AMDGPULegalizerInfo.cpp

Go to the documentation of this file.

//===- 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.
//===----------------------------------------------------------------------===//
 
#include "AMDGPULegalizerInfo.h"
 
#include "AMDGPU.h"
#include "AMDGPUGlobalISelUtils.h"
#include "AMDGPUInstrInfo.h"
#include "AMDGPUTargetMachine.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
#include "SIRegisterInfo.h"
#include "Utils/AMDGPUBaseInfo.h"
#include "llvm/ADT/ScopeExit.h"
#include "llvm/BinaryFormat/ELF.h"
#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.h"
#include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
#include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
#include "llvm/CodeGen/GlobalISel/Utils.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/IR/DiagnosticInfo.h"
#include "llvm/IR/IntrinsicsAMDGPU.h"
#include "llvm/IR/IntrinsicsR600.h"
 
#define DEBUG_TYPE "amdgpu-legalinfo"
 
using namespace llvm;
using namespace LegalizeActions;
using namespace LegalizeMutations;
using namespace LegalityPredicates;
using namespace MIPatternMatch;
 
// Hack until load/store selection patterns support any tuple of legal types.
static cl::opt<bool> EnableNewLegality(
  "amdgpu-global-isel-new-legality",
  cl::desc("Use GlobalISel desired legality, rather than try to use"
           "rules compatible with selection patterns"),
  cl::init(false),
  cl::ReallyHidden);
 
static constexpr unsigned MaxRegisterSize = 1024;
 
// 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.changeElementCount(ElementCount::getFixed(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);
}
 
/// \returns true if this is an odd sized vector which should widen by adding an
/// additional element. This is mostly to handle <3 x s16> -> <4 x s16>. This
/// excludes s1 vectors, which should always be scalarized.
static LegalityPredicate isSmallOddVector(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    if (!Ty.isVector())
      return false;
 
    const LLT EltTy = Ty.getElementType();
    const unsigned EltSize = EltTy.getSizeInBits();
    return Ty.getNumElements() % 2 != 0 &&
           EltSize > 1 && EltSize < 32 &&
           Ty.getSizeInBits() % 32 != 0;
  };
}
 
static LegalityPredicate sizeIsMultipleOf32(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    return 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::pair(TypeIdx,
                     LLT::fixed_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::pair(TypeIdx, LLT::scalarOrVector(
                                  ElementCount::getFixed(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::pair(TypeIdx, LLT::fixed_vector(NewNumElts, EltTy));
  };
}
 
// Increase the number of vector elements to reach the next legal RegClass.
static LegalizeMutation moreElementsToNextExistingRegClass(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    const unsigned NumElts = Ty.getNumElements();
    const unsigned EltSize = Ty.getElementType().getSizeInBits();
    const unsigned MaxNumElts = MaxRegisterSize / EltSize;
 
    assert(EltSize == 32 || EltSize == 64);
    assert(Ty.getSizeInBits() < MaxRegisterSize);
 
    unsigned NewNumElts;
    // Find the nearest legal RegClass that is larger than the current type.
    for (NewNumElts = NumElts; NewNumElts < MaxNumElts; ++NewNumElts) {
      if (SIRegisterInfo::getSGPRClassForBitWidth(NewNumElts * EltSize))
        break;
    }
 
    return std::pair(TypeIdx, LLT::fixed_vector(NewNumElts, EltSize));
  };
}
 
static LLT getBufferRsrcScalarType(const LLT Ty) {
  if (!Ty.isVector())
    return LLT::scalar(128);
  const ElementCount NumElems = Ty.getElementCount();
  return LLT::vector(NumElems, LLT::scalar(128));
}
 
static LLT getBufferRsrcRegisterType(const LLT Ty) {
  if (!Ty.isVector())
    return LLT::fixed_vector(4, LLT::scalar(32));
  const unsigned NumElems = Ty.getElementCount().getFixedValue();
  return LLT::fixed_vector(NumElems * 4, LLT::scalar(32));
}
 
static LLT getBitcastRegisterType(const LLT Ty) {
  const unsigned Size = Ty.getSizeInBits();
 
  if (Size <= 32) {
    // <2 x s8> -> s16
    // <4 x s8> -> s32
    return LLT::scalar(Size);
  }
 
  return LLT::scalarOrVector(ElementCount::getFixed(Size / 32), 32);
}
 
static LegalizeMutation bitcastToRegisterType(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    return std::pair(TypeIdx, getBitcastRegisterType(Ty));
  };
}
 
static LegalizeMutation bitcastToVectorElement32(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    unsigned Size = Ty.getSizeInBits();
    assert(Size % 32 == 0);
    return std::pair(
        TypeIdx, LLT::scalarOrVector(ElementCount::getFixed(Size / 32), 32));
  };
}
 
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;
  };
}
 
static bool isRegisterSize(unsigned Size) {
  return Size % 32 == 0 && Size <= MaxRegisterSize;
}
 
static bool isRegisterVectorElementType(LLT EltTy) {
  const int EltSize = EltTy.getSizeInBits();
  return EltSize == 16 || EltSize % 32 == 0;
}
 
static bool isRegisterVectorType(LLT Ty) {
  const int EltSize = Ty.getElementType().getSizeInBits();
  return EltSize == 32 || EltSize == 64 ||
         (EltSize == 16 && Ty.getNumElements() % 2 == 0) ||
         EltSize == 128 || EltSize == 256;
}
 
// TODO: replace all uses of isRegisterType with isRegisterClassType
static bool isRegisterType(LLT Ty) {
  if (!isRegisterSize(Ty.getSizeInBits()))
    return false;
 
  if (Ty.isVector())
    return isRegisterVectorType(Ty);
 
  return true;
}
 
// Any combination of 32 or 64-bit elements up the maximum register size, and
// multiples of v2s16.
static LegalityPredicate isRegisterType(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    return isRegisterType(Query.Types[TypeIdx]);
  };
}
 
// RegisterType that doesn't have a corresponding RegClass.
// TODO: Once `isRegisterType` is replaced with `isRegisterClassType` this
// should be removed.
static LegalityPredicate isIllegalRegisterType(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    LLT Ty = Query.Types[TypeIdx];
    return isRegisterType(Ty) &&
           !SIRegisterInfo::getSGPRClassForBitWidth(Ty.getSizeInBits());
  };
}
 
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 const LLT S1 = LLT::scalar(1);
static const LLT S8 = LLT::scalar(8);
static const LLT S16 = LLT::scalar(16);
static const LLT S32 = LLT::scalar(32);
static const LLT F32 = LLT::float32();
static const LLT S64 = LLT::scalar(64);
static const LLT F64 = LLT::float64();
static const LLT S96 = LLT::scalar(96);
static const LLT S128 = LLT::scalar(128);
static const LLT S160 = LLT::scalar(160);
static const LLT S224 = LLT::scalar(224);
static const LLT S256 = LLT::scalar(256);
static const LLT S512 = LLT::scalar(512);
static const LLT MaxScalar = LLT::scalar(MaxRegisterSize);
 
static const LLT V2S8 = LLT::fixed_vector(2, 8);
static const LLT V2S16 = LLT::fixed_vector(2, 16);
static const LLT V4S16 = LLT::fixed_vector(4, 16);
static const LLT V6S16 = LLT::fixed_vector(6, 16);
static const LLT V8S16 = LLT::fixed_vector(8, 16);
static const LLT V10S16 = LLT::fixed_vector(10, 16);
static const LLT V12S16 = LLT::fixed_vector(12, 16);
static const LLT V16S16 = LLT::fixed_vector(16, 16);
 
static const LLT V2F16 = LLT::fixed_vector(2, LLT::float16());
static const LLT V2BF16 = V2F16; // FIXME
 
static const LLT V2S32 = LLT::fixed_vector(2, 32);
static const LLT V3S32 = LLT::fixed_vector(3, 32);
static const LLT V4S32 = LLT::fixed_vector(4, 32);
static const LLT V5S32 = LLT::fixed_vector(5, 32);
static const LLT V6S32 = LLT::fixed_vector(6, 32);
static const LLT V7S32 = LLT::fixed_vector(7, 32);
static const LLT V8S32 = LLT::fixed_vector(8, 32);
static const LLT V9S32 = LLT::fixed_vector(9, 32);
static const LLT V10S32 = LLT::fixed_vector(10, 32);
static const LLT V11S32 = LLT::fixed_vector(11, 32);
static const LLT V12S32 = LLT::fixed_vector(12, 32);
static const LLT V16S32 = LLT::fixed_vector(16, 32);
static const LLT V32S32 = LLT::fixed_vector(32, 32);
 
static const LLT V2S64 = LLT::fixed_vector(2, 64);
static const LLT V3S64 = LLT::fixed_vector(3, 64);
static const LLT V4S64 = LLT::fixed_vector(4, 64);
static const LLT V5S64 = LLT::fixed_vector(5, 64);
static const LLT V6S64 = LLT::fixed_vector(6, 64);
static const LLT V7S64 = LLT::fixed_vector(7, 64);
static const LLT V8S64 = LLT::fixed_vector(8, 64);
static const LLT V16S64 = LLT::fixed_vector(16, 64);
 
static const LLT V2S128 = LLT::fixed_vector(2, 128);
static const LLT V4S128 = LLT::fixed_vector(4, 128);
 
static std::initializer_list<LLT> AllScalarTypes = {S32,  S64,  S96,  S128,
                                                    S160, S224, S256, S512};
 
static std::initializer_list<LLT> AllS16Vectors{
    V2S16, V4S16, V6S16, V8S16, V10S16, V12S16, V16S16, V2S128, V4S128};
 
static std::initializer_list<LLT> AllS32Vectors = {
    V2S32, V3S32,  V4S32,  V5S32,  V6S32,  V7S32, V8S32,
    V9S32, V10S32, V11S32, V12S32, V16S32, V32S32};
 
static std::initializer_list<LLT> AllS64Vectors = {V2S64, V3S64, V4S64, V5S64,
                                                   V6S64, V7S64, V8S64, V16S64};
 
// Checks whether a type is in the list of legal register types.
static bool isRegisterClassType(LLT Ty) {
  if (Ty.isPointerOrPointerVector())
    Ty = Ty.changeElementType(LLT::scalar(Ty.getScalarSizeInBits()));
 
  return is_contained(AllS32Vectors, Ty) || is_contained(AllS64Vectors, Ty) ||
         is_contained(AllScalarTypes, Ty) || is_contained(AllS16Vectors, Ty);
}
 
static LegalityPredicate isRegisterClassType(unsigned TypeIdx) {
  return [TypeIdx](const LegalityQuery &Query) {
    return isRegisterClassType(Query.Types[TypeIdx]);
  };
}
 
// If we have a truncating store or an extending load with a data size larger
// than 32-bits, we need to reduce to a 32-bit type.
static LegalityPredicate isWideScalarExtLoadTruncStore(unsigned TypeIdx) {
  return [=](const LegalityQuery &Query) {
    const LLT Ty = Query.Types[TypeIdx];
    return !Ty.isVector() && Ty.getSizeInBits() > 32 &&
           Query.MMODescrs[0].MemoryTy.getSizeInBits() < Ty.getSizeInBits();
  };
}
 
// 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.
static unsigned maxSizeForAddrSpace(const GCNSubtarget &ST, unsigned AS,
                                    bool IsLoad, bool IsAtomic) {
  switch (AS) {
  case AMDGPUAS::PRIVATE_ADDRESS:
    // FIXME: Private element size.
    return ST.enableFlatScratch() ? 128 : 32;
  case AMDGPUAS::LOCAL_ADDRESS:
    return ST.useDS128() ? 128 : 64;
  case AMDGPUAS::GLOBAL_ADDRESS:
  case AMDGPUAS::CONSTANT_ADDRESS:
  case AMDGPUAS::CONSTANT_ADDRESS_32BIT:
  case AMDGPUAS::BUFFER_RESOURCE:
    // 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.
    return IsLoad ? 512 : 128;
  default:
    // FIXME: Flat addresses may contextually need to be split to 32-bit parts
    // if they may alias scratch depending on the subtarget.  This needs to be
    // moved to custom handling to use addressMayBeAccessedAsPrivate
    return ST.hasMultiDwordFlatScratchAddressing() || IsAtomic ? 128 : 32;
  }
}
 
static bool isLoadStoreSizeLegal(const GCNSubtarget &ST,
                                 const LegalityQuery &Query) {
  const LLT Ty = Query.Types[0];
 
  // Handle G_LOAD, G_ZEXTLOAD, G_SEXTLOAD
  const bool IsLoad = Query.Opcode != AMDGPU::G_STORE;
 
  unsigned RegSize = Ty.getSizeInBits();
  uint64_t MemSize = Query.MMODescrs[0].MemoryTy.getSizeInBits();
  uint64_t AlignBits = Query.MMODescrs[0].AlignInBits;
  unsigned AS = Query.Types[1].getAddressSpace();
 
  // All of these need to be custom lowered to cast the pointer operand.
  if (AS == AMDGPUAS::CONSTANT_ADDRESS_32BIT)
    return false;
 
  // Do not handle extending vector loads.
  if (Ty.isVector() && MemSize != RegSize)
    return false;
 
  // TODO: We should be able to widen loads if the alignment is high enough, but
  // we also need to modify the memory access size.
#if 0
  // Accept widening loads based on alignment.
  if (IsLoad && MemSize < Size)
    MemSize = std::max(MemSize, Align);
#endif
 
  // Only 1-byte and 2-byte to 32-bit extloads are valid.
  if (MemSize != RegSize && RegSize != 32)
    return false;
 
  if (MemSize > maxSizeForAddrSpace(ST, AS, IsLoad,
                                    Query.MMODescrs[0].Ordering !=
                                        AtomicOrdering::NotAtomic))
    return false;
 
  switch (MemSize) {
  case 8:
  case 16:
  case 32:
  case 64:
  case 128:
    break;
  case 96:
    if (!ST.hasDwordx3LoadStores())
      return false;
    break;
  case 256:
  case 512:
    // These may contextually need to be broken down.
    break;
  default:
    return false;
  }
 
  assert(RegSize >= MemSize);
 
  if (AlignBits < MemSize) {
    const SITargetLowering *TLI = ST.getTargetLowering();
    if (!TLI->allowsMisalignedMemoryAccessesImpl(MemSize, AS,
                                                 Align(AlignBits / 8)))
      return false;
  }
 
  return true;
}
 
// The newer buffer intrinsic forms take their resource arguments as
// pointers in address space 8, aka s128 values. However, in order to not break
// SelectionDAG, the underlying operations have to continue to take v4i32
// arguments. Therefore, we convert resource pointers - or vectors of them
// to integer values here.
static bool hasBufferRsrcWorkaround(const LLT Ty) {
  if (Ty.isPointer() && Ty.getAddressSpace() == AMDGPUAS::BUFFER_RESOURCE)
    return true;
  if (Ty.isVector()) {
    const LLT ElemTy = Ty.getElementType();
    return hasBufferRsrcWorkaround(ElemTy);
  }
  return false;
}
 
// The current selector can't handle <6 x s16>, <8 x s16>, s96, s128 etc, so
// workaround this. Eventually it should ignore the type for loads and only care
// about the size. Return true in cases where we will workaround this for now by
// bitcasting.
static bool loadStoreBitcastWorkaround(const LLT Ty) {
  if (EnableNewLegality)
    return false;
 
  const unsigned Size = Ty.getSizeInBits();
  if (Size <= 64)
    return false;
  // Address space 8 pointers get their own workaround.
  if (hasBufferRsrcWorkaround(Ty))
    return false;
  if (!Ty.isVector())
    return true;
 
  if (Ty.isPointerVector())
    return true;
 
  unsigned EltSize = Ty.getScalarSizeInBits();
  return EltSize != 32 && EltSize != 64;
}
 
static bool isLoadStoreLegal(const GCNSubtarget &ST, const LegalityQuery &Query) {
  const LLT Ty = Query.Types[0];
  return isRegisterType(Ty) && isLoadStoreSizeLegal(ST, Query) &&
         !hasBufferRsrcWorkaround(Ty) && !loadStoreBitcastWorkaround(Ty);
}
 
/// Return true if a load or store of the type should be lowered with a bitcast
/// to a different type.
static bool shouldBitcastLoadStoreType(const GCNSubtarget &ST, const LLT Ty,
                                       const LLT MemTy) {
  const unsigned MemSizeInBits = MemTy.getSizeInBits();
  const unsigned Size = Ty.getSizeInBits();
  if (Size != MemSizeInBits)
    return Size <= 32 && Ty.isVector();
 
  if (loadStoreBitcastWorkaround(Ty) && isRegisterType(Ty))
    return true;
 
  // Don't try to handle bitcasting vector ext loads for now.
  return Ty.isVector() && (!MemTy.isVector() || MemTy == Ty) &&
         (Size <= 32 || isRegisterSize(Size)) &&
         !isRegisterVectorElementType(Ty.getElementType());
}
 
/// Return true if we should legalize a load by widening an odd sized memory
/// access up to the alignment. Note this case when the memory access itself
/// changes, not the size of the result register.
static bool shouldWidenLoad(const GCNSubtarget &ST, LLT MemoryTy,
                            uint64_t AlignInBits, unsigned AddrSpace,
                            unsigned Opcode) {
  unsigned SizeInBits = MemoryTy.getSizeInBits();
  // We don't want to widen cases that are naturally legal.
  if (isPowerOf2_32(SizeInBits))
    return false;
 
  // If we have 96-bit memory operations, we shouldn't touch them. Note we may
  // end up widening these for a scalar load during RegBankSelect, if we don't
  // have 96-bit scalar loads.
  if (SizeInBits == 96 && ST.hasDwordx3LoadStores())
    return false;
 
  if (SizeInBits >= maxSizeForAddrSpace(ST, AddrSpace, Opcode, false))
    return false;
 
  // A load is known dereferenceable up to the alignment, so it's legal to widen
  // to it.
  //
  // TODO: Could check dereferenceable for less aligned cases.
  unsigned RoundedSize = NextPowerOf2(SizeInBits);
  if (AlignInBits < RoundedSize)
    return false;
 
  // Do not widen if it would introduce a slow unaligned load.
  const SITargetLowering *TLI = ST.getTargetLowering();
  unsigned Fast = 0;
  return TLI->allowsMisalignedMemoryAccessesImpl(
             RoundedSize, AddrSpace, Align(AlignInBits / 8),
             MachineMemOperand::MOLoad, &Fast) &&
         Fast;
}
 
static bool shouldWidenLoad(const GCNSubtarget &ST, const LegalityQuery &Query,
                            unsigned Opcode) {
  if (Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic)
    return false;
 
  return shouldWidenLoad(ST, Query.MMODescrs[0].MemoryTy,
                         Query.MMODescrs[0].AlignInBits,
                         Query.Types[1].getAddressSpace(), Opcode);
}
 
/// Mutates IR (typicaly a load instruction) to use a <4 x s32> as the initial
/// type of the operand `idx` and then to transform it to a `p8` via bitcasts
/// and inttoptr. In addition, handle vectors of p8. Returns the new type.
static LLT castBufferRsrcFromV4I32(MachineInstr &MI, MachineIRBuilder &B,
                                   MachineRegisterInfo &MRI, unsigned Idx) {
  MachineOperand &MO = MI.getOperand(Idx);
 
  const LLT PointerTy = MRI.getType(MO.getReg());
 
  // Paranoidly prevent us from doing this multiple times.
  if (!hasBufferRsrcWorkaround(PointerTy))
    return PointerTy;
 
  const LLT ScalarTy = getBufferRsrcScalarType(PointerTy);
  const LLT VectorTy = getBufferRsrcRegisterType(PointerTy);
  if (!PointerTy.isVector()) {
    // Happy path: (4 x s32) -> (s32, s32, s32, s32) -> (p8)
    const unsigned NumParts = PointerTy.getSizeInBits() / 32;
    const LLT S32 = LLT::scalar(32);
 
    Register VectorReg = MRI.createGenericVirtualRegister(VectorTy);
    std::array<Register, 4> VectorElems;
    B.setInsertPt(B.getMBB(), ++B.getInsertPt());
    for (unsigned I = 0; I < NumParts; ++I)
      VectorElems[I] =
          B.buildExtractVectorElementConstant(S32, VectorReg, I).getReg(0);
    B.buildMergeValues(MO, VectorElems);
    MO.setReg(VectorReg);
    return VectorTy;
  }
  Register BitcastReg = MRI.createGenericVirtualRegister(VectorTy);
  B.setInsertPt(B.getMBB(), ++B.getInsertPt());
  auto Scalar = B.buildBitcast(ScalarTy, BitcastReg);
  B.buildIntToPtr(MO, Scalar);
  MO.setReg(BitcastReg);
 
  return VectorTy;
}
 
/// Cast a buffer resource (an address space 8 pointer) into a 4xi32, which is
/// the form in which the value must be in order to be passed to the low-level
/// representations used for MUBUF/MTBUF intrinsics. This is a hack, which is
/// needed in order to account for the fact that we can't define a register
/// class for s128 without breaking SelectionDAG.
static Register castBufferRsrcToV4I32(Register Pointer, MachineIRBuilder &B) {
  MachineRegisterInfo &MRI = *B.getMRI();
  const LLT PointerTy = MRI.getType(Pointer);
  const LLT ScalarTy = getBufferRsrcScalarType(PointerTy);
  const LLT VectorTy = getBufferRsrcRegisterType(PointerTy);
 
  if (!PointerTy.isVector()) {
    // Special case: p8 -> (s32, s32, s32, s32) -> (4xs32)
    SmallVector<Register, 4> PointerParts;
    const unsigned NumParts = PointerTy.getSizeInBits() / 32;
    auto Unmerged = B.buildUnmerge(LLT::scalar(32), Pointer);
    for (unsigned I = 0; I < NumParts; ++I)
      PointerParts.push_back(Unmerged.getReg(I));
    return B.buildBuildVector(VectorTy, PointerParts).getReg(0);
  }
  Register Scalar = B.buildPtrToInt(ScalarTy, Pointer).getReg(0);
  return B.buildBitcast(VectorTy, Scalar).getReg(0);
}
 
static void castBufferRsrcArgToV4I32(MachineInstr &MI, MachineIRBuilder &B,
                                     unsigned Idx) {
  MachineOperand &MO = MI.getOperand(Idx);
 
  const LLT PointerTy = B.getMRI()->getType(MO.getReg());
  // Paranoidly prevent us from doing this multiple times.
  if (!hasBufferRsrcWorkaround(PointerTy))
    return;
  MO.setReg(castBufferRsrcToV4I32(MO.getReg(), B));
}
 
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 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 BufferFatPtr = GetAddrSpacePtr(AMDGPUAS::BUFFER_FAT_POINTER);
  const LLT RsrcPtr = GetAddrSpacePtr(AMDGPUAS::BUFFER_RESOURCE);
  const LLT BufferStridedPtr =
      GetAddrSpacePtr(AMDGPUAS::BUFFER_STRIDED_POINTER);
 
  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> AddrSpaces128 = {RsrcPtr};
 
  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;
 
  // s1 for VCC branches, s32 for SCC branches.
  getActionDefinitionsBuilder(G_BRCOND).legalFor({S1, S32});
 
  // TODO: All multiples of 32, vectors of pointers, all v2s16 pairs, more
  // elements for v3s16
  getActionDefinitionsBuilder(G_PHI)
      .legalFor({S32, S64, V2S16, S16, V4S16, S1, S128, S256})
      .legalFor(AllS32Vectors)
      .legalFor(AllS64Vectors)
      .legalFor(AddrSpaces64)
      .legalFor(AddrSpaces32)
      .legalFor(AddrSpaces128)
      .legalIf(isPointer(0))
      .clampScalar(0, S16, S256)
      .widenScalarToNextPow2(0, 32)
      .clampMaxNumElements(0, S32, 16)
      .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
      .scalarize(0);
 
  if (ST.hasVOP3PInsts() && ST.hasAddNoCarry() && ST.hasIntClamp()) {
    // Full set of gfx9 features.
    if (ST.hasScalarAddSub64()) {
      getActionDefinitionsBuilder({G_ADD, G_SUB})
          .legalFor({S64, S32, S16, V2S16})
          .clampMaxNumElementsStrict(0, S16, 2)
          .scalarize(0)
          .minScalar(0, S16)
          .widenScalarToNextMultipleOf(0, 32)
          .maxScalar(0, S32);
    } else {
      getActionDefinitionsBuilder({G_ADD, G_SUB})
          .legalFor({S32, S16, V2S16})
          .clampMaxNumElementsStrict(0, S16, 2)
          .scalarize(0)
          .minScalar(0, S16)
          .widenScalarToNextMultipleOf(0, 32)
          .maxScalar(0, S32);
    }
 
    if (ST.hasScalarSMulU64()) {
      getActionDefinitionsBuilder(G_MUL)
          .legalFor({S64, S32, S16, V2S16})
          .clampMaxNumElementsStrict(0, S16, 2)
          .scalarize(0)
          .minScalar(0, S16)
          .widenScalarToNextMultipleOf(0, 32)
          .custom();
    } else {
      getActionDefinitionsBuilder(G_MUL)
          .legalFor({S32, S16, V2S16})
          .clampMaxNumElementsStrict(0, S16, 2)
          .scalarize(0)
          .minScalar(0, S16)
          .widenScalarToNextMultipleOf(0, 32)
          .custom();
    }
    assert(ST.hasMad64_32());
 
    getActionDefinitionsBuilder({G_UADDSAT, G_USUBSAT, G_SADDSAT, G_SSUBSAT})
      .legalFor({S32, S16, V2S16}) // Clamp modifier
      .minScalarOrElt(0, S16)
      .clampMaxNumElementsStrict(0, S16, 2)
      .scalarize(0)
      .widenScalarToNextPow2(0, 32)
      .lower();
  } else if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder({G_ADD, G_SUB})
      .legalFor({S32, S16})
      .minScalar(0, S16)
      .widenScalarToNextMultipleOf(0, 32)
      .maxScalar(0, S32)
      .scalarize(0);
 
    getActionDefinitionsBuilder(G_MUL)
      .legalFor({S32, S16})
      .scalarize(0)
      .minScalar(0, S16)
      .widenScalarToNextMultipleOf(0, 32)
      .custom();
    assert(ST.hasMad64_32());
 
    // Technically the saturating operations require clamp bit support, but this
    // was introduced at the same time as 16-bit operations.
    getActionDefinitionsBuilder({G_UADDSAT, G_USUBSAT})
      .legalFor({S32, S16}) // Clamp modifier
      .minScalar(0, S16)
      .scalarize(0)
      .widenScalarToNextPow2(0, 16)
      .lower();
 
    // We're just lowering this, but it helps get a better result to try to
    // coerce to the desired type first.
    getActionDefinitionsBuilder({G_SADDSAT, G_SSUBSAT})
      .minScalar(0, S16)
      .scalarize(0)
      .lower();
  } else {
    getActionDefinitionsBuilder({G_ADD, G_SUB})
      .legalFor({S32})
      .widenScalarToNextMultipleOf(0, 32)
      .clampScalar(0, S32, S32)
      .scalarize(0);
 
    auto &Mul = getActionDefinitionsBuilder(G_MUL)
      .legalFor({S32})
      .scalarize(0)
      .minScalar(0, S32)
      .widenScalarToNextMultipleOf(0, 32);
 
    if (ST.hasMad64_32())
      Mul.custom();
    else
      Mul.maxScalar(0, S32);
 
    if (ST.hasIntClamp()) {
      getActionDefinitionsBuilder({G_UADDSAT, G_USUBSAT})
        .legalFor({S32}) // Clamp modifier.
        .scalarize(0)
        .minScalarOrElt(0, S32)
        .lower();
    } else {
      // Clamp bit support was added in VI, along with 16-bit operations.
      getActionDefinitionsBuilder({G_UADDSAT, G_USUBSAT})
        .minScalar(0, S32)
        .scalarize(0)
        .lower();
    }
 
    // FIXME: DAG expansion gets better results. The widening uses the smaller
    // range values and goes for the min/max lowering directly.
    getActionDefinitionsBuilder({G_SADDSAT, G_SSUBSAT})
      .minScalar(0, S32)
      .scalarize(0)
      .lower();
  }
 
  getActionDefinitionsBuilder(
      {G_SDIV, G_UDIV, G_SREM, G_UREM, G_SDIVREM, G_UDIVREM})
      .customFor({S32, S64})
      .clampScalar(0, S32, S64)
      .widenScalarToNextPow2(0, 32)
      .scalarize(0);
 
  auto &Mulh = getActionDefinitionsBuilder({G_UMULH, G_SMULH})
                   .legalFor({S32})
                   .maxScalar(0, S32);
 
  if (ST.hasVOP3PInsts()) {
    Mulh
      .clampMaxNumElements(0, S8, 2)
      .lowerFor({V2S8});
  }
 
  Mulh
    .scalarize(0)
    .lower();
 
  // 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}})
      .clampScalar(0, S32, S32)
      .scalarize(0);
 
  getActionDefinitionsBuilder(G_BITCAST)
      // Don't worry about the size constraint.
      .legalIf(all(isRegisterClassType(0), isRegisterClassType(1)))
      .lower();
 
  getActionDefinitionsBuilder(G_CONSTANT)
    .legalFor({S1, S32, S64, S16, GlobalPtr,
               LocalPtr, ConstantPtr, PrivatePtr, FlatPtr })
    .legalIf(isPointer(0))
    .clampScalar(0, S32, S64)
    .widenScalarToNextPow2(0);
 
  getActionDefinitionsBuilder(G_FCONSTANT)
    .legalFor({S32, S64, S16})
    .clampScalar(0, S16, S64);
 
  getActionDefinitionsBuilder({G_IMPLICIT_DEF, G_FREEZE})
      .legalIf(isRegisterType(0))
      // s1 and s16 are special cases because they have legal operations on
      // them, but don't really occupy registers in the normal way.
      .legalFor({S1, S16})
      .moreElementsIf(isSmallOddVector(0), oneMoreElement(0))
      .clampScalarOrElt(0, S32, MaxScalar)
      .widenScalarToNextPow2(0, 32)
      .clampMaxNumElements(0, S32, 16);
 
  getActionDefinitionsBuilder(G_FRAME_INDEX).legalFor({PrivatePtr});
 
  // If the amount is divergent, we have to do a wave reduction to get the
  // maximum value, so this is expanded during RegBankSelect.
  getActionDefinitionsBuilder(G_DYN_STACKALLOC)
    .legalFor({{PrivatePtr, S32}});
 
  getActionDefinitionsBuilder(G_STACKSAVE)
    .customFor({PrivatePtr});
  getActionDefinitionsBuilder(G_STACKRESTORE)
    .legalFor({PrivatePtr});
 
  getActionDefinitionsBuilder({G_GET_FPENV, G_SET_FPENV}).customFor({S64});
 
  getActionDefinitionsBuilder(G_GLOBAL_VALUE)
    .customIf(typeIsNot(0, PrivatePtr));
 
  getActionDefinitionsBuilder(G_BLOCK_ADDR).legalFor({CodePtr});
 
  auto &FPOpActions = getActionDefinitionsBuilder(
    { G_FADD, G_FMUL, G_FMA, G_FCANONICALIZE,
      G_STRICT_FADD, G_STRICT_FMUL, G_STRICT_FMA})
    .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});
  }
 
  if (ST.hasPackedFP32Ops()) {
    FPOpActions.legalFor({V2S32});
    FPOpActions.clampMaxNumElementsStrict(0, S32, 2);
  }
 
  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.clampMaxNumElementsStrict(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)
    .clampMaxNumElementsStrict(0, S16, 2)
    .scalarize(0)
    .clampScalar(0, S16, S64);
 
  if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder(G_FSQRT)
      .legalFor({S16})
      .customFor({S32, S64})
      .scalarize(0)
      .unsupported();
    getActionDefinitionsBuilder(G_FFLOOR)
      .legalFor({S32, S64, S16})
      .scalarize(0)
      .clampScalar(0, S16, S64);
 
    getActionDefinitionsBuilder({G_FLDEXP, G_STRICT_FLDEXP})
      .legalFor({{S32, S32}, {S64, S32}, {S16, S16}})
      .scalarize(0)
      .maxScalarIf(typeIs(0, S16), 1, S16)
      .clampScalar(1, S32, S32)
      .lower();
 
    getActionDefinitionsBuilder(G_FFREXP)
      .customFor({{S32, S32}, {S64, S32}, {S16, S16}, {S16, S32}})
      .scalarize(0)
      .lower();
  } else {
    getActionDefinitionsBuilder(G_FSQRT)
      .customFor({S32, S64, S16})
      .scalarize(0)
      .unsupported();
 
 
    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_FLDEXP, G_STRICT_FLDEXP})
      .legalFor({{S32, S32}, {S64, S32}})
      .scalarize(0)
      .clampScalar(0, S32, S64)
      .clampScalar(1, S32, S32)
      .lower();
 
    getActionDefinitionsBuilder(G_FFREXP)
      .customFor({{S32, S32}, {S64, S32}})
      .scalarize(0)
      .minScalar(0, S32)
      .clampScalar(1, S32, S32)
      .lower();
  }
 
  getActionDefinitionsBuilder(G_FPTRUNC)
    .legalFor({{S32, S64}, {S16, S32}})
    .scalarize(0)
    .lower();
 
  getActionDefinitionsBuilder(G_FPEXT)
    .legalFor({{S64, S32}, {S32, S16}})
    .narrowScalarFor({{S64, S16}}, changeTo(0, S32))
    .scalarize(0);
 
  auto &FSubActions = getActionDefinitionsBuilder({G_FSUB, G_STRICT_FSUB});
  if (ST.has16BitInsts()) {
    FSubActions
      // Use actual fsub instruction
      .legalFor({S32, S16})
      // Must use fadd + fneg
      .lowerFor({S64, V2S16});
  } else {
    FSubActions
      // Use actual fsub instruction
      .legalFor({S32})
      // Must use fadd + fneg
      .lowerFor({S64, S16, V2S16});
  }
 
  FSubActions
    .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() && ST.hasMadMacF32Insts())
    FMad.customFor({S32, S16});
  else if (ST.hasMadMacF32Insts())
    FMad.customFor({S32});
  else if (ST.hasMadF16())
    FMad.customFor({S16});
  FMad.scalarize(0)
      .lower();
 
  auto &FRem = getActionDefinitionsBuilder(G_FREM);
  if (ST.has16BitInsts()) {
    FRem.customFor({S16, S32, S64});
  } else {
    FRem.minScalar(0, S32)
        .customFor({S32, S64});
  }
  FRem.scalarize(0);
 
  // 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}})
                    .lowerIf(typeIs(1, S1))
                    .customFor({{S32, S64}, {S64, S64}});
  if (ST.has16BitInsts())
    IToFP.legalFor({{S16, S16}});
  IToFP.clampScalar(1, S32, S64)
       .minScalar(0, S32)
       .scalarize(0)
       .widenScalarToNextPow2(1);
 
  auto &FPToI = getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
    .legalFor({{S32, S32}, {S32, S64}, {S32, S16}})
    .customFor({{S64, S32}, {S64, S64}})
    .narrowScalarFor({{S64, S16}}, changeTo(0, S32));
  if (ST.has16BitInsts())
    FPToI.legalFor({{S16, S16}});
  else
    FPToI.minScalar(1, S32);
 
  FPToI.minScalar(0, S32)
       .widenScalarToNextPow2(0, 32)
       .scalarize(0)
       .lower();
 
  getActionDefinitionsBuilder({G_LROUND, G_LLROUND})
      .clampScalar(0, S16, S64)
      .scalarize(0)
      .lower();
 
  getActionDefinitionsBuilder(G_INTRINSIC_FPTRUNC_ROUND)
      .customFor({S16, S32})
      .scalarize(0)
      .lower();
 
  // Lower G_FNEARBYINT and G_FRINT into G_INTRINSIC_ROUNDEVEN
  getActionDefinitionsBuilder({G_INTRINSIC_ROUND, G_FRINT, G_FNEARBYINT})
      .scalarize(0)
      .lower();
 
  getActionDefinitionsBuilder({G_INTRINSIC_LRINT, G_INTRINSIC_LLRINT})
      .clampScalar(0, S16, S64)
      .scalarize(0)
      .lower();
 
  if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder(
        {G_INTRINSIC_TRUNC, G_FCEIL, G_INTRINSIC_ROUNDEVEN})
        .legalFor({S16, S32, S64})
        .clampScalar(0, S16, S64)
        .scalarize(0);
  } else if (ST.getGeneration() >= AMDGPUSubtarget::SEA_ISLANDS) {
    getActionDefinitionsBuilder(
        {G_INTRINSIC_TRUNC, G_FCEIL, G_INTRINSIC_ROUNDEVEN})
        .legalFor({S32, S64})
        .clampScalar(0, S32, S64)
        .scalarize(0);
  } else {
    getActionDefinitionsBuilder(
        {G_INTRINSIC_TRUNC, G_FCEIL, G_INTRINSIC_ROUNDEVEN})
        .legalFor({S32})
        .customFor({S64})
        .clampScalar(0, S32, S64)
        .scalarize(0);
  }
 
  getActionDefinitionsBuilder(G_PTR_ADD)
      .unsupportedFor({BufferFatPtr, BufferStridedPtr, RsrcPtr})
      .legalIf(all(isPointer(0), sameSize(0, 1)))
      .scalarize(0)
      .scalarSameSizeAs(1, 0);
 
  getActionDefinitionsBuilder(G_PTRMASK)
    .legalIf(all(sameSize(0, 1), typeInSet(1, {S64, S32})))
    .scalarSameSizeAs(1, 0)
    .scalarize(0);
 
  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)));
 
  auto &FCmpBuilder =
      getActionDefinitionsBuilder(G_FCMP).legalForCartesianProduct(
          {S1}, ST.has16BitInsts() ? FPTypes16 : FPTypesBase);
 
  if (ST.hasSALUFloatInsts())
    FCmpBuilder.legalForCartesianProduct({S32}, {S16, S32});
 
  FCmpBuilder
    .widenScalarToNextPow2(1)
    .clampScalar(1, S32, S64)
    .scalarize(0);
 
  // FIXME: fpow has a selection pattern that should move to custom lowering.
  auto &ExpOps = getActionDefinitionsBuilder(G_FPOW);
  if (ST.has16BitInsts())
    ExpOps.customFor({{S32}, {S16}});
  else
    ExpOps.customFor({S32});
  ExpOps.clampScalar(0, MinScalarFPTy, S32)
        .scalarize(0);
 
  getActionDefinitionsBuilder(G_FPOWI)
    .clampScalar(0, MinScalarFPTy, S32)
    .lower();
 
  auto &Log2Ops = getActionDefinitionsBuilder({G_FLOG2, G_FEXP2});
  Log2Ops.customFor({S32});
  if (ST.has16BitInsts())
    Log2Ops.legalFor({S16});
  else
    Log2Ops.customFor({S16});
  Log2Ops.scalarize(0)
    .lower();
 
  auto &LogOps =
      getActionDefinitionsBuilder({G_FLOG, G_FLOG10, G_FEXP, G_FEXP10});
  LogOps.customFor({S32, S16});
  LogOps.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)
    .widenScalarToNextPow2(1, 32)
    .clampScalar(1, S32, S64)
    .scalarize(0)
    .widenScalarToNextPow2(0, 32);
 
  // If no 16 bit instr is available, lower into different instructions.
  if (ST.has16BitInsts())
    getActionDefinitionsBuilder(G_IS_FPCLASS)
        .legalForCartesianProduct({S1}, FPTypes16)
        .widenScalarToNextPow2(1)
        .scalarize(0)
        .lower();
  else
    getActionDefinitionsBuilder(G_IS_FPCLASS)
        .legalForCartesianProduct({S1}, FPTypesBase)
        .lowerFor({S1, S16})
        .widenScalarToNextPow2(1)
        .scalarize(0)
        .lower();
 
  // 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)
    .custom();
 
  // The 64-bit versions produce 32-bit results, but only on the SALU.
  getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF)
      .legalFor({{S32, S32}, {S32, S64}})
      .customIf(scalarNarrowerThan(1, 32))
      .clampScalar(0, S32, S32)
      .clampScalar(1, S32, S64)
      .scalarize(0)
      .widenScalarToNextPow2(0, 32)
      .widenScalarToNextPow2(1, 32);
 
  getActionDefinitionsBuilder(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);
 
  // S64 is only legal on SALU, and needs to be broken into 32-bit elements in
  // RegBankSelect.
  getActionDefinitionsBuilder(G_BITREVERSE)
    .legalFor({S32, S64})
    .clampScalar(0, S32, S64)
    .scalarize(0)
    .widenScalarToNextPow2(0);
 
  if (ST.has16BitInsts()) {
    getActionDefinitionsBuilder(G_BSWAP)
      .legalFor({S16, S32, V2S16})
      .clampMaxNumElementsStrict(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, G_ABS})
        .legalFor({S32, S16, V2S16})
        .clampMaxNumElements(0, S16, 2)
        .minScalar(0, S16)
        .widenScalarToNextPow2(0)
        .scalarize(0)
        .lower();
    } else {
      getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX, G_ABS})
        .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(scalarNarrowerThan(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, G_ABS})
      .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::pair(
                           1, LLT::scalar(Query.Types[0].getSizeInBits()));
                     })
      .narrowScalarIf(largerThan(1, 0), [](const LegalityQuery &Query) {
        return std::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::pair(
                           0, LLT::scalar(Query.Types[1].getSizeInBits()));
                     })
      .narrowScalarIf(largerThan(0, 1), [](const LegalityQuery &Query) {
        return std::pair(0, LLT::scalar(Query.Types[1].getSizeInBits()));
      });
 
  getActionDefinitionsBuilder(G_ADDRSPACE_CAST)
    .scalarize(0)
    .custom();
 
  const auto needToSplitMemOp = [=](const LegalityQuery &Query,
                                    bool IsLoad) -> bool {
    const LLT DstTy = Query.Types[0];
 
    // Split vector extloads.
    unsigned MemSize = Query.MMODescrs[0].MemoryTy.getSizeInBits();
 
    if (DstTy.isVector() && DstTy.getSizeInBits() > MemSize)
      return true;
 
    const LLT PtrTy = Query.Types[1];
    unsigned AS = PtrTy.getAddressSpace();
    if (MemSize > maxSizeForAddrSpace(ST, AS, IsLoad,
                                      Query.MMODescrs[0].Ordering !=
                                          AtomicOrdering::NotAtomic))
      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;
    }
 
    return false;
  };
 
  unsigned GlobalAlign32 = ST.hasUnalignedBufferAccessEnabled() ? 0 : 32;
  unsigned GlobalAlign16 = ST.hasUnalignedBufferAccessEnabled() ? 0 : 16;
  unsigned GlobalAlign8 = ST.hasUnalignedBufferAccessEnabled() ? 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);
    // Explicitly list some common cases.
    // TODO: Does this help compile time at all?
    Actions.legalForTypesWithMemDesc({{S32, GlobalPtr, S32, GlobalAlign32},
                                      {V2S32, GlobalPtr, V2S32, GlobalAlign32},
                                      {V4S32, GlobalPtr, V4S32, GlobalAlign32},
                                      {S64, GlobalPtr, S64, GlobalAlign32},
                                      {V2S64, GlobalPtr, V2S64, GlobalAlign32},
                                      {V2S16, GlobalPtr, V2S16, GlobalAlign32},
                                      {S32, GlobalPtr, S8, GlobalAlign8},
                                      {S32, GlobalPtr, S16, GlobalAlign16},
 
                                      {S32, LocalPtr, S32, 32},
                                      {S64, LocalPtr, S64, 32},
                                      {V2S32, LocalPtr, V2S32, 32},
                                      {S32, LocalPtr, S8, 8},
                                      {S32, LocalPtr, S16, 16},
                                      {V2S16, LocalPtr, S32, 32},
 
                                      {S32, PrivatePtr, S32, 32},
                                      {S32, PrivatePtr, S8, 8},
                                      {S32, PrivatePtr, S16, 16},
                                      {V2S16, PrivatePtr, S32, 32},
 
                                      {S32, ConstantPtr, S32, GlobalAlign32},
                                      {V2S32, ConstantPtr, V2S32, GlobalAlign32},
                                      {V4S32, ConstantPtr, V4S32, GlobalAlign32},
                                      {S64, ConstantPtr, S64, GlobalAlign32},
                                      {V2S32, ConstantPtr, V2S32, GlobalAlign32}});
    Actions.legalIf(
      [=](const LegalityQuery &Query) -> bool {
        return isLoadStoreLegal(ST, Query);
      });
 
    // The custom pointers (fat pointers, buffer resources) don't work with load
    // and store at this level. Fat pointers should have been lowered to
    // intrinsics before the translation to MIR.
    Actions.unsupportedIf(
        typeInSet(1, {BufferFatPtr, BufferStridedPtr, RsrcPtr}));
 
    // Address space 8 pointers are handled by a 4xs32 load, bitcast, and
    // ptrtoint. This is needed to account for the fact that we can't have i128
    // as a register class for SelectionDAG reasons.
    Actions.customIf([=](const LegalityQuery &Query) -> bool {
      return hasBufferRsrcWorkaround(Query.Types[0]);
    });
 
    // Constant 32-bit is handled by addrspacecasting the 32-bit pointer to
    // 64-bits.
    //
    // TODO: Should generalize bitcast action into coerce, which will also cover
    // inserting addrspacecasts.
    Actions.customIf(typeIs(1, Constant32Ptr));
 
    // Turn any illegal element vectors into something easier to deal
    // with. These will ultimately produce 32-bit scalar shifts to extract the
    // parts anyway.
    //
    // For odd 16-bit element vectors, prefer to split those into pieces with
    // 16-bit vector parts.
    Actions.bitcastIf(
      [=](const LegalityQuery &Query) -> bool {
        return shouldBitcastLoadStoreType(ST, Query.Types[0],
                                          Query.MMODescrs[0].MemoryTy);
      }, bitcastToRegisterType(0));
 
    if (!IsStore) {
      // Widen suitably aligned loads by loading extra bytes. The standard
      // legalization actions can't properly express widening memory operands.
      Actions.customIf([=](const LegalityQuery &Query) -> bool {
        return shouldWidenLoad(ST, Query, G_LOAD);
      });
    }
 
    // FIXME: load/store narrowing should be moved to lower action
    Actions
        .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].MemoryTy.getSizeInBits();
 
              // Split extloads.
              if (DstSize > MemSize)
                return std::pair(0, LLT::scalar(MemSize));
 
              unsigned MaxSize = maxSizeForAddrSpace(
                  ST, PtrTy.getAddressSpace(), Op == G_LOAD,
                  Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic);
              if (MemSize > MaxSize)
                return std::pair(0, LLT::scalar(MaxSize));
 
              uint64_t Align = Query.MMODescrs[0].AlignInBits;
              return std::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(
                  ST, PtrTy.getAddressSpace(), Op == G_LOAD,
                  Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic);
 
              // 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.
              unsigned MemSize = Query.MMODescrs[0].MemoryTy.getSizeInBits();
              if (MemSize > MaxSize) {
                unsigned NumElts = DstTy.getNumElements();
                unsigned EltSize = EltTy.getSizeInBits();
 
                if (MaxSize % EltSize == 0) {
                  return std::pair(
                      0, LLT::scalarOrVector(
                             ElementCount::getFixed(MaxSize / EltSize), EltTy));
                }
 
                unsigned NumPieces = MemSize / 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::pair(0, EltTy);
 
                return std::pair(0,
                                 LLT::fixed_vector(NumElts / NumPieces, EltTy));
              }
 
              // FIXME: We could probably handle weird extending loads better.
              if (DstTy.getSizeInBits() > MemSize)
                return std::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 = llvm::bit_floor(DstSize);
                return std::pair(
                    0, LLT::scalarOrVector(
                           ElementCount::getFixed(FloorSize / EltSize), EltTy));
              }
 
              // May need relegalization for the scalars.
              return std::pair(0, EltTy);
            })
    .minScalar(0, S32)
    .narrowScalarIf(isWideScalarExtLoadTruncStore(0), changeTo(0, S32))
    .widenScalarToNextPow2(0)
    .moreElementsIf(vectorSmallerThan(0, 32), moreEltsToNext32Bit(0))
    .lower();
  }
 
  // FIXME: Unaligned accesses not lowered.
  auto &ExtLoads = getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
                       .legalForTypesWithMemDesc({{S32, GlobalPtr, S8, 8},
                                                  {S32, GlobalPtr, S16, 2 * 8},
                                                  {S32, LocalPtr, S8, 8},
                                                  {S32, LocalPtr, S16, 16},
                                                  {S32, PrivatePtr, S8, 8},
                                                  {S32, PrivatePtr, S16, 16},
                                                  {S32, ConstantPtr, S8, 8},
                                                  {S32, ConstantPtr, S16, 2 * 8}})
                       .legalIf(
                         [=](const LegalityQuery &Query) -> bool {
                           return isLoadStoreLegal(ST, Query);
                         });
 
  if (ST.hasFlatAddressSpace()) {
    ExtLoads.legalForTypesWithMemDesc(
        {{S32, FlatPtr, S8, 8}, {S32, FlatPtr, S16, 16}});
  }
 
  // Constant 32-bit is handled by addrspacecasting the 32-bit pointer to
  // 64-bits.
  //
  // TODO: Should generalize bitcast action into coerce, which will also cover
  // inserting addrspacecasts.
  ExtLoads.customIf(typeIs(1, Constant32Ptr));
 
  ExtLoads.clampScalar(0, S32, S32)
          .widenScalarToNextPow2(0)
          .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, G_ATOMICRMW_UINC_WRAP, G_ATOMICRMW_UDEC_WRAP})
    .legalFor({{S32, GlobalPtr}, {S32, LocalPtr},
               {S64, GlobalPtr}, {S64, LocalPtr},
               {S32, RegionPtr}, {S64, RegionPtr}});
  if (ST.hasFlatAddressSpace()) {
    Atomics.legalFor({{S32, FlatPtr}, {S64, FlatPtr}});
  }
 
  // TODO: v2bf16 operations, and fat buffer pointer support.
  auto &Atomic = getActionDefinitionsBuilder(G_ATOMICRMW_FADD);
  if (ST.hasLDSFPAtomicAddF32()) {
    Atomic.legalFor({{S32, LocalPtr}, {S32, RegionPtr}});
    if (ST.hasLdsAtomicAddF64())
      Atomic.legalFor({{S64, LocalPtr}});
    if (ST.hasAtomicDsPkAdd16Insts())
      Atomic.legalFor({{V2F16, LocalPtr}, {V2BF16, LocalPtr}});
  }
  if (ST.hasAtomicFaddInsts())
    Atomic.legalFor({{S32, GlobalPtr}});
  if (ST.hasFlatAtomicFaddF32Inst())
    Atomic.legalFor({{S32, FlatPtr}});
 
  if (ST.hasGFX90AInsts()) {
    // These are legal with some caveats, and should have undergone expansion in
    // the IR in most situations
    // TODO: Move atomic expansion into legalizer
    Atomic.legalFor({
        {S32, GlobalPtr},
        {S64, GlobalPtr},
        {S64, FlatPtr}
      });
  }
 
  if (ST.hasAtomicBufferGlobalPkAddF16NoRtnInsts() ||
      ST.hasAtomicBufferGlobalPkAddF16Insts())
    Atomic.legalFor({{V2F16, GlobalPtr}, {V2F16, BufferFatPtr}});
  if (ST.hasAtomicGlobalPkAddBF16Inst())
    Atomic.legalFor({{V2BF16, GlobalPtr}});
  if (ST.hasAtomicFlatPkAdd16Insts())
    Atomic.legalFor({{V2F16, FlatPtr}, {V2BF16, FlatPtr}});
 
 
  // Most of the legalization work here is done by AtomicExpand. We could
  // probably use a simpler legality rule that just assumes anything is OK.
  auto &AtomicFMinFMax =
    getActionDefinitionsBuilder({G_ATOMICRMW_FMIN, G_ATOMICRMW_FMAX})
    .legalFor({{F32, LocalPtr}, {F64, LocalPtr}});
 
  if (ST.hasAtomicFMinFMaxF32GlobalInsts())
    AtomicFMinFMax.legalFor({{F32, GlobalPtr},{F32, BufferFatPtr}});
  if (ST.hasAtomicFMinFMaxF64GlobalInsts())
    AtomicFMinFMax.legalFor({{F64, GlobalPtr}, {F64, BufferFatPtr}});
  if (ST.hasAtomicFMinFMaxF32FlatInsts())
    AtomicFMinFMax.legalFor({F32, FlatPtr});
  if (ST.hasAtomicFMinFMaxF64FlatInsts())
    AtomicFMinFMax.legalFor({F64, FlatPtr});
 
  // 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::fixed_vector(2, LocalPtr),
                                 LLT::fixed_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.widenScalarToNextPow2(0, 16);
    Shifts.clampScalar(0, S16, S64);
 
    getActionDefinitionsBuilder({G_SSHLSAT, G_USHLSAT})
      .minScalar(0, S16)
      .scalarize(0)
      .lower();
  } 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.widenScalarToNextPow2(0, 32);
    Shifts.clampScalar(0, S32, S64);
 
    getActionDefinitionsBuilder({G_SSHLSAT, G_USHLSAT})
      .minScalar(0, S32)
      .scalarize(0)
      .lower();
  }
  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];
          const unsigned EltSize = EltTy.getSizeInBits();
          const bool isLegalVecType =
              !!SIRegisterInfo::getSGPRClassForBitWidth(VecTy.getSizeInBits());
          // Address space 8 pointers are 128-bit wide values, but the logic
          // below will try to bitcast them to 2N x s64, which will fail.
          // Therefore, as an intermediate step, wrap extracts/insertions from a
          // ptrtoint-ing the vector and scalar arguments (or inttoptring the
          // extraction result) in order to produce a vector operation that can
          // be handled by the logic below.
          if (EltTy.isPointer() && EltSize > 64)
            return true;
          return (EltSize == 32 || EltSize == 64) &&
                  VecTy.getSizeInBits() % 32 == 0 &&
                  VecTy.getSizeInBits() <= MaxRegisterSize &&
                  IdxTy.getSizeInBits() == 32 &&
                  isLegalVecType;
        })
      .bitcastIf(all(sizeIsMultipleOf32(VecTypeIdx), scalarOrEltNarrowerThan(VecTypeIdx, 32)),
                 bitcastToVectorElement32(VecTypeIdx))
      //.bitcastIf(vectorSmallerThan(1, 32), bitcastToScalar(1))
      .bitcastIf(
        all(sizeIsMultipleOf32(VecTypeIdx), scalarOrEltWiderThan(VecTypeIdx, 64)),
        [=](const LegalityQuery &Query) {
          // For > 64-bit element types, try to turn this into a 64-bit
          // element vector since we may be able to do better indexing
          // if this is scalar. If not, fall back to 32.
          const LLT EltTy = Query.Types[EltTypeIdx];
          const LLT VecTy = Query.Types[VecTypeIdx];
          const unsigned DstEltSize = EltTy.getSizeInBits();
          const unsigned VecSize = VecTy.getSizeInBits();
 
          const unsigned TargetEltSize = DstEltSize % 64 == 0 ? 64 : 32;
          return std::pair(
              VecTypeIdx,
              LLT::fixed_vector(VecSize / TargetEltSize, TargetEltSize));
        })
      .clampScalar(EltTypeIdx, S32, S64)
      .clampScalar(VecTypeIdx, S32, S64)
      .clampScalar(IdxTypeIdx, S32, S32)
      .clampMaxNumElements(VecTypeIdx, S32, 32)
      // TODO: Clamp elements for 64-bit vectors?
      .moreElementsIf(
        isIllegalRegisterType(VecTypeIdx),
        moreElementsToNextExistingRegClass(VecTypeIdx))
      // It should only be necessary with variable indexes.
      // As a last resort, lower to the stack
      .lower();
  }
 
  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)))
      .lowerIf([=](const LegalityQuery &Query) {
          // Sub-vector(or single element) insert and extract.
          // TODO: verify immediate offset here since lower only works with
          // whole elements.
          const LLT BigTy = Query.Types[BigTyIdx];
          return BigTy.isVector();
        })
      // 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))
    .moreElementsIf(
      isIllegalRegisterType(0),
      moreElementsToNextExistingRegClass(0));
 
  if (ST.hasScalarPackInsts()) {
    BuildVector
      // FIXME: Should probably widen s1 vectors straight to s32
      .minScalarOrElt(0, S16)
      .minScalar(1, S16);
 
    getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC)
      .legalFor({V2S16, S32})
      .lower();
  } 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(all(isRegisterType(0), isRegisterType(1)))
    .clampMaxNumElements(0, S32, 32)
    .clampMaxNumElements(1, S16, 2) // TODO: Make 4?
    .clampMaxNumElements(0, S16, 64);
 
  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 (!llvm::has_single_bit<uint32_t>(EltTy.getSizeInBits()))
          return true;
      }
      return false;
    };
 
    auto &Builder = getActionDefinitionsBuilder(Op)
      .legalIf(all(isRegisterType(0), isRegisterType(1)))
      .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(scalarNarrowerThan(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, MaxScalar);
 
    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 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::pair(BigTyIdx, LLT::scalar(NewSizeInBits));
      })
      // 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.
      .clampMaxNumElementsStrict(0, S16, 2);
  } 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_ROTR, G_ROTL})
    .scalarize(0)
    .lower();
 
  // TODO: Only Try to form v2s16 with legal packed instructions.
  getActionDefinitionsBuilder(G_FSHR)
    .legalFor({{S32, S32}})
    .lowerFor({{V2S16, V2S16}})
    .clampMaxNumElementsStrict(0, S16, 2)
    .scalarize(0)
    .lower();
 
  if (ST.hasVOP3PInsts()) {
    getActionDefinitionsBuilder(G_FSHL)
      .lowerFor({{V2S16, V2S16}})
      .clampMaxNumElementsStrict(0, S16, 2)
      .scalarize(0)
      .lower();
  } else {
    getActionDefinitionsBuilder(G_FSHL)
      .scalarize(0)
      .lower();
  }
 
  getActionDefinitionsBuilder(G_READCYCLECOUNTER)
    .legalFor({S64});
 
  getActionDefinitionsBuilder(G_READSTEADYCOUNTER).legalFor({S64});
 
  getActionDefinitionsBuilder(G_FENCE)
    .alwaysLegal();
 
  getActionDefinitionsBuilder({G_SMULO, G_UMULO})
      .scalarize(0)
      .minScalar(0, S32)
      .lower();
 
  getActionDefinitionsBuilder({G_SBFX, G_UBFX})
      .legalFor({{S32, S32}, {S64, S32}})
      .clampScalar(1, S32, S32)
      .clampScalar(0, S32, S64)
      .widenScalarToNextPow2(0)
      .scalarize(0);
 
  getActionDefinitionsBuilder(
      {// TODO: Verify V_BFI_B32 is generated from expanded bit ops
       G_FCOPYSIGN,
 
       G_ATOMIC_CMPXCHG_WITH_SUCCESS, G_ATOMICRMW_NAND, G_ATOMICRMW_FSUB,
       G_READ_REGISTER, G_WRITE_REGISTER,
 
       G_SADDO, G_SSUBO})
      .lower();
 
  if (ST.hasIEEEMinMax()) {
    getActionDefinitionsBuilder({G_FMINIMUM, G_FMAXIMUM})
        .legalFor(FPTypesPK16)
        .clampMaxNumElements(0, S16, 2)
        .scalarize(0);
  } else {
    // TODO: Implement
    getActionDefinitionsBuilder({G_FMINIMUM, G_FMAXIMUM}).lower();
  }
 
  getActionDefinitionsBuilder({G_MEMCPY, G_MEMCPY_INLINE, G_MEMMOVE, G_MEMSET})
      .lower();
 
  getActionDefinitionsBuilder({G_TRAP, G_DEBUGTRAP}).custom();
 
  getActionDefinitionsBuilder({G_VASTART, G_VAARG, G_BRJT, G_JUMP_TABLE,
        G_INDEXED_LOAD, G_INDEXED_SEXTLOAD,
        G_INDEXED_ZEXTLOAD, G_INDEXED_STORE})
    .unsupported();
 
  getActionDefinitionsBuilder(G_PREFETCH).alwaysLegal();
 
  getLegacyLegalizerInfo().computeTables();
  verify(*ST.getInstrInfo());
}
 
bool AMDGPULegalizerInfo::legalizeCustom(
    LegalizerHelper &Helper, MachineInstr &MI,
    LostDebugLocObserver &LocObserver) const {
  MachineIRBuilder &B = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *B.getMRI();
 
  switch (MI.getOpcode()) {
  case TargetOpcode::G_ADDRSPACE_CAST:
    return legalizeAddrSpaceCast(MI, MRI, B);
  case TargetOpcode::G_INTRINSIC_ROUNDEVEN:
    return legalizeFroundeven(MI, MRI, B);
  case TargetOpcode::G_FCEIL:
    return legalizeFceil(MI, MRI, B);
  case TargetOpcode::G_FREM:
    return legalizeFrem(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(Helper, MI);
  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_FSIN:
  case TargetOpcode::G_FCOS:
    return legalizeSinCos(MI, MRI, B);
  case TargetOpcode::G_GLOBAL_VALUE:
    return legalizeGlobalValue(MI, MRI, B);
  case TargetOpcode::G_LOAD:
  case TargetOpcode::G_SEXTLOAD:
  case TargetOpcode::G_ZEXTLOAD:
    return legalizeLoad(Helper, MI);
  case TargetOpcode::G_STORE:
    return legalizeStore(Helper, MI);
  case TargetOpcode::G_FMAD:
    return legalizeFMad(MI, MRI, B);
  case TargetOpcode::G_FDIV:
    return legalizeFDIV(MI, MRI, B);
  case TargetOpcode::G_FFREXP:
    return legalizeFFREXP(MI, MRI, B);
  case TargetOpcode::G_FSQRT:
    return legalizeFSQRT(MI, MRI, B);
  case TargetOpcode::G_UDIV:
  case TargetOpcode::G_UREM:
  case TargetOpcode::G_UDIVREM:
    return legalizeUnsignedDIV_REM(MI, MRI, B);
  case TargetOpcode::G_SDIV:
  case TargetOpcode::G_SREM:
  case TargetOpcode::G_SDIVREM:
    return legalizeSignedDIV_REM(MI, MRI, B);
  case TargetOpcode::G_ATOMIC_CMPXCHG:
    return legalizeAtomicCmpXChg(MI, MRI, B);
  case TargetOpcode::G_FLOG2:
    return legalizeFlog2(MI, B);
  case TargetOpcode::G_FLOG:
  case TargetOpcode::G_FLOG10:
    return legalizeFlogCommon(MI, B);
  case TargetOpcode::G_FEXP2:
    return legalizeFExp2(MI, B);
  case TargetOpcode::G_FEXP:
  case TargetOpcode::G_FEXP10:
    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:
  case TargetOpcode::G_BUILD_VECTOR_TRUNC:
    return legalizeBuildVector(MI, MRI, B);
  case TargetOpcode::G_MUL:
    return legalizeMul(Helper, MI);
  case TargetOpcode::G_CTLZ:
  case TargetOpcode::G_CTTZ:
    return legalizeCTLZ_CTTZ(MI, MRI, B);
  case TargetOpcode::G_CTLZ_ZERO_UNDEF:
    return legalizeCTLZ_ZERO_UNDEF(MI, MRI, B);
  case TargetOpcode::G_INTRINSIC_FPTRUNC_ROUND:
    return legalizeFPTruncRound(MI, B);
  case TargetOpcode::G_STACKSAVE:
    return legalizeStackSave(MI, B);
  case TargetOpcode::G_GET_FPENV:
    return legalizeGetFPEnv(MI, MRI, B);
  case TargetOpcode::G_SET_FPENV:
    return legalizeSetFPEnv(MI, MRI, B);
  case TargetOpcode::G_TRAP:
    return legalizeTrap(MI, MRI, B);
  case TargetOpcode::G_DEBUGTRAP:
    return legalizeDebugTrap(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);
  const LLT S64 = LLT::scalar(64);
 
  assert(AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::PRIVATE_ADDRESS);
 
  if (ST.hasApertureRegs()) {
    // Note: this register is somewhat broken. When used as a 32-bit operand,
    // it only returns zeroes. The real value is in the upper 32 bits.
    // Thus, we must emit extract the high 32 bits.
    const unsigned ApertureRegNo = (AS == AMDGPUAS::LOCAL_ADDRESS)
                                       ? AMDGPU::SRC_SHARED_BASE
                                       : AMDGPU::SRC_PRIVATE_BASE;
    // FIXME: It would be more natural to emit a COPY here, but then copy
    // coalescing would kick in and it would think it's okay to use the "HI"
    // subregister (instead of extracting the HI 32 bits) which is an artificial
    // (unusable) register.
    //  Register TableGen definitions would need an overhaul to get rid of the
    //  artificial "HI" aperture registers and prevent this kind of issue from
    //  happening.
    Register Dst = MRI.createGenericVirtualRegister(S64);
    MRI.setRegClass(Dst, &AMDGPU::SReg_64RegClass);
    B.buildInstr(AMDGPU::S_MOV_B64, {Dst}, {Register(ApertureRegNo)});
    return B.buildUnmerge(S32, Dst).getReg(1);
  }
 
  // TODO: can we be smarter about machine pointer info?
  MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS);
  Register LoadAddr = MRI.createGenericVirtualRegister(
    LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64));
  // For code object version 5, private_base and shared_base are passed through
  // implicit kernargs.
  if (AMDGPU::getAMDHSACodeObjectVersion(*MF.getFunction().getParent()) >=
      AMDGPU::AMDHSA_COV5) {
    AMDGPUTargetLowering::ImplicitParameter Param =
        AS == AMDGPUAS::LOCAL_ADDRESS ? AMDGPUTargetLowering::SHARED_BASE
                                      : AMDGPUTargetLowering::PRIVATE_BASE;
    uint64_t Offset =
        ST.getTargetLowering()->getImplicitParameterOffset(B.getMF(), Param);
 
    Register KernargPtrReg = MRI.createGenericVirtualRegister(
        LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64));
 
    if (!loadInputValue(KernargPtrReg, B,
                        AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR))
      return Register();
 
    MachineMemOperand *MMO = MF.getMachineMemOperand(
        PtrInfo,
        MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
            MachineMemOperand::MOInvariant,
        LLT::scalar(32), commonAlignment(Align(64), Offset));
 
    // Pointer address
    B.buildPtrAdd(LoadAddr, KernargPtrReg,
                  B.buildConstant(LLT::scalar(64), Offset).getReg(0));
    // Load address
    return B.buildLoad(S32, LoadAddr, *MMO).getReg(0);
  }
 
  Register QueuePtr = MRI.createGenericVirtualRegister(
    LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64));
 
  if (!loadInputValue(QueuePtr, B, AMDGPUFunctionArgInfo::QUEUE_PTR))
    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;
 
  MachineMemOperand *MMO = MF.getMachineMemOperand(
      PtrInfo,
      MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
          MachineMemOperand::MOInvariant,
      LLT::scalar(32), commonAlignment(Align(64), StructOffset));
 
  B.buildPtrAdd(LoadAddr, QueuePtr,
                B.buildConstant(LLT::scalar(64), StructOffset).getReg(0));
  return B.buildLoad(S32, LoadAddr, *MMO).getReg(0);
}
 
/// Return true if the value is a known valid address, such that a null check is
/// not necessary.
static bool isKnownNonNull(Register Val, MachineRegisterInfo &MRI,
                           const AMDGPUTargetMachine &TM, unsigned AddrSpace) {
  MachineInstr *Def = MRI.getVRegDef(Val);
  switch (Def->getOpcode()) {
  case AMDGPU::G_FRAME_INDEX:
  case AMDGPU::G_GLOBAL_VALUE:
  case AMDGPU::G_BLOCK_ADDR:
    return true;
  case AMDGPU::G_CONSTANT: {
    const ConstantInt *CI = Def->getOperand(1).getCImm();
    return CI->getSExtValue() != TM.getNullPointerValue(AddrSpace);
  }
  default:
    return false;
  }
 
  return false;
}
 
bool AMDGPULegalizerInfo::legalizeAddrSpaceCast(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
  MachineFunction &MF = B.getMF();
 
  // MI can either be a G_ADDRSPACE_CAST or a
  // G_INTRINSIC @llvm.amdgcn.addrspacecast.nonnull
  assert(MI.getOpcode() == TargetOpcode::G_ADDRSPACE_CAST ||
         (isa<GIntrinsic>(MI) && cast<GIntrinsic>(MI).getIntrinsicID() ==
                                     Intrinsic::amdgcn_addrspacecast_nonnull));
 
  const LLT S32 = LLT::scalar(32);
  Register Dst = MI.getOperand(0).getReg();
  Register Src = isa<GIntrinsic>(MI) ? MI.getOperand(2).getReg()
                                     : 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());
 
  if (TM.isNoopAddrSpaceCast(SrcAS, DestAS)) {
    MI.setDesc(B.getTII().get(TargetOpcode::G_BITCAST));
    return true;
  }
 
  if (SrcAS == AMDGPUAS::FLAT_ADDRESS &&
      (DestAS == AMDGPUAS::LOCAL_ADDRESS ||
       DestAS == AMDGPUAS::PRIVATE_ADDRESS)) {
    // For llvm.amdgcn.addrspacecast.nonnull we can always assume non-null, for
    // G_ADDRSPACE_CAST we need to guess.
    if (isa<GIntrinsic>(MI) || isKnownNonNull(Src, MRI, TM, SrcAS)) {
      // Extract low 32-bits of the pointer.
      B.buildExtract(Dst, Src, 0);
      MI.eraseFromParent();
      return true;
    }
 
    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 (DestAS == AMDGPUAS::FLAT_ADDRESS &&
      (SrcAS == AMDGPUAS::LOCAL_ADDRESS ||
       SrcAS == AMDGPUAS::PRIVATE_ADDRESS)) {
    Register ApertureReg = getSegmentAperture(SrcAS, MRI, B);
    if (!ApertureReg.isValid())
      return false;
 
    // 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.buildMergeLikeInstr(DstTy, {SrcAsInt, ApertureReg});
 
    // For llvm.amdgcn.addrspacecast.nonnull we can always assume non-null, for
    // G_ADDRSPACE_CAST we need to guess.
    if (isa<GIntrinsic>(MI) || isKnownNonNull(Src, MRI, TM, SrcAS)) {
      B.buildCopy(Dst, BuildPtr);
      MI.eraseFromParent();
      return true;
    }
 
    auto SegmentNull = B.buildConstant(SrcTy, TM.getNullPointerValue(SrcAS));
    auto FlatNull = B.buildConstant(DstTy, TM.getNullPointerValue(DestAS));
 
    auto CmpRes = B.buildICmp(CmpInst::ICMP_NE, LLT::scalar(1), Src,
                              SegmentNull.getReg(0));
 
    B.buildSelect(Dst, CmpRes, BuildPtr, FlatNull);
 
    MI.eraseFromParent();
    return true;
  }
 
  if (DestAS == AMDGPUAS::CONSTANT_ADDRESS_32BIT &&
      SrcTy.getSizeInBits() == 64) {
    // Truncate.
    B.buildExtract(Dst, Src, 0);
    MI.eraseFromParent();
    return true;
  }
 
  if (SrcAS == AMDGPUAS::CONSTANT_ADDRESS_32BIT &&
      DstTy.getSizeInBits() == 64) {
    const SIMachineFunctionInfo *Info = MF.getInfo<SIMachineFunctionInfo>();
    uint32_t AddrHiVal = Info->get32BitAddressHighBits();
    auto PtrLo = B.buildPtrToInt(S32, Src);
    auto HighAddr = B.buildConstant(S32, AddrHiVal);
    B.buildMergeLikeInstr(Dst, {PtrLo, HighAddr});
    MI.eraseFromParent();
    return true;
  }
 
  DiagnosticInfoUnsupported InvalidAddrSpaceCast(
      MF.getFunction(), "invalid addrspacecast", B.getDebugLoc());
 
  LLVMContext &Ctx = MF.getFunction().getContext();
  Ctx.diagnose(InvalidAddrSpaceCast);
  B.buildUndef(Dst);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFroundeven(MachineInstr &MI,
                                             MachineRegisterInfo &MRI,
                                             MachineIRBuilder &B) const {
  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);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFceil(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
 
  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);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFrem(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
    Register DstReg = MI.getOperand(0).getReg();
    Register Src0Reg = MI.getOperand(1).getReg();
    Register Src1Reg = MI.getOperand(2).getReg();
    auto Flags = MI.getFlags();
    LLT Ty = MRI.getType(DstReg);
 
    auto Div = B.buildFDiv(Ty, Src0Reg, Src1Reg, Flags);
    auto Trunc = B.buildIntrinsicTrunc(Ty, Div, Flags);
    auto Neg = B.buildFNeg(Ty, Trunc, Flags);
    B.buildFMA(DstReg, Neg, Src1Reg, Src0Reg, Flags);
    MI.eraseFromParent();
    return true;
}
 
static MachineInstrBuilder extractF64Exponent(Register 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})
                     .addUse(Hi)
                     .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 {
  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.buildMergeLikeInstr(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);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeITOFP(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B, bool Signed) const {
 
  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);
 
  auto Unmerge = B.buildUnmerge({S32, S32}, Src);
  auto ThirtyTwo = B.buildConstant(S32, 32);
 
  if (MRI.getType(Dst) == S64) {
    auto CvtHi = Signed ? B.buildSITOFP(S64, Unmerge.getReg(1))
                        : B.buildUITOFP(S64, Unmerge.getReg(1));
 
    auto CvtLo = B.buildUITOFP(S64, Unmerge.getReg(0));
    auto LdExp = B.buildFLdexp(S64, CvtHi, ThirtyTwo);
 
    // TODO: Should this propagate fast-math-flags?
    B.buildFAdd(Dst, LdExp, CvtLo);
    MI.eraseFromParent();
    return true;
  }
 
  assert(MRI.getType(Dst) == S32);
 
  auto One = B.buildConstant(S32, 1);
 
  MachineInstrBuilder ShAmt;
  if (Signed) {
    auto ThirtyOne = B.buildConstant(S32, 31);
    auto X = B.buildXor(S32, Unmerge.getReg(0), Unmerge.getReg(1));
    auto OppositeSign = B.buildAShr(S32, X, ThirtyOne);
    auto MaxShAmt = B.buildAdd(S32, ThirtyTwo, OppositeSign);
    auto LS = B.buildIntrinsic(Intrinsic::amdgcn_sffbh, {S32})
                  .addUse(Unmerge.getReg(1));
    auto LS2 = B.buildSub(S32, LS, One);
    ShAmt = B.buildUMin(S32, LS2, MaxShAmt);
  } else
    ShAmt = B.buildCTLZ(S32, Unmerge.getReg(1));
  auto Norm = B.buildShl(S64, Src, ShAmt);
  auto Unmerge2 = B.buildUnmerge({S32, S32}, Norm);
  auto Adjust = B.buildUMin(S32, One, Unmerge2.getReg(0));
  auto Norm2 = B.buildOr(S32, Unmerge2.getReg(1), Adjust);
  auto FVal = Signed ? B.buildSITOFP(S32, Norm2) : B.buildUITOFP(S32, Norm2);
  auto Scale = B.buildSub(S32, ThirtyTwo, ShAmt);
  B.buildFLdexp(Dst, FVal, Scale);
  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 {
 
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();
 
  const LLT S64 = LLT::scalar(64);
  const LLT S32 = LLT::scalar(32);
 
  const LLT SrcLT = MRI.getType(Src);
  assert((SrcLT == S32 || SrcLT == S64) && MRI.getType(Dst) == S64);
 
  unsigned Flags = MI.getFlags();
 
  // The basic idea of converting a floating point number into a pair of 32-bit
  // integers is illustrated as follows:
  //
  //     tf := trunc(val);
  //    hif := floor(tf * 2^-32);
  //    lof := tf - hif * 2^32; // lof is always positive due to floor.
  //     hi := fptoi(hif);
  //     lo := fptoi(lof);
  //
  auto Trunc = B.buildIntrinsicTrunc(SrcLT, Src, Flags);
  MachineInstrBuilder Sign;
  if (Signed && SrcLT == S32) {
    // However, a 32-bit floating point number has only 23 bits mantissa and
    // it's not enough to hold all the significant bits of `lof` if val is
    // negative. To avoid the loss of precision, We need to take the absolute
    // value after truncating and flip the result back based on the original
    // signedness.
    Sign = B.buildAShr(S32, Src, B.buildConstant(S32, 31));
    Trunc = B.buildFAbs(S32, Trunc, Flags);
  }
  MachineInstrBuilder K0, K1;
  if (SrcLT == S64) {
    K0 = B.buildFConstant(
        S64, llvm::bit_cast<double>(UINT64_C(/*2^-32*/ 0x3df0000000000000)));
    K1 = B.buildFConstant(
        S64, llvm::bit_cast<double>(UINT64_C(/*-2^32*/ 0xc1f0000000000000)));
  } else {
    K0 = B.buildFConstant(
        S32, llvm::bit_cast<float>(UINT32_C(/*2^-32*/ 0x2f800000)));
    K1 = B.buildFConstant(
        S32, llvm::bit_cast<float>(UINT32_C(/*-2^32*/ 0xcf800000)));
  }
 
  auto Mul = B.buildFMul(SrcLT, Trunc, K0, Flags);
  auto FloorMul = B.buildFFloor(SrcLT, Mul, Flags);
  auto Fma = B.buildFMA(SrcLT, FloorMul, K1, Trunc, Flags);
 
  auto Hi = (Signed && SrcLT == S64) ? B.buildFPTOSI(S32, FloorMul)
                                     : B.buildFPTOUI(S32, FloorMul);
  auto Lo = B.buildFPTOUI(S32, Fma);
 
  if (Signed && SrcLT == S32) {
    // Flip the result based on the signedness, which is either all 0s or 1s.
    Sign = B.buildMergeLikeInstr(S64, {Sign, Sign});
    // r := xor({lo, hi}, sign) - sign;
    B.buildSub(Dst, B.buildXor(S64, B.buildMergeLikeInstr(S64, {Lo, Hi}), Sign),
               Sign);
  } else
    B.buildMergeLikeInstr(Dst, {Lo, Hi});
  MI.eraseFromParent();
 
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeMinNumMaxNum(LegalizerHelper &Helper,
                                               MachineInstr &MI) const {
  MachineFunction &MF = Helper.MIRBuilder.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;
 
  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
 
  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));
 
  // Other legalization maps vector<? x [type bigger than 64 bits]> via bitcasts
  // but we can't go directly to that logic becasue you can't bitcast a vector
  // of pointers to a vector of integers. Therefore, introduce an intermediate
  // vector of integers using ptrtoint (and inttoptr on the output) in order to
  // drive the legalization forward.
  if (EltTy.isPointer() && EltTy.getSizeInBits() > 64) {
    LLT IntTy = LLT::scalar(EltTy.getSizeInBits());
    LLT IntVecTy = VecTy.changeElementType(IntTy);
 
    auto IntVec = B.buildPtrToInt(IntVecTy, Vec);
    auto IntElt = B.buildExtractVectorElement(IntTy, IntVec, MI.getOperand(2));
    B.buildIntToPtr(Dst, IntElt);
 
    MI.eraseFromParent();
    return true;
  }
 
  // FIXME: Artifact combiner probably should have replaced the truncated
  // constant before this, so we shouldn't need
  // getIConstantVRegValWithLookThrough.
  std::optional<ValueAndVReg> MaybeIdxVal =
      getIConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI);
  if (!MaybeIdxVal) // Dynamic case will be selected to register indexing.
    return true;
  const uint64_t IdxVal = MaybeIdxVal->Value.getZExtValue();
 
  if (IdxVal < VecTy.getNumElements()) {
    auto Unmerge = B.buildUnmerge(EltTy, Vec);
    B.buildCopy(Dst, Unmerge.getReg(IdxVal));
  } 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
 
  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));
 
  // Other legalization maps vector<? x [type bigger than 64 bits]> via bitcasts
  // but we can't go directly to that logic becasue you can't bitcast a vector
  // of pointers to a vector of integers. Therefore, make the pointer vector
  // into an equivalent vector of integers with ptrtoint, insert the ptrtoint'd
  // new value, and then inttoptr the result vector back. This will then allow
  // the rest of legalization to take over.
  if (EltTy.isPointer() && EltTy.getSizeInBits() > 64) {
    LLT IntTy = LLT::scalar(EltTy.getSizeInBits());
    LLT IntVecTy = VecTy.changeElementType(IntTy);
 
    auto IntVecSource = B.buildPtrToInt(IntVecTy, Vec);
    auto IntIns = B.buildPtrToInt(IntTy, Ins);
    auto IntVecDest = B.buildInsertVectorElement(IntVecTy, IntVecSource, IntIns,
                                                 MI.getOperand(3));
    B.buildIntToPtr(Dst, IntVecDest);
    MI.eraseFromParent();
    return true;
  }
 
  // FIXME: Artifact combiner probably should have replaced the truncated
  // constant before this, so we shouldn't need
  // getIConstantVRegValWithLookThrough.
  std::optional<ValueAndVReg> MaybeIdxVal =
      getIConstantVRegValWithLookThrough(MI.getOperand(3).getReg(), MRI);
  if (!MaybeIdxVal) // Dynamic case will be selected to register indexing.
    return true;
 
  const uint64_t IdxVal = MaybeIdxVal->Value.getZExtValue();
 
  unsigned NumElts = VecTy.getNumElements();
  if (IdxVal < NumElts) {
    SmallVector<Register, 8> SrcRegs;
    for (unsigned i = 0; i < NumElts; ++i)
      SrcRegs.push_back(MRI.createGenericVirtualRegister(EltTy));
    B.buildUnmerge(SrcRegs, Vec);
 
    SrcRegs[IdxVal] = MI.getOperand(2).getReg();
    B.buildMergeLikeInstr(Dst, SrcRegs);
  } else {
    B.buildUndef(Dst);
  }
 
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeSinCos(
  MachineInstr &MI, MachineRegisterInfo &MRI,
  MachineIRBuilder &B) const {
 
  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 * numbers::inv_pi);
  if (ST.hasTrigReducedRange()) {
    auto MulVal = B.buildFMul(Ty, SrcReg, OneOver2Pi, Flags);
    TrigVal = B.buildIntrinsic(Intrinsic::amdgcn_fract, {Ty})
                  .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, ArrayRef<Register>(DstReg))
      .addUse(TrigVal)
      .setMIFlags(Flags);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::buildPCRelGlobalAddress(Register DstReg, LLT PtrTy,
                                                  MachineIRBuilder &B,
                                                  const GlobalValue *GV,
                                                  int64_t Offset,
                                                  unsigned GAFlags) const {
  assert(isInt<32>(Offset + 4) && "32-bit offset is expected!");
  // 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.
 
  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, GAFlags);
  if (GAFlags == SIInstrInfo::MO_NONE)
    MIB.addImm(0);
  else
    MIB.addGlobalAddress(GV, Offset, GAFlags + 1);
 
  if (!B.getMRI()->getRegClassOrNull(PCReg))
    B.getMRI()->setRegClass(PCReg, &AMDGPU::SReg_64RegClass);
 
  if (PtrTy.getSizeInBits() == 32)
    B.buildExtract(DstReg, PCReg, 0);
  return true;
}
 
// Emit a ABS32_LO / ABS32_HI relocation stub.
void AMDGPULegalizerInfo::buildAbsGlobalAddress(
    Register DstReg, LLT PtrTy, MachineIRBuilder &B, const GlobalValue *GV,
    MachineRegisterInfo &MRI) const {
  bool RequiresHighHalf = PtrTy.getSizeInBits() != 32;
 
  LLT S32 = LLT::scalar(32);
 
  // Use the destination directly, if and only if we store the lower address
  // part only and we don't have a register class being set.
  Register AddrLo = !RequiresHighHalf && !MRI.getRegClassOrNull(DstReg)
                        ? DstReg
                        : MRI.createGenericVirtualRegister(S32);
 
  if (!MRI.getRegClassOrNull(AddrLo))
    MRI.setRegClass(AddrLo, &AMDGPU::SReg_32RegClass);
 
  // Write the lower half.
  B.buildInstr(AMDGPU::S_MOV_B32)
      .addDef(AddrLo)
      .addGlobalAddress(GV, 0, SIInstrInfo::MO_ABS32_LO);
 
  // If required, write the upper half as well.
  if (RequiresHighHalf) {
    assert(PtrTy.getSizeInBits() == 64 &&
           "Must provide a 64-bit pointer type!");
 
    Register AddrHi = MRI.createGenericVirtualRegister(S32);
    MRI.setRegClass(AddrHi, &AMDGPU::SReg_32RegClass);
 
    B.buildInstr(AMDGPU::S_MOV_B32)
        .addDef(AddrHi)
        .addGlobalAddress(GV, 0, SIInstrInfo::MO_ABS32_HI);
 
    // Use the destination directly, if and only if we don't have a register
    // class being set.
    Register AddrDst = !MRI.getRegClassOrNull(DstReg)
                           ? DstReg
                           : MRI.createGenericVirtualRegister(LLT::scalar(64));
 
    if (!MRI.getRegClassOrNull(AddrDst))
      MRI.setRegClass(AddrDst, &AMDGPU::SReg_64RegClass);
 
    B.buildMergeValues(AddrDst, {AddrLo, AddrHi});
 
    // If we created a new register for the destination, cast the result into
    // the final output.
    if (AddrDst != DstReg)
      B.buildCast(DstReg, AddrDst);
  } else if (AddrLo != DstReg) {
    // If we created a new register for the destination, cast the result into
    // the final output.
    B.buildCast(DstReg, AddrLo);
  }
}
 
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>();
 
  if (AS == AMDGPUAS::LOCAL_ADDRESS || AS == AMDGPUAS::REGION_ADDRESS) {
    if (!MFI->isModuleEntryFunction() &&
        GV->getName() != "llvm.amdgcn.module.lds") {
      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.buildTrap();
      B.buildUndef(DstReg);
      MI.eraseFromParent();
      return true;
    }
 
    // TODO: We could emit code to handle the initialization somewhere.
    // We ignore the initializer for now and legalize it to allow selection.
    // The initializer will anyway get errored out during assembly emission.
    const SITargetLowering *TLI = ST.getTargetLowering();
    if (!TLI->shouldUseLDSConstAddress(GV)) {
      MI.getOperand(1).setTargetFlags(SIInstrInfo::MO_ABS32_LO);
      return true; // Leave in place;
    }
 
    if (AS == AMDGPUAS::LOCAL_ADDRESS && GV->hasExternalLinkage()) {
      Type *Ty = GV->getValueType();
      // HIP uses an unsized array `extern __shared__ T s[]` or similar
      // zero-sized type in other languages to declare the dynamic shared
      // memory which size is not known at the compile time. They will be
      // allocated by the runtime and placed directly after the static
      // allocated ones. They all share the same offset.
      if (B.getDataLayout().getTypeAllocSize(Ty).isZero()) {
        // Adjust alignment for that dynamic shared memory array.
        MFI->setDynLDSAlign(MF.getFunction(), *cast<GlobalVariable>(GV));
        LLT S32 = LLT::scalar(32);
        auto Sz = B.buildIntrinsic(Intrinsic::amdgcn_groupstaticsize, {S32});
        B.buildIntToPtr(DstReg, Sz);
        MI.eraseFromParent();
        return true;
      }
    }
 
    B.buildConstant(DstReg, MFI->allocateLDSGlobal(B.getDataLayout(),
                                                   *cast<GlobalVariable>(GV)));
    MI.eraseFromParent();
    return true;
  }
 
  if (ST.isAmdPalOS() || ST.isMesa3DOS()) {
    buildAbsGlobalAddress(DstReg, Ty, B, GV, MRI);
    MI.eraseFromParent();
    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);
 
  LLT LoadTy = Ty.getSizeInBits() == 32 ? PtrTy : Ty;
  MachineMemOperand *GOTMMO = MF.getMachineMemOperand(
      MachinePointerInfo::getGOT(MF),
      MachineMemOperand::MOLoad | MachineMemOperand::MODereferenceable |
          MachineMemOperand::MOInvariant,
      LoadTy, Align(8));
 
  buildPCRelGlobalAddress(GOTAddr, PtrTy, B, GV, 0, SIInstrInfo::MO_GOTPCREL32);
 
  if (Ty.getSizeInBits() == 32) {
    // Truncate if this is a 32-bit constant address.
    auto Load = B.buildLoad(PtrTy, GOTAddr, *GOTMMO);
    B.buildExtract(DstReg, Load, 0);
  } else
    B.buildLoad(DstReg, GOTAddr, *GOTMMO);
 
  MI.eraseFromParent();
  return true;
}
 
static LLT widenToNextPowerOf2(LLT Ty) {
  if (Ty.isVector())
    return Ty.changeElementCount(
        ElementCount::getFixed(PowerOf2Ceil(Ty.getNumElements())));
  return LLT::scalar(PowerOf2Ceil(Ty.getSizeInBits()));
}
 
bool AMDGPULegalizerInfo::legalizeLoad(LegalizerHelper &Helper,
                                       MachineInstr &MI) const {
  MachineIRBuilder &B = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *B.getMRI();
  GISelChangeObserver &Observer = Helper.Observer;
 
  Register PtrReg = MI.getOperand(1).getReg();
  LLT PtrTy = MRI.getType(PtrReg);
  unsigned AddrSpace = PtrTy.getAddressSpace();
 
  if (AddrSpace == AMDGPUAS::CONSTANT_ADDRESS_32BIT) {
    LLT ConstPtr = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
    auto Cast = B.buildAddrSpaceCast(ConstPtr, PtrReg);
    Observer.changingInstr(MI);
    MI.getOperand(1).setReg(Cast.getReg(0));
    Observer.changedInstr(MI);
    return true;
  }
 
  if (MI.getOpcode() != AMDGPU::G_LOAD)
    return false;
 
  Register ValReg = MI.getOperand(0).getReg();
  LLT ValTy = MRI.getType(ValReg);
 
  if (hasBufferRsrcWorkaround(ValTy)) {
    Observer.changingInstr(MI);
    castBufferRsrcFromV4I32(MI, B, MRI, 0);
    Observer.changedInstr(MI);
    return true;
  }
 
  MachineMemOperand *MMO = *MI.memoperands_begin();
  const unsigned ValSize = ValTy.getSizeInBits();
  const LLT MemTy = MMO->getMemoryType();
  const Align MemAlign = MMO->getAlign();
  const unsigned MemSize = MemTy.getSizeInBits();
  const uint64_t AlignInBits = 8 * MemAlign.value();
 
  // Widen non-power-of-2 loads to the alignment if needed
  if (shouldWidenLoad(ST, MemTy, AlignInBits, AddrSpace, MI.getOpcode())) {
    const unsigned WideMemSize = PowerOf2Ceil(MemSize);
 
    // This was already the correct extending load result type, so just adjust
    // the memory type.
    if (WideMemSize == ValSize) {
      MachineFunction &MF = B.getMF();
 
      MachineMemOperand *WideMMO =
          MF.getMachineMemOperand(MMO, 0, WideMemSize / 8);
      Observer.changingInstr(MI);
      MI.setMemRefs(MF, {WideMMO});
      Observer.changedInstr(MI);
      return true;
    }
 
    // Don't bother handling edge case that should probably never be produced.
    if (ValSize > WideMemSize)
      return false;
 
    LLT WideTy = widenToNextPowerOf2(ValTy);
 
    Register WideLoad;
    if (!WideTy.isVector()) {
      WideLoad = B.buildLoadFromOffset(WideTy, PtrReg, *MMO, 0).getReg(0);
      B.buildTrunc(ValReg, WideLoad).getReg(0);
    } else {
      // Extract the subvector.
 
      if (isRegisterType(ValTy)) {
        // If this a case where G_EXTRACT is legal, use it.
        // (e.g. <3 x s32> -> <4 x s32>)
        WideLoad = B.buildLoadFromOffset(WideTy, PtrReg, *MMO, 0).getReg(0);
        B.buildExtract(ValReg, WideLoad, 0);
      } else {
        // For cases where the widened type isn't a nice register value, unmerge
        // from a widened register (e.g. <3 x s16> -> <4 x s16>)
        WideLoad = B.buildLoadFromOffset(WideTy, PtrReg, *MMO, 0).getReg(0);
        B.buildDeleteTrailingVectorElements(ValReg, WideLoad);
      }
    }
 
    MI.eraseFromParent();
    return true;
  }
 
  return false;
}
 
bool AMDGPULegalizerInfo::legalizeStore(LegalizerHelper &Helper,
                                        MachineInstr &MI) const {
  MachineIRBuilder &B = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *B.getMRI();
  GISelChangeObserver &Observer = Helper.Observer;
 
  Register DataReg = MI.getOperand(0).getReg();
  LLT DataTy = MRI.getType(DataReg);
 
  if (hasBufferRsrcWorkaround(DataTy)) {
    Observer.changingInstr(MI);
    castBufferRsrcArgToV4I32(MI, B, 0);
    Observer.changedInstr(MI);
    return true;
  }
  return false;
}
 
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::float32() &&
      MFI->getMode().FP32Denormals == DenormalMode::getPreserveSign())
    return true;
  if (Ty == LLT::float16() &&
      MFI->getMode().FP64FP16Denormals == DenormalMode::getPreserveSign())
    return true;
 
  MachineIRBuilder HelperBuilder(MI);
  GISelObserverWrapper DummyObserver;
  LegalizerHelper Helper(MF, DummyObserver, HelperBuilder);
  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(AMDGPU::isFlatGlobalAddrSpace(MRI.getType(PtrReg).getAddressSpace()) &&
         "this should not have been custom lowered");
 
  LLT ValTy = MRI.getType(CmpVal);
  LLT VecTy = LLT::fixed_vector(2, ValTy);
 
  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;
}
 
/// Return true if it's known that \p Src can never be an f32 denormal value.
static bool valueIsKnownNeverF32Denorm(const MachineRegisterInfo &MRI,
                                       Register Src) {
  const MachineInstr *DefMI = MRI.getVRegDef(Src);
  switch (DefMI->getOpcode()) {
  case TargetOpcode::G_INTRINSIC: {
    switch (cast<GIntrinsic>(DefMI)->getIntrinsicID()) {
    case Intrinsic::amdgcn_frexp_mant:
      return true;
    default:
      break;
    }
 
    break;
  }
  case TargetOpcode::G_FFREXP: {
    if (DefMI->getOperand(0).getReg() == Src)
      return true;
    break;
  }
  case TargetOpcode::G_FPEXT: {
    return MRI.getType(DefMI->getOperand(1).getReg()) == LLT::scalar(16);
  }
  default:
    return false;
  }
 
  return false;
}
 
static bool allowApproxFunc(const MachineFunction &MF, unsigned Flags) {
  if (Flags & MachineInstr::FmAfn)
    return true;
  const auto &Options = MF.getTarget().Options;
  return Options.UnsafeFPMath || Options.ApproxFuncFPMath;
}
 
static bool needsDenormHandlingF32(const MachineFunction &MF, Register Src,
                                   unsigned Flags) {
  return !valueIsKnownNeverF32Denorm(MF.getRegInfo(), Src) &&
         MF.getDenormalMode(APFloat::IEEEsingle()).Input !=
             DenormalMode::PreserveSign;
}
 
std::pair<Register, Register>
AMDGPULegalizerInfo::getScaledLogInput(MachineIRBuilder &B, Register Src,
                                       unsigned Flags) const {
  if (!needsDenormHandlingF32(B.getMF(), Src, Flags))
    return {};
 
  const LLT F32 = LLT::scalar(32);
  auto SmallestNormal = B.buildFConstant(
      F32, APFloat::getSmallestNormalized(APFloat::IEEEsingle()));
  auto IsLtSmallestNormal =
      B.buildFCmp(CmpInst::FCMP_OLT, LLT::scalar(1), Src, SmallestNormal);
 
  auto Scale32 = B.buildFConstant(F32, 0x1.0p+32);
  auto One = B.buildFConstant(F32, 1.0);
  auto ScaleFactor =
      B.buildSelect(F32, IsLtSmallestNormal, Scale32, One, Flags);
  auto ScaledInput = B.buildFMul(F32, Src, ScaleFactor, Flags);
 
  return {ScaledInput.getReg(0), IsLtSmallestNormal.getReg(0)};
}
 
bool AMDGPULegalizerInfo::legalizeFlog2(MachineInstr &MI,
                                        MachineIRBuilder &B) const {
  // v_log_f32 is good enough for OpenCL, except it doesn't handle denormals.
  // If we have to handle denormals, scale up the input and adjust the result.
 
  // scaled = x * (is_denormal ? 0x1.0p+32 : 1.0)
  // log2 = amdgpu_log2 - (is_denormal ? 32.0 : 0.0)
 
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();
  LLT Ty = B.getMRI()->getType(Dst);
  unsigned Flags = MI.getFlags();
 
  if (Ty == LLT::scalar(16)) {
    const LLT F32 = LLT::scalar(32);
    // Nothing in half is a denormal when promoted to f32.
    auto Ext = B.buildFPExt(F32, Src, Flags);
    auto Log2 = B.buildIntrinsic(Intrinsic::amdgcn_log, {F32})
                    .addUse(Ext.getReg(0))
                    .setMIFlags(Flags);
    B.buildFPTrunc(Dst, Log2, Flags);
    MI.eraseFromParent();
    return true;
  }
 
  assert(Ty == LLT::scalar(32));
 
  auto [ScaledInput, IsLtSmallestNormal] = getScaledLogInput(B, Src, Flags);
  if (!ScaledInput) {
    B.buildIntrinsic(Intrinsic::amdgcn_log, {MI.getOperand(0)})
        .addUse(Src)
        .setMIFlags(Flags);
    MI.eraseFromParent();
    return true;
  }
 
  auto Log2 = B.buildIntrinsic(Intrinsic::amdgcn_log, {Ty})
                  .addUse(ScaledInput)
                  .setMIFlags(Flags);
 
  auto ThirtyTwo = B.buildFConstant(Ty, 32.0);
  auto Zero = B.buildFConstant(Ty, 0.0);
  auto ResultOffset =
      B.buildSelect(Ty, IsLtSmallestNormal, ThirtyTwo, Zero, Flags);
  B.buildFSub(Dst, Log2, ResultOffset, Flags);
 
  MI.eraseFromParent();
  return true;
}
 
static Register getMad(MachineIRBuilder &B, LLT Ty, Register X, Register Y,
                       Register Z, unsigned Flags) {
  auto FMul = B.buildFMul(Ty, X, Y, Flags);
  return B.buildFAdd(Ty, FMul, Z, Flags).getReg(0);
}
 
bool AMDGPULegalizerInfo::legalizeFlogCommon(MachineInstr &MI,
                                             MachineIRBuilder &B) const {
  const bool IsLog10 = MI.getOpcode() == TargetOpcode::G_FLOG10;
  assert(IsLog10 || MI.getOpcode() == TargetOpcode::G_FLOG);
 
  MachineRegisterInfo &MRI = *B.getMRI();
  Register Dst = MI.getOperand(0).getReg();
  Register X = MI.getOperand(1).getReg();
  unsigned Flags = MI.getFlags();
  const LLT Ty = MRI.getType(X);
  MachineFunction &MF = B.getMF();
 
  const LLT F32 = LLT::scalar(32);
  const LLT F16 = LLT::scalar(16);
 
  const AMDGPUTargetMachine &TM =
      static_cast<const AMDGPUTargetMachine &>(MF.getTarget());
 
  if (Ty == F16 || MI.getFlag(MachineInstr::FmAfn) ||
      TM.Options.ApproxFuncFPMath || TM.Options.UnsafeFPMath) {
    if (Ty == F16 && !ST.has16BitInsts()) {
      Register LogVal = MRI.createGenericVirtualRegister(F32);
      auto PromoteSrc = B.buildFPExt(F32, X);
      legalizeFlogUnsafe(B, LogVal, PromoteSrc.getReg(0), IsLog10, Flags);
      B.buildFPTrunc(Dst, LogVal);
    } else {
      legalizeFlogUnsafe(B, Dst, X, IsLog10, Flags);
    }
 
    MI.eraseFromParent();
    return true;
  }
 
  auto [ScaledInput, IsScaled] = getScaledLogInput(B, X, Flags);
  if (ScaledInput)
    X = ScaledInput;
 
  auto Y =
      B.buildIntrinsic(Intrinsic::amdgcn_log, {Ty}).addUse(X).setMIFlags(Flags);
 
  Register R;
  if (ST.hasFastFMAF32()) {
    // c+cc are ln(2)/ln(10) to more than 49 bits
    const float c_log10 = 0x1.344134p-2f;
    const float cc_log10 = 0x1.09f79ep-26f;
 
    // c + cc is ln(2) to more than 49 bits
    const float c_log = 0x1.62e42ep-1f;
    const float cc_log = 0x1.efa39ep-25f;
 
    auto C = B.buildFConstant(Ty, IsLog10 ? c_log10 : c_log);
    auto CC = B.buildFConstant(Ty, IsLog10 ? cc_log10 : cc_log);
 
    R = B.buildFMul(Ty, Y, C, Flags).getReg(0);
    auto NegR = B.buildFNeg(Ty, R, Flags);
    auto FMA0 = B.buildFMA(Ty, Y, C, NegR, Flags);
    auto FMA1 = B.buildFMA(Ty, Y, CC, FMA0, Flags);
    R = B.buildFAdd(Ty, R, FMA1, Flags).getReg(0);
  } else {
    // ch+ct is ln(2)/ln(10) to more than 36 bits
    const float ch_log10 = 0x1.344000p-2f;
    const float ct_log10 = 0x1.3509f6p-18f;
 
    // ch + ct is ln(2) to more than 36 bits
    const float ch_log = 0x1.62e000p-1f;
    const float ct_log = 0x1.0bfbe8p-15f;
 
    auto CH = B.buildFConstant(Ty, IsLog10 ? ch_log10 : ch_log);
    auto CT = B.buildFConstant(Ty, IsLog10 ? ct_log10 : ct_log);
 
    auto MaskConst = B.buildConstant(Ty, 0xfffff000);
    auto YH = B.buildAnd(Ty, Y, MaskConst);
    auto YT = B.buildFSub(Ty, Y, YH, Flags);
    auto YTCT = B.buildFMul(Ty, YT, CT, Flags);
 
    Register Mad0 =
        getMad(B, Ty, YH.getReg(0), CT.getReg(0), YTCT.getReg(0), Flags);
    Register Mad1 = getMad(B, Ty, YT.getReg(0), CH.getReg(0), Mad0, Flags);
    R = getMad(B, Ty, YH.getReg(0), CH.getReg(0), Mad1, Flags);
  }
 
  const bool IsFiniteOnly =
      (MI.getFlag(MachineInstr::FmNoNans) || TM.Options.NoNaNsFPMath) &&
      (MI.getFlag(MachineInstr::FmNoInfs) || TM.Options.NoInfsFPMath);
 
  if (!IsFiniteOnly) {
    // Expand isfinite(x) => fabs(x) < inf
    auto Inf = B.buildFConstant(Ty, APFloat::getInf(APFloat::IEEEsingle()));
    auto Fabs = B.buildFAbs(Ty, Y);
    auto IsFinite =
        B.buildFCmp(CmpInst::FCMP_OLT, LLT::scalar(1), Fabs, Inf, Flags);
    R = B.buildSelect(Ty, IsFinite, R, Y, Flags).getReg(0);
  }
 
  if (ScaledInput) {
    auto Zero = B.buildFConstant(Ty, 0.0);
    auto ShiftK =
        B.buildFConstant(Ty, IsLog10 ? 0x1.344136p+3f : 0x1.62e430p+4f);
    auto Shift = B.buildSelect(Ty, IsScaled, ShiftK, Zero, Flags);
    B.buildFSub(Dst, R, Shift, Flags);
  } else {
    B.buildCopy(Dst, R);
  }
 
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFlogUnsafe(MachineIRBuilder &B, Register Dst,
                                             Register Src, bool IsLog10,
                                             unsigned Flags) const {
  const double Log2BaseInverted =
      IsLog10 ? numbers::ln2 / numbers::ln10 : numbers::ln2;
 
  LLT Ty = B.getMRI()->getType(Dst);
 
  if (Ty == LLT::scalar(32)) {
    auto [ScaledInput, IsScaled] = getScaledLogInput(B, Src, Flags);
    if (ScaledInput) {
      auto LogSrc = B.buildIntrinsic(Intrinsic::amdgcn_log, {Ty})
                        .addUse(Src)
                        .setMIFlags(Flags);
      auto ScaledResultOffset = B.buildFConstant(Ty, -32.0 * Log2BaseInverted);
      auto Zero = B.buildFConstant(Ty, 0.0);
      auto ResultOffset =
          B.buildSelect(Ty, IsScaled, ScaledResultOffset, Zero, Flags);
      auto Log2Inv = B.buildFConstant(Ty, Log2BaseInverted);
 
      if (ST.hasFastFMAF32())
        B.buildFMA(Dst, LogSrc, Log2Inv, ResultOffset, Flags);
      else {
        auto Mul = B.buildFMul(Ty, LogSrc, Log2Inv, Flags);
        B.buildFAdd(Dst, Mul, ResultOffset, Flags);
      }
 
      return true;
    }
  }
 
  auto Log2Operand = Ty == LLT::scalar(16)
                         ? B.buildFLog2(Ty, Src, Flags)
                         : B.buildIntrinsic(Intrinsic::amdgcn_log, {Ty})
                               .addUse(Src)
                               .setMIFlags(Flags);
  auto Log2BaseInvertedOperand = B.buildFConstant(Ty, Log2BaseInverted);
  B.buildFMul(Dst, Log2Operand, Log2BaseInvertedOperand, Flags);
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFExp2(MachineInstr &MI,
                                        MachineIRBuilder &B) const {
  // v_exp_f32 is good enough for OpenCL, except it doesn't handle denormals.
  // If we have to handle denormals, scale up the input and adjust the result.
 
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();
  unsigned Flags = MI.getFlags();
  LLT Ty = B.getMRI()->getType(Dst);
  const LLT F16 = LLT::scalar(16);
  const LLT F32 = LLT::scalar(32);
 
  if (Ty == F16) {
    // Nothing in half is a denormal when promoted to f32.
    auto Ext = B.buildFPExt(F32, Src, Flags);
    auto Log2 = B.buildIntrinsic(Intrinsic::amdgcn_exp2, {F32})
                    .addUse(Ext.getReg(0))
                    .setMIFlags(Flags);
    B.buildFPTrunc(Dst, Log2, Flags);
    MI.eraseFromParent();
    return true;
  }
 
  assert(Ty == F32);
 
  if (!needsDenormHandlingF32(B.getMF(), Src, Flags)) {
    B.buildIntrinsic(Intrinsic::amdgcn_exp2, ArrayRef<Register>{Dst})
        .addUse(Src)
        .setMIFlags(Flags);
    MI.eraseFromParent();
    return true;
  }
 
  // bool needs_scaling = x < -0x1.f80000p+6f;
  // v_exp_f32(x + (s ? 0x1.0p+6f : 0.0f)) * (s ? 0x1.0p-64f : 1.0f);
 
  // -nextafter(128.0, -1)
  auto RangeCheckConst = B.buildFConstant(Ty, -0x1.f80000p+6f);
  auto NeedsScaling = B.buildFCmp(CmpInst::FCMP_OLT, LLT::scalar(1), Src,
                                  RangeCheckConst, Flags);
 
  auto SixtyFour = B.buildFConstant(Ty, 0x1.0p+6f);
  auto Zero = B.buildFConstant(Ty, 0.0);
  auto AddOffset = B.buildSelect(F32, NeedsScaling, SixtyFour, Zero, Flags);
  auto AddInput = B.buildFAdd(F32, Src, AddOffset, Flags);
 
  auto Exp2 = B.buildIntrinsic(Intrinsic::amdgcn_exp2, {Ty})
                  .addUse(AddInput.getReg(0))
                  .setMIFlags(Flags);
 
  auto TwoExpNeg64 = B.buildFConstant(Ty, 0x1.0p-64f);
  auto One = B.buildFConstant(Ty, 1.0);
  auto ResultScale = B.buildSelect(F32, NeedsScaling, TwoExpNeg64, One, Flags);
  B.buildFMul(Dst, Exp2, ResultScale, Flags);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFExpUnsafe(MachineIRBuilder &B, Register Dst,
                                             Register X, unsigned Flags) const {
  LLT Ty = B.getMRI()->getType(Dst);
  LLT F32 = LLT::scalar(32);
 
  if (Ty != F32 || !needsDenormHandlingF32(B.getMF(), X, Flags)) {
    auto Log2E = B.buildFConstant(Ty, numbers::log2e);
    auto Mul = B.buildFMul(Ty, X, Log2E, Flags);
 
    if (Ty == F32) {
      B.buildIntrinsic(Intrinsic::amdgcn_exp2, ArrayRef<Register>{Dst})
        .addUse(Mul.getReg(0))
        .setMIFlags(Flags);
    } else {
      B.buildFExp2(Dst, Mul.getReg(0), Flags);
    }
 
    return true;
  }
 
  auto Threshold = B.buildFConstant(Ty, -0x1.5d58a0p+6f);
  auto NeedsScaling =
      B.buildFCmp(CmpInst::FCMP_OLT, LLT::scalar(1), X, Threshold, Flags);
  auto ScaleOffset = B.buildFConstant(Ty, 0x1.0p+6f);
  auto ScaledX = B.buildFAdd(Ty, X, ScaleOffset, Flags);
  auto AdjustedX = B.buildSelect(Ty, NeedsScaling, ScaledX, X, Flags);
 
  auto Log2E = B.buildFConstant(Ty, numbers::log2e);
  auto ExpInput = B.buildFMul(Ty, AdjustedX, Log2E, Flags);
 
  auto Exp2 = B.buildIntrinsic(Intrinsic::amdgcn_exp2, {Ty})
    .addUse(ExpInput.getReg(0))
    .setMIFlags(Flags);
 
  auto ResultScaleFactor = B.buildFConstant(Ty, 0x1.969d48p-93f);
  auto AdjustedResult = B.buildFMul(Ty, Exp2, ResultScaleFactor, Flags);
  B.buildSelect(Dst, NeedsScaling, AdjustedResult, Exp2, Flags);
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFExp(MachineInstr &MI,
                                       MachineIRBuilder &B) const {
  Register Dst = MI.getOperand(0).getReg();
  Register X = MI.getOperand(1).getReg();
  const unsigned Flags = MI.getFlags();
  MachineFunction &MF = B.getMF();
  MachineRegisterInfo &MRI = *B.getMRI();
  LLT Ty = MRI.getType(Dst);
  const LLT F16 = LLT::scalar(16);
  const LLT F32 = LLT::scalar(32);
  const bool IsExp10 = MI.getOpcode() == TargetOpcode::G_FEXP10;
 
  if (Ty == F16) {
    // v_exp_f16 (fmul x, log2e)
    if (allowApproxFunc(MF, Flags)) {
      // TODO: Does this really require fast?
      legalizeFExpUnsafe(B, Dst, X, Flags);
      MI.eraseFromParent();
      return true;
    }
 
    // exp(f16 x) ->
    //   fptrunc (v_exp_f32 (fmul (fpext x), log2e))
 
    // Nothing in half is a denormal when promoted to f32.
    auto Ext = B.buildFPExt(F32, X, Flags);
    Register Lowered = MRI.createGenericVirtualRegister(F32);
    legalizeFExpUnsafe(B, Lowered, Ext.getReg(0), Flags);
    B.buildFPTrunc(Dst, Lowered, Flags);
    MI.eraseFromParent();
    return true;
  }
 
  assert(Ty == F32);
 
  // TODO: Interpret allowApproxFunc as ignoring DAZ. This is currently copying
  // library behavior. Also, is known-not-daz source sufficient?
  if (allowApproxFunc(MF, Flags)) {
    legalizeFExpUnsafe(B, Dst, X, Flags);
    MI.eraseFromParent();
    return true;
  }
 
  //    Algorithm:
  //
  //    e^x = 2^(x/ln(2)) = 2^(x*(64/ln(2))/64)
  //
  //    x*(64/ln(2)) = n + f, |f| <= 0.5, n is integer
  //    n = 64*m + j,   0 <= j < 64
  //
  //    e^x = 2^((64*m + j + f)/64)
  //        = (2^m) * (2^(j/64)) * 2^(f/64)
  //        = (2^m) * (2^(j/64)) * e^(f*(ln(2)/64))
  //
  //    f = x*(64/ln(2)) - n
  //    r = f*(ln(2)/64) = x - n*(ln(2)/64)
  //
  //    e^x = (2^m) * (2^(j/64)) * e^r
  //
  //    (2^(j/64)) is precomputed
  //
  //    e^r = 1 + r + (r^2)/2! + (r^3)/3! + (r^4)/4! + (r^5)/5!
  //    e^r = 1 + q
  //
  //    q = r + (r^2)/2! + (r^3)/3! + (r^4)/4! + (r^5)/5!
  //
  //    e^x = (2^m) * ( (2^(j/64)) + q*(2^(j/64)) )
  const unsigned FlagsNoContract = Flags & ~MachineInstr::FmContract;
  Register PH, PL;
 
  if (ST.hasFastFMAF32()) {
    const float c_exp = numbers::log2ef;
    const float cc_exp = 0x1.4ae0bep-26f; // c+cc are 49 bits
    const float c_exp10 = 0x1.a934f0p+1f;
    const float cc_exp10 = 0x1.2f346ep-24f;
 
    auto C = B.buildFConstant(Ty, IsExp10 ? c_exp10 : c_exp);
    PH = B.buildFMul(Ty, X, C, Flags).getReg(0);
    auto NegPH = B.buildFNeg(Ty, PH, Flags);
    auto FMA0 = B.buildFMA(Ty, X, C, NegPH, Flags);
 
    auto CC = B.buildFConstant(Ty, IsExp10 ? cc_exp10 : cc_exp);
    PL = B.buildFMA(Ty, X, CC, FMA0, Flags).getReg(0);
  } else {
    const float ch_exp = 0x1.714000p+0f;
    const float cl_exp = 0x1.47652ap-12f; // ch + cl are 36 bits
 
    const float ch_exp10 = 0x1.a92000p+1f;
    const float cl_exp10 = 0x1.4f0978p-11f;
 
    auto MaskConst = B.buildConstant(Ty, 0xfffff000);
    auto XH = B.buildAnd(Ty, X, MaskConst);
    auto XL = B.buildFSub(Ty, X, XH, Flags);
 
    auto CH = B.buildFConstant(Ty, IsExp10 ? ch_exp10 : ch_exp);
    PH = B.buildFMul(Ty, XH, CH, Flags).getReg(0);
 
    auto CL = B.buildFConstant(Ty, IsExp10 ? cl_exp10 : cl_exp);
    auto XLCL = B.buildFMul(Ty, XL, CL, Flags);
 
    Register Mad0 =
        getMad(B, Ty, XL.getReg(0), CH.getReg(0), XLCL.getReg(0), Flags);
    PL = getMad(B, Ty, XH.getReg(0), CL.getReg(0), Mad0, Flags);
  }
 
  auto E = B.buildIntrinsicRoundeven(Ty, PH, Flags);
 
  // It is unsafe to contract this fsub into the PH multiply.
  auto PHSubE = B.buildFSub(Ty, PH, E, FlagsNoContract);
  auto A = B.buildFAdd(Ty, PHSubE, PL, Flags);
  auto IntE = B.buildFPTOSI(LLT::scalar(32), E);
 
  auto Exp2 = B.buildIntrinsic(Intrinsic::amdgcn_exp2, {Ty})
                  .addUse(A.getReg(0))
                  .setMIFlags(Flags);
  auto R = B.buildFLdexp(Ty, Exp2, IntE, Flags);
 
  auto UnderflowCheckConst =
      B.buildFConstant(Ty, IsExp10 ? -0x1.66d3e8p+5f : -0x1.9d1da0p+6f);
  auto Zero = B.buildFConstant(Ty, 0.0);
  auto Underflow =
      B.buildFCmp(CmpInst::FCMP_OLT, LLT::scalar(1), X, UnderflowCheckConst);
 
  R = B.buildSelect(Ty, Underflow, Zero, R);
 
  const auto &Options = MF.getTarget().Options;
 
  if (!(Flags & MachineInstr::FmNoInfs) && !Options.NoInfsFPMath) {
    auto OverflowCheckConst =
        B.buildFConstant(Ty, IsExp10 ? 0x1.344136p+5f : 0x1.62e430p+6f);
 
    auto Overflow =
        B.buildFCmp(CmpInst::FCMP_OGT, LLT::scalar(1), X, OverflowCheckConst);
    auto Inf = B.buildFConstant(Ty, APFloat::getInf(APFloat::IEEEsingle()));
    R = B.buildSelect(Ty, Overflow, Inf, R, Flags);
  }
 
  B.buildCopy(Dst, R);
  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);
  const LLT F16 = LLT::float16();
  const LLT F32 = LLT::float32();
 
  if (Ty == F32) {
    auto Log = B.buildFLog2(F32, Src0, Flags);
    auto Mul = B.buildIntrinsic(Intrinsic::amdgcn_fmul_legacy, {F32})
                   .addUse(Log.getReg(0))
                   .addUse(Src1)
                   .setMIFlags(Flags);
    B.buildFExp2(Dst, Mul, Flags);
  } else if (Ty == F16) {
    // There's no f16 fmul_legacy, so we need to convert for it.
    auto Log = B.buildFLog2(F16, Src0, Flags);
    auto Ext0 = B.buildFPExt(F32, Log, Flags);
    auto Ext1 = B.buildFPExt(F32, Src1, Flags);
    auto Mul = B.buildIntrinsic(Intrinsic::amdgcn_fmul_legacy, {F32})
                   .addUse(Ext0.getReg(0))
                   .addUse(Ext1.getReg(0))
                   .setMIFlags(Flags);
    B.buildFExp2(Dst, B.buildFPTrunc(F16, 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 {
 
  const LLT S1 = LLT::scalar(1);
  const LLT F64 = LLT::float64();
  Register Dst = MI.getOperand(0).getReg();
  Register OrigSrc = MI.getOperand(1).getReg();
  unsigned Flags = MI.getFlags();
  assert(ST.hasFractBug() && MRI.getType(Dst) == F64 &&
         "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, {F64})
                   .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(F64, llvm::bit_cast<double>(0x3fefffffffffffff));
 
  Register Min = MRI.createGenericVirtualRegister(F64);
 
  // 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(F64, IsNan, ModSrc, Min, Flags).getReg(0);
  }
 
  auto NegFract = B.buildFNeg(F64, 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);
  const LLT S16 = LLT::scalar(16);
  assert(MRI.getType(Dst) == LLT::fixed_vector(2, 16));
 
  Register Src0 = MI.getOperand(1).getReg();
  Register Src1 = MI.getOperand(2).getReg();
 
  if (MI.getOpcode() == AMDGPU::G_BUILD_VECTOR_TRUNC) {
    assert(MRI.getType(Src0) == S32);
    Src0 = B.buildTrunc(S16, MI.getOperand(1).getReg()).getReg(0);
    Src1 = B.buildTrunc(S16, MI.getOperand(2).getReg()).getReg(0);
  }
 
  auto Merge = B.buildMergeLikeInstr(S32, {Src0, Src1});
  B.buildBitcast(Dst, Merge);
 
  MI.eraseFromParent();
  return true;
}
 
// Build a big integer multiply or multiply-add using MAD_64_32 instructions.
//
// Source and accumulation registers must all be 32-bits.
//
// TODO: When the multiply is uniform, we should produce a code sequence
// that is better suited to instruction selection on the SALU. Instead of
// the outer loop going over parts of the result, the outer loop should go
// over parts of one of the factors. This should result in instruction
// selection that makes full use of S_ADDC_U32 instructions.
void AMDGPULegalizerInfo::buildMultiply(LegalizerHelper &Helper,
                                        MutableArrayRef<Register> Accum,
                                        ArrayRef<Register> Src0,
                                        ArrayRef<Register> Src1,
                                        bool UsePartialMad64_32,
                                        bool SeparateOddAlignedProducts) const {
  // Use (possibly empty) vectors of S1 registers to represent the set of
  // carries from one pair of positions to the next.
  using Carry = SmallVector<Register, 2>;
 
  MachineIRBuilder &B = Helper.MIRBuilder;
  GISelKnownBits &KB = *Helper.getKnownBits();
 
  const LLT S1 = LLT::scalar(1);
  const LLT S32 = LLT::scalar(32);
  const LLT S64 = LLT::scalar(64);
 
  Register Zero32;
  Register Zero64;
 
  auto getZero32 = [&]() -> Register {
    if (!Zero32)
      Zero32 = B.buildConstant(S32, 0).getReg(0);
    return Zero32;
  };
  auto getZero64 = [&]() -> Register {
    if (!Zero64)
      Zero64 = B.buildConstant(S64, 0).getReg(0);
    return Zero64;
  };
 
  SmallVector<bool, 2> Src0KnownZeros, Src1KnownZeros;
  for (unsigned i = 0; i < Src0.size(); ++i) {
    Src0KnownZeros.push_back(KB.getKnownBits(Src0[i]).isZero());
    Src1KnownZeros.push_back(KB.getKnownBits(Src1[i]).isZero());
  }
 
  // Merge the given carries into the 32-bit LocalAccum, which is modified
  // in-place.
  //
  // Returns the carry-out, which is a single S1 register or null.
  auto mergeCarry =
      [&](Register &LocalAccum, const Carry &CarryIn) -> Register {
        if (CarryIn.empty())
          return Register();
 
        bool HaveCarryOut = true;
        Register CarryAccum;
        if (CarryIn.size() == 1) {
          if (!LocalAccum) {
            LocalAccum = B.buildZExt(S32, CarryIn[0]).getReg(0);
            return Register();
          }
 
          CarryAccum = getZero32();
        } else {
          CarryAccum = B.buildZExt(S32, CarryIn[0]).getReg(0);
          for (unsigned i = 1; i + 1 < CarryIn.size(); ++i) {
            CarryAccum =
                B.buildUAdde(S32, S1, CarryAccum, getZero32(), CarryIn[i])
                    .getReg(0);
          }
 
          if (!LocalAccum) {
            LocalAccum = getZero32();
            HaveCarryOut = false;
          }
        }
 
        auto Add =
            B.buildUAdde(S32, S1, CarryAccum, LocalAccum, CarryIn.back());
        LocalAccum = Add.getReg(0);
        return HaveCarryOut ? Add.getReg(1) : Register();
      };
 
  // Build a multiply-add chain to compute
  //
  //   LocalAccum + (partial products at DstIndex)
  //       + (opportunistic subset of CarryIn)
  //
  // LocalAccum is an array of one or two 32-bit registers that are updated
  // in-place. The incoming registers may be null.
  //
  // In some edge cases, carry-ins can be consumed "for free". In that case,
  // the consumed carry bits are removed from CarryIn in-place.
  auto buildMadChain =
      [&](MutableArrayRef<Register> LocalAccum, unsigned DstIndex, Carry &CarryIn)
          -> Carry {
        assert((DstIndex + 1 < Accum.size() && LocalAccum.size() == 2) ||
               (DstIndex + 1 >= Accum.size() && LocalAccum.size() == 1));
 
        Carry CarryOut;
        unsigned j0 = 0;
 
        // Use plain 32-bit multiplication for the most significant part of the
        // result by default.
        if (LocalAccum.size() == 1 &&
            (!UsePartialMad64_32 || !CarryIn.empty())) {
          do {
            // Skip multiplication if one of the operands is 0
            unsigned j1 = DstIndex - j0;
            if (Src0KnownZeros[j0] || Src1KnownZeros[j1]) {
              ++j0;
              continue;
            }
            auto Mul = B.buildMul(S32, Src0[j0], Src1[j1]);
            if (!LocalAccum[0] || KB.getKnownBits(LocalAccum[0]).isZero()) {
              LocalAccum[0] = Mul.getReg(0);
            } else {
              if (CarryIn.empty()) {
                LocalAccum[0] = B.buildAdd(S32, LocalAccum[0], Mul).getReg(0);
              } else {
                LocalAccum[0] =
                    B.buildUAdde(S32, S1, LocalAccum[0], Mul, CarryIn.back())
                        .getReg(0);
                CarryIn.pop_back();
              }
            }
            ++j0;
          } while (j0 <= DstIndex && (!UsePartialMad64_32 || !CarryIn.empty()));
        }
 
        // Build full 64-bit multiplies.
        if (j0 <= DstIndex) {
          bool HaveSmallAccum = false;
          Register Tmp;
 
          if (LocalAccum[0]) {
            if (LocalAccum.size() == 1) {
              Tmp = B.buildAnyExt(S64, LocalAccum[0]).getReg(0);
              HaveSmallAccum = true;
            } else if (LocalAccum[1]) {
              Tmp = B.buildMergeLikeInstr(S64, LocalAccum).getReg(0);
              HaveSmallAccum = false;
            } else {
              Tmp = B.buildZExt(S64, LocalAccum[0]).getReg(0);
              HaveSmallAccum = true;
            }
          } else {
            assert(LocalAccum.size() == 1 || !LocalAccum[1]);
            Tmp = getZero64();
            HaveSmallAccum = true;
          }
 
          do {
            unsigned j1 = DstIndex - j0;
            if (Src0KnownZeros[j0] || Src1KnownZeros[j1]) {
              ++j0;
              continue;
            }
            auto Mad = B.buildInstr(AMDGPU::G_AMDGPU_MAD_U64_U32, {S64, S1},
                                    {Src0[j0], Src1[j1], Tmp});
            Tmp = Mad.getReg(0);
            if (!HaveSmallAccum)
              CarryOut.push_back(Mad.getReg(1));
            HaveSmallAccum = false;
 
            ++j0;
          } while (j0 <= DstIndex);
 
          auto Unmerge = B.buildUnmerge(S32, Tmp);
          LocalAccum[0] = Unmerge.getReg(0);
          if (LocalAccum.size() > 1)
            LocalAccum[1] = Unmerge.getReg(1);
        }
 
        return CarryOut;
      };
 
  // Outer multiply loop, iterating over destination parts from least
  // significant to most significant parts.
  //
  // The columns of the following diagram correspond to the destination parts
  // affected by one iteration of the outer loop (ignoring boundary
  // conditions).
  //
  //   Dest index relative to 2 * i:      1 0 -1
  //                                      ------
  //   Carries from previous iteration:     e o
  //   Even-aligned partial product sum:  E E .
  //   Odd-aligned partial product sum:     O O
  //
  // 'o' is OddCarry, 'e' is EvenCarry.
  // EE and OO are computed from partial products via buildMadChain and use
  // accumulation where possible and appropriate.
  //
  Register SeparateOddCarry;
  Carry EvenCarry;
  Carry OddCarry;
 
  for (unsigned i = 0; i <= Accum.size() / 2; ++i) {
    Carry OddCarryIn = std::move(OddCarry);
    Carry EvenCarryIn = std::move(EvenCarry);
    OddCarry.clear();
    EvenCarry.clear();
 
    // Partial products at offset 2 * i.
    if (2 * i < Accum.size()) {
      auto LocalAccum = Accum.drop_front(2 * i).take_front(2);
      EvenCarry = buildMadChain(LocalAccum, 2 * i, EvenCarryIn);
    }
 
    // Partial products at offset 2 * i - 1.
    if (i > 0) {
      if (!SeparateOddAlignedProducts) {
        auto LocalAccum = Accum.drop_front(2 * i - 1).take_front(2);
        OddCarry = buildMadChain(LocalAccum, 2 * i - 1, OddCarryIn);
      } else {
        bool IsHighest = 2 * i >= Accum.size();
        Register SeparateOddOut[2];
        auto LocalAccum = MutableArrayRef(SeparateOddOut)
                              .take_front(IsHighest ? 1 : 2);
        OddCarry = buildMadChain(LocalAccum, 2 * i - 1, OddCarryIn);
 
        MachineInstr *Lo;
 
        if (i == 1) {
          if (!IsHighest)
            Lo = B.buildUAddo(S32, S1, Accum[2 * i - 1], SeparateOddOut[0]);
          else
            Lo = B.buildAdd(S32, Accum[2 * i - 1], SeparateOddOut[0]);
        } else {
          Lo = B.buildUAdde(S32, S1, Accum[2 * i - 1], SeparateOddOut[0],
                            SeparateOddCarry);
        }
        Accum[2 * i - 1] = Lo->getOperand(0).getReg();
 
        if (!IsHighest) {
          auto Hi = B.buildUAdde(S32, S1, Accum[2 * i], SeparateOddOut[1],
                                Lo->getOperand(1).getReg());
          Accum[2 * i] = Hi.getReg(0);
          SeparateOddCarry = Hi.getReg(1);
        }
      }
    }
 
    // Add in the carries from the previous iteration
    if (i > 0) {
      if (Register CarryOut = mergeCarry(Accum[2 * i - 1], OddCarryIn))
        EvenCarryIn.push_back(CarryOut);
 
      if (2 * i < Accum.size()) {
        if (Register CarryOut = mergeCarry(Accum[2 * i], EvenCarryIn))
          OddCarry.push_back(CarryOut);
      }
    }
  }
}
 
// Custom narrowing of wide multiplies using wide multiply-add instructions.
//
// TODO: If the multiply is followed by an addition, we should attempt to
// integrate it to make better use of V_MAD_U64_U32's multiply-add capabilities.
bool AMDGPULegalizerInfo::legalizeMul(LegalizerHelper &Helper,
                                      MachineInstr &MI) const {
  assert(ST.hasMad64_32());
  assert(MI.getOpcode() == TargetOpcode::G_MUL);
 
  MachineIRBuilder &B = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *B.getMRI();
 
  Register DstReg = MI.getOperand(0).getReg();
  Register Src0 = MI.getOperand(1).getReg();
  Register Src1 = MI.getOperand(2).getReg();
 
  LLT Ty = MRI.getType(DstReg);
  assert(Ty.isScalar());
 
  unsigned Size = Ty.getSizeInBits();
  unsigned NumParts = Size / 32;
  assert((Size % 32) == 0);
  assert(NumParts >= 2);
 
  // Whether to use MAD_64_32 for partial products whose high half is
  // discarded. This avoids some ADD instructions but risks false dependency
  // stalls on some subtargets in some cases.
  const bool UsePartialMad64_32 = ST.getGeneration() < AMDGPUSubtarget::GFX10;
 
  // Whether to compute odd-aligned partial products separately. This is
  // advisable on subtargets where the accumulator of MAD_64_32 must be placed
  // in an even-aligned VGPR.
  const bool SeparateOddAlignedProducts = ST.hasFullRate64Ops();
 
  LLT S32 = LLT::scalar(32);
  SmallVector<Register, 2> Src0Parts, Src1Parts;
  for (unsigned i = 0; i < NumParts; ++i) {
    Src0Parts.push_back(MRI.createGenericVirtualRegister(S32));
    Src1Parts.push_back(MRI.createGenericVirtualRegister(S32));
  }
  B.buildUnmerge(Src0Parts, Src0);
  B.buildUnmerge(Src1Parts, Src1);
 
  SmallVector<Register, 2> AccumRegs(NumParts);
  buildMultiply(Helper, AccumRegs, Src0Parts, Src1Parts, UsePartialMad64_32,
                SeparateOddAlignedProducts);
 
  B.buildMergeLikeInstr(DstReg, AccumRegs);
  MI.eraseFromParent();
  return true;
}
 
// Legalize ctlz/cttz to ffbh/ffbl instead of the default legalization to
// ctlz/cttz_zero_undef. This allows us to fix up the result for the zero input
// case with a single min instruction instead of a compare+select.
bool AMDGPULegalizerInfo::legalizeCTLZ_CTTZ(MachineInstr &MI,
                                            MachineRegisterInfo &MRI,
                                            MachineIRBuilder &B) const {
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();
  LLT DstTy = MRI.getType(Dst);
  LLT SrcTy = MRI.getType(Src);
 
  unsigned NewOpc = MI.getOpcode() == AMDGPU::G_CTLZ
                        ? AMDGPU::G_AMDGPU_FFBH_U32
                        : AMDGPU::G_AMDGPU_FFBL_B32;
  auto Tmp = B.buildInstr(NewOpc, {DstTy}, {Src});
  B.buildUMin(Dst, Tmp, B.buildConstant(DstTy, SrcTy.getSizeInBits()));
 
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeCTLZ_ZERO_UNDEF(MachineInstr &MI,
                                                  MachineRegisterInfo &MRI,
                                                  MachineIRBuilder &B) const {
  Register Dst = MI.getOperand(0).getReg();
  Register Src = MI.getOperand(1).getReg();
  LLT SrcTy = MRI.getType(Src);
  TypeSize NumBits = SrcTy.getSizeInBits();
 
  assert(NumBits < 32u);
 
  auto ShiftAmt = B.buildConstant(S32, 32u - NumBits);
  auto Extend = B.buildAnyExt(S32, {Src}).getReg(0u);
  auto Shift = B.buildShl(S32, Extend, ShiftAmt);
  auto Ctlz = B.buildInstr(AMDGPU::G_AMDGPU_FFBH_U32, {S32}, {Shift});
  B.buildTrunc(Dst, Ctlz);
  MI.eraseFromParent();
  return true;
}
 
// Check that this is a G_XOR x, -1
static bool isNot(const MachineRegisterInfo &MRI, const MachineInstr &MI) {
  if (MI.getOpcode() != TargetOpcode::G_XOR)
    return false;
  auto ConstVal = getIConstantVRegSExtVal(MI.getOperand(2).getReg(), MRI);
  return ConstVal && *ConstVal == -1;
}
 
// Return the use branch instruction, otherwise null if the usage is invalid.
static MachineInstr *
verifyCFIntrinsic(MachineInstr &MI, MachineRegisterInfo &MRI, MachineInstr *&Br,
                  MachineBasicBlock *&UncondBrTarget, bool &Negated) {
  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 (isNot(MRI, *UseMI)) {
    Register NegatedCond = UseMI->getOperand(0).getReg();
    if (!MRI.hasOneNonDBGUse(NegatedCond))
      return nullptr;
 
    // We're deleting the def of this value, so we need to remove it.
    eraseInstr(*UseMI, MRI);
 
    UseMI = &*MRI.use_instr_nodbg_begin(NegatedCond);
    Negated = true;
  }
 
  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;
}
 
bool AMDGPULegalizerInfo::loadInputValue(Register DstReg, MachineIRBuilder &B,
                                         const ArgDescriptor *Arg,
                                         const TargetRegisterClass *ArgRC,
                                         LLT ArgTy) const {
  MCRegister SrcReg = Arg->getRegister();
  assert(Register::isPhysicalRegister(SrcReg) && "Physical register expected");
  assert(DstReg.isVirtual() && "Virtual register expected");
 
  Register LiveIn = getFunctionLiveInPhysReg(B.getMF(), B.getTII(), SrcReg,
                                             *ArgRC, B.getDebugLoc(), ArgTy);
  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 = llvm::countr_zero<unsigned>(Mask);
 
    Register AndMaskSrc = LiveIn;
 
    // TODO: Avoid clearing the high bits if we know workitem id y/z are always
    // 0.
    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::loadInputValue(
    Register DstReg, MachineIRBuilder &B,
    AMDGPUFunctionArgInfo::PreloadedValue ArgType) const {
  const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
  const ArgDescriptor *Arg = nullptr;
  const TargetRegisterClass *ArgRC;
  LLT ArgTy;
 
  CallingConv::ID CC = B.getMF().getFunction().getCallingConv();
  const ArgDescriptor WorkGroupIDX =
      ArgDescriptor::createRegister(AMDGPU::TTMP9);
  // If GridZ is not programmed in an entry function then the hardware will set
  // it to all zeros, so there is no need to mask the GridY value in the low
  // order bits.
  const ArgDescriptor WorkGroupIDY = ArgDescriptor::createRegister(
      AMDGPU::TTMP7,
      AMDGPU::isEntryFunctionCC(CC) && !MFI->hasWorkGroupIDZ() ? ~0u : 0xFFFFu);
  const ArgDescriptor WorkGroupIDZ =
      ArgDescriptor::createRegister(AMDGPU::TTMP7, 0xFFFF0000u);
  if (ST.hasArchitectedSGPRs() &&
      (AMDGPU::isCompute(CC) || CC == CallingConv::AMDGPU_Gfx)) {
    switch (ArgType) {
    case AMDGPUFunctionArgInfo::WORKGROUP_ID_X:
      Arg = &WorkGroupIDX;
      ArgRC = &AMDGPU::SReg_32RegClass;
      ArgTy = LLT::scalar(32);
      break;
    case AMDGPUFunctionArgInfo::WORKGROUP_ID_Y:
      Arg = &WorkGroupIDY;
      ArgRC = &AMDGPU::SReg_32RegClass;
      ArgTy = LLT::scalar(32);
      break;
    case AMDGPUFunctionArgInfo::WORKGROUP_ID_Z:
      Arg = &WorkGroupIDZ;
      ArgRC = &AMDGPU::SReg_32RegClass;
      ArgTy = LLT::scalar(32);
      break;
    default:
      break;
    }
  }
 
  if (!Arg)
    std::tie(Arg, ArgRC, ArgTy) = MFI->getPreloadedValue(ArgType);
 
  if (!Arg) {
    if (ArgType == AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR) {
      // The intrinsic may appear when we have a 0 sized kernarg segment, in which
      // case the pointer argument may be missing and we use null.
      B.buildConstant(DstReg, 0);
      return true;
    }
 
    // It's undefined behavior if a function marked with the amdgpu-no-*
    // attributes uses the corresponding intrinsic.
    B.buildUndef(DstReg);
    return true;
  }
 
  if (!Arg->isRegister() || !Arg->getRegister().isValid())
    return false; // TODO: Handle these
  return loadInputValue(DstReg, B, Arg, ArgRC, ArgTy);
}
 
bool AMDGPULegalizerInfo::legalizePreloadedArgIntrin(
    MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B,
    AMDGPUFunctionArgInfo::PreloadedValue ArgType) const {
  if (!loadInputValue(MI.getOperand(0).getReg(), B, ArgType))
    return false;
 
  MI.eraseFromParent();
  return true;
}
 
static bool replaceWithConstant(MachineIRBuilder &B, MachineInstr &MI,
                                int64_t C) {
  B.buildConstant(MI.getOperand(0).getReg(), C);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeWorkitemIDIntrinsic(
    MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &B,
    unsigned Dim, AMDGPUFunctionArgInfo::PreloadedValue ArgType) const {
  unsigned MaxID = ST.getMaxWorkitemID(B.getMF().getFunction(), Dim);
  if (MaxID == 0)
    return replaceWithConstant(B, MI, 0);
 
  const SIMachineFunctionInfo *MFI = B.getMF().getInfo<SIMachineFunctionInfo>();
  const ArgDescriptor *Arg;
  const TargetRegisterClass *ArgRC;
  LLT ArgTy;
  std::tie(Arg, ArgRC, ArgTy) = MFI->getPreloadedValue(ArgType);
 
  Register DstReg = MI.getOperand(0).getReg();
  if (!Arg) {
    // It's undefined behavior if a function marked with the amdgpu-no-*
    // attributes uses the corresponding intrinsic.
    B.buildUndef(DstReg);
    MI.eraseFromParent();
    return true;
  }
 
  if (Arg->isMasked()) {
    // Don't bother inserting AssertZext for packed IDs since we're emitting the
    // masking operations anyway.
    //
    // TODO: We could assert the top bit is 0 for the source copy.
    if (!loadInputValue(DstReg, B, ArgType))
      return false;
  } else {
    Register TmpReg = MRI.createGenericVirtualRegister(LLT::scalar(32));
    if (!loadInputValue(TmpReg, B, ArgType))
      return false;
    B.buildAssertZExt(DstReg, TmpReg, llvm::bit_width(MaxID));
  }
 
  MI.eraseFromParent();
  return true;
}
 
Register AMDGPULegalizerInfo::getKernargParameterPtr(MachineIRBuilder &B,
                                                     int64_t Offset) const {
  LLT PtrTy = LLT::pointer(AMDGPUAS::CONSTANT_ADDRESS, 64);
  Register KernArgReg = B.getMRI()->createGenericVirtualRegister(PtrTy);
 
  // TODO: If we passed in the base kernel offset we could have a better
  // alignment than 4, but we don't really need it.
  if (!loadInputValue(KernArgReg, B,
                      AMDGPUFunctionArgInfo::KERNARG_SEGMENT_PTR))
    llvm_unreachable("failed to find kernarg segment ptr");
 
  auto COffset = B.buildConstant(LLT::scalar(64), Offset);
  // TODO: Should get nuw
  return B.buildPtrAdd(PtrTy, KernArgReg, COffset).getReg(0);
}
 
/// Legalize a value that's loaded from kernel arguments. This is only used by
/// legacy intrinsics.
bool AMDGPULegalizerInfo::legalizeKernargMemParameter(MachineInstr &MI,
                                                      MachineIRBuilder &B,
                                                      uint64_t Offset,
                                                      Align Alignment) const {
  Register DstReg = MI.getOperand(0).getReg();
 
  assert(B.getMRI()->getType(DstReg) == LLT::scalar(32) &&
         "unexpected kernarg parameter type");
 
  Register Ptr = getKernargParameterPtr(B, Offset);
  MachinePointerInfo PtrInfo(AMDGPUAS::CONSTANT_ADDRESS);
  B.buildLoad(DstReg, Ptr, PtrInfo, Align(4),
              MachineMemOperand::MODereferenceable |
                  MachineMemOperand::MOInvariant);
  MI.eraseFromParent();
  return true;
}
 
bool AMDGPULegalizerInfo::legalizeFDIV(MachineInstr &MI,
                                       MachineRegisterInfo &MRI,
                                       MachineIRBuilder &B) const {
  Register Dst = MI.getOperand(0).getReg();
  LLT DstTy =