AArch64LegalizerInfo.cpp

Go to the documentation of this file.

//===- 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 "AArch64Subtarget.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/CodeGen/GlobalISel/GenericMachineInstrs.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/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetOpcodes.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 v2s8 = LLT::fixed_vector(2, 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);
 
  const LLT nxv16s8 = LLT::scalable_vector(16, s8);
  const LLT nxv8s16 = LLT::scalable_vector(8, s16);
  const LLT nxv4s32 = LLT::scalable_vector(4, s32);
  const LLT nxv2s64 = LLT::scalable_vector(2, s64);
 
  const LLT bf16 = LLT::bfloat16();
  const LLT v4bf16 = LLT::fixed_vector(4, bf16);
  const LLT v8bf16 = LLT::fixed_vector(8, bf16);
 
  const LLT f16 = LLT::float16();
  const LLT v4f16 = LLT::fixed_vector(4, f16);
  const LLT v8f16 = LLT::fixed_vector(8, f16);
 
  const LLT f32 = LLT::float32();
  const LLT v2f32 = LLT::fixed_vector(2, f32);
  const LLT v4f32 = LLT::fixed_vector(4, f32);
 
  const LLT f64 = LLT::float64();
  const LLT v2f64 = LLT::fixed_vector(2, f64);
 
  const LLT f128 = LLT::float128();
 
  const LLT i8 = LLT::integer(8);
  const LLT v8i8 = LLT::fixed_vector(8, i8);
  const LLT v16i8 = LLT::fixed_vector(16, i8);
 
  const LLT i16 = LLT::integer(16);
  const LLT v8i16 = LLT::fixed_vector(8, i16);
  const LLT v4i16 = LLT::fixed_vector(4, i16);
 
  const LLT i32 = LLT::integer(32);
  const LLT v2i32 = LLT::fixed_vector(2, i32);
  const LLT v4i32 = LLT::fixed_vector(4, i32);
 
  const LLT i64 = LLT::integer(64);
  const LLT v2i64 = LLT::fixed_vector(2, i64);
 
  const LLT i128 = LLT::integer(128);
 
  const LLT nxv16i8 = LLT::scalable_vector(16, i8);
  const LLT nxv8i16 = LLT::scalable_vector(8, i16);
  const LLT nxv4i32 = LLT::scalable_vector(4, i32);
  const LLT nxv2i64 = LLT::scalable_vector(2, i64);
 
  std::initializer_list<LLT> PackedVectorAllTypeList = {/* Begin 128bit types */
                                                        v16s8, v8s16, v4s32,
                                                        v2s64, v2p0,
                                                        /* End 128bit types */
                                                        /* Begin 64bit types */
                                                        v8s8, v4s16, v2s32};
  std::initializer_list<LLT> ScalarAndPtrTypesList = {s8, s16, s32, s64, p0};
  SmallVector<LLT, 8> PackedVectorAllTypesVec(PackedVectorAllTypeList);
  SmallVector<LLT, 8> ScalarAndPtrTypesVec(ScalarAndPtrTypesList);
 
  const TargetMachine &TM = ST.getTargetLowering()->getTargetMachine();
 
  // FIXME: support subtargets which have neon/fp-armv8 disabled.
  if (!ST.hasNEON() || !ST.hasFPARMv8())
    return;
 
  // Some instructions only support s16 if the subtarget has full 16-bit FP
  // support.
  const bool HasFP16 = ST.hasFullFP16();
  const bool HasCSSC = ST.hasCSSC();
  const bool HasRCPC3 = ST.hasRCPC3();
  const bool HasSVE = ST.hasSVE();
 
  getActionDefinitionsBuilder(
      {G_IMPLICIT_DEF, G_FREEZE, G_CONSTANT_FOLD_BARRIER})
      .legalFor({p0, s8, s16, s32, s64, s128})
      .legalFor({v2s8, v4s8, v8s8, v16s8, v2s16, v4s16, v8s16, v2s32, v4s32,
                 v2s64, v2p0})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s8, s64)
      .moreElementsToNextPow2(0)
      .widenVectorEltsToVectorMinSize(0, 64)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .clampMaxNumElements(0, p0, 2)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0);
 
  getActionDefinitionsBuilder(G_PHI)
      .legalFor({p0, s16, s32, s64})
      .legalFor(PackedVectorAllTypeList)
      .widenScalarToNextPow2(0)
      .moreElementsToNextPow2(0)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .clampScalar(0, s16, s64)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .clampMaxNumElements(0, p0, 2)
      .widenScalarOrEltToNextPow2OrMinSize(0, 8);
 
  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);
 
  getActionDefinitionsBuilder({G_ADD, G_SUB, G_AND, G_OR, G_XOR})
      .legalFor({i32, i64, v8i8, v16i8, v4i16, v8i16, v2i32, v4i32, v2i64})
      .legalFor(HasSVE, {nxv16i8, nxv8i16, nxv4i32, nxv2i64})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64)
      .clampMaxNumElements(0, s8, 16)
      .clampMaxNumElements(0, s16, 8)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() <= 2;
          },
          0, s32)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() <= 4;
          },
          0, s16)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() <= 16;
          },
          0, s8)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .moreElementsToNextPow2(0);
 
  getActionDefinitionsBuilder(G_MUL)
      .legalFor({i32, i64, v8i8, v16i8, v4i16, v8i16, v2i32, v4i32, v2i64})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s32, s64)
      .clampMaxNumElements(0, s8, 16)
      .clampMaxNumElements(0, s16, 8)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() <= 2;
          },
          0, s32)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() <= 4;
          },
          0, s16)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() <= 16;
          },
          0, s8)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .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({
          {i32, i32},
          {i32, i64},
          {i64, i64},
          {v8i8, v8i8},
          {v16i8, v16i8},
          {v4i16, v4i16},
          {v8i16, v8i16},
          {v2i32, v2i32},
          {v4i32, v4i32},
          {v2i64, v2i64},
      })
      .widenScalarToNextPow2(1)
      .widenScalarToNextPow2(0)
      .clampScalar(1, s32, s64)
      .clampScalar(0, s32, s64)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0)
      .minScalarSameAs(1, 0)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .minScalarEltSameAsIf(isVector(0), 1, 0)
      .maxScalarEltSameAsIf(isVector(0), 1, 0);
 
  getActionDefinitionsBuilder(G_PTR_ADD)
      .legalFor({{p0, i64}, {v2p0, v2i64}})
      .clampScalarOrElt(1, s64, s64)
      .clampNumElements(0, v2p0, v2p0);
 
  getActionDefinitionsBuilder(G_PTRMASK).legalFor({{p0, s64}});
 
  getActionDefinitionsBuilder({G_SDIV, G_UDIV})
      .legalFor({i32, i64})
      .libcallFor({i128})
      .clampScalar(0, s32, s64)
      .widenScalarToNextPow2(0)
      .scalarize(0);
 
  getActionDefinitionsBuilder({G_SREM, G_UREM, G_SDIVREM, G_UDIVREM})
      .lowerFor({i8, i16, i32, i64, v2i32, v4i32, v2i64})
      .libcallFor({i128})
      .widenScalarOrEltToNextPow2(0)
      .minScalarOrElt(0, s32)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .scalarize(0);
 
  getActionDefinitionsBuilder({G_SMULO, G_UMULO})
      .widenScalarToNextPow2(0, /*Min = */ 32)
      .clampScalar(0, s32, s64)
      .lower();
 
  getActionDefinitionsBuilder({G_SMULH, G_UMULH})
      .legalFor({i64, v16i8, v8i16, v4i32})
      .lower();
 
  getActionDefinitionsBuilder(
      {G_SMULFIX, G_UMULFIX, G_SMULFIXSAT, G_UMULFIXSAT})
      .lower();
 
  getActionDefinitionsBuilder({G_SMIN, G_SMAX, G_UMIN, G_UMAX})
      .legalFor({v8i8, v16i8, v4i16, v8i16, v2i32, v4i32})
      .legalFor(HasCSSC, {i32, i64})
      .minScalar(HasCSSC, 0, s32)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .lower();
 
  // FIXME: Legal vector types are only legal with NEON.
  getActionDefinitionsBuilder(G_ABS)
      .legalFor(HasCSSC, {i32, i64})
      .legalFor({v16i8, v8i16, v4i32, v2i64, v2p0, v8i8, v4i16, v2i32})
      .customIf([=](const LegalityQuery &Q) {
        // TODO: Fix suboptimal codegen for 128+ bit types.
        LLT SrcTy = Q.Types[0];
        return SrcTy.isScalar() && SrcTy.getSizeInBits() < 128;
      })
      .widenScalarIf(
          [=](const LegalityQuery &Query) { return Query.Types[0] == v4s8; },
          [=](const LegalityQuery &Query) { return std::make_pair(0, v4i16); })
      .widenScalarIf(
          [=](const LegalityQuery &Query) { return Query.Types[0] == v2s16; },
          [=](const LegalityQuery &Query) { return std::make_pair(0, v2i32); })
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0)
      .lower();
 
  getActionDefinitionsBuilder(
      {G_ABDS, G_ABDU, G_UAVGFLOOR, G_UAVGCEIL, G_SAVGFLOOR, G_SAVGCEIL})
      .legalFor({v8i8, v16i8, v4i16, v8i16, v2i32, v4i32})
      .lower();
 
  getActionDefinitionsBuilder(
      {G_SADDE, G_SSUBE, G_UADDE, G_USUBE, G_SADDO, G_SSUBO, G_UADDO, G_USUBO})
      .legalFor({{i32, i32}, {i64, i32}})
      .clampScalar(0, s32, s64)
      .clampScalar(1, s32, s64)
      .widenScalarToNextPow2(0);
 
  getActionDefinitionsBuilder({G_FSHL, G_FSHR})
      .customFor({{i32, i32}, {i32, i64}, {i64, i64}})
      .lower();
 
  getActionDefinitionsBuilder(G_ROTR)
      .legalFor({{i32, i64}, {i64, i64}})
      .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}});
 
  auto always = [=](const LegalityQuery &Q) { return true; };
  getActionDefinitionsBuilder(G_CTPOP)
      .legalFor(HasCSSC, {{i32, i32}, {i64, i64}})
      .legalFor({{v8i8, v8i8}, {v16i8, v16i8}})
      .customFor(!HasCSSC, {{s32, s32}, {s64, s64}})
      .customFor({{s128, s128},
                  {v4s16, v4s16},
                  {v8s16, v8s16},
                  {v2s32, v2s32},
                  {v4s32, v4s32},
                  {v2s64, v2s64}})
      .clampScalar(0, s32, s128)
      .widenScalarToNextPow2(0)
      .widenScalarOrEltToNextPow2OrMinSize(0, 8)
      .minScalarEltSameAsIf(always, 1, 0)
      .maxScalarEltSameAsIf(always, 1, 0)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0);
 
  getActionDefinitionsBuilder({G_CTLZ, G_CTLS})
      .legalFor({{i32, i32},
                 {i64, i64},
                 {v8i8, v8i8},
                 {v16i8, v16i8},
                 {v4i16, v4i16},
                 {v8i16, v8i16},
                 {v2i32, v2i32},
                 {v4i32, v4i32}})
      .widenScalarToNextPow2(1, /*Min=*/32)
      .clampScalar(1, s32, s64)
      .widenScalarOrEltToNextPow2OrMinSize(1, /*Min=*/8)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .moreElementsToNextPow2(0)
      .scalarizeIf(scalarOrEltWiderThan(0, 32), 0)
      .scalarSameSizeAs(0, 1);
 
  getActionDefinitionsBuilder(G_INSERT_SUBVECTOR).lower();
 
  getActionDefinitionsBuilder(G_CTLZ_ZERO_POISON).lower();
 
  getActionDefinitionsBuilder(G_CTTZ)
      .lowerIf(isVector(0))
      .widenScalarToNextPow2(1, /*Min=*/32)
      .clampScalar(1, s32, s64)
      .scalarSameSizeAs(0, 1)
      .legalFor(HasCSSC, {s32, s64})
      .customFor(!HasCSSC, {s32, s64});
 
  getActionDefinitionsBuilder(G_CTTZ_ZERO_POISON).lower();
 
  getActionDefinitionsBuilder(G_BITREVERSE)
      .legalFor({i32, i64, v8i8, v16i8})
      .widenScalarToNextPow2(0, /*Min = */ 32)
      .widenScalarOrEltToNextPow2OrMinSize(0, 8)
      .clampScalar(0, s32, s64)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .moreElementsToNextPow2(0)
      .lower();
 
  getActionDefinitionsBuilder(G_CLMUL).legalFor({v8i8, v16i8});
 
  getActionDefinitionsBuilder(G_BSWAP)
      .legalFor({i32, i64, v4i16, v8i16, v2i32, v4i32, v2i64})
      .widenScalarOrEltToNextPow2(0, 16)
      .clampScalar(0, s32, s64)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0);
 
  getActionDefinitionsBuilder({G_UADDSAT, G_SADDSAT, G_USUBSAT, G_SSUBSAT})
      .legalFor({v8i8, v16i8, v4i16, v8i16, v2i32, v4i32, v2i64})
      .legalFor(HasSVE, {nxv16i8, nxv8i16, nxv4i32, nxv2i64})
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .moreElementsToNextPow2(0)
      .lower();
 
  getActionDefinitionsBuilder(
      {G_FADD, G_FSUB, G_FMUL, G_FDIV, G_FMA, G_FSQRT, G_FMAXNUM, G_FMINNUM,
       G_FMAXIMUM, G_FMINIMUM, G_FCEIL, G_FFLOOR, G_FRINT, G_FNEARBYINT,
       G_INTRINSIC_TRUNC, G_INTRINSIC_ROUND, G_INTRINSIC_ROUNDEVEN})
      .legalFor({f32, f64, v2f32, v4f32, v2f64})
      .legalFor(HasFP16, {f16, v4f16, v8f16})
      .libcallFor({f128})
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .widenScalarIf(
          [=](const LegalityQuery &Q) {
            return (!HasFP16 && Q.Types[0].getScalarType().isFloat16()) ||
                   Q.Types[0].getScalarType().isBFloat16();
          },
          changeElementTo(0, f32))
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0);
 
  getActionDefinitionsBuilder({G_FABS, G_FNEG})
      .legalFor({f32, f64, v2f32, v4f32, v2f64})
      .legalFor(HasFP16, {f16, bf16, v4f16, v4bf16, v8f16, v8bf16})
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .lowerIf(scalarOrEltWiderThan(0, 64))
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .moreElementsToNextPow2(0)
      .lowerFor({f16, bf16, v4f16, v4bf16, v8f16, v8bf16});
 
  getActionDefinitionsBuilder({G_FREM, G_FCOS, G_FSIN, G_FPOW, G_FLOG, G_FLOG2,
                               G_FLOG10, G_FTAN, G_FEXP, G_FEXP2, G_FEXP10,
                               G_FACOS, G_FASIN, G_FATAN, G_FATAN2, G_FCOSH,
                               G_FSINH, G_FTANH, G_FMODF})
      .libcallFor({f32, f64, f128})
      .widenScalarFor({f16, bf16}, changeElementTo(0, f32))
      .scalarize(0);
  getActionDefinitionsBuilder({G_FPOWI, G_FLDEXP})
      .libcallFor({{f32, i32}, {f64, i32}, {f128, i32}})
      .widenScalarFor({f16, bf16}, changeElementTo(0, f32))
      .scalarize(0);
 
  getActionDefinitionsBuilder({G_LROUND, G_INTRINSIC_LRINT})
      .legalFor({{i32, f32}, {i32, f64}, {i64, f32}, {i64, f64}})
      .legalFor(HasFP16, {{i32, f16}, {i64, f16}})
      .minScalar(1, s32)
      .libcallFor({{s64, s128}})
      .lower();
  getActionDefinitionsBuilder({G_LLROUND, G_INTRINSIC_LLRINT})
      .legalFor({{i64, f32}, {i64, f64}})
      .legalFor(HasFP16, {{i64, f16}})
      .minScalar(0, s64)
      .minScalar(1, s32)
      .libcallFor({{s64, s128}})
      .lower();
 
  // TODO: Custom legalization for mismatched types.
  getActionDefinitionsBuilder(G_FCOPYSIGN)
      .moreElementsIf(
          [](const LegalityQuery &Query) { return Query.Types[0].isScalar(); },
          [=](const LegalityQuery &Query) {
            const LLT Ty = Query.Types[0];
            return std::pair(0, LLT::fixed_vector(Ty == s16 ? 4 : 2, Ty));
          })
      .lower();
 
  getActionDefinitionsBuilder(G_FMAD).lower();
 
  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];
    return ValTy.isPointerVector() && ValTy.getAddressSpace() == 0;
  };
 
  getActionDefinitionsBuilder(G_LOAD)
      .customIf([=](const LegalityQuery &Query) {
        return HasRCPC3 && Query.Types[0] == s128 &&
               Query.MMODescrs[0].Ordering == AtomicOrdering::Acquire;
      })
      .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}, {s64, p0, s32, 8}})
      .legalForTypesWithMemDesc({
          // SVE vscale x 128 bit base sizes
          {nxv16s8, p0, nxv16s8, 8},
          {nxv8s16, p0, nxv8s16, 8},
          {nxv4s32, p0, nxv4s32, 8},
          {nxv2s64, p0, nxv2s64, 8},
      })
      .widenScalarToNextPow2(0, /* MinSize = */ 8)
      .clampMaxNumElements(0, s8, 16)
      .clampMaxNumElements(0, s16, 8)
      .clampMaxNumElements(0, s32, 4)
      .clampMaxNumElements(0, s64, 2)
      .clampMaxNumElements(0, p0, 2)
      .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))
      // TODO: Use BITCAST for v2i8, v2i16 after G_TRUNC gets sorted out
      .bitcastIf(typeInSet(0, {v4s8}),
                 [=](const LegalityQuery &Query) {
                   const LLT VecTy = Query.Types[0];
                   return std::pair(0, LLT::integer(VecTy.getSizeInBits()));
                 })
      .customIf(IsPtrVecPred)
      .scalarizeIf(typeInSet(0, {v2s16, v2s8}), 0)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0);
 
  getActionDefinitionsBuilder(G_STORE)
      .customIf([=](const LegalityQuery &Query) {
        return HasRCPC3 && Query.Types[0] == s128 &&
               Query.MMODescrs[0].Ordering == AtomicOrdering::Release;
      })
      .customIf([=](const LegalityQuery &Query) {
        return Query.Types[0] == s128 &&
               Query.MMODescrs[0].Ordering != AtomicOrdering::NotAtomic;
      })
      .widenScalarIf(
          all(scalarNarrowerThan(0, 32),
              atomicOrderingAtLeastOrStrongerThan(0, AtomicOrdering::Release)),
          changeElementSizeTo(0, s32))
      .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}})
      .legalForTypesWithMemDesc({
          // SVE vscale x 128 bit base sizes
          // TODO: Add nxv2p0. Consider bitcastIf.
          //       See #92130
          // https://github.com/llvm/llvm-project/pull/92130#discussion_r1616888461
          {nxv16s8, p0, nxv16s8, 8},
          {nxv8s16, p0, nxv8s16, 8},
          {nxv4s32, p0, nxv4s32, 8},
          {nxv2s64, p0, nxv2s64, 8},
      })
      .clampScalar(0, s8, s64)
      .minScalarOrElt(0, s8)
      .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()
      // TODO: Use BITCAST for v2i8, v2i16 after G_TRUNC gets sorted out
      .bitcastIf(all(typeInSet(0, {v4s8}),
                     LegalityPredicate([=](const LegalityQuery &Query) {
                       return Query.Types[0].getSizeInBits() ==
                              Query.MMODescrs[0].MemoryTy.getSizeInBits();
                     })),
                 [=](const LegalityQuery &Query) {
                   const LLT VecTy = Query.Types[0];
                   return std::pair(0, LLT::integer(VecTy.getSizeInBits()));
                 })
      .customIf(IsPtrVecPred)
      .scalarizeIf(typeInSet(0, {v2s16, v2s8}), 0)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .lower();
 
  getActionDefinitionsBuilder(G_INDEXED_STORE)
      // Idx 0 == Ptr, Idx 1 == Val
      // TODO: we can implement legalizations but as of now these are
      // generated in a very specific way.
      .legalForTypesWithMemDesc({
          {p0, s8, s8, 8},
          {p0, s16, s16, 8},
          {p0, s32, s8, 8},
          {p0, s32, s16, 8},
          {p0, s32, s32, 8},
          {p0, s64, s64, 8},
          {p0, p0, p0, 8},
          {p0, v8s8, v8s8, 8},
          {p0, v16s8, v16s8, 8},
          {p0, v4s16, v4s16, 8},
          {p0, v8s16, v8s16, 8},
          {p0, v2s32, v2s32, 8},
          {p0, v4s32, v4s32, 8},
          {p0, v2s64, v2s64, 8},
          {p0, v2p0, v2p0, 8},
          {p0, s128, s128, 8},
      })
      .unsupported();
 
  auto IndexedLoadBasicPred = [=](const LegalityQuery &Query) {
    LLT LdTy = Query.Types[0];
    LLT PtrTy = Query.Types[1];
    if (!llvm::is_contained(PackedVectorAllTypesVec, LdTy) &&
        !llvm::is_contained(ScalarAndPtrTypesVec, LdTy) && LdTy != s128)
      return false;
    if (PtrTy != p0)
      return false;
    return true;
  };
  getActionDefinitionsBuilder(G_INDEXED_LOAD)
      .unsupportedIf(
          atomicOrderingAtLeastOrStrongerThan(0, AtomicOrdering::Unordered))
      .legalIf(IndexedLoadBasicPred)
      .unsupported();
  getActionDefinitionsBuilder({G_INDEXED_SEXTLOAD, G_INDEXED_ZEXTLOAD})
      .unsupportedIf(
          atomicOrderingAtLeastOrStrongerThan(0, AtomicOrdering::Unordered))
      .legalIf(all(typeInSet(0, {s16, s32, s64}),
                   LegalityPredicate([=](const LegalityQuery &Q) {
                     LLT LdTy = Q.Types[0];
                     LLT PtrTy = Q.Types[1];
                     LLT MemTy = Q.MMODescrs[0].MemoryTy;
                     if (PtrTy != p0)
                       return false;
                     if (LdTy == s16)
                       return MemTy == s8;
                     if (LdTy == s32)
                       return MemTy == s8 || MemTy == s16;
                     if (LdTy == s64)
                       return MemTy == s8 || MemTy == s16 || MemTy == s32;
                     return false;
                   })))
      .unsupported();
 
  // Constants
  getActionDefinitionsBuilder(G_CONSTANT)
      .legalFor({p0, s8, s16, s32, s64})
      .widenScalarToNextPow2(0)
      .clampScalar(0, s8, s64);
  getActionDefinitionsBuilder(G_FCONSTANT)
      .legalFor({s16, s32, s64, s128});
 
  // FIXME: fix moreElementsToNextPow2
  getActionDefinitionsBuilder(G_ICMP)
      .legalFor({{i32, i32}, {i32, i64}, {i32, p0}})
      .widenScalarOrEltToNextPow2(1)
      .minScalarOrElt(1, s8)
      .clampScalar(1, s32, s64)
      .clampScalar(0, s32, s32)
      .scalarizeIf(scalarOrEltWiderThan(1, 64), 1)
      .minScalarEltSameAsIf(
          [=](const LegalityQuery &Query) {
            const LLT &Ty = Query.Types[0];
            const LLT &SrcTy = Query.Types[1];
            return Ty.isVector() && !SrcTy.isPointerVector() &&
                   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].isPointerVector();
          },
          0, s64)
      .moreElementsToNextPow2(1)
      .clampNumElements(1, v8s8, v16s8)
      .clampNumElements(1, v4s16, v8s16)
      .clampNumElements(1, v2s32, v4s32)
      .clampNumElements(1, v2s64, v2s64)
      .clampNumElements(1, v2p0, v2p0)
      .customIf(isVector(0));
 
  getActionDefinitionsBuilder(G_FCMP)
      .legalFor({{i32, f32},
                 {i32, f64},
                 {v4i32, v4f32},
                 {v2i32, v2f32},
                 {v2i64, v2f64}})
      .legalFor(HasFP16, {{i32, f16}, {v4i16, v4f16}, {v8i16, v8f16}})
      .widenScalarOrEltToNextPow2(1)
      .clampScalar(0, s32, s32)
      .widenScalarIf(
          [=](const LegalityQuery &Q) {
            return (!HasFP16 && Q.Types[1].getScalarType().isFloat16()) ||
                   Q.Types[1].getScalarType().isBFloat16();
          },
          changeElementTo(1, f32))
      .scalarizeIf(scalarOrEltWiderThan(1, 64), 1)
      .minScalarEltSameAsIf(
          [=](const LegalityQuery &Query) {
            const LLT &Ty = Query.Types[0];
            const LLT &SrcTy = Query.Types[1];
            return Ty.isVector() && !SrcTy.isPointerVector() &&
                   Ty.getElementType() != SrcTy.getElementType();
          },
          0, 1)
      .clampNumElements(1, v4s16, v8s16)
      .clampNumElements(1, v2s32, v4s32)
      .clampMaxNumElements(1, s64, 2)
      .moreElementsToNextPow2(1)
      .libcallFor({{s32, s128}});
 
  // Extensions
  auto ExtLegalFunc = [=](const LegalityQuery &Query) {
    unsigned DstSize = Query.Types[0].getSizeInBits();
 
    // Handle legal vectors using legalFor
    if (Query.Types[0].isVector())
      return false;
 
    if (DstSize < 8 || DstSize >= 128 || !isPowerOf2_32(DstSize))
      return false; // Extending to a scalar s128 needs narrowing.
 
    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)
      .legalFor({{v8s16, v8s8}, {v4s32, v4s16}, {v2s64, v2s32}})
      .clampScalar(0, s64, s64) // Just for s128, others are handled above.
      .moreElementsToNextPow2(0)
      .clampMaxNumElements(1, s8, 8)
      .clampMaxNumElements(1, s16, 4)
      .clampMaxNumElements(1, s32, 2)
      // Tries to convert a large EXTEND into two smaller EXTENDs
      .lowerIf([=](const LegalityQuery &Query) {
        return (Query.Types[0].getScalarSizeInBits() >
                Query.Types[1].getScalarSizeInBits() * 2) &&
               Query.Types[0].isVector() &&
               (Query.Types[1].getScalarSizeInBits() == 8 ||
                Query.Types[1].getScalarSizeInBits() == 16);
      })
      .clampMinNumElements(1, s8, 8)
      .clampMinNumElements(1, s16, 4)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0);
 
  getActionDefinitionsBuilder(G_TRUNC)
      .legalFor({{v8s8, v8s16}, {v4s16, v4s32}, {v2s32, v2s64}})
      .moreElementsToNextPow2(0)
      .clampMaxNumElements(0, s8, 8)
      .clampMaxNumElements(0, s16, 4)
      .clampMaxNumElements(0, s32, 2)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) { return Query.Types[0].isVector(); },
          0, s8)
      .lowerIf([=](const LegalityQuery &Query) {
        LLT DstTy = Query.Types[0];
        LLT SrcTy = Query.Types[1];
        return DstTy.isVector() && SrcTy.getSizeInBits() > 128 &&
               DstTy.getScalarSizeInBits() * 2 <= SrcTy.getScalarSizeInBits();
      })
      .clampMinNumElements(0, s8, 8)
      .clampMinNumElements(0, s16, 4)
      .alwaysLegal();
 
  getActionDefinitionsBuilder({G_TRUNC_SSAT_S, G_TRUNC_SSAT_U, G_TRUNC_USAT_U})
      .legalFor({{v8i8, v8i16}, {v4i16, v4i32}, {v2i32, v2i64}})
      .clampNumElements(0, v8s8, v8s8)
      .clampNumElements(0, v4s16, v4s16)
      .clampNumElements(0, v2s32, v2s32)
      .lower();
 
  getActionDefinitionsBuilder(G_SEXT_INREG)
      .legalFor({i32, i64, v8i8, v16i8, v4i16, v8i16, v2i32, v4i32, v2i64})
      .maxScalar(0, s64)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .lower();
 
  // FP conversions
  getActionDefinitionsBuilder(G_FPTRUNC)
      .legalFor(
          {{f16, f32}, {f16, f64}, {f32, f64}, {v4f16, v4f32}, {v2f32, v2f64}})
      .legalFor(ST.hasBF16(), {{bf16, f32}, {v4bf16, v4f32}})
      .libcallFor({{f16, f128}, {f32, f128}, {f64, f128}})
      .moreElementsToNextPow2(1)
      .customIf([](const LegalityQuery &Q) {
        LLT DstTy = Q.Types[0];
        LLT SrcTy = Q.Types[1];
        return SrcTy.getScalarSizeInBits() == 64 &&
               DstTy.getScalarSizeInBits() == 16;
      })
      .lowerFor({{bf16, f32}, {v4bf16, v4f32}})
      // Clamp based on input
      .clampNumElements(1, v4s32, v4s32)
      .clampNumElements(1, v2s64, v2s64)
      .scalarize(0);
 
  getActionDefinitionsBuilder(G_FPEXT)
      .legalFor({{f32, f16},
                 {f64, f16},
                 {f32, bf16},
                 {f64, f32},
                 {v4f32, v4f16},
                 {v4f32, v4bf16},
                 {v2f64, v2f32}})
      .libcallFor({{f128, f64}, {f128, f32}, {f128, f16}})
      .moreElementsToNextPow2(0)
      .widenScalarIf(
          [](const LegalityQuery &Q) {
            LLT DstTy = Q.Types[0];
            LLT SrcTy = Q.Types[1];
            return SrcTy.isVector() && DstTy.isVector() &&
                   SrcTy.getScalarSizeInBits() == 16 &&
                   DstTy.getScalarSizeInBits() == 64;
          },
          changeElementTo(1, f32))
      .clampNumElements(0, v4s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .scalarize(0);
 
  // Conversions
  getActionDefinitionsBuilder({G_FPTOSI, G_FPTOUI})
      .legalFor({{i32, f32},
                 {i64, f32},
                 {i32, f64},
                 {i64, f64},
                 {v2i32, v2f32},
                 {v4i32, v4f32},
                 {v2i64, v2f64}})
      .legalFor(HasFP16,
                {{i32, f16}, {i64, f16}, {v4i16, v4f16}, {v8i16, v8f16}})
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .scalarizeIf(scalarOrEltWiderThan(1, 64), 1)
      // The range of a fp16 value fits into an i17, so we can lower the width
      // to i64.
      .narrowScalarIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1] == f16 && Query.Types[0].getSizeInBits() > 64;
          },
          changeTo(0, i64))
      .moreElementsToNextPow2(0)
      .widenScalarOrEltToNextPow2OrMinSize(0)
      .minScalar(0, s32)
      .widenScalarIf(
          [HasFP16](const LegalityQuery &Query) {
            return (!HasFP16 && Query.Types[1].getScalarType().isFloat16()) ||
                   Query.Types[1].getScalarType().isBFloat16();
          },
          changeElementTo(1, f32))
      .widenScalarIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].getScalarSizeInBits() <= 64 &&
                   Query.Types[0].getScalarSizeInBits() >
                       Query.Types[1].getScalarSizeInBits();
          },
          LegalizeMutations::changeElementSizeTo(1, 0))
      .widenScalarIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].getScalarSizeInBits() <= 64 &&
                   Query.Types[0].getScalarSizeInBits() <
                       Query.Types[1].getScalarSizeInBits();
          },
          LegalizeMutations::changeElementSizeTo(0, 1))
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .libcallFor(
          {{i32, f128}, {i64, f128}, {i128, f128}, {i128, f32}, {i128, f64}});
 
  getActionDefinitionsBuilder({G_FPTOSI_SAT, G_FPTOUI_SAT})
      .legalFor({{i32, f32},
                 {i64, f32},
                 {i32, f64},
                 {i64, f64},
                 {v2i32, v2f32},
                 {v4i32, v4f32},
                 {v2i64, v2f64}})
      .legalFor(
          HasFP16,
          {{i16, f16}, {i32, f16}, {i64, f16}, {v4i16, v4f16}, {v8i16, v8f16}})
      // Handle types larger than i64 by scalarizing/lowering.
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .scalarizeIf(scalarOrEltWiderThan(1, 64), 1)
      // The range of a fp16 value fits into an i17, so we can lower the width
      // to i64.
      .narrowScalarIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1] == f16 && Query.Types[0].getSizeInBits() > 64;
          },
          changeTo(0, i64))
      .lowerIf(::any(scalarWiderThan(0, 64), scalarWiderThan(1, 64)), 0)
      .moreElementsToNextPow2(0)
      .widenScalarToNextPow2(0, /*MinSize=*/32)
      .minScalar(0, s32)
      .widenScalarIf(
          [HasFP16](const LegalityQuery &Query) {
            return (!HasFP16 && Query.Types[1].getScalarType().isFloat16()) ||
                   Query.Types[1].getScalarType().isBFloat16();
          },
          changeElementTo(1, f32))
      .widenScalarIf(
          [=](const LegalityQuery &Query) {
            unsigned ITySize = Query.Types[0].getScalarSizeInBits();
            return (ITySize == 16 || ITySize == 32 || ITySize == 64) &&
                   ITySize > Query.Types[1].getScalarSizeInBits();
          },
          LegalizeMutations::changeElementSizeTo(1, 0))
      .widenScalarIf(
          [=](const LegalityQuery &Query) {
            unsigned FTySize = Query.Types[1].getScalarSizeInBits();
            return (FTySize == 16 || FTySize == 32 || FTySize == 64) &&
                   Query.Types[0].getScalarSizeInBits() < FTySize;
          },
          LegalizeMutations::changeElementSizeTo(0, 1))
      .widenScalarOrEltToNextPow2(0)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2);
 
  getActionDefinitionsBuilder({G_SITOFP, G_UITOFP})
      .legalFor({{f32, i32},
                 {f64, i32},
                 {f32, i64},
                 {f64, i64},
                 {v2f32, v2i32},
                 {v4f32, v4i32},
                 {v2f64, v2i64}})
      .legalFor(HasFP16,
                {{f16, i32}, {f16, i64}, {v4f16, v4i16}, {v8f16, v8i16}})
      .unsupportedIf([&](const LegalityQuery &Query) {
        return Query.Types[0].getScalarType().isBFloat16();
      })
      .scalarizeIf(scalarOrEltWiderThan(1, 64), 1)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .moreElementsToNextPow2(1)
      .widenScalarOrEltToNextPow2OrMinSize(1)
      .minScalar(1, f32)
      .lowerIf([](const LegalityQuery &Query) {
        return Query.Types[1].isVector() &&
               Query.Types[1].getScalarSizeInBits() == 64 &&
               Query.Types[0].getScalarSizeInBits() == 16;
      })
      .widenScalarOrEltToNextPow2OrMinSize(0, /*MinSize=*/HasFP16 ? 16 : 32)
      .scalarizeIf(
          // v2i64->v2f32 needs to scalarize to avoid double-rounding issues.
          [](const LegalityQuery &Query) {
            return Query.Types[0].getScalarSizeInBits() == 32 &&
                   Query.Types[1].getScalarSizeInBits() == 64;
          },
          0)
      .widenScalarIf(
          [](const LegalityQuery &Query) {
            return Query.Types[1].getScalarSizeInBits() <= 64 &&
                   Query.Types[0].getScalarSizeInBits() <
                       Query.Types[1].getScalarSizeInBits();
          },
          LegalizeMutations::changeElementSizeTo(0, 1))
      .widenScalarIf(
          [](const LegalityQuery &Query) {
            return Query.Types[0].getScalarSizeInBits() <= 64 &&
                   Query.Types[0].getScalarSizeInBits() >
                       Query.Types[1].getScalarSizeInBits();
          },
          LegalizeMutations::changeElementSizeTo(1, 0))
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .libcallFor({{f16, i128},
                   {f32, i128},
                   {f64, i128},
                   {f128, i128},
                   {f128, i32},
                   {f128, i64}});
 
  // Control-flow
  getActionDefinitionsBuilder(G_BR).alwaysLegal();
  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)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .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_PTRAUTH_GLOBAL_VALUE)
      .legalIf(all(typeIs(0, p0), typeIs(1, p0)));
 
  getActionDefinitionsBuilder(G_PTRTOINT)
      .legalFor({{i64, p0}, {v2i64, v2p0}})
      .widenScalarToNextPow2(0, 64)
      .clampScalar(0, s64, s64)
      .clampMaxNumElements(0, s64, 2);
 
  getActionDefinitionsBuilder(G_INTTOPTR)
      .unsupportedIf([&](const LegalityQuery &Query) {
        return Query.Types[0].getSizeInBits() != Query.Types[1].getSizeInBits();
      })
      .legalFor({{p0, i64}, {v2p0, v2i64}})
      .clampMaxNumElements(1, s64, 2);
 
  // 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)
      .legalForCartesianProduct({s16})
      // Keeping 32-bit instructions legal to prevent regression in some tests
      .legalForCartesianProduct({s32, v2s16, v4s8})
      .legalForCartesianProduct({s64, v8s8, v4s16, v2s32})
      .legalForCartesianProduct({s128, v16s8, v8s16, v4s32, v2s64, v2p0})
      .customIf([=](const LegalityQuery &Query) {
        // Handle casts from i1 vectors to scalars.
        LLT DstTy = Query.Types[0];
        LLT SrcTy = Query.Types[1];
        return DstTy.isScalar() && SrcTy.isVector() &&
               SrcTy.getScalarSizeInBits() == 1;
      })
      .lowerIf([=](const LegalityQuery &Query) {
        return Query.Types[0].isVector() != Query.Types[1].isVector();
      })
      .moreElementsToNextPow2(0)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .lower();
 
  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)));
 
  bool UseOutlineAtomics = ST.outlineAtomics() && !ST.hasLSE();
 
  getActionDefinitionsBuilder(G_ATOMIC_CMPXCHG)
      .legalFor(!UseOutlineAtomics, {{s32, p0}, {s64, p0}})
      .customFor(!UseOutlineAtomics, {{s128, p0}})
      .libcallFor(UseOutlineAtomics,
                  {{s8, p0}, {s16, p0}, {s32, p0}, {s64, p0}, {s128, p0}})
      .clampScalar(0, s32, s64);
 
  getActionDefinitionsBuilder({G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD,
                               G_ATOMICRMW_SUB, G_ATOMICRMW_AND, G_ATOMICRMW_OR,
                               G_ATOMICRMW_XOR})
      .legalFor(!UseOutlineAtomics, {{s32, p0}, {s64, p0}})
      .libcallFor(UseOutlineAtomics,
                  {{s8, p0}, {s16, p0}, {s32, p0}, {s64, p0}})
      .clampScalar(0, s32, s64);
 
  // Do not outline these atomics operations, as per comment in
  // AArch64ISelLowering.cpp's shouldExpandAtomicRMWInIR().
  getActionDefinitionsBuilder(
      {G_ATOMICRMW_MIN, G_ATOMICRMW_MAX, G_ATOMICRMW_UMIN, G_ATOMICRMW_UMAX})
      .legalIf(all(typeInSet(0, {s32, s64}), typeIs(1, p0)))
      .clampScalar(0, s32, s64);
 
  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;
          }
        });
  }
 
  // TODO : nxv4s16, nxv2s16, nxv2s32
  getActionDefinitionsBuilder(G_EXTRACT_VECTOR_ELT)
      .legalFor(HasSVE, {{s16, nxv16s8, s64},
                         {s16, nxv8s16, s64},
                         {s32, nxv4s32, s64},
                         {s64, nxv2s64, s64}})
      .unsupportedIf([=](const LegalityQuery &Query) {
        const LLT &EltTy = Query.Types[1].getElementType();
        if (Query.Types[1].isScalableVector())
          return false;
        return Query.Types[0] != EltTy;
      })
      .minScalar(2, s64)
      .customIf([=](const LegalityQuery &Query) {
        const LLT &VecTy = Query.Types[1];
        return VecTy == v8s8 || VecTy == v16s8 || VecTy == v2s16 ||
               VecTy == v4s16 || VecTy == v8s16 || VecTy == v2s32 ||
               VecTy == v4s32 || VecTy == v2s64 || 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].isFixedVector() &&
                   Query.Types[1].getNumElements() <= 2;
          },
          0, s64)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].isFixedVector() &&
                   Query.Types[1].getNumElements() <= 4;
          },
          0, s32)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].isFixedVector() &&
                   Query.Types[1].getNumElements() <= 8;
          },
          0, s16)
      .minScalarOrEltIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].isFixedVector() &&
                   Query.Types[1].getNumElements() <= 16;
          },
          0, s8)
      .minScalarOrElt(0, s8) // Worst case, we need at least s8.
      .moreElementsToNextPow2(1)
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .clampMaxNumElements(1, s16, 8)
      .clampMaxNumElements(1, s8, 16)
      .clampMaxNumElements(1, p0, 2)
      .scalarizeIf(scalarOrEltWiderThan(1, 64), 1);
 
  getActionDefinitionsBuilder(G_INSERT_VECTOR_ELT)
      .legalIf(
          typeInSet(0, {v8s8, v16s8, v4s16, v8s16, v2s32, v4s32, v2s64, v2p0}))
      .legalFor(HasSVE, {{nxv16s8, s32, s64},
                         {nxv8s16, s32, s64},
                         {nxv4s32, s32, s64},
                         {nxv2s64, s64, s64}})
      .moreElementsToNextPow2(0)
      .widenVectorEltsToVectorMinSize(0, 64)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v2s32, v4s32)
      .clampMaxNumElements(0, s64, 2)
      .clampMaxNumElements(0, p0, 2)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0);
 
  getActionDefinitionsBuilder(G_BUILD_VECTOR)
      .legalFor({{v8s8, s8},
                 {v16s8, s8},
                 {v4s16, s16},
                 {v8s16, s16},
                 {v2s32, s32},
                 {v4s32, s32},
                 {v2s64, s64},
                 {v2p0, p0}})
      .clampNumElements(0, v4s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .minScalarOrElt(0, s8)
      .widenVectorEltsToVectorMinSize(0, 64)
      .widenScalarOrEltToNextPow2(0)
      .minScalarSameAs(1, 0);
 
  getActionDefinitionsBuilder(G_BUILD_VECTOR_TRUNC).lower();
 
  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;
        return llvm::is_contained(
            {v8s8, v16s8, v4s16, v8s16, v2s32, v4s32, v2s64}, DstTy);
      })
      .moreElementsIf(
          [](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() >
                   Query.Types[1].getNumElements();
          },
          changeTo(1, 0))
      .moreElementsToNextPow2(0)
      .moreElementsIf(
          [](const LegalityQuery &Query) {
            return Query.Types[0].getNumElements() <
                   Query.Types[1].getNumElements();
          },
          changeTo(0, 1))
      .widenScalarOrEltToNextPow2OrMinSize(0, 8)
      .clampNumElements(0, v8s8, v16s8)
      .clampNumElements(0, v4s16, v8s16)
      .clampNumElements(0, v4s32, v4s32)
      .clampNumElements(0, v2s64, v2s64)
      .scalarizeIf(scalarOrEltWiderThan(0, 64), 0)
      .bitcastIf(isPointerVector(0), [=](const LegalityQuery &Query) {
        // Bitcast pointers vector to i64.
        const LLT DstTy = Query.Types[0];
        return std::pair(
            0, LLT::vector(DstTy.getElementCount(), LLT::integer(64)));
      });
 
  getActionDefinitionsBuilder(G_CONCAT_VECTORS)
      .legalFor({{v16s8, v8s8}, {v8s16, v4s16}, {v4s32, v2s32}})
      .customIf([=](const LegalityQuery &Query) {
        return Query.Types[0].isFixedVector() &&
               Query.Types[0].getScalarSizeInBits() < 8;
      })
      .bitcastIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[0].isFixedVector() &&
                   Query.Types[1].isFixedVector() &&
                   Query.Types[0].getScalarSizeInBits() >= 8 &&
                   isPowerOf2_64(Query.Types[0].getScalarSizeInBits()) &&
                   Query.Types[0].getSizeInBits() <= 128 &&
                   Query.Types[1].getSizeInBits() <= 64;
          },
          [=](const LegalityQuery &Query) {
            const LLT DstTy = Query.Types[0];
            const LLT SrcTy = Query.Types[1];
            return std::pair(
                0, DstTy.changeElementSize(SrcTy.getSizeInBits())
                       .changeElementCount(
                           DstTy.getElementCount().divideCoefficientBy(
                               SrcTy.getNumElements())));
          });
 
  getActionDefinitionsBuilder(G_EXTRACT_SUBVECTOR)
      .legalFor({{v8s8, v16s8}, {v4s16, v8s16}, {v2s32, v4s32}})
      .widenScalarOrEltToNextPow2(0)
      .clampMaxNumElements(0, s8, 16)
      .clampMaxNumElements(0, s16, 8)
      .clampMaxNumElements(0, s32, 4)
      .clampNumElements(1, v8s8, v16s8)
      .clampNumElements(1, v4s16, v8s16)
      .clampNumElements(1, v2s32, v4s32)
      .lower()
      .immIdx(0); // Inform verifier imm idx 0 is handled.
 
  // TODO: {nxv16s8, s8}, {nxv8s16, s16}
  getActionDefinitionsBuilder(G_SPLAT_VECTOR)
      .legalFor(HasSVE, {{nxv4s32, s32}, {nxv2s64, s64}});
 
  getActionDefinitionsBuilder(G_JUMP_TABLE).legalFor({p0});
 
  getActionDefinitionsBuilder(G_BRJT).legalFor({{p0, s64}});
 
  getActionDefinitionsBuilder({G_TRAP, G_DEBUGTRAP, G_UBSANTRAP}).alwaysLegal();
 
  getActionDefinitionsBuilder(G_DYN_STACKALLOC).custom();
 
  getActionDefinitionsBuilder({G_STACKSAVE, G_STACKRESTORE}).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});
 
    getActionDefinitionsBuilder(G_MEMSET_INLINE)
        .legalForCartesianProduct({p0}, {s64}, {s64})
        .customForCartesianProduct({p0}, {s8}, {s64});
  } else {
    getActionDefinitionsBuilder({G_BZERO, G_MEMCPY, G_MEMMOVE, G_MEMSET})
        .libcall();
  }
 
  // For fadd reductions we have pairwise operations available. We treat the
  // usual legal types as legal and handle the lowering to pairwise instructions
  // later.
  getActionDefinitionsBuilder(G_VECREDUCE_FADD)
      .legalFor({{f32, v2f32}, {f32, v4f32}, {f64, v2f64}})
      .legalFor(HasFP16, {{f16, v4f16}, {f16, v8f16}})
      .widenScalarIf(
          [HasFP16](const LegalityQuery &Query) {
            return (!HasFP16 && Query.Types[0].getScalarType().isFloat16()) ||
                   Query.Types[0].getScalarType().isBFloat16();
          },
          changeElementTo(0, f32))
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .clampMaxNumElements(1, s16, 8)
      .moreElementsToNextPow2(1)
      .scalarize(1)
      .lower();
 
  // For fmul reductions we need to split up into individual operations. We
  // clamp to 128 bit vectors then to 64bit vectors to produce a cascade of
  // smaller types, followed by scalarizing what remains.
  getActionDefinitionsBuilder(G_VECREDUCE_FMUL)
      .widenScalarIf(
          [HasFP16](const LegalityQuery &Query) {
            return (!HasFP16 && Query.Types[0].getScalarType().isFloat16()) ||
                   Query.Types[0].getScalarType().isBFloat16();
          },
          changeElementTo(0, f32))
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .clampMaxNumElements(1, s16, 8)
      .clampMaxNumElements(1, s32, 2)
      .clampMaxNumElements(1, s16, 4)
      .scalarize(1)
      .lower();
 
  getActionDefinitionsBuilder({G_VECREDUCE_SEQ_FADD, G_VECREDUCE_SEQ_FMUL})
      .scalarize(2)
      .lower();
 
  getActionDefinitionsBuilder(G_VECREDUCE_ADD)
      .legalFor({{i8, v8i8},
                 {i8, v16i8},
                 {i16, v4i16},
                 {i16, v8i16},
                 {i32, v2i32},
                 {i32, v4i32},
                 {i64, v2i64}})
      .moreElementsToNextPow2(1)
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .clampMaxNumElements(1, s16, 8)
      .clampMaxNumElements(1, s8, 16)
      .widenVectorEltsToVectorMinSize(1, 64)
      .scalarize(1);
 
  getActionDefinitionsBuilder({G_VECREDUCE_FMIN, G_VECREDUCE_FMAX,
                               G_VECREDUCE_FMINIMUM, G_VECREDUCE_FMAXIMUM})
      .legalFor({{f32, v2f32}, {f32, v4f32}, {f64, v2f64}})
      .legalFor(HasFP16, {{f16, v4f16}, {f16, v8f16}})
      .widenScalarIf(
          [HasFP16](const LegalityQuery &Query) {
            return (!HasFP16 && Query.Types[0].getScalarType().isFloat16()) ||
                   Query.Types[0].getScalarType().isBFloat16();
          },
          changeElementTo(0, f32))
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .clampMaxNumElements(1, s16, 8)
      .scalarize(1)
      .lower();
 
  getActionDefinitionsBuilder(G_VECREDUCE_MUL)
      .clampMaxNumElements(1, s32, 2)
      .clampMaxNumElements(1, s16, 4)
      .clampMaxNumElements(1, s8, 8)
      .scalarize(1)
      .lower();
 
  getActionDefinitionsBuilder(
      {G_VECREDUCE_SMIN, G_VECREDUCE_SMAX, G_VECREDUCE_UMIN, G_VECREDUCE_UMAX})
      .legalFor({{i8, v8i8},
                 {i8, v16i8},
                 {i16, v4i16},
                 {i16, v8i16},
                 {i32, v2i32},
                 {i32, v4i32}})
      .moreElementsIf(
          [=](const LegalityQuery &Query) {
            return Query.Types[1].isVector() &&
                   Query.Types[1].getElementType() != s8 &&
                   Query.Types[1].getNumElements() & 1;
          },
          LegalizeMutations::moreElementsToNextPow2(1))
      .clampMaxNumElements(1, s64, 2)
      .clampMaxNumElements(1, s32, 4)
      .clampMaxNumElements(1, s16, 8)
      .clampMaxNumElements(1, s8, 16)
      .scalarize(1)
      .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();
 
  // TODO: Update this to correct handling when adding AArch64/SVE support.
  getActionDefinitionsBuilder(G_VECTOR_COMPRESS).lower();
 
  // Access to floating-point environment.
  getActionDefinitionsBuilder({G_GET_FPENV, G_SET_FPENV, G_RESET_FPENV,
                               G_GET_FPMODE, G_SET_FPMODE, G_RESET_FPMODE})
      .libcall();
 
  getActionDefinitionsBuilder({G_GET_ROUNDING, G_SET_ROUNDING})
      .customFor({s32});
 
  getActionDefinitionsBuilder(G_IS_FPCLASS).lower();
 
  getActionDefinitionsBuilder(G_PREFETCH).custom();
 
  getActionDefinitionsBuilder({G_SCMP, G_UCMP}).lower();
 
  getActionDefinitionsBuilder({G_INTRINSIC, G_INTRINSIC_W_SIDE_EFFECTS})
      .alwaysLegal();
  getActionDefinitionsBuilder(G_FENCE).alwaysLegal();
  getActionDefinitionsBuilder(G_INVOKE_REGION_START).alwaysLegal();
 
  verify(*ST.getInstrInfo());
}
 
bool AArch64LegalizerInfo::legalizeCustom(
    LegalizerHelper &Helper, MachineInstr &MI,
    LostDebugLocObserver &LocObserver) 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_SBFX:
  case TargetOpcode::G_UBFX:
    return legalizeBitfieldExtract(MI, MRI, Helper);
  case TargetOpcode::G_FSHL:
  case TargetOpcode::G_FSHR:
    return legalizeFunnelShift(MI, MRI, MIRBuilder, Observer, 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:
  case TargetOpcode::G_MEMSET_INLINE:
    return legalizeMemOps(MI, Helper);
  case TargetOpcode::G_EXTRACT_VECTOR_ELT:
    return legalizeExtractVectorElt(MI, MRI, Helper);
  case TargetOpcode::G_DYN_STACKALLOC:
    return legalizeDynStackAlloc(MI, Helper);
  case TargetOpcode::G_PREFETCH:
    return legalizePrefetch(MI, Helper);
  case TargetOpcode::G_ABS:
    return Helper.lowerAbsToCNeg(MI);
  case TargetOpcode::G_ICMP:
    return legalizeICMP(MI, MRI, MIRBuilder);
  case TargetOpcode::G_BITCAST:
    return legalizeBitcast(MI, Helper);
  case TargetOpcode::G_CONCAT_VECTORS:
    return legalizeConcatVectors(MI, MRI, MIRBuilder);
  case TargetOpcode::G_FPTRUNC:
    // In order to lower f16 to f64 properly, we need to use f32 as an
    // intermediary
    return legalizeFptrunc(MI, MIRBuilder, MRI);
  case TargetOpcode::G_GET_ROUNDING:
    return legalizeGetRounding(MI, MIRBuilder, MRI, Helper);
  case TargetOpcode::G_SET_ROUNDING:
    return legalizeSetRounding(MI, MIRBuilder, MRI, Helper);
  }
 
  llvm_unreachable("expected switch to return");
}
 
bool AArch64LegalizerInfo::legalizeBitcast(MachineInstr &MI,
                                           LegalizerHelper &Helper) const {
  assert(MI.getOpcode() == TargetOpcode::G_BITCAST && "Unexpected opcode");
  auto [DstReg, DstTy, SrcReg, SrcTy] = MI.getFirst2RegLLTs();
  // We're trying to handle casts from i1 vectors to scalars but reloading from
  // stack.
  if (!DstTy.isScalar() || !SrcTy.isVector() ||
      SrcTy.getElementType() != LLT::scalar(1))
    return false;
 
  Helper.createStackStoreLoad(DstReg, SrcReg);
  MI.eraseFromParent();
  return true;
}
 
bool AArch64LegalizerInfo::legalizeFunnelShift(MachineInstr &MI,
                                               MachineRegisterInfo &MRI,
                                               MachineIRBuilder &MIRBuilder,
                                               GISelChangeObserver &Observer,
                                               LegalizerHelper &Helper) const {
  assert(MI.getOpcode() == TargetOpcode::G_FSHL ||
         MI.getOpcode() == TargetOpcode::G_FSHR);
 
  // Keep as G_FSHR if shift amount is a G_CONSTANT, else use generic
  // lowering
  Register ShiftNo = MI.getOperand(3).getReg();
  LLT ShiftTy = MRI.getType(ShiftNo);
  auto VRegAndVal = getIConstantVRegValWithLookThrough(ShiftNo, MRI);
 
  // Adjust shift amount according to Opcode (FSHL/FSHR)
  // Convert FSHL to FSHR
  LLT OperationTy = MRI.getType(MI.getOperand(0).getReg());
  APInt BitWidth(ShiftTy.getSizeInBits(), OperationTy.getSizeInBits(), false);
 
  // Lower non-constant shifts and leave zero shifts to the optimizer.
  if (!VRegAndVal || VRegAndVal->Value.urem(BitWidth) == 0)
    return (Helper.lowerFunnelShiftAsShifts(MI) ==
            LegalizerHelper::LegalizeResult::Legalized);
 
  APInt Amount = VRegAndVal->Value.urem(BitWidth);
 
  Amount = MI.getOpcode() == TargetOpcode::G_FSHL ? BitWidth - Amount : Amount;
 
  // If the instruction is G_FSHR, has a 64-bit G_CONSTANT for shift amount
  // in the range of 0 <-> BitWidth, it is legal
  if (ShiftTy.getSizeInBits() == 64 && MI.getOpcode() == TargetOpcode::G_FSHR &&
      VRegAndVal->Value.ult(BitWidth))
    return true;
 
  // Cast the ShiftNumber to a 64-bit type
  auto Cast64 = MIRBuilder.buildConstant(LLT::integer(64), Amount.zext(64));
 
  if (MI.getOpcode() == TargetOpcode::G_FSHR) {
    Observer.changingInstr(MI);
    MI.getOperand(3).setReg(Cast64.getReg(0));
    Observer.changedInstr(MI);
  }
  // If Opcode is FSHL, remove the FSHL instruction and create a FSHR
  // instruction
  else if (MI.getOpcode() == TargetOpcode::G_FSHL) {
    MIRBuilder.buildInstr(TargetOpcode::G_FSHR, {MI.getOperand(0).getReg()},
                          {MI.getOperand(1).getReg(), MI.getOperand(2).getReg(),
                           Cast64.getReg(0)});
    MI.eraseFromParent();
  }
  return true;
}
 
bool AArch64LegalizerInfo::legalizeICMP(MachineInstr &MI,
                                        MachineRegisterInfo &MRI,
                                        MachineIRBuilder &MIRBuilder) const {
  Register DstReg = MI.getOperand(0).getReg();
  Register SrcReg1 = MI.getOperand(2).getReg();
  Register SrcReg2 = MI.getOperand(3).getReg();
  LLT DstTy = MRI.getType(DstReg);
  LLT SrcTy = MRI.getType(SrcReg1);
 
  // Check the vector types are legal
  if (DstTy.getScalarSizeInBits() != SrcTy.getScalarSizeInBits() ||
      DstTy.getNumElements() != SrcTy.getNumElements() ||
      (DstTy.getSizeInBits() != 64 && DstTy.getSizeInBits() != 128))
    return false;
 
  // Lowers G_ICMP NE => G_ICMP EQ to allow better pattern matching for
  // following passes
  CmpInst::Predicate Pred = (CmpInst::Predicate)MI.getOperand(1).getPredicate();
  if (Pred != CmpInst::ICMP_NE)
    return true;
  Register CmpReg =
      MIRBuilder
          .buildICmp(CmpInst::ICMP_EQ, MRI.getType(DstReg), SrcReg1, SrcReg2)
          .getReg(0);
  MIRBuilder.buildNot(DstReg, CmpReg);
 
  MI.eraseFromParent();
  return true;
}
 
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.buildZExt(LLT::integer(64), AmtReg);
  Helper.Observer.changingInstr(MI);
  MI.getOperand(2).setReg(NewAmt.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);
  // Don't modify an intrinsic call.
  if (GlobalOp.isSymbol())
    return true;
  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
  // 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 {
  MachineIRBuilder &MIB = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *MIB.getMRI();
 
  auto LowerUnaryOp = [&MI, &MIB](unsigned Opcode) {
    MIB.buildInstr(Opcode, {MI.getOperand(0)}, {MI.getOperand(2)});
    MI.eraseFromParent();
    return true;
  };
  auto LowerBinOp = [&MI, &MIB](unsigned Opcode) {
    MIB.buildInstr(Opcode, {MI.getOperand(0)},
                   {MI.getOperand(2), MI.getOperand(3)});
    MI.eraseFromParent();
    return true;
  };
  auto LowerTriOp = [&MI, &MIB](unsigned Opcode) {
    MIB.buildInstr(Opcode, {MI.getOperand(0)},
                   {MI.getOperand(2), MI.getOperand(3), MI.getOperand(4)});
    MI.eraseFromParent();
    return true;
  };
 
  Intrinsic::ID IntrinsicID = cast<GIntrinsic>(MI).getIntrinsicID();
  switch (IntrinsicID) {
  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::integer(VaListSize * 8));
    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: {
    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);
    // Anyext the value being set to 64 bit (only the bottom 8 bits are read by
    // the instruction).
    auto &Value = MI.getOperand(3);
    Register ExtValueReg = MIB.buildAnyExt(LLT::integer(64), Value).getReg(0);
    Value.setReg(ExtValueReg);
    return true;
  }
  case Intrinsic::aarch64_prefetch: {
    auto &AddrVal = MI.getOperand(1);
 
    int64_t IsWrite = MI.getOperand(2).getImm();
    int64_t Target = MI.getOperand(3).getImm();
    int64_t IsStream = MI.getOperand(4).getImm();
    int64_t IsData = MI.getOperand(5).getImm();
 
    unsigned PrfOp = (IsWrite << 4) |    // Load/Store bit
                     (!IsData << 3) |    // IsDataCache bit
                     (Target << 1) |     // Cache level bits
                     (unsigned)IsStream; // Stream bit
 
    MIB.buildInstr(AArch64::G_AARCH64_PREFETCH).addImm(PrfOp).add(AddrVal);
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_range_prefetch: {
    auto &AddrVal = MI.getOperand(1);
 
    int64_t IsWrite = MI.getOperand(2).getImm();
    int64_t IsStream = MI.getOperand(3).getImm();
    unsigned PrfOp = (IsStream << 2) | IsWrite;
 
    MIB.buildInstr(AArch64::G_AARCH64_RANGE_PREFETCH)
        .addImm(PrfOp)
        .add(AddrVal)
        .addUse(MI.getOperand(4).getReg()); // Metadata
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_prefetch_ir: {
    auto &AddrVal = MI.getOperand(1);
    MIB.buildInstr(AArch64::G_AARCH64_PREFETCH).addImm(24).add(AddrVal);
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_uaddv:
  case Intrinsic::aarch64_neon_saddv:
  case Intrinsic::aarch64_neon_umaxv:
  case Intrinsic::aarch64_neon_smaxv:
  case Intrinsic::aarch64_neon_uminv:
  case Intrinsic::aarch64_neon_sminv: {
    bool IsSigned = IntrinsicID == Intrinsic::aarch64_neon_saddv ||
                    IntrinsicID == Intrinsic::aarch64_neon_smaxv ||
                    IntrinsicID == Intrinsic::aarch64_neon_sminv;
 
    auto OldDst = MI.getOperand(0).getReg();
    auto OldDstTy = MRI.getType(OldDst);
    LLT NewDstTy = MRI.getType(MI.getOperand(2).getReg()).getElementType();
    if (OldDstTy == NewDstTy)
      return true;
 
    auto NewDst = MRI.createGenericVirtualRegister(NewDstTy);
 
    Helper.Observer.changingInstr(MI);
    MI.getOperand(0).setReg(NewDst);
    Helper.Observer.changedInstr(MI);
 
    MIB.setInsertPt(MIB.getMBB(), ++MIB.getInsertPt());
    MIB.buildExtOrTrunc(IsSigned ? TargetOpcode::G_SEXT : TargetOpcode::G_ZEXT,
                        OldDst, NewDst);
 
    return true;
  }
  case Intrinsic::aarch64_neon_uaddlp:
  case Intrinsic::aarch64_neon_saddlp: {
    unsigned Opc = IntrinsicID == Intrinsic::aarch64_neon_uaddlp
                       ? AArch64::G_UADDLP
                       : AArch64::G_SADDLP;
    MIB.buildInstr(Opc, {MI.getOperand(0)}, {MI.getOperand(2)});
    MI.eraseFromParent();
 
    return true;
  }
  case Intrinsic::aarch64_neon_uaddlv:
  case Intrinsic::aarch64_neon_saddlv: {
    unsigned Opc = IntrinsicID == Intrinsic::aarch64_neon_uaddlv
                       ? AArch64::G_UADDLV
                       : AArch64::G_SADDLV;
    Register DstReg = MI.getOperand(0).getReg();
    Register SrcReg = MI.getOperand(2).getReg();
    LLT DstTy = MRI.getType(DstReg);
 
    LLT MidTy, ExtTy;
    if (DstTy.isScalar() && DstTy.getScalarSizeInBits() <= 32) {
      ExtTy = LLT::integer(32);
      MidTy = LLT::fixed_vector(4, ExtTy);
    } else {
      ExtTy = LLT::integer(64);
      MidTy = LLT::fixed_vector(2, ExtTy);
    }
 
    Register MidReg =
        MIB.buildInstr(Opc, {MidTy}, {SrcReg})->getOperand(0).getReg();
    Register ZeroReg =
        MIB.buildConstant(LLT::integer(64), 0)->getOperand(0).getReg();
    Register ExtReg = MIB.buildInstr(AArch64::G_EXTRACT_VECTOR_ELT, {ExtTy},
                                     {MidReg, ZeroReg})
                          .getReg(0);
 
    if (DstTy.getScalarSizeInBits() < 32)
      MIB.buildTrunc(DstReg, ExtReg);
    else
      MIB.buildCopy(DstReg, ExtReg);
 
    MI.eraseFromParent();
 
    return true;
  }
  case Intrinsic::aarch64_neon_smax:
    return LowerBinOp(TargetOpcode::G_SMAX);
  case Intrinsic::aarch64_neon_smin:
    return LowerBinOp(TargetOpcode::G_SMIN);
  case Intrinsic::aarch64_neon_umax:
    return LowerBinOp(TargetOpcode::G_UMAX);
  case Intrinsic::aarch64_neon_umin:
    return LowerBinOp(TargetOpcode::G_UMIN);
  case Intrinsic::aarch64_neon_fmax:
    return LowerBinOp(TargetOpcode::G_FMAXIMUM);
  case Intrinsic::aarch64_neon_fmin:
    return LowerBinOp(TargetOpcode::G_FMINIMUM);
  case Intrinsic::aarch64_neon_fmaxnm:
    return LowerBinOp(TargetOpcode::G_FMAXNUM);
  case Intrinsic::aarch64_neon_fminnm:
    return LowerBinOp(TargetOpcode::G_FMINNUM);
  case Intrinsic::aarch64_neon_pmul:
    return LowerBinOp(TargetOpcode::G_CLMUL);
  case Intrinsic::aarch64_neon_pmull:
  case Intrinsic::aarch64_neon_pmull64:
    return LowerBinOp(AArch64::G_PMULL);
  case Intrinsic::aarch64_neon_smull:
    return LowerBinOp(AArch64::G_SMULL);
  case Intrinsic::aarch64_neon_umull:
    return LowerBinOp(AArch64::G_UMULL);
  case Intrinsic::aarch64_neon_sabd:
    return LowerBinOp(TargetOpcode::G_ABDS);
  case Intrinsic::aarch64_neon_uabd:
    return LowerBinOp(TargetOpcode::G_ABDU);
  case Intrinsic::aarch64_neon_uhadd:
    return LowerBinOp(TargetOpcode::G_UAVGFLOOR);
  case Intrinsic::aarch64_neon_urhadd:
    return LowerBinOp(TargetOpcode::G_UAVGCEIL);
  case Intrinsic::aarch64_neon_shadd:
    return LowerBinOp(TargetOpcode::G_SAVGFLOOR);
  case Intrinsic::aarch64_neon_srhadd:
    return LowerBinOp(TargetOpcode::G_SAVGCEIL);
  case Intrinsic::aarch64_neon_sqshrn: {
    if (!MRI.getType(MI.getOperand(0).getReg()).isVector())
      return true;
    // Create right shift instruction. Store the output register in Shr.
    auto Shr = MIB.buildInstr(AArch64::G_VASHR,
                              {MRI.getType(MI.getOperand(2).getReg())},
                              {MI.getOperand(2), MI.getOperand(3).getImm()});
    // Build the narrow intrinsic, taking in Shr.
    MIB.buildInstr(TargetOpcode::G_TRUNC_SSAT_S, {MI.getOperand(0)}, {Shr});
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_sqshrun: {
    if (!MRI.getType(MI.getOperand(0).getReg()).isVector())
      return true;
    // Create right shift instruction. Store the output register in Shr.
    auto Shr = MIB.buildInstr(AArch64::G_VASHR,
                              {MRI.getType(MI.getOperand(2).getReg())},
                              {MI.getOperand(2), MI.getOperand(3).getImm()});
    // Build the narrow intrinsic, taking in Shr.
    MIB.buildInstr(TargetOpcode::G_TRUNC_SSAT_U, {MI.getOperand(0)}, {Shr});
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_sqrshrn: {
    if (!MRI.getType(MI.getOperand(0).getReg()).isVector())
      return true;
    // Create right shift instruction. Store the output register in Shr.
    auto Shr = MIB.buildInstr(AArch64::G_SRSHR_I,
                              {MRI.getType(MI.getOperand(2).getReg())},
                              {MI.getOperand(2), MI.getOperand(3).getImm()});
    // Build the narrow intrinsic, taking in Shr.
    MIB.buildInstr(TargetOpcode::G_TRUNC_SSAT_S, {MI.getOperand(0)}, {Shr});
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_sqrshrun: {
    if (!MRI.getType(MI.getOperand(0).getReg()).isVector())
      return true;
    // Create right shift instruction. Store the output register in Shr.
    auto Shr = MIB.buildInstr(AArch64::G_SRSHR_I,
                              {MRI.getType(MI.getOperand(2).getReg())},
                              {MI.getOperand(2), MI.getOperand(3).getImm()});
    // Build the narrow intrinsic, taking in Shr.
    MIB.buildInstr(TargetOpcode::G_TRUNC_SSAT_U, {MI.getOperand(0)}, {Shr});
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_uqrshrn: {
    if (!MRI.getType(MI.getOperand(0).getReg()).isVector())
      return true;
    // Create right shift instruction. Store the output register in Shr.
    auto Shr = MIB.buildInstr(AArch64::G_URSHR_I,
                              {MRI.getType(MI.getOperand(2).getReg())},
                              {MI.getOperand(2), MI.getOperand(3).getImm()});
    // Build the narrow intrinsic, taking in Shr.
    MIB.buildInstr(TargetOpcode::G_TRUNC_USAT_U, {MI.getOperand(0)}, {Shr});
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_uqshrn: {
    if (!MRI.getType(MI.getOperand(0).getReg()).isVector())
      return true;
    // Create right shift instruction. Store the output register in Shr.
    auto Shr = MIB.buildInstr(AArch64::G_VLSHR,
                              {MRI.getType(MI.getOperand(2).getReg())},
                              {MI.getOperand(2), MI.getOperand(3).getImm()});
    // Build the narrow intrinsic, taking in Shr.
    MIB.buildInstr(TargetOpcode::G_TRUNC_USAT_U, {MI.getOperand(0)}, {Shr});
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_sqshlu: {
    // Check if last operand is constant vector dup
    auto ShiftAmount =
        isConstantOrConstantSplatVector(MI.getOperand(3).getReg(), MRI);
    if (ShiftAmount) {
      // If so, create a new intrinsic with the correct shift amount
      MIB.buildInstr(AArch64::G_SQSHLU_I, {MI.getOperand(0)},
                     {MI.getOperand(2)})
          .addImm(ShiftAmount->getSExtValue());
      MI.eraseFromParent();
      return true;
    }
    return false;
  }
  case Intrinsic::aarch64_neon_vsli: {
    MIB.buildInstr(
        AArch64::G_SLI, {MI.getOperand(0)},
        {MI.getOperand(2), MI.getOperand(3), MI.getOperand(4).getImm()});
    MI.eraseFromParent();
    break;
  }
  case Intrinsic::aarch64_neon_vsri: {
    MIB.buildInstr(
        AArch64::G_SRI, {MI.getOperand(0)},
        {MI.getOperand(2), MI.getOperand(3), MI.getOperand(4).getImm()});
    MI.eraseFromParent();
    break;
  }
  case Intrinsic::aarch64_neon_abs: {
    // Lower the intrinsic to G_ABS.
    MIB.buildInstr(TargetOpcode::G_ABS, {MI.getOperand(0)}, {MI.getOperand(2)});
    MI.eraseFromParent();
    return true;
  }
  case Intrinsic::aarch64_neon_addhn:
    return LowerBinOp(AArch64::G_ADDHN);
  case Intrinsic::aarch64_neon_sqadd: {
    if (MRI.getType(MI.getOperand(0).getReg()).isVector())
      return LowerBinOp(TargetOpcode::G_SADDSAT);
    break;
  }
  case Intrinsic::aarch64_neon_sqsub: {
    if (MRI.getType(MI.getOperand(0).getReg()).isVector())
      return LowerBinOp(TargetOpcode::G_SSUBSAT);
    break;
  }
  case Intrinsic::aarch64_neon_uqadd: {
    if (MRI.getType(MI.getOperand(0).getReg()).isVector())
      return LowerBinOp(TargetOpcode::G_UADDSAT);
    break;
  }
  case Intrinsic::aarch64_neon_uqsub: {
    if (MRI.getType(MI.getOperand(0).getReg()).isVector())
      return LowerBinOp(TargetOpcode::G_USUBSAT);
    break;
  }
  case Intrinsic::aarch64_neon_udot:
    return LowerTriOp(AArch64::G_UDOT);
  case Intrinsic::aarch64_neon_sdot:
    return LowerTriOp(AArch64::G_SDOT);
  case Intrinsic::aarch64_neon_usdot:
    return LowerTriOp(AArch64::G_USDOT);
  case Intrinsic::aarch64_neon_sqxtn:
    return LowerUnaryOp(TargetOpcode::G_TRUNC_SSAT_S);
  case Intrinsic::aarch64_neon_sqxtun:
    return LowerUnaryOp(TargetOpcode::G_TRUNC_SSAT_U);
  case Intrinsic::aarch64_neon_uqxtn:
    return LowerUnaryOp(TargetOpcode::G_TRUNC_USAT_U);
  case Intrinsic::aarch64_neon_fcvtzu:
    return LowerUnaryOp(TargetOpcode::G_FPTOUI_SAT);
  case Intrinsic::aarch64_neon_fcvtzs:
    return LowerUnaryOp(TargetOpcode::G_FPTOSI_SAT);
  case Intrinsic::aarch64_neon_cls:
    return LowerUnaryOp(TargetOpcode::G_CTLS);
 
  case Intrinsic::vector_reverse:
    // TODO: Add support for vector_reverse
    return false;
  }
 
  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();
  LLT AmtRegEltTy = MRI.getType(AmtReg).getScalarType();
  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(AmtRegEltTy.changeElementSize(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)) {
 
    AtomicOrdering Ordering = (*MI.memoperands_begin())->getSuccessOrdering();
    bool IsLoad = MI.getOpcode() == TargetOpcode::G_LOAD;
    bool IsLoadAcquire = IsLoad && Ordering == AtomicOrdering::Acquire;
    bool IsStoreRelease = !IsLoad && Ordering == AtomicOrdering::Release;
    bool IsRcpC3 =
        ST->hasLSE2() && ST->hasRCPC3() && (IsLoadAcquire || IsStoreRelease);
 
    LLT s64 = LLT::integer(64);
 
    unsigned Opcode;
    if (IsRcpC3) {
      Opcode = IsLoad ? AArch64::LDIAPPX : AArch64::STILPX;
    } else {
      // For LSE2, loads/stores should have been converted to monotonic and had
      // a fence inserted after them.
      assert(Ordering == AtomicOrdering::Monotonic ||
             Ordering == AtomicOrdering::Unordered);
      assert(ST->hasLSE2() && "ldp/stp not single copy atomic without +lse2");
 
      Opcode = IsLoad ? AArch64::LDPXi : AArch64::STPXi;
    }
 
    MachineInstrBuilder NewI;
    if (IsLoad) {
      NewI = MIRBuilder.buildInstr(Opcode, {s64, s64}, {});
      MIRBuilder.buildMergeLikeInstr(
          ValReg, {NewI->getOperand(0), NewI->getOperand(1)});
    } else {
      auto Split = MIRBuilder.buildUnmerge(s64, MI.getOperand(0));
      NewI = MIRBuilder.buildInstr(
          Opcode, {}, {Split->getOperand(0), Split->getOperand(1)});
    }
 
    if (IsRcpC3) {
      NewI.addUse(MI.getOperand(1).getReg());
    } else {
      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.isPointerVector() ||
      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(), LLT::integer(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::integer(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 {
  // When there is no integer popcount instruction (FEAT_CSSC isn't available),
  // 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
 
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
  Register Dst = MI.getOperand(0).getReg();
  Register Val = MI.getOperand(1).getReg();
  LLT Ty = MRI.getType(Val);
 
  LLT i64 = LLT::integer(64);
  LLT i32 = LLT::integer(32);
  LLT i16 = LLT::integer(16);
  LLT i8 = LLT::integer(8);
  unsigned Size = Ty.getSizeInBits();
 
  assert(Ty == MRI.getType(Dst) &&
         "Expected src and dst to have the same type!");
 
  if (ST->hasCSSC() && Ty.isScalar() && Size == 128) {
 
    auto Split = MIRBuilder.buildUnmerge(i64, Val);
    auto CTPOP1 = MIRBuilder.buildCTPOP(i64, Split->getOperand(0));
    auto CTPOP2 = MIRBuilder.buildCTPOP(i64, Split->getOperand(1));
    auto Add = MIRBuilder.buildAdd(i64, CTPOP1, CTPOP2);
 
    MIRBuilder.buildZExt(Dst, Add);
    MI.eraseFromParent();
    return true;
  }
 
  if (!ST->hasNEON() ||
      MI.getMF()->getFunction().hasFnAttribute(Attribute::NoImplicitFloat)) {
    // Use generic lowering when custom lowering is not possible.
    return Ty.isScalar() && (Size == 32 || Size == 64) &&
           Helper.lowerBitCount(MI) ==
               LegalizerHelper::LegalizeResult::Legalized;
  }
 
  // 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, i8) : LLT::fixed_vector(8, i8);
  if (Ty.isScalar()) {
    assert((Size == 32 || Size == 64 || Size == 128) && "Expected only 32, 64, or 128 bit scalars!");
    if (Size == 32) {
      Val = MIRBuilder.buildZExt(i64, 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.
  if (ST->hasDotProd() && Ty.isVector() && Ty.getNumElements() >= 2 &&
      Ty.getScalarSizeInBits() != 16) {
    LLT Dt = Ty == LLT::fixed_vector(2, i64) ? LLT::fixed_vector(4, i32) : Ty;
    auto Zeros = MIRBuilder.buildConstant(Dt, 0);
    auto Ones = MIRBuilder.buildConstant(VTy, 1);
    MachineInstrBuilder Sum;
 
    if (Ty == LLT::fixed_vector(2, i64)) {
      auto UDOT =
          MIRBuilder.buildInstr(AArch64::G_UDOT, {Dt}, {Zeros, Ones, CTPOP});
      Sum = MIRBuilder.buildInstr(AArch64::G_UADDLP, {Ty}, {UDOT});
    } else if (Ty == LLT::fixed_vector(4, i32)) {
      Sum = MIRBuilder.buildInstr(AArch64::G_UDOT, {Dt}, {Zeros, Ones, CTPOP});
    } else if (Ty == LLT::fixed_vector(2, i32)) {
      Sum = MIRBuilder.buildInstr(AArch64::G_UDOT, {Dt}, {Zeros, Ones, CTPOP});
    } else {
      llvm_unreachable("unexpected vector shape");
    }
 
    Sum->getOperand(0).setReg(Dst);
    MI.eraseFromParent();
    return true;
  }
 
  Register HSum = CTPOP.getReg(0);
  unsigned Opc;
  SmallVector<LLT> HAddTys;
  if (Ty.isScalar()) {
    Opc = Intrinsic::aarch64_neon_uaddlv;
    HAddTys.push_back(i32);
  } else if (Ty == LLT::fixed_vector(8, i16)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(8, i16));
  } else if (Ty == LLT::fixed_vector(4, i32)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(8, i16));
    HAddTys.push_back(LLT::fixed_vector(4, i32));
  } else if (Ty == LLT::fixed_vector(2, i64)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(8, i16));
    HAddTys.push_back(LLT::fixed_vector(4, i32));
    HAddTys.push_back(LLT::fixed_vector(2, i64));
  } else if (Ty == LLT::fixed_vector(4, i16)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(4, i16));
  } else if (Ty == LLT::fixed_vector(2, i32)) {
    Opc = Intrinsic::aarch64_neon_uaddlp;
    HAddTys.push_back(LLT::fixed_vector(4, i16));
    HAddTys.push_back(LLT::fixed_vector(2, i32));
  } else
    llvm_unreachable("unexpected vector shape");
  MachineInstrBuilder UADD;
  for (LLT HTy : HAddTys) {
    UADD = MIRBuilder.buildIntrinsic(Opc, {HTy}).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 i64 = LLT::integer(64);
  auto Addr = MI.getOperand(1).getReg();
  auto DesiredI = MIRBuilder.buildUnmerge({i64, i64}, MI.getOperand(2));
  auto NewI = MIRBuilder.buildUnmerge({i64, i64}, MI.getOperand(3));
  auto DstLo = MRI.createGenericVirtualRegister(i64);
  auto DstHi = MRI.createGenericVirtualRegister(i64);
 
  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::integer(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.buildMergeLikeInstr(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 ||
      MI.getOpcode() == TargetOpcode::G_MEMSET_INLINE) {
    // Anyext the value being set to 64 bit (only the bottom 8 bits are read by
    // the instruction).
    auto &Value = MI.getOperand(1);
    Register ExtValueReg =
        MIRBuilder.buildAnyExt(LLT::integer(64), Value).getReg(0);
    Value.setReg(ExtValueReg);
    return true;
  }
 
  return false;
}
 
bool AArch64LegalizerInfo::legalizeExtractVectorElt(
    MachineInstr &MI, MachineRegisterInfo &MRI, LegalizerHelper &Helper) const {
  const GExtractVectorElement *Element = cast<GExtractVectorElement>(&MI);
  auto VRegAndVal =
      getIConstantVRegValWithLookThrough(Element->getIndexReg(), MRI);
  if (VRegAndVal)
    return true;
  LLT VecTy = MRI.getType(Element->getVectorReg());
  if (VecTy.isScalableVector())
    return true;
  return Helper.lowerExtractInsertVectorElt(MI) !=
         LegalizerHelper::LegalizeResult::UnableToLegalize;
}
 
bool AArch64LegalizerInfo::legalizeDynStackAlloc(
    MachineInstr &MI, LegalizerHelper &Helper) const {
  MachineFunction &MF = *MI.getParent()->getParent();
  MachineIRBuilder &MIRBuilder = Helper.MIRBuilder;
  MachineRegisterInfo &MRI = *MIRBuilder.getMRI();
 
  // If stack probing is not enabled for this function, use the default
  // lowering.
  if (!MF.getFunction().hasFnAttribute("probe-stack") ||
      MF.getFunction().getFnAttribute("probe-stack").getValueAsString() !=
          "inline-asm") {
    Helper.lowerDynStackAlloc(MI);
    return true;
  }
 
  Register Dst = MI.getOperand(0).getReg();
  Register AllocSize = MI.getOperand(1).getReg();
  Align Alignment = assumeAligned(MI.getOperand(2).getImm());
 
  assert(MRI.getType(Dst) == LLT::pointer(0, 64) &&
         "Unexpected type for dynamic alloca");
  assert(MRI.getType(AllocSize) == LLT::scalar(64) &&
         "Unexpected type for dynamic alloca");
 
  LLT PtrTy = MRI.getType(Dst);
  Register SPReg =
      Helper.getTargetLowering().getStackPointerRegisterToSaveRestore();
  Register SPTmp =
      Helper.getDynStackAllocTargetPtr(SPReg, AllocSize, Alignment, PtrTy);
  auto NewMI =
      MIRBuilder.buildInstr(AArch64::PROBED_STACKALLOC_DYN, {}, {SPTmp});
  MRI.setRegClass(NewMI.getReg(0), &AArch64::GPR64commonRegClass);
  MIRBuilder.setInsertPt(*NewMI->getParent(), NewMI);
  MIRBuilder.buildCopy(Dst, SPTmp);
 
  MI.eraseFromParent();
  return true;
}
 
bool AArch64LegalizerInfo::legalizePrefetch(MachineInstr &MI,
                                            LegalizerHelper &Helper) const {
  MachineIRBuilder &MIB = Helper.MIRBuilder;
  auto &AddrVal = MI.getOperand(0);
 
  int64_t IsWrite = MI.getOperand(1).getImm();
  int64_t Locality = MI.getOperand(2).getImm();
  int64_t IsData = MI.getOperand(3).getImm();
 
  bool IsStream = Locality == 0;
  if (Locality != 0) {
    assert(Locality <= 3 && "Prefetch locality out-of-range");
    // The locality degree is the opposite of the cache speed.
    // Put the number the other way around.
    // The encoding starts at 0 for level 1
    Locality = 3 - Locality;
  }
 
  unsigned PrfOp = (IsWrite << 4) | (!IsData << 3) | (Locality << 1) | IsStream;
 
  MIB.buildInstr(AArch64::G_AARCH64_PREFETCH).addImm(PrfOp).add(AddrVal);
  MI.eraseFromParent();
  return true;
}
 
bool AArch64LegalizerInfo::legalizeConcatVectors(
    MachineInstr &MI, MachineRegisterInfo &MRI,
    MachineIRBuilder &MIRBuilder) const {
  // Widen sub-byte element vectors to byte-sized elements before concatenating.
  // This is analogous to SDAG's integer type promotion for sub-byte types.
  auto &Concat = cast<GConcatVectors>(MI);
  Register DstReg = Concat.getReg(0);
  LLT DstTy = MRI.getType(DstReg);
  assert(DstTy.getScalarSizeInBits() < 8 && "Expected dst ty to be < 8b");
 
  unsigned WideEltSize =
      std::max(8u, (unsigned)PowerOf2Ceil(DstTy.getScalarSizeInBits()));
  LLT SrcTy = MRI.getType(Concat.getSourceReg(0));
  LLT WideSrcTy = SrcTy.changeElementSize(WideEltSize);
  LLT WideDstTy = DstTy.changeElementSize(WideEltSize);
 
  SmallVector<Register> WideSrcs;
  for (unsigned I = 0; I < Concat.getNumSources(); ++I) {
    auto Wide = MIRBuilder.buildAnyExt(WideSrcTy, Concat.getSourceReg(I));
    WideSrcs.push_back(Wide.getReg(0));
  }
 
  auto WideConcat = MIRBuilder.buildConcatVectors(WideDstTy, WideSrcs);
  MIRBuilder.buildTrunc(DstReg, WideConcat);
  MI.eraseFromParent();
  return true;
}
 
bool AArch64LegalizerInfo::legalizeFptrunc(MachineInstr &MI,
                                           MachineIRBuilder &MIRBuilder,
                                           MachineRegisterInfo &MRI) const {
  auto [Dst, DstTy, Src, SrcTy] = MI.getFirst2RegLLTs();
 
  // This function legalizes f64 -> bf16 and f64 -> f16 truncations via f64 ->
  // f32 G_FPTRUNC_ODD and f32 -> [b]f16 G_FPTRUNC, which apparently avoids the
  // usual double-rounding issue that could be present from using twin
  // G_FPTRUNC.
 
  if (DstTy.isBFloat16() && SrcTy.isFloat64()) {
    auto Mid = MIRBuilder.buildInstr(AArch64::G_FPTRUNC_ODD, {LLT::float32()},
                                     {Src}, MI.getFlags());
    MIRBuilder.buildInstr(AArch64::G_FPTRUNC, {Dst}, {Mid}, MI.getFlags());
    MI.eraseFromParent();
    return true;
  }
 
  assert(SrcTy.isFixedVector() && isPowerOf2_32(SrcTy.getNumElements()) &&
         "Expected a power of 2 elements");
 
  // We must mutate types here as FPTrunc may be used on a IEEE floating point
  // or a brainfloat.
  LLT v2s16 = DstTy.changeElementCount(2);
  LLT v4s16 = DstTy.changeElementCount(4);
  LLT v2s32 = SrcTy.changeElementCount(2).changeElementSize(32);
  LLT v4s32 = SrcTy.changeElementCount(4).changeElementSize(32);
  LLT v2s64 = SrcTy.changeElementCount(2);
 
  SmallVector<Register> RegsToUnmergeTo;
  SmallVector<Register> TruncOddDstRegs;
  SmallVector<Register> RegsToMerge;
 
  unsigned ElemCount = SrcTy.getNumElements();
 
  // Find the biggest size chunks we can work with
  int StepSize = ElemCount % 4 ? 2 : 4;
 
  // If we have a power of 2 greater than 2, we need to first unmerge into
  // enough pieces
  if (ElemCount <= 2)
    RegsToUnmergeTo.push_back(Src);
  else {
    for (unsigned i = 0; i < ElemCount / 2; ++i)
      RegsToUnmergeTo.push_back(MRI.createGenericVirtualRegister(v2s64));
 
    MIRBuilder.buildUnmerge(RegsToUnmergeTo, Src);
  }
 
  // Create all of the round-to-odd instructions and store them
  for (auto SrcReg : RegsToUnmergeTo) {
    Register Mid = MIRBuilder
                       .buildInstr(AArch64::G_FPTRUNC_ODD, {v2s32}, {SrcReg},
                                   MI.getFlags())
                       .getReg(0);
    TruncOddDstRegs.push_back(Mid);
  }
 
  // Truncate 4s32 to 4s16 if we can to reduce instruction count, otherwise
  // truncate 2s32 to 2s16.
  unsigned Index = 0;
  for (unsigned LoopIter = 0; LoopIter < ElemCount / StepSize; ++LoopIter) {
    if (StepSize == 4) {
      Register ConcatDst =
          MIRBuilder
              .buildMergeLikeInstr(
                  {v4s32}, {TruncOddDstRegs[Index++], TruncOddDstRegs[Index++]})
              .getReg(0);
 
      RegsToMerge.push_back(
          MIRBuilder.buildFPTrunc(v4s16, ConcatDst, MI.getFlags()).getReg(0));
    } else {
      RegsToMerge.push_back(
          MIRBuilder
              .buildFPTrunc(v2s16, TruncOddDstRegs[Index++], MI.getFlags())
              .getReg(0));
    }
  }
 
  // If there is only one register, replace the destination
  if (RegsToMerge.size() == 1) {
    MRI.replaceRegWith(Dst, RegsToMerge.pop_back_val());
    MI.eraseFromParent();
    return true;
  }
 
  // Merge the rest of the instructions & replace the register
  Register Fin = MIRBuilder.buildMergeLikeInstr(DstTy, RegsToMerge).getReg(0);
  MRI.replaceRegWith(Dst, Fin);
  MI.eraseFromParent();
  return true;
}
 
bool AArch64LegalizerInfo::legalizeGetRounding(MachineInstr &MI,
                                               MachineIRBuilder &MIRBuilder,
                                               MachineRegisterInfo &MRI,
                                               LegalizerHelper &Helper) const {
  const LLT I32 = LLT::integer(32);
  const LLT I64 = LLT::integer(64);
 
  Register Dst = MI.getOperand(0).getReg();
  Register FPCR64 = MRI.createGenericVirtualRegister(I64);
  MachineInstrBuilder GetFPCR =
      MIRBuilder.buildIntrinsic(Intrinsic::aarch64_get_fpcr, ArrayRef{FPCR64});
 
  // AArch64 rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3,
  // 3->0, so we add one to the FPCR bits for the rounding mode.
  // Instead of shifting and then adding as `((FPCR >> 22) + 1) & 0b11` which
  // generates 3 instructions, we increment the rounding mode with
  // `(FPCR + (1 << 22))` and extract the bits. The shift and addition is done
  // in one instruction as `add .., .., #1024, lsl #12`, so overall we generate
  // one less instruction.
  auto FPCR32 = MIRBuilder.buildTrunc(I32, GetFPCR);
  auto One = MIRBuilder.buildConstant(I32, 1U << 22);
  auto Added = MIRBuilder.buildAdd(I32, FPCR32, One);
  auto LSB = MIRBuilder.buildConstant(I32, 22);
  auto Width = MIRBuilder.buildConstant(I32, 2);
  MIRBuilder.buildInstr(TargetOpcode::G_UBFX, {Dst}, {Added, LSB, Width});
 
  MI.eraseFromParent();
  return true;
}
 
bool AArch64LegalizerInfo::legalizeSetRounding(MachineInstr &MI,
                                               MachineIRBuilder &MIRBuilder,
                                               MachineRegisterInfo &MRI,
                                               LegalizerHelper &Helper) const {
  const LLT I32 = LLT::integer(32);
  const LLT I64 = LLT::integer(64);
 
  // AArch64 rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3,
  // 3->0, so calculate the new value of FPCR[23:22] as `((arg - 1) & 3) << 22`.
  Register RM = MI.getOperand(0).getReg();
  auto One = MIRBuilder.buildConstant(I32, 1);
  auto Subtracted = MIRBuilder.buildSub(I32, RM, One);
  auto Mask = MIRBuilder.buildConstant(I32, 0b11);
  auto Masked = MIRBuilder.buildAnd(I32, Subtracted, Mask);
  auto ShiftAmount = MIRBuilder.buildConstant(I32, 22);
  auto Shifted = MIRBuilder.buildShl(I32, Masked, ShiftAmount);
 
  // Get current value of FPCR.
  MachineInstrBuilder GetFPCR =
      MIRBuilder.buildIntrinsic(Intrinsic::aarch64_get_fpcr, {I64});
 
  // (FPCR & ~Mask) | Shifted
  auto FPCRMask = MIRBuilder.buildConstant(I64, ~((int64_t)0b11 << 22));
  auto FPCRMasked = MIRBuilder.buildAnd(I64, GetFPCR, FPCRMask);
  auto ShiftedS64 = MIRBuilder.buildZExt(I64, Shifted);
  auto FPCRUpdated = MIRBuilder.buildOr(I64, FPCRMasked, ShiftedS64);
 
  // Write new FPCR.
  MIRBuilder.buildIntrinsic(Intrinsic::aarch64_set_fpcr, ArrayRef<Register>())
      .addUse(FPCRUpdated.getReg(0));
 
  MI.eraseFromParent();
 
  return true;
}