//===- AArch64LegalizerInfo.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
/// AArch64.
/// \todo This should be generated by TableGen.
//===----------------------------------------------------------------------===//

#include "AArch64LegalizerInfo.h"
#include "AArch64RegisterBankInfo.h"
#include "AArch64Subtarget.h"
#include "llvm/CodeGen/GlobalISel/LegalizerHelper.h"
#include "llvm/CodeGen/GlobalISel/LegalizerInfo.h"
#include "llvm/CodeGen/GlobalISel/MIPatternMatch.h"
#include "llvm/CodeGen/GlobalISel/MachineIRBuilder.h"
#include "llvm/CodeGen/GlobalISel/Utils.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsAArch64.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/MathExtras.h"
#include <initializer_list>

#define DEBUG_TYPE "aarch64-legalinfo"

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

AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
    : ST(&ST) {
  using namespace TargetOpcode;
  const LLT p0 = LLT::pointer(0, 64);
  const LLT s8 = LLT::scalar(8);
  const LLT s16 = LLT::scalar(16);
  const LLT s32 = LLT::scalar(32);
  const LLT s64 = LLT::scalar(64);
  const LLT s128 = LLT::scalar(128);
  const LLT v16s8 = LLT::fixed_vector(16, 8);
  const LLT v8s8 = LLT::fixed_vector(8, 8);
  const LLT v4s8 = LLT::fixed_vector(4, 8);
  const LLT v8s16 = LLT::fixed_vector(8, 16);
  const LLT v4s16 = LLT::fixed_vector(4, 16);
  const LLT v2s16 = LLT::fixed_vector(2, 16);
  const LLT v2s32 = LLT::fixed_vector(2, 32);
  const LLT v4s32 = LLT::fixed_vector(4, 32);
  const LLT v2s64 = LLT::fixed_vector(2, 64);
  const LLT v2p0 = LLT::fixed_vector(2, p0);

  std::initializer_list<LLT> PackedVectorAllTypeList = {/* Begin 128bit types */
                                                        v16s8, v8s16, v4s32,
                                                        v2s64, v2p0,
                                                        /* End 128bit types */
                                                        /* Begin 64bit types */
                                                        v8s8, v4s16, v2s32};

  const TargetMachine &TM = ST.getTargetLowering()->getTargetMachine();

  // FIXME: support subtargets which have neon/fp-armv8 disabled.
  if (!ST.hasNEON() || !ST.hasFPARMv8()) {
    getLegacyLegalizerInfo().computeTables();
    return;
  }

  // Some instructions only support s16 if the subtarget has full 16-bit FP
  // support.
  const bool HasFP16 = ST.hasFullFP16();
  const LLT &MinFPScalar = HasFP16 ? s16 : s32;

  getActionDefinitionsBuilder({G_IMPLICIT_DEF, G_FREEZE})
      .legalFor({p0, s8, s16, s32, s64})
      .legalFor(PackedVectorAllTypeList)
      .widenScalarToNextPow2(0)
      .clampScalar(0, s8, s64)
      .fewerElementsIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].isVector() &&
                   (Query.Types[0].getElementType() != s64 ||
                    Query.Types[0].getNumElements() != 2);
          },
          [=](const LegalityQuery &Query) {
            LLT EltTy = Query.Types[0].getElementType();
            if (EltTy == s64)
              return std::make_pair(0, LLT::fixed_vector(2, 64));
            return std::make_pair(0, EltTy);
          });

  getActionDefinitionsBuilder(G_PHI)
      .legalFor({p0, s16, s32, s64})
      .legalFor(PackedVectorAllTypeList)
      .widenScalarToNextPow2(0)
      .clampScalar(0, s16, s64)
      // Maximum: sN * k = 128
      .clampMaxNumElements(0, s8, 16)
      .clampMaxNumElements(0, s16, 8)
      .clampMaxNumElements(0, s32, 4)
      .clampMaxNumElements(0, s64, 2)
      .clampMaxNumElements(0, p0, 2);

  getActionDefinitionsBuilder(G_BSWAP)
      .legalFor({s32, s64, v4s32, v2s32, v2s64})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64);

  getActionDefinitionsBuilder({G_ADD, G_SUB, G_MUL, G_AND, G_OR, G_XOR})
      .legalFor({s32, s64, v2s32, v4s32, v4s16, v8s16, v16s8, v8s8})
      .scalarizeIf(
          [=](const LegalityQuery &Query) {
            return Query.Opcode == G_MUL && Query.Types[0] == v2s64;
          },
          0)
      .legalFor({v2s64})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0);

  getActionDefinitionsBuilder({G_SHL, G_ASHR, G_LSHR})
      .customIf([=](const LegalityQuery &Query) {
        const auto &SrcTy = Query.Types[0];
        const auto &AmtTy = Query.Types[1];
        return !SrcTy.isVector() && SrcTy.getSizeInBits() == 32 &&
               AmtTy.getSizeInBits() == 32;
      })
      .legalFor({
          {s32, s32},
          {s32, s64},
          {s64, s64},
          {v8s8, v8s8},
          {v16s8, v16s8},
          {v4s16, v4s16},
          {v8s16, v8s16},
          {v2s32, v2s32},
          {v4s32, v4s32},
          {v2s64, v2s64},
      })
      .widenScalarToNextPow2(0)
      .clampScalar(1, s32, s64)
      .clampScalar(0, s32, s64)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0)
      .minScalarSameAs(1, 0);

  getActionDefinitionsBuilder(G_PTR_ADD)
      .legalFor({{p0, s64}, {v2p0, v2s64}})
      .clampScalar(1, s64, s64);

  getActionDefinitionsBuilder(G_PTRMASK).legalFor({{p0, s64}});

  getActionDefinitionsBuilder({G_SDIV, G_UDIV})
      .legalFor({s32, s64})
      .libcallFor({s128})
      .clampScalar(0, s32, s64)
      .widenScalarToNextPow2(0)
      .scalarize(0);

  getActionDefinitionsBuilder({G_SREM, G_UREM, G_SDIVREM, G_UDIVREM})
      .lowerFor({s8, s16, s32, s64, v2s64, v4s32, v2s32})
      .widenScalarOrEltToNextPow2(0)
      .clampScalarOrElt(0, s32, s64)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0);


  getActionDefinitionsBuilder({G_SMULO, G_UMULO})
      .widenScalarToNextPow2(0, /*Min = */ 32)
      .clampScalar(0, s32, s64)
      .lower();

  getActionDefinitionsBuilder({G_SMULH, G_UMULH})
      .legalFor({s64, v8s16, v16s8, v4s32})
      .lower();

  getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
      .legalFor({v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      // FIXME: This sholdn't be needed as v2s64 types are going to
      // be expanded anyway, but G_ICMP doesn't support splitting vectors yet
      .clampNumElements(0, v2s64, v2s64)
      .lower();

  getActionDefinitionsBuilder(
      {G_SADDE, G_SSUBE, G_UADDE, G_USUBE, G_SADDO, G_SSUBO, G_UADDO, G_USUBO})
      .legalFor({{s32, s32}, {s64, s32}})
      .clampScalar(0, s32, s64)
       .clampScalar(1, s32, s64)
      .widenScalarToNextPow2(0);

  getActionDefinitionsBuilder({G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FNEG})
      .legalFor({MinFPScalar, s32, s64, v2s64, v4s32, v2s32})
      .clampScalar(0, MinFPScalar, s64)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64);

  getActionDefinitionsBuilder(G_FREM).libcallFor({s32, s64});

  getActionDefinitionsBuilder({G_FCEIL, G_FABS, G_FSQRT, G_FFLOOR, G_FRINT,
                               G_FMA, G_INTRINSIC_TRUNC, G_INTRINSIC_ROUND,
                               G_FNEARBYINT, G_INTRINSIC_LRINT})
      // If we don't have full FP16 support, then scalarize the elements of
      // vectors containing fp16 types.
      .fewerElementsIf(
          [=, &ST](const LegalityQuery &Query) {
            const auto &Ty = Query.Types[0];
            return Ty.isVector() && Ty.getElementType() == s16 &&
                   !ST.hasFullFP16();
          },
          [=](const LegalityQuery &Query) { return std::make_pair(0, s16); })
      // If we don't have full FP16 support, then widen s16 to s32 if we
      // encounter it.
      .widenScalarIf(
          [=, &ST](const LegalityQuery &Query) {
            return Query.Types[0] == s16 && !ST.hasFullFP16();
          },
          [=](const LegalityQuery &Query) { return std::make_pair(0, s32); })
      .legalFor({s16, s32, s64, v2s32, v4s32, v2s64, v2s16, v4s16, v8s16});

  getActionDefinitionsBuilder(
      {G_FCOS, G_FSIN, G_FLOG10, G_FLOG, G_FLOG2, G_FEXP, G_FEXP2, G_FPOW})
      // We need a call for these, so we always need to scalarize.
      .scalarize(0)
      // Regardless of FP16 support, widen 16-bit elements to 32-bits.
      .minScalar(0, s32)
      .libcallFor({s32, s64, v2s32, v4s32, v2s64});

  getActionDefinitionsBuilder(G_INSERT)
      .legalIf(all(typeInSet(0, {s32, s64, p0}),
                   typeInSet(1, {s8, s16, s32}), smallerThan(1, 0)))
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64)
      .widenScalarToNextPow2(1)
      .minScalar(1, s8)
      .maxScalarIf(typeInSet(0, {s32}), 1, s16)
      .maxScalarIf(typeInSet(0, {s64, p0}), 1, s32);

  getActionDefinitionsBuilder(G_EXTRACT)
      .legalIf(all(typeInSet(0, {s16, s32, s64, p0}),
                   typeInSet(1, {s32, s64, s128, p0}), smallerThan(0, 1)))
      .widenScalarToNextPow2(1)
      .clampScalar(1, s32, s128)
      .widenScalarToNextPow2(0)
      .minScalar(0, s16)
      .maxScalarIf(typeInSet(1, {s32}), 0, s16)
      .maxScalarIf(typeInSet(1, {s64, p0}), 0, s32)
      .maxScalarIf(typeInSet(1, {s128}), 0, s64);


  for (unsigned Op : {G_SEXTLOAD, G_ZEXTLOAD}) {
    auto &Actions =  getActionDefinitionsBuilder(Op);

    if (Op == G_SEXTLOAD)
      Actions.lowerIf(atomicOrderingAtLeastOrStrongerThan(0, AtomicOrdering::Unordered));

    // Atomics have zero extending behavior.
    Actions
      .legalForTypesWithMemDesc({{s32, p0, s8, 8},
                                 {s32, p0, s16, 8},
                                 {s32, p0, s32, 8},
                                 {s64, p0, s8, 2},
                                 {s64, p0, s16, 2},
                                 {s64, p0, s32, 4},
                                 {s64, p0, s64, 8},
                                 {p0, p0, s64, 8},
                                 {v2s32, p0, s64, 8}})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64)
      // TODO: We could support sum-of-pow2's but the lowering code doesn't know
      //       how to do that yet.
      .unsupportedIfMemSizeNotPow2()
      // Lower anything left over into G_*EXT and G_LOAD
      .lower();
  }

  auto IsPtrVecPred = [=](const LegalityQuery &Query) {
    const LLT &ValTy = Query.Types[0];
    if (!ValTy.isVector())
      return false;
    const LLT EltTy = ValTy.getElementType();
    return EltTy.isPointer() && EltTy.getAddressSpace() == 0;
  };

  getActionDefinitionsBuilder(G_LOAD)
      .customIf([=](const LegalityQuery &Query) {
        return Query.Types[0] == s128 &&
               Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic;
      })
      .legalForTypesWithMemDesc({{s8, p0, s8, 8},
                                 {s16, p0, s16, 8},
                                 {s32, p0, s32, 8},
                                 {s64, p0, s64, 8},
                                 {p0, p0, s64, 8},
                                 {s128, p0, s128, 8},
                                 {v8s8, p0, s64, 8},
                                 {v16s8, p0, s128, 8},
                                 {v4s16, p0, s64, 8},
                                 {v8s16, p0, s128, 8},
                                 {v2s32, p0, s64, 8},
                                 {v4s32, p0, s128, 8},
                                 {v2s64, p0, s128, 8}})
      // These extends are also legal
      .legalForTypesWithMemDesc({{s32, p0, s8, 8}, {s32, p0, s16, 8}})
      .widenScalarToNextPow2(0, /* MinSize = */8)
      .lowerIfMemSizeNotByteSizePow2()
      .clampScalar(0, s8, s64)
      .narrowScalarIf([=](const LegalityQuery &Query) {
        // Clamp extending load results to 32-bits.
        return Query.Types[0].isScalar() &&
          Query.Types[0] != Query.MMODescrs[0].MemoryTy &&
          Query.Types[0].getSizeInBits() > 32;
        },
        changeTo(0, s32))
      .clampMaxNumElements(0, s8, 16)
      .clampMaxNumElements(0, s16, 8)
      .clampMaxNumElements(0, s32, 4)
      .clampMaxNumElements(0, s64, 2)
      .clampMaxNumElements(0, p0, 2)
      .customIf(IsPtrVecPred)
      .scalarizeIf(typeIs(0, v2s16), 0);

  getActionDefinitionsBuilder(G_STORE)
      .customIf([=](const LegalityQuery &Query) {
        return Query.Types[0] == s128 &&
               Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic;
      })
      .legalForTypesWithMemDesc({{s8, p0, s8, 8},
                                 {s16, p0, s8, 8}, // truncstorei8 from s16
                                 {s32, p0, s8, 8}, // truncstorei8 from s32
                                 {s64, p0, s8, 8}, // truncstorei8 from s64
                                 {s16, p0, s16, 8},
                                 {s32, p0, s16, 8}, // truncstorei16 from s32
                                 {s64, p0, s16, 8}, // truncstorei16 from s64
                                 {s32, p0, s8, 8},
                                 {s32, p0, s16, 8},
                                 {s32, p0, s32, 8},
                                 {s64, p0, s64, 8},
                                 {s64, p0, s32, 8}, // truncstorei32 from s64
                                 {p0, p0, s64, 8},
                                 {s128, p0, s128, 8},
                                 {v16s8, p0, s128, 8},
                                 {v8s8, p0, s64, 8},
                                 {v4s16, p0, s64, 8},
                                 {v8s16, p0, s128, 8},
                                 {v2s32, p0, s64, 8},
                                 {v4s32, p0, s128, 8},
                                 {v2s64, p0, s128, 8}})
      .clampScalar(0, s8, s64)
      .lowerIf([=](const LegalityQuery &Query) {
        return Query.Types[0].isScalar() &&
               Query.Types[0] != Query.MMODescrs[0].MemoryTy;
      })
      // Maximum: sN * k = 128
      .clampMaxNumElements(0, s8, 16)
      .clampMaxNumElements(0, s16, 8)
      .clampMaxNumElements(0, s32, 4)
      .clampMaxNumElements(0, s64, 2)
      .clampMaxNumElements(0, p0, 2)
      .lowerIfMemSizeNotPow2()
      .customIf(IsPtrVecPred)
      .scalarizeIf(typeIs(0, v2s16), 0);

  // Constants
  getActionDefinitionsBuilder(G_CONSTANT)
      .legalFor({p0, s8, s16, s32, s64})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s8, s64);
  getActionDefinitionsBuilder(G_FCONSTANT)
      .legalIf([=](const LegalityQuery &Query) {
        const auto &Ty = Query.Types[0];
        if (HasFP16 && Ty == s16)
          return true;
        return Ty == s32 || Ty == s64 || Ty == s128;
      })
      .clampScalar(0, MinFPScalar, s128);

  getActionDefinitionsBuilder({G_ICMP, G_FCMP})
      .legalFor({{s32, s32},
                 {s32, s64},
                 {s32, p0},
                 {v4s32, v4s32},
                 {v2s32, v2s32},
                 {v2s64, v2s64},
                 {v2s64, v2p0},
                 {v4s16, v4s16},
                 {v8s16, v8s16},
                 {v8s8, v8s8},
                 {v16s8, v16s8}})
      .widenScalarOrEltToNextPow2(1)
      .clampScalar(1, s32, s64)
      .clampScalar(0, s32, s32)
      .minScalarEltSameAsIf(
          [=](const LegalityQuery &Query) {
            const LLT &Ty = Query.Types[0];
            const LLT &SrcTy = Query.Types[1];
            return Ty.isVector() && !SrcTy.getElementType().isPointer() &&
                   Ty.getElementType() != SrcTy.getElementType();
          },
          0, 1)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) { return Query.Types[1] == v2s16; },
          1, s32)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) { return Query.Types[1] == v2p0; }, 0,
          s64)
      .clampNumElements(0, v2s32, v4s32);

  // Extensions
  auto ExtLegalFunc = [=](const LegalityQuery &Query) {
    unsigned DstSize = Query.Types[0].getSizeInBits();

    if (DstSize == 128 && !Query.Types[0].isVector())
      return false; // Extending to a scalar s128 needs narrowing.

    // Make sure that we have something that will fit in a register, and
    // make sure it's a power of 2.
    if (DstSize < 8 || DstSize > 128 || !isPowerOf2_32(DstSize))
      return false;

    const LLT &SrcTy = Query.Types[1];

    // Make sure we fit in a register otherwise. Don't bother checking that
    // the source type is below 128 bits. We shouldn't be allowing anything
    // through which is wider than the destination in the first place.
    unsigned SrcSize = SrcTy.getSizeInBits();
    if (SrcSize < 8 || !isPowerOf2_32(SrcSize))
      return false;

    return true;
  };
  getActionDefinitionsBuilder({G_ZEXT, G_SEXT, G_ANYEXT})
      .legalIf(ExtLegalFunc)
      .clampScalar(0, s64, s64); // Just for s128, others are handled above.

  getActionDefinitionsBuilder(G_TRUNC)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) { return Query.Types[0].isVector(); },
          0, s8)
      .customIf([=](const LegalityQuery &Query) {
        LLT DstTy = Query.Types[0];
        LLT SrcTy = Query.Types[1];
        return DstTy == v8s8 && SrcTy.getSizeInBits() > 128;
      })
      .alwaysLegal();

  getActionDefinitionsBuilder(G_SEXT_INREG).legalFor({s32, s64}).lower();

  // FP conversions
  getActionDefinitionsBuilder(G_FPTRUNC)
      .legalFor(
          {{s16, s32}, {s16, s64}, {s32, s64}, {v4s16, v4s32}, {v2s32, v2s64}})
      .clampMaxNumElements(0, s32, 2);
  getActionDefinitionsBuilder(G_FPEXT)
      .legalFor(
          {{s32, s16}, {s64, s16}, {s64, s32}, {v4s32, v4s16}, {v2s64, v2s32}})
      .clampMaxNumElements(0, s64, 2);

  // Conversions
  getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
      .legalForCartesianProduct({s32, s64, v2s64, v4s32, v2s32})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64)
      .widenScalarToNextPow2(1)
      .clampScalar(1, s32, s64);

  getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
      .legalForCartesianProduct({s32, s64, v2s64, v4s32, v2s32})
      .clampScalar(1, s32, s64)
      .minScalarSameAs(1, 0)
      .clampScalar(0, s32, s64)
      .widenScalarToNextPow2(0);

  // Control-flow
  getActionDefinitionsBuilder(G_BRCOND)
    .legalFor({s32})
    .clampScalar(0, s32, s32);
  getActionDefinitionsBuilder(G_BRINDIRECT).legalFor({p0});

  getActionDefinitionsBuilder(G_SELECT)
      .legalFor({{s32, s32}, {s64, s32}, {p0, s32}})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64)
      .clampScalar(1, s32, s32)
      .minScalarEltSameAsIf(all(isVector(0), isVector(1)), 1, 0)
      .lowerIf(isVector(0));

  // Pointer-handling
  getActionDefinitionsBuilder(G_FRAME_INDEX).legalFor({p0});

  if (TM.getCodeModel() == CodeModel::Small)
    getActionDefinitionsBuilder(G_GLOBAL_VALUE).custom();
  else
    getActionDefinitionsBuilder(G_GLOBAL_VALUE).legalFor({p0});

  getActionDefinitionsBuilder(G_PTRTOINT)
      .legalForCartesianProduct({s8, s16, s32, s64}, {p0})
      .legalFor({{v2s64, v2p0}})
      .maxScalar(0, s64)
      .widenScalarToNextPow2(0, /*Min*/ 8);

  getActionDefinitionsBuilder(G_INTTOPTR)
      .unsupportedIf([&](const LegalityQuery &Query) {
        return Query.Types[0].getSizeInBits() != Query.Types[1].getSizeInBits();
      })
      .legalFor({{p0, s64}, {v2p0, v2s64}});

  // Casts for 32 and 64-bit width type are just copies.
  // Same for 128-bit width type, except they are on the FPR bank.
  getActionDefinitionsBuilder(G_BITCAST)
      // FIXME: This is wrong since G_BITCAST is not allowed to change the
      // number of bits but it's what the previous code described and fixing
      // it breaks tests.
      .legalForCartesianProduct({s8, s16, s32, s64, s128, v16s8, v8s8, v4s8,
                                 v8s16, v4s16, v2s16, v4s32, v2s32, v2s64,
                                 v2p0});

  getActionDefinitionsBuilder(G_VASTART).legalFor({p0});

  // va_list must be a pointer, but most sized types are pretty easy to handle
  // as the destination.
  getActionDefinitionsBuilder(G_VAARG)
      .customForCartesianProduct({s8, s16, s32, s64, p0}, {p0})
      .clampScalar(0, s8, s64)
      .widenScalarToNextPow2(0, /*Min*/ 8);

  getActionDefinitionsBuilder(G_ATOMIC_CMPXCHG_WITH_SUCCESS)
      .lowerIf(
          all(typeInSet(0, {s8, s16, s32, s64, s128}), typeIs(2, p0)));

  getActionDefinitionsBuilder(G_ATOMIC_CMPXCHG)
      .customIf([](const LegalityQuery &Query) {
        return Query.Types[0].getSizeInBits() == 128;
      })
      .clampScalar(0, s32, s64)
      .legalIf(all(typeInSet(0, {s32, s64}), typeIs(1, p0)));

  getActionDefinitionsBuilder(
      {G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD, G_ATOMICRMW_SUB, G_ATOMICRMW_AND,
       G_ATOMICRMW_OR, G_ATOMICRMW_XOR, G_ATOMICRMW_MIN, G_ATOMICRMW_MAX,
       G_ATOMICRMW_UMIN, G_ATOMICRMW_UMAX})
      .clampScalar(0, s32, s64)
      .legalIf(all(typeInSet(0, {s32, s64}), typeIs(1, p0)));

  getActionDefinitionsBuilder(G_BLOCK_ADDR).legalFor({p0});

  // 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;
    getActionDefinitionsBuilder(Op)
        .widenScalarToNextPow2(LitTyIdx, 8)
        .widenScalarToNextPow2(BigTyIdx, 32)
        .clampScalar(LitTyIdx, s8, s64)
        .clampScalar(BigTyIdx, s32, s128)
        .legalIf([=](const LegalityQuery &Q) {
          switch (Q.Types[BigTyIdx].getSizeInBits()) {
          case 32:
          case 64:
          case 128:
            break;
          default:
            return false;
          }
          switch (Q.Types[LitTyIdx].getSizeInBits()) {
          case 8:
          case 16:
          case 32:
          case 64:
            return true;
          default:
            return false;
          }
        });
  }

  getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
      .unsupportedIf([=](const LegalityQuery &Query) {
        const LLT &EltTy = Query.Types[1].getElementType();
        return Query.Types[0] != EltTy;
      })
      .minScalar(2, s64)
      .legalIf([=](const LegalityQuery &Query) {
        const LLT &VecTy = Query.Types[1];
        return VecTy == v2s16 || VecTy == v4s16 || VecTy == v8s16 ||
               VecTy == v4s32 || VecTy == v2s64 || VecTy == v2s32 ||
               VecTy == v8s8 || VecTy == v16s8 || VecTy == v2s32 ||
               VecTy == v2p0;
      })
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            // We want to promote to <M x s1> to <M x s64> if that wouldn't
            // cause the total vec size to be > 128b.
            return Query.Types[1].getNumElements() <= 2;
          },
          0, s64)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].getNumElements() <= 4;
          },
          0, s32)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].getNumElements() <= 8;
          },
          0, s16)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].getNumElements() <= 16;
          },
          0, s8)
      .minScalarOrElt(0, s8) // Worst case, we need at least s8.
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .clampMaxNumElements(1, s16, 8)
      .clampMaxNumElements(1, p0, 2);

  getActionDefinitionsBuilder(G_INSERT_VECTOR_ELT)
      .legalIf(typeInSet(0, {v8s16, v2s32, v4s32, v2s64}));

  getActionDefinitionsBuilder(G_BUILD_VECTOR)
      .legalFor({{v8s8, s8},
                 {v16s8, s8},
                 {v2s16, s16},
                 {v4s16, s16},
                 {v8s16, s16},
                 {v2s32, s32},
                 {v4s32, s32},
                 {v2p0, p0},
                 {v2s64, s64}})
      .clampNumElements(0, v4s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .minScalarOrElt(0, s8)
      .minScalarSameAs(1, 0);

  getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC).lower();

  getActionDefinitionsBuilder(G_CTLZ)
      .legalForCartesianProduct(
          {s32, s64, v8s8, v16s8, v4s16, v8s16, v2s32, v4s32})
      .scalarize(1);
  getActionDefinitionsBuilder(G_CTLZ_ZERO_UNDEF).lower();

  // TODO: Custom lowering for v2s32, v4s32, v2s64.
  getActionDefinitionsBuilder(G_BITREVERSE)
      .legalFor({s32, s64, v8s8, v16s8})
      .widenScalarToNextPow2(0, /*Min = */ 32)
      .clampScalar(0, s32, s64);

  getActionDefinitionsBuilder(G_CTTZ_ZERO_UNDEF).lower();

  getActionDefinitionsBuilder(G_CTTZ)
      .lowerIf(isVector(0))
      .clampScalar(0, s32, s64)
      .scalarSameSizeAs(1, 0)
      .customFor({s32, s64});

  getActionDefinitionsBuilder(G_SHUFFLE_VECTOR)
      .legalIf([=](const LegalityQuery &Query) {
        const LLT &DstTy = Query.Types[0];
        const LLT &SrcTy = Query.Types[1];
        // For now just support the TBL2 variant which needs the source vectors
        // to be the same size as the dest.
        if (DstTy != SrcTy)
          return false;
        for (auto &Ty : {v2s32, v4s32, v2s64, v2p0, v16s8, v8s16}) {
          if (DstTy == Ty)
            return true;
        }
        return false;
      })
      // G_SHUFFLE_VECTOR can have scalar sources (from 1 x s vectors), we
      // just want those lowered into G_BUILD_VECTOR
      .lowerIf([=](const LegalityQuery &Query) {
        return !Query.Types[1].isVector();
      })
      .moreElementsToNextPow2(0)
      .clampNumElements(0, v4s32, v4s32)
      .clampNumElements(0, v2s64, v2s64);

  getActionDefinitionsBuilder(G_CONCAT_VECTORS)
      .legalFor({{v4s32, v2s32}, {v8s16, v4s16}, {v16s8, v8s8}});

  getActionDefinitionsBuilder(G_JUMP_TABLE).legalFor({{p0}, {s64}});

  getActionDefinitionsBuilder(G_BRJT).legalIf([=](const LegalityQuery &Query) {
    return Query.Types[0] == p0 && Query.Types[1] == s64;
  });

  getActionDefinitionsBuilder(G_DYN_STACKALLOC).lower();

  if (ST.hasMOPS()) {
    // G_BZERO is not supported. Currently it is only emitted by
    // PreLegalizerCombiner for G_MEMSET with zero constant.
    getActionDefinitionsBuilder(G_BZERO).unsupported();

    getActionDefinitionsBuilder(G_MEMSET)
        .legalForCartesianProduct({p0}, {s64}, {s64})
        .customForCartesianProduct({p0}, {s8}, {s64})
        .immIdx(0); // Inform verifier imm idx 0 is handled.

    getActionDefinitionsBuilder({G_MEMCPY, G_MEMMOVE})
        .legalForCartesianProduct({p0}, {p0}, {s64})
        .immIdx(0); // Inform verifier imm idx 0 is handled.

    // G_MEMCPY_INLINE does not have a tailcall immediate
    getActionDefinitionsBuilder(G_MEMCPY_INLINE)
        .legalForCartesianProduct({p0}, {p0}, {s64});

  } else {
    getActionDefinitionsBuilder({G_BZERO, G_MEMCPY, G_MEMMOVE, G_MEMSET})
        .libcall();
  }

  // FIXME: Legal types are only legal with NEON.
  getActionDefinitionsBuilder(G_ABS)
      .lowerIf(isScalar(0))
      .legalFor(PackedVectorAllTypeList);

  getActionDefinitionsBuilder(G_VECREDUCE_FADD)
      // We only have FADDP to do reduction-like operations. Lower the rest.
      .legalFor({{s32, v2s32}, {s64, v2s64}})
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 2)
      .lower();

  getActionDefinitionsBuilder(G_VECREDUCE_ADD)
      .legalFor(
          {{s8, v16s8}, {s16, v8s16}, {s32, v4s32}, {s32, v2s32}, {s64, v2s64}})
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .lower();

  getActionDefinitionsBuilder(
      {G_VECREDUCE_OR, G_VECREDUCE_AND, G_VECREDUCE_XOR})
      // Try to break down into smaller vectors as long as they're at least 64
      // bits. This lets us use vector operations for some parts of the
      // reduction.
      .fewerElementsIf(
          [=](const LegalityQuery &Q) {
            LLT SrcTy = Q.Types[1];
            if (SrcTy.isScalar())
              return false;
            if (!isPowerOf2_32(SrcTy.getNumElements()))
              return false;
            // We can usually perform 64b vector operations.
            return SrcTy.getSizeInBits() > 64;
          },
          [=](const LegalityQuery &Q) {
            LLT SrcTy = Q.Types[1];
            return std::make_pair(1, SrcTy.divide(2));
          })
      .scalarize(1)
      .lower();

  getActionDefinitionsBuilder({G_UADDSAT, G_USUBSAT})
      .lowerIf([=](const LegalityQuery &Q) { return Q.Types[0].isScalar(); });

  getActionDefinitionsBuilder({G_FSHL, G_FSHR}).lower();

  getActionDefinitionsBuilder(G_ROTR)
      .legalFor({{s32, s64}, {s64, s64}})
      .customIf([=](const LegalityQuery &Q) {
        return Q.Types[0].isScalar() && Q.Types[1].getScalarSizeInBits() < 64;
      })
      .lower();
  getActionDefinitionsBuilder(G_ROTL).lower();

  getActionDefinitionsBuilder({G_SBFX, G_UBFX})
      .customFor({{s32, s32}, {s64, s64}});

  // TODO: Use generic lowering when custom lowering is not possible.
  auto always = [=](const LegalityQuery &Q) { return true; };
  getActionDefinitionsBuilder(G_CTPOP)
      .legalFor({{v8s8, v8s8}, {v16s8, v16s8}})
      .clampScalar(0, s32, s128)
      .widenScalarToNextPow2(0)
      .minScalarEltSameAsIf(always, 1, 0)
      .maxScalarEltSameAsIf(always, 1, 0)
      .customFor({{s32, s32},
                  {s64, s64},
                  {s128, s128},
                  {v2s64, v2s64},
                  {v2s32, v2s32},
                  {v4s32, v4s32},
                  {v4s16, v4s16},
                  {v8s16, v8s16}});

  // TODO: Vector types.
  getActionDefinitionsBuilder({G_SADDSAT, G_SSUBSAT}).lowerIf(isScalar(0));

  // TODO: Vector types.
  getActionDefinitionsBuilder({G_FMAXNUM, G_FMINNUM})
      .legalFor({MinFPScalar, s32, s64})
      .libcallFor({s128})
      .minScalar(0, MinFPScalar);

  // TODO: Vector types.
  getActionDefinitionsBuilder({G_FMAXIMUM, G_FMINIMUM})
      .legalFor({MinFPScalar, s32, s64})
      .minScalar(0, MinFPScalar);

  // TODO: Libcall support for s128.
  // TODO: s16 should be legal with full FP16 support.
  getActionDefinitionsBuilder({G_LROUND, G_LLROUND})
      .legalFor({{s64, s32}, {s64, s64}});

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

bool AArch64LegalizerInfo::legalizeCustom(LegalizerHelper &Helper,
                                          MachineInstr &MI) const {
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
  GISelChangeObserver &Observer = Helper.Observer;
  switch (MI.getOpcode()) {
  default:
    // No idea what to do.
    return false;
  case TargetOpcode::G_VAARG:
    return legalizeVaArg(MI, MRI, MIRBuilder);
  case TargetOpcode::G_LOAD:
  case TargetOpcode::G_STORE:
    return legalizeLoadStore(MI, MRI, MIRBuilder, Observer);
  case TargetOpcode::G_SHL:
  case TargetOpcode::G_ASHR:
  case TargetOpcode::G_LSHR:
    return legalizeShlAshrLshr(MI, MRI, MIRBuilder, Observer);
  case TargetOpcode::G_GLOBAL_VALUE:
    return legalizeSmallCMGlobalValue(MI, MRI, MIRBuilder, Observer);
  case TargetOpcode::G_TRUNC:
    return legalizeVectorTrunc(MI, Helper);
  case TargetOpcode::G_SBFX:
  case TargetOpcode::G_UBFX:
    return legalizeBitfieldExtract(MI, MRI, Helper);
  case TargetOpcode::G_ROTR:
    return legalizeRotate(MI, MRI, Helper);
  case TargetOpcode::G_CTPOP:
    return legalizeCTPOP(MI, MRI, Helper);
  case TargetOpcode::G_ATOMIC_CMPXCHG:
    return legalizeAtomicCmpxchg128(MI, MRI, Helper);
  case TargetOpcode::G_CTTZ:
    return legalizeCTTZ(MI, Helper);
  case TargetOpcode::G_BZERO:
  case TargetOpcode::G_MEMCPY:
  case TargetOpcode::G_MEMMOVE:
  case TargetOpcode::G_MEMSET:
    return legalizeMemOps(MI, Helper);
  }

  llvm_unreachable("expected switch to return");
}

bool AArch64LegalizerInfo::legalizeRotate(MachineInstr &MI,
                                          MachineRegisterInfo &MRI,
                                          LegalizerHelper &Helper) const {
  // To allow for imported patterns to match, we ensure that the rotate amount
  // is 64b with an extension.
  Register AmtReg = MI.getOperand(2).getReg();
  LLT AmtTy = MRI.getType(AmtReg);
  (void)AmtTy;
  assert(AmtTy.isScalar() && "Expected a scalar rotate");
  assert(AmtTy.getSizeInBits() < 64 && "Expected this rotate to be legal");
  auto NewAmt = Helper.MIRBuilder.buildSExt(LLT::scalar(64), AmtReg);
  Helper.Observer.changingInstr(MI);
  MI.getOperand(2).setReg(NewAmt.getReg(0));
  Helper.Observer.changedInstr(MI);
  return true;
}

static void extractParts(Register Reg, MachineRegisterInfo &MRI,
                         MachineIRBuilder &MIRBuilder, LLT Ty, int NumParts,
                         SmallVectorImpl<Register> &VRegs) {
  for (int I = 0; I < NumParts; ++I)
    VRegs.push_back(MRI.createGenericVirtualRegister(Ty));
  MIRBuilder.buildUnmerge(VRegs, Reg);
}

bool AArch64LegalizerInfo::legalizeVectorTrunc(
    MachineInstr &MI, LegalizerHelper &Helper) const {
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
  // Similar to how operand splitting is done in SelectiondDAG, we can handle
  // %res(v8s8) = G_TRUNC %in(v8s32) by generating:
  //   %inlo(<4x s32>), %inhi(<4 x s32>) = G_UNMERGE %in(<8 x s32>)
  //   %lo16(<4 x s16>) = G_TRUNC %inlo
  //   %hi16(<4 x s16>) = G_TRUNC %inhi
  //   %in16(<8 x s16>) = G_CONCAT_VECTORS %lo16, %hi16
  //   %res(<8 x s8>) = G_TRUNC %in16

  Register DstReg = MI.getOperand(0).getReg();
  Register SrcReg = MI.getOperand(1).getReg();
  LLT DstTy = MRI.getType(DstReg);
  LLT SrcTy = MRI.getType(SrcReg);
  assert(isPowerOf2_32(DstTy.getSizeInBits()) &&
         isPowerOf2_32(SrcTy.getSizeInBits()));

  // Split input type.
  LLT SplitSrcTy =
      SrcTy.changeElementCount(SrcTy.getElementCount().divideCoefficientBy(2));
  // First, split the source into two smaller vectors.
  SmallVector<Register, 2> SplitSrcs;
  extractParts(SrcReg, MRI, MIRBuilder, SplitSrcTy, 2, SplitSrcs);

  // Truncate the splits into intermediate narrower elements.
  LLT InterTy = SplitSrcTy.changeElementSize(DstTy.getScalarSizeInBits() * 2);
  for (unsigned I = 0; I < SplitSrcs.size(); ++I)
    SplitSrcs[I] = MIRBuilder.buildTrunc(InterTy, SplitSrcs[I]).getReg(0);

  auto Concat = MIRBuilder.buildConcatVectors(
      DstTy.changeElementSize(DstTy.getScalarSizeInBits() * 2), SplitSrcs);

  Helper.Observer.changingInstr(MI);
  MI.getOperand(1).setReg(Concat.getReg(0));
  Helper.Observer.changedInstr(MI);
  return true;
}

bool AArch64LegalizerInfo::legalizeSmallCMGlobalValue(
    MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
    GISelChangeObserver &Observer) const {
  assert(MI.getOpcode() == TargetOpcode::G_GLOBAL_VALUE);
  // We do this custom legalization to convert G_GLOBAL_VALUE into target ADRP +
  // G_ADD_LOW instructions.
  // By splitting this here, we can optimize accesses in the small code model by
  // folding in the G_ADD_LOW into the load/store offset.
  auto &GlobalOp = MI.getOperand(1);
  const auto* GV = GlobalOp.getGlobal();
  if (GV->isThreadLocal())
    return true; // Don't want to modify TLS vars.

  auto &TM = ST->getTargetLowering()->getTargetMachine();
  unsigned OpFlags = ST->ClassifyGlobalReference(GV, TM);

  if (OpFlags & AArch64II::MO_GOT)
    return true;

  auto Offset = GlobalOp.getOffset();
  Register DstReg = MI.getOperand(0).getReg();
  auto ADRP = MIRBuilder.buildInstr(AArch64::ADRP, {LLT::pointer(0, 64)}, {})
                  .addGlobalAddress(GV, Offset, OpFlags | AArch64II::MO_PAGE);
  // Set the regclass on the dest reg too.
  MRI.setRegClass(ADRP.getReg(0), &AArch64::GPR64RegClass);

  // MO_TAGGED on the page indicates a tagged address. Set the tag now. We do so
  // by creating a MOVK that sets bits 48-63 of the register to (global address
  // + 0x100000000 - PC) >> 48. The additional 0x100000000 offset here is to
  // prevent an incorrect tag being generated during relocation when the the
  // global appears before the code section. Without the offset, a global at
  // `0x0f00'0000'0000'1000` (i.e. at `0x1000` with tag `0xf`) that's referenced
  // by code at `0x2000` would result in `0x0f00'0000'0000'1000 - 0x2000 =
  // 0x0eff'ffff'ffff'f000`, meaning the tag would be incorrectly set to `0xe`
  // instead of `0xf`.
  // This assumes that we're in the small code model so we can assume a binary
  // size of <= 4GB, which makes the untagged PC relative offset positive. The
  // binary must also be loaded into address range [0, 2^48). Both of these
  // properties need to be ensured at runtime when using tagged addresses.
  if (OpFlags & AArch64II::MO_TAGGED) {
    assert(!Offset &&
           "Should not have folded in an offset for a tagged global!");
    ADRP = MIRBuilder.buildInstr(AArch64::MOVKXi, {LLT::pointer(0, 64)}, {ADRP})
               .addGlobalAddress(GV, 0x100000000,
                                 AArch64II::MO_PREL | AArch64II::MO_G3)
               .addImm(48);
    MRI.setRegClass(ADRP.getReg(0), &AArch64::GPR64RegClass);
  }

  MIRBuilder.buildInstr(AArch64::G_ADD_LOW, {DstReg}, {ADRP})
      .addGlobalAddress(GV, Offset,
                        OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
  MI.eraseFromParent();
  return true;
}

bool AArch64LegalizerInfo::legalizeIntrinsic(LegalizerHelper &Helper,
                                             MachineInstr &MI) const {
  switch (MI.getIntrinsicID()) {
  case Intrinsic::vacopy: {
    unsigned PtrSize = ST->isTargetILP32() ? 4 : 8;
    unsigned VaListSize =
      (ST->isTargetDarwin() || ST->isTargetWindows())
          ? PtrSize
          : ST->isTargetILP32() ? 20 : 32;

    MachineFunction &MF = *MI.getMF();
    auto Val = MF.getRegInfo().createGenericVirtualRegister(
        LLT::scalar(VaListSize * 8));
    MachineIRBuilder MIB(MI);
    MIB.buildLoad(Val, MI.getOperand(2),
                  *MF.getMachineMemOperand(MachinePointerInfo(),
                                           MachineMemOperand::MOLoad,
                                           VaListSize, Align(PtrSize)));
    MIB.buildStore(Val, MI.getOperand(1),
                   *MF.getMachineMemOperand(MachinePointerInfo(),
                                            MachineMemOperand::MOStore,
                                            VaListSize, Align(PtrSize)));
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::get_dynamic_area_offset: {
    MachineIRBuilder &MIB = Helper.MIRBuilder;
    MIB.buildConstant(MI.getOperand(0).getReg(), 0);
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_mops_memset_tag: {
    assert(MI.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS);
    // Zext the value to 64 bit
    MachineIRBuilder MIB(MI);
    auto &Value = MI.getOperand(3);
    Register ZExtValueReg = MIB.buildAnyExt(LLT::scalar(64), Value).getReg(0);
    Value.setReg(ZExtValueReg);
    return true;
  }
  }

  return true;
}

bool AArch64LegalizerInfo::legalizeShlAshrLshr(
    MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
    GISelChangeObserver &Observer) const {
  assert(MI.getOpcode() == TargetOpcode::G_ASHR ||
         MI.getOpcode() == TargetOpcode::G_LSHR ||
         MI.getOpcode() == TargetOpcode::G_SHL);
  // If the shift amount is a G_CONSTANT, promote it to a 64 bit type so the
  // imported patterns can select it later. Either way, it will be legal.
  Register AmtReg = MI.getOperand(2).getReg();
  auto VRegAndVal = getIConstantVRegValWithLookThrough(AmtReg, MRI);
  if (!VRegAndVal)
    return true;
  // Check the shift amount is in range for an immediate form.
  int64_t Amount = VRegAndVal->Value.getSExtValue();
  if (Amount > 31)
    return true; // This will have to remain a register variant.
  auto ExtCst = MIRBuilder.buildConstant(LLT::scalar(64), Amount);
  Observer.changingInstr(MI);
  MI.getOperand(2).setReg(ExtCst.getReg(0));
  Observer.changedInstr(MI);
  return true;
}

static void matchLDPSTPAddrMode(Register Root, Register &Base, int &Offset,
                                MachineRegisterInfo &MRI) {
  Base = Root;
  Offset = 0;

  Register NewBase;
  int64_t NewOffset;
  if (mi_match(Root, MRI, m_GPtrAdd(m_Reg(NewBase), m_ICst(NewOffset))) &&
      isShiftedInt<7, 3>(NewOffset)) {
    Base = NewBase;
    Offset = NewOffset;
  }
}

// FIXME: This should be removed and replaced with the generic bitcast legalize
// action.
bool AArch64LegalizerInfo::legalizeLoadStore(
    MachineInstr &MI, MachineRegisterInfo &MRI, MachineIRBuilder &MIRBuilder,
    GISelChangeObserver &Observer) const {
  assert(MI.getOpcode() == TargetOpcode::G_STORE ||
         MI.getOpcode() == TargetOpcode::G_LOAD);
  // Here we just try to handle vector loads/stores where our value type might
  // have pointer elements, which the SelectionDAG importer can't handle. To
  // allow the existing patterns for s64 to fire for p0, we just try to bitcast
  // the value to use s64 types.

  // Custom legalization requires the instruction, if not deleted, must be fully
  // legalized. In order to allow further legalization of the inst, we create
  // a new instruction and erase the existing one.

  Register ValReg = MI.getOperand(0).getReg();
  const LLT ValTy = MRI.getType(ValReg);

  if (ValTy == LLT::scalar(128)) {
    assert((*MI.memoperands_begin())->getSuccessOrdering() ==
               AtomicOrdering::Monotonic ||
           (*MI.memoperands_begin())->getSuccessOrdering() ==
               AtomicOrdering::Unordered);
    assert(ST->hasLSE2() && "ldp/stp not single copy atomic without +lse2");
    LLT s64 = LLT::scalar(64);
    MachineInstrBuilder NewI;
    if (MI.getOpcode() == TargetOpcode::G_LOAD) {
      NewI = MIRBuilder.buildInstr(AArch64::LDPXi, {s64, s64}, {});
      MIRBuilder.buildMerge(ValReg, {NewI->getOperand(0), NewI->getOperand(1)});
    } else {
      auto Split = MIRBuilder.buildUnmerge(s64, MI.getOperand(0));
      NewI = MIRBuilder.buildInstr(
          AArch64::STPXi, {}, {Split->getOperand(0), Split->getOperand(1)});
    }
    Register Base;
    int Offset;
    matchLDPSTPAddrMode(MI.getOperand(1).getReg(), Base, Offset, MRI);
    NewI.addUse(Base);
    NewI.addImm(Offset / 8);

    NewI.cloneMemRefs(MI);
    constrainSelectedInstRegOperands(*NewI, *ST->getInstrInfo(),
                                     *MRI.getTargetRegisterInfo(),
                                     *ST->getRegBankInfo());
    MI.eraseFromParent();
    return true;
  }

  if (!ValTy.isVector() || !ValTy.getElementType().isPointer() ||
      ValTy.getElementType().getAddressSpace() != 0) {
    LLVM_DEBUG(dbgs() << "Tried to do custom legalization on wrong load/store");
    return false;
  }

  unsigned PtrSize = ValTy.getElementType().getSizeInBits();
  const LLT NewTy = LLT::vector(ValTy.getElementCount(), PtrSize);
  auto &MMO = **MI.memoperands_begin();
  MMO.setType(NewTy);

  if (MI.getOpcode() == TargetOpcode::G_STORE) {
    auto Bitcast = MIRBuilder.buildBitcast(NewTy, ValReg);
    MIRBuilder.buildStore(Bitcast.getReg(0), MI.getOperand(1), MMO);
  } else {
    auto NewLoad = MIRBuilder.buildLoad(NewTy, MI.getOperand(1), MMO);
    MIRBuilder.buildBitcast(ValReg, NewLoad);
  }
  MI.eraseFromParent();
  return true;
}

bool AArch64LegalizerInfo::legalizeVaArg(MachineInstr &MI,
                                         MachineRegisterInfo &MRI,
                                         MachineIRBuilder &MIRBuilder) const {
  MachineFunction &MF = MIRBuilder.getMF();
  Align Alignment(MI.getOperand(2).getImm());
  Register Dst = MI.getOperand(0).getReg();
  Register ListPtr = MI.getOperand(1).getReg();

  LLT PtrTy = MRI.getType(ListPtr);
  LLT IntPtrTy = LLT::scalar(PtrTy.getSizeInBits());

  const unsigned PtrSize = PtrTy.getSizeInBits() / 8;
  const Align PtrAlign = Align(PtrSize);
  auto List = MIRBuilder.buildLoad(
      PtrTy, ListPtr,
      *MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
                               PtrTy, PtrAlign));

  MachineInstrBuilder DstPtr;
  if (Alignment > PtrAlign) {
    // Realign the list to the actual required alignment.
    auto AlignMinus1 =
        MIRBuilder.buildConstant(IntPtrTy, Alignment.value() - 1);
    auto ListTmp = MIRBuilder.buildPtrAdd(PtrTy, List, AlignMinus1.getReg(0));
    DstPtr = MIRBuilder.buildMaskLowPtrBits(PtrTy, ListTmp, Log2(Alignment));
  } else
    DstPtr = List;

  LLT ValTy = MRI.getType(Dst);
  uint64_t ValSize = ValTy.getSizeInBits() / 8;
  MIRBuilder.buildLoad(
      Dst, DstPtr,
      *MF.getMachineMemOperand(MachinePointerInfo(), MachineMemOperand::MOLoad,
                               ValTy, std::max(Alignment, PtrAlign)));

  auto Size = MIRBuilder.buildConstant(IntPtrTy, alignTo(ValSize, PtrAlign));

  auto NewList = MIRBuilder.buildPtrAdd(PtrTy, DstPtr, Size.getReg(0));

  MIRBuilder.buildStore(NewList, ListPtr,
                        *MF.getMachineMemOperand(MachinePointerInfo(),
                                                 MachineMemOperand::MOStore,
                                                 PtrTy, PtrAlign));

  MI.eraseFromParent();
  return true;
}

bool AArch64LegalizerInfo::legalizeBitfieldExtract(
    MachineInstr &MI, MachineRegisterInfo &MRI, LegalizerHelper &Helper) const {
  // Only legal if we can select immediate forms.
  // TODO: Lower this otherwise.
  return getIConstantVRegValWithLookThrough(MI.getOperand(2).getReg(), MRI) &&
         getIConstantVRegValWithLookThrough(MI.getOperand(3).getReg(), MRI);
}

bool AArch64LegalizerInfo::legalizeCTPOP(MachineInstr &MI,
                                         MachineRegisterInfo &MRI,
                                         LegalizerHelper &Helper) const {
  // While there is no integer popcount instruction, it can
  // be more efficiently lowered to the following sequence that uses
  // AdvSIMD registers/instructions as long as the copies to/from
  // the AdvSIMD registers are cheap.
  //  FMOV    D0, X0        // copy 64-bit int to vector, high bits zero'd
  //  CNT     V0.8B, V0.8B  // 8xbyte pop-counts
  //  ADDV    B0, V0.8B     // sum 8xbyte pop-counts
  //  UMOV    X0, V0.B[0]   // copy byte result back to integer reg
  //
  // For 128 bit vector popcounts, we lower to the following sequence:
  //  cnt.16b   v0, v0  // v8s16, v4s32, v2s64
  //  uaddlp.8h v0, v0  // v8s16, v4s32, v2s64
  //  uaddlp.4s v0, v0  //        v4s32, v2s64
  //  uaddlp.2d v0, v0  //               v2s64
  //
  // For 64 bit vector popcounts, we lower to the following sequence:
  //  cnt.8b    v0, v0  // v4s16, v2s32
  //  uaddlp.4h v0, v0  // v4s16, v2s32
  //  uaddlp.2s v0, v0  //        v2s32

  if (!ST->hasNEON() ||
      MI.getMF()->getFunction().hasFnAttribute(Attribute::NoImplicitFloat))
    return false;
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
  Register Dst = MI.getOperand(0).getReg();
  Register Val = MI.getOperand(1).getReg();
  LLT Ty = MRI.getType(Val);

  assert(Ty == MRI.getType(Dst) &&
         "Expected src and dst to have the same type!");
  unsigned Size = Ty.getSizeInBits();

  // Pre-conditioning: widen Val up to the nearest vector type.
  // s32,s64,v4s16,v2s32 -> v8i8
  // v8s16,v4s32,v2s64 -> v16i8
  LLT VTy = Size == 128 ? LLT::fixed_vector(16, 8) : LLT::fixed_vector(8, 8);
  if (Ty.isScalar()) {
    assert((Size == 32 || Size == 64 || Size == 128) && "Expected only 32, 64, or 128 bit scalars!");
    if (Size == 32) {
      Val = MIRBuilder.buildZExt(LLT::scalar(64), Val).getReg(0);
    }
  }
  Val = MIRBuilder.buildBitcast(VTy, Val).getReg(0);

  // Count bits in each byte-sized lane.
  auto CTPOP = MIRBuilder.buildCTPOP(VTy, Val);

  // Sum across lanes.
  Register HSum = CTPOP.getReg(0);
  unsigned Opc;
  SmallVector<LLT> HAddTys;
  if (Ty.isScalar()) {
    Opc = Intrinsic::aarch64_neon_uaddlv;
    HAddTys.push_back(LLT::scalar(32));
  } else if (Ty == LLT::fixed_vector(8, 16)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(8, 16));
  } else if (Ty == LLT::fixed_vector(4, 32)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(8, 16));
    HAddTys.push_back(LLT::fixed_vector(4, 32));
  } else if (Ty == LLT::fixed_vector(2, 64)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(8, 16));
    HAddTys.push_back(LLT::fixed_vector(4, 32));
    HAddTys.push_back(LLT::fixed_vector(2, 64));
  } else if (Ty == LLT::fixed_vector(4, 16)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(4, 16));
  } else if (Ty == LLT::fixed_vector(2, 32)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(4, 16));
    HAddTys.push_back(LLT::fixed_vector(2, 32));
  } else
    llvm_unreachable("unexpected vector shape");
  MachineInstrBuilder UADD;
  for (LLT HTy : HAddTys) {
    UADD = MIRBuilder.buildIntrinsic(Opc, {HTy}, /*HasSideEffects =*/false)
                     .addUse(HSum);
    HSum = UADD.getReg(0);
  }

  // Post-conditioning.
  if (Ty.isScalar() && (Size == 64 || Size == 128))
    MIRBuilder.buildZExt(Dst, UADD);
  else
    UADD->getOperand(0).setReg(Dst);
  MI.eraseFromParent();
  return true;
}

bool AArch64LegalizerInfo::legalizeAtomicCmpxchg128(
    MachineInstr &MI, MachineRegisterInfo &MRI, LegalizerHelper &Helper) const {
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
  LLT s64 = LLT::scalar(64);
  auto Addr = MI.getOperand(1).getReg();
  auto DesiredI = MIRBuilder.buildUnmerge({s64, s64}, MI.getOperand(2));
  auto NewI = MIRBuilder.buildUnmerge({s64, s64}, MI.getOperand(3));
  auto DstLo = MRI.createGenericVirtualRegister(s64);
  auto DstHi = MRI.createGenericVirtualRegister(s64);

  MachineInstrBuilder CAS;
  if (ST->hasLSE()) {
    // We have 128-bit CASP instructions taking XSeqPair registers, which are
    // s128. We need the merge/unmerge to bracket the expansion and pair up with
    // the rest of the MIR so we must reassemble the extracted registers into a
    // 128-bit known-regclass one with code like this:
    //
    //     %in1 = REG_SEQUENCE Lo, Hi    ; One for each input
    //     %out = CASP %in1, ...
    //     %OldLo = G_EXTRACT %out, 0
    //     %OldHi = G_EXTRACT %out, 64
    auto Ordering = (*MI.memoperands_begin())->getMergedOrdering();
    unsigned Opcode;
    switch (Ordering) {
    case AtomicOrdering::Acquire:
      Opcode = AArch64::CASPAX;
      break;
    case AtomicOrdering::Release:
      Opcode = AArch64::CASPLX;
      break;
    case AtomicOrdering::AcquireRelease:
    case AtomicOrdering::SequentiallyConsistent:
      Opcode = AArch64::CASPALX;
      break;
    default:
      Opcode = AArch64::CASPX;
      break;
    }

    LLT s128 = LLT::scalar(128);
    auto CASDst = MRI.createGenericVirtualRegister(s128);
    auto CASDesired = MRI.createGenericVirtualRegister(s128);
    auto CASNew = MRI.createGenericVirtualRegister(s128);
    MIRBuilder.buildInstr(TargetOpcode::REG_SEQUENCE, {CASDesired}, {})
        .addUse(DesiredI->getOperand(0).getReg())
        .addImm(AArch64::sube64)
        .addUse(DesiredI->getOperand(1).getReg())
        .addImm(AArch64::subo64);
    MIRBuilder.buildInstr(TargetOpcode::REG_SEQUENCE, {CASNew}, {})
        .addUse(NewI->getOperand(0).getReg())
        .addImm(AArch64::sube64)
        .addUse(NewI->getOperand(1).getReg())
        .addImm(AArch64::subo64);

    CAS = MIRBuilder.buildInstr(Opcode, {CASDst}, {CASDesired, CASNew, Addr});

    MIRBuilder.buildExtract({DstLo}, {CASDst}, 0);
    MIRBuilder.buildExtract({DstHi}, {CASDst}, 64);
  } else {
    // The -O0 CMP_SWAP_128 is friendlier to generate code for because LDXP/STXP
    // can take arbitrary registers so it just has the normal GPR64 operands the
    // rest of AArch64 is expecting.
    auto Ordering = (*MI.memoperands_begin())->getMergedOrdering();
    unsigned Opcode;
    switch (Ordering) {
    case AtomicOrdering::Acquire:
      Opcode = AArch64::CMP_SWAP_128_ACQUIRE;
      break;
    case AtomicOrdering::Release:
      Opcode = AArch64::CMP_SWAP_128_RELEASE;
      break;
    case AtomicOrdering::AcquireRelease:
    case AtomicOrdering::SequentiallyConsistent:
      Opcode = AArch64::CMP_SWAP_128;
      break;
    default:
      Opcode = AArch64::CMP_SWAP_128_MONOTONIC;
      break;
    }

    auto Scratch = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
    CAS = MIRBuilder.buildInstr(Opcode, {DstLo, DstHi, Scratch},
                                {Addr, DesiredI->getOperand(0),
                                 DesiredI->getOperand(1), NewI->getOperand(0),
                                 NewI->getOperand(1)});
  }

  CAS.cloneMemRefs(MI);
  constrainSelectedInstRegOperands(*CAS, *ST->getInstrInfo(),
                                   *MRI.getTargetRegisterInfo(),
                                   *ST->getRegBankInfo());

  MIRBuilder.buildMerge(MI.getOperand(0), {DstLo, DstHi});
  MI.eraseFromParent();
  return true;
}

bool AArch64LegalizerInfo::legalizeCTTZ(MachineInstr &MI,
                                        LegalizerHelper &Helper) const {
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
  LLT Ty = MRI.getType(MI.getOperand(1).getReg());
  auto BitReverse = MIRBuilder.buildBitReverse(Ty, MI.getOperand(1));
  MIRBuilder.buildCTLZ(MI.getOperand(0).getReg(), BitReverse);
  MI.eraseFromParent();
  return true;
}

bool AArch64LegalizerInfo::legalizeMemOps(MachineInstr &MI,
                                          LegalizerHelper &Helper) const {
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;

  // Tagged version MOPSMemorySetTagged is legalised in legalizeIntrinsic
  if (MI.getOpcode() == TargetOpcode::G_MEMSET) {
    // Zext the value operand to 64 bit
    auto &Value = MI.getOperand(1);
    Register ZExtValueReg =
        MIRBuilder.buildAnyExt(LLT::scalar(64), Value).getReg(0);
    Value.setReg(ZExtValueReg);
    return true;
  }

  return false;
}