = getConstantVRegValWithLookThrough(
    MI.getOperand(3).getReg(), MRI);
  if (!IdxVal) // Dynamic case will be selected to register indexing.
    return true;

  Register Dst = MI.getOperand(0).getReg();
  Register Vec = MI.getOperand(1).getReg();
  Register Ins = MI.getOperand(2).getReg();

  LLT VecTy = MRI.getType(Vec);
  LLT EltTy = VecTy.getElementType();
  assert(EltTy == MRI.getType(Ins));

  B.setInstr(MI);

  if (IdxVal->Value < VecTy.getNumElements())
    B.buildInsert(Dst, Vec, Ins, IdxVal->Value * EltTy.getSizeInBits());
  else
    B.buildUndef(Dst);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeShuffleVector(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  const LLT V2S16 = LLT::vector(2, 16);

  Register Dst = MI.getOperand(0).getReg();
  Register Src0 = MI.getOperand(1).getReg();
  LLT DstTy = MRI.getType(Dst);
  LLT SrcTy = MRI.getType(Src0);

  if (SrcTy == V2S16 && DstTy == V2S16 &&
      AMDGPU::isLegalVOP3PShuffleMask(MI.getOperand(3).getShuffleMask()))
    return true;

  MachineIRBuilder HelperBuilder(MI);
  GISelObserverWrapper DummyObserver;
  LegalizerHelper Helper(B.getMF(), DummyObserver, HelperBuilder);
  HelperBuilder.setInstr(MI);
  return Helper.lowerShuffleVector(MI) == LegalizerHelper::Legalized;
}

bool AMDGPULegalizerInfo::legalizeSinCos(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  B.setInstr(MI);

  Register DstReg = MI.getOperand(0).getReg();
  Register SrcReg = MI.getOperand(1).getReg();
  LLT Ty = MRI.getType(DstReg);
  unsigned Flags = MI.getFlags();

  Register TrigVal;
  auto OneOver2Pi = B.buildFConstant(Ty, 0.5 / M_PI);
  if (ST.hasTrigReducedRange()) {
    auto MulVal = B.buildFMul(Ty, SrcReg, OneOver2Pi, Flags);
    TrigVal = B.buildIntrinsic(Intrinsic::amdgcn_fract, {Ty}, false)
      .addUse(MulVal.getReg(0))
      .setMIFlags(Flags).getReg(0);
  } else
    TrigVal = B.buildFMul(Ty, SrcReg, OneOver2Pi, Flags).getReg(0);

  Intrinsic::ID TrigIntrin = MI.getOpcode() == AMDGPU::G_FSIN ?
    Intrinsic::amdgcn_sin : Intrinsic::amdgcn_cos;
  B.buildIntrinsic(TrigIntrin, makeArrayRef<Register>(DstReg), false)
    .addUse(TrigVal)
    .setMIFlags(Flags);
  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::buildPCRelGlobalAddress(
  Register DstReg, LLT PtrTy,
  MachineIRBuilder &B, const GlobalValue *GV,
  unsigned Offset, unsigned GAFlags) const {
  // In order to support pc-relative addressing, SI_PC_ADD_REL_OFFSET is lowered
  // to the following code sequence:
  //
  // For constant address space:
  //   s_getpc_b64 s[0:1]
  //   s_add_u32 s0, s0, $symbol
  //   s_addc_u32 s1, s1, 0
  //
  //   s_getpc_b64 returns the address of the s_add_u32 instruction and then
  //   a fixup or relocation is emitted to replace $symbol with a literal
  //   constant, which is a pc-relative offset from the encoding of the $symbol
  //   operand to the global variable.
  //
  // For global address space:
  //   s_getpc_b64 s[0:1]
  //   s_add_u32 s0, s0, $symbol@{gotpc}rel32@lo
  //   s_addc_u32 s1, s1, $symbol@{gotpc}rel32@hi
  //
  //   s_getpc_b64 returns the address of the s_add_u32 instruction and then
  //   fixups or relocations are emitted to replace $symbol@*@lo and
  //   $symbol@*@hi with lower 32 bits and higher 32 bits of a literal constant,
  //   which is a 64-bit pc-relative offset from the encoding of the $symbol
  //   operand to the global variable.
  //
  // What we want here is an offset from the value returned by s_getpc
  // (which is the address of the s_add_u32 instruction) to the global
  // variable, but since the encoding of $symbol starts 4 bytes after the start
  // of the s_add_u32 instruction, we end up with an offset that is 4 bytes too
  // small. This requires us to add 4 to the global variable offset in order to
  // compute the correct address.

  LLT ConstPtrTy = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);

  Register PCReg = PtrTy.getSizeInBits() != 32 ? DstReg :
    B.getMRI()->createGenericVirtualRegister(ConstPtrTy);

  MachineInstrBuilder MIB = B.buildInstr(AMDGPU::SI_PC_ADD_REL_OFFSET)
    .addDef(PCReg);

  MIB.addGlobalAddress(GV, Offset + 4, GAFlags);
  if (GAFlags == SIInstrInfo::MO_NONE)
    MIB.addImm(0);
  else
    MIB.addGlobalAddress(GV, Offset + 4, GAFlags + 1);

  B.getMRI()->setRegClass(PCReg, &AMDGPU::SReg_64RegClass);

  if (PtrTy.getSizeInBits() == 32)
    B.buildExtract(DstReg, PCReg, 0);
  return true;
 }

bool AMDGPULegalizerInfo::legalizeGlobalValue(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  Register DstReg = MI.getOperand(0).getReg();
  LLT Ty = MRI.getType(DstReg);
  unsigned AS = Ty.getAddressSpace();

  const GlobalValue *GV = MI.getOperand(1).getGlobal();
  MachineFunction &MF = B.getMF();
  SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
  B.setInstr(MI);

  if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) {
    if (!MFI->isEntryFunction()) {
      const Function &Fn = MF.getFunction();
      DiagnosticInfoUnsupported BadLDSDecl(
        Fn, "local memory global used by non-kernel function", MI.getDebugLoc(),
        DS_Warning);
      Fn.getContext().diagnose(BadLDSDecl);

      // We currently don't have a way to correctly allocate LDS objects that
      // aren't directly associated with a kernel. We do force inlining of
      // functions that use local objects. However, if these dead functions are
      // not eliminated, we don't want a compile time error. Just emit a warning
      // and a trap, since there should be no callable path here.
      B.buildIntrinsic(Intrinsic::trap, ArrayRef<Register>(), true);
      B.buildUndef(DstReg);
      MI.eraseFromParent();
      return true;
    }

    // TODO: We could emit code to handle the initialization somewhere.
    if (!AMDGPUTargetLowering::hasDefinedInitializer(GV)) {
      const SITargetLowering *TLI = ST.getTargetLowering();
      if (!TLI->shouldUseLDSConstAddress(GV)) {
        MI.getOperand(1).setTargetFlags(SIInstrInfo::MO_ABS32_LO);
        return true; // Leave in place;
      }

      B.buildConstant(DstReg, MFI->allocateLDSGlobal(B.getDataLayout(), *GV));
      MI.eraseFromParent();
      return true;
    }

    const Function &Fn = MF.getFunction();
    DiagnosticInfoUnsupported BadInit(
      Fn, "unsupported initializer for address space", MI.getDebugLoc());
    Fn.getContext().diagnose(BadInit);
    return true;
  }

  const SITargetLowering *TLI = ST.getTargetLowering();

  if (TLI->shouldEmitFixup(GV)) {
    buildPCRelGlobalAddress(DstReg, Ty, B, GV, 0);
    MI.eraseFromParent();
    return true;
  }

  if (TLI->shouldEmitPCReloc(GV)) {
    buildPCRelGlobalAddress(DstReg, Ty, B, GV, 0, SIInstrInfo::MO_REL32);
    MI.eraseFromParent();
    return true;
  }

  LLT PtrTy = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
  Register GOTAddr = MRI.createGenericVirtualRegister(PtrTy);

  MachineMemOperand *GOTMMO = MF.getMachineMemOperand(
      MachinePointerInfo::getGOT(MF),
      MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
          MachineMemOperand::MOInvariant,
      8 /*Size*/, Align(8));

  buildPCRelGlobalAddress(GOTAddr, PtrTy, B, GV, 0, SIInstrInfo::MO_GOTPCREL32);

  if (Ty.getSizeInBits() == 32) {
    // Truncate if this is a 32-bit constant adrdess.
    auto Load = B.buildLoad(PtrTy, GOTAddr, *GOTMMO);
    B.buildExtract(DstReg, Load, 0);
  } else
    B.buildLoad(DstReg, GOTAddr, *GOTMMO);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeLoad(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B, GISelChangeObserver &Observer) const {
  B.setInstr(MI);
  LLT ConstPtr = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
  auto Cast = B.buildAddrSpaceCast(ConstPtr, MI.getOperand(1).getReg());
  Observer.changingInstr(MI);
  MI.getOperand(1).setReg(Cast.getReg(0));
  Observer.changedInstr(MI);
  return true;
}

bool AMDGPULegalizerInfo::legalizeFMad(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  LLT Ty = MRI.getType(MI.getOperand(0).getReg());
  assert(Ty.isScalar());

  MachineFunction &MF = B.getMF();
  const SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();

  // TODO: Always legal with future ftz flag.
  // FIXME: Do we need just output?
  if (Ty == LLT::scalar(32) && !MFI->getMode().allFP32Denormals())
    return true;
  if (Ty == LLT::scalar(16) && !MFI->getMode().allFP64FP16Denormals())
    return true;

  MachineIRBuilder HelperBuilder(MI);
  GISelObserverWrapper DummyObserver;
  LegalizerHelper Helper(MF, DummyObserver, HelperBuilder);
  HelperBuilder.setInstr(MI);
  return Helper.lowerFMad(MI) == LegalizerHelper::Legalized;
}

bool AMDGPULegalizerInfo::legalizeAtomicCmpXChg(
  MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B) const {
  Register DstReg = MI.getOperand(0).getReg();
  Register PtrReg = MI.getOperand(1).getReg();
  Register CmpVal = MI.getOperand(2).getReg();
  Register NewVal = MI.getOperand(3).getReg();

  assert(SITargetLowering::isFlatGlobalAddrSpace(
           MRI.getType(PtrReg).getAddressSpace()) &&
         "this should not have been custom lowered");

  LLT ValTy = MRI.getType(CmpVal);
  LLT VecTy = LLT::vector(2, ValTy);

  B.setInstr(MI);
  Register PackedVal = B.buildBuildVector(VecTy, { NewVal, CmpVal }).getReg(0);

  B.buildInstr(AMDGPU::G_AMDGPU_ATOMIC_CMPXCHG)
    .addDef(DstReg)
    .addUse(PtrReg)
    .addUse(PackedVal)
    .setMemRefs(MI.memoperands());

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeFlog(
  MachineInstr &MI, MachineIRBuilder &B, double Log2BaseInverted) const {
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();
  LLT Ty = B.getMRI()->getType(Dst);
  unsigned Flags = MI.getFlags();
  B.setInstr(MI);

  auto Log2Operand = B.buildFLog2(Ty, Src, Flags);
  auto Log2BaseInvertedOperand = B.buildFConstant(Ty, Log2BaseInverted);

  B.buildFMul(Dst, Log2Operand, Log2BaseInvertedOperand, Flags);
  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeFExp(MachineInstr &MI,
                                       MachineIRBuilder &B) const {
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();
  unsigned Flags = MI.getFlags();
  LLT Ty = B.getMRI()->getType(Dst);
  B.setInstr(MI);

  auto K = B.buildFConstant(Ty, numbers::log2e);
  auto Mul = B.buildFMul(Ty, Src, K, Flags);
  B.buildFExp2(Dst, Mul, Flags);
  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeFPow(MachineInstr &MI,
                                       MachineIRBuilder &B) const {
  Register Dst = MI.getOperand(0).getReg();
  Register Src0 = MI.getOperand(1).getReg();
  Register Src1 = MI.getOperand(2).getReg();
  unsigned Flags = MI.getFlags();
  LLT Ty = B.getMRI()->getType(Dst);
  B.setInstr(MI);
  const LLT S16 = LLT::scalar(16);
  const LLT S32 = LLT::scalar(32);

  if (Ty == S32) {
    auto Log = B.buildFLog2(S32, Src0, Flags);
    auto Mul = B.buildIntrinsic(Intrinsic::amdgcn_fmul_legacy, {S32}, false)
      .addUse(Log.getReg(0))
      .addUse(Src1)
      .setMIFlags(Flags);
    B.buildFExp2(Dst, Mul, Flags);
  } else if (Ty == S16) {
    // There's no f16 fmul_legacy, so we need to convert for it.
    auto Log = B.buildFLog2(S16, Src0, Flags);
    auto Ext0 = B.buildFPExt(S32, Log, Flags);
    auto Ext1 = B.buildFPExt(S32, Src1, Flags);
    auto Mul = B.buildIntrinsic(Intrinsic::amdgcn_fmul_legacy, {S32}, false)
      .addUse(Ext0.getReg(0))
      .addUse(Ext1.getReg(0))
      .setMIFlags(Flags);

    B.buildFExp2(Dst, B.buildFPTrunc(S16, Mul), Flags);
  } else
    return false;

  MI.eraseFromParent();
  return true;
}

// Find a source register, ignoring any possible source modifiers.
static Register stripAnySourceMods(Register OrigSrc, MachineRegisterInfo &MRI) {
  Register ModSrc = OrigSrc;
  if (MachineInstr *SrcFNeg = getOpcodeDef(AMDGPU::G_FNEG, ModSrc, MRI)) {
    ModSrc = SrcFNeg->getOperand(1).getReg();
    if (MachineInstr *SrcFAbs = getOpcodeDef(AMDGPU::G_FABS, ModSrc, MRI))
      ModSrc = SrcFAbs->getOperand(1).getReg();
  } else if (MachineInstr *SrcFAbs = getOpcodeDef(AMDGPU::G_FABS, ModSrc, MRI))
    ModSrc = SrcFAbs->getOperand(1).getReg();
  return ModSrc;
}

bool AMDGPULegalizerInfo::legalizeFFloor(MachineInstr &MI,
                                         MachineRegisterInfo &MRI,
                                         MachineIRBuilder &B) const {
  B.setInstr(MI);

  const LLT S1 = LLT::scalar(1);
  const LLT S64 = LLT::scalar(64);
  Register Dst = MI.getOperand(0).getReg();
  Register OrigSrc = MI.getOperand(1).getReg();
  unsigned Flags = MI.getFlags();
  assert(ST.hasFractBug() && MRI.getType(Dst) == S64 &&
         "this should not have been custom lowered");

  // V_FRACT is buggy on SI, so the F32 version is never used and (x-floor(x))
  // is used instead. However, SI doesn't have V_FLOOR_F64, so the most
  // efficient way to implement it is using V_FRACT_F64. The workaround for the
  // V_FRACT bug is:
  //    fract(x) = isnan(x) ? x : min(V_FRACT(x), 0.99999999999999999)
  //
  // Convert floor(x) to (x - fract(x))

  auto Fract = B.buildIntrinsic(Intrinsic::amdgcn_fract, {S64}, false)
    .addUse(OrigSrc)
    .setMIFlags(Flags);

  // Give source modifier matching some assistance before obscuring a foldable
  // pattern.

  // TODO: We can avoid the neg on the fract? The input sign to fract
  // shouldn't matter?
  Register ModSrc = stripAnySourceMods(OrigSrc, MRI);

  auto Const = B.buildFConstant(S64, BitsToDouble(0x3fefffffffffffff));

  Register Min = MRI.createGenericVirtualRegister(S64);

  // We don't need to concern ourselves with the snan handling difference, so
  // use the one which will directly select.
  const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
  if (MFI->getMode().IEEE)
    B.buildFMinNumIEEE(Min, Fract, Const, Flags);
  else
    B.buildFMinNum(Min, Fract, Const, Flags);

  Register CorrectedFract = Min;
  if (!MI.getFlag(MachineInstr::FmNoNans)) {
    auto IsNan = B.buildFCmp(CmpInst::FCMP_ORD, S1, ModSrc, ModSrc, Flags);
    CorrectedFract = B.buildSelect(S64, IsNan, ModSrc, Min, Flags).getReg(0);
  }

  auto NegFract = B.buildFNeg(S64, CorrectedFract, Flags);
  B.buildFAdd(Dst, OrigSrc, NegFract, Flags);

  MI.eraseFromParent();
  return true;
}

// Turn an illegal packed v2s16 build vector into bit operations.
// TODO: This should probably be a bitcast action in LegalizerHelper.
bool AMDGPULegalizerInfo::legalizeBuildVector(
  MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B) const {
  Register Dst = MI.getOperand(0).getReg();
  const LLT S32 = LLT::scalar(32);
  assert(MRI.getType(Dst) == LLT::vector(2, 16));

  Register Src0 = MI.getOperand(1).getReg();
  Register Src1 = MI.getOperand(2).getReg();
  assert(MRI.getType(Src0) == LLT::scalar(16));

  B.setInstr(MI);
  auto Merge = B.buildMerge(S32, {Src0, Src1});
  B.buildBitcast(Dst, Merge);

  MI.eraseFromParent();
  return true;
}

// Return the use branch instruction, otherwise null if the usage is invalid.
static MachineInstr *verifyCFIntrinsic(MachineInstr &MI,
                                       MachineRegisterInfo &MRI,
                                       MachineInstr *&Br,
                                       MachineBasicBlock *&UncondBrTarget) {
  Register CondDef = MI.getOperand(0).getReg();
  if (!MRI.hasOneNonDBGUse(CondDef))
    return nullptr;

  MachineBasicBlock *Parent = MI.getParent();
  MachineInstr &UseMI = *MRI.use_instr_nodbg_begin(CondDef);
  if (UseMI.getParent() != Parent ||
      UseMI.getOpcode() != AMDGPU::G_BRCOND)
    return nullptr;

  // Make sure the cond br is followed by a G_BR, or is the last instruction.
  MachineBasicBlock::iterator Next = std::next(UseMI.getIterator());
  if (Next == Parent->end()) {
    MachineFunction::iterator NextMBB = std::next(Parent->getIterator());
    if (NextMBB == Parent->getParent()->end()) // Illegal intrinsic use.
      return nullptr;
    UncondBrTarget = &*NextMBB;
  } else {
    if (Next->getOpcode() != AMDGPU::G_BR)
      return nullptr;
    Br = &*Next;
    UncondBrTarget = Br->getOperand(0).getMBB();
  }

  return &UseMI;
}

Register AMDGPULegalizerInfo::insertLiveInCopy(MachineIRBuilder &B,
                                               MachineRegisterInfo &MRI,
                                               Register LiveIn,
                                               Register PhyReg) const {
  assert(PhyReg.isPhysical() && "Physical register expected");

  // Insert the live-in copy, if required, by defining destination virtual
  // register.
  // FIXME: It seems EmitLiveInCopies isn't called anywhere?
  if (!MRI.getVRegDef(LiveIn)) {
    // FIXME: Should have scoped insert pt
    MachineBasicBlock &OrigInsBB = B.getMBB();
    auto OrigInsPt = B.getInsertPt();

    MachineBasicBlock &EntryMBB = B.getMF().front();
    EntryMBB.addLiveIn(PhyReg);
    B.setInsertPt(EntryMBB, EntryMBB.begin());
    B.buildCopy(LiveIn, PhyReg);

    B.setInsertPt(OrigInsBB, OrigInsPt);
  }

  return LiveIn;
}

Register AMDGPULegalizerInfo::getLiveInRegister(MachineIRBuilder &B,
                                                MachineRegisterInfo &MRI,
                                                Register PhyReg, LLT Ty,
                                                bool InsertLiveInCopy) const {
  assert(PhyReg.isPhysical() && "Physical register expected");

  // Get or create virtual live-in regester
  Register LiveIn = MRI.getLiveInVirtReg(PhyReg);
  if (!LiveIn) {
    LiveIn = MRI.createGenericVirtualRegister(Ty);
    MRI.addLiveIn(PhyReg, LiveIn);
  }

  // When the actual true copy required is from virtual register to physical
  // register (to be inserted later), live-in copy insertion from physical
  // to register virtual register is not required
  if (!InsertLiveInCopy)
    return LiveIn;

  return insertLiveInCopy(B, MRI, LiveIn, PhyReg);
}

const ArgDescriptor *AMDGPULegalizerInfo::getArgDescriptor(
    MachineIRBuilder &B, AMDGPUFunctionArgInfo::PreloadedValue ArgType) const {
  const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
  const ArgDescriptor *Arg;
  const TargetRegisterClass *RC;
  std::tie(Arg, RC) = MFI->getPreloadedValue(ArgType);
  if (!Arg) {
    LLVM_DEBUG(dbgs() << "Required arg register missing\n");
    return nullptr;
  }
  return Arg;
}

bool AMDGPULegalizerInfo::loadInputValue(Register DstReg, MachineIRBuilder &B,
                                         const ArgDescriptor *Arg) const {
  if (!Arg->isRegister() || !Arg->getRegister().isValid())
    return false; // TODO: Handle these

  Register SrcReg = Arg->getRegister();
  assert(SrcReg.isPhysical() && "Physical register expected");
  assert(DstReg.isVirtual() && "Virtual register expected");

  MachineRegisterInfo &MRI = *B.getMRI();

  LLT Ty = MRI.getType(DstReg);
  Register LiveIn = getLiveInRegister(B, MRI, SrcReg, Ty);

  if (Arg->isMasked()) {
    // TODO: Should we try to emit this once in the entry block?
    const LLT S32 = LLT::scalar(32);
    const unsigned Mask = Arg->getMask();
    const unsigned Shift = countTrailingZeros<unsigned>(Mask);

    Register AndMaskSrc = LiveIn;

    if (Shift != 0) {
      auto ShiftAmt = B.buildConstant(S32, Shift);
      AndMaskSrc = B.buildLShr(S32, LiveIn, ShiftAmt).getReg(0);
    }

    B.buildAnd(DstReg, AndMaskSrc, B.buildConstant(S32, Mask >> Shift));
  } else {
    B.buildCopy(DstReg, LiveIn);
  }

  return true;
}

bool AMDGPULegalizerInfo::legalizePreloadedArgIntrin(
    MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B,
    AMDGPUFunctionArgInfo::PreloadedValue ArgType) const {
  B.setInstr(MI);

  const ArgDescriptor *Arg = getArgDescriptor(B, ArgType);
  if (!Arg)
    return false;

  if (!loadInputValue(MI.getOperand(0).getReg(), B, Arg))
    return false;

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeFDIV(MachineInstr &MI,
                                       MachineRegisterInfo &MRI,
                                       MachineIRBuilder &B) const {
  B.setInstr(MI);
  Register Dst = MI.getOperand(0).getReg();
  LLT DstTy = MRI.getType(Dst);
  LLT S16 = LLT::scalar(16);
  LLT S32 = LLT::scalar(32);
  LLT S64 = LLT::scalar(64);

  if (legalizeFastUnsafeFDIV(MI, MRI, B))
    return true;

  if (DstTy == S16)
    return legalizeFDIV16(MI, MRI, B);
  if (DstTy == S32)
    return legalizeFDIV32(MI, MRI, B);
  if (DstTy == S64)
    return legalizeFDIV64(MI, MRI, B);

  return false;
}

static Register buildDivRCP(MachineIRBuilder &B, Register Src) {
  const LLT S32 = LLT::scalar(32);

  auto Cvt0 = B.buildUITOFP(S32, Src);
  auto RcpIFlag = B.buildInstr(AMDGPU::G_AMDGPU_RCP_IFLAG, {S32}, {Cvt0});
  auto FPUIntMaxPlus1 = B.buildFConstant(S32, BitsToFloat(0x4f800000));
  auto Mul = B.buildFMul(S32, RcpIFlag, FPUIntMaxPlus1);
  return B.buildFPTOUI(S32, Mul).getReg(0);
}

void AMDGPULegalizerInfo::legalizeUDIV_UREM32Impl(MachineIRBuilder &B,
                                                  Register DstReg,
                                                  Register Num,
                                                  Register Den,
                                                  bool IsRem) const {
  const LLT S1 = LLT::scalar(1);
  const LLT S32 = LLT::scalar(32);

  // RCP =  URECIP(Den) = 2^32 / Den + e
  // e is rounding error.
  auto RCP = buildDivRCP(B, Den);

  // RCP_LO = mul(RCP, Den)
  auto RCP_LO = B.buildMul(S32, RCP, Den);

  // RCP_HI = mulhu (RCP, Den) */
  auto RCP_HI = B.buildUMulH(S32, RCP, Den);

  // NEG_RCP_LO = -RCP_LO
  auto Zero = B.buildConstant(S32, 0);
  auto NEG_RCP_LO = B.buildSub(S32, Zero, RCP_LO);

  // ABS_RCP_LO = (RCP_HI == 0 ? NEG_RCP_LO : RCP_LO)
  auto CmpRcpHiZero = B.buildICmp(CmpInst::ICMP_EQ, S1, RCP_HI, Zero);
  auto ABS_RCP_LO = B.buildSelect(S32, CmpRcpHiZero, NEG_RCP_LO, RCP_LO);

  // Calculate the rounding error from the URECIP instruction
  // E = mulhu(ABS_RCP_LO, RCP)
  auto E = B.buildUMulH(S32, ABS_RCP_LO, RCP);

  // RCP_A_E = RCP + E
  auto RCP_A_E = B.buildAdd(S32, RCP, E);

  // RCP_S_E = RCP - E
  auto RCP_S_E = B.buildSub(S32, RCP, E);

  // Tmp0 = (RCP_HI == 0 ? RCP_A_E : RCP_SUB_E)
  auto Tmp0 = B.buildSelect(S32, CmpRcpHiZero, RCP_A_E, RCP_S_E);

  // Quotient = mulhu(Tmp0, Num)stmp
  auto Quotient = B.buildUMulH(S32, Tmp0, Num);

  // Num_S_Remainder = Quotient * Den
  auto Num_S_Remainder = B.buildMul(S32, Quotient, Den);

  // Remainder = Num - Num_S_Remainder
  auto Remainder = B.buildSub(S32, Num, Num_S_Remainder);

  // Remainder_GE_Den = Remainder >= Den
  auto Remainder_GE_Den = B.buildICmp(CmpInst::ICMP_UGE, S1, Remainder, Den);

  // Remainder_GE_Zero = Num >= Num_S_Remainder;
  auto Remainder_GE_Zero = B.buildICmp(CmpInst::ICMP_UGE, S1,
                                       Num, Num_S_Remainder);

  // Tmp1 = Remainder_GE_Den & Remainder_GE_Zero
  auto Tmp1 = B.buildAnd(S1, Remainder_GE_Den, Remainder_GE_Zero);

  // Calculate Division result:

  // Quotient_A_One = Quotient + 1
  auto One = B.buildConstant(S32, 1);
  auto Quotient_A_One = B.buildAdd(S32, Quotient, One);

  // Quotient_S_One = Quotient - 1
  auto Quotient_S_One = B.buildSub(S32, Quotient, One);

  // Div = (Tmp1 == 0 ? Quotient_A_One : Quotient)
  auto Div = B.buildSelect(S32, Tmp1, Quotient, Quotient_A_One);

  // Div = (Remainder_GE_Zero ? Div : Quotient_S_One)
  if (IsRem) {
    Div = B.buildSelect(S32, Remainder_GE_Zero, Div, Quotient_S_One);

    // Calculate Rem result:
    auto Remainder_S_Den = B.buildSub(S32, Remainder, Den);

    // Remainder_A_Den = Remainder + Den
    auto Remainder_A_Den = B.buildAdd(S32, Remainder, Den);

    // Rem = (Tmp1 ? Remainder_S_Den : Remainder)
    auto Rem = B.buildSelect(S32, Tmp1, Remainder_S_Den, Remainder);

    // Rem = (Remainder_GE_Zero ? Rem : Remainder_A_Den)
    B.buildSelect(DstReg, Remainder_GE_Zero, Rem, Remainder_A_Den);
  } else {
    B.buildSelect(DstReg, Remainder_GE_Zero, Div, Quotient_S_One);
  }
}

bool AMDGPULegalizerInfo::legalizeUDIV_UREM32(MachineInstr &MI,
                                              MachineRegisterInfo &MRI,
                                              MachineIRBuilder &B) const {
  B.setInstr(MI);
  const bool IsRem = MI.getOpcode() == AMDGPU::G_UREM;
  Register DstReg = MI.getOperand(0).getReg();
  Register Num = MI.getOperand(1).getReg();
  Register Den = MI.getOperand(2).getReg();
  legalizeUDIV_UREM32Impl(B, DstReg, Num, Den, IsRem);
  MI.eraseFromParent();
  return true;
}

// Build integer reciprocal sequence arounud V_RCP_IFLAG_F32
//
// Return lo, hi of result
//
// %cvt.lo = G_UITOFP Val.lo
// %cvt.hi = G_UITOFP Val.hi
// %mad = G_FMAD %cvt.hi, 2**32, %cvt.lo
// %rcp = G_AMDGPU_RCP_IFLAG %mad
// %mul1 = G_FMUL %rcp, 0x5f7ffffc
// %mul2 = G_FMUL %mul1, 2**(-32)
// %trunc = G_INTRINSIC_TRUNC %mul2
// %mad2 = G_FMAD %trunc, -(2**32), %mul1
// return {G_FPTOUI %mad2, G_FPTOUI %trunc}
static std::pair<Register, Register> emitReciprocalU64(MachineIRBuilder &B,
                                                       Register Val) {
  const LLT S32 = LLT::scalar(32);
  auto Unmerge = B.buildUnmerge(S32, Val);

  auto CvtLo = B.buildUITOFP(S32, Unmerge.getReg(0));
  auto CvtHi = B.buildUITOFP(S32, Unmerge.getReg(1));

  auto Mad = B.buildFMAD(S32, CvtHi, // 2**32
                         B.buildFConstant(S32, BitsToFloat(0x4f800000)), CvtLo);

  auto Rcp = B.buildInstr(AMDGPU::G_AMDGPU_RCP_IFLAG, {S32}, {Mad});
  auto Mul1 =
      B.buildFMul(S32, Rcp, B.buildFConstant(S32, BitsToFloat(0x5f7ffffc)));

  // 2**(-32)
  auto Mul2 =
      B.buildFMul(S32, Mul1, B.buildFConstant(S32, BitsToFloat(0x2f800000)));
  auto Trunc = B.buildIntrinsicTrunc(S32, Mul2);

  // -(2**32)
  auto Mad2 = B.buildFMAD(S32, Trunc,
                          B.buildFConstant(S32, BitsToFloat(0xcf800000)), Mul1);

  auto ResultLo = B.buildFPTOUI(S32, Mad2);
  auto ResultHi = B.buildFPTOUI(S32, Trunc);

  return {ResultLo.getReg(0), ResultHi.getReg(0)};
}

bool AMDGPULegalizerInfo::legalizeUDIV_UREM64(MachineInstr &MI,
                                              MachineRegisterInfo &MRI,
                                              MachineIRBuilder &B) const {
  B.setInstr(MI);

  const bool IsDiv = MI.getOpcode() == TargetOpcode::G_UDIV;
  const LLT S32 = LLT::scalar(32);
  const LLT S64 = LLT::scalar(64);
  const LLT S1 = LLT::scalar(1);
  Register Numer = MI.getOperand(1).getReg();
  Register Denom = MI.getOperand(2).getReg();
  Register RcpLo, RcpHi;

  std::tie(RcpLo, RcpHi) = emitReciprocalU64(B, Denom);

  auto Rcp = B.buildMerge(S64, {RcpLo, RcpHi});

  auto Zero64 = B.buildConstant(S64, 0);
  auto NegDenom = B.buildSub(S64, Zero64, Denom);

  auto MulLo1 = B.buildMul(S64, NegDenom, Rcp);
  auto MulHi1 = B.buildUMulH(S64, Rcp, MulLo1);

  auto UnmergeMulHi1 = B.buildUnmerge(S32, MulHi1);
  Register MulHi1_Lo = UnmergeMulHi1.getReg(0);
  Register MulHi1_Hi = UnmergeMulHi1.getReg(1);

  auto Add1_Lo = B.buildUAddo(S32, S1, RcpLo, MulHi1_Lo);
  auto Add1_Hi = B.buildUAdde(S32, S1, RcpHi, MulHi1_Hi, Add1_Lo.getReg(1));
  auto Add1_HiNc = B.buildAdd(S32, RcpHi, MulHi1_Hi);
  auto Add1 = B.buildMerge(S64, {Add1_Lo, Add1_Hi});

  auto MulLo2 = B.buildMul(S64, NegDenom, Add1);
  auto MulHi2 = B.buildUMulH(S64, Add1, MulLo2);
  auto UnmergeMulHi2 = B.buildUnmerge(S32, MulHi2);
  Register MulHi2_Lo = UnmergeMulHi2.getReg(0);
  Register MulHi2_Hi = UnmergeMulHi2.getReg(1);

  auto Zero32 = B.buildConstant(S32, 0);
  auto Add2_Lo = B.buildUAddo(S32, S1, Add1_Lo, MulHi2_Lo);
  auto Add2_HiC =
      B.buildUAdde(S32, S1, Add1_HiNc, MulHi2_Hi, Add1_Lo.getReg(1));
  auto Add2_Hi = B.buildUAdde(S32, S1, Add2_HiC, Zero32, Add2_Lo.getReg(1));
  auto Add2 = B.buildMerge(S64, {Add2_Lo, Add2_Hi});

  auto UnmergeNumer = B.buildUnmerge(S32, Numer);
  Register NumerLo = UnmergeNumer.getReg(0);
  Register NumerHi = UnmergeNumer.getReg(1);

  auto MulHi3 = B.buildUMulH(S64, Numer, Add2);
  auto Mul3 = B.buildMul(S64, Denom, MulHi3);
  auto UnmergeMul3 = B.buildUnmerge(S32, Mul3);
  Register Mul3_Lo = UnmergeMul3.getReg(0);
  Register Mul3_Hi = UnmergeMul3.getReg(1);
  auto Sub1_Lo = B.buildUSubo(S32, S1, NumerLo, Mul3_Lo);
  auto Sub1_Hi = B.buildUSube(S32, S1, NumerHi, Mul3_Hi, Sub1_Lo.getReg(1));
  auto Sub1_Mi = B.buildSub(S32, NumerHi, Mul3_Hi);
  auto Sub1 = B.buildMerge(S64, {Sub1_Lo, Sub1_Hi});

  auto UnmergeDenom = B.buildUnmerge(S32, Denom);
  Register DenomLo = UnmergeDenom.getReg(0);
  Register DenomHi = UnmergeDenom.getReg(1);

  auto CmpHi = B.buildICmp(CmpInst::ICMP_UGE, S1, Sub1_Hi, DenomHi);
  auto C1 = B.buildSExt(S32, CmpHi);

  auto CmpLo = B.buildICmp(CmpInst::ICMP_UGE, S1, Sub1_Lo, DenomLo);
  auto C2 = B.buildSExt(S32, CmpLo);

  auto CmpEq = B.buildICmp(CmpInst::ICMP_EQ, S1, Sub1_Hi, DenomHi);
  auto C3 = B.buildSelect(S32, CmpEq, C2, C1);

  // TODO: Here and below portions of the code can be enclosed into if/endif.
  // Currently control flow is unconditional and we have 4 selects after
  // potential endif to substitute PHIs.

  // if C3 != 0 ...
  auto Sub2_Lo = B.buildUSubo(S32, S1, Sub1_Lo, DenomLo);
  auto Sub2_Mi = B.buildUSube(S32, S1, Sub1_Mi, DenomHi, Sub1_Lo.getReg(1));
  auto Sub2_Hi = B.buildUSube(S32, S1, Sub2_Mi, Zero32, Sub2_Lo.getReg(1));
  auto Sub2 = B.buildMerge(S64, {Sub2_Lo, Sub2_Hi});

  auto One64 = B.buildConstant(S64, 1);
  auto Add3 = B.buildAdd(S64, MulHi3, One64);

  auto C4 =
      B.buildSExt(S32, B.buildICmp(CmpInst::ICMP_UGE, S1, Sub2_Hi, DenomHi));
  auto C5 =
      B.buildSExt(S32, B.buildICmp(CmpInst::ICMP_UGE, S1, Sub2_Lo, DenomLo));
  auto C6 = B.buildSelect(
      S32, B.buildICmp(CmpInst::ICMP_EQ, S1, Sub2_Hi, DenomHi), C5, C4);

  // if (C6 != 0)
  auto Add4 = B.buildAdd(S64, Add3, One64);
  auto Sub3_Lo = B.buildUSubo(S32, S1, Sub2_Lo, DenomLo);

  auto Sub3_Mi = B.buildUSube(S32, S1, Sub2_Mi, DenomHi, Sub2_Lo.getReg(1));
  auto Sub3_Hi = B.buildUSube(S32, S1, Sub3_Mi, Zero32, Sub3_Lo.getReg(1));
  auto Sub3 = B.buildMerge(S64, {Sub3_Lo, Sub3_Hi});

  // endif C6
  // endif C3

  if (IsDiv) {
    auto Sel1 = B.buildSelect(
        S64, B.buildICmp(CmpInst::ICMP_NE, S1, C6, Zero32), Add4, Add3);
    B.buildSelect(MI.getOperand(0),
                  B.buildICmp(CmpInst::ICMP_NE, S1, C3, Zero32), Sel1, MulHi3);
  } else {
    auto Sel2 = B.buildSelect(
        S64, B.buildICmp(CmpInst::ICMP_NE, S1, C6, Zero32), Sub3, Sub2);
    B.buildSelect(MI.getOperand(0),
                  B.buildICmp(CmpInst::ICMP_NE, S1, C3, Zero32), Sel2, Sub1);
  }

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeUDIV_UREM(MachineInstr &MI,
                                            MachineRegisterInfo &MRI,
                                            MachineIRBuilder &B) const {
  LLT Ty = MRI.getType(MI.getOperand(0).getReg());
  if (Ty == LLT::scalar(32))
    return legalizeUDIV_UREM32(MI, MRI, B);
  if (Ty == LLT::scalar(64))
    return legalizeUDIV_UREM64(MI, MRI, B);
  return false;
}

bool AMDGPULegalizerInfo::legalizeSDIV_SREM32(MachineInstr &MI,
                                              MachineRegisterInfo &MRI,
                                              MachineIRBuilder &B) const {
  B.setInstr(MI);
  const LLT S32 = LLT::scalar(32);

  const bool IsRem = MI.getOpcode() == AMDGPU::G_SREM;
  Register DstReg = MI.getOperand(0).getReg();
  Register LHS = MI.getOperand(1).getReg();
  Register RHS = MI.getOperand(2).getReg();

  auto ThirtyOne = B.buildConstant(S32, 31);
  auto LHSign = B.buildAShr(S32, LHS, ThirtyOne);
  auto RHSign = B.buildAShr(S32, LHS, ThirtyOne);

  LHS = B.buildAdd(S32, LHS, LHSign).getReg(0);
  RHS = B.buildAdd(S32, RHS, RHSign).getReg(0);

  LHS = B.buildXor(S32, LHS, LHSign).getReg(0);
  RHS = B.buildXor(S32, RHS, RHSign).getReg(0);

  Register UDivRem = MRI.createGenericVirtualRegister(S32);
  legalizeUDIV_UREM32Impl(B, UDivRem, LHS, RHS, IsRem);

  if (IsRem) {
    auto RSign = LHSign; // Remainder sign is the same as LHS
    UDivRem = B.buildXor(S32, UDivRem, RSign).getReg(0);
    B.buildSub(DstReg, UDivRem, RSign);
  } else {
    auto DSign = B.buildXor(S32, LHSign, RHSign);
    UDivRem = B.buildXor(S32, UDivRem, DSign).getReg(0);
    B.buildSub(DstReg, UDivRem, DSign);
  }

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeSDIV_SREM(MachineInstr &MI,
                                            MachineRegisterInfo &MRI,
                                            MachineIRBuilder &B) const {
  if (MRI.getType(MI.getOperand(0).getReg()) == LLT::scalar(32))
    return legalizeSDIV_SREM32(MI, MRI, B);
  return false;
}

bool AMDGPULegalizerInfo::legalizeFastUnsafeFDIV(MachineInstr &MI,
                                                 MachineRegisterInfo &MRI,
                                                 MachineIRBuilder &B) const {
  Register Res = MI.getOperand(0).getReg();
  Register LHS = MI.getOperand(1).getReg();
  Register RHS = MI.getOperand(2).getReg();

  uint16_t Flags = MI.getFlags();

  LLT ResTy = MRI.getType(Res);
  LLT S32 = LLT::scalar(32);
  LLT S64 = LLT::scalar(64);

  const MachineFunction &MF = B.getMF();
  bool Unsafe =
    MF.getTarget().Options.UnsafeFPMath || MI.getFlag(MachineInstr::FmArcp);

  if (!MF.getTarget().Options.UnsafeFPMath && ResTy == S64)
    return false;

  if (!Unsafe && ResTy == S32 &&
      MF.getInfo<SIMachineFunctionInfo>()->getMode().allFP32Denormals())
    return false;

  if (auto CLHS = getConstantFPVRegVal(LHS, MRI)) {
    // 1 / x -> RCP(x)
    if (CLHS->isExactlyValue(1.0)) {
      B.buildIntrinsic(Intrinsic::amdgcn_rcp, Res, false)
        .addUse(RHS)
        .setMIFlags(Flags);

      MI.eraseFromParent();
      return true;
    }

    // -1 / x -> RCP( FNEG(x) )
    if (CLHS->isExactlyValue(-1.0)) {
      auto FNeg = B.buildFNeg(ResTy, RHS, Flags);
      B.buildIntrinsic(Intrinsic::amdgcn_rcp, Res, false)
        .addUse(FNeg.getReg(0))
        .setMIFlags(Flags);

      MI.eraseFromParent();
      return true;
    }
  }

  // x / y -> x * (1.0 / y)
  if (Unsafe) {
    auto RCP = B.buildIntrinsic(Intrinsic::amdgcn_rcp, {ResTy}, false)
      .addUse(RHS)
      .setMIFlags(Flags);
    B.buildFMul(Res, LHS, RCP, Flags);

    MI.eraseFromParent();
    return true;
  }

  return false;
}

bool AMDGPULegalizerInfo::legalizeFDIV16(MachineInstr &MI,
                                         MachineRegisterInfo &MRI,
                                         MachineIRBuilder &B) const {
  B.setInstr(MI);
  Register Res = MI.getOperand(0).getReg();
  Register LHS = MI.getOperand(1).getReg();
  Register RHS = MI.getOperand(2).getReg();

  uint16_t Flags = MI.getFlags();

  LLT S16 = LLT::scalar(16);
  LLT S32 = LLT::scalar(32);

  auto LHSExt = B.buildFPExt(S32, LHS, Flags);
  auto RHSExt = B.buildFPExt(S32, RHS, Flags);

  auto RCP = B.buildIntrinsic(Intrinsic::amdgcn_rcp, {S32}, false)
    .addUse(RHSExt.getReg(0))
    .setMIFlags(Flags);

  auto QUOT = B.buildFMul(S32, LHSExt, RCP, Flags);
  auto RDst = B.buildFPTrunc(S16, QUOT, Flags);

  B.buildIntrinsic(Intrinsic::amdgcn_div_fixup, Res, false)
    .addUse(RDst.getReg(0))
    .addUse(RHS)
    .addUse(LHS)
    .setMIFlags(Flags);

  MI.eraseFromParent();
  return true;
}

// Enable or disable FP32 denorm mode. When 'Enable' is true, emit instructions
// to enable denorm mode. When 'Enable' is false, disable denorm mode.
static void toggleSPDenormMode(bool Enable,
                               MachineIRBuilder &B,
                               const GCNSubtarget &ST,
                               AMDGPU::SIModeRegisterDefaults Mode) {
  // Set SP denorm mode to this value.
  unsigned SPDenormMode =
    Enable ? FP_DENORM_FLUSH_NONE : Mode.fpDenormModeSPValue();

  if (ST.hasDenormModeInst()) {
    // Preserve default FP64FP16 denorm mode while updating FP32 mode.
    uint32_t DPDenormModeDefault = Mode.fpDenormModeDPValue();

    uint32_t NewDenormModeValue = SPDenormMode | (DPDenormModeDefault << 2);
    B.buildInstr(AMDGPU::S_DENORM_MODE)
      .addImm(NewDenormModeValue);

  } else {
    // Select FP32 bit field in mode register.
    unsigned SPDenormModeBitField = AMDGPU::Hwreg::ID_MODE |
                                    (4 << AMDGPU::Hwreg::OFFSET_SHIFT_) |
                                    (1 << AMDGPU::Hwreg::WIDTH_M1_SHIFT_);

    B.buildInstr(AMDGPU::S_SETREG_IMM32_B32)
      .addImm(SPDenormMode)
      .addImm(SPDenormModeBitField);
  }
}

bool AMDGPULegalizerInfo::legalizeFDIV32(MachineInstr &MI,
                                         MachineRegisterInfo &MRI,
                                         MachineIRBuilder &B) const {
  B.setInstr(MI);
  Register Res = MI.getOperand(0).getReg();
  Register LHS = MI.getOperand(1).getReg();
  Register RHS = MI.getOperand(2).getReg();
  const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
  AMDGPU::SIModeRegisterDefaults Mode = MFI->getMode();

  uint16_t Flags = MI.getFlags();

  LLT S32 = LLT::scalar(32);
  LLT S1 = LLT::scalar(1);

  auto One = B.buildFConstant(S32, 1.0f);

  auto DenominatorScaled =
    B.buildIntrinsic(Intrinsic::amdgcn_div_scale, {S32, S1}, false)
      .addUse(LHS)
      .addUse(RHS)
      .addImm(0)
      .setMIFlags(Flags);
  auto NumeratorScaled =
    B.buildIntrinsic(Intrinsic::amdgcn_div_scale, {S32, S1}, false)
      .addUse(LHS)
      .addUse(RHS)
      .addImm(1)
      .setMIFlags(Flags);

  auto ApproxRcp = B.buildIntrinsic(Intrinsic::amdgcn_rcp, {S32}, false)
    .addUse(DenominatorScaled.getReg(0))
    .setMIFlags(Flags);
  auto NegDivScale0 = B.buildFNeg(S32, DenominatorScaled, Flags);

  // FIXME: Doesn't correctly model the FP mode switch, and the FP operations
  // aren't modeled as reading it.
  if (!Mode.allFP32Denormals())
    toggleSPDenormMode(true, B, ST, Mode);

  auto Fma0 = B.buildFMA(S32, NegDivScale0, ApproxRcp, One, Flags);
  auto Fma1 = B.buildFMA(S32, Fma0, ApproxRcp, ApproxRcp, Flags);
  auto Mul = B.buildFMul(S32, NumeratorScaled, Fma1, Flags);
  auto Fma2 = B.buildFMA(S32, NegDivScale0, Mul, NumeratorScaled, Flags);
  auto Fma3 = B.buildFMA(S32, Fma2, Fma1, Mul, Flags);
  auto Fma4 = B.buildFMA(S32, NegDivScale0, Fma3, NumeratorScaled, Flags);

  if (!Mode.allFP32Denormals())
    toggleSPDenormMode(false, B, ST, Mode);

  auto Fmas = B.buildIntrinsic(Intrinsic::amdgcn_div_fmas, {S32}, false)
    .addUse(Fma4.getReg(0))
    .addUse(Fma1.getReg(0))
    .addUse(Fma3.getReg(0))
    .addUse(NumeratorScaled.getReg(1))
    .setMIFlags(Flags);

  B.buildIntrinsic(Intrinsic::amdgcn_div_fixup, Res, false)
    .addUse(Fmas.getReg(0))
    .addUse(RHS)
    .addUse(LHS)
    .setMIFlags(Flags);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeFDIV64(MachineInstr &MI,
                                         MachineRegisterInfo &MRI,
                                         MachineIRBuilder &B) const {
  B.setInstr(MI);
  Register Res = MI.getOperand(0).getReg();
  Register LHS = MI.getOperand(1).getReg();
  Register RHS = MI.getOperand(2).getReg();

  uint16_t Flags = MI.getFlags();

  LLT S64 = LLT::scalar(64);
  LLT S1 = LLT::scalar(1);

  auto One = B.buildFConstant(S64, 1.0);

  auto DivScale0 = B.buildIntrinsic(Intrinsic::amdgcn_div_scale, {S64, S1}, false)
    .addUse(LHS)
    .addUse(RHS)
    .addImm(0)
    .setMIFlags(Flags);

  auto NegDivScale0 = B.buildFNeg(S64, DivScale0.getReg(0), Flags);

  auto Rcp = B.buildIntrinsic(Intrinsic::amdgcn_rcp, {S64}, false)
    .addUse(DivScale0.getReg(0))
    .setMIFlags(Flags);

  auto Fma0 = B.buildFMA(S64, NegDivScale0, Rcp, One, Flags);
  auto Fma1 = B.buildFMA(S64, Rcp, Fma0, Rcp, Flags);
  auto Fma2 = B.buildFMA(S64, NegDivScale0, Fma1, One, Flags);

  auto DivScale1 = B.buildIntrinsic(Intrinsic::amdgcn_div_scale, {S64, S1}, false)
    .addUse(LHS)
    .addUse(RHS)
    .addImm(1)
    .setMIFlags(Flags);

  auto Fma3 = B.buildFMA(S64, Fma1, Fma2, Fma1, Flags);
  auto Mul = B.buildFMul(S64, DivScale1.getReg(0), Fma3, Flags);
  auto Fma4 = B.buildFMA(S64, NegDivScale0, Mul, DivScale1.getReg(0), Flags);

  Register Scale;
  if (!ST.hasUsableDivScaleConditionOutput()) {
    // Workaround a hardware bug on SI where the condition output from div_scale
    // is not usable.

    LLT S32 = LLT::scalar(32);

    auto NumUnmerge = B.buildUnmerge(S32, LHS);
    auto DenUnmerge = B.buildUnmerge(S32, RHS);
    auto Scale0Unmerge = B.buildUnmerge(S32, DivScale0);
    auto Scale1Unmerge = B.buildUnmerge(S32, DivScale1);

    auto CmpNum = B.buildICmp(ICmpInst::ICMP_EQ, S1, NumUnmerge.getReg(1),
                              Scale1Unmerge.getReg(1));
    auto CmpDen = B.buildICmp(ICmpInst::ICMP_EQ, S1, DenUnmerge.getReg(1),
                              Scale0Unmerge.getReg(1));
    Scale = B.buildXor(S1, CmpNum, CmpDen).getReg(0);
  } else {
    Scale = DivScale1.getReg(1);
  }

  auto Fmas = B.buildIntrinsic(Intrinsic::amdgcn_div_fmas, {S64}, false)
    .addUse(Fma4.getReg(0))
    .addUse(Fma3.getReg(0))
    .addUse(Mul.getReg(0))
    .addUse(Scale)
    .setMIFlags(Flags);

  B.buildIntrinsic(Intrinsic::amdgcn_div_fixup, makeArrayRef(Res), false)
    .addUse(Fmas.getReg(0))
    .addUse(RHS)
    .addUse(LHS)
    .setMIFlags(Flags);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeFDIVFastIntrin(MachineInstr &MI,
                                                 MachineRegisterInfo &MRI,
                                                 MachineIRBuilder &B) const {
  B.setInstr(MI);
  Register Res = MI.getOperand(0).getReg();
  Register LHS = MI.getOperand(2).getReg();
  Register RHS = MI.getOperand(3).getReg();
  uint16_t Flags = MI.getFlags();

  LLT S32 = LLT::scalar(32);
  LLT S1 = LLT::scalar(1);

  auto Abs = B.buildFAbs(S32, RHS, Flags);
  const APFloat C0Val(1.0f);

  auto C0 = B.buildConstant(S32, 0x6f800000);
  auto C1 = B.buildConstant(S32, 0x2f800000);
  auto C2 = B.buildConstant(S32, FloatToBits(1.0f));

  auto CmpRes = B.buildFCmp(CmpInst::FCMP_OGT, S1, Abs, C0, Flags);
  auto Sel = B.buildSelect(S32, CmpRes, C1, C2, Flags);

  auto Mul0 = B.buildFMul(S32, RHS, Sel, Flags);

  auto RCP = B.buildIntrinsic(Intrinsic::amdgcn_rcp, {S32}, false)
    .addUse(Mul0.getReg(0))
    .setMIFlags(Flags);

  auto Mul1 = B.buildFMul(S32, LHS, RCP, Flags);

  B.buildFMul(Res, Sel, Mul1, Flags);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeImplicitArgPtr(MachineInstr &MI,
                                                 MachineRegisterInfo &MRI,
                                                 MachineIRBuilder &B) const {
  const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
  if (!MFI->isEntryFunction()) {
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::IMPLICIT_ARG_PTR);
  }

  B.setInstr(MI);

  uint64_t Offset =
    ST.getTargetLowering()->getImplicitParameterOffset(
      B.getMF(), AMDGPUTargetLowering::FIRST_IMPLICIT);
  Register DstReg = MI.getOperand(0).getReg();
  LLT DstTy = MRI.getType(DstReg);
  LLT IdxTy = LLT::scalar(DstTy.getSizeInBits());

  const ArgDescriptor *Arg;
  const TargetRegisterClass *RC;
  std::tie(Arg, RC)
    = MFI->getPreloadedValue(AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
  if (!Arg)
    return false;

  Register KernargPtrReg = MRI.createGenericVirtualRegister(DstTy);
  if (!loadInputValue(KernargPtrReg, B, Arg))
    return false;

  B.buildPtrAdd(DstReg, KernargPtrReg, B.buildConstant(IdxTy, Offset).getReg(0));
  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeIsAddrSpace(MachineInstr &MI,
                                              MachineRegisterInfo &MRI,
                                              MachineIRBuilder &B,
                                              unsigned AddrSpace) const {
  B.setInstr(MI);
  Register ApertureReg = getSegmentAperture(AddrSpace, MRI, B);
  auto Hi32 = B.buildExtract(LLT::scalar(32), MI.getOperand(2).getReg(), 32);
  B.buildICmp(ICmpInst::ICMP_EQ, MI.getOperand(0), Hi32, ApertureReg);
  MI.eraseFromParent();
  return true;
}

// The raw.(t)buffer and struct.(t)buffer intrinsics have two offset args:
// offset (the offset that is included in bounds checking and swizzling, to be
// split between the instruction's voffset and immoffset fields) and soffset
// (the offset that is excluded from bounds checking and swizzling, to go in
// the instruction's soffset field).  This function takes the first kind of
// offset and figures out how to split it between voffset and immoffset.
std::tuple<Register, unsigned, unsigned>
AMDGPULegalizerInfo::splitBufferOffsets(MachineIRBuilder &B,
                                        Register OrigOffset) const {
  const unsigned MaxImm = 4095;
  Register BaseReg;
  unsigned TotalConstOffset;
  MachineInstr *OffsetDef;
  const LLT S32 = LLT::scalar(32);

  std::tie(BaseReg, TotalConstOffset, OffsetDef)
    = AMDGPU::getBaseWithConstantOffset(*B.getMRI(), OrigOffset);

  unsigned ImmOffset = TotalConstOffset;

  // If the immediate value is too big for the immoffset field, put the value
  // and -4096 into the immoffset field so that the value that is copied/added
  // for the voffset field is a multiple of 4096, and it stands more chance
  // of being CSEd with the copy/add for another similar load/store.
  // However, do not do that rounding down to a multiple of 4096 if that is a
  // negative number, as it appears to be illegal to have a negative offset
  // in the vgpr, even if adding the immediate offset makes it positive.
  unsigned Overflow = ImmOffset & ~MaxImm;
  ImmOffset -= Overflow;
  if ((int32_t)Overflow < 0) {
    Overflow += ImmOffset;
    ImmOffset = 0;
  }

  if (Overflow != 0) {
    if (!BaseReg) {
      BaseReg = B.buildConstant(S32, Overflow).getReg(0);
    } else {
      auto OverflowVal = B.buildConstant(S32, Overflow);
      BaseReg = B.buildAdd(S32, BaseReg, OverflowVal).getReg(0);
    }
  }

  if (!BaseReg)
    BaseReg = B.buildConstant(S32, 0).getReg(0);

  return std::make_tuple(BaseReg, ImmOffset, TotalConstOffset);
}

/// Handle register layout difference for f16 images for some subtargets.
Register AMDGPULegalizerInfo::handleD16VData(MachineIRBuilder &B,
                                             MachineRegisterInfo &MRI,
                                             Register Reg) const {
  if (!ST.hasUnpackedD16VMem())
    return Reg;

  const LLT S16 = LLT::scalar(16);
  const LLT S32 = LLT::scalar(32);
  LLT StoreVT = MRI.getType(Reg);
  assert(StoreVT.isVector() && StoreVT.getElementType() == S16);

  auto Unmerge = B.buildUnmerge(S16, Reg);

  SmallVector<Register, 4> WideRegs;
  for (int I = 0, E = Unmerge->getNumOperands() - 1; I != E; ++I)
    WideRegs.push_back(B.buildAnyExt(S32, Unmerge.getReg(I)).getReg(0));

  int NumElts = StoreVT.getNumElements();

  return B.buildBuildVector(LLT::vector(NumElts, S32), WideRegs).getReg(0);
}

Register AMDGPULegalizerInfo::fixStoreSourceType(
  MachineIRBuilder &B, Register VData, bool IsFormat) const {
  MachineRegisterInfo *MRI = B.getMRI();
  LLT Ty = MRI->getType(VData);

  const LLT S16 = LLT::scalar(16);

  // Fixup illegal register types for i8 stores.
  if (Ty == LLT::scalar(8) || Ty == S16) {
    Register AnyExt = B.buildAnyExt(LLT::scalar(32), VData).getReg(0);
    return AnyExt;
  }

  if (Ty.isVector()) {
    if (Ty.getElementType() == S16 && Ty.getNumElements() <= 4) {
      if (IsFormat)
        return handleD16VData(B, *MRI, VData);
    }
  }

  return VData;
}

bool AMDGPULegalizerInfo::legalizeBufferStore(MachineInstr &MI,
                                              MachineRegisterInfo &MRI,
                                              MachineIRBuilder &B,
                                              bool IsTyped,
                                              bool IsFormat) const {
  B.setInstr(MI);

  Register VData = MI.getOperand(1).getReg();
  LLT Ty = MRI.getType(VData);
  LLT EltTy = Ty.getScalarType();
  const bool IsD16 = IsFormat && (EltTy.getSizeInBits() == 16);
  const LLT S32 = LLT::scalar(32);

  VData = fixStoreSourceType(B, VData, IsFormat);
  Register RSrc = MI.getOperand(2).getReg();

  MachineMemOperand *MMO = *MI.memoperands_begin();
  const int MemSize = MMO->getSize();

  unsigned ImmOffset;
  unsigned TotalOffset;

  // The typed intrinsics add an immediate after the registers.
  const unsigned NumVIndexOps = IsTyped ? 8 : 7;

  // The struct intrinsic variants add one additional operand over raw.
  const bool HasVIndex = MI.getNumOperands() == NumVIndexOps;
  Register VIndex;
  int OpOffset = 0;
  if (HasVIndex) {
    VIndex = MI.getOperand(3).getReg();
    OpOffset = 1;
  }

  Register VOffset = MI.getOperand(3 + OpOffset).getReg();
  Register SOffset = MI.getOperand(4 + OpOffset).getReg();

  unsigned Format = 0;
  if (IsTyped) {
    Format = MI.getOperand(5 + OpOffset).getImm();
    ++OpOffset;
  }

  unsigned AuxiliaryData = MI.getOperand(5 + OpOffset).getImm();

  std::tie(VOffset, ImmOffset, TotalOffset) = splitBufferOffsets(B, VOffset);
  if (TotalOffset != 0)
    MMO = B.getMF().getMachineMemOperand(MMO, TotalOffset, MemSize);

  unsigned Opc;
  if (IsTyped) {
    Opc = IsD16 ? AMDGPU::G_AMDGPU_TBUFFER_STORE_FORMAT_D16 :
                  AMDGPU::G_AMDGPU_TBUFFER_STORE_FORMAT;
  } else if (IsFormat) {
    Opc = IsD16 ? AMDGPU::G_AMDGPU_BUFFER_STORE_FORMAT_D16 :
                  AMDGPU::G_AMDGPU_BUFFER_STORE_FORMAT;
  } else {
    switch (MemSize) {
    case 1:
      Opc = AMDGPU::G_AMDGPU_BUFFER_STORE_BYTE;
      break;
    case 2:
      Opc = AMDGPU::G_AMDGPU_BUFFER_STORE_SHORT;
      break;
    default:
      Opc = AMDGPU::G_AMDGPU_BUFFER_STORE;
      break;
    }
  }

  if (!VIndex)
    VIndex = B.buildConstant(S32, 0).getReg(0);

  auto MIB = B.buildInstr(Opc)
    .addUse(VData)              // vdata
    .addUse(RSrc)               // rsrc
    .addUse(VIndex)             // vindex
    .addUse(VOffset)            // voffset
    .addUse(SOffset)            // soffset
    .addImm(ImmOffset);         // offset(imm)

  if (IsTyped)
    MIB.addImm(Format);

  MIB.addImm(AuxiliaryData)      // cachepolicy, swizzled buffer(imm)
     .addImm(HasVIndex ? -1 : 0) // idxen(imm)
     .addMemOperand(MMO);

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeBufferLoad(MachineInstr &MI,
                                             MachineRegisterInfo &MRI,
                                             MachineIRBuilder &B,
                                             bool IsFormat,
                                             bool IsTyped) const {
  B.setInstr(MI);

  // FIXME: Verifier should enforce 1 MMO for these intrinsics.
  MachineMemOperand *MMO = *MI.memoperands_begin();
  const int MemSize = MMO->getSize();
  const LLT S32 = LLT::scalar(32);

  Register Dst = MI.getOperand(0).getReg();
  Register RSrc = MI.getOperand(2).getReg();

  // The typed intrinsics add an immediate after the registers.
  const unsigned NumVIndexOps = IsTyped ? 8 : 7;

  // The struct intrinsic variants add one additional operand over raw.
  const bool HasVIndex = MI.getNumOperands() == NumVIndexOps;
  Register VIndex;
  int OpOffset = 0;
  if (HasVIndex) {
    VIndex = MI.getOperand(3).getReg();
    OpOffset = 1;
  }

  Register VOffset = MI.getOperand(3 + OpOffset).getReg();
  Register SOffset = MI.getOperand(4 + OpOffset).getReg();

  unsigned Format = 0;
  if (IsTyped) {
    Format = MI.getOperand(5 + OpOffset).getImm();
    ++OpOffset;
  }

  unsigned AuxiliaryData = MI.getOperand(5 + OpOffset).getImm();
  unsigned ImmOffset;
  unsigned TotalOffset;

  LLT Ty = MRI.getType(Dst);
  LLT EltTy = Ty.getScalarType();
  const bool IsD16 = IsFormat && (EltTy.getSizeInBits() == 16);
  const bool Unpacked = ST.hasUnpackedD16VMem();

  std::tie(VOffset, ImmOffset, TotalOffset) = splitBufferOffsets(B, VOffset);
  if (TotalOffset != 0)
    MMO = B.getMF().getMachineMemOperand(MMO, TotalOffset, MemSize);

  unsigned Opc;

  if (IsTyped) {
    Opc = IsD16 ? AMDGPU::G_AMDGPU_TBUFFER_LOAD_FORMAT_D16 :
                  AMDGPU::G_AMDGPU_TBUFFER_LOAD_FORMAT;
  } else if (IsFormat) {
    Opc = IsD16 ? AMDGPU::G_AMDGPU_BUFFER_LOAD_FORMAT_D16 :
                  AMDGPU::G_AMDGPU_BUFFER_LOAD_FORMAT;
  } else {
    switch (MemSize) {
    case 1:
      Opc = AMDGPU::G_AMDGPU_BUFFER_LOAD_UBYTE;
      break;
    case 2:
      Opc = AMDGPU::G_AMDGPU_BUFFER_LOAD_USHORT;
      break;
    default:
      Opc = AMDGPU::G_AMDGPU_BUFFER_LOAD;
      break;
    }
  }

  Register LoadDstReg;

  bool IsExtLoad = (!IsD16 && MemSize < 4) || (IsD16 && !Ty.isVector());
  LLT UnpackedTy = Ty.changeElementSize(32);

  if (IsExtLoad)
    LoadDstReg = B.getMRI()->createGenericVirtualRegister(S32);
  else if (Unpacked && IsD16 && Ty.isVector())
    LoadDstReg = B.getMRI()->createGenericVirtualRegister(UnpackedTy);
  else
    LoadDstReg = Dst;

  if (!VIndex)
    VIndex = B.buildConstant(S32, 0).getReg(0);

  auto MIB = B.buildInstr(Opc)
    .addDef(LoadDstReg)         // vdata
    .addUse(RSrc)               // rsrc
    .addUse(VIndex)             // vindex
    .addUse(VOffset)            // voffset
    .addUse(SOffset)            // soffset
    .addImm(ImmOffset);         // offset(imm)

  if (IsTyped)
    MIB.addImm(Format);

  MIB.addImm(AuxiliaryData)      // cachepolicy, swizzled buffer(imm)
     .addImm(HasVIndex ? -1 : 0) // idxen(imm)
     .addMemOperand(MMO);

  if (LoadDstReg != Dst) {
    B.setInsertPt(B.getMBB(), ++B.getInsertPt());

    // Widen result for extending loads was widened.
    if (IsExtLoad)
      B.buildTrunc(Dst, LoadDstReg);
    else {
      // Repack to original 16-bit vector result
      // FIXME: G_TRUNC should work, but legalization currently fails
      auto Unmerge = B.buildUnmerge(S32, LoadDstReg);
      SmallVector<Register, 4> Repack;
      for (unsigned I = 0, N = Unmerge->getNumOperands() - 1; I != N; ++I)
        Repack.push_back(B.buildTrunc(EltTy, Unmerge.getReg(I)).getReg(0));
      B.buildMerge(Dst, Repack);
    }
  }

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeAtomicIncDec(MachineInstr &MI,
                                               MachineIRBuilder &B,
                                               bool IsInc) const {
  B.setInstr(MI);
  unsigned Opc = IsInc ? AMDGPU::G_AMDGPU_ATOMIC_INC :
                         AMDGPU::G_AMDGPU_ATOMIC_DEC;
  B.buildInstr(Opc)
    .addDef(MI.getOperand(0).getReg())
    .addUse(MI.getOperand(2).getReg())
    .addUse(MI.getOperand(3).getReg())
    .cloneMemRefs(MI);
  MI.eraseFromParent();
  return true;
}

static unsigned getBufferAtomicPseudo(Intrinsic::ID IntrID) {
  switch (IntrID) {
  case Intrinsic::amdgcn_raw_buffer_atomic_swap:
  case Intrinsic::amdgcn_struct_buffer_atomic_swap:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SWAP;
  case Intrinsic::amdgcn_raw_buffer_atomic_add:
  case Intrinsic::amdgcn_struct_buffer_atomic_add:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_ADD;
  case Intrinsic::amdgcn_raw_buffer_atomic_sub:
  case Intrinsic::amdgcn_struct_buffer_atomic_sub:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SUB;
  case Intrinsic::amdgcn_raw_buffer_atomic_smin:
  case Intrinsic::amdgcn_struct_buffer_atomic_smin:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SMIN;
  case Intrinsic::amdgcn_raw_buffer_atomic_umin:
  case Intrinsic::amdgcn_struct_buffer_atomic_umin:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_UMIN;
  case Intrinsic::amdgcn_raw_buffer_atomic_smax:
  case Intrinsic::amdgcn_struct_buffer_atomic_smax:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_SMAX;
  case Intrinsic::amdgcn_raw_buffer_atomic_umax:
  case Intrinsic::amdgcn_struct_buffer_atomic_umax:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_UMAX;
  case Intrinsic::amdgcn_raw_buffer_atomic_and:
  case Intrinsic::amdgcn_struct_buffer_atomic_and:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_AND;
  case Intrinsic::amdgcn_raw_buffer_atomic_or:
  case Intrinsic::amdgcn_struct_buffer_atomic_or:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_OR;
  case Intrinsic::amdgcn_raw_buffer_atomic_xor:
  case Intrinsic::amdgcn_struct_buffer_atomic_xor:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_XOR;
  case Intrinsic::amdgcn_raw_buffer_atomic_inc:
  case Intrinsic::amdgcn_struct_buffer_atomic_inc:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_INC;
  case Intrinsic::amdgcn_raw_buffer_atomic_dec:
  case Intrinsic::amdgcn_struct_buffer_atomic_dec:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_DEC;
  case Intrinsic::amdgcn_raw_buffer_atomic_cmpswap:
  case Intrinsic::amdgcn_struct_buffer_atomic_cmpswap:
    return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_CMPSWAP;
  default:
    llvm_unreachable("unhandled atomic opcode");
  }
}

bool AMDGPULegalizerInfo::legalizeBufferAtomic(MachineInstr &MI,
                                               MachineIRBuilder &B,
                                               Intrinsic::ID IID) const {
  B.setInstr(MI);

  const bool IsCmpSwap = IID == Intrinsic::amdgcn_raw_buffer_atomic_cmpswap ||
                         IID == Intrinsic::amdgcn_struct_buffer_atomic_cmpswap;

  Register Dst = MI.getOperand(0).getReg();
  Register VData = MI.getOperand(2).getReg();

  Register CmpVal;
  int OpOffset = 0;

  if (IsCmpSwap) {
    CmpVal = MI.getOperand(3 + OpOffset).getReg();
    ++OpOffset;
  }

  Register RSrc = MI.getOperand(3 + OpOffset).getReg();
  const unsigned NumVIndexOps = IsCmpSwap ? 9 : 8;

  // The struct intrinsic variants add one additional operand over raw.
  const bool HasVIndex = MI.getNumOperands() == NumVIndexOps;
  Register VIndex;
  if (HasVIndex) {
    VIndex = MI.getOperand(4 + OpOffset).getReg();
    ++OpOffset;
  }

  Register VOffset = MI.getOperand(4 + OpOffset).getReg();
  Register SOffset = MI.getOperand(5 + OpOffset).getReg();
  unsigned AuxiliaryData = MI.getOperand(6 + OpOffset).getImm();

  MachineMemOperand *MMO = *MI.memoperands_begin();

  unsigned ImmOffset;
  unsigned TotalOffset;
  std::tie(VOffset, ImmOffset, TotalOffset) = splitBufferOffsets(B, VOffset);
  if (TotalOffset != 0)
    MMO = B.getMF().getMachineMemOperand(MMO, TotalOffset, MMO->getSize());

  if (!VIndex)
    VIndex = B.buildConstant(LLT::scalar(32), 0).getReg(0);

  auto MIB = B.buildInstr(getBufferAtomicPseudo(IID))
    .addDef(Dst)
    .addUse(VData); // vdata

  if (IsCmpSwap)
    MIB.addReg(CmpVal);

  MIB.addUse(RSrc)               // rsrc
     .addUse(VIndex)             // vindex
     .addUse(VOffset)            // voffset
     .addUse(SOffset)            // soffset
     .addImm(ImmOffset)          // offset(imm)
     .addImm(AuxiliaryData)      // cachepolicy, swizzled buffer(imm)
     .addImm(HasVIndex ? -1 : 0) // idxen(imm)
     .addMemOperand(MMO);

  MI.eraseFromParent();
  return true;
}

/// Turn a set of s16 typed registers in \p A16AddrRegs into a dword sized
/// vector with s16 typed elements.
static void packImageA16AddressToDwords(MachineIRBuilder &B, MachineInstr &MI,
                                        SmallVectorImpl<Register> &PackedAddrs,
                                        int AddrIdx, int DimIdx, int NumVAddrs,
                                        int NumGradients) {
  const LLT S16 = LLT::scalar(16);
  const LLT V2S16 = LLT::vector(2, 16);

  for (int I = AddrIdx; I < AddrIdx + NumVAddrs; ++I) {
    MachineOperand &SrcOp = MI.getOperand(I);
    if (!SrcOp.isReg())
      continue; // _L to _LZ may have eliminated this.

    Register AddrReg = SrcOp.getReg();

    if (I < DimIdx) {
      AddrReg = B.buildBitcast(V2S16, AddrReg).getReg(0);
      PackedAddrs.push_back(AddrReg);
    } else {
      // Dz/dh, dz/dv and the last odd coord are packed with undef. Also, in 1D,
      // derivatives dx/dh and dx/dv are packed with undef.
      if (((I + 1) >= (AddrIdx + NumVAddrs)) ||
          ((NumGradients / 2) % 2 == 1 &&
           (I == DimIdx + (NumGradients / 2) - 1 ||
            I == DimIdx + NumGradients - 1)) ||
          // Check for _L to _LZ optimization
          !MI.getOperand(I + 1).isReg()) {
        PackedAddrs.push_back(
            B.buildBuildVector(V2S16, {AddrReg, B.buildUndef(S16).getReg(0)})
                .getReg(0));
      } else {
        PackedAddrs.push_back(
            B.buildBuildVector(V2S16, {AddrReg, MI.getOperand(I + 1).getReg()})
                .getReg(0));
        ++I;
      }
    }
  }
}

/// Convert from separate vaddr components to a single vector address register,
/// and replace the remaining operands with $noreg.
static void convertImageAddrToPacked(MachineIRBuilder &B, MachineInstr &MI,
                                     int DimIdx, int NumVAddrs) {
  const LLT S32 = LLT::scalar(32);

  SmallVector<Register, 8> AddrRegs;
  for (int I = 0; I != NumVAddrs; ++I) {
    MachineOperand &SrcOp = MI.getOperand(DimIdx + I);
    if (SrcOp.isReg()) {
      AddrRegs.push_back(SrcOp.getReg());
      assert(B.getMRI()->getType(SrcOp.getReg()) == S32);
    }
  }

  int NumAddrRegs = AddrRegs.size();
  if (NumAddrRegs != 1) {
    // Round up to 8 elements for v5-v7
    // FIXME: Missing intermediate sized register classes and instructions.
    if (NumAddrRegs > 4 && !isPowerOf2_32(NumAddrRegs)) {
      const int RoundedNumRegs = NextPowerOf2(NumAddrRegs);
      auto Undef = B.buildUndef(S32);
      AddrRegs.append(RoundedNumRegs - NumAddrRegs, Undef.getReg(0));
      NumAddrRegs = RoundedNumRegs;
    }

    auto VAddr = B.buildBuildVector(LLT::vector(NumAddrRegs, 32), AddrRegs);
    MI.getOperand(DimIdx).setReg(VAddr.getReg(0));
  }

  for (int I = 1; I != NumVAddrs; ++I) {
    MachineOperand &SrcOp = MI.getOperand(DimIdx + I);
    if (SrcOp.isReg())
      MI.getOperand(DimIdx + I).setReg(AMDGPU::NoRegister);
  }
}

/// Rewrite image intrinsics to use register layouts expected by the subtarget.
///
/// Depending on the subtarget, load/store with 16-bit element data need to be
/// rewritten to use the low half of 32-bit registers, or directly use a packed
/// layout. 16-bit addresses should also sometimes be packed into 32-bit
/// registers.
///
/// We don't want to directly select image instructions just yet, but also want
/// to exposes all register repacking to the legalizer/combiners. We also don't
/// want a selected instrution entering RegBankSelect. In order to avoid
/// defining a multitude of intermediate image instructions, directly hack on
/// the intrinsic's arguments. In cases like a16 addreses, this requires padding
/// now unnecessary arguments with $noreg.
bool AMDGPULegalizerInfo::legalizeImageIntrinsic(
    MachineInstr &MI, MachineIRBuilder &B,
    GISelChangeObserver &Observer,
    const AMDGPU::ImageDimIntrinsicInfo *ImageDimIntr) const {
  B.setInstr(MI);

  const int NumDefs = MI.getNumExplicitDefs();
  bool IsTFE = NumDefs == 2;
  // We are only processing the operands of d16 image operations on subtargets
  // that use the unpacked register layout, or need to repack the TFE result.

  // TODO: Do we need to guard against already legalized intrinsics?
  const AMDGPU::MIMGBaseOpcodeInfo *BaseOpcode =
    AMDGPU::getMIMGBaseOpcodeInfo(ImageDimIntr->BaseOpcode);

  MachineRegisterInfo *MRI = B.getMRI();
  const LLT S32 = LLT::scalar(32);
  const LLT S16 = LLT::scalar(16);
  const LLT V2S16 = LLT::vector(2, 16);

  // Index of first address argument
  const int AddrIdx = getImageVAddrIdxBegin(BaseOpcode, NumDefs);

  // Check for 16 bit addresses and pack if true.
  int DimIdx = AddrIdx + BaseOpcode->NumExtraArgs;
  LLT AddrTy = MRI->getType(MI.getOperand(DimIdx).getReg());
  const bool IsA16 = AddrTy == S16;

  int NumVAddrs, NumGradients;
  std::tie(NumVAddrs, NumGradients) = getImageNumVAddr(ImageDimIntr, BaseOpcode);
  const int DMaskIdx = BaseOpcode->Atomic ? -1 :
    getDMaskIdx(BaseOpcode, NumDefs);
  unsigned DMask = 0;

  int DMaskLanes = 0;
  if (!BaseOpcode->Atomic) {
    DMask = MI.getOperand(DMaskIdx).getImm();
    if (BaseOpcode->Gather4) {
      DMaskLanes = 4;
    } else if (DMask != 0) {
      DMaskLanes = countPopulation(DMask);
    } else if (!IsTFE && !BaseOpcode->Store) {
      // If dmask is 0, this is a no-op load. This can be eliminated.
      B.buildUndef(MI.getOperand(0));
      MI.eraseFromParent();
      return true;
    }
  }

  Observer.changingInstr(MI);
  auto ChangedInstr = make_scope_exit([&] { Observer.changedInstr(MI); });

  unsigned NewOpcode = NumDefs == 0 ?
    AMDGPU::G_AMDGPU_INTRIN_IMAGE_STORE : AMDGPU::G_AMDGPU_INTRIN_IMAGE_LOAD;

  // Track that we legalized this
  MI.setDesc(B.getTII().get(NewOpcode));

  // Expecting to get an error flag since TFC is on - and dmask is 0 Force
  // dmask to be at least 1 otherwise the instruction will fail
  if (IsTFE && DMask == 0) {
    DMask = 0x1;
    DMaskLanes = 1;
    MI.getOperand(DMaskIdx).setImm(DMask);
  }

  if (BaseOpcode->Atomic) {
    Register VData0 = MI.getOperand(2).getReg();
    LLT Ty = MRI->getType(VData0);

    // TODO: Allow atomic swap and bit ops for v2s16/v4s16
    if (Ty.isVector())
      return false;

    if (BaseOpcode->AtomicX2) {
      Register VData1 = MI.getOperand(3).getReg();
      // The two values are packed in one register.
      LLT PackedTy = LLT::vector(2, Ty);
      auto Concat = B.buildBuildVector(PackedTy, {VData0, VData1});
      MI.getOperand(2).setReg(Concat.getReg(0));
      MI.getOperand(3).setReg(AMDGPU::NoRegister);
    }
  }

  int CorrectedNumVAddrs = NumVAddrs;

  // Optimize _L to _LZ when _L is zero
  if (const AMDGPU::MIMGLZMappingInfo *LZMappingInfo =
        AMDGPU::getMIMGLZMappingInfo(ImageDimIntr->BaseOpcode)) {
    const ConstantFP *ConstantLod;
    const int LodIdx = AddrIdx + NumVAddrs - 1;

    if (mi_match(MI.getOperand(LodIdx).getReg(), *MRI, m_GFCst(ConstantLod))) {
      if (ConstantLod->isZero() || ConstantLod->isNegative()) {
        // Set new opcode to _lz variant of _l, and change the intrinsic ID.
        ImageDimIntr = AMDGPU::getImageDimInstrinsicByBaseOpcode(
          LZMappingInfo->LZ, ImageDimIntr->Dim);

        // The starting indexes should remain in the same place.
        --NumVAddrs;
        --CorrectedNumVAddrs;

        MI.getOperand(MI.getNumExplicitDefs()).setIntrinsicID(
          static_cast<Intrinsic::ID>(ImageDimIntr->Intr));
        MI.RemoveOperand(LodIdx);
      }
    }
  }

  // Optimize _mip away, when 'lod' is zero
  if (AMDGPU::getMIMGMIPMappingInfo(ImageDimIntr->BaseOpcode)) {
    int64_t ConstantLod;
    const int LodIdx = AddrIdx + NumVAddrs - 1;

    if (mi_match(MI.getOperand(LodIdx).getReg(), *MRI, m_ICst(ConstantLod))) {
      if (ConstantLod == 0) {
        // TODO: Change intrinsic opcode and remove operand instead or replacing
        // it with 0, as the _L to _LZ handling is done above.
        MI.getOperand(LodIdx).ChangeToImmediate(0);
        --CorrectedNumVAddrs;
      }
    }
  }

  // If the register allocator cannot place the address registers contiguously
  // without introducing moves, then using the non-sequential address encoding
  // is always preferable, since it saves VALU instructions and is usually a
  // wash in terms of code size or even better.
  //
  // However, we currently have no way of hinting to the register allocator
  // that MIMG addresses should be placed contiguously when it is possible to
  // do so, so force non-NSA for the common 2-address case as a heuristic.
  //
  // SIShrinkInstructions will convert NSA encodings to non-NSA after register
  // allocation when possible.
  const bool UseNSA = CorrectedNumVAddrs >= 3 && ST.hasNSAEncoding();

  // Rewrite the addressing register layout before doing anything else.
  if (IsA16) {
    // FIXME: this feature is missing from gfx10. When that is fixed, this check
    // should be introduced.
    if (!ST.hasR128A16() && !ST.hasGFX10A16())
      return false;

    if (NumVAddrs > 1) {
      SmallVector<Register, 4> PackedRegs;
      packImageA16AddressToDwords(B, MI, PackedRegs, AddrIdx, DimIdx, NumVAddrs,
                                  NumGradients);

      if (!UseNSA && PackedRegs.size() > 1) {
        LLT PackedAddrTy = LLT::vector(2 * PackedRegs.size(), 16);
        auto Concat = B.buildConcatVectors(PackedAddrTy, PackedRegs);
        PackedRegs[0] = Concat.getReg(0);
        PackedRegs.resize(1);
      }

      const int NumPacked = PackedRegs.size();
      for (int I = 0; I != NumVAddrs; ++I) {
        MachineOperand &SrcOp = MI.getOperand(AddrIdx + I);
        if (!SrcOp.isReg()) {
          assert(SrcOp.isImm() && SrcOp.getImm() == 0);
          continue;
        }

        assert(SrcOp.getReg() != AMDGPU::NoRegister);

        if (I < NumPacked)
          SrcOp.setReg(PackedRegs[I]);
        else
          SrcOp.setReg(AMDGPU::NoRegister);
      }
    }
  } else if (!UseNSA && NumVAddrs > 1) {
    convertImageAddrToPacked(B, MI, AddrIdx, NumVAddrs);
  }


  if (BaseOpcode->Store) { // No TFE for stores?
    // TODO: Handle dmask trim
    Register VData = MI.getOperand(1).getReg();
    LLT Ty = MRI->getType(VData);
    if (!Ty.isVector() || Ty.getElementType() != S16)
      return true;

    B.setInstr(MI);

    Register RepackedReg = handleD16VData(B, *MRI, VData);
    if (RepackedReg != VData) {
      MI.getOperand(1).setReg(RepackedReg);
    }

    return true;
  }

  Register DstReg = MI.getOperand(0).getReg();
  LLT Ty = MRI->getType(DstReg);
  const LLT EltTy = Ty.getScalarType();
  const bool IsD16 = Ty.getScalarType() == S16;
  const int NumElts = Ty.isVector() ? Ty.getNumElements() : 1;

  // Confirm that the return type is large enough for the dmask specified
  if (NumElts < DMaskLanes)
    return false;

  if (NumElts > 4 || DMaskLanes > 4)
    return false;

  const unsigned AdjustedNumElts = DMaskLanes == 0 ? 1 : DMaskLanes;
  const LLT AdjustedTy = Ty.changeNumElements(AdjustedNumElts);

  // The raw dword aligned data component of the load. The only legal cases
  // where this matters should be when using the packed D16 format, for
  // s16 -> <2 x s16>, and <3 x s16> -> <4 x s16>,
  LLT RoundedTy;

  // S32 vector to to cover all data, plus TFE result element.
  LLT TFETy;

  // Register type to use for each loaded component. Will be S32 or V2S16.
  LLT RegTy;

  if (IsD16 && ST.hasUnpackedD16VMem()) {
    RoundedTy = LLT::scalarOrVector(AdjustedNumElts, 32);
    TFETy = LLT::vector(AdjustedNumElts + 1, 32);
    RegTy = S32;
  } else {
    unsigned EltSize = EltTy.getSizeInBits();
    unsigned RoundedElts = (AdjustedTy.getSizeInBits() + 31) / 32;
    unsigned RoundedSize = 32 * RoundedElts;
    RoundedTy = LLT::scalarOrVector(RoundedSize / EltSize, EltSize);
    TFETy = LLT::vector(RoundedSize / 32 + 1, S32);
    RegTy = !IsTFE && EltSize == 16 ? V2S16 : S32;
  }

  // The return type does not need adjustment.
  // TODO: Should we change s16 case to s32 or <2 x s16>?
  if (!IsTFE && (RoundedTy == Ty || !Ty.isVector()))
    return true;

  Register Dst1Reg;

  // Insert after the instruction.
  B.setInsertPt(*MI.getParent(), ++MI.getIterator());

  // TODO: For TFE with d16, if we used a TFE type that was a multiple of <2 x
  // s16> instead of s32, we would only need 1 bitcast instead of multiple.
  const LLT LoadResultTy = IsTFE ? TFETy : RoundedTy;
  const int ResultNumRegs = LoadResultTy.getSizeInBits() / 32;

  Register NewResultReg = MRI->createGenericVirtualRegister(LoadResultTy);

  MI.getOperand(0).setReg(NewResultReg);

  // In the IR, TFE is supposed to be used with a 2 element struct return
  // type. The intruction really returns these two values in one contiguous
  // register, with one additional dword beyond the loaded data. Rewrite the
  // return type to use a single register result.

  if (IsTFE) {
    Dst1Reg = MI.getOperand(1).getReg();
    if (MRI->getType(Dst1Reg) != S32)
      return false;

    // TODO: Make sure the TFE operand bit is set.
    MI.RemoveOperand(1);

    // Handle the easy case that requires no repack instructions.
    if (Ty == S32) {
      B.buildUnmerge({DstReg, Dst1Reg}, NewResultReg);
      return true;
    }
  }

  // Now figure out how to copy the new result register back into the old
  // result.
  SmallVector<Register, 5> ResultRegs(ResultNumRegs, Dst1Reg);

  const int NumDataRegs = IsTFE ? ResultNumRegs - 1  : ResultNumRegs;

  if (ResultNumRegs == 1) {
    assert(!IsTFE);
    ResultRegs[0] = NewResultReg;
  } else {
    // We have to repack into a new vector of some kind.
    for (int I = 0; I != NumDataRegs; ++I)
      ResultRegs[I] = MRI->createGenericVirtualRegister(RegTy);
    B.buildUnmerge(ResultRegs, NewResultReg);

    // Drop the final TFE element to get the data part. The TFE result is
    // directly written to the right place already.
    if (IsTFE)
      ResultRegs.resize(NumDataRegs);
  }

  // For an s16 scalar result, we form an s32 result with a truncate regardless
  // of packed vs. unpacked.
  if (IsD16 && !Ty.isVector()) {
    B.buildTrunc(DstReg, ResultRegs[0]);
    return true;
  }

  // Avoid a build/concat_vector of 1 entry.
  if (Ty == V2S16 && NumDataRegs == 1 && !ST.hasUnpackedD16VMem()) {
    B.buildBitcast(DstReg, ResultRegs[0]);
    return true;
  }

  assert(Ty.isVector());

  if (IsD16) {
    // For packed D16 results with TFE enabled, all the data components are
    // S32. Cast back to the expected type.
    //
    // TODO: We don't really need to use load s32 elements. We would only need one
    // cast for the TFE result if a multiple of v2s16 was used.
    if (RegTy != V2S16 && !ST.hasUnpackedD16VMem()) {
      for (Register &Reg : ResultRegs)
        Reg = B.buildBitcast(V2S16, Reg).getReg(0);
    } else if (ST.hasUnpackedD16VMem()) {
      for (Register &Reg : ResultRegs)
        Reg = B.buildTrunc(S16, Reg).getReg(0);
    }
  }

  auto padWithUndef = [&](LLT Ty, int NumElts) {
    if (NumElts == 0)
      return;
    Register Undef = B.buildUndef(Ty).getReg(0);
    for (int I = 0; I != NumElts; ++I)
      ResultRegs.push_back(Undef);
  };

  // Pad out any elements eliminated due to the dmask.
  LLT ResTy = MRI->getType(ResultRegs[0]);
  if (!ResTy.isVector()) {
    padWithUndef(ResTy, NumElts - ResultRegs.size());
    B.buildBuildVector(DstReg, ResultRegs);
    return true;
  }

  assert(!ST.hasUnpackedD16VMem() && ResTy == V2S16);
  const int RegsToCover = (Ty.getSizeInBits() + 31) / 32;

  // Deal with the one annoying legal case.
  const LLT V3S16 = LLT::vector(3, 16);
  if (Ty == V3S16) {
    padWithUndef(ResTy, RegsToCover - ResultRegs.size() + 1);
    auto Concat = B.buildConcatVectors(LLT::vector(6, 16), ResultRegs);
    B.buildUnmerge({DstReg, MRI->createGenericVirtualRegister(V3S16)}, Concat);
    return true;
  }

  padWithUndef(ResTy, RegsToCover - ResultRegs.size());
  B.buildConcatVectors(DstReg, ResultRegs);
  return true;
}

bool AMDGPULegalizerInfo::legalizeSBufferLoad(
  MachineInstr &MI, MachineIRBuilder &B,
  GISelChangeObserver &Observer) const {
  Register Dst = MI.getOperand(0).getReg();
  LLT Ty = B.getMRI()->getType(Dst);
  unsigned Size = Ty.getSizeInBits();
  MachineFunction &MF = B.getMF();

  Observer.changingInstr(MI);

  // FIXME: We don't really need this intermediate instruction. The intrinsic
  // should be fixed to have a memory operand. Since it's readnone, we're not
  // allowed to add one.
  MI.setDesc(B.getTII().get(AMDGPU::G_AMDGPU_S_BUFFER_LOAD));
  MI.RemoveOperand(1); // Remove intrinsic ID

  // FIXME: When intrinsic definition is fixed, this should have an MMO already.
  // TODO: Should this use datalayout alignment?
  const unsigned MemSize = (Size + 7) / 8;
  const Align MemAlign(4);
  MachineMemOperand *MMO = MF.getMachineMemOperand(
      MachinePointerInfo(),
      MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
          MachineMemOperand::MOInvariant,
      MemSize, MemAlign);
  MI.addMemOperand(MF, MMO);

  // There are no 96-bit result scalar loads, but widening to 128-bit should
  // always be legal. We may need to restore this to a 96-bit result if it turns
  // out this needs to be converted to a vector load during RegBankSelect.
  if (!isPowerOf2_32(Size)) {
    LegalizerHelper Helper(MF, *this, Observer, B);
    B.setInstr(MI);

    if (Ty.isVector())
      Helper.moreElementsVectorDst(MI, getPow2VectorType(Ty), 0);
    else
      Helper.widenScalarDst(MI, getPow2ScalarType(Ty), 0);
  }

  Observer.changedInstr(MI);
  return true;
}

bool AMDGPULegalizerInfo::legalizeTrapIntrinsic(MachineInstr &MI,
                                                MachineRegisterInfo &MRI,
                                                MachineIRBuilder &B) const {
  B.setInstr(MI);

  // Is non-HSA path or trap-handler disabled? then, insert s_endpgm instruction
  if (ST.getTrapHandlerAbi() != GCNSubtarget::TrapHandlerAbiHsa ||
      !ST.isTrapHandlerEnabled()) {
    B.buildInstr(AMDGPU::S_ENDPGM).addImm(0);
  } else {
    // Pass queue pointer to trap handler as input, and insert trap instruction
    // Reference: https://llvm.org/docs/AMDGPUUsage.html#trap-handler-abi
    const ArgDescriptor *Arg =
        getArgDescriptor(B, AMDGPUFunctionArgInfo::QUEUE_PTR);
    if (!Arg)
      return false;
    MachineRegisterInfo &MRI = *B.getMRI();
    Register SGPR01(AMDGPU::SGPR0_SGPR1);
    Register LiveIn = getLiveInRegister(
        B, MRI, SGPR01, LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64),
        /*InsertLiveInCopy=*/false);
    if (!loadInputValue(LiveIn, B, Arg))
      return false;
    B.buildCopy(SGPR01, LiveIn);
    B.buildInstr(AMDGPU::S_TRAP)
        .addImm(GCNSubtarget::TrapIDLLVMTrap)
        .addReg(SGPR01, RegState::Implicit);
  }

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeDebugTrapIntrinsic(
    MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B) const {
  B.setInstr(MI);

  // Is non-HSA path or trap-handler disabled? then, report a warning
  // accordingly
  if (ST.getTrapHandlerAbi() != GCNSubtarget::TrapHandlerAbiHsa ||
      !ST.isTrapHandlerEnabled()) {
    DiagnosticInfoUnsupported NoTrap(B.getMF().getFunction(),
                                     "debugtrap handler not supported",
                                     MI.getDebugLoc(), DS_Warning);
    LLVMContext &Ctx = B.getMF().getFunction().getContext();
    Ctx.diagnose(NoTrap);
  } else {
    // Insert debug-trap instruction
    B.buildInstr(AMDGPU::S_TRAP).addImm(GCNSubtarget::TrapIDLLVMDebugTrap);
  }

  MI.eraseFromParent();
  return true;
}

bool AMDGPULegalizerInfo::legalizeIntrinsic(MachineInstr &MI,
                                            MachineIRBuilder &B,
                                            GISelChangeObserver &Observer) const {
  MachineRegisterInfo &MRI = *B.getMRI();

  // Replace the use G_BRCOND with the exec manipulate and branch pseudos.
  auto IntrID = MI.getIntrinsicID();
  switch (IntrID) {
  case Intrinsic::amdgcn_if:
  case Intrinsic::amdgcn_else: {
    MachineInstr *Br = nullptr;
    MachineBasicBlock *UncondBrTarget = nullptr;
    if (MachineInstr *BrCond = verifyCFIntrinsic(MI, MRI, Br, UncondBrTarget)) {
      const SIRegisterInfo *TRI
        = static_cast<const SIRegisterInfo *>(MRI.getTargetRegisterInfo());

      B.setInstr(*BrCond);
      Register Def = MI.getOperand(1).getReg();
      Register Use = MI.getOperand(3).getReg();

      MachineBasicBlock *CondBrTarget = BrCond->getOperand(1).getMBB();
      if (IntrID == Intrinsic::amdgcn_if) {
        B.buildInstr(AMDGPU::SI_IF)
          .addDef(Def)
          .addUse(Use)
          .addMBB(UncondBrTarget);
      } else {
        B.buildInstr(AMDGPU::SI_ELSE)
          .addDef(Def)
          .addUse(Use)
          .addMBB(UncondBrTarget)
          .addImm(0);
      }

      if (Br) {
        Br->getOperand(0).setMBB(CondBrTarget);
      } else {
        // The IRTranslator skips inserting the G_BR for fallthrough cases, but
        // since we're swapping branch targets it needs to be reinserted.
        // FIXME: IRTranslator should probably not do this
        B.buildBr(*CondBrTarget);
      }

      MRI.setRegClass(Def, TRI->getWaveMaskRegClass());
      MRI.setRegClass(Use, TRI->getWaveMaskRegClass());
      MI.eraseFromParent();
      BrCond->eraseFromParent();
      return true;
    }

    return false;
  }
  case Intrinsic::amdgcn_loop: {
    MachineInstr *Br = nullptr;
    MachineBasicBlock *UncondBrTarget = nullptr;
    if (MachineInstr *BrCond = verifyCFIntrinsic(MI, MRI, Br, UncondBrTarget)) {
      const SIRegisterInfo *TRI
        = static_cast<const SIRegisterInfo *>(MRI.getTargetRegisterInfo());

      B.setInstr(*BrCond);

      MachineBasicBlock *CondBrTarget = BrCond->getOperand(1).getMBB();
      Register Reg = MI.getOperand(2).getReg();
      B.buildInstr(AMDGPU::SI_LOOP)
        .addUse(Reg)
        .addMBB(UncondBrTarget);

      if (Br)
        Br->getOperand(0).setMBB(CondBrTarget);
      else
        B.buildBr(*CondBrTarget);

      MI.eraseFromParent();
      BrCond->eraseFromParent();
      MRI.setRegClass(Reg, TRI->getWaveMaskRegClass());
      return true;
    }

    return false;
  }
  case Intrinsic::amdgcn_kernarg_segment_ptr:
    if (!AMDGPU::isKernel(B.getMF().getFunction().getCallingConv())) {
      B.setInstr(MI);
      // This only makes sense to call in a kernel, so just lower to null.
      B.buildConstant(MI.getOperand(0).getReg(), 0);
      MI.eraseFromParent();
      return true;
    }

    return legalizePreloadedArgIntrin(
      MI, MRI, B, AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR);
  case Intrinsic::amdgcn_implicitarg_ptr:
    return legalizeImplicitArgPtr(MI, MRI, B);
  case Intrinsic::amdgcn_workitem_id_x:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::WORKITEM_ID_X);
  case Intrinsic::amdgcn_workitem_id_y:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::WORKITEM_ID_Y);
  case Intrinsic::amdgcn_workitem_id_z:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::WORKITEM_ID_Z);
  case Intrinsic::amdgcn_workgroup_id_x:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::WORKGROUP_ID_X);
  case Intrinsic::amdgcn_workgroup_id_y:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::WORKGROUP_ID_Y);
  case Intrinsic::amdgcn_workgroup_id_z:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::WORKGROUP_ID_Z);
  case Intrinsic::amdgcn_dispatch_ptr:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::DISPATCH_PTR);
  case Intrinsic::amdgcn_queue_ptr:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::QUEUE_PTR);
  case Intrinsic::amdgcn_implicit_buffer_ptr:
    return legalizePreloadedArgIntrin(
      MI, MRI, B, AMDGPUFunctionArgInfo::IMPLICIT_BUFFER_PTR);
  case Intrinsic::amdgcn_dispatch_id:
    return legalizePreloadedArgIntrin(MI, MRI, B,
                                      AMDGPUFunctionArgInfo::DISPATCH_ID);
  case Intrinsic::amdgcn_fdiv_fast:
    return legalizeFDIVFastIntrin(MI, MRI, B);
  case Intrinsic::amdgcn_is_shared:
    return legalizeIsAddrSpace(MI, MRI, B, AMDGPUAS::LOCAL_ADDRESS);
  case Intrinsic::amdgcn_is_private:
    return legalizeIsAddrSpace(MI, MRI, B, AMDGPUAS::PRIVATE_ADDRESS);
  case Intrinsic::amdgcn_wavefrontsize: {
    B.setInstr(MI);
    B.buildConstant(MI.getOperand(0), ST.getWavefrontSize());
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::amdgcn_s_buffer_load:
    return legalizeSBufferLoad(MI, B, Observer);
  case Intrinsic::amdgcn_raw_buffer_store:
  case Intrinsic::amdgcn_struct_buffer_store:
    return legalizeBufferStore(MI, MRI, B, false, false);
  case Intrinsic::amdgcn_raw_buffer_store_format:
  case Intrinsic::amdgcn_struct_buffer_store_format:
    return legalizeBufferStore(MI, MRI, B, false, true);
  case Intrinsic::amdgcn_raw_tbuffer_store:
  case Intrinsic::amdgcn_struct_tbuffer_store:
    return legalizeBufferStore(MI, MRI, B, true, true);
  case Intrinsic::amdgcn_raw_buffer_load:
  case Intrinsic::amdgcn_struct_buffer_load:
    return legalizeBufferLoad(MI, MRI, B, false, false);
  case Intrinsic::amdgcn_raw_buffer_load_format:
  case Intrinsic::amdgcn_struct_buffer_load_format:
    return legalizeBufferLoad(MI, MRI, B, true, false);
  case Intrinsic::amdgcn_raw_tbuffer_load:
  case Intrinsic::amdgcn_struct_tbuffer_load:
    return legalizeBufferLoad(MI, MRI, B, true, true);
  case Intrinsic::amdgcn_raw_buffer_atomic_swap:
  case Intrinsic::amdgcn_struct_buffer_atomic_swap:
  case Intrinsic::amdgcn_raw_buffer_atomic_add:
  case Intrinsic::amdgcn_struct_buffer_atomic_add:
  case Intrinsic::amdgcn_raw_buffer_atomic_sub:
  case Intrinsic::amdgcn_struct_buffer_atomic_sub:
  case Intrinsic::amdgcn_raw_buffer_atomic_smin:
  case Intrinsic::amdgcn_struct_buffer_atomic_smin:
  case Intrinsic::amdgcn_raw_buffer_atomic_umin:
  case Intrinsic::amdgcn_struct_buffer_atomic_umin:
  case Intrinsic::amdgcn_raw_buffer_atomic_smax:
  case Intrinsic::amdgcn_struct_buffer_atomic_smax:
  case Intrinsic::amdgcn_raw_buffer_atomic_umax:
  case Intrinsic::amdgcn_struct_buffer_atomic_umax:
  case Intrinsic::amdgcn_raw_buffer_atomic_and:
  case Intrinsic::amdgcn_struct_buffer_atomic_and:
  case Intrinsic::amdgcn_raw_buffer_atomic_or:
  case Intrinsic::amdgcn_struct_buffer_atomic_or:
  case Intrinsic::amdgcn_raw_buffer_atomic_xor:
  case Intrinsic::amdgcn_struct_buffer_atomic_xor:
  case Intrinsic::amdgcn_raw_buffer_atomic_inc:
  case Intrinsic::amdgcn_struct_buffer_atomic_inc:
  case Intrinsic::amdgcn_raw_buffer_atomic_dec:
  case Intrinsic::amdgcn_struct_buffer_atomic_dec:
  case Intrinsic::amdgcn_raw_buffer_atomic_cmpswap:
  case Intrinsic::amdgcn_struct_buffer_atomic_cmpswap:
    return legalizeBufferAtomic(MI, B, IntrID);
  case Intrinsic::amdgcn_atomic_inc:
    return legalizeAtomicIncDec(MI, B, true);
  case Intrinsic::amdgcn_atomic_dec:
    return legalizeAtomicIncDec(MI, B, false);
  case Intrinsic::trap:
    return legalizeTrapIntrinsic(MI, MRI, B);
  case Intrinsic::debugtrap:
    return legalizeDebugTrapIntrinsic(MI, MRI, B);
  default: {
    if (const AMDGPU::ImageDimIntrinsicInfo *ImageDimIntr =
            AMDGPU::getImageDimIntrinsicInfo(IntrID))
      return legalizeImageIntrinsic(MI, B, Observer, ImageDimIntr);
    return true;
  }
  }

  return true;
}