//===-- AArch64ISelLowering.cpp - AArch64 DAG Lowering Implementation  ----===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// This file implements the AArch64TargetLowering class.
//
//===----------------------------------------------------------------------===//

#include "AArch64ISelLowering.h"
#include "AArch64CallingConvention.h"
#include "AArch64ExpandImm.h"
#include "AArch64MachineFunctionInfo.h"
#include "AArch64PerfectShuffle.h"
#include "AArch64RegisterInfo.h"
#include "AArch64Subtarget.h"
#include "MCTargetDesc/AArch64AddressingModes.h"
#include "Utils/AArch64BaseInfo.h"
#include "Utils/AArch64SMEAttributes.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/ArrayRef.h"
#include "llvm/ADT/STLExtras.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/SmallVectorExtras.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringRef.h"
#include "llvm/ADT/Twine.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/MemoryLocation.h"
#include "llvm/Analysis/ObjCARCUtil.h"
#include "llvm/Analysis/OptimizationRemarkEmitter.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/ValueTracking.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/ComplexDeinterleavingPass.h"
#include "llvm/CodeGen/GlobalISel/Utils.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/CodeGen/TargetCallingConv.h"
#include "llvm/CodeGen/TargetInstrInfo.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/CodeGenTypes/MachineValueType.h"
#include "llvm/IR/Attributes.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsAArch64.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Use.h"
#include "llvm/IR/Value.h"
#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/InstructionCost.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/SipHash.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/TargetParser/Triple.h"
#include <algorithm>
#include <bitset>
#include <cassert>
#include <cctype>
#include <cstdint>
#include <cstdlib>
#include <iterator>
#include <limits>
#include <optional>
#include <tuple>
#include <utility>
#include <vector>

using namespace llvm;
using namespace llvm::PatternMatch;

#define DEBUG_TYPE "aarch64-lower"

STATISTIC(NumTailCalls, "Number of tail calls");
STATISTIC(NumOptimizedImms, "Number of times immediates were optimized");

// FIXME: The necessary dtprel relocations don't seem to be supported
// well in the GNU bfd and gold linkers at the moment. Therefore, by
// default, for now, fall back to GeneralDynamic code generation.
cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
    "aarch64-elf-ldtls-generation", cl::Hidden,
    cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
    cl::init(false));

static cl::opt<bool>
EnableOptimizeLogicalImm("aarch64-enable-logical-imm", cl::Hidden,
                         cl::desc("Enable AArch64 logical imm instruction "
                                  "optimization"),
                         cl::init(true));

// Temporary option added for the purpose of testing functionality added
// to DAGCombiner.cpp in D92230. It is expected that this can be removed
// in future when both implementations will be based off MGATHER rather
// than the GLD1 nodes added for the SVE gather load intrinsics.
static cl::opt<bool>
EnableCombineMGatherIntrinsics("aarch64-enable-mgather-combine", cl::Hidden,
                                cl::desc("Combine extends of AArch64 masked "
                                         "gather intrinsics"),
                                cl::init(true));

static cl::opt<bool> EnableExtToTBL("aarch64-enable-ext-to-tbl", cl::Hidden,
                                    cl::desc("Combine ext and trunc to TBL"),
                                    cl::init(true));

// All of the XOR, OR and CMP use ALU ports, and data dependency will become the
// bottleneck after this transform on high end CPU. So this max leaf node
// limitation is guard cmp+ccmp will be profitable.
static cl::opt<unsigned> MaxXors("aarch64-max-xors", cl::init(16), cl::Hidden,
                                 cl::desc("Maximum of xors"));

// By turning this on, we will not fallback to DAG ISel when encountering
// scalable vector types for all instruction, even if SVE is not yet supported
// with some instructions.
// See [AArch64TargetLowering::fallbackToDAGISel] for implementation details.
cl::opt<bool> EnableSVEGISel(
    "aarch64-enable-gisel-sve", cl::Hidden,
    cl::desc("Enable / disable SVE scalable vectors in Global ISel"),
    cl::init(false));

// TODO: This option should be removed once we switch to always using PTRADD in
// the SelectionDAG.
static cl::opt<bool> UseFEATCPACodegen(
    "aarch64-use-featcpa-codegen", cl::Hidden,
    cl::desc("Generate ISD::PTRADD nodes for pointer arithmetic in "
             "SelectionDAG for FEAT_CPA"),
    cl::init(false));

/// Value type used for condition codes.
static const MVT MVT_CC = MVT::i32;

static const MCPhysReg GPRArgRegs[] = {AArch64::X0, AArch64::X1, AArch64::X2,
                                       AArch64::X3, AArch64::X4, AArch64::X5,
                                       AArch64::X6, AArch64::X7};
static const MCPhysReg FPRArgRegs[] = {AArch64::Q0, AArch64::Q1, AArch64::Q2,
                                       AArch64::Q3, AArch64::Q4, AArch64::Q5,
                                       AArch64::Q6, AArch64::Q7};

ArrayRef<MCPhysReg> llvm::AArch64::getGPRArgRegs() { return GPRArgRegs; }

ArrayRef<MCPhysReg> llvm::AArch64::getFPRArgRegs() { return FPRArgRegs; }

static inline EVT getPackedSVEVectorVT(EVT VT) {
  switch (VT.getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("unexpected element type for vector");
  case MVT::i8:
    return MVT::nxv16i8;
  case MVT::i16:
    return MVT::nxv8i16;
  case MVT::i32:
    return MVT::nxv4i32;
  case MVT::i64:
    return MVT::nxv2i64;
  case MVT::f16:
    return MVT::nxv8f16;
  case MVT::f32:
    return MVT::nxv4f32;
  case MVT::f64:
    return MVT::nxv2f64;
  case MVT::bf16:
    return MVT::nxv8bf16;
  }
}

// NOTE: Currently there's only a need to return integer vector types. If this
// changes then just add an extra "type" parameter.
static inline EVT getPackedSVEVectorVT(ElementCount EC) {
  switch (EC.getKnownMinValue()) {
  default:
    llvm_unreachable("unexpected element count for vector");
  case 16:
    return MVT::nxv16i8;
  case 8:
    return MVT::nxv8i16;
  case 4:
    return MVT::nxv4i32;
  case 2:
    return MVT::nxv2i64;
  }
}

static inline EVT getPromotedVTForPredicate(EVT VT) {
  assert(VT.isScalableVector() && (VT.getVectorElementType() == MVT::i1) &&
         "Expected scalable predicate vector type!");
  switch (VT.getVectorMinNumElements()) {
  default:
    llvm_unreachable("unexpected element count for vector");
  case 2:
    return MVT::nxv2i64;
  case 4:
    return MVT::nxv4i32;
  case 8:
    return MVT::nxv8i16;
  case 16:
    return MVT::nxv16i8;
  }
}

/// Returns true if VT's elements occupy the lowest bit positions of its
/// associated register class without any intervening space.
///
/// For example, nxv2f16, nxv4f16 and nxv8f16 are legal types that belong to the
/// same register class, but only nxv8f16 can be treated as a packed vector.
static inline bool isPackedVectorType(EVT VT, SelectionDAG &DAG) {
  assert(VT.isVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
         "Expected legal vector type!");
  return VT.isFixedLengthVector() ||
         VT.getSizeInBits().getKnownMinValue() == AArch64::SVEBitsPerBlock;
}

// Returns true for ####_MERGE_PASSTHRU opcodes, whose operands have a leading
// predicate and end with a passthru value matching the result type.
static bool isMergePassthruOpcode(unsigned Opc) {
  switch (Opc) {
  default:
    return false;
  case AArch64ISD::BITREVERSE_MERGE_PASSTHRU:
  case AArch64ISD::BSWAP_MERGE_PASSTHRU:
  case AArch64ISD::REVH_MERGE_PASSTHRU:
  case AArch64ISD::REVW_MERGE_PASSTHRU:
  case AArch64ISD::REVD_MERGE_PASSTHRU:
  case AArch64ISD::CTLZ_MERGE_PASSTHRU:
  case AArch64ISD::CTPOP_MERGE_PASSTHRU:
  case AArch64ISD::DUP_MERGE_PASSTHRU:
  case AArch64ISD::ABS_MERGE_PASSTHRU:
  case AArch64ISD::NEG_MERGE_PASSTHRU:
  case AArch64ISD::FNEG_MERGE_PASSTHRU:
  case AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU:
  case AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU:
  case AArch64ISD::FCEIL_MERGE_PASSTHRU:
  case AArch64ISD::FFLOOR_MERGE_PASSTHRU:
  case AArch64ISD::FNEARBYINT_MERGE_PASSTHRU:
  case AArch64ISD::FRINT_MERGE_PASSTHRU:
  case AArch64ISD::FROUND_MERGE_PASSTHRU:
  case AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU:
  case AArch64ISD::FTRUNC_MERGE_PASSTHRU:
  case AArch64ISD::FP_ROUND_MERGE_PASSTHRU:
  case AArch64ISD::FP_EXTEND_MERGE_PASSTHRU:
  case AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU:
  case AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU:
  case AArch64ISD::FCVTX_MERGE_PASSTHRU:
  case AArch64ISD::FCVTZU_MERGE_PASSTHRU:
  case AArch64ISD::FCVTZS_MERGE_PASSTHRU:
  case AArch64ISD::FSQRT_MERGE_PASSTHRU:
  case AArch64ISD::FRECPX_MERGE_PASSTHRU:
  case AArch64ISD::FABS_MERGE_PASSTHRU:
    return true;
  }
}

// Returns true if inactive lanes are known to be zeroed by construction.
static bool isZeroingInactiveLanes(SDValue Op) {
  switch (Op.getOpcode()) {
  default:
    return false;
  // We guarantee i1 splat_vectors to zero the other lanes
  case ISD::SPLAT_VECTOR:
  case ISD::GET_ACTIVE_LANE_MASK:
  case AArch64ISD::PTRUE:
  case AArch64ISD::SETCC_MERGE_ZERO:
    return true;
  case ISD::INTRINSIC_WO_CHAIN:
    switch (Op.getConstantOperandVal(0)) {
    default:
      return false;
    case Intrinsic::aarch64_sve_ptrue:
    case Intrinsic::aarch64_sve_pnext:
    case Intrinsic::aarch64_sve_cmpeq:
    case Intrinsic::aarch64_sve_cmpne:
    case Intrinsic::aarch64_sve_cmpge:
    case Intrinsic::aarch64_sve_cmpgt:
    case Intrinsic::aarch64_sve_cmphs:
    case Intrinsic::aarch64_sve_cmphi:
    case Intrinsic::aarch64_sve_cmpeq_wide:
    case Intrinsic::aarch64_sve_cmpne_wide:
    case Intrinsic::aarch64_sve_cmpge_wide:
    case Intrinsic::aarch64_sve_cmpgt_wide:
    case Intrinsic::aarch64_sve_cmplt_wide:
    case Intrinsic::aarch64_sve_cmple_wide:
    case Intrinsic::aarch64_sve_cmphs_wide:
    case Intrinsic::aarch64_sve_cmphi_wide:
    case Intrinsic::aarch64_sve_cmplo_wide:
    case Intrinsic::aarch64_sve_cmpls_wide:
    case Intrinsic::aarch64_sve_fcmpeq:
    case Intrinsic::aarch64_sve_fcmpne:
    case Intrinsic::aarch64_sve_fcmpge:
    case Intrinsic::aarch64_sve_fcmpgt:
    case Intrinsic::aarch64_sve_fcmpuo:
    case Intrinsic::aarch64_sve_facgt:
    case Intrinsic::aarch64_sve_facge:
    case Intrinsic::aarch64_sve_whilege:
    case Intrinsic::aarch64_sve_whilegt:
    case Intrinsic::aarch64_sve_whilehi:
    case Intrinsic::aarch64_sve_whilehs:
    case Intrinsic::aarch64_sve_whilele:
    case Intrinsic::aarch64_sve_whilelo:
    case Intrinsic::aarch64_sve_whilels:
    case Intrinsic::aarch64_sve_whilelt:
    case Intrinsic::aarch64_sve_match:
    case Intrinsic::aarch64_sve_nmatch:
    case Intrinsic::aarch64_sve_whilege_x2:
    case Intrinsic::aarch64_sve_whilegt_x2:
    case Intrinsic::aarch64_sve_whilehi_x2:
    case Intrinsic::aarch64_sve_whilehs_x2:
    case Intrinsic::aarch64_sve_whilele_x2:
    case Intrinsic::aarch64_sve_whilelo_x2:
    case Intrinsic::aarch64_sve_whilels_x2:
    case Intrinsic::aarch64_sve_whilelt_x2:
      return true;
    }
  }
}

static std::tuple<SDValue, SDValue>
extractPtrauthBlendDiscriminators(SDValue Disc, SelectionDAG *DAG) {
  SDLoc DL(Disc);
  SDValue AddrDisc;
  SDValue ConstDisc;

  // If this is a blend, remember the constant and address discriminators.
  // Otherwise, it's either a constant discriminator, or a non-blended
  // address discriminator.
  if (Disc->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
      Disc->getConstantOperandVal(0) == Intrinsic::ptrauth_blend) {
    AddrDisc = Disc->getOperand(1);
    ConstDisc = Disc->getOperand(2);
  } else {
    ConstDisc = Disc;
  }

  // If the constant discriminator (either the blend RHS, or the entire
  // discriminator value) isn't a 16-bit constant, bail out, and let the
  // discriminator be computed separately.
  const auto *ConstDiscN = dyn_cast<ConstantSDNode>(ConstDisc);
  if (!ConstDiscN || !isUInt<16>(ConstDiscN->getZExtValue()))
    return std::make_tuple(DAG->getTargetConstant(0, DL, MVT::i64), Disc);

  // If there's no address discriminator, use NoRegister, which we'll later
  // replace with XZR, or directly use a Z variant of the inst. when available.
  if (!AddrDisc)
    AddrDisc = DAG->getRegister(AArch64::NoRegister, MVT::i64);

  return std::make_tuple(
      DAG->getTargetConstant(ConstDiscN->getZExtValue(), DL, MVT::i64),
      AddrDisc);
}

AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
                                             const AArch64Subtarget &STI)
    : TargetLowering(TM), Subtarget(&STI) {
  // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
  // we have to make something up. Arbitrarily, choose ZeroOrOne.
  setBooleanContents(ZeroOrOneBooleanContent);
  // When comparing vectors the result sets the different elements in the
  // vector to all-one or all-zero.
  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);

  // Set up the register classes.
  addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
  addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);

  if (Subtarget->hasLS64()) {
    addRegisterClass(MVT::i64x8, &AArch64::GPR64x8ClassRegClass);
    setOperationAction(ISD::LOAD, MVT::i64x8, Custom);
    setOperationAction(ISD::STORE, MVT::i64x8, Custom);
  }

  if (Subtarget->hasFPARMv8()) {
    addRegisterClass(MVT::aarch64mfp8, &AArch64::FPR8RegClass);
    addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
    addRegisterClass(MVT::bf16, &AArch64::FPR16RegClass);
    addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
    addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
    addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
  }

  if (Subtarget->hasNEON()) {
    addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
    addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);

    addDRType(MVT::v2f32);
    addDRType(MVT::v8i8);
    addDRType(MVT::v4i16);
    addDRType(MVT::v2i32);
    addDRType(MVT::v1i64);
    addDRType(MVT::v1f64);
    addDRType(MVT::v4f16);
    addDRType(MVT::v4bf16);

    addQRType(MVT::v4f32);
    addQRType(MVT::v2f64);
    addQRType(MVT::v16i8);
    addQRType(MVT::v8i16);
    addQRType(MVT::v4i32);
    addQRType(MVT::v2i64);
    addQRType(MVT::v8f16);
    addQRType(MVT::v8bf16);
  }

  if (Subtarget->isSVEorStreamingSVEAvailable()) {
    // Add legal sve predicate types
    addRegisterClass(MVT::nxv1i1, &AArch64::PPRRegClass);
    addRegisterClass(MVT::nxv2i1, &AArch64::PPRRegClass);
    addRegisterClass(MVT::nxv4i1, &AArch64::PPRRegClass);
    addRegisterClass(MVT::nxv8i1, &AArch64::PPRRegClass);
    addRegisterClass(MVT::nxv16i1, &AArch64::PPRRegClass);

    // Add legal sve data types
    addRegisterClass(MVT::nxv16i8, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv8i16, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv4i32, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv2i64, &AArch64::ZPRRegClass);

    addRegisterClass(MVT::nxv2f16, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv4f16, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv8f16, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv2f32, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv4f32, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv2f64, &AArch64::ZPRRegClass);

    addRegisterClass(MVT::nxv2bf16, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv4bf16, &AArch64::ZPRRegClass);
    addRegisterClass(MVT::nxv8bf16, &AArch64::ZPRRegClass);

    if (Subtarget->useSVEForFixedLengthVectors()) {
      for (MVT VT : MVT::integer_fixedlen_vector_valuetypes())
        if (useSVEForFixedLengthVectorVT(VT))
          addRegisterClass(VT, &AArch64::ZPRRegClass);

      for (MVT VT : MVT::fp_fixedlen_vector_valuetypes())
        if (useSVEForFixedLengthVectorVT(VT))
          addRegisterClass(VT, &AArch64::ZPRRegClass);
    }
  }

  if (Subtarget->hasSVE2p1() || Subtarget->hasSME2()) {
    addRegisterClass(MVT::aarch64svcount, &AArch64::PPRRegClass);
    setOperationPromotedToType(ISD::LOAD, MVT::aarch64svcount, MVT::nxv16i1);
    setOperationPromotedToType(ISD::STORE, MVT::aarch64svcount, MVT::nxv16i1);

    setOperationAction(ISD::SELECT, MVT::aarch64svcount, Custom);
    setOperationAction(ISD::SELECT_CC, MVT::aarch64svcount, Expand);
  }

  // Compute derived properties from the register classes
  computeRegisterProperties(Subtarget->getRegisterInfo());

  // Provide all sorts of operation actions
  setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
  setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
  setOperationAction(ISD::SETCC, MVT::i32, Custom);
  setOperationAction(ISD::SETCC, MVT::i64, Custom);
  setOperationAction(ISD::SETCC, MVT::bf16, Custom);
  setOperationAction(ISD::SETCC, MVT::f16, Custom);
  setOperationAction(ISD::SETCC, MVT::f32, Custom);
  setOperationAction(ISD::SETCC, MVT::f64, Custom);
  setOperationAction(ISD::STRICT_FSETCC, MVT::bf16, Custom);
  setOperationAction(ISD::STRICT_FSETCC, MVT::f16, Custom);
  setOperationAction(ISD::STRICT_FSETCC, MVT::f32, Custom);
  setOperationAction(ISD::STRICT_FSETCC, MVT::f64, Custom);
  setOperationAction(ISD::STRICT_FSETCCS, MVT::f16, Custom);
  setOperationAction(ISD::STRICT_FSETCCS, MVT::f32, Custom);
  setOperationAction(ISD::STRICT_FSETCCS, MVT::f64, Custom);
  setOperationAction(ISD::BITREVERSE, MVT::i32, Legal);
  setOperationAction(ISD::BITREVERSE, MVT::i64, Legal);
  setOperationAction(ISD::BRCOND, MVT::Other, Custom);
  setOperationAction(ISD::BR_CC, MVT::i32, Custom);
  setOperationAction(ISD::BR_CC, MVT::i64, Custom);
  setOperationAction(ISD::BR_CC, MVT::f16, Custom);
  setOperationAction(ISD::BR_CC, MVT::f32, Custom);
  setOperationAction(ISD::BR_CC, MVT::f64, Custom);
  setOperationAction(ISD::SELECT, MVT::i32, Custom);
  setOperationAction(ISD::SELECT, MVT::i64, Custom);
  if (Subtarget->hasFPARMv8()) {
    setOperationAction(ISD::SELECT, MVT::f16, Custom);
    setOperationAction(ISD::SELECT, MVT::bf16, Custom);
  }
  setOperationAction(ISD::SELECT, MVT::f32, Custom);
  setOperationAction(ISD::SELECT, MVT::f64, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::i64, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f16, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::bf16, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
  setOperationAction(ISD::BR_JT, MVT::Other, Custom);
  setOperationAction(ISD::JumpTable, MVT::i64, Custom);
  setOperationAction(ISD::BRIND, MVT::Other, Custom);
  setOperationAction(ISD::SETCCCARRY, MVT::i64, Custom);

  setOperationAction(ISD::PtrAuthGlobalAddress, MVT::i64, Custom);

  setOperationAction(ISD::SHL_PARTS, MVT::i64, Custom);
  setOperationAction(ISD::SRA_PARTS, MVT::i64, Custom);
  setOperationAction(ISD::SRL_PARTS, MVT::i64, Custom);

  setOperationAction(ISD::FREM, MVT::f32, Expand);
  setOperationAction(ISD::FREM, MVT::f64, Expand);
  setOperationAction(ISD::FREM, MVT::f80, Expand);

  setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);

  // Custom lowering hooks are needed for XOR
  // to fold it into CSINC/CSINV.
  setOperationAction(ISD::XOR, MVT::i32, Custom);
  setOperationAction(ISD::XOR, MVT::i64, Custom);

  setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom);
  setOperationAction(ISD::ADDRSPACECAST, MVT::i64, Custom);

  // Virtually no operation on f128 is legal, but LLVM can't expand them when
  // there's a valid register class, so we need custom operations in most cases.
  setOperationAction(ISD::FABS, MVT::f128, Expand);
  setOperationAction(ISD::FADD, MVT::f128, LibCall);
  setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
  setOperationAction(ISD::FCOS, MVT::f128, Expand);
  setOperationAction(ISD::FDIV, MVT::f128, LibCall);
  setOperationAction(ISD::FMA, MVT::f128, Expand);
  setOperationAction(ISD::FMUL, MVT::f128, LibCall);
  setOperationAction(ISD::FNEG, MVT::f128, Expand);
  setOperationAction(ISD::FPOW, MVT::f128, Expand);
  setOperationAction(ISD::FREM, MVT::f128, Expand);
  setOperationAction(ISD::FRINT, MVT::f128, Expand);
  setOperationAction(ISD::FSIN, MVT::f128, Expand);
  setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
  setOperationAction(ISD::FSQRT, MVT::f128, Expand);
  setOperationAction(ISD::FSUB, MVT::f128, LibCall);
  setOperationAction(ISD::FTAN, MVT::f128, Expand);
  setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
  setOperationAction(ISD::SETCC, MVT::f128, Custom);
  setOperationAction(ISD::STRICT_FSETCC, MVT::f128, Custom);
  setOperationAction(ISD::STRICT_FSETCCS, MVT::f128, Custom);
  setOperationAction(ISD::BR_CC, MVT::f128, Custom);
  setOperationAction(ISD::SELECT, MVT::f128, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
  setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);
  // FIXME: f128 FMINIMUM and FMAXIMUM (including STRICT versions) currently
  // aren't handled.

  // Lowering for many of the conversions is actually specified by the non-f128
  // type. The LowerXXX function will be trivial when f128 isn't involved.
  setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
  setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
  setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom);
  setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom);
  setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i128, Custom);
  setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
  setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
  setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom);
  setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Custom);
  setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i128, Custom);
  setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
  setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
  setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom);
  setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom);
  setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
  setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
  setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom);
  setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom);
  setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i128, Custom);
  if (Subtarget->hasFPARMv8()) {
    setOperationAction(ISD::FP_ROUND, MVT::f16, Custom);
    setOperationAction(ISD::FP_ROUND, MVT::bf16, Custom);
  }
  setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
  setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
  if (Subtarget->hasFPARMv8()) {
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Custom);
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::bf16, Custom);
  }
  setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Custom);
  setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Custom);

  setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i32, Custom);
  setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i64, Custom);
  setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i32, Custom);
  setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i64, Custom);

  // Variable arguments.
  setOperationAction(ISD::VASTART, MVT::Other, Custom);
  setOperationAction(ISD::VAARG, MVT::Other, Custom);
  setOperationAction(ISD::VACOPY, MVT::Other, Custom);
  setOperationAction(ISD::VAEND, MVT::Other, Expand);

  // Variable-sized objects.
  setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
  setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);

  // Lowering Funnel Shifts to EXTR
  setOperationAction(ISD::FSHR, MVT::i32, Custom);
  setOperationAction(ISD::FSHR, MVT::i64, Custom);
  setOperationAction(ISD::FSHL, MVT::i32, Custom);
  setOperationAction(ISD::FSHL, MVT::i64, Custom);

  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Custom);

  // Constant pool entries
  setOperationAction(ISD::ConstantPool, MVT::i64, Custom);

  // BlockAddress
  setOperationAction(ISD::BlockAddress, MVT::i64, Custom);

  // AArch64 lacks both left-rotate and popcount instructions.
  setOperationAction(ISD::ROTL, MVT::i32, Expand);
  setOperationAction(ISD::ROTL, MVT::i64, Expand);
  for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
    setOperationAction(ISD::ROTL, VT, Expand);
    setOperationAction(ISD::ROTR, VT, Expand);
  }

  // AArch64 doesn't have i32 MULH{S|U}.
  setOperationAction(ISD::MULHU, MVT::i32, Expand);
  setOperationAction(ISD::MULHS, MVT::i32, Expand);

  // AArch64 doesn't have {U|S}MUL_LOHI.
  setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
  setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
  setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
  setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);

  if (Subtarget->hasCSSC()) {
    setOperationAction(ISD::CTPOP, MVT::i32, Legal);
    setOperationAction(ISD::CTPOP, MVT::i64, Legal);
    setOperationAction(ISD::CTPOP, MVT::i128, Expand);

    setOperationAction(ISD::PARITY, MVT::i128, Expand);

    setOperationAction(ISD::CTTZ, MVT::i32, Legal);
    setOperationAction(ISD::CTTZ, MVT::i64, Legal);
    setOperationAction(ISD::CTTZ, MVT::i128, Expand);

    setOperationAction(ISD::ABS, MVT::i32, Legal);
    setOperationAction(ISD::ABS, MVT::i64, Legal);

    setOperationAction(ISD::SMAX, MVT::i32, Legal);
    setOperationAction(ISD::SMAX, MVT::i64, Legal);
    setOperationAction(ISD::UMAX, MVT::i32, Legal);
    setOperationAction(ISD::UMAX, MVT::i64, Legal);

    setOperationAction(ISD::SMIN, MVT::i32, Legal);
    setOperationAction(ISD::SMIN, MVT::i64, Legal);
    setOperationAction(ISD::UMIN, MVT::i32, Legal);
    setOperationAction(ISD::UMIN, MVT::i64, Legal);
  } else {
    setOperationAction(ISD::CTPOP, MVT::i32, Custom);
    setOperationAction(ISD::CTPOP, MVT::i64, Custom);
    setOperationAction(ISD::CTPOP, MVT::i128, Custom);

    setOperationAction(ISD::PARITY, MVT::i64, Custom);
    setOperationAction(ISD::PARITY, MVT::i128, Custom);

    setOperationAction(ISD::ABS, MVT::i32, Custom);
    setOperationAction(ISD::ABS, MVT::i64, Custom);
  }

  setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
  setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
  for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
    setOperationAction(ISD::SDIVREM, VT, Expand);
    setOperationAction(ISD::UDIVREM, VT, Expand);
  }
  setOperationAction(ISD::SREM, MVT::i32, Expand);
  setOperationAction(ISD::SREM, MVT::i64, Expand);
  setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
  setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
  setOperationAction(ISD::UREM, MVT::i32, Expand);
  setOperationAction(ISD::UREM, MVT::i64, Expand);

  // Custom lower Add/Sub/Mul with overflow.
  setOperationAction(ISD::SADDO, MVT::i32, Custom);
  setOperationAction(ISD::SADDO, MVT::i64, Custom);
  setOperationAction(ISD::UADDO, MVT::i32, Custom);
  setOperationAction(ISD::UADDO, MVT::i64, Custom);
  setOperationAction(ISD::SSUBO, MVT::i32, Custom);
  setOperationAction(ISD::SSUBO, MVT::i64, Custom);
  setOperationAction(ISD::USUBO, MVT::i32, Custom);
  setOperationAction(ISD::USUBO, MVT::i64, Custom);
  setOperationAction(ISD::SMULO, MVT::i32, Custom);
  setOperationAction(ISD::SMULO, MVT::i64, Custom);
  setOperationAction(ISD::UMULO, MVT::i32, Custom);
  setOperationAction(ISD::UMULO, MVT::i64, Custom);

  setOperationAction(ISD::UADDO_CARRY, MVT::i32, Custom);
  setOperationAction(ISD::UADDO_CARRY, MVT::i64, Custom);
  setOperationAction(ISD::USUBO_CARRY, MVT::i32, Custom);
  setOperationAction(ISD::USUBO_CARRY, MVT::i64, Custom);
  setOperationAction(ISD::SADDO_CARRY, MVT::i32, Custom);
  setOperationAction(ISD::SADDO_CARRY, MVT::i64, Custom);
  setOperationAction(ISD::SSUBO_CARRY, MVT::i32, Custom);
  setOperationAction(ISD::SSUBO_CARRY, MVT::i64, Custom);

  setOperationAction(ISD::FSIN, MVT::f32, Expand);
  setOperationAction(ISD::FSIN, MVT::f64, Expand);
  setOperationAction(ISD::FCOS, MVT::f32, Expand);
  setOperationAction(ISD::FCOS, MVT::f64, Expand);
  setOperationAction(ISD::FPOW, MVT::f32, Expand);
  setOperationAction(ISD::FPOW, MVT::f64, Expand);
  setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
  setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
  if (Subtarget->hasFullFP16()) {
    setOperationAction(ISD::FCOPYSIGN, MVT::f16, Custom);
    setOperationAction(ISD::FCOPYSIGN, MVT::bf16, Custom);
  } else {
    setOperationAction(ISD::FCOPYSIGN, MVT::f16, Promote);
    setOperationAction(ISD::FCOPYSIGN, MVT::bf16, Promote);
  }

  for (auto Op : {ISD::FREM,          ISD::FPOW,         ISD::FPOWI,
                  ISD::FCOS,          ISD::FSIN,         ISD::FSINCOS,
                  ISD::FSINCOSPI,     ISD::FMODF,        ISD::FACOS,
                  ISD::FASIN,         ISD::FATAN,        ISD::FATAN2,
                  ISD::FCOSH,         ISD::FSINH,        ISD::FTANH,
                  ISD::FTAN,          ISD::FEXP,         ISD::FEXP2,
                  ISD::FEXP10,        ISD::FLOG,         ISD::FLOG2,
                  ISD::FLOG10,        ISD::STRICT_FREM,  ISD::STRICT_FPOW,
                  ISD::STRICT_FPOWI,  ISD::STRICT_FCOS,  ISD::STRICT_FSIN,
                  ISD::STRICT_FACOS,  ISD::STRICT_FASIN, ISD::STRICT_FATAN,
                  ISD::STRICT_FATAN2, ISD::STRICT_FCOSH, ISD::STRICT_FSINH,
                  ISD::STRICT_FTANH,  ISD::STRICT_FEXP,  ISD::STRICT_FEXP2,
                  ISD::STRICT_FLOG,   ISD::STRICT_FLOG2, ISD::STRICT_FLOG10,
                  ISD::STRICT_FTAN}) {
    setOperationAction(Op, MVT::f16, Promote);
    setOperationAction(Op, MVT::v4f16, Expand);
    setOperationAction(Op, MVT::v8f16, Expand);
    setOperationAction(Op, MVT::bf16, Promote);
    setOperationAction(Op, MVT::v4bf16, Expand);
    setOperationAction(Op, MVT::v8bf16, Expand);
  }

  // Legalize fcanonicalize to circumvent default expansion
  setOperationAction(ISD::FCANONICALIZE, {MVT::f32, MVT::f64}, Legal);
  if (Subtarget->hasFullFP16()) {
    setOperationAction(ISD::FCANONICALIZE, MVT::f16, Legal);
  }

  // fpextend from f16 or bf16 to f32 is legal
  setOperationAction(ISD::FP_EXTEND, MVT::f32, Legal);
  setOperationAction(ISD::FP_EXTEND, MVT::v4f32, Legal);
  setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Legal);
  setOperationAction(ISD::STRICT_FP_EXTEND, MVT::v4f32, Legal);
  // fpextend from bf16 to f64 needs to be split into two fpextends
  setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
  setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Custom);

  auto LegalizeNarrowFP = [this](MVT ScalarVT) {
    for (auto Op : {
             ISD::SETCC,
             ISD::SELECT_CC,
             ISD::BR_CC,
             ISD::FADD,
             ISD::FSUB,
             ISD::FMUL,
             ISD::FDIV,
             ISD::FMA,
             ISD::FCEIL,
             ISD::FSQRT,
             ISD::FFLOOR,
             ISD::FNEARBYINT,
             ISD::FRINT,
             ISD::FROUND,
             ISD::FROUNDEVEN,
             ISD::FTRUNC,
             ISD::FMINNUM,
             ISD::FMAXNUM,
             ISD::FMINIMUM,
             ISD::FMAXIMUM,
             ISD::FMINIMUMNUM,
             ISD::FMAXIMUMNUM,
             ISD::FCANONICALIZE,
             ISD::STRICT_FADD,
             ISD::STRICT_FSUB,
             ISD::STRICT_FMUL,
             ISD::STRICT_FDIV,
             ISD::STRICT_FMA,
             ISD::STRICT_FCEIL,
             ISD::STRICT_FFLOOR,
             ISD::STRICT_FSQRT,
             ISD::STRICT_FRINT,
             ISD::STRICT_FNEARBYINT,
             ISD::STRICT_FROUND,
             ISD::STRICT_FTRUNC,
             ISD::STRICT_FROUNDEVEN,
             ISD::STRICT_FMINNUM,
             ISD::STRICT_FMAXNUM,
             ISD::STRICT_FMINIMUM,
             ISD::STRICT_FMAXIMUM,
         })
      setOperationAction(Op, ScalarVT, Promote);

    for (auto Op : {ISD::FNEG, ISD::FABS})
      setOperationAction(Op, ScalarVT, Legal);

    // Round-to-integer need custom lowering for fp16, as Promote doesn't work
    // because the result type is integer.
    for (auto Op : {ISD::LROUND, ISD::LLROUND, ISD::LRINT, ISD::LLRINT,
                    ISD::STRICT_LROUND, ISD::STRICT_LLROUND, ISD::STRICT_LRINT,
                    ISD::STRICT_LLRINT})
      setOperationAction(Op, ScalarVT, Custom);

    // promote v4f16 to v4f32 when that is known to be safe.
    auto V4Narrow = MVT::getVectorVT(ScalarVT, 4);
    setOperationPromotedToType(ISD::FADD,       V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FSUB,       V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FMUL,       V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FDIV,       V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FCEIL,      V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FFLOOR,     V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FROUND,     V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FTRUNC,     V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FROUNDEVEN, V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FRINT,      V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FNEARBYINT, V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::FCANONICALIZE, V4Narrow, MVT::v4f32);
    setOperationPromotedToType(ISD::SETCC,         V4Narrow, MVT::v4f32);

    setOperationAction(ISD::FABS,        V4Narrow, Legal);
    setOperationAction(ISD::FNEG,        V4Narrow, Legal);
    setOperationAction(ISD::FMA,         V4Narrow, Expand);
    setOperationAction(ISD::BR_CC,       V4Narrow, Expand);
    setOperationAction(ISD::SELECT,      V4Narrow, Expand);
    setOperationAction(ISD::SELECT_CC,   V4Narrow, Expand);
    setOperationAction(ISD::FCOPYSIGN,   V4Narrow, Custom);
    setOperationAction(ISD::FSQRT,       V4Narrow, Expand);

    auto V8Narrow = MVT::getVectorVT(ScalarVT, 8);
    setOperationPromotedToType(ISD::FCANONICALIZE, V8Narrow, MVT::v8f32);
    setOperationPromotedToType(ISD::SETCC,         V8Narrow, MVT::v8f32);

    setOperationAction(ISD::FABS,        V8Narrow, Legal);
    setOperationAction(ISD::FADD,        V8Narrow, Legal);
    setOperationAction(ISD::FCEIL,       V8Narrow, Legal);
    setOperationAction(ISD::FCOPYSIGN,   V8Narrow, Custom);
    setOperationAction(ISD::FDIV,        V8Narrow, Legal);
    setOperationAction(ISD::FFLOOR,      V8Narrow, Legal);
    setOperationAction(ISD::FMA,         V8Narrow, Expand);
    setOperationAction(ISD::FMUL,        V8Narrow, Legal);
    setOperationAction(ISD::FNEARBYINT,  V8Narrow, Legal);
    setOperationAction(ISD::FNEG,        V8Narrow, Legal);
    setOperationAction(ISD::FROUND,      V8Narrow, Legal);
    setOperationAction(ISD::FROUNDEVEN,  V8Narrow, Legal);
    setOperationAction(ISD::FRINT,       V8Narrow, Legal);
    setOperationAction(ISD::FSQRT,       V8Narrow, Expand);
    setOperationAction(ISD::FSUB,        V8Narrow, Legal);
    setOperationAction(ISD::FTRUNC,      V8Narrow, Legal);
    setOperationAction(ISD::BR_CC,       V8Narrow, Expand);
    setOperationAction(ISD::SELECT,      V8Narrow, Expand);
    setOperationAction(ISD::SELECT_CC,   V8Narrow, Expand);
    setOperationAction(ISD::FP_EXTEND,   V8Narrow, Expand);
  };

  if (!Subtarget->hasFullFP16()) {
    LegalizeNarrowFP(MVT::f16);
  }
  LegalizeNarrowFP(MVT::bf16);
  setOperationAction(ISD::FP_ROUND, MVT::v4f32, Custom);
  setOperationAction(ISD::FP_ROUND, MVT::v4bf16, Custom);

  // AArch64 has implementations of a lot of rounding-like FP operations.
  // clang-format off
  for (auto Op :
       {ISD::FFLOOR,          ISD::FNEARBYINT,      ISD::FCEIL,
        ISD::FRINT,           ISD::FTRUNC,          ISD::FROUND,
        ISD::FROUNDEVEN,      ISD::FMINNUM,         ISD::FMAXNUM,
        ISD::FMINIMUM,        ISD::FMAXIMUM,        ISD::LROUND,
        ISD::LLROUND,         ISD::LRINT,           ISD::LLRINT,
        ISD::FMINNUM_IEEE,    ISD::FMAXNUM_IEEE,
        ISD::STRICT_FFLOOR,   ISD::STRICT_FCEIL,    ISD::STRICT_FNEARBYINT,
        ISD::STRICT_FRINT,    ISD::STRICT_FTRUNC,   ISD::STRICT_FROUNDEVEN,
        ISD::STRICT_FROUND,   ISD::STRICT_FMINNUM,  ISD::STRICT_FMAXNUM,
        ISD::STRICT_FMINIMUM, ISD::STRICT_FMAXIMUM, ISD::STRICT_LROUND,
        ISD::STRICT_LLROUND,  ISD::STRICT_LRINT,    ISD::STRICT_LLRINT}) {
    for (MVT Ty : {MVT::f32, MVT::f64})
      setOperationAction(Op, Ty, Legal);
    if (Subtarget->hasFullFP16())
      setOperationAction(Op, MVT::f16, Legal);
  }
  // clang-format on

  // Basic strict FP operations are legal
  for (auto Op : {ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
                  ISD::STRICT_FDIV, ISD::STRICT_FMA, ISD::STRICT_FSQRT}) {
    for (MVT Ty : {MVT::f32, MVT::f64})
      setOperationAction(Op, Ty, Legal);
    if (Subtarget->hasFullFP16())
      setOperationAction(Op, MVT::f16, Legal);
  }

  setOperationAction(ISD::PREFETCH, MVT::Other, Custom);

  setOperationAction(ISD::GET_ROUNDING, MVT::i32, Custom);
  setOperationAction(ISD::SET_ROUNDING, MVT::Other, Custom);
  setOperationAction(ISD::GET_FPMODE, MVT::i32, Custom);
  setOperationAction(ISD::SET_FPMODE, MVT::i32, Custom);
  setOperationAction(ISD::RESET_FPMODE, MVT::Other, Custom);

  setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
  if (!Subtarget->hasLSE() && !Subtarget->outlineAtomics()) {
    setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, LibCall);
  } else {
    setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Expand);
  }
  setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i32, Custom);
  setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);

  // Generate outline atomics library calls only if LSE was not specified for
  // subtarget
  if (Subtarget->outlineAtomics() && !Subtarget->hasLSE()) {
    setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, LibCall);
    setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, LibCall);
    setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, LibCall);
    setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, LibCall);
    setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, LibCall);
    setOperationAction(ISD::ATOMIC_SWAP, MVT::i8, LibCall);
    setOperationAction(ISD::ATOMIC_SWAP, MVT::i16, LibCall);
    setOperationAction(ISD::ATOMIC_SWAP, MVT::i32, LibCall);
    setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i8, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i8, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i8, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_CLR, MVT::i64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i8, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, LibCall);
  }

  if (Subtarget->outlineAtomics() && !Subtarget->hasLSFE()) {
    setOperationAction(ISD::ATOMIC_LOAD_FADD, MVT::f16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FADD, MVT::f32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FADD, MVT::f64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FADD, MVT::bf16, LibCall);

    setOperationAction(ISD::ATOMIC_LOAD_FMAX, MVT::f16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMAX, MVT::f32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMAX, MVT::f64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMAX, MVT::bf16, LibCall);

    setOperationAction(ISD::ATOMIC_LOAD_FMIN, MVT::f16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMIN, MVT::f32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMIN, MVT::f64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMIN, MVT::bf16, LibCall);

    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::f16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::f32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::f64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMAXIMUM, MVT::bf16, LibCall);

    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::f16, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::f32, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::f64, LibCall);
    setOperationAction(ISD::ATOMIC_LOAD_FMINIMUM, MVT::bf16, LibCall);
  }

  if (Subtarget->hasLSE128()) {
    // Custom lowering because i128 is not legal. Must be replaced by 2x64
    // values. ATOMIC_LOAD_AND also needs op legalisation to emit LDCLRP.
    setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i128, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i128, Custom);
    setOperationAction(ISD::ATOMIC_SWAP, MVT::i128, Custom);
  }

  // 128-bit loads and stores can be done without expanding
  setOperationAction(ISD::LOAD, MVT::i128, Custom);
  setOperationAction(ISD::STORE, MVT::i128, Custom);

  // Aligned 128-bit loads and stores are single-copy atomic according to the
  // v8.4a spec. LRCPC3 introduces 128-bit STILP/LDIAPP but still requires LSE2.
  if (Subtarget->hasLSE2()) {
    setOperationAction(ISD::ATOMIC_LOAD, MVT::i128, Custom);
    setOperationAction(ISD::ATOMIC_STORE, MVT::i128, Custom);
  }

  // 256 bit non-temporal stores can be lowered to STNP. Do this as part of the
  // custom lowering, as there are no un-paired non-temporal stores and
  // legalization will break up 256 bit inputs.
  setOperationAction(ISD::STORE, MVT::v32i8, Custom);
  setOperationAction(ISD::STORE, MVT::v16i16, Custom);
  setOperationAction(ISD::STORE, MVT::v16f16, Custom);
  setOperationAction(ISD::STORE, MVT::v16bf16, Custom);
  setOperationAction(ISD::STORE, MVT::v8i32, Custom);
  setOperationAction(ISD::STORE, MVT::v8f32, Custom);
  setOperationAction(ISD::STORE, MVT::v4f64, Custom);
  setOperationAction(ISD::STORE, MVT::v4i64, Custom);

  // 256 bit non-temporal loads can be lowered to LDNP. This is done using
  // custom lowering, as there are no un-paired non-temporal loads legalization
  // will break up 256 bit inputs.
  setOperationAction(ISD::LOAD, MVT::v32i8, Custom);
  setOperationAction(ISD::LOAD, MVT::v16i16, Custom);
  setOperationAction(ISD::LOAD, MVT::v16f16, Custom);
  setOperationAction(ISD::LOAD, MVT::v16bf16, Custom);
  setOperationAction(ISD::LOAD, MVT::v8i32, Custom);
  setOperationAction(ISD::LOAD, MVT::v8f32, Custom);
  setOperationAction(ISD::LOAD, MVT::v4f64, Custom);
  setOperationAction(ISD::LOAD, MVT::v4i64, Custom);

  // Lower READCYCLECOUNTER using an mrs from CNTVCT_EL0.
  setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);

  if (getLibcallName(RTLIB::SINCOS_STRET_F32) != nullptr &&
      getLibcallName(RTLIB::SINCOS_STRET_F64) != nullptr) {
    // Issue __sincos_stret if available.
    setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
    setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
  } else {
    setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
    setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
  }

  // Make floating-point constants legal for the large code model, so they don't
  // become loads from the constant pool.
  if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
    setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
    setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
  }

  // AArch64 does not have floating-point extending loads, i1 sign-extending
  // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
  for (MVT VT : MVT::fp_valuetypes()) {
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::bf16, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
  }
  for (MVT VT : MVT::integer_valuetypes())
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);

  for (MVT WideVT : MVT::fp_valuetypes()) {
    for (MVT NarrowVT : MVT::fp_valuetypes()) {
      if (WideVT.getScalarSizeInBits() > NarrowVT.getScalarSizeInBits()) {
        setTruncStoreAction(WideVT, NarrowVT, Expand);
      }
    }
  }

  if (Subtarget->hasFPARMv8()) {
    setOperationAction(ISD::BITCAST, MVT::i16, Custom);
    setOperationAction(ISD::BITCAST, MVT::f16, Custom);
    setOperationAction(ISD::BITCAST, MVT::bf16, Custom);
  }

  // Indexed loads and stores are supported.
  for (unsigned im = (unsigned)ISD::PRE_INC;
       im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
    setIndexedLoadAction(im, MVT::i8, Legal);
    setIndexedLoadAction(im, MVT::i16, Legal);
    setIndexedLoadAction(im, MVT::i32, Legal);
    setIndexedLoadAction(im, MVT::i64, Legal);
    setIndexedLoadAction(im, MVT::f64, Legal);
    setIndexedLoadAction(im, MVT::f32, Legal);
    setIndexedLoadAction(im, MVT::f16, Legal);
    setIndexedLoadAction(im, MVT::bf16, Legal);
    setIndexedStoreAction(im, MVT::i8, Legal);
    setIndexedStoreAction(im, MVT::i16, Legal);
    setIndexedStoreAction(im, MVT::i32, Legal);
    setIndexedStoreAction(im, MVT::i64, Legal);
    setIndexedStoreAction(im, MVT::f64, Legal);
    setIndexedStoreAction(im, MVT::f32, Legal);
    setIndexedStoreAction(im, MVT::f16, Legal);
    setIndexedStoreAction(im, MVT::bf16, Legal);
  }

  // Trap.
  setOperationAction(ISD::TRAP, MVT::Other, Legal);
  setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
  setOperationAction(ISD::UBSANTRAP, MVT::Other, Legal);

  // We combine OR nodes for bitfield operations.
  setTargetDAGCombine(ISD::OR);
  // Try to create BICs for vector ANDs.
  setTargetDAGCombine(ISD::AND);

  // llvm.init.trampoline and llvm.adjust.trampoline
  setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
  setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);

  // Vector add and sub nodes may conceal a high-half opportunity.
  // Also, try to fold ADD into CSINC/CSINV..
  setTargetDAGCombine({ISD::ADD, ISD::ABS, ISD::SUB, ISD::XOR, ISD::SINT_TO_FP,
                       ISD::UINT_TO_FP});

  setTargetDAGCombine({ISD::FP_TO_SINT, ISD::FP_TO_UINT, ISD::FP_TO_SINT_SAT,
                       ISD::FP_TO_UINT_SAT, ISD::FADD});

  // Try and combine setcc with csel
  setTargetDAGCombine(ISD::SETCC);

  setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);

  setTargetDAGCombine({ISD::ANY_EXTEND, ISD::ZERO_EXTEND, ISD::SIGN_EXTEND,
                       ISD::SIGN_EXTEND_INREG, ISD::CONCAT_VECTORS,
                       ISD::EXTRACT_SUBVECTOR, ISD::INSERT_SUBVECTOR,
                       ISD::STORE, ISD::BUILD_VECTOR});
  setTargetDAGCombine(ISD::SMIN);
  setTargetDAGCombine(ISD::TRUNCATE);
  setTargetDAGCombine(ISD::LOAD);

  setTargetDAGCombine(ISD::MSTORE);

  setTargetDAGCombine(ISD::MUL);

  setTargetDAGCombine({ISD::SELECT, ISD::VSELECT});

  setTargetDAGCombine({ISD::INTRINSIC_VOID, ISD::INTRINSIC_W_CHAIN,
                       ISD::INSERT_VECTOR_ELT, ISD::EXTRACT_VECTOR_ELT,
                       ISD::VECREDUCE_ADD, ISD::STEP_VECTOR});

  setTargetDAGCombine(
      {ISD::MGATHER, ISD::MSCATTER, ISD::EXPERIMENTAL_VECTOR_HISTOGRAM});

  setTargetDAGCombine(ISD::FP_EXTEND);

  setTargetDAGCombine(ISD::GlobalAddress);

  setTargetDAGCombine(ISD::CTLZ);

  setTargetDAGCombine(ISD::GET_ACTIVE_LANE_MASK);

  setTargetDAGCombine(ISD::VECREDUCE_AND);
  setTargetDAGCombine(ISD::VECREDUCE_OR);
  setTargetDAGCombine(ISD::VECREDUCE_XOR);

  setTargetDAGCombine(ISD::SCALAR_TO_VECTOR);

  setTargetDAGCombine(ISD::SHL);

  // In case of strict alignment, avoid an excessive number of byte wide stores.
  MaxStoresPerMemsetOptSize = 8;
  MaxStoresPerMemset =
      Subtarget->requiresStrictAlign() ? MaxStoresPerMemsetOptSize : 32;

  MaxGluedStoresPerMemcpy = 4;
  MaxStoresPerMemcpyOptSize = 4;
  MaxStoresPerMemcpy =
      Subtarget->requiresStrictAlign() ? MaxStoresPerMemcpyOptSize : 16;

  MaxStoresPerMemmoveOptSize = 4;
  MaxStoresPerMemmove =
      Subtarget->requiresStrictAlign() ? MaxStoresPerMemmoveOptSize : 16;

  MaxLoadsPerMemcmpOptSize = 4;
  MaxLoadsPerMemcmp =
      Subtarget->requiresStrictAlign() ? MaxLoadsPerMemcmpOptSize : 8;

  setStackPointerRegisterToSaveRestore(AArch64::SP);

  setSchedulingPreference(Sched::Hybrid);

  EnableExtLdPromotion = true;

  // Set required alignment.
  setMinFunctionAlignment(Align(4));
  // Set preferred alignments.

  // Don't align loops on Windows. The SEH unwind info generation needs to
  // know the exact length of functions before the alignments have been
  // expanded.
  if (!Subtarget->isTargetWindows())
    setPrefLoopAlignment(STI.getPrefLoopAlignment());
  setMaxBytesForAlignment(STI.getMaxBytesForLoopAlignment());
  setPrefFunctionAlignment(STI.getPrefFunctionAlignment());

  // Only change the limit for entries in a jump table if specified by
  // the sub target, but not at the command line.
  unsigned MaxJT = STI.getMaximumJumpTableSize();
  if (MaxJT && getMaximumJumpTableSize() == UINT_MAX)
    setMaximumJumpTableSize(MaxJT);

  setHasExtractBitsInsn(true);

  setMaxDivRemBitWidthSupported(128);

  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
  if (Subtarget->hasSME())
    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i1, Custom);

  if (Subtarget->isNeonAvailable()) {
    // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
    // silliness like this:
    // clang-format off
    for (auto Op :
         {ISD::SELECT,            ISD::SELECT_CC,      ISD::FATAN2,
          ISD::BR_CC,             ISD::FADD,           ISD::FSUB,
          ISD::FMUL,              ISD::FDIV,           ISD::FMA,
          ISD::FNEG,              ISD::FABS,           ISD::FCEIL,
          ISD::FSQRT,             ISD::FFLOOR,         ISD::FNEARBYINT,
          ISD::FSIN,              ISD::FCOS,           ISD::FTAN,
          ISD::FASIN,             ISD::FACOS,          ISD::FATAN,
          ISD::FSINH,             ISD::FCOSH,          ISD::FTANH,
          ISD::FPOW,              ISD::FLOG,           ISD::FLOG2,
          ISD::FLOG10,            ISD::FEXP,           ISD::FEXP2,
          ISD::FEXP10,            ISD::FRINT,          ISD::FROUND,
          ISD::FROUNDEVEN,        ISD::FTRUNC,         ISD::FMINNUM,
          ISD::FMAXNUM,           ISD::FMINIMUM,       ISD::FMAXIMUM,
          ISD::FMAXNUM_IEEE,      ISD::FMINNUM_IEEE,
          ISD::STRICT_FADD,       ISD::STRICT_FSUB,    ISD::STRICT_FMUL,
          ISD::STRICT_FDIV,       ISD::STRICT_FMA,     ISD::STRICT_FCEIL,
          ISD::STRICT_FFLOOR,     ISD::STRICT_FSQRT,   ISD::STRICT_FRINT,
          ISD::STRICT_FNEARBYINT, ISD::STRICT_FROUND,  ISD::STRICT_FTRUNC,
          ISD::STRICT_FROUNDEVEN, ISD::STRICT_FMINNUM, ISD::STRICT_FMAXNUM,
          ISD::STRICT_FMINIMUM,   ISD::STRICT_FMAXIMUM})
      setOperationAction(Op, MVT::v1f64, Expand);
    // clang-format on

    for (auto Op :
         {ISD::FP_TO_SINT, ISD::FP_TO_UINT, ISD::SINT_TO_FP, ISD::UINT_TO_FP,
          ISD::FP_ROUND, ISD::FP_TO_SINT_SAT, ISD::FP_TO_UINT_SAT, ISD::MUL,
          ISD::STRICT_FP_TO_SINT, ISD::STRICT_FP_TO_UINT,
          ISD::STRICT_SINT_TO_FP, ISD::STRICT_UINT_TO_FP, ISD::STRICT_FP_ROUND})
      setOperationAction(Op, MVT::v1i64, Expand);

    // AArch64 doesn't have a direct vector ->f32 conversion instructions for
    // elements smaller than i32, so promote the input to i32 first.
    setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v4i8, MVT::v4i32);
    setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v4i8, MVT::v4i32);

    // Similarly, there is no direct i32 -> f64 vector conversion instruction.
    // Or, direct i32 -> f16 vector conversion.  Set it so custom, so the
    // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
    for (auto Op : {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::STRICT_SINT_TO_FP,
                    ISD::STRICT_UINT_TO_FP})
      for (auto VT : {MVT::v2i32, MVT::v2i64, MVT::v4i32})
        setOperationAction(Op, VT, Custom);

    if (Subtarget->hasFullFP16()) {
      setOperationAction(ISD::ConstantFP, MVT::f16, Legal);
      setOperationAction(ISD::ConstantFP, MVT::bf16, Legal);

      setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Custom);
      setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
      setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Custom);
      setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
      setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
      setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
      setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Custom);
      setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
    } else {
      // when AArch64 doesn't have fullfp16 support, promote the input
      // to i32 first.
      setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v8i8, MVT::v8i32);
      setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v8i8, MVT::v8i32);
      setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v16i8, MVT::v16i32);
      setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v16i8, MVT::v16i32);
      setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v4i16, MVT::v4i32);
      setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v4i16, MVT::v4i32);
      setOperationPromotedToType(ISD::SINT_TO_FP, MVT::v8i16, MVT::v8i32);
      setOperationPromotedToType(ISD::UINT_TO_FP, MVT::v8i16, MVT::v8i32);
    }

    setOperationAction(ISD::CTLZ,       MVT::v1i64, Expand);
    setOperationAction(ISD::CTLZ,       MVT::v2i64, Expand);
    setOperationAction(ISD::BITREVERSE, MVT::v8i8, Legal);
    setOperationAction(ISD::BITREVERSE, MVT::v16i8, Legal);
    setOperationAction(ISD::BITREVERSE, MVT::v2i32, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::v4i32, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::v1i64, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::v2i64, Custom);
    for (auto VT : {MVT::v1i64, MVT::v2i64}) {
      setOperationAction(ISD::UMAX, VT, Custom);
      setOperationAction(ISD::SMAX, VT, Custom);
      setOperationAction(ISD::UMIN, VT, Custom);
      setOperationAction(ISD::SMIN, VT, Custom);
    }

    // Custom handling for some quad-vector types to detect MULL.
    setOperationAction(ISD::MUL, MVT::v8i16, Custom);
    setOperationAction(ISD::MUL, MVT::v4i32, Custom);
    setOperationAction(ISD::MUL, MVT::v2i64, Custom);
    setOperationAction(ISD::MUL, MVT::v4i16, Custom);
    setOperationAction(ISD::MUL, MVT::v2i32, Custom);
    setOperationAction(ISD::MUL, MVT::v1i64, Custom);

    // Saturates
    for (MVT VT : { MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64,
                    MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) {
      setOperationAction(ISD::SADDSAT, VT, Legal);
      setOperationAction(ISD::UADDSAT, VT, Legal);
      setOperationAction(ISD::SSUBSAT, VT, Legal);
      setOperationAction(ISD::USUBSAT, VT, Legal);
    }

    for (MVT VT : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v16i8, MVT::v8i16,
                   MVT::v4i32}) {
      setOperationAction(ISD::AVGFLOORS, VT, Legal);
      setOperationAction(ISD::AVGFLOORU, VT, Legal);
      setOperationAction(ISD::AVGCEILS, VT, Legal);
      setOperationAction(ISD::AVGCEILU, VT, Legal);
      setOperationAction(ISD::ABDS, VT, Legal);
      setOperationAction(ISD::ABDU, VT, Legal);
    }

    // Vector reductions
    for (MVT VT : { MVT::v4f16, MVT::v2f32,
                    MVT::v8f16, MVT::v4f32, MVT::v2f64 }) {
      if (VT.getVectorElementType() != MVT::f16 || Subtarget->hasFullFP16()) {
        setOperationAction(ISD::VECREDUCE_FMAX, VT, Legal);
        setOperationAction(ISD::VECREDUCE_FMIN, VT, Legal);
        setOperationAction(ISD::VECREDUCE_FMAXIMUM, VT, Legal);
        setOperationAction(ISD::VECREDUCE_FMINIMUM, VT, Legal);

        setOperationAction(ISD::VECREDUCE_FADD, VT, Legal);
      }
    }
    if (Subtarget->hasFullFP16())
      setOperationAction(ISD::VECREDUCE_FADD, MVT::v2f16, Custom);

    for (MVT VT : { MVT::v8i8, MVT::v4i16, MVT::v2i32,
                    MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
      setOperationAction(ISD::VECREDUCE_ADD, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
    }
    setOperationAction(ISD::VECREDUCE_ADD, MVT::v2i64, Custom);
    setOperationAction(ISD::VECREDUCE_AND, MVT::v2i64, Custom);
    setOperationAction(ISD::VECREDUCE_OR, MVT::v2i64, Custom);
    setOperationAction(ISD::VECREDUCE_XOR, MVT::v2i64, Custom);

    setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
    setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
    // Likewise, narrowing and extending vector loads/stores aren't handled
    // directly.
    for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
      setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);

      if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32) {
        setOperationAction(ISD::MULHS, VT, Legal);
        setOperationAction(ISD::MULHU, VT, Legal);
      } else {
        setOperationAction(ISD::MULHS, VT, Expand);
        setOperationAction(ISD::MULHU, VT, Expand);
      }
      setOperationAction(ISD::SMUL_LOHI, VT, Expand);
      setOperationAction(ISD::UMUL_LOHI, VT, Expand);

      setOperationAction(ISD::BSWAP, VT, Expand);
      setOperationAction(ISD::CTTZ, VT, Expand);

      for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
        setTruncStoreAction(VT, InnerVT, Expand);
        setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
      }
    }

    for (auto Op :
         {ISD::FFLOOR, ISD::FNEARBYINT, ISD::FCEIL, ISD::FRINT, ISD::FTRUNC,
          ISD::FROUND, ISD::FROUNDEVEN, ISD::FMAXNUM_IEEE, ISD::FMINNUM_IEEE,
          ISD::STRICT_FFLOOR, ISD::STRICT_FNEARBYINT, ISD::STRICT_FCEIL,
          ISD::STRICT_FRINT, ISD::STRICT_FTRUNC, ISD::STRICT_FROUND,
          ISD::STRICT_FROUNDEVEN}) {
      for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64})
        setOperationAction(Op, Ty, Legal);
      if (Subtarget->hasFullFP16())
        for (MVT Ty : {MVT::v4f16, MVT::v8f16})
          setOperationAction(Op, Ty, Legal);
    }

    // LRINT and LLRINT.
    for (auto Op : {ISD::LRINT, ISD::LLRINT}) {
      for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64})
        setOperationAction(Op, Ty, Custom);
      if (Subtarget->hasFullFP16())
        for (MVT Ty : {MVT::v4f16, MVT::v8f16})
          setOperationAction(Op, Ty, Custom);
    }

    setTruncStoreAction(MVT::v4i16, MVT::v4i8, Custom);

    setOperationAction(ISD::BITCAST, MVT::i2, Custom);
    setOperationAction(ISD::BITCAST, MVT::i4, Custom);
    setOperationAction(ISD::BITCAST, MVT::i8, Custom);
    setOperationAction(ISD::BITCAST, MVT::i16, Custom);

    setOperationAction(ISD::BITCAST, MVT::v2i8, Custom);
    setOperationAction(ISD::BITCAST, MVT::v2i16, Custom);
    setOperationAction(ISD::BITCAST, MVT::v4i8, Custom);

    setLoadExtAction(ISD::EXTLOAD,  MVT::v4i16, MVT::v4i8, Custom);
    setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, MVT::v4i8, Custom);
    setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i16, MVT::v4i8, Custom);
    setLoadExtAction(ISD::EXTLOAD,  MVT::v4i32, MVT::v4i8, Custom);
    setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Custom);
    setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Custom);

    // ADDP custom lowering
    for (MVT VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 })
      setOperationAction(ISD::ADD, VT, Custom);
    // FADDP custom lowering
    for (MVT VT : { MVT::v16f16, MVT::v8f32, MVT::v4f64 })
      setOperationAction(ISD::FADD, VT, Custom);

    if (Subtarget->hasDotProd()) {
      static const unsigned MLAOps[] = {ISD::PARTIAL_REDUCE_SMLA,
                                        ISD::PARTIAL_REDUCE_UMLA};

      setPartialReduceMLAAction(MLAOps, MVT::v4i32, MVT::v16i8, Legal);
      setPartialReduceMLAAction(MLAOps, MVT::v2i32, MVT::v8i8, Legal);
      setPartialReduceMLAAction(MLAOps, MVT::v2i64, MVT::v16i8, Custom);

      if (Subtarget->hasMatMulInt8()) {
        setPartialReduceMLAAction(ISD::PARTIAL_REDUCE_SUMLA, MVT::v4i32,
                                  MVT::v16i8, Legal);
        setPartialReduceMLAAction(ISD::PARTIAL_REDUCE_SUMLA, MVT::v2i64,
                                  MVT::v16i8, Custom);

        setPartialReduceMLAAction(ISD::PARTIAL_REDUCE_SUMLA, MVT::v2i32,
                                  MVT::v8i8, Legal);
      }
    }

  } else /* !isNeonAvailable */ {
    for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
      for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
        setOperationAction(Op, VT, Expand);

      if (VT.is128BitVector() || VT.is64BitVector()) {
        setOperationAction(ISD::LOAD, VT, Legal);
        setOperationAction(ISD::STORE, VT, Legal);
        setOperationAction(ISD::BITCAST, VT,
                           Subtarget->isLittleEndian() ? Legal : Expand);
      }
      for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
        setTruncStoreAction(VT, InnerVT, Expand);
        setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
      }
    }
  }

  for (MVT VT : {MVT::v8i16, MVT::v4i32, MVT::v2i64}) {
    setOperationAction(ISD::TRUNCATE_SSAT_S, VT, Legal);
    setOperationAction(ISD::TRUNCATE_SSAT_U, VT, Legal);
    setOperationAction(ISD::TRUNCATE_USAT_U, VT, Legal);
  }

  if (Subtarget->hasSME()) {
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
  }

  // FIXME: Move lowering for more nodes here if those are common between
  // SVE and SME.
  if (Subtarget->isSVEorStreamingSVEAvailable()) {
    for (auto VT :
         {MVT::nxv16i1, MVT::nxv8i1, MVT::nxv4i1, MVT::nxv2i1, MVT::nxv1i1}) {
      setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
    }
    for (auto VT : {MVT::nxv16i1, MVT::nxv8i1, MVT::nxv4i1, MVT::nxv2i1}) {
      setOperationAction(ISD::VECTOR_FIND_LAST_ACTIVE, VT, Legal);
      setOperationAction(ISD::GET_ACTIVE_LANE_MASK, VT, Legal);
    }

    if (Subtarget->hasSVE2p1() ||
        (Subtarget->hasSME2() && Subtarget->isStreaming()))
      setOperationAction(ISD::GET_ACTIVE_LANE_MASK, MVT::nxv32i1, Custom);

    for (auto VT : {MVT::v16i8, MVT::v8i8, MVT::v4i16, MVT::v2i32})
      setOperationAction(ISD::GET_ACTIVE_LANE_MASK, VT, Custom);
  }

  if (Subtarget->isSVEorStreamingSVEAvailable()) {
    for (auto VT : {MVT::nxv16i8, MVT::nxv8i16, MVT::nxv4i32, MVT::nxv2i64}) {
      setOperationAction(ISD::BITREVERSE, VT, Custom);
      setOperationAction(ISD::BSWAP, VT, Custom);
      setOperationAction(ISD::CTLZ, VT, Custom);
      setOperationAction(ISD::CTPOP, VT, Custom);
      setOperationAction(ISD::CTTZ, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::UINT_TO_FP, VT, Custom);
      setOperationAction(ISD::SINT_TO_FP, VT, Custom);
      setOperationAction(ISD::FP_TO_UINT, VT, Custom);
      setOperationAction(ISD::FP_TO_SINT, VT, Custom);
      setOperationAction(ISD::MLOAD, VT, Custom);
      setOperationAction(ISD::MUL, VT, Custom);
      setOperationAction(ISD::MULHS, VT, Custom);
      setOperationAction(ISD::MULHU, VT, Custom);
      setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
      setOperationAction(ISD::VECTOR_SPLICE, VT, Custom);
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SETCC, VT, Custom);
      setOperationAction(ISD::SDIV, VT, Custom);
      setOperationAction(ISD::UDIV, VT, Custom);
      setOperationAction(ISD::SMIN, VT, Custom);
      setOperationAction(ISD::UMIN, VT, Custom);
      setOperationAction(ISD::SMAX, VT, Custom);
      setOperationAction(ISD::UMAX, VT, Custom);
      setOperationAction(ISD::SHL, VT, Custom);
      setOperationAction(ISD::SRL, VT, Custom);
      setOperationAction(ISD::SRA, VT, Custom);
      setOperationAction(ISD::ABS, VT, Custom);
      setOperationAction(ISD::ABDS, VT, Custom);
      setOperationAction(ISD::ABDU, VT, Custom);
      setOperationAction(ISD::VECREDUCE_ADD, VT, Custom);
      setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom);
      setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);

      setOperationAction(ISD::UMUL_LOHI, VT, Expand);
      setOperationAction(ISD::SMUL_LOHI, VT, Expand);
      setOperationAction(ISD::SELECT_CC, VT, Expand);
      setOperationAction(ISD::ROTL, VT, Expand);
      setOperationAction(ISD::ROTR, VT, Expand);

      setOperationAction(ISD::SADDSAT, VT, Legal);
      setOperationAction(ISD::UADDSAT, VT, Legal);
      setOperationAction(ISD::SSUBSAT, VT, Legal);
      setOperationAction(ISD::USUBSAT, VT, Legal);
      setOperationAction(ISD::UREM, VT, Expand);
      setOperationAction(ISD::SREM, VT, Expand);
      setOperationAction(ISD::SDIVREM, VT, Expand);
      setOperationAction(ISD::UDIVREM, VT, Expand);

      setOperationAction(ISD::AVGFLOORS, VT, Custom);
      setOperationAction(ISD::AVGFLOORU, VT, Custom);
      setOperationAction(ISD::AVGCEILS, VT, Custom);
      setOperationAction(ISD::AVGCEILU, VT, Custom);

      if (!Subtarget->isLittleEndian())
        setOperationAction(ISD::BITCAST, VT, Custom);

      if (Subtarget->hasSVE2() ||
          (Subtarget->hasSME() && Subtarget->isStreaming()))
        // For SLI/SRI.
        setOperationAction(ISD::OR, VT, Custom);
    }

    // Illegal unpacked integer vector types.
    for (auto VT : {MVT::nxv8i8, MVT::nxv4i16, MVT::nxv2i32}) {
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
    }

    // Type legalize unpacked bitcasts.
    for (auto VT : {MVT::nxv2i16, MVT::nxv4i16, MVT::nxv2i32})
      setOperationAction(ISD::BITCAST, VT, Custom);

    for (auto VT :
         { MVT::nxv2i8, MVT::nxv2i16, MVT::nxv2i32, MVT::nxv2i64, MVT::nxv4i8,
           MVT::nxv4i16, MVT::nxv4i32, MVT::nxv8i8, MVT::nxv8i16 })
      setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Legal);

    for (auto VT :
         {MVT::nxv16i1, MVT::nxv8i1, MVT::nxv4i1, MVT::nxv2i1, MVT::nxv1i1}) {
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SETCC, VT, Custom);
      setOperationAction(ISD::TRUNCATE, VT, Custom);
      setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);

      setOperationAction(ISD::SELECT_CC, VT, Expand);
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);

      // There are no legal MVT::nxv16f## based types.
      if (VT != MVT::nxv16i1) {
        setOperationAction(ISD::FP_TO_SINT, VT, Custom);
        setOperationAction(ISD::FP_TO_UINT, VT, Custom);
        setOperationAction(ISD::SINT_TO_FP, VT, Custom);
        setOperationAction(ISD::UINT_TO_FP, VT, Custom);
      }
    }

    // NEON doesn't support masked loads/stores, but SME and SVE do.
    for (auto VT :
         {MVT::v4f16, MVT::v8f16, MVT::v2f32, MVT::v4f32, MVT::v1f64,
          MVT::v2f64, MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16,
          MVT::v2i32, MVT::v4i32, MVT::v1i64, MVT::v2i64}) {
      setOperationAction(ISD::MLOAD, VT, Custom);
      setOperationAction(ISD::MSTORE, VT, Custom);
    }

    // Firstly, exclude all scalable vector extending loads/truncating stores,
    // include both integer and floating scalable vector.
    for (MVT VT : MVT::scalable_vector_valuetypes()) {
      for (MVT InnerVT : MVT::scalable_vector_valuetypes()) {
        setTruncStoreAction(VT, InnerVT, Expand);
        setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
      }
    }

    // Then, selectively enable those which we directly support.
    setTruncStoreAction(MVT::nxv2i64, MVT::nxv2i8, Legal);
    setTruncStoreAction(MVT::nxv2i64, MVT::nxv2i16, Legal);
    setTruncStoreAction(MVT::nxv2i64, MVT::nxv2i32, Legal);
    setTruncStoreAction(MVT::nxv4i32, MVT::nxv4i8, Legal);
    setTruncStoreAction(MVT::nxv4i32, MVT::nxv4i16, Legal);
    setTruncStoreAction(MVT::nxv8i16, MVT::nxv8i8, Legal);
    for (auto Op : {ISD::ZEXTLOAD, ISD::SEXTLOAD, ISD::EXTLOAD}) {
      setLoadExtAction(Op, MVT::nxv2i64, MVT::nxv2i8, Legal);
      setLoadExtAction(Op, MVT::nxv2i64, MVT::nxv2i16, Legal);
      setLoadExtAction(Op, MVT::nxv2i64, MVT::nxv2i32, Legal);
      setLoadExtAction(Op, MVT::nxv4i32, MVT::nxv4i8, Legal);
      setLoadExtAction(Op, MVT::nxv4i32, MVT::nxv4i16, Legal);
      setLoadExtAction(Op, MVT::nxv8i16, MVT::nxv8i8, Legal);
    }

    // SVE supports truncating stores of 64 and 128-bit vectors
    setTruncStoreAction(MVT::v2i64, MVT::v2i8, Custom);
    setTruncStoreAction(MVT::v2i64, MVT::v2i16, Custom);
    setTruncStoreAction(MVT::v2i64, MVT::v2i32, Custom);
    setTruncStoreAction(MVT::v2i32, MVT::v2i8, Custom);
    setTruncStoreAction(MVT::v2i32, MVT::v2i16, Custom);

    for (auto VT : {MVT::nxv2f16, MVT::nxv4f16, MVT::nxv8f16, MVT::nxv2f32,
                    MVT::nxv4f32, MVT::nxv2f64}) {
      setOperationAction(ISD::BITCAST, VT, Custom);
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::MLOAD, VT, Custom);
      setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SETCC, VT, Custom);
      setOperationAction(ISD::FADD, VT, Custom);
      setOperationAction(ISD::FCOPYSIGN, VT, Custom);
      setOperationAction(ISD::FDIV, VT, Custom);
      setOperationAction(ISD::FMA, VT, Custom);
      setOperationAction(ISD::FMAXIMUM, VT, Custom);
      setOperationAction(ISD::FMAXNUM, VT, Custom);
      setOperationAction(ISD::FMINIMUM, VT, Custom);
      setOperationAction(ISD::FMINNUM, VT, Custom);
      setOperationAction(ISD::FMUL, VT, Custom);
      setOperationAction(ISD::FNEG, VT, Custom);
      setOperationAction(ISD::FSUB, VT, Custom);
      setOperationAction(ISD::FCEIL, VT, Custom);
      setOperationAction(ISD::FFLOOR, VT, Custom);
      setOperationAction(ISD::FNEARBYINT, VT, Custom);
      setOperationAction(ISD::FRINT, VT, Custom);
      setOperationAction(ISD::LRINT, VT, Custom);
      setOperationAction(ISD::LLRINT, VT, Custom);
      setOperationAction(ISD::FROUND, VT, Custom);
      setOperationAction(ISD::FROUNDEVEN, VT, Custom);
      setOperationAction(ISD::FTRUNC, VT, Custom);
      setOperationAction(ISD::FSQRT, VT, Custom);
      setOperationAction(ISD::FABS, VT, Custom);
      setOperationAction(ISD::FP_EXTEND, VT, Custom);
      setOperationAction(ISD::FP_ROUND, VT, Custom);
      setOperationAction(ISD::VECREDUCE_FADD, VT, Custom);
      setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_FMAXIMUM, VT, Custom);
      setOperationAction(ISD::VECREDUCE_FMINIMUM, VT, Custom);
      setOperationAction(ISD::VECTOR_SPLICE, VT, Custom);
      setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);

      setOperationAction(ISD::SELECT_CC, VT, Expand);
      setOperationAction(ISD::FREM, VT, Expand);
      setOperationAction(ISD::FPOW, VT, Expand);
      setOperationAction(ISD::FPOWI, VT, Expand);
      setOperationAction(ISD::FCOS, VT, Expand);
      setOperationAction(ISD::FSIN, VT, Expand);
      setOperationAction(ISD::FSINCOS, VT, Expand);
      setOperationAction(ISD::FTAN, VT, Expand);
      setOperationAction(ISD::FACOS, VT, Expand);
      setOperationAction(ISD::FASIN, VT, Expand);
      setOperationAction(ISD::FATAN, VT, Expand);
      setOperationAction(ISD::FATAN2, VT, Expand);
      setOperationAction(ISD::FCOSH, VT, Expand);
      setOperationAction(ISD::FSINH, VT, Expand);
      setOperationAction(ISD::FTANH, VT, Expand);
      setOperationAction(ISD::FEXP, VT, Expand);
      setOperationAction(ISD::FEXP2, VT, Expand);
      setOperationAction(ISD::FEXP10, VT, Expand);
      setOperationAction(ISD::FLOG, VT, Expand);
      setOperationAction(ISD::FLOG2, VT, Expand);
      setOperationAction(ISD::FLOG10, VT, Expand);

      setCondCodeAction(ISD::SETO, VT, Expand);
      setCondCodeAction(ISD::SETOLT, VT, Expand);
      setCondCodeAction(ISD::SETLT, VT, Expand);
      setCondCodeAction(ISD::SETOLE, VT, Expand);
      setCondCodeAction(ISD::SETLE, VT, Expand);
      setCondCodeAction(ISD::SETULT, VT, Expand);
      setCondCodeAction(ISD::SETULE, VT, Expand);
      setCondCodeAction(ISD::SETUGE, VT, Expand);
      setCondCodeAction(ISD::SETUGT, VT, Expand);
      setCondCodeAction(ISD::SETUEQ, VT, Expand);
      setCondCodeAction(ISD::SETONE, VT, Expand);
    }

    for (auto VT : {MVT::nxv2bf16, MVT::nxv4bf16, MVT::nxv8bf16}) {
      setOperationAction(ISD::BITCAST, VT, Custom);
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
      setOperationAction(ISD::FABS, VT, Custom);
      setOperationAction(ISD::FCOPYSIGN, VT, Custom);
      setOperationAction(ISD::FNEG, VT, Custom);
      setOperationAction(ISD::FP_EXTEND, VT, Custom);
      setOperationAction(ISD::FP_ROUND, VT, Custom);
      setOperationAction(ISD::MLOAD, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::SELECT, VT, Custom);
      setOperationAction(ISD::SELECT_CC, VT, Expand);
      setOperationAction(ISD::SPLAT_VECTOR, VT, Legal);
      setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
      setOperationAction(ISD::VECTOR_SPLICE, VT, Custom);

      if (Subtarget->hasSVEB16B16()) {
        setOperationAction(ISD::FADD, VT, Legal);
        setOperationAction(ISD::FMA, VT, Custom);
        setOperationAction(ISD::FMAXIMUM, VT, Custom);
        setOperationAction(ISD::FMAXNUM, VT, Custom);
        setOperationAction(ISD::FMINIMUM, VT, Custom);
        setOperationAction(ISD::FMINNUM, VT, Custom);
        setOperationAction(ISD::FMUL, VT, Legal);
        setOperationAction(ISD::FSUB, VT, Legal);
      }
    }

    for (auto Opcode :
         {ISD::FCEIL, ISD::FDIV, ISD::FFLOOR, ISD::FNEARBYINT, ISD::FRINT,
          ISD::FROUND, ISD::FROUNDEVEN, ISD::FSQRT, ISD::FTRUNC, ISD::SETCC,
          ISD::VECREDUCE_FADD, ISD::VECREDUCE_FMAX, ISD::VECREDUCE_FMAXIMUM,
          ISD::VECREDUCE_FMIN, ISD::VECREDUCE_FMINIMUM}) {
      setOperationPromotedToType(Opcode, MVT::nxv2bf16, MVT::nxv2f32);
      setOperationPromotedToType(Opcode, MVT::nxv4bf16, MVT::nxv4f32);
      setOperationPromotedToType(Opcode, MVT::nxv8bf16, MVT::nxv8f32);
    }

    if (!Subtarget->hasSVEB16B16()) {
      for (auto Opcode : {ISD::FADD, ISD::FMA, ISD::FMAXIMUM, ISD::FMAXNUM,
                          ISD::FMINIMUM, ISD::FMINNUM, ISD::FMUL, ISD::FSUB}) {
        setOperationPromotedToType(Opcode, MVT::nxv2bf16, MVT::nxv2f32);
        setOperationPromotedToType(Opcode, MVT::nxv4bf16, MVT::nxv4f32);
        setOperationPromotedToType(Opcode, MVT::nxv8bf16, MVT::nxv8f32);
      }
    }

    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i8, Custom);
    setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i16, Custom);

    // NEON doesn't support integer divides, but SVE does
    for (auto VT : {MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16, MVT::v2i32,
                    MVT::v4i32, MVT::v1i64, MVT::v2i64}) {
      setOperationAction(ISD::SDIV, VT, Custom);
      setOperationAction(ISD::UDIV, VT, Custom);
    }

    // NEON doesn't support 64-bit vector integer muls, but SVE does.
    setOperationAction(ISD::MUL, MVT::v1i64, Custom);
    setOperationAction(ISD::MUL, MVT::v2i64, Custom);

    // NOTE: Currently this has to happen after computeRegisterProperties rather
    // than the preferred option of combining it with the addRegisterClass call.
    if (Subtarget->useSVEForFixedLengthVectors()) {
      for (MVT VT : MVT::integer_fixedlen_vector_valuetypes()) {
        if (useSVEForFixedLengthVectorVT(
                VT, /*OverrideNEON=*/!Subtarget->isNeonAvailable()))
          addTypeForFixedLengthSVE(VT);
      }
      for (MVT VT : MVT::fp_fixedlen_vector_valuetypes()) {
        if (useSVEForFixedLengthVectorVT(
                VT, /*OverrideNEON=*/!Subtarget->isNeonAvailable()))
          addTypeForFixedLengthSVE(VT);
      }

      // 64bit results can mean a bigger than NEON input.
      for (auto VT : {MVT::v8i8, MVT::v4i16})
        setOperationAction(ISD::TRUNCATE, VT, Custom);
      setOperationAction(ISD::FP_ROUND, MVT::v4f16, Custom);

      // 128bit results imply a bigger than NEON input.
      for (auto VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
        setOperationAction(ISD::TRUNCATE, VT, Custom);
      for (auto VT : {MVT::v8f16, MVT::v4f32})
        setOperationAction(ISD::FP_ROUND, VT, Custom);

      // These operations are not supported on NEON but SVE can do them.
      setOperationAction(ISD::BITREVERSE, MVT::v1i64, Custom);
      setOperationAction(ISD::CTLZ, MVT::v1i64, Custom);
      setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
      setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
      setOperationAction(ISD::MULHS, MVT::v1i64, Custom);
      setOperationAction(ISD::MULHS, MVT::v2i64, Custom);
      setOperationAction(ISD::MULHU, MVT::v1i64, Custom);
      setOperationAction(ISD::MULHU, MVT::v2i64, Custom);
      setOperationAction(ISD::SMAX, MVT::v1i64, Custom);
      setOperationAction(ISD::SMAX, MVT::v2i64, Custom);
      setOperationAction(ISD::SMIN, MVT::v1i64, Custom);
      setOperationAction(ISD::SMIN, MVT::v2i64, Custom);
      setOperationAction(ISD::UMAX, MVT::v1i64, Custom);
      setOperationAction(ISD::UMAX, MVT::v2i64, Custom);
      setOperationAction(ISD::UMIN, MVT::v1i64, Custom);
      setOperationAction(ISD::UMIN, MVT::v2i64, Custom);
      setOperationAction(ISD::VECREDUCE_SMAX, MVT::v2i64, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN, MVT::v2i64, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX, MVT::v2i64, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN, MVT::v2i64, Custom);

      // Int operations with no NEON support.
      for (auto VT : {MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16,
                      MVT::v2i32, MVT::v4i32, MVT::v2i64}) {
        setOperationAction(ISD::BITREVERSE, VT, Custom);
        setOperationAction(ISD::CTTZ, VT, Custom);
        setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
        setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
        setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
        setOperationAction(ISD::MULHS, VT, Custom);
        setOperationAction(ISD::MULHU, VT, Custom);
      }

      // Use SVE for vectors with more than 2 elements.
      for (auto VT : {MVT::v4f16, MVT::v8f16, MVT::v4f32})
        setOperationAction(ISD::VECREDUCE_FADD, VT, Custom);
    }

    setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv2i1, MVT::nxv2i64);
    setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv4i1, MVT::nxv4i32);
    setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv8i1, MVT::nxv8i16);
    setOperationPromotedToType(ISD::VECTOR_SPLICE, MVT::nxv16i1, MVT::nxv16i8);

    setOperationAction(ISD::VSCALE, MVT::i32, Custom);

    for (auto VT : {MVT::v16i1, MVT::v8i1, MVT::v4i1, MVT::v2i1})
      setOperationAction(ISD::INTRINSIC_WO_CHAIN, VT, Custom);
  }

  // Handle partial reduction operations
  if (Subtarget->isSVEorStreamingSVEAvailable()) {
    // Mark known legal pairs as 'Legal' (these will expand to UDOT or SDOT).
    // Other pairs will default to 'Expand'.
    static const unsigned MLAOps[] = {ISD::PARTIAL_REDUCE_SMLA,
                                      ISD::PARTIAL_REDUCE_UMLA};
    setPartialReduceMLAAction(MLAOps, MVT::nxv2i64, MVT::nxv8i16, Legal);
    setPartialReduceMLAAction(MLAOps, MVT::nxv4i32, MVT::nxv16i8, Legal);

    setPartialReduceMLAAction(MLAOps, MVT::nxv2i64, MVT::nxv16i8, Custom);

    if (Subtarget->hasMatMulInt8()) {
      setPartialReduceMLAAction(ISD::PARTIAL_REDUCE_SUMLA, MVT::nxv4i32,
                                MVT::nxv16i8, Legal);
      setPartialReduceMLAAction(ISD::PARTIAL_REDUCE_SUMLA, MVT::nxv2i64,
                                MVT::nxv16i8, Custom);
    }

    // Wide add types
    if (Subtarget->hasSVE2() || Subtarget->hasSME()) {
      setPartialReduceMLAAction(MLAOps, MVT::nxv2i64, MVT::nxv4i32, Legal);
      setPartialReduceMLAAction(MLAOps, MVT::nxv4i32, MVT::nxv8i16, Legal);
      setPartialReduceMLAAction(MLAOps, MVT::nxv8i16, MVT::nxv16i8, Legal);
    }
  }

  // Handle operations that are only available in non-streaming SVE mode.
  if (Subtarget->isSVEAvailable()) {
    for (auto VT : {MVT::nxv16i8,  MVT::nxv8i16, MVT::nxv4i32,  MVT::nxv2i64,
                    MVT::nxv2f16,  MVT::nxv4f16, MVT::nxv8f16,  MVT::nxv2f32,
                    MVT::nxv4f32,  MVT::nxv2f64, MVT::nxv2bf16, MVT::nxv4bf16,
                    MVT::nxv8bf16, MVT::v4f16,   MVT::v8f16,    MVT::v2f32,
                    MVT::v4f32,    MVT::v1f64,   MVT::v2f64,    MVT::v8i8,
                    MVT::v16i8,    MVT::v4i16,   MVT::v8i16,    MVT::v2i32,
                    MVT::v4i32,    MVT::v1i64,   MVT::v2i64}) {
      setOperationAction(ISD::MGATHER, VT, Custom);
      setOperationAction(ISD::MSCATTER, VT, Custom);
    }

    for (auto VT : {MVT::nxv2f16, MVT::nxv4f16, MVT::nxv8f16, MVT::nxv2f32,
                    MVT::nxv4f32, MVT::nxv2f64, MVT::v4f16, MVT::v8f16,
                    MVT::v2f32, MVT::v4f32, MVT::v2f64})
      setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom);

    // We can lower types that have <vscale x {2|4}> elements to compact.
    for (auto VT :
         {MVT::nxv2i8, MVT::nxv2i16, MVT::nxv2i32, MVT::nxv2i64, MVT::nxv2f32,
          MVT::nxv2f64, MVT::nxv4i8, MVT::nxv4i16, MVT::nxv4i32, MVT::nxv4f32})
      setOperationAction(ISD::VECTOR_COMPRESS, VT, Custom);

    // If we have SVE, we can use SVE logic for legal (or smaller than legal)
    // NEON vectors in the lowest bits of the SVE register.
    for (auto VT : {MVT::v2i8, MVT::v2i16, MVT::v2i32, MVT::v2i64, MVT::v2f32,
                    MVT::v2f64, MVT::v4i8, MVT::v4i16, MVT::v4i32, MVT::v4f32})
      setOperationAction(ISD::VECTOR_COMPRESS, VT, Custom);

    // Histcnt is SVE2 only
    if (Subtarget->hasSVE2()) {
      setOperationAction(ISD::EXPERIMENTAL_VECTOR_HISTOGRAM, MVT::nxv4i32,
                         Custom);
      setOperationAction(ISD::EXPERIMENTAL_VECTOR_HISTOGRAM, MVT::nxv2i64,
                         Custom);

      static const unsigned MLAOps[] = {ISD::PARTIAL_REDUCE_SMLA,
                                        ISD::PARTIAL_REDUCE_UMLA};
      // Must be lowered to SVE instructions.
      setPartialReduceMLAAction(MLAOps, MVT::v2i64, MVT::v4i32, Custom);
      setPartialReduceMLAAction(MLAOps, MVT::v2i64, MVT::v8i16, Custom);
      setPartialReduceMLAAction(MLAOps, MVT::v2i64, MVT::v16i8, Custom);
      setPartialReduceMLAAction(MLAOps, MVT::v4i32, MVT::v8i16, Custom);
      setPartialReduceMLAAction(MLAOps, MVT::v4i32, MVT::v16i8, Custom);
      setPartialReduceMLAAction(MLAOps, MVT::v8i16, MVT::v16i8, Custom);
    }
  }

  if (Subtarget->hasMOPS() && Subtarget->hasMTE()) {
    // Only required for llvm.aarch64.mops.memset.tag
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i8, Custom);
  }

  setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);

  if (Subtarget->hasSVE()) {
    setOperationAction(ISD::FLDEXP, MVT::f64, Custom);
    setOperationAction(ISD::FLDEXP, MVT::f32, Custom);
    setOperationAction(ISD::FLDEXP, MVT::f16, Custom);
    setOperationAction(ISD::FLDEXP, MVT::bf16, Custom);
  }

  PredictableSelectIsExpensive = Subtarget->predictableSelectIsExpensive();

  IsStrictFPEnabled = true;
  setMaxAtomicSizeInBitsSupported(128);

  // On MSVC, both 32-bit and 64-bit, ldexpf(f32) is not defined.  MinGW has
  // it, but it's just a wrapper around ldexp.
  if (Subtarget->isTargetWindows()) {
    for (ISD::NodeType Op : {ISD::FLDEXP, ISD::STRICT_FLDEXP, ISD::FFREXP})
      if (isOperationExpand(Op, MVT::f32))
        setOperationAction(Op, MVT::f32, Promote);
  }

  // LegalizeDAG currently can't expand fp16 LDEXP/FREXP on targets where i16
  // isn't legal.
  for (ISD::NodeType Op : {ISD::FLDEXP, ISD::STRICT_FLDEXP, ISD::FFREXP})
    if (isOperationExpand(Op, MVT::f16))
      setOperationAction(Op, MVT::f16, Promote);
}

void AArch64TargetLowering::addTypeForNEON(MVT VT) {
  assert(VT.isVector() && "VT should be a vector type");

  if (VT.isFloatingPoint()) {
    MVT PromoteTo = EVT(VT).changeVectorElementTypeToInteger().getSimpleVT();
    setOperationPromotedToType(ISD::LOAD, VT, PromoteTo);
    setOperationPromotedToType(ISD::STORE, VT, PromoteTo);
  }

  // Mark vector float intrinsics as expand.
  if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
    setOperationAction(ISD::FSIN, VT, Expand);
    setOperationAction(ISD::FCOS, VT, Expand);
    setOperationAction(ISD::FTAN, VT, Expand);
    setOperationAction(ISD::FASIN, VT, Expand);
    setOperationAction(ISD::FACOS, VT, Expand);
    setOperationAction(ISD::FATAN, VT, Expand);
    setOperationAction(ISD::FATAN2, VT, Expand);
    setOperationAction(ISD::FSINH, VT, Expand);
    setOperationAction(ISD::FCOSH, VT, Expand);
    setOperationAction(ISD::FTANH, VT, Expand);
    setOperationAction(ISD::FPOW, VT, Expand);
    setOperationAction(ISD::FLOG, VT, Expand);
    setOperationAction(ISD::FLOG2, VT, Expand);
    setOperationAction(ISD::FLOG10, VT, Expand);
    setOperationAction(ISD::FEXP, VT, Expand);
    setOperationAction(ISD::FEXP2, VT, Expand);
    setOperationAction(ISD::FEXP10, VT, Expand);
  }

  // But we do support custom-lowering for FCOPYSIGN.
  if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
      ((VT == MVT::v4bf16 || VT == MVT::v8bf16 || VT == MVT::v4f16 ||
        VT == MVT::v8f16) &&
       Subtarget->hasFullFP16()))
    setOperationAction(ISD::FCOPYSIGN, VT, Custom);

  setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
  setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
  setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom);
  setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
  setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
  setOperationAction(ISD::SRA, VT, Custom);
  setOperationAction(ISD::SRL, VT, Custom);
  setOperationAction(ISD::SHL, VT, Custom);
  setOperationAction(ISD::OR, VT, Custom);
  setOperationAction(ISD::SETCC, VT, Custom);
  setOperationAction(ISD::CONCAT_VECTORS, VT, Legal);

  setOperationAction(ISD::SELECT, VT, Expand);
  setOperationAction(ISD::SELECT_CC, VT, Expand);
  setOperationAction(ISD::VSELECT, VT, Expand);
  for (MVT InnerVT : MVT::all_valuetypes())
    setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);

  // CNT supports only B element sizes, then use UADDLP to widen.
  if (VT != MVT::v8i8 && VT != MVT::v16i8)
    setOperationAction(ISD::CTPOP, VT, Custom);

  setOperationAction(ISD::UDIV, VT, Expand);
  setOperationAction(ISD::SDIV, VT, Expand);
  setOperationAction(ISD::UREM, VT, Expand);
  setOperationAction(ISD::SREM, VT, Expand);
  setOperationAction(ISD::FREM, VT, Expand);

  for (unsigned Opcode :
       {ISD::FP_TO_SINT, ISD::FP_TO_UINT, ISD::FP_TO_SINT_SAT,
        ISD::FP_TO_UINT_SAT, ISD::STRICT_FP_TO_SINT, ISD::STRICT_FP_TO_UINT})
    setOperationAction(Opcode, VT, Custom);

  if (!VT.isFloatingPoint())
    setOperationAction(ISD::ABS, VT, Legal);

  // [SU][MIN|MAX] are available for all NEON types apart from i64.
  if (!VT.isFloatingPoint() && VT != MVT::v2i64 && VT != MVT::v1i64)
    for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
      setOperationAction(Opcode, VT, Legal);

  // F[MIN|MAX][NUM|NAN] and simple strict operations are available for all FP
  // NEON types.
  if (VT.isFloatingPoint() &&
      VT.getVectorElementType() != MVT::bf16 &&
      (VT.getVectorElementType() != MVT::f16 || Subtarget->hasFullFP16()))
    for (unsigned Opcode :
         {ISD::FMINIMUM, ISD::FMAXIMUM, ISD::FMINNUM, ISD::FMAXNUM,
          ISD::FMINNUM_IEEE, ISD::FMAXNUM_IEEE, ISD::STRICT_FMINIMUM,
          ISD::STRICT_FMAXIMUM, ISD::STRICT_FMINNUM, ISD::STRICT_FMAXNUM,
          ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
          ISD::STRICT_FDIV, ISD::STRICT_FMA, ISD::STRICT_FSQRT})
      setOperationAction(Opcode, VT, Legal);

  // Strict fp extend and trunc are legal
  if (VT.isFloatingPoint() && VT.getScalarSizeInBits() != 16)
    setOperationAction(ISD::STRICT_FP_EXTEND, VT, Legal);
  if (VT.isFloatingPoint() && VT.getScalarSizeInBits() != 64)
    setOperationAction(ISD::STRICT_FP_ROUND, VT, Legal);

  // FIXME: We could potentially make use of the vector comparison instructions
  // for STRICT_FSETCC and STRICT_FSETCSS, but there's a number of
  // complications:
  //  * FCMPEQ/NE are quiet comparisons, the rest are signalling comparisons,
  //    so we would need to expand when the condition code doesn't match the
  //    kind of comparison.
  //  * Some kinds of comparison require more than one FCMXY instruction so
  //    would need to be expanded instead.
  //  * The lowering of the non-strict versions involves target-specific ISD
  //    nodes so we would likely need to add strict versions of all of them and
  //    handle them appropriately.
  setOperationAction(ISD::STRICT_FSETCC, VT, Expand);
  setOperationAction(ISD::STRICT_FSETCCS, VT, Expand);

  // When little-endian we can use ordinary d and q register loads/stores for
  // vector types, but when big-endian we need to use structure load/store which
  // only allow post-index addressing.
  if (Subtarget->isLittleEndian()) {
    for (unsigned im = (unsigned)ISD::PRE_INC;
         im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
      setIndexedLoadAction(im, VT, Legal);
      setIndexedStoreAction(im, VT, Legal);
    }
  } else {
    setIndexedLoadAction(ISD::POST_INC, VT, Legal);
    setIndexedStoreAction(ISD::POST_INC, VT, Legal);
  }

  if (Subtarget->hasD128()) {
    setOperationAction(ISD::READ_REGISTER, MVT::i128, Custom);
    setOperationAction(ISD::WRITE_REGISTER, MVT::i128, Custom);
  }

  if (VT.isInteger()) {
    // Let common code emit inverted variants of compares we do support.
    setCondCodeAction(ISD::SETNE, VT, Expand);
    setCondCodeAction(ISD::SETLE, VT, Expand);
    setCondCodeAction(ISD::SETLT, VT, Expand);
    setCondCodeAction(ISD::SETULE, VT, Expand);
    setCondCodeAction(ISD::SETULT, VT, Expand);
  }
}

bool AArch64TargetLowering::shouldExpandGetActiveLaneMask(EVT ResVT,
                                                          EVT OpVT) const {
  // Only SVE has a 1:1 mapping from intrinsic -> instruction (whilelo).
  if (!Subtarget->isSVEorStreamingSVEAvailable() ||
      ResVT.getVectorElementType() != MVT::i1)
    return true;

  // Only support illegal types if the result is scalable and min elements > 1.
  if (ResVT.getVectorMinNumElements() == 1 ||
      (ResVT.isFixedLengthVector() && (ResVT.getVectorNumElements() > 16 ||
                                       (OpVT != MVT::i32 && OpVT != MVT::i64))))
    return true;

  // 32 & 64 bit operands are supported. We can promote anything < 64 bits,
  // but anything larger should be expanded.
  if (OpVT.getFixedSizeInBits() > 64)
    return true;

  return false;
}

bool AArch64TargetLowering::shouldExpandPartialReductionIntrinsic(
    const IntrinsicInst *I) const {
  assert(I->getIntrinsicID() ==
             Intrinsic::experimental_vector_partial_reduce_add &&
         "Unexpected intrinsic!");
  return true;
}

bool AArch64TargetLowering::shouldExpandCttzElements(EVT VT) const {
  if (!Subtarget->isSVEorStreamingSVEAvailable())
    return true;

  // We can only use the BRKB + CNTP sequence with legal predicate types. We can
  // also support fixed-width predicates.
  return VT != MVT::nxv16i1 && VT != MVT::nxv8i1 && VT != MVT::nxv4i1 &&
         VT != MVT::nxv2i1 && VT != MVT::v16i1 && VT != MVT::v8i1 &&
         VT != MVT::v4i1 && VT != MVT::v2i1;
}

bool AArch64TargetLowering::shouldExpandVectorMatch(EVT VT,
                                                    unsigned SearchSize) const {
  // MATCH is SVE2 and only available in non-streaming mode.
  if (!Subtarget->hasSVE2() || !Subtarget->isSVEAvailable())
    return true;
  // Furthermore, we can only use it for 8-bit or 16-bit elements.
  if (VT == MVT::nxv8i16 || VT == MVT::v8i16)
    return SearchSize != 8;
  if (VT == MVT::nxv16i8 || VT == MVT::v16i8 || VT == MVT::v8i8)
    return SearchSize != 8 && SearchSize != 16;
  return true;
}

void AArch64TargetLowering::addTypeForFixedLengthSVE(MVT VT) {
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  // By default everything must be expanded.
  for (unsigned Op = 0; Op < ISD::BUILTIN_OP_END; ++Op)
    setOperationAction(Op, VT, Expand);

  if (VT.isFloatingPoint()) {
    setCondCodeAction(ISD::SETO, VT, Expand);
    setCondCodeAction(ISD::SETOLT, VT, Expand);
    setCondCodeAction(ISD::SETOLE, VT, Expand);
    setCondCodeAction(ISD::SETULT, VT, Expand);
    setCondCodeAction(ISD::SETULE, VT, Expand);
    setCondCodeAction(ISD::SETUGE, VT, Expand);
    setCondCodeAction(ISD::SETUGT, VT, Expand);
    setCondCodeAction(ISD::SETUEQ, VT, Expand);
    setCondCodeAction(ISD::SETONE, VT, Expand);
  }

  TargetLoweringBase::LegalizeAction Default =
      VT == MVT::v1f64 ? Expand : Custom;

  // Mark integer truncating stores/extending loads as having custom lowering
  if (VT.isInteger()) {
    MVT InnerVT = VT.changeVectorElementType(MVT::i8);
    while (InnerVT != VT) {
      setTruncStoreAction(VT, InnerVT, Default);
      setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Default);
      setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Default);
      setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Default);
      InnerVT = InnerVT.changeVectorElementType(
          MVT::getIntegerVT(2 * InnerVT.getScalarSizeInBits()));
    }
  }

  // Mark floating-point truncating stores/extending loads as having custom
  // lowering
  if (VT.isFloatingPoint()) {
    MVT InnerVT = VT.changeVectorElementType(MVT::f16);
    while (InnerVT != VT) {
      setTruncStoreAction(VT, InnerVT, Custom);
      setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Default);
      InnerVT = InnerVT.changeVectorElementType(
          MVT::getFloatingPointVT(2 * InnerVT.getScalarSizeInBits()));
    }
  }

  bool PreferNEON = VT.is64BitVector() || VT.is128BitVector();
  bool PreferSVE = !PreferNEON && Subtarget->isSVEAvailable();

  static const unsigned MLAOps[] = {ISD::PARTIAL_REDUCE_SMLA,
                                    ISD::PARTIAL_REDUCE_UMLA};
  unsigned NumElts = VT.getVectorNumElements();
  if (VT.getVectorElementType() == MVT::i64) {
    setPartialReduceMLAAction(MLAOps, VT,
                              MVT::getVectorVT(MVT::i8, NumElts * 8), Custom);
    setPartialReduceMLAAction(MLAOps, VT,
                              MVT::getVectorVT(MVT::i16, NumElts * 4), Custom);
    setPartialReduceMLAAction(MLAOps, VT,
                              MVT::getVectorVT(MVT::i32, NumElts * 2), Custom);
  } else if (VT.getVectorElementType() == MVT::i32) {
    setPartialReduceMLAAction(MLAOps, VT,
                              MVT::getVectorVT(MVT::i8, NumElts * 4), Custom);
    setPartialReduceMLAAction(MLAOps, VT,
                              MVT::getVectorVT(MVT::i16, NumElts * 2), Custom);
  } else if (VT.getVectorElementType() == MVT::i16) {
    setPartialReduceMLAAction(MLAOps, VT,
                              MVT::getVectorVT(MVT::i8, NumElts * 2), Custom);
  }
  if (Subtarget->hasMatMulInt8()) {
    if (VT.getVectorElementType() == MVT::i32)
      setPartialReduceMLAAction(ISD::PARTIAL_REDUCE_SUMLA, VT,
                                MVT::getVectorVT(MVT::i8, NumElts * 4), Custom);
    else if (VT.getVectorElementType() == MVT::i64)
      setPartialReduceMLAAction(ISD::PARTIAL_REDUCE_SUMLA, VT,
                                MVT::getVectorVT(MVT::i8, NumElts * 8), Custom);
  }

  // Lower fixed length vector operations to scalable equivalents.
  setOperationAction(ISD::ABDS, VT, Default);
  setOperationAction(ISD::ABDU, VT, Default);
  setOperationAction(ISD::ABS, VT, Default);
  setOperationAction(ISD::ADD, VT, Default);
  setOperationAction(ISD::AND, VT, Default);
  setOperationAction(ISD::ANY_EXTEND, VT, Default);
  setOperationAction(ISD::BITCAST, VT, PreferNEON ? Legal : Default);
  setOperationAction(ISD::BITREVERSE, VT, Default);
  setOperationAction(ISD::BSWAP, VT, Default);
  setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
  setOperationAction(ISD::CONCAT_VECTORS, VT, Default);
  setOperationAction(ISD::CTLZ, VT, Default);
  setOperationAction(ISD::CTPOP, VT, Default);
  setOperationAction(ISD::CTTZ, VT, Default);
  setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Default);
  setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Default);
  setOperationAction(ISD::FABS, VT, Default);
  setOperationAction(ISD::FADD, VT, Default);
  setOperationAction(ISD::FCEIL, VT, Default);
  setOperationAction(ISD::FCOPYSIGN, VT, Default);
  setOperationAction(ISD::FDIV, VT, Default);
  setOperationAction(ISD::FFLOOR, VT, Default);
  setOperationAction(ISD::FMA, VT, Default);
  setOperationAction(ISD::FMAXIMUM, VT, Default);
  setOperationAction(ISD::FMAXNUM, VT, Default);
  setOperationAction(ISD::FMINIMUM, VT, Default);
  setOperationAction(ISD::FMINNUM, VT, Default);
  setOperationAction(ISD::FMUL, VT, Default);
  setOperationAction(ISD::FNEARBYINT, VT, Default);
  setOperationAction(ISD::FNEG, VT, Default);
  setOperationAction(ISD::FP_EXTEND, VT, Default);
  setOperationAction(ISD::FP_ROUND, VT, Default);
  setOperationAction(ISD::FP_TO_SINT, VT, Default);
  setOperationAction(ISD::FP_TO_UINT, VT, Default);
  setOperationAction(ISD::FRINT, VT, Default);
  setOperationAction(ISD::LRINT, VT, Default);
  setOperationAction(ISD::LLRINT, VT, Default);
  setOperationAction(ISD::FROUND, VT, Default);
  setOperationAction(ISD::FROUNDEVEN, VT, Default);
  setOperationAction(ISD::FSQRT, VT, Default);
  setOperationAction(ISD::FSUB, VT, Default);
  setOperationAction(ISD::FTRUNC, VT, Default);
  setOperationAction(ISD::GET_ACTIVE_LANE_MASK, VT, Default);
  setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Default);
  setOperationAction(ISD::LOAD, VT, PreferNEON ? Legal : Default);
  setOperationAction(ISD::MGATHER, VT, PreferSVE ? Default : Expand);
  setOperationAction(ISD::MLOAD, VT, Default);
  setOperationAction(ISD::MSCATTER, VT, PreferSVE ? Default : Expand);
  setOperationAction(ISD::MSTORE, VT, Default);
  setOperationAction(ISD::MUL, VT, Default);
  setOperationAction(ISD::MULHS, VT, Default);
  setOperationAction(ISD::MULHU, VT, Default);
  setOperationAction(ISD::OR, VT, Default);
  setOperationAction(ISD::SCALAR_TO_VECTOR, VT, PreferNEON ? Legal : Expand);
  setOperationAction(ISD::SDIV, VT, Default);
  setOperationAction(ISD::SELECT, VT, Default);
  setOperationAction(ISD::SETCC, VT, Default);
  setOperationAction(ISD::SHL, VT, Default);
  setOperationAction(ISD::SIGN_EXTEND, VT, Default);
  setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Default);
  setOperationAction(ISD::SINT_TO_FP, VT, Default);
  setOperationAction(ISD::SMAX, VT, Default);
  setOperationAction(ISD::SMIN, VT, Default);
  setOperationAction(ISD::SPLAT_VECTOR, VT, Default);
  setOperationAction(ISD::SRA, VT, Default);
  setOperationAction(ISD::SRL, VT, Default);
  setOperationAction(ISD::STORE, VT, PreferNEON ? Legal : Default);
  setOperationAction(ISD::SUB, VT, Default);
  setOperationAction(ISD::TRUNCATE, VT, Default);
  setOperationAction(ISD::UDIV, VT, Default);
  setOperationAction(ISD::UINT_TO_FP, VT, Default);
  setOperationAction(ISD::UMAX, VT, Default);
  setOperationAction(ISD::UMIN, VT, Default);
  setOperationAction(ISD::VECREDUCE_ADD, VT, Default);
  setOperationAction(ISD::VECREDUCE_AND, VT, Default);
  setOperationAction(ISD::VECREDUCE_FADD, VT, Default);
  setOperationAction(ISD::VECREDUCE_FMAX, VT, Default);
  setOperationAction(ISD::VECREDUCE_FMIN, VT, Default);
  setOperationAction(ISD::VECREDUCE_FMAXIMUM, VT, Default);
  setOperationAction(ISD::VECREDUCE_FMINIMUM, VT, Default);
  setOperationAction(ISD::VECREDUCE_OR, VT, Default);
  setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, PreferSVE ? Default : Expand);
  setOperationAction(ISD::VECREDUCE_SMAX, VT, Default);
  setOperationAction(ISD::VECREDUCE_SMIN, VT, Default);
  setOperationAction(ISD::VECREDUCE_UMAX, VT, Default);
  setOperationAction(ISD::VECREDUCE_UMIN, VT, Default);
  setOperationAction(ISD::VECREDUCE_XOR, VT, Default);
  setOperationAction(ISD::VECTOR_SHUFFLE, VT, Default);
  setOperationAction(ISD::VECTOR_SPLICE, VT, Default);
  setOperationAction(ISD::VSELECT, VT, Default);
  setOperationAction(ISD::XOR, VT, Default);
  setOperationAction(ISD::ZERO_EXTEND, VT, Default);
}

void AArch64TargetLowering::addDRType(MVT VT) {
  addRegisterClass(VT, &AArch64::FPR64RegClass);
  if (Subtarget->isNeonAvailable())
    addTypeForNEON(VT);
}

void AArch64TargetLowering::addQRType(MVT VT) {
  addRegisterClass(VT, &AArch64::FPR128RegClass);
  if (Subtarget->isNeonAvailable())
    addTypeForNEON(VT);
}

EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &,
                                              LLVMContext &C, EVT VT) const {
  if (!VT.isVector())
    return MVT::i32;
  if (VT.isScalableVector())
    return EVT::getVectorVT(C, MVT::i1, VT.getVectorElementCount());
  return VT.changeVectorElementTypeToInteger();
}

// isIntImmediate - This method tests to see if the node is a constant
// operand. If so Imm will receive the value.
static bool isIntImmediate(const SDNode *N, uint64_t &Imm) {
  if (const ConstantSDNode *C = dyn_cast<const ConstantSDNode>(N)) {
    Imm = C->getZExtValue();
    return true;
  }
  return false;
}

bool isVectorizedBinOp(unsigned Opcode) {
  switch (Opcode) {
  case AArch64ISD::SQDMULH:
    return true;
  default:
    return false;
  }
}

// isOpcWithIntImmediate - This method tests to see if the node is a specific
// opcode and that it has a immediate integer right operand.
// If so Imm will receive the value.
static bool isOpcWithIntImmediate(const SDNode *N, unsigned Opc,
                                  uint64_t &Imm) {
  return N->getOpcode() == Opc &&
         isIntImmediate(N->getOperand(1).getNode(), Imm);
}

static bool optimizeLogicalImm(SDValue Op, unsigned Size, uint64_t Imm,
                               const APInt &Demanded,
                               TargetLowering::TargetLoweringOpt &TLO,
                               unsigned NewOpc) {
  uint64_t OldImm = Imm, NewImm, Enc;
  uint64_t Mask = ((uint64_t)(-1LL) >> (64 - Size)), OrigMask = Mask;

  // Return if the immediate is already all zeros, all ones, a bimm32 or a
  // bimm64.
  if (Imm == 0 || Imm == Mask ||
      AArch64_AM::isLogicalImmediate(Imm & Mask, Size))
    return false;

  unsigned EltSize = Size;
  uint64_t DemandedBits = Demanded.getZExtValue();

  // Clear bits that are not demanded.
  Imm &= DemandedBits;

  while (true) {
    // The goal here is to set the non-demanded bits in a way that minimizes
    // the number of switching between 0 and 1. In order to achieve this goal,
    // we set the non-demanded bits to the value of the preceding demanded bits.
    // For example, if we have an immediate 0bx10xx0x1 ('x' indicates a
    // non-demanded bit), we copy bit0 (1) to the least significant 'x',
    // bit2 (0) to 'xx', and bit6 (1) to the most significant 'x'.
    // The final result is 0b11000011.
    uint64_t NonDemandedBits = ~DemandedBits;
    uint64_t InvertedImm = ~Imm & DemandedBits;
    uint64_t RotatedImm =
        ((InvertedImm << 1) | (InvertedImm >> (EltSize - 1) & 1)) &
        NonDemandedBits;
    uint64_t Sum = RotatedImm + NonDemandedBits;
    bool Carry = NonDemandedBits & ~Sum & (1ULL << (EltSize - 1));
    uint64_t Ones = (Sum + Carry) & NonDemandedBits;
    NewImm = (Imm | Ones) & Mask;

    // If NewImm or its bitwise NOT is a shifted mask, it is a bitmask immediate
    // or all-ones or all-zeros, in which case we can stop searching. Otherwise,
    // we halve the element size and continue the search.
    if (isShiftedMask_64(NewImm) || isShiftedMask_64(~(NewImm | ~Mask)))
      break;

    // We cannot shrink the element size any further if it is 2-bits.
    if (EltSize == 2)
      return false;

    EltSize /= 2;
    Mask >>= EltSize;
    uint64_t Hi = Imm >> EltSize, DemandedBitsHi = DemandedBits >> EltSize;

    // Return if there is mismatch in any of the demanded bits of Imm and Hi.
    if (((Imm ^ Hi) & (DemandedBits & DemandedBitsHi) & Mask) != 0)
      return false;

    // Merge the upper and lower halves of Imm and DemandedBits.
    Imm |= Hi;
    DemandedBits |= DemandedBitsHi;
  }

  ++NumOptimizedImms;

  // Replicate the element across the register width.
  while (EltSize < Size) {
    NewImm |= NewImm << EltSize;
    EltSize *= 2;
  }

  (void)OldImm;
  assert(((OldImm ^ NewImm) & Demanded.getZExtValue()) == 0 &&
         "demanded bits should never be altered");
  assert(OldImm != NewImm && "the new imm shouldn't be equal to the old imm");

  // Create the new constant immediate node.
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue New;

  // If the new constant immediate is all-zeros or all-ones, let the target
  // independent DAG combine optimize this node.
  if (NewImm == 0 || NewImm == OrigMask) {
    New = TLO.DAG.getNode(Op.getOpcode(), DL, VT, Op.getOperand(0),
                          TLO.DAG.getConstant(NewImm, DL, VT));
  // Otherwise, create a machine node so that target independent DAG combine
  // doesn't undo this optimization.
  } else {
    Enc = AArch64_AM::encodeLogicalImmediate(NewImm, Size);
    SDValue EncConst = TLO.DAG.getTargetConstant(Enc, DL, VT);
    New = SDValue(
        TLO.DAG.getMachineNode(NewOpc, DL, VT, Op.getOperand(0), EncConst), 0);
  }

  return TLO.CombineTo(Op, New);
}

bool AArch64TargetLowering::targetShrinkDemandedConstant(
    SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts,
    TargetLoweringOpt &TLO) const {
  // Delay this optimization to as late as possible.
  if (!TLO.LegalOps)
    return false;

  if (!EnableOptimizeLogicalImm)
    return false;

  EVT VT = Op.getValueType();
  if (VT.isVector())
    return false;

  unsigned Size = VT.getSizeInBits();

  if (Size != 32 && Size != 64)
    return false;

  // Exit early if we demand all bits.
  if (DemandedBits.popcount() == Size)
    return false;

  unsigned NewOpc;
  switch (Op.getOpcode()) {
  default:
    return false;
  case ISD::AND:
    NewOpc = Size == 32 ? AArch64::ANDWri : AArch64::ANDXri;
    break;
  case ISD::OR:
    NewOpc = Size == 32 ? AArch64::ORRWri : AArch64::ORRXri;
    break;
  case ISD::XOR:
    NewOpc = Size == 32 ? AArch64::EORWri : AArch64::EORXri;
    break;
  }
  ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
  if (!C)
    return false;
  uint64_t Imm = C->getZExtValue();
  return optimizeLogicalImm(Op, Size, Imm, DemandedBits, TLO, NewOpc);
}

/// computeKnownBitsForTargetNode - Determine which of the bits specified in
/// Mask are known to be either zero or one and return them Known.
void AArch64TargetLowering::computeKnownBitsForTargetNode(
    const SDValue Op, KnownBits &Known, const APInt &DemandedElts,
    const SelectionDAG &DAG, unsigned Depth) const {
  switch (Op.getOpcode()) {
  default:
    break;
  case AArch64ISD::DUP: {
    SDValue SrcOp = Op.getOperand(0);
    Known = DAG.computeKnownBits(SrcOp, Depth + 1);
    if (SrcOp.getValueSizeInBits() != Op.getScalarValueSizeInBits()) {
      assert(SrcOp.getValueSizeInBits() > Op.getScalarValueSizeInBits() &&
             "Expected DUP implicit truncation");
      Known = Known.trunc(Op.getScalarValueSizeInBits());
    }
    break;
  }
  case AArch64ISD::CSEL: {
    KnownBits Known2;
    Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1);
    Known2 = DAG.computeKnownBits(Op->getOperand(1), Depth + 1);
    Known = Known.intersectWith(Known2);
    break;
  }
  case AArch64ISD::BICi: {
    // Compute the bit cleared value.
    APInt Mask =
        ~(Op->getConstantOperandAPInt(1) << Op->getConstantOperandAPInt(2))
             .trunc(Known.getBitWidth());
    Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1);
    Known &= KnownBits::makeConstant(Mask);
    break;
  }
  case AArch64ISD::VLSHR: {
    KnownBits Known2;
    Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1);
    Known2 = DAG.computeKnownBits(Op->getOperand(1), Depth + 1);
    Known = KnownBits::lshr(Known, Known2);
    break;
  }
  case AArch64ISD::VASHR: {
    KnownBits Known2;
    Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1);
    Known2 = DAG.computeKnownBits(Op->getOperand(1), Depth + 1);
    Known = KnownBits::ashr(Known, Known2);
    break;
  }
  case AArch64ISD::VSHL: {
    KnownBits Known2;
    Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1);
    Known2 = DAG.computeKnownBits(Op->getOperand(1), Depth + 1);
    Known = KnownBits::shl(Known, Known2);
    break;
  }
  case AArch64ISD::MOVI: {
    Known = KnownBits::makeConstant(
        APInt(Known.getBitWidth(), Op->getConstantOperandVal(0)));
    break;
  }
  case AArch64ISD::MOVIshift: {
    Known = KnownBits::makeConstant(
        APInt(Known.getBitWidth(), Op->getConstantOperandVal(0)
                                       << Op->getConstantOperandVal(1)));
    break;
  }
  case AArch64ISD::LOADgot:
  case AArch64ISD::ADDlow: {
    if (!Subtarget->isTargetILP32())
      break;
    // In ILP32 mode all valid pointers are in the low 4GB of the address-space.
    Known.Zero = APInt::getHighBitsSet(64, 32);
    break;
  }
  case AArch64ISD::ASSERT_ZEXT_BOOL: {
    Known = DAG.computeKnownBits(Op->getOperand(0), Depth + 1);
    Known.Zero |= APInt(Known.getBitWidth(), 0xFE);
    break;
  }
  case ISD::INTRINSIC_W_CHAIN: {
    Intrinsic::ID IntID =
        static_cast<Intrinsic::ID>(Op->getConstantOperandVal(1));
    switch (IntID) {
    default: return;
    case Intrinsic::aarch64_ldaxr:
    case Intrinsic::aarch64_ldxr: {
      unsigned BitWidth = Known.getBitWidth();
      EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
      unsigned MemBits = VT.getScalarSizeInBits();
      Known.Zero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
      return;
    }
    }
    break;
  }
  case ISD::INTRINSIC_WO_CHAIN:
  case ISD::INTRINSIC_VOID: {
    unsigned IntNo = Op.getConstantOperandVal(0);
    switch (IntNo) {
    default:
      break;
    case Intrinsic::aarch64_neon_uaddlv: {
      MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
      unsigned BitWidth = Known.getBitWidth();
      if (VT == MVT::v8i8 || VT == MVT::v16i8) {
        unsigned Bound = (VT == MVT::v8i8) ?  11 : 12;
        assert(BitWidth >= Bound && "Unexpected width!");
        APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - Bound);
        Known.Zero |= Mask;
      }
      break;
    }
    case Intrinsic::aarch64_neon_umaxv:
    case Intrinsic::aarch64_neon_uminv: {
      // Figure out the datatype of the vector operand. The UMINV instruction
      // will zero extend the result, so we can mark as known zero all the
      // bits larger than the element datatype. 32-bit or larget doesn't need
      // this as those are legal types and will be handled by isel directly.
      MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
      unsigned BitWidth = Known.getBitWidth();
      if (VT == MVT::v8i8 || VT == MVT::v16i8) {
        assert(BitWidth >= 8 && "Unexpected width!");
        APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
        Known.Zero |= Mask;
      } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
        assert(BitWidth >= 16 && "Unexpected width!");
        APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
        Known.Zero |= Mask;
      }
      break;
    } break;
    }
  }
  }
}

unsigned AArch64TargetLowering::ComputeNumSignBitsForTargetNode(
    SDValue Op, const APInt &DemandedElts, const SelectionDAG &DAG,
    unsigned Depth) const {
  EVT VT = Op.getValueType();
  unsigned VTBits = VT.getScalarSizeInBits();
  unsigned Opcode = Op.getOpcode();
  switch (Opcode) {
  case AArch64ISD::FCMEQ:
  case AArch64ISD::FCMGE:
  case AArch64ISD::FCMGT:
    // Compares return either 0 or all-ones
    return VTBits;
  case AArch64ISD::VASHR: {
    unsigned Tmp =
        DAG.ComputeNumSignBits(Op.getOperand(0), DemandedElts, Depth + 1);
    return std::min<uint64_t>(Tmp + Op.getConstantOperandVal(1), VTBits);
  }
  }

  return 1;
}

MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
                                                  EVT) const {
  return MVT::i64;
}

bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(
    EVT VT, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags,
    unsigned *Fast) const {

  // Allow SVE loads/stores where the alignment >= the size of the element type,
  // even with +strict-align. Predicated SVE loads/stores (e.g. ld1/st1), used
  // for stores that come from IR, only require element-size alignment (even if
  // unaligned accesses are disabled). Without this, these will be forced to
  // have 16-byte alignment with +strict-align (and fail to lower as we don't
  // yet support TLI.expandUnalignedLoad() and TLI.expandUnalignedStore()).
  if (VT.isScalableVector()) {
    unsigned ElementSizeBits = VT.getScalarSizeInBits();
    if (ElementSizeBits % 8 == 0 && Alignment >= Align(ElementSizeBits / 8))
      return true;
  }

  if (Subtarget->requiresStrictAlign())
    return false;

  if (Fast) {
    // Some CPUs are fine with unaligned stores except for 128-bit ones.
    *Fast = !Subtarget->isMisaligned128StoreSlow() || VT.getStoreSize() != 16 ||
            // See comments in performSTORECombine() for more details about
            // these conditions.

            // Code that uses clang vector extensions can mark that it
            // wants unaligned accesses to be treated as fast by
            // underspecifying alignment to be 1 or 2.
            Alignment <= 2 ||

            // Disregard v2i64. Memcpy lowering produces those and splitting
            // them regresses performance on micro-benchmarks and olden/bh.
            VT == MVT::v2i64;
  }
  return true;
}

// Same as above but handling LLTs instead.
bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(
    LLT Ty, unsigned AddrSpace, Align Alignment, MachineMemOperand::Flags Flags,
    unsigned *Fast) const {
  if (Subtarget->requiresStrictAlign())
    return false;

  if (Fast) {
    // Some CPUs are fine with unaligned stores except for 128-bit ones.
    *Fast = !Subtarget->isMisaligned128StoreSlow() ||
            Ty.getSizeInBytes() != 16 ||
            // See comments in performSTORECombine() for more details about
            // these conditions.

            // Code that uses clang vector extensions can mark that it
            // wants unaligned accesses to be treated as fast by
            // underspecifying alignment to be 1 or 2.
            Alignment <= 2 ||

            // Disregard v2i64. Memcpy lowering produces those and splitting
            // them regresses performance on micro-benchmarks and olden/bh.
            Ty == LLT::fixed_vector(2, 64);
  }
  return true;
}

FastISel *
AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
                                      const TargetLibraryInfo *libInfo) const {
  return AArch64::createFastISel(funcInfo, libInfo);
}

MachineBasicBlock *
AArch64TargetLowering::EmitF128CSEL(MachineInstr &MI,
                                    MachineBasicBlock *MBB) const {
  // We materialise the F128CSEL pseudo-instruction as some control flow and a
  // phi node:

  // OrigBB:
  //     [... previous instrs leading to comparison ...]
  //     b.ne TrueBB
  //     b EndBB
  // TrueBB:
  //     ; Fallthrough
  // EndBB:
  //     Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]

  MachineFunction *MF = MBB->getParent();
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  const BasicBlock *LLVM_BB = MBB->getBasicBlock();
  DebugLoc DL = MI.getDebugLoc();
  MachineFunction::iterator It = ++MBB->getIterator();

  Register DestReg = MI.getOperand(0).getReg();
  Register IfTrueReg = MI.getOperand(1).getReg();
  Register IfFalseReg = MI.getOperand(2).getReg();
  unsigned CondCode = MI.getOperand(3).getImm();
  bool NZCVKilled = MI.getOperand(4).isKill();

  MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
  MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
  MF->insert(It, TrueBB);
  MF->insert(It, EndBB);

  // Transfer rest of current basic-block to EndBB
  EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
                MBB->end());
  EndBB->transferSuccessorsAndUpdatePHIs(MBB);

  BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
  BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
  MBB->addSuccessor(TrueBB);
  MBB->addSuccessor(EndBB);

  // TrueBB falls through to the end.
  TrueBB->addSuccessor(EndBB);

  if (!NZCVKilled) {
    TrueBB->addLiveIn(AArch64::NZCV);
    EndBB->addLiveIn(AArch64::NZCV);
  }

  BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
      .addReg(IfTrueReg)
      .addMBB(TrueBB)
      .addReg(IfFalseReg)
      .addMBB(MBB);

  MI.eraseFromParent();
  return EndBB;
}

MachineBasicBlock *AArch64TargetLowering::EmitLoweredCatchRet(
       MachineInstr &MI, MachineBasicBlock *BB) const {
  assert(!isAsynchronousEHPersonality(classifyEHPersonality(
             BB->getParent()->getFunction().getPersonalityFn())) &&
         "SEH does not use catchret!");
  return BB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitDynamicProbedAlloc(MachineInstr &MI,
                                              MachineBasicBlock *MBB) const {
  MachineFunction &MF = *MBB->getParent();
  MachineBasicBlock::iterator MBBI = MI.getIterator();
  const AArch64InstrInfo &TII =
      *MF.getSubtarget<AArch64Subtarget>().getInstrInfo();
  Register TargetReg = MI.getOperand(0).getReg();
  MachineBasicBlock::iterator NextInst =
      TII.probedStackAlloc(MBBI, TargetReg, false);

  MI.eraseFromParent();
  return NextInst->getParent();
}

MachineBasicBlock *
AArch64TargetLowering::EmitTileLoad(unsigned Opc, unsigned BaseReg,
                                    MachineInstr &MI,
                                    MachineBasicBlock *BB) const {
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  MachineInstrBuilder MIB = BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(Opc));

  MIB.addReg(BaseReg + MI.getOperand(0).getImm(), RegState::Define);
  MIB.add(MI.getOperand(1)); // slice index register
  MIB.add(MI.getOperand(2)); // slice index offset
  MIB.add(MI.getOperand(3)); // pg
  MIB.add(MI.getOperand(4)); // base
  MIB.add(MI.getOperand(5)); // offset

  MI.eraseFromParent(); // The pseudo is gone now.
  return BB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitFill(MachineInstr &MI, MachineBasicBlock *BB) const {
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  MachineInstrBuilder MIB =
      BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::LDR_ZA));

  MIB.addReg(AArch64::ZA, RegState::Define);
  MIB.add(MI.getOperand(0)); // Vector select register
  MIB.add(MI.getOperand(1)); // Vector select offset
  MIB.add(MI.getOperand(2)); // Base
  MIB.add(MI.getOperand(1)); // Offset, same as vector select offset

  MI.eraseFromParent(); // The pseudo is gone now.
  return BB;
}

MachineBasicBlock *AArch64TargetLowering::EmitZTInstr(MachineInstr &MI,
                                                      MachineBasicBlock *BB,
                                                      unsigned Opcode,
                                                      bool Op0IsDef) const {
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  MachineInstrBuilder MIB;

  MIB = BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(Opcode))
            .addReg(MI.getOperand(0).getReg(), Op0IsDef ? RegState::Define : 0);
  for (unsigned I = 1; I < MI.getNumOperands(); ++I)
    MIB.add(MI.getOperand(I));

  MI.eraseFromParent(); // The pseudo is gone now.
  return BB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitZAInstr(unsigned Opc, unsigned BaseReg,
                                   MachineInstr &MI,
                                   MachineBasicBlock *BB) const {
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  MachineInstrBuilder MIB = BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(Opc));
  unsigned StartIdx = 0;

  bool HasTile = BaseReg != AArch64::ZA;
  bool HasZPROut = HasTile && MI.getOperand(0).isReg();
  if (HasZPROut) {
    MIB.add(MI.getOperand(StartIdx)); // Output ZPR
    ++StartIdx;
  }
  if (HasTile) {
    MIB.addReg(BaseReg + MI.getOperand(StartIdx).getImm(),
               RegState::Define);                           // Output ZA Tile
    MIB.addReg(BaseReg + MI.getOperand(StartIdx).getImm()); // Input Za Tile
    StartIdx++;
  } else {
    // Avoids all instructions with mnemonic za.<sz>[Reg, Imm,
    if (MI.getOperand(0).isReg() && !MI.getOperand(1).isImm()) {
      MIB.add(MI.getOperand(StartIdx)); // Output ZPR
      ++StartIdx;
    }
    MIB.addReg(BaseReg, RegState::Define).addReg(BaseReg);
  }
  for (unsigned I = StartIdx; I < MI.getNumOperands(); ++I)
    MIB.add(MI.getOperand(I));

  MI.eraseFromParent(); // The pseudo is gone now.
  return BB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitZero(MachineInstr &MI, MachineBasicBlock *BB) const {
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  MachineInstrBuilder MIB =
      BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::ZERO_M));
  MIB.add(MI.getOperand(0)); // Mask

  unsigned Mask = MI.getOperand(0).getImm();
  for (unsigned I = 0; I < 8; I++) {
    if (Mask & (1 << I))
      MIB.addDef(AArch64::ZAD0 + I, RegState::ImplicitDefine);
  }

  MI.eraseFromParent(); // The pseudo is gone now.
  return BB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitInitTPIDR2Object(MachineInstr &MI,
                                            MachineBasicBlock *BB) const {
  MachineFunction *MF = BB->getParent();
  MachineFrameInfo &MFI = MF->getFrameInfo();
  AArch64FunctionInfo *FuncInfo = MF->getInfo<AArch64FunctionInfo>();
  TPIDR2Object &TPIDR2 = FuncInfo->getTPIDR2Obj();
  if (TPIDR2.Uses > 0) {
    const TargetInstrInfo *TII = Subtarget->getInstrInfo();
    // Store the buffer pointer to the TPIDR2 stack object.
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::STRXui))
        .addReg(MI.getOperand(0).getReg())
        .addFrameIndex(TPIDR2.FrameIndex)
        .addImm(0);
    // Set the reserved bytes (10-15) to zero
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::STRHHui))
        .addReg(AArch64::WZR)
        .addFrameIndex(TPIDR2.FrameIndex)
        .addImm(5);
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::STRWui))
        .addReg(AArch64::WZR)
        .addFrameIndex(TPIDR2.FrameIndex)
        .addImm(3);
  } else
    MFI.RemoveStackObject(TPIDR2.FrameIndex);

  BB->remove_instr(&MI);
  return BB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitAllocateZABuffer(MachineInstr &MI,
                                            MachineBasicBlock *BB) const {
  MachineFunction *MF = BB->getParent();
  MachineFrameInfo &MFI = MF->getFrameInfo();
  AArch64FunctionInfo *FuncInfo = MF->getInfo<AArch64FunctionInfo>();
  // TODO This function grows the stack with a subtraction, which doesn't work
  // on Windows. Some refactoring to share the functionality in
  // LowerWindowsDYNAMIC_STACKALLOC will be required once the Windows ABI
  // supports SME
  assert(!MF->getSubtarget<AArch64Subtarget>().isTargetWindows() &&
         "Lazy ZA save is not yet supported on Windows");

  TPIDR2Object &TPIDR2 = FuncInfo->getTPIDR2Obj();

  if (TPIDR2.Uses > 0) {
    const TargetInstrInfo *TII = Subtarget->getInstrInfo();
    MachineRegisterInfo &MRI = MF->getRegInfo();

    // The SUBXrs below won't always be emitted in a form that accepts SP
    // directly
    Register SP = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY), SP)
        .addReg(AArch64::SP);

    // Allocate a lazy-save buffer object of the size given, normally SVL * SVL
    auto Size = MI.getOperand(1).getReg();
    auto Dest = MI.getOperand(0).getReg();
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::MSUBXrrr), Dest)
        .addReg(Size)
        .addReg(Size)
        .addReg(SP);
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY),
            AArch64::SP)
        .addReg(Dest);

    // We have just allocated a variable sized object, tell this to PEI.
    MFI.CreateVariableSizedObject(Align(16), nullptr);
  }

  BB->remove_instr(&MI);
  return BB;
}

// TODO: Find a way to merge this with EmitAllocateZABuffer.
MachineBasicBlock *
AArch64TargetLowering::EmitAllocateSMESaveBuffer(MachineInstr &MI,
                                                 MachineBasicBlock *BB) const {
  MachineFunction *MF = BB->getParent();
  MachineFrameInfo &MFI = MF->getFrameInfo();
  AArch64FunctionInfo *FuncInfo = MF->getInfo<AArch64FunctionInfo>();
  assert(!MF->getSubtarget<AArch64Subtarget>().isTargetWindows() &&
         "Lazy ZA save is not yet supported on Windows");

  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  if (FuncInfo->isSMESaveBufferUsed()) {
    // Allocate a buffer object of the size given by MI.getOperand(1).
    auto Size = MI.getOperand(1).getReg();
    auto Dest = MI.getOperand(0).getReg();
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::SUBXrx64), AArch64::SP)
        .addReg(AArch64::SP)
        .addReg(Size)
        .addImm(AArch64_AM::getArithExtendImm(AArch64_AM::UXTX, 0));
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY), Dest)
        .addReg(AArch64::SP);

    // We have just allocated a variable sized object, tell this to PEI.
    MFI.CreateVariableSizedObject(Align(16), nullptr);
  } else
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::IMPLICIT_DEF),
            MI.getOperand(0).getReg());

  BB->remove_instr(&MI);
  return BB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitGetSMESaveSize(MachineInstr &MI,
                                          MachineBasicBlock *BB) const {
  // If the buffer is used, emit a call to __arm_sme_state_size()
  MachineFunction *MF = BB->getParent();
  AArch64FunctionInfo *FuncInfo = MF->getInfo<AArch64FunctionInfo>();
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  if (FuncInfo->isSMESaveBufferUsed()) {
    const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(AArch64::BL))
        .addExternalSymbol("__arm_sme_state_size")
        .addReg(AArch64::X0, RegState::ImplicitDefine)
        .addRegMask(TRI->getCallPreservedMask(
            *MF, CallingConv::
                     AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1));
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY),
            MI.getOperand(0).getReg())
        .addReg(AArch64::X0);
  } else
    BuildMI(*BB, MI, MI.getDebugLoc(), TII->get(TargetOpcode::COPY),
            MI.getOperand(0).getReg())
        .addReg(AArch64::XZR);
  BB->remove_instr(&MI);
  return BB;
}

MachineBasicBlock *AArch64TargetLowering::EmitInstrWithCustomInserter(
    MachineInstr &MI, MachineBasicBlock *BB) const {

  int SMEOrigInstr = AArch64::getSMEPseudoMap(MI.getOpcode());
  if (SMEOrigInstr != -1) {
    const TargetInstrInfo *TII = Subtarget->getInstrInfo();
    uint64_t SMEMatrixType =
        TII->get(MI.getOpcode()).TSFlags & AArch64::SMEMatrixTypeMask;
    switch (SMEMatrixType) {
    case (AArch64::SMEMatrixArray):
      return EmitZAInstr(SMEOrigInstr, AArch64::ZA, MI, BB);
    case (AArch64::SMEMatrixTileB):
      return EmitZAInstr(SMEOrigInstr, AArch64::ZAB0, MI, BB);
    case (AArch64::SMEMatrixTileH):
      return EmitZAInstr(SMEOrigInstr, AArch64::ZAH0, MI, BB);
    case (AArch64::SMEMatrixTileS):
      return EmitZAInstr(SMEOrigInstr, AArch64::ZAS0, MI, BB);
    case (AArch64::SMEMatrixTileD):
      return EmitZAInstr(SMEOrigInstr, AArch64::ZAD0, MI, BB);
    case (AArch64::SMEMatrixTileQ):
      return EmitZAInstr(SMEOrigInstr, AArch64::ZAQ0, MI, BB);
    }
  }

  switch (MI.getOpcode()) {
  default:
#ifndef NDEBUG
    MI.dump();
#endif
    llvm_unreachable("Unexpected instruction for custom inserter!");
  case AArch64::InitTPIDR2Obj:
    return EmitInitTPIDR2Object(MI, BB);
  case AArch64::AllocateZABuffer:
    return EmitAllocateZABuffer(MI, BB);
  case AArch64::AllocateSMESaveBuffer:
    return EmitAllocateSMESaveBuffer(MI, BB);
  case AArch64::GetSMESaveSize:
    return EmitGetSMESaveSize(MI, BB);
  case AArch64::F128CSEL:
    return EmitF128CSEL(MI, BB);
  case TargetOpcode::STATEPOINT:
    // STATEPOINT is a pseudo instruction which has no implicit defs/uses
    // while bl call instruction (where statepoint will be lowered at the end)
    // has implicit def. This def is early-clobber as it will be set at
    // the moment of the call and earlier than any use is read.
    // Add this implicit dead def here as a workaround.
    MI.addOperand(*MI.getMF(),
                  MachineOperand::CreateReg(
                      AArch64::LR, /*isDef*/ true,
                      /*isImp*/ true, /*isKill*/ false, /*isDead*/ true,
                      /*isUndef*/ false, /*isEarlyClobber*/ true));
    [[fallthrough]];
  case TargetOpcode::STACKMAP:
  case TargetOpcode::PATCHPOINT:
    return emitPatchPoint(MI, BB);

  case TargetOpcode::PATCHABLE_EVENT_CALL:
  case TargetOpcode::PATCHABLE_TYPED_EVENT_CALL:
    return BB;

  case AArch64::CATCHRET:
    return EmitLoweredCatchRet(MI, BB);

  case AArch64::PROBED_STACKALLOC_DYN:
    return EmitDynamicProbedAlloc(MI, BB);

  case AArch64::LD1_MXIPXX_H_PSEUDO_B:
    return EmitTileLoad(AArch64::LD1_MXIPXX_H_B, AArch64::ZAB0, MI, BB);
  case AArch64::LD1_MXIPXX_H_PSEUDO_H:
    return EmitTileLoad(AArch64::LD1_MXIPXX_H_H, AArch64::ZAH0, MI, BB);
  case AArch64::LD1_MXIPXX_H_PSEUDO_S:
    return EmitTileLoad(AArch64::LD1_MXIPXX_H_S, AArch64::ZAS0, MI, BB);
  case AArch64::LD1_MXIPXX_H_PSEUDO_D:
    return EmitTileLoad(AArch64::LD1_MXIPXX_H_D, AArch64::ZAD0, MI, BB);
  case AArch64::LD1_MXIPXX_H_PSEUDO_Q:
    return EmitTileLoad(AArch64::LD1_MXIPXX_H_Q, AArch64::ZAQ0, MI, BB);
  case AArch64::LD1_MXIPXX_V_PSEUDO_B:
    return EmitTileLoad(AArch64::LD1_MXIPXX_V_B, AArch64::ZAB0, MI, BB);
  case AArch64::LD1_MXIPXX_V_PSEUDO_H:
    return EmitTileLoad(AArch64::LD1_MXIPXX_V_H, AArch64::ZAH0, MI, BB);
  case AArch64::LD1_MXIPXX_V_PSEUDO_S:
    return EmitTileLoad(AArch64::LD1_MXIPXX_V_S, AArch64::ZAS0, MI, BB);
  case AArch64::LD1_MXIPXX_V_PSEUDO_D:
    return EmitTileLoad(AArch64::LD1_MXIPXX_V_D, AArch64::ZAD0, MI, BB);
  case AArch64::LD1_MXIPXX_V_PSEUDO_Q:
    return EmitTileLoad(AArch64::LD1_MXIPXX_V_Q, AArch64::ZAQ0, MI, BB);
  case AArch64::LDR_ZA_PSEUDO:
    return EmitFill(MI, BB);
  case AArch64::LDR_TX_PSEUDO:
    return EmitZTInstr(MI, BB, AArch64::LDR_TX, /*Op0IsDef=*/true);
  case AArch64::STR_TX_PSEUDO:
    return EmitZTInstr(MI, BB, AArch64::STR_TX, /*Op0IsDef=*/false);
  case AArch64::ZERO_M_PSEUDO:
    return EmitZero(MI, BB);
  case AArch64::ZERO_T_PSEUDO:
    return EmitZTInstr(MI, BB, AArch64::ZERO_T, /*Op0IsDef=*/true);
  case AArch64::MOVT_TIZ_PSEUDO:
    return EmitZTInstr(MI, BB, AArch64::MOVT_TIZ, /*Op0IsDef=*/true);
  }
}

//===----------------------------------------------------------------------===//
// AArch64 Lowering private implementation.
//===----------------------------------------------------------------------===//

//===----------------------------------------------------------------------===//
// Lowering Code
//===----------------------------------------------------------------------===//

// Forward declarations of SVE fixed length lowering helpers
static EVT getContainerForFixedLengthVector(SelectionDAG &DAG, EVT VT);
static SDValue convertToScalableVector(SelectionDAG &DAG, EVT VT, SDValue V);
static SDValue convertFromScalableVector(SelectionDAG &DAG, EVT VT, SDValue V);
static SDValue convertFixedMaskToScalableVector(SDValue Mask,
                                                SelectionDAG &DAG);
static SDValue getPredicateForVector(SelectionDAG &DAG, SDLoc &DL, EVT VT);
static SDValue getPredicateForScalableVector(SelectionDAG &DAG, SDLoc &DL,
                                             EVT VT);

/// isZerosVector - Check whether SDNode N is a zero-filled vector.
static bool isZerosVector(const SDNode *N) {
  // Look through a bit convert.
  while (N->getOpcode() == ISD::BITCAST)
    N = N->getOperand(0).getNode();

  if (ISD::isConstantSplatVectorAllZeros(N))
    return true;

  if (N->getOpcode() != AArch64ISD::DUP)
    return false;

  auto Opnd0 = N->getOperand(0);
  return isNullConstant(Opnd0) || isNullFPConstant(Opnd0);
}

/// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
/// CC
static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
  switch (CC) {
  default:
    llvm_unreachable("Unknown condition code!");
  case ISD::SETNE:
    return AArch64CC::NE;
  case ISD::SETEQ:
    return AArch64CC::EQ;
  case ISD::SETGT:
    return AArch64CC::GT;
  case ISD::SETGE:
    return AArch64CC::GE;
  case ISD::SETLT:
    return AArch64CC::LT;
  case ISD::SETLE:
    return AArch64CC::LE;
  case ISD::SETUGT:
    return AArch64CC::HI;
  case ISD::SETUGE:
    return AArch64CC::HS;
  case ISD::SETULT:
    return AArch64CC::LO;
  case ISD::SETULE:
    return AArch64CC::LS;
  }
}

/// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
static void changeFPCCToAArch64CC(ISD::CondCode CC,
                                  AArch64CC::CondCode &CondCode,
                                  AArch64CC::CondCode &CondCode2) {
  CondCode2 = AArch64CC::AL;
  switch (CC) {
  default:
    llvm_unreachable("Unknown FP condition!");
  case ISD::SETEQ:
  case ISD::SETOEQ:
    CondCode = AArch64CC::EQ;
    break;
  case ISD::SETGT:
  case ISD::SETOGT:
    CondCode = AArch64CC::GT;
    break;
  case ISD::SETGE:
  case ISD::SETOGE:
    CondCode = AArch64CC::GE;
    break;
  case ISD::SETOLT:
    CondCode = AArch64CC::MI;
    break;
  case ISD::SETOLE:
    CondCode = AArch64CC::LS;
    break;
  case ISD::SETONE:
    CondCode = AArch64CC::MI;
    CondCode2 = AArch64CC::GT;
    break;
  case ISD::SETO:
    CondCode = AArch64CC::VC;
    break;
  case ISD::SETUO:
    CondCode = AArch64CC::VS;
    break;
  case ISD::SETUEQ:
    CondCode = AArch64CC::EQ;
    CondCode2 = AArch64CC::VS;
    break;
  case ISD::SETUGT:
    CondCode = AArch64CC::HI;
    break;
  case ISD::SETUGE:
    CondCode = AArch64CC::PL;
    break;
  case ISD::SETLT:
  case ISD::SETULT:
    CondCode = AArch64CC::LT;
    break;
  case ISD::SETLE:
  case ISD::SETULE:
    CondCode = AArch64CC::LE;
    break;
  case ISD::SETNE:
  case ISD::SETUNE:
    CondCode = AArch64CC::NE;
    break;
  }
}

/// Convert a DAG fp condition code to an AArch64 CC.
/// This differs from changeFPCCToAArch64CC in that it returns cond codes that
/// should be AND'ed instead of OR'ed.
static void changeFPCCToANDAArch64CC(ISD::CondCode CC,
                                     AArch64CC::CondCode &CondCode,
                                     AArch64CC::CondCode &CondCode2) {
  CondCode2 = AArch64CC::AL;
  switch (CC) {
  default:
    changeFPCCToAArch64CC(CC, CondCode, CondCode2);
    assert(CondCode2 == AArch64CC::AL);
    break;
  case ISD::SETONE:
    // (a one b)
    // == ((a olt b) || (a ogt b))
    // == ((a ord b) && (a une b))
    CondCode = AArch64CC::VC;
    CondCode2 = AArch64CC::NE;
    break;
  case ISD::SETUEQ:
    // (a ueq b)
    // == ((a uno b) || (a oeq b))
    // == ((a ule b) && (a uge b))
    CondCode = AArch64CC::PL;
    CondCode2 = AArch64CC::LE;
    break;
  }
}

/// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
/// CC usable with the vector instructions. Fewer operations are available
/// without a real NZCV register, so we have to use less efficient combinations
/// to get the same effect.
static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
                                        AArch64CC::CondCode &CondCode,
                                        AArch64CC::CondCode &CondCode2,
                                        bool &Invert) {
  Invert = false;
  switch (CC) {
  default:
    // Mostly the scalar mappings work fine.
    changeFPCCToAArch64CC(CC, CondCode, CondCode2);
    break;
  case ISD::SETUO:
    Invert = true;
    [[fallthrough]];
  case ISD::SETO:
    CondCode = AArch64CC::MI;
    CondCode2 = AArch64CC::GE;
    break;
  case ISD::SETUEQ:
  case ISD::SETULT:
  case ISD::SETULE:
  case ISD::SETUGT:
  case ISD::SETUGE:
    // All of the compare-mask comparisons are ordered, but we can switch
    // between the two by a double inversion. E.g. ULE == !OGT.
    Invert = true;
    changeFPCCToAArch64CC(getSetCCInverse(CC, /* FP inverse */ MVT::f32),
                          CondCode, CondCode2);
    break;
  }
}

static bool isLegalArithImmed(uint64_t C) {
  // Matches AArch64DAGToDAGISel::SelectArithImmed().
  bool IsLegal = (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
  LLVM_DEBUG(dbgs() << "Is imm " << C
                    << " legal: " << (IsLegal ? "yes\n" : "no\n"));
  return IsLegal;
}

bool isLegalCmpImmed(APInt C) {
  // Works for negative immediates too, as it can be written as an ADDS
  // instruction with a negated immediate.
  return isLegalArithImmed(C.abs().getZExtValue());
}

static bool isSafeSignedCMN(SDValue Op, SelectionDAG &DAG) {
  // 0 - INT_MIN sign wraps, so no signed wrap means cmn is safe.
  if (Op->getFlags().hasNoSignedWrap())
    return true;

  // We can still figure out if the second operand is safe to use
  // in a CMN instruction by checking if it is known to be not the minimum
  // signed value. If it is not, then we can safely use CMN.
  // Note: We can eventually remove this check and simply rely on
  // Op->getFlags().hasNoSignedWrap() once SelectionDAG/ISelLowering
  // consistently sets them appropriately when making said nodes.

  KnownBits KnownSrc = DAG.computeKnownBits(Op.getOperand(1));
  return !KnownSrc.getSignedMinValue().isMinSignedValue();
}

// Can a (CMP op1, (sub 0, op2) be turned into a CMN instruction on
// the grounds that "op1 - (-op2) == op1 + op2" ? Not always, the C and V flags
// can be set differently by this operation. It comes down to whether
// "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
// everything is fine. If not then the optimization is wrong. Thus general
// comparisons are only valid if op2 != 0 and op2 != INT_MIN.
//
// So, finally, the only LLVM-native comparisons that don't mention C or V
// are the ones that aren't unsigned comparisons. They're the only ones we can
// safely use CMN for in the absence of information about op2.
static bool isCMN(SDValue Op, ISD::CondCode CC, SelectionDAG &DAG) {
  return Op.getOpcode() == ISD::SUB && isNullConstant(Op.getOperand(0)) &&
         (isIntEqualitySetCC(CC) ||
          (isUnsignedIntSetCC(CC) && DAG.isKnownNeverZero(Op.getOperand(1))) ||
          (isSignedIntSetCC(CC) && isSafeSignedCMN(Op, DAG)));
}

static SDValue emitStrictFPComparison(SDValue LHS, SDValue RHS, const SDLoc &DL,
                                      SelectionDAG &DAG, SDValue Chain,
                                      bool IsSignaling) {
  EVT VT = LHS.getValueType();
  assert(VT != MVT::f128);

  const bool FullFP16 = DAG.getSubtarget<AArch64Subtarget>().hasFullFP16();

  if ((VT == MVT::f16 && !FullFP16) || VT == MVT::bf16) {
    LHS = DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {MVT::f32, MVT::Other},
                      {Chain, LHS});
    RHS = DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {MVT::f32, MVT::Other},
                      {LHS.getValue(1), RHS});
    Chain = RHS.getValue(1);
  }
  unsigned Opcode =
      IsSignaling ? AArch64ISD::STRICT_FCMPE : AArch64ISD::STRICT_FCMP;
  return DAG.getNode(Opcode, DL, {MVT::i32, MVT::Other}, {Chain, LHS, RHS});
}

static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
                              const SDLoc &DL, SelectionDAG &DAG) {
  EVT VT = LHS.getValueType();
  const bool FullFP16 = DAG.getSubtarget<AArch64Subtarget>().hasFullFP16();

  if (VT.isFloatingPoint()) {
    assert(VT != MVT::f128);
    if ((VT == MVT::f16 && !FullFP16) || VT == MVT::bf16) {
      LHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, LHS);
      RHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, RHS);
    }
    return DAG.getNode(AArch64ISD::FCMP, DL, MVT::i32, LHS, RHS);
  }

  // The CMP instruction is just an alias for SUBS, and representing it as
  // SUBS means that it's possible to get CSE with subtract operations.
  // A later phase can perform the optimization of setting the destination
  // register to WZR/XZR if it ends up being unused.
  unsigned Opcode = AArch64ISD::SUBS;

  if (isCMN(RHS, CC, DAG)) {
    // Can we combine a (CMP op1, (sub 0, op2) into a CMN instruction ?
    Opcode = AArch64ISD::ADDS;
    RHS = RHS.getOperand(1);
  } else if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) &&
             isIntEqualitySetCC(CC)) {
    // As we are looking for EQ/NE compares, the operands can be commuted ; can
    // we combine a (CMP (sub 0, op1), op2) into a CMN instruction ?
    Opcode = AArch64ISD::ADDS;
    LHS = LHS.getOperand(1);
  } else if (isNullConstant(RHS) && !isUnsignedIntSetCC(CC)) {
    if (LHS.getOpcode() == ISD::AND) {
      // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
      // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
      // of the signed comparisons.
      const SDValue ANDSNode =
          DAG.getNode(AArch64ISD::ANDS, DL, DAG.getVTList(VT, MVT_CC),
                      LHS.getOperand(0), LHS.getOperand(1));
      // Replace all users of (and X, Y) with newly generated (ands X, Y)
      DAG.ReplaceAllUsesWith(LHS, ANDSNode);
      return ANDSNode.getValue(1);
    } else if (LHS.getOpcode() == AArch64ISD::ANDS) {
      // Use result of ANDS
      return LHS.getValue(1);
    }
  }

  return DAG.getNode(Opcode, DL, DAG.getVTList(VT, MVT_CC), LHS, RHS)
      .getValue(1);
}

/// \defgroup AArch64CCMP CMP;CCMP matching
///
/// These functions deal with the formation of CMP;CCMP;... sequences.
/// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
/// a comparison. They set the NZCV flags to a predefined value if their
/// predicate is false. This allows to express arbitrary conjunctions, for
/// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B)))"
/// expressed as:
///   cmp A
///   ccmp B, inv(CB), CA
///   check for CB flags
///
/// This naturally lets us implement chains of AND operations with SETCC
/// operands. And we can even implement some other situations by transforming
/// them:
///   - We can implement (NEG SETCC) i.e. negating a single comparison by
///     negating the flags used in a CCMP/FCCMP operations.
///   - We can negate the result of a whole chain of CMP/CCMP/FCCMP operations
///     by negating the flags we test for afterwards. i.e.
///     NEG (CMP CCMP CCCMP ...) can be implemented.
///   - Note that we can only ever negate all previously processed results.
///     What we can not implement by flipping the flags to test is a negation
///     of two sub-trees (because the negation affects all sub-trees emitted so
///     far, so the 2nd sub-tree we emit would also affect the first).
/// With those tools we can implement some OR operations:
///   - (OR (SETCC A) (SETCC B)) can be implemented via:
///     NEG (AND (NEG (SETCC A)) (NEG (SETCC B)))
///   - After transforming OR to NEG/AND combinations we may be able to use NEG
///     elimination rules from earlier to implement the whole thing as a
///     CCMP/FCCMP chain.
///
/// As complete example:
///     or (or (setCA (cmp A)) (setCB (cmp B)))
///        (and (setCC (cmp C)) (setCD (cmp D)))"
/// can be reassociated to:
///     or (and (setCC (cmp C)) setCD (cmp D))
//         (or (setCA (cmp A)) (setCB (cmp B)))
/// can be transformed to:
///     not (and (not (and (setCC (cmp C)) (setCD (cmp D))))
///              (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
/// which can be implemented as:
///   cmp C
///   ccmp D, inv(CD), CC
///   ccmp A, CA, inv(CD)
///   ccmp B, CB, inv(CA)
///   check for CB flags
///
/// A counterexample is "or (and A B) (and C D)" which translates to
/// not (and (not (and (not A) (not B))) (not (and (not C) (not D)))), we
/// can only implement 1 of the inner (not) operations, but not both!
/// @{

/// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate.
static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
                                         ISD::CondCode CC, SDValue CCOp,
                                         AArch64CC::CondCode Predicate,
                                         AArch64CC::CondCode OutCC,
                                         const SDLoc &DL, SelectionDAG &DAG) {
  unsigned Opcode = 0;
  const bool FullFP16 = DAG.getSubtarget<AArch64Subtarget>().hasFullFP16();

  if (LHS.getValueType().isFloatingPoint()) {
    assert(LHS.getValueType() != MVT::f128);
    if ((LHS.getValueType() == MVT::f16 && !FullFP16) ||
        LHS.getValueType() == MVT::bf16) {
      LHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, LHS);
      RHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, RHS);
    }
    Opcode = AArch64ISD::FCCMP;
  } else if (ConstantSDNode *Const = dyn_cast<ConstantSDNode>(RHS)) {
    APInt Imm = Const->getAPIntValue();
    if (Imm.isNegative() && Imm.sgt(-32)) {
      Opcode = AArch64ISD::CCMN;
      RHS = DAG.getConstant(Imm.abs(), DL, Const->getValueType(0));
    }
  } else if (isCMN(RHS, CC, DAG)) {
    Opcode = AArch64ISD::CCMN;
    RHS = RHS.getOperand(1);
  } else if (LHS.getOpcode() == ISD::SUB && isNullConstant(LHS.getOperand(0)) &&
             isIntEqualitySetCC(CC)) {
    // As we are looking for EQ/NE compares, the operands can be commuted ; can
    // we combine a (CCMP (sub 0, op1), op2) into a CCMN instruction ?
    Opcode = AArch64ISD::CCMN;
    LHS = LHS.getOperand(1);
  }
  if (Opcode == 0)
    Opcode = AArch64ISD::CCMP;

  SDValue Condition = DAG.getConstant(Predicate, DL, MVT_CC);
  AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
  unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
  SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);
  return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
}

/// Returns true if @p Val is a tree of AND/OR/SETCC operations that can be
/// expressed as a conjunction. See \ref AArch64CCMP.
/// \param CanNegate    Set to true if we can negate the whole sub-tree just by
///                     changing the conditions on the SETCC tests.
///                     (this means we can call emitConjunctionRec() with
///                      Negate==true on this sub-tree)
/// \param MustBeFirst  Set to true if this subtree needs to be negated and we
///                     cannot do the negation naturally. We are required to
///                     emit the subtree first in this case.
/// \param WillNegate   Is true if are called when the result of this
///                     subexpression must be negated. This happens when the
///                     outer expression is an OR. We can use this fact to know
///                     that we have a double negation (or (or ...) ...) that
///                     can be implemented for free.
static bool canEmitConjunction(const SDValue Val, bool &CanNegate,
                               bool &MustBeFirst, bool WillNegate,
                               unsigned Depth = 0) {
  if (!Val.hasOneUse())
    return false;
  unsigned Opcode = Val->getOpcode();
  if (Opcode == ISD::SETCC) {
    if (Val->getOperand(0).getValueType() == MVT::f128)
      return false;
    CanNegate = true;
    MustBeFirst = false;
    return true;
  }
  // Protect against exponential runtime and stack overflow.
  if (Depth > 6)
    return false;
  if (Opcode == ISD::AND || Opcode == ISD::OR) {
    bool IsOR = Opcode == ISD::OR;
    SDValue O0 = Val->getOperand(0);
    SDValue O1 = Val->getOperand(1);
    bool CanNegateL;
    bool MustBeFirstL;
    if (!canEmitConjunction(O0, CanNegateL, MustBeFirstL, IsOR, Depth+1))
      return false;
    bool CanNegateR;
    bool MustBeFirstR;
    if (!canEmitConjunction(O1, CanNegateR, MustBeFirstR, IsOR, Depth+1))
      return false;

    if (MustBeFirstL && MustBeFirstR)
      return false;

    if (IsOR) {
      // For an OR expression we need to be able to naturally negate at least
      // one side or we cannot do the transformation at all.
      if (!CanNegateL && !CanNegateR)
        return false;
      // If we the result of the OR will be negated and we can naturally negate
      // the leafs, then this sub-tree as a whole negates naturally.
      CanNegate = WillNegate && CanNegateL && CanNegateR;
      // If we cannot naturally negate the whole sub-tree, then this must be
      // emitted first.
      MustBeFirst = !CanNegate;
    } else {
      assert(Opcode == ISD::AND && "Must be OR or AND");
      // We cannot naturally negate an AND operation.
      CanNegate = false;
      MustBeFirst = MustBeFirstL || MustBeFirstR;
    }
    return true;
  }
  return false;
}

/// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
/// of CCMP/CFCMP ops. See @ref AArch64CCMP.
/// Tries to transform the given i1 producing node @p Val to a series compare
/// and conditional compare operations. @returns an NZCV flags producing node
/// and sets @p OutCC to the flags that should be tested or returns SDValue() if
/// transformation was not possible.
/// \p Negate is true if we want this sub-tree being negated just by changing
/// SETCC conditions.
static SDValue emitConjunctionRec(SelectionDAG &DAG, SDValue Val,
    AArch64CC::CondCode &OutCC, bool Negate, SDValue CCOp,
    AArch64CC::CondCode Predicate) {
  // We're at a tree leaf, produce a conditional comparison operation.
  unsigned Opcode = Val->getOpcode();
  if (Opcode == ISD::SETCC) {
    SDValue LHS = Val->getOperand(0);
    SDValue RHS = Val->getOperand(1);
    ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
    bool isInteger = LHS.getValueType().isInteger();
    if (Negate)
      CC = getSetCCInverse(CC, LHS.getValueType());
    SDLoc DL(Val);
    // Determine OutCC and handle FP special case.
    if (isInteger) {
      OutCC = changeIntCCToAArch64CC(CC);
    } else {
      assert(LHS.getValueType().isFloatingPoint());
      AArch64CC::CondCode ExtraCC;
      changeFPCCToANDAArch64CC(CC, OutCC, ExtraCC);
      // Some floating point conditions can't be tested with a single condition
      // code. Construct an additional comparison in this case.
      if (ExtraCC != AArch64CC::AL) {
        SDValue ExtraCmp;
        if (!CCOp.getNode())
          ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
        else
          ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, Predicate,
                                               ExtraCC, DL, DAG);
        CCOp = ExtraCmp;
        Predicate = ExtraCC;
      }
    }

    // Produce a normal comparison if we are first in the chain
    if (!CCOp)
      return emitComparison(LHS, RHS, CC, DL, DAG);
    // Otherwise produce a ccmp.
    return emitConditionalComparison(LHS, RHS, CC, CCOp, Predicate, OutCC, DL,
                                     DAG);
  }
  assert(Val->hasOneUse() && "Valid conjunction/disjunction tree");

  bool IsOR = Opcode == ISD::OR;

  SDValue LHS = Val->getOperand(0);
  bool CanNegateL;
  bool MustBeFirstL;
  bool ValidL = canEmitConjunction(LHS, CanNegateL, MustBeFirstL, IsOR);
  assert(ValidL && "Valid conjunction/disjunction tree");
  (void)ValidL;

  SDValue RHS = Val->getOperand(1);
  bool CanNegateR;
  bool MustBeFirstR;
  bool ValidR = canEmitConjunction(RHS, CanNegateR, MustBeFirstR, IsOR);
  assert(ValidR && "Valid conjunction/disjunction tree");
  (void)ValidR;

  // Swap sub-tree that must come first to the right side.
  if (MustBeFirstL) {
    assert(!MustBeFirstR && "Valid conjunction/disjunction tree");
    std::swap(LHS, RHS);
    std::swap(CanNegateL, CanNegateR);
    std::swap(MustBeFirstL, MustBeFirstR);
  }

  bool NegateR;
  bool NegateAfterR;
  bool NegateL;
  bool NegateAfterAll;
  if (Opcode == ISD::OR) {
    // Swap the sub-tree that we can negate naturally to the left.
    if (!CanNegateL) {
      assert(CanNegateR && "at least one side must be negatable");
      assert(!MustBeFirstR && "invalid conjunction/disjunction tree");
      assert(!Negate);
      std::swap(LHS, RHS);
      NegateR = false;
      NegateAfterR = true;
    } else {
      // Negate the left sub-tree if possible, otherwise negate the result.
      NegateR = CanNegateR;
      NegateAfterR = !CanNegateR;
    }
    NegateL = true;
    NegateAfterAll = !Negate;
  } else {
    assert(Opcode == ISD::AND && "Valid conjunction/disjunction tree");
    assert(!Negate && "Valid conjunction/disjunction tree");

    NegateL = false;
    NegateR = false;
    NegateAfterR = false;
    NegateAfterAll = false;
  }

  // Emit sub-trees.
  AArch64CC::CondCode RHSCC;
  SDValue CmpR = emitConjunctionRec(DAG, RHS, RHSCC, NegateR, CCOp, Predicate);
  if (NegateAfterR)
    RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
  SDValue CmpL = emitConjunctionRec(DAG, LHS, OutCC, NegateL, CmpR, RHSCC);
  if (NegateAfterAll)
    OutCC = AArch64CC::getInvertedCondCode(OutCC);
  return CmpL;
}

/// Emit expression as a conjunction (a series of CCMP/CFCMP ops).
/// In some cases this is even possible with OR operations in the expression.
/// See \ref AArch64CCMP.
/// \see emitConjunctionRec().
static SDValue emitConjunction(SelectionDAG &DAG, SDValue Val,
                               AArch64CC::CondCode &OutCC) {
  bool DummyCanNegate;
  bool DummyMustBeFirst;
  if (!canEmitConjunction(Val, DummyCanNegate, DummyMustBeFirst, false))
    return SDValue();

  return emitConjunctionRec(DAG, Val, OutCC, false, SDValue(), AArch64CC::AL);
}

/// @}

/// Returns how profitable it is to fold a comparison's operand's shift and/or
/// extension operations.
static unsigned getCmpOperandFoldingProfit(SDValue Op) {
  auto isSupportedExtend = [&](SDValue V) {
    if (V.getOpcode() == ISD::SIGN_EXTEND_INREG)
      return true;

    if (V.getOpcode() == ISD::AND)
      if (ConstantSDNode *MaskCst = dyn_cast<ConstantSDNode>(V.getOperand(1))) {
        uint64_t Mask = MaskCst->getZExtValue();
        return (Mask == 0xFF || Mask == 0xFFFF || Mask == 0xFFFFFFFF);
      }

    return false;
  };

  if (!Op.hasOneUse())
    return 0;

  if (isSupportedExtend(Op))
    return 1;

  unsigned Opc = Op.getOpcode();
  if (Opc == ISD::SHL || Opc == ISD::SRL || Opc == ISD::SRA)
    if (ConstantSDNode *ShiftCst = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
      uint64_t Shift = ShiftCst->getZExtValue();
      if (isSupportedExtend(Op.getOperand(0)))
        return (Shift <= 4) ? 2 : 1;
      EVT VT = Op.getValueType();
      if ((VT == MVT::i32 && Shift <= 31) || (VT == MVT::i64 && Shift <= 63))
        return 1;
    }

  return 0;
}

// emitComparison() converts comparison with one or negative one to comparison
// with 0. Note that this only works for signed comparisons because of how ANDS
// works.
static bool shouldBeAdjustedToZero(SDValue LHS, APInt C, ISD::CondCode &CC) {
  // Only works for ANDS and AND.
  if (LHS.getOpcode() != ISD::AND && LHS.getOpcode() != AArch64ISD::ANDS)
    return false;

  if (C.isOne() && (CC == ISD::SETLT || CC == ISD::SETGE)) {
    CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
    return true;
  }

  if (C.isAllOnes() && (CC == ISD::SETLE || CC == ISD::SETGT)) {
    CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
    return true;
  }

  return false;
}

static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
                             SDValue &AArch64cc, SelectionDAG &DAG,
                             const SDLoc &DL) {
  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
    EVT VT = RHS.getValueType();
    APInt C = RHSC->getAPIntValue();
    // shouldBeAdjustedToZero is a special case to better fold with
    // emitComparison().
    if (shouldBeAdjustedToZero(LHS, C, CC)) {
      // Adjust the constant to zero.
      // CC has already been adjusted.
      RHS = DAG.getConstant(0, DL, VT);
    } else if (!isLegalCmpImmed(C)) {
      // Constant does not fit, try adjusting it by one?
      switch (CC) {
      default:
        break;
      case ISD::SETLT:
      case ISD::SETGE:
        if (!C.isMinSignedValue()) {
          APInt CMinusOne = C - 1;
          if (isLegalCmpImmed(CMinusOne)) {
            CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
            RHS = DAG.getConstant(CMinusOne, DL, VT);
          }
        }
        break;
      case ISD::SETULT:
      case ISD::SETUGE:
        if (!C.isZero()) {
          APInt CMinusOne = C - 1;
          if (isLegalCmpImmed(CMinusOne)) {
            CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
            RHS = DAG.getConstant(CMinusOne, DL, VT);
          }
        }
        break;
      case ISD::SETLE:
      case ISD::SETGT:
        if (!C.isMaxSignedValue()) {
          APInt CPlusOne = C + 1;
          if (isLegalCmpImmed(CPlusOne)) {
            CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
            RHS = DAG.getConstant(CPlusOne, DL, VT);
          }
        }
        break;
      case ISD::SETULE:
      case ISD::SETUGT:
        if (!C.isAllOnes()) {
          APInt CPlusOne = C + 1;
          if (isLegalCmpImmed(CPlusOne)) {
            CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
            RHS = DAG.getConstant(CPlusOne, DL, VT);
          }
        }
        break;
      }
    }
  }

  // Comparisons are canonicalized so that the RHS operand is simpler than the
  // LHS one, the extreme case being when RHS is an immediate. However, AArch64
  // can fold some shift+extend operations on the RHS operand, so swap the
  // operands if that can be done.
  //
  // For example:
  //    lsl     w13, w11, #1
  //    cmp     w13, w12
  // can be turned into:
  //    cmp     w12, w11, lsl #1
  if (!isa<ConstantSDNode>(RHS) || !isLegalCmpImmed(RHS->getAsAPIntVal())) {
    bool LHSIsCMN = isCMN(LHS, CC, DAG);
    bool RHSIsCMN = isCMN(RHS, CC, DAG);
    SDValue TheLHS = LHSIsCMN ? LHS.getOperand(1) : LHS;
    SDValue TheRHS = RHSIsCMN ? RHS.getOperand(1) : RHS;

    if (getCmpOperandFoldingProfit(TheLHS) + (LHSIsCMN ? 1 : 0) >
        getCmpOperandFoldingProfit(TheRHS) + (RHSIsCMN ? 1 : 0)) {
      std::swap(LHS, RHS);
      CC = ISD::getSetCCSwappedOperands(CC);
    }
  }

  SDValue Cmp;
  AArch64CC::CondCode AArch64CC;
  if (isIntEqualitySetCC(CC) && isa<ConstantSDNode>(RHS)) {
    const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);

    // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
    // For the i8 operand, the largest immediate is 255, so this can be easily
    // encoded in the compare instruction. For the i16 operand, however, the
    // largest immediate cannot be encoded in the compare.
    // Therefore, use a sign extending load and cmn to avoid materializing the
    // -1 constant. For example,
    // movz w1, #65535
    // ldrh w0, [x0, #0]
    // cmp w0, w1
    // >
    // ldrsh w0, [x0, #0]
    // cmn w0, #1
    // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
    // if and only if (sext LHS) == (sext RHS). The checks are in place to
    // ensure both the LHS and RHS are truly zero extended and to make sure the
    // transformation is profitable.
    if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
        cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
        cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
        LHS.getNode()->hasNUsesOfValue(1, 0)) {
      int16_t ValueofRHS = RHS->getAsZExtVal();
      if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
        SDValue SExt =
            DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, LHS.getValueType(), LHS,
                        DAG.getValueType(MVT::i16));
        Cmp = emitComparison(
            SExt, DAG.getSignedConstant(ValueofRHS, DL, RHS.getValueType()), CC,
            DL, DAG);
        AArch64CC = changeIntCCToAArch64CC(CC);
      }
    }

    if (!Cmp && (RHSC->isZero() || RHSC->isOne())) {
      if ((Cmp = emitConjunction(DAG, LHS, AArch64CC))) {
        if ((CC == ISD::SETNE) ^ RHSC->isZero())
          AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
      }
    }
  }

  if (!Cmp) {
    Cmp = emitComparison(LHS, RHS, CC, DL, DAG);
    AArch64CC = changeIntCCToAArch64CC(CC);
  }
  AArch64cc = DAG.getConstant(AArch64CC, DL, MVT_CC);
  return Cmp;
}

static std::pair<SDValue, SDValue>
getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
  assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
         "Unsupported value type");
  SDValue Value, Overflow;
  SDLoc DL(Op);
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  unsigned Opc = 0;
  switch (Op.getOpcode()) {
  default:
    llvm_unreachable("Unknown overflow instruction!");
  case ISD::SADDO:
    Opc = AArch64ISD::ADDS;
    CC = AArch64CC::VS;
    break;
  case ISD::UADDO:
    Opc = AArch64ISD::ADDS;
    CC = AArch64CC::HS;
    break;
  case ISD::SSUBO:
    Opc = AArch64ISD::SUBS;
    CC = AArch64CC::VS;
    break;
  case ISD::USUBO:
    Opc = AArch64ISD::SUBS;
    CC = AArch64CC::LO;
    break;
  // Multiply needs a little bit extra work.
  case ISD::SMULO:
  case ISD::UMULO: {
    CC = AArch64CC::NE;
    bool IsSigned = Op.getOpcode() == ISD::SMULO;
    if (Op.getValueType() == MVT::i32) {
      // Extend to 64-bits, then perform a 64-bit multiply.
      unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
      LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
      RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
      SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
      Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Mul);

      // Check that the result fits into a 32-bit integer.
      SDVTList VTs = DAG.getVTList(MVT::i64, MVT_CC);
      if (IsSigned) {
        // cmp xreg, wreg, sxtw
        SDValue SExtMul = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Value);
        Overflow =
            DAG.getNode(AArch64ISD::SUBS, DL, VTs, Mul, SExtMul).getValue(1);
      } else {
        // tst xreg, #0xffffffff00000000
        SDValue UpperBits = DAG.getConstant(0xFFFFFFFF00000000, DL, MVT::i64);
        Overflow =
            DAG.getNode(AArch64ISD::ANDS, DL, VTs, Mul, UpperBits).getValue(1);
      }
      break;
    }
    assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
    // For the 64 bit multiply
    Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
    if (IsSigned) {
      SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
      SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
                                      DAG.getConstant(63, DL, MVT::i64));
      // It is important that LowerBits is last, otherwise the arithmetic
      // shift will not be folded into the compare (SUBS).
      SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
      Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
                     .getValue(1);
    } else {
      SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
      SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
      Overflow =
          DAG.getNode(AArch64ISD::SUBS, DL, VTs,
                      DAG.getConstant(0, DL, MVT::i64),
                      UpperBits).getValue(1);
    }
    break;
  }
  } // switch (...)

  if (Opc) {
    SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);

    // Emit the AArch64 operation with overflow check.
    Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
    Overflow = Value.getValue(1);
  }
  return std::make_pair(Value, Overflow);
}

SDValue AArch64TargetLowering::LowerXOR(SDValue Op, SelectionDAG &DAG) const {
  if (useSVEForFixedLengthVectorVT(Op.getValueType(),
                                   !Subtarget->isNeonAvailable()))
    return LowerToScalableOp(Op, DAG);

  SDValue Sel = Op.getOperand(0);
  SDValue Other = Op.getOperand(1);
  SDLoc DL(Sel);

  // If the operand is an overflow checking operation, invert the condition
  // code and kill the Not operation. I.e., transform:
  // (xor (overflow_op_bool, 1))
  //   -->
  // (csel 1, 0, invert(cc), overflow_op_bool)
  // ... which later gets transformed to just a cset instruction with an
  // inverted condition code, rather than a cset + eor sequence.
  if (isOneConstant(Other) && ISD::isOverflowIntrOpRes(Sel)) {
    // Only lower legal XALUO ops.
    if (!DAG.getTargetLoweringInfo().isTypeLegal(Sel->getValueType(0)))
      return SDValue();

    SDValue TVal = DAG.getConstant(1, DL, MVT::i32);
    SDValue FVal = DAG.getConstant(0, DL, MVT::i32);
    AArch64CC::CondCode CC;
    SDValue Value, Overflow;
    std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Sel.getValue(0), DAG);
    SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), DL, MVT::i32);
    return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
                       CCVal, Overflow);
  }
  // If neither operand is a SELECT_CC, give up.
  if (Sel.getOpcode() != ISD::SELECT_CC)
    std::swap(Sel, Other);
  if (Sel.getOpcode() != ISD::SELECT_CC)
    return Op;

  // The folding we want to perform is:
  // (xor x, (select_cc a, b, cc, 0, -1) )
  //   -->
  // (csel x, (xor x, -1), cc ...)
  //
  // The latter will get matched to a CSINV instruction.

  ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
  SDValue LHS = Sel.getOperand(0);
  SDValue RHS = Sel.getOperand(1);
  SDValue TVal = Sel.getOperand(2);
  SDValue FVal = Sel.getOperand(3);

  // FIXME: This could be generalized to non-integer comparisons.
  if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
    return Op;

  ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
  ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);

  // The values aren't constants, this isn't the pattern we're looking for.
  if (!CFVal || !CTVal)
    return Op;

  // We can commute the SELECT_CC by inverting the condition.  This
  // might be needed to make this fit into a CSINV pattern.
  if (CTVal->isAllOnes() && CFVal->isZero()) {
    std::swap(TVal, FVal);
    std::swap(CTVal, CFVal);
    CC = ISD::getSetCCInverse(CC, LHS.getValueType());
  }

  // If the constants line up, perform the transform!
  if (CTVal->isZero() && CFVal->isAllOnes()) {
    SDValue CCVal;
    SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, DL);

    FVal = Other;
    TVal = DAG.getNode(ISD::XOR, DL, Other.getValueType(), Other,
                       DAG.getAllOnesConstant(DL, Other.getValueType()));

    return DAG.getNode(AArch64ISD::CSEL, DL, Sel.getValueType(), FVal, TVal,
                       CCVal, Cmp);
  }

  return Op;
}

// If Invert is false, sets 'C' bit of NZCV to 0 if value is 0, else sets 'C'
// bit to 1. If Invert is true, sets 'C' bit of NZCV to 1 if value is 0, else
// sets 'C' bit to 0.
static SDValue valueToCarryFlag(SDValue Value, SelectionDAG &DAG, bool Invert) {
  SDLoc DL(Value);
  EVT VT = Value.getValueType();
  SDValue Op0 = Invert ? DAG.getConstant(0, DL, VT) : Value;
  SDValue Op1 = Invert ? Value : DAG.getConstant(1, DL, VT);
  SDValue Cmp =
      DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::Glue), Op0, Op1);
  return Cmp.getValue(1);
}

// If Invert is false, value is 1 if 'C' bit of NZCV is 1, else 0.
// If Invert is true, value is 0 if 'C' bit of NZCV is 1, else 1.
static SDValue carryFlagToValue(SDValue Glue, EVT VT, SelectionDAG &DAG,
                                bool Invert) {
  assert(Glue.getResNo() == 1);
  SDLoc DL(Glue);
  SDValue Zero = DAG.getConstant(0, DL, VT);
  SDValue One = DAG.getConstant(1, DL, VT);
  unsigned Cond = Invert ? AArch64CC::LO : AArch64CC::HS;
  SDValue CC = DAG.getConstant(Cond, DL, MVT::i32);
  return DAG.getNode(AArch64ISD::CSEL, DL, VT, One, Zero, CC, Glue);
}

// Value is 1 if 'V' bit of NZCV is 1, else 0
static SDValue overflowFlagToValue(SDValue Glue, EVT VT, SelectionDAG &DAG) {
  assert(Glue.getResNo() == 1);
  SDLoc DL(Glue);
  SDValue Zero = DAG.getConstant(0, DL, VT);
  SDValue One = DAG.getConstant(1, DL, VT);
  SDValue CC = DAG.getConstant(AArch64CC::VS, DL, MVT::i32);
  return DAG.getNode(AArch64ISD::CSEL, DL, VT, One, Zero, CC, Glue);
}

// This lowering is inefficient, but it will get cleaned up by
// `foldOverflowCheck`
static SDValue lowerADDSUBO_CARRY(SDValue Op, SelectionDAG &DAG,
                                  unsigned Opcode, bool IsSigned) {
  EVT VT0 = Op.getValue(0).getValueType();
  EVT VT1 = Op.getValue(1).getValueType();

  if (VT0 != MVT::i32 && VT0 != MVT::i64)
    return SDValue();

  bool InvertCarry = Opcode == AArch64ISD::SBCS;
  SDValue OpLHS = Op.getOperand(0);
  SDValue OpRHS = Op.getOperand(1);
  SDValue OpCarryIn = valueToCarryFlag(Op.getOperand(2), DAG, InvertCarry);

  SDLoc DL(Op);
  SDVTList VTs = DAG.getVTList(VT0, VT1);

  SDValue Sum = DAG.getNode(Opcode, DL, DAG.getVTList(VT0, MVT::Glue), OpLHS,
                            OpRHS, OpCarryIn);

  SDValue OutFlag =
      IsSigned ? overflowFlagToValue(Sum.getValue(1), VT1, DAG)
               : carryFlagToValue(Sum.getValue(1), VT1, DAG, InvertCarry);

  return DAG.getNode(ISD::MERGE_VALUES, DL, VTs, Sum, OutFlag);
}

static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
  // Let legalize expand this if it isn't a legal type yet.
  if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
    return SDValue();

  SDLoc DL(Op);
  AArch64CC::CondCode CC;
  // The actual operation that sets the overflow or carry flag.
  SDValue Value, Overflow;
  std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);

  // We use 0 and 1 as false and true values.
  SDValue TVal = DAG.getConstant(1, DL, MVT::i32);
  SDValue FVal = DAG.getConstant(0, DL, MVT::i32);

  // We use an inverted condition, because the conditional select is inverted
  // too. This will allow it to be selected to a single instruction:
  // CSINC Wd, WZR, WZR, invert(cond).
  SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), DL, MVT::i32);
  Overflow =
      DAG.getNode(AArch64ISD::CSEL, DL, MVT::i32, FVal, TVal, CCVal, Overflow);

  SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
  return DAG.getNode(ISD::MERGE_VALUES, DL, VTs, Value, Overflow);
}

// Prefetch operands are:
// 1: Address to prefetch
// 2: bool isWrite
// 3: int locality (0 = no locality ... 3 = extreme locality)
// 4: bool isDataCache
static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  unsigned IsWrite = Op.getConstantOperandVal(2);
  unsigned Locality = Op.getConstantOperandVal(3);
  unsigned IsData = Op.getConstantOperandVal(4);

  bool IsStream = !Locality;
  // When the locality number is set
  if (Locality) {
    // The front-end should have filtered out the out-of-range values
    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;
  }

  // built the mask value encoding the expected behavior.
  unsigned PrfOp = (IsWrite << 4) |     // Load/Store bit
                   (!IsData << 3) |     // IsDataCache bit
                   (Locality << 1) |    // Cache level bits
                   (unsigned)IsStream;  // Stream bit
  return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
                     DAG.getTargetConstant(PrfOp, DL, MVT::i32),
                     Op.getOperand(1));
}

// Converts SETCC (AND X Y) Z ULT -> SETCC (AND X (Y & ~(Z - 1)) 0 EQ when Y is
// a power of 2. This is then lowered to ANDS X (Y & ~(Z - 1)) instead of SUBS
// (AND X Y) Z which produces a better opt with EmitComparison
static void simplifySetCCIntoEq(ISD::CondCode &CC, SDValue &LHS, SDValue &RHS,
                                SelectionDAG &DAG, const SDLoc DL) {
  if (CC == ISD::SETULT && LHS.getOpcode() == ISD::AND && LHS->hasOneUse()) {
    ConstantSDNode *LHSConstOp = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
    ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(RHS);
    if (LHSConstOp && RHSConst) {
      uint64_t LHSConstValue = LHSConstOp->getZExtValue();
      uint64_t RHSConstant = RHSConst->getZExtValue();
      if (isPowerOf2_64(RHSConstant)) {
        uint64_t NewMaskValue = LHSConstValue & ~(RHSConstant - 1);
        LHS =
            DAG.getNode(ISD::AND, DL, LHS.getValueType(), LHS.getOperand(0),
                        DAG.getConstant(NewMaskValue, DL, LHS.getValueType()));
        RHS = DAG.getConstant(0, DL, RHS.getValueType());
        CC = ISD::SETEQ;
      }
    }
  }
}

SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
                                              SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  if (VT.isScalableVector()) {
    SDValue SrcVal = Op.getOperand(0);

    if (VT == MVT::nxv2f64 && SrcVal.getValueType() == MVT::nxv2bf16) {
      // Break conversion in two with the first part converting to f32 and the
      // second using native f32->VT instructions.
      SDLoc DL(Op);
      return DAG.getNode(ISD::FP_EXTEND, DL, VT,
                         DAG.getNode(ISD::FP_EXTEND, DL, MVT::nxv2f32, SrcVal));
    }

    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FP_EXTEND_MERGE_PASSTHRU);
  }

  if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()))
    return LowerFixedLengthFPExtendToSVE(Op, DAG);

  bool IsStrict = Op->isStrictFPOpcode();
  SDValue Op0 = Op.getOperand(IsStrict ? 1 : 0);
  EVT Op0VT = Op0.getValueType();
  if (VT == MVT::f64) {
    // FP16->FP32 extends are legal for v32 and v4f32.
    if (Op0VT == MVT::f32 || Op0VT == MVT::f16)
      return Op;
    // Split bf16->f64 extends into two fpextends.
    if (Op0VT == MVT::bf16 && IsStrict) {
      SDValue Ext1 =
          DAG.getNode(ISD::STRICT_FP_EXTEND, SDLoc(Op), {MVT::f32, MVT::Other},
                      {Op0, Op.getOperand(0)});
      return DAG.getNode(ISD::STRICT_FP_EXTEND, SDLoc(Op), {VT, MVT::Other},
                         {Ext1, Ext1.getValue(1)});
    }
    if (Op0VT == MVT::bf16)
      return DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), VT,
                         DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), MVT::f32, Op0));
    return SDValue();
  }

  assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
  return SDValue();
}

SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
                                             SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  bool IsStrict = Op->isStrictFPOpcode();
  SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0);
  EVT SrcVT = SrcVal.getValueType();
  bool Trunc = Op.getConstantOperandVal(IsStrict ? 2 : 1) == 1;

  if (VT.isScalableVector()) {
    // Let common code split the operation.
    if (SrcVT == MVT::nxv8f32)
      return Op;

    if (VT.getScalarType() != MVT::bf16)
      return LowerToPredicatedOp(Op, DAG, AArch64ISD::FP_ROUND_MERGE_PASSTHRU);

    SDLoc DL(Op);
    constexpr EVT I32 = MVT::nxv4i32;
    auto ImmV = [&](int I) -> SDValue { return DAG.getConstant(I, DL, I32); };

    SDValue NaN;
    SDValue Narrow;

    if (SrcVT == MVT::nxv2f32 || SrcVT == MVT::nxv4f32) {
      if (Subtarget->hasBF16())
        return LowerToPredicatedOp(Op, DAG,
                                   AArch64ISD::FP_ROUND_MERGE_PASSTHRU);

      Narrow = getSVESafeBitCast(I32, SrcVal, DAG);

      // Set the quiet bit.
      if (!DAG.isKnownNeverSNaN(SrcVal))
        NaN = DAG.getNode(ISD::OR, DL, I32, Narrow, ImmV(0x400000));
    } else if (SrcVT == MVT::nxv2f64 &&
               (Subtarget->hasSVE2() || Subtarget->isStreamingSVEAvailable())) {
      // Round to float without introducing rounding errors and try again.
      SDValue Pg = getPredicateForVector(DAG, DL, MVT::nxv2f32);
      Narrow = DAG.getNode(AArch64ISD::FCVTX_MERGE_PASSTHRU, DL, MVT::nxv2f32,
                           Pg, SrcVal, DAG.getUNDEF(MVT::nxv2f32));

      SmallVector<SDValue, 3> NewOps;
      if (IsStrict)
        NewOps.push_back(Op.getOperand(0));
      NewOps.push_back(Narrow);
      NewOps.push_back(Op.getOperand(IsStrict ? 2 : 1));
      return DAG.getNode(Op.getOpcode(), DL, VT, NewOps, Op->getFlags());
    } else
      return SDValue();

    if (!Trunc) {
      SDValue Lsb = DAG.getNode(ISD::SRL, DL, I32, Narrow, ImmV(16));
      Lsb = DAG.getNode(ISD::AND, DL, I32, Lsb, ImmV(1));
      SDValue RoundingBias = DAG.getNode(ISD::ADD, DL, I32, Lsb, ImmV(0x7fff));
      Narrow = DAG.getNode(ISD::ADD, DL, I32, Narrow, RoundingBias);
    }

    // Don't round if we had a NaN, we don't want to turn 0x7fffffff into
    // 0x80000000.
    if (NaN) {
      EVT I1 = I32.changeElementType(MVT::i1);
      EVT CondVT = VT.changeElementType(MVT::i1);
      SDValue IsNaN = DAG.getSetCC(DL, CondVT, SrcVal, SrcVal, ISD::SETUO);
      IsNaN = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, I1, IsNaN);
      Narrow = DAG.getSelect(DL, I32, IsNaN, NaN, Narrow);
    }

    // Now that we have rounded, shift the bits into position.
    Narrow = DAG.getNode(ISD::SRL, DL, I32, Narrow, ImmV(16));
    return getSVESafeBitCast(VT, Narrow, DAG);
  }

  if (useSVEForFixedLengthVectorVT(SrcVT, !Subtarget->isNeonAvailable()))
    return LowerFixedLengthFPRoundToSVE(Op, DAG);

  // Expand cases where the result type is BF16 but we don't have hardware
  // instructions to lower it.
  if (VT.getScalarType() == MVT::bf16 &&
      !((Subtarget->hasNEON() || Subtarget->hasSME()) &&
        Subtarget->hasBF16())) {
    SDLoc DL(Op);
    SDValue Narrow = SrcVal;
    SDValue NaN;
    EVT I32 = SrcVT.changeElementType(MVT::i32);
    EVT F32 = SrcVT.changeElementType(MVT::f32);
    if (SrcVT.getScalarType() == MVT::f32) {
      bool NeverSNaN = DAG.isKnownNeverSNaN(Narrow);
      Narrow = DAG.getNode(ISD::BITCAST, DL, I32, Narrow);
      if (!NeverSNaN) {
        // Set the quiet bit.
        NaN = DAG.getNode(ISD::OR, DL, I32, Narrow,
                          DAG.getConstant(0x400000, DL, I32));
      }
    } else if (SrcVT.getScalarType() == MVT::f64) {
      Narrow = DAG.getNode(AArch64ISD::FCVTXN, DL, F32, Narrow);
      Narrow = DAG.getNode(ISD::BITCAST, DL, I32, Narrow);
    } else {
      return SDValue();
    }
    if (!Trunc) {
      SDValue One = DAG.getConstant(1, DL, I32);
      SDValue Lsb = DAG.getNode(ISD::SRL, DL, I32, Narrow,
                                DAG.getShiftAmountConstant(16, I32, DL));
      Lsb = DAG.getNode(ISD::AND, DL, I32, Lsb, One);
      SDValue RoundingBias =
          DAG.getNode(ISD::ADD, DL, I32, DAG.getConstant(0x7fff, DL, I32), Lsb);
      Narrow = DAG.getNode(ISD::ADD, DL, I32, Narrow, RoundingBias);
    }

    // Don't round if we had a NaN, we don't want to turn 0x7fffffff into
    // 0x80000000.
    if (NaN) {
      SDValue IsNaN = DAG.getSetCC(
          DL, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), SrcVT),
          SrcVal, SrcVal, ISD::SETUO);
      Narrow = DAG.getSelect(DL, I32, IsNaN, NaN, Narrow);
    }

    // Now that we have rounded, shift the bits into position.
    Narrow = DAG.getNode(ISD::SRL, DL, I32, Narrow,
                         DAG.getShiftAmountConstant(16, I32, DL));
    if (VT.isVector()) {
      EVT I16 = I32.changeVectorElementType(MVT::i16);
      Narrow = DAG.getNode(ISD::TRUNCATE, DL, I16, Narrow);
      return DAG.getNode(ISD::BITCAST, DL, VT, Narrow);
    }
    Narrow = DAG.getNode(ISD::BITCAST, DL, F32, Narrow);
    SDValue Result = DAG.getTargetExtractSubreg(AArch64::hsub, DL, VT, Narrow);
    return IsStrict ? DAG.getMergeValues({Result, Op.getOperand(0)}, DL)
                    : Result;
  }

  if (SrcVT != MVT::f128) {
    // Expand cases where the input is a vector bigger than NEON.
    if (useSVEForFixedLengthVectorVT(SrcVT))
      return SDValue();

    // It's legal except when f128 is involved
    return Op;
  }

  return SDValue();
}

SDValue AArch64TargetLowering::LowerVectorFP_TO_INT(SDValue Op,
                                                    SelectionDAG &DAG) const {
  // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
  // Any additional optimization in this function should be recorded
  // in the cost tables.
  bool IsStrict = Op->isStrictFPOpcode();
  EVT InVT = Op.getOperand(IsStrict ? 1 : 0).getValueType();
  EVT VT = Op.getValueType();

  assert(!(IsStrict && VT.isScalableVector()) &&
         "Unimplemented SVE support for STRICT_FP_to_INT!");

  // f16 conversions are promoted to f32 when full fp16 is not supported.
  if ((InVT.getVectorElementType() == MVT::f16 && !Subtarget->hasFullFP16()) ||
      InVT.getVectorElementType() == MVT::bf16) {
    EVT NewVT = VT.changeElementType(MVT::f32);
    SDLoc DL(Op);
    if (IsStrict) {
      SDValue Ext = DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {NewVT, MVT::Other},
                                {Op.getOperand(0), Op.getOperand(1)});
      return DAG.getNode(Op.getOpcode(), DL, {VT, MVT::Other},
                         {Ext.getValue(1), Ext.getValue(0)});
    }
    return DAG.getNode(
        Op.getOpcode(), DL, Op.getValueType(),
        DAG.getNode(ISD::FP_EXTEND, DL, NewVT, Op.getOperand(0)));
  }

  if (VT.isScalableVector()) {
    if (VT.getVectorElementType() == MVT::i1) {
      SDLoc DL(Op);
      EVT CvtVT = getPromotedVTForPredicate(VT);
      SDValue Cvt = DAG.getNode(Op.getOpcode(), DL, CvtVT, Op.getOperand(0));
      SDValue Zero = DAG.getConstant(0, DL, CvtVT);
      return DAG.getSetCC(DL, VT, Cvt, Zero, ISD::SETNE);
    }

    // Let common code split the operation.
    if (InVT == MVT::nxv8f32)
      return Op;

    unsigned Opcode = Op.getOpcode() == ISD::FP_TO_UINT
                          ? AArch64ISD::FCVTZU_MERGE_PASSTHRU
                          : AArch64ISD::FCVTZS_MERGE_PASSTHRU;
    return LowerToPredicatedOp(Op, DAG, Opcode);
  }

  if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()) ||
      useSVEForFixedLengthVectorVT(InVT, !Subtarget->isNeonAvailable()))
    return LowerFixedLengthFPToIntToSVE(Op, DAG);

  uint64_t VTSize = VT.getFixedSizeInBits();
  uint64_t InVTSize = InVT.getFixedSizeInBits();
  if (VTSize < InVTSize) {
    SDLoc DL(Op);
    if (IsStrict) {
      InVT = InVT.changeVectorElementTypeToInteger();
      SDValue Cv = DAG.getNode(Op.getOpcode(), DL, {InVT, MVT::Other},
                               {Op.getOperand(0), Op.getOperand(1)});
      SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, VT, Cv);
      return DAG.getMergeValues({Trunc, Cv.getValue(1)}, DL);
    }
    SDValue Cv =
        DAG.getNode(Op.getOpcode(), DL, InVT.changeVectorElementTypeToInteger(),
                    Op.getOperand(0));
    return DAG.getNode(ISD::TRUNCATE, DL, VT, Cv);
  }

  if (VTSize > InVTSize) {
    SDLoc DL(Op);
    MVT ExtVT =
        MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
                         VT.getVectorNumElements());
    if (IsStrict) {
      SDValue Ext = DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {ExtVT, MVT::Other},
                                {Op.getOperand(0), Op.getOperand(1)});
      return DAG.getNode(Op.getOpcode(), DL, {VT, MVT::Other},
                         {Ext.getValue(1), Ext.getValue(0)});
    }
    SDValue Ext = DAG.getNode(ISD::FP_EXTEND, DL, ExtVT, Op.getOperand(0));
    return DAG.getNode(Op.getOpcode(), DL, VT, Ext);
  }

  // Use a scalar operation for conversions between single-element vectors of
  // the same size.
  if (InVT.getVectorNumElements() == 1) {
    SDLoc DL(Op);
    SDValue Extract = DAG.getNode(
        ISD::EXTRACT_VECTOR_ELT, DL, InVT.getScalarType(),
        Op.getOperand(IsStrict ? 1 : 0), DAG.getConstant(0, DL, MVT::i64));
    EVT ScalarVT = VT.getScalarType();
    if (IsStrict)
      return DAG.getNode(Op.getOpcode(), DL, {ScalarVT, MVT::Other},
                         {Op.getOperand(0), Extract});
    return DAG.getNode(Op.getOpcode(), DL, ScalarVT, Extract);
  }

  // Type changing conversions are illegal.
  return Op;
}

SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
                                              SelectionDAG &DAG) const {
  bool IsStrict = Op->isStrictFPOpcode();
  SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0);

  if (SrcVal.getValueType().isVector())
    return LowerVectorFP_TO_INT(Op, DAG);

  // f16 conversions are promoted to f32 when full fp16 is not supported.
  if ((SrcVal.getValueType() == MVT::f16 && !Subtarget->hasFullFP16()) ||
      SrcVal.getValueType() == MVT::bf16) {
    SDLoc DL(Op);
    if (IsStrict) {
      SDValue Ext =
          DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {MVT::f32, MVT::Other},
                      {Op.getOperand(0), SrcVal});
      return DAG.getNode(Op.getOpcode(), DL, {Op.getValueType(), MVT::Other},
                         {Ext.getValue(1), Ext.getValue(0)});
    }
    return DAG.getNode(Op.getOpcode(), DL, Op.getValueType(),
                       DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, SrcVal));
  }

  if (SrcVal.getValueType() != MVT::f128) {
    // It's legal except when f128 is involved
    return Op;
  }

  return SDValue();
}

SDValue
AArch64TargetLowering::LowerVectorFP_TO_INT_SAT(SDValue Op,
                                                SelectionDAG &DAG) const {
  // AArch64 FP-to-int conversions saturate to the destination element size, so
  // we can lower common saturating conversions to simple instructions.
  SDValue SrcVal = Op.getOperand(0);
  EVT SrcVT = SrcVal.getValueType();
  EVT DstVT = Op.getValueType();
  EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT();

  uint64_t SrcElementWidth = SrcVT.getScalarSizeInBits();
  uint64_t DstElementWidth = DstVT.getScalarSizeInBits();
  uint64_t SatWidth = SatVT.getScalarSizeInBits();
  assert(SatWidth <= DstElementWidth &&
         "Saturation width cannot exceed result width");

  // TODO: Consider lowering to SVE operations, as in LowerVectorFP_TO_INT.
  // Currently, the `llvm.fpto[su]i.sat.*` intrinsics don't accept scalable
  // types, so this is hard to reach.
  if (DstVT.isScalableVector())
    return SDValue();

  EVT SrcElementVT = SrcVT.getVectorElementType();

  // In the absence of FP16 support, promote f16 to f32 and saturate the result.
  SDLoc DL(Op);
  SDValue SrcVal2;
  if ((SrcElementVT == MVT::f16 &&
       (!Subtarget->hasFullFP16() || DstElementWidth > 16)) ||
      SrcElementVT == MVT::bf16) {
    MVT F32VT = MVT::getVectorVT(MVT::f32, SrcVT.getVectorNumElements());
    SrcVal = DAG.getNode(ISD::FP_EXTEND, DL, F32VT, SrcVal);
    // If we are extending to a v8f32, split into two v4f32 to produce legal
    // types.
    if (F32VT.getSizeInBits() > 128) {
      std::tie(SrcVal, SrcVal2) = DAG.SplitVector(SrcVal, DL);
      F32VT = F32VT.getHalfNumVectorElementsVT();
    }
    SrcVT = F32VT;
    SrcElementVT = MVT::f32;
    SrcElementWidth = 32;
  } else if (SrcElementVT != MVT::f64 && SrcElementVT != MVT::f32 &&
             SrcElementVT != MVT::f16 && SrcElementVT != MVT::bf16)
    return SDValue();

  // Expand to f64 if we are saturating to i64, to help keep the lanes the same
  // width and produce a fcvtzu.
  if (SatWidth == 64 && SrcElementWidth < 64) {
    MVT F64VT = MVT::getVectorVT(MVT::f64, SrcVT.getVectorNumElements());
    SrcVal = DAG.getNode(ISD::FP_EXTEND, DL, F64VT, SrcVal);
    SrcVT = F64VT;
    SrcElementVT = MVT::f64;
    SrcElementWidth = 64;
  }
  // Cases that we can emit directly.
  if (SrcElementWidth == DstElementWidth && SrcElementWidth == SatWidth) {
    SDValue Res = DAG.getNode(Op.getOpcode(), DL, DstVT, SrcVal,
                              DAG.getValueType(DstVT.getScalarType()));
    if (SrcVal2) {
      SDValue Res2 = DAG.getNode(Op.getOpcode(), DL, DstVT, SrcVal2,
                                 DAG.getValueType(DstVT.getScalarType()));
      return DAG.getNode(ISD::CONCAT_VECTORS, DL, DstVT, Res, Res2);
    }
    return Res;
  }

  // Otherwise we emit a cvt that saturates to a higher BW, and saturate the
  // result. This is only valid if the legal cvt is larger than the saturate
  // width. For double, as we don't have MIN/MAX, it can be simpler to scalarize
  // (at least until sqxtn is selected).
  if (SrcElementWidth < SatWidth || SrcElementVT == MVT::f64)
    return SDValue();

  EVT IntVT = SrcVT.changeVectorElementTypeToInteger();
  SDValue NativeCvt = DAG.getNode(Op.getOpcode(), DL, IntVT, SrcVal,
                                  DAG.getValueType(IntVT.getScalarType()));
  SDValue NativeCvt2 =
      SrcVal2 ? DAG.getNode(Op.getOpcode(), DL, IntVT, SrcVal2,
                            DAG.getValueType(IntVT.getScalarType()))
              : SDValue();
  SDValue Sat, Sat2;
  if (Op.getOpcode() == ISD::FP_TO_SINT_SAT) {
    SDValue MinC = DAG.getConstant(
        APInt::getSignedMaxValue(SatWidth).sext(SrcElementWidth), DL, IntVT);
    SDValue Min = DAG.getNode(ISD::SMIN, DL, IntVT, NativeCvt, MinC);
    SDValue Min2 = SrcVal2 ? DAG.getNode(ISD::SMIN, DL, IntVT, NativeCvt2, MinC) : SDValue();
    SDValue MaxC = DAG.getConstant(
        APInt::getSignedMinValue(SatWidth).sext(SrcElementWidth), DL, IntVT);
    Sat = DAG.getNode(ISD::SMAX, DL, IntVT, Min, MaxC);
    Sat2 = SrcVal2 ? DAG.getNode(ISD::SMAX, DL, IntVT, Min2, MaxC) : SDValue();
  } else {
    SDValue MinC = DAG.getConstant(
        APInt::getAllOnes(SatWidth).zext(SrcElementWidth), DL, IntVT);
    Sat = DAG.getNode(ISD::UMIN, DL, IntVT, NativeCvt, MinC);
    Sat2 = SrcVal2 ? DAG.getNode(ISD::UMIN, DL, IntVT, NativeCvt2, MinC) : SDValue();
  }

  if (SrcVal2)
    Sat = DAG.getNode(ISD::CONCAT_VECTORS, DL,
                      IntVT.getDoubleNumVectorElementsVT(*DAG.getContext()),
                      Sat, Sat2);

  return DAG.getNode(ISD::TRUNCATE, DL, DstVT, Sat);
}

SDValue AArch64TargetLowering::LowerFP_TO_INT_SAT(SDValue Op,
                                                  SelectionDAG &DAG) const {
  // AArch64 FP-to-int conversions saturate to the destination register size, so
  // we can lower common saturating conversions to simple instructions.
  SDValue SrcVal = Op.getOperand(0);
  EVT SrcVT = SrcVal.getValueType();

  if (SrcVT.isVector())
    return LowerVectorFP_TO_INT_SAT(Op, DAG);

  EVT DstVT = Op.getValueType();
  EVT SatVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
  uint64_t SatWidth = SatVT.getScalarSizeInBits();
  uint64_t DstWidth = DstVT.getScalarSizeInBits();
  assert(SatWidth <= DstWidth && "Saturation width cannot exceed result width");

  // In the absence of FP16 support, promote f16 to f32 and saturate the result.
  if ((SrcVT == MVT::f16 && !Subtarget->hasFullFP16()) || SrcVT == MVT::bf16) {
    SrcVal = DAG.getNode(ISD::FP_EXTEND, SDLoc(Op), MVT::f32, SrcVal);
    SrcVT = MVT::f32;
  } else if (SrcVT != MVT::f64 && SrcVT != MVT::f32 && SrcVT != MVT::f16 &&
             SrcVT != MVT::bf16)
    return SDValue();

  SDLoc DL(Op);
  // Cases that we can emit directly.
  if ((SrcVT == MVT::f64 || SrcVT == MVT::f32 ||
       (SrcVT == MVT::f16 && Subtarget->hasFullFP16())) &&
      DstVT == SatVT && (DstVT == MVT::i64 || DstVT == MVT::i32))
    return DAG.getNode(Op.getOpcode(), DL, DstVT, SrcVal,
                       DAG.getValueType(DstVT));

  // Otherwise we emit a cvt that saturates to a higher BW, and saturate the
  // result. This is only valid if the legal cvt is larger than the saturate
  // width.
  if (DstWidth < SatWidth)
    return SDValue();

  SDValue NativeCvt =
      DAG.getNode(Op.getOpcode(), DL, DstVT, SrcVal, DAG.getValueType(DstVT));
  SDValue Sat;
  if (Op.getOpcode() == ISD::FP_TO_SINT_SAT) {
    SDValue MinC = DAG.getConstant(
        APInt::getSignedMaxValue(SatWidth).sext(DstWidth), DL, DstVT);
    SDValue Min = DAG.getNode(ISD::SMIN, DL, DstVT, NativeCvt, MinC);
    SDValue MaxC = DAG.getConstant(
        APInt::getSignedMinValue(SatWidth).sext(DstWidth), DL, DstVT);
    Sat = DAG.getNode(ISD::SMAX, DL, DstVT, Min, MaxC);
  } else {
    SDValue MinC = DAG.getConstant(
        APInt::getAllOnes(SatWidth).zext(DstWidth), DL, DstVT);
    Sat = DAG.getNode(ISD::UMIN, DL, DstVT, NativeCvt, MinC);
  }

  return DAG.getNode(ISD::TRUNCATE, DL, DstVT, Sat);
}

SDValue AArch64TargetLowering::LowerVectorXRINT(SDValue Op,
                                                SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDValue Src = Op.getOperand(0);
  SDLoc DL(Op);

  assert(VT.isVector() && "Expected vector type");

  EVT CastVT =
      VT.changeVectorElementType(Src.getValueType().getVectorElementType());

  // Round the floating-point value into a floating-point register with the
  // current rounding mode.
  SDValue FOp = DAG.getNode(ISD::FRINT, DL, CastVT, Src);

  // Truncate the rounded floating point to an integer.
  return DAG.getNode(ISD::FP_TO_SINT_SAT, DL, VT, FOp,
                     DAG.getValueType(VT.getVectorElementType()));
}

SDValue AArch64TargetLowering::LowerVectorINT_TO_FP(SDValue Op,
                                                    SelectionDAG &DAG) const {
  // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
  // Any additional optimization in this function should be recorded
  // in the cost tables.
  bool IsStrict = Op->isStrictFPOpcode();
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue In = Op.getOperand(IsStrict ? 1 : 0);
  EVT InVT = In.getValueType();
  unsigned Opc = Op.getOpcode();
  bool IsSigned = Opc == ISD::SINT_TO_FP || Opc == ISD::STRICT_SINT_TO_FP;

  assert(!(IsStrict && VT.isScalableVector()) &&
         "Unimplemented SVE support for ISD:::STRICT_INT_TO_FP!");

  // NOTE: i1->bf16 does not require promotion to f32.
  if (VT.isScalableVector() && InVT.getVectorElementType() == MVT::i1) {
    SDValue FalseVal = DAG.getConstantFP(0.0, DL, VT);
    SDValue TrueVal = IsSigned ? DAG.getConstantFP(-1.0, DL, VT)
                               : DAG.getConstantFP(1.0, DL, VT);
    return DAG.getNode(ISD::VSELECT, DL, VT, In, TrueVal, FalseVal);
  }

  // Promote bf16 conversions to f32.
  if (VT.getVectorElementType() == MVT::bf16) {
    EVT F32 = VT.changeElementType(MVT::f32);
    if (IsStrict) {
      SDValue Val = DAG.getNode(Op.getOpcode(), DL, {F32, MVT::Other},
                                {Op.getOperand(0), In});
      return DAG.getNode(ISD::STRICT_FP_ROUND, DL,
                         {Op.getValueType(), MVT::Other},
                         {Val.getValue(1), Val.getValue(0),
                          DAG.getIntPtrConstant(0, DL, /*isTarget=*/true)});
    }
    return DAG.getNode(ISD::FP_ROUND, DL, Op.getValueType(),
                       DAG.getNode(Op.getOpcode(), DL, F32, In),
                       DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
  }

  if (VT.isScalableVector()) {
    // Let common code split the operation.
    if (VT == MVT::nxv8f32)
      return Op;

    unsigned Opcode = IsSigned ? AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU
                               : AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU;
    return LowerToPredicatedOp(Op, DAG, Opcode);
  }

  if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()) ||
      useSVEForFixedLengthVectorVT(InVT, !Subtarget->isNeonAvailable()))
    return LowerFixedLengthIntToFPToSVE(Op, DAG);

  uint64_t VTSize = VT.getFixedSizeInBits();
  uint64_t InVTSize = InVT.getFixedSizeInBits();
  if (VTSize < InVTSize) {
    // AArch64 doesn't have a direct vector instruction to convert
    // fixed point to floating point AND narrow it at the same time.
    // Additional rounding when the target is f32/f64 causes double
    // rounding issues. Conversion to f16 is fine due to narrow width.
    bool IsTargetf32 = VT.getVectorElementType() == MVT::f32;
    bool IsTargetf16 = false;
    if (Op.hasOneUse() &&
        Op->user_begin()->getOpcode() == ISD::CONCAT_VECTORS) {
      // Some vector types are split during legalization into half, followed by
      // concatenation, followed by rounding to the original vector type. If we
      // end up resolving to f16 type, we shouldn't worry about rounding errors.
      SDNode *U = *Op->user_begin();
      if (U->hasOneUse() && U->user_begin()->getOpcode() == ISD::FP_ROUND) {
        EVT TmpVT = U->user_begin()->getValueType(0);
        if (TmpVT.getScalarType() == MVT::f16)
          IsTargetf16 = true;
      }
    }

    if (IsTargetf32 && !IsTargetf16) {
      return !IsStrict ? DAG.UnrollVectorOp(Op.getNode()) : SDValue();
    }

    MVT CastVT =
        MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
                         InVT.getVectorNumElements());
    if (IsStrict) {
      In = DAG.getNode(Opc, DL, {CastVT, MVT::Other}, {Op.getOperand(0), In});
      return DAG.getNode(ISD::STRICT_FP_ROUND, DL, {VT, MVT::Other},
                         {In.getValue(1), In.getValue(0),
                          DAG.getIntPtrConstant(0, DL, /*isTarget=*/true)});
    }
    In = DAG.getNode(Opc, DL, CastVT, In);
    return DAG.getNode(ISD::FP_ROUND, DL, VT, In,
                       DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
  }

  if (VTSize > InVTSize) {
    unsigned CastOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
    EVT CastVT = VT.changeVectorElementTypeToInteger();
    In = DAG.getNode(CastOpc, DL, CastVT, In);
    if (IsStrict)
      return DAG.getNode(Opc, DL, {VT, MVT::Other}, {Op.getOperand(0), In});
    return DAG.getNode(Opc, DL, VT, In);
  }

  // Use a scalar operation for conversions between single-element vectors of
  // the same size.
  if (VT.getVectorNumElements() == 1) {
    SDValue Extract =
        DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, InVT.getScalarType(), In,
                    DAG.getConstant(0, DL, MVT::i64));
    EVT ScalarVT = VT.getScalarType();
    if (IsStrict)
      return DAG.getNode(Op.getOpcode(), DL, {ScalarVT, MVT::Other},
                         {Op.getOperand(0), Extract});
    return DAG.getNode(Op.getOpcode(), DL, ScalarVT, Extract);
  }

  return Op;
}

SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
                                            SelectionDAG &DAG) const {
  if (Op.getValueType().isVector())
    return LowerVectorINT_TO_FP(Op, DAG);

  bool IsStrict = Op->isStrictFPOpcode();
  SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0);

  bool IsSigned = Op->getOpcode() == ISD::STRICT_SINT_TO_FP ||
                  Op->getOpcode() == ISD::SINT_TO_FP;

  auto IntToFpViaPromotion = [&](EVT PromoteVT) {
    SDLoc DL(Op);
    if (IsStrict) {
      SDValue Val = DAG.getNode(Op.getOpcode(), DL, {PromoteVT, MVT::Other},
                                {Op.getOperand(0), SrcVal});
      return DAG.getNode(ISD::STRICT_FP_ROUND, DL,
                         {Op.getValueType(), MVT::Other},
                         {Val.getValue(1), Val.getValue(0),
                          DAG.getIntPtrConstant(0, DL, /*isTarget=*/true)});
    }
    return DAG.getNode(ISD::FP_ROUND, DL, Op.getValueType(),
                       DAG.getNode(Op.getOpcode(), DL, PromoteVT, SrcVal),
                       DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
  };

  if (Op.getValueType() == MVT::bf16) {
    unsigned MaxWidth = IsSigned
                            ? DAG.ComputeMaxSignificantBits(SrcVal)
                            : DAG.computeKnownBits(SrcVal).countMaxActiveBits();
    // bf16 conversions are promoted to f32 when converting from i16.
    if (MaxWidth <= 24) {
      return IntToFpViaPromotion(MVT::f32);
    }

    // bf16 conversions are promoted to f64 when converting from i32.
    if (MaxWidth <= 53) {
      return IntToFpViaPromotion(MVT::f64);
    }

    // We need to be careful about i64 -> bf16.
    // Consider an i32 22216703.
    // This number cannot be represented exactly as an f32 and so a itofp will
    // turn it into 22216704.0 fptrunc to bf16 will turn this into 22282240.0
    // However, the correct bf16 was supposed to be 22151168.0
    // We need to use sticky rounding to get this correct.
    if (SrcVal.getValueType() == MVT::i64) {
      SDLoc DL(Op);
      // This algorithm is equivalent to the following:
      // uint64_t SrcHi = SrcVal & ~0xfffull;
      // uint64_t SrcLo = SrcVal &  0xfffull;
      // uint64_t Highest = SrcVal >> 53;
      // bool HasHighest = Highest != 0;
      // uint64_t ToRound = HasHighest ? SrcHi : SrcVal;
      // double  Rounded = static_cast<double>(ToRound);
      // uint64_t RoundedBits = std::bit_cast<uint64_t>(Rounded);
      // uint64_t HasLo = SrcLo != 0;
      // bool NeedsAdjustment = HasHighest & HasLo;
      // uint64_t AdjustedBits = RoundedBits | uint64_t{NeedsAdjustment};
      // double Adjusted = std::bit_cast<double>(AdjustedBits);
      // return static_cast<__bf16>(Adjusted);
      //
      // Essentially, what happens is that SrcVal either fits perfectly in a
      // double-precision value or it is too big. If it is sufficiently small,
      // we should just go u64 -> double -> bf16 in a naive way. Otherwise, we
      // ensure that u64 -> double has no rounding error by only using the 52
      // MSB of the input. The low order bits will get merged into a sticky bit
      // which will avoid issues incurred by double rounding.

      // Signed conversion is more or less like so:
      // copysign((__bf16)abs(SrcVal), SrcVal)
      SDValue SignBit;
      if (IsSigned) {
        SignBit = DAG.getNode(ISD::AND, DL, MVT::i64, SrcVal,
                              DAG.getConstant(1ull << 63, DL, MVT::i64));
        SrcVal = DAG.getNode(ISD::ABS, DL, MVT::i64, SrcVal);
      }
      SDValue SrcHi = DAG.getNode(ISD::AND, DL, MVT::i64, SrcVal,
                                  DAG.getConstant(~0xfffull, DL, MVT::i64));
      SDValue SrcLo = DAG.getNode(ISD::AND, DL, MVT::i64, SrcVal,
                                  DAG.getConstant(0xfffull, DL, MVT::i64));
      SDValue Highest =
          DAG.getNode(ISD::SRL, DL, MVT::i64, SrcVal,
                      DAG.getShiftAmountConstant(53, MVT::i64, DL));
      SDValue Zero64 = DAG.getConstant(0, DL, MVT::i64);
      SDValue ToRound =
          DAG.getSelectCC(DL, Highest, Zero64, SrcHi, SrcVal, ISD::SETNE);
      SDValue Rounded =
          IsStrict ? DAG.getNode(Op.getOpcode(), DL, {MVT::f64, MVT::Other},
                                 {Op.getOperand(0), ToRound})
                   : DAG.getNode(Op.getOpcode(), DL, MVT::f64, ToRound);

      SDValue RoundedBits = DAG.getNode(ISD::BITCAST, DL, MVT::i64, Rounded);
      if (SignBit) {
        RoundedBits = DAG.getNode(ISD::OR, DL, MVT::i64, RoundedBits, SignBit);
      }

      SDValue HasHighest = DAG.getSetCC(
          DL,
          getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
          Highest, Zero64, ISD::SETNE);

      SDValue HasLo = DAG.getSetCC(
          DL,
          getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
          SrcLo, Zero64, ISD::SETNE);

      SDValue NeedsAdjustment =
          DAG.getNode(ISD::AND, DL, HasLo.getValueType(), HasHighest, HasLo);
      NeedsAdjustment = DAG.getZExtOrTrunc(NeedsAdjustment, DL, MVT::i64);

      SDValue AdjustedBits =
          DAG.getNode(ISD::OR, DL, MVT::i64, RoundedBits, NeedsAdjustment);
      SDValue Adjusted = DAG.getNode(ISD::BITCAST, DL, MVT::f64, AdjustedBits);
      return IsStrict
                 ? DAG.getNode(
                       ISD::STRICT_FP_ROUND, DL,
                       {Op.getValueType(), MVT::Other},
                       {Rounded.getValue(1), Adjusted,
                        DAG.getIntPtrConstant(0, DL, /*isTarget=*/true)})
                 : DAG.getNode(ISD::FP_ROUND, DL, Op.getValueType(), Adjusted,
                               DAG.getIntPtrConstant(0, DL, /*isTarget=*/true));
    }
  }

  // f16 conversions are promoted to f32 when full fp16 is not supported.
  if (Op.getValueType() == MVT::f16 && !Subtarget->hasFullFP16()) {
    return IntToFpViaPromotion(MVT::f32);
  }

  // i128 conversions are libcalls.
  if (SrcVal.getValueType() == MVT::i128)
    return SDValue();

  // Other conversions are legal, unless it's to the completely software-based
  // fp128.
  if (Op.getValueType() != MVT::f128)
    return Op;
  return SDValue();
}

SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
                                            SelectionDAG &DAG) const {
  // For iOS, we want to call an alternative entry point: __sincos_stret,
  // which returns the values in two S / D registers.
  SDLoc DL(Op);
  SDValue Arg = Op.getOperand(0);
  EVT ArgVT = Arg.getValueType();
  Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());

  ArgListTy Args;
  ArgListEntry Entry;

  Entry.Node = Arg;
  Entry.Ty = ArgTy;
  Entry.IsSExt = false;
  Entry.IsZExt = false;
  Args.push_back(Entry);

  RTLIB::Libcall LC = ArgVT == MVT::f64 ? RTLIB::SINCOS_STRET_F64
                                        : RTLIB::SINCOS_STRET_F32;
  const char *LibcallName = getLibcallName(LC);
  SDValue Callee =
      DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));

  StructType *RetTy = StructType::get(ArgTy, ArgTy);
  TargetLowering::CallLoweringInfo CLI(DAG);
  CallingConv::ID CC = getLibcallCallingConv(LC);
  CLI.setDebugLoc(DL)
      .setChain(DAG.getEntryNode())
      .setLibCallee(CC, RetTy, Callee, std::move(Args));

  std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
  return CallResult.first;
}

static MVT getSVEContainerType(EVT ContentTy);

SDValue AArch64TargetLowering::LowerBITCAST(SDValue Op,
                                            SelectionDAG &DAG) const {
  EVT OpVT = Op.getValueType();
  EVT ArgVT = Op.getOperand(0).getValueType();

  if (useSVEForFixedLengthVectorVT(OpVT))
    return LowerFixedLengthBitcastToSVE(Op, DAG);

  if (OpVT.isScalableVector()) {
    assert(isTypeLegal(OpVT) && "Unexpected result type!");

    // Handle type legalisation first.
    if (!isTypeLegal(ArgVT)) {
      assert(OpVT.isFloatingPoint() && !ArgVT.isFloatingPoint() &&
             "Expected int->fp bitcast!");

      // Bitcasting between unpacked vector types of different element counts is
      // not a NOP because the live elements are laid out differently.
      //                01234567
      // e.g. nxv2i32 = XX??XX??
      //      nxv4f16 = X?X?X?X?
      if (OpVT.getVectorElementCount() != ArgVT.getVectorElementCount())
        return SDValue();

      SDValue ExtResult =
          DAG.getNode(ISD::ANY_EXTEND, SDLoc(Op), getSVEContainerType(ArgVT),
                      Op.getOperand(0));
      return getSVESafeBitCast(OpVT, ExtResult, DAG);
    }

    // Bitcasts between legal types with the same element count are legal.
    if (OpVT.getVectorElementCount() == ArgVT.getVectorElementCount())
      return Op;

    // getSVESafeBitCast does not support casting between unpacked types.
    if (!isPackedVectorType(OpVT, DAG))
      return SDValue();

    return getSVESafeBitCast(OpVT, Op.getOperand(0), DAG);
  }

  if (OpVT != MVT::f16 && OpVT != MVT::bf16)
    return SDValue();

  // Bitcasts between f16 and bf16 are legal.
  if (ArgVT == MVT::f16 || ArgVT == MVT::bf16)
    return Op;

  assert(ArgVT == MVT::i16);
  SDLoc DL(Op);

  Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
  Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
  return DAG.getTargetExtractSubreg(AArch64::hsub, DL, OpVT, Op);
}

// Returns lane if Op extracts from a two-element vector and lane is constant
// (i.e., extractelt(<2 x Ty> %v, ConstantLane)), and std::nullopt otherwise.
static std::optional<uint64_t>
getConstantLaneNumOfExtractHalfOperand(SDValue &Op) {
  SDNode *OpNode = Op.getNode();
  if (OpNode->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
    return std::nullopt;

  EVT VT = OpNode->getOperand(0).getValueType();
  ConstantSDNode *C = dyn_cast<ConstantSDNode>(OpNode->getOperand(1));
  if (!VT.isFixedLengthVector() || VT.getVectorNumElements() != 2 || !C)
    return std::nullopt;

  return C->getZExtValue();
}

static bool isExtendedBUILD_VECTOR(SDValue N, SelectionDAG &DAG,
                                   bool isSigned) {
  EVT VT = N.getValueType();

  if (N.getOpcode() != ISD::BUILD_VECTOR)
    return false;

  for (const SDValue &Elt : N->op_values()) {
    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
      unsigned EltSize = VT.getScalarSizeInBits();
      unsigned HalfSize = EltSize / 2;
      if (isSigned) {
        if (!isIntN(HalfSize, C->getSExtValue()))
          return false;
      } else {
        if (!isUIntN(HalfSize, C->getZExtValue()))
          return false;
      }
      continue;
    }
    return false;
  }

  return true;
}

static SDValue skipExtensionForVectorMULL(SDValue N, SelectionDAG &DAG) {
  EVT VT = N.getValueType();
  assert(VT.is128BitVector() && "Unexpected vector MULL size");
  EVT HalfVT = EVT::getVectorVT(
      *DAG.getContext(),
      VT.getScalarType().getHalfSizedIntegerVT(*DAG.getContext()),
      VT.getVectorElementCount());
  return DAG.getNode(ISD::TRUNCATE, SDLoc(N), HalfVT, N);
}

static bool isSignExtended(SDValue N, SelectionDAG &DAG) {
  return N.getOpcode() == ISD::SIGN_EXTEND ||
         N.getOpcode() == ISD::ANY_EXTEND ||
         isExtendedBUILD_VECTOR(N, DAG, true);
}

static bool isZeroExtended(SDValue N, SelectionDAG &DAG) {
  return N.getOpcode() == ISD::ZERO_EXTEND ||
         N.getOpcode() == ISD::ANY_EXTEND ||
         isExtendedBUILD_VECTOR(N, DAG, false);
}

static bool isAddSubSExt(SDValue N, SelectionDAG &DAG) {
  unsigned Opcode = N.getOpcode();
  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
    SDValue N0 = N.getOperand(0);
    SDValue N1 = N.getOperand(1);
    return N0->hasOneUse() && N1->hasOneUse() &&
      isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
  }
  return false;
}

static bool isAddSubZExt(SDValue N, SelectionDAG &DAG) {
  unsigned Opcode = N.getOpcode();
  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
    SDValue N0 = N.getOperand(0);
    SDValue N1 = N.getOperand(1);
    return N0->hasOneUse() && N1->hasOneUse() &&
      isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
  }
  return false;
}

SDValue AArch64TargetLowering::LowerGET_ROUNDING(SDValue Op,
                                                 SelectionDAG &DAG) const {
  // The rounding mode is in bits 23:22 of the FPSCR.
  // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
  // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
  // so that the shift + and get folded into a bitfield extract.
  SDLoc DL(Op);

  SDValue Chain = Op.getOperand(0);
  SDValue FPCR_64 = DAG.getNode(
      ISD::INTRINSIC_W_CHAIN, DL, {MVT::i64, MVT::Other},
      {Chain, DAG.getConstant(Intrinsic::aarch64_get_fpcr, DL, MVT::i64)});
  Chain = FPCR_64.getValue(1);
  SDValue FPCR_32 = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, FPCR_64);
  SDValue FltRounds = DAG.getNode(ISD::ADD, DL, MVT::i32, FPCR_32,
                                  DAG.getConstant(1U << 22, DL, MVT::i32));
  SDValue RMODE = DAG.getNode(ISD::SRL, DL, MVT::i32, FltRounds,
                              DAG.getConstant(22, DL, MVT::i32));
  SDValue AND = DAG.getNode(ISD::AND, DL, MVT::i32, RMODE,
                            DAG.getConstant(3, DL, MVT::i32));
  return DAG.getMergeValues({AND, Chain}, DL);
}

SDValue AArch64TargetLowering::LowerSET_ROUNDING(SDValue Op,
                                                 SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Chain = Op->getOperand(0);
  SDValue RMValue = Op->getOperand(1);

  // The rounding mode is in bits 23:22 of the FPCR.
  // The llvm.set.rounding argument value to the rounding mode in FPCR mapping
  // is 0->3, 1->0, 2->1, 3->2. The formula we use to implement this is
  // ((arg - 1) & 3) << 22).
  //
  // The argument of llvm.set.rounding must be within the segment [0, 3], so
  // NearestTiesToAway (4) is not handled here. It is responsibility of the code
  // generated llvm.set.rounding to ensure this condition.

  // Calculate new value of FPCR[23:22].
  RMValue = DAG.getNode(ISD::SUB, DL, MVT::i32, RMValue,
                        DAG.getConstant(1, DL, MVT::i32));
  RMValue = DAG.getNode(ISD::AND, DL, MVT::i32, RMValue,
                        DAG.getConstant(0x3, DL, MVT::i32));
  RMValue =
      DAG.getNode(ISD::SHL, DL, MVT::i32, RMValue,
                  DAG.getConstant(AArch64::RoundingBitsPos, DL, MVT::i32));
  RMValue = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, RMValue);

  // Get current value of FPCR.
  SDValue Ops[] = {
      Chain, DAG.getTargetConstant(Intrinsic::aarch64_get_fpcr, DL, MVT::i64)};
  SDValue FPCR =
      DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL, {MVT::i64, MVT::Other}, Ops);
  Chain = FPCR.getValue(1);
  FPCR = FPCR.getValue(0);

  // Put new rounding mode into FPSCR[23:22].
  const int RMMask = ~(AArch64::Rounding::rmMask << AArch64::RoundingBitsPos);
  FPCR = DAG.getNode(ISD::AND, DL, MVT::i64, FPCR,
                     DAG.getConstant(RMMask, DL, MVT::i64));
  FPCR = DAG.getNode(ISD::OR, DL, MVT::i64, FPCR, RMValue);
  SDValue Ops2[] = {
      Chain, DAG.getTargetConstant(Intrinsic::aarch64_set_fpcr, DL, MVT::i64),
      FPCR};
  return DAG.getNode(ISD::INTRINSIC_VOID, DL, MVT::Other, Ops2);
}

SDValue AArch64TargetLowering::LowerGET_FPMODE(SDValue Op,
                                               SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Chain = Op->getOperand(0);

  // Get current value of FPCR.
  SDValue Ops[] = {
      Chain, DAG.getTargetConstant(Intrinsic::aarch64_get_fpcr, DL, MVT::i64)};
  SDValue FPCR =
      DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL, {MVT::i64, MVT::Other}, Ops);
  Chain = FPCR.getValue(1);
  FPCR = FPCR.getValue(0);

  // Truncate FPCR to 32 bits.
  SDValue Result = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, FPCR);

  return DAG.getMergeValues({Result, Chain}, DL);
}

SDValue AArch64TargetLowering::LowerSET_FPMODE(SDValue Op,
                                               SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Chain = Op->getOperand(0);
  SDValue Mode = Op->getOperand(1);

  // Extend the specified value to 64 bits.
  SDValue FPCR = DAG.getZExtOrTrunc(Mode, DL, MVT::i64);

  // Set new value of FPCR.
  SDValue Ops2[] = {
      Chain, DAG.getConstant(Intrinsic::aarch64_set_fpcr, DL, MVT::i64), FPCR};
  return DAG.getNode(ISD::INTRINSIC_VOID, DL, MVT::Other, Ops2);
}

SDValue AArch64TargetLowering::LowerRESET_FPMODE(SDValue Op,
                                                 SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Chain = Op->getOperand(0);

  // Get current value of FPCR.
  SDValue Ops[] = {
      Chain, DAG.getTargetConstant(Intrinsic::aarch64_get_fpcr, DL, MVT::i64)};
  SDValue FPCR =
      DAG.getNode(ISD::INTRINSIC_W_CHAIN, DL, {MVT::i64, MVT::Other}, Ops);
  Chain = FPCR.getValue(1);
  FPCR = FPCR.getValue(0);

  // Clear bits that are not reserved.
  SDValue FPSCRMasked = DAG.getNode(
      ISD::AND, DL, MVT::i64, FPCR,
      DAG.getConstant(AArch64::ReservedFPControlBits, DL, MVT::i64));

  // Set new value of FPCR.
  SDValue Ops2[] = {Chain,
                    DAG.getConstant(Intrinsic::aarch64_set_fpcr, DL, MVT::i64),
                    FPSCRMasked};
  return DAG.getNode(ISD::INTRINSIC_VOID, DL, MVT::Other, Ops2);
}

static unsigned selectUmullSmull(SDValue &N0, SDValue &N1, SelectionDAG &DAG,
                                 SDLoc DL, bool &IsMLA) {
  bool IsN0SExt = isSignExtended(N0, DAG);
  bool IsN1SExt = isSignExtended(N1, DAG);
  if (IsN0SExt && IsN1SExt)
    return AArch64ISD::SMULL;

  bool IsN0ZExt = isZeroExtended(N0, DAG);
  bool IsN1ZExt = isZeroExtended(N1, DAG);

  if (IsN0ZExt && IsN1ZExt)
    return AArch64ISD::UMULL;

  // Select UMULL if we can replace the other operand with an extend.
  EVT VT = N0.getValueType();
  unsigned EltSize = VT.getScalarSizeInBits();
  APInt Mask = APInt::getHighBitsSet(EltSize, EltSize / 2);
  if (IsN0ZExt || IsN1ZExt) {
    if (DAG.MaskedValueIsZero(IsN0ZExt ? N1 : N0, Mask))
      return AArch64ISD::UMULL;
  } else if (VT == MVT::v2i64 && DAG.MaskedValueIsZero(N0, Mask) &&
             DAG.MaskedValueIsZero(N1, Mask)) {
    // For v2i64 we look more aggressively at both operands being zero, to avoid
    // scalarization.
    return AArch64ISD::UMULL;
  }

  if (IsN0SExt || IsN1SExt) {
    if (DAG.ComputeNumSignBits(IsN0SExt ? N1 : N0) > EltSize / 2)
      return AArch64ISD::SMULL;
  } else if (VT == MVT::v2i64 && DAG.ComputeNumSignBits(N0) > EltSize / 2 &&
             DAG.ComputeNumSignBits(N1) > EltSize / 2) {
    return AArch64ISD::SMULL;
  }

  if (!IsN1SExt && !IsN1ZExt)
    return 0;

  // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
  // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
  if (IsN1SExt && isAddSubSExt(N0, DAG)) {
    IsMLA = true;
    return AArch64ISD::SMULL;
  }
  if (IsN1ZExt && isAddSubZExt(N0, DAG)) {
    IsMLA = true;
    return AArch64ISD::UMULL;
  }
  if (IsN0ZExt && isAddSubZExt(N1, DAG)) {
    std::swap(N0, N1);
    IsMLA = true;
    return AArch64ISD::UMULL;
  }
  return 0;
}

SDValue AArch64TargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();

  bool OverrideNEON = !Subtarget->isNeonAvailable();
  if (VT.isScalableVector() || useSVEForFixedLengthVectorVT(VT, OverrideNEON))
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::MUL_PRED);

  // Multiplications are only custom-lowered for 128-bit and 64-bit vectors so
  // that VMULL can be detected.  Otherwise v2i64 multiplications are not legal.
  assert((VT.is128BitVector() || VT.is64BitVector()) && VT.isInteger() &&
         "unexpected type for custom-lowering ISD::MUL");
  SDValue N0 = Op.getOperand(0);
  SDValue N1 = Op.getOperand(1);
  bool isMLA = false;
  EVT OVT = VT;
  if (VT.is64BitVector()) {
    if (N0.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
        isNullConstant(N0.getOperand(1)) &&
        N1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
        isNullConstant(N1.getOperand(1))) {
      N0 = N0.getOperand(0);
      N1 = N1.getOperand(0);
      VT = N0.getValueType();
    } else {
      if (VT == MVT::v1i64) {
        if (Subtarget->hasSVE())
          return LowerToPredicatedOp(Op, DAG, AArch64ISD::MUL_PRED);
        // Fall through to expand this.  It is not legal.
        return SDValue();
      } else
        // Other vector multiplications are legal.
        return Op;
    }
  }

  SDLoc DL(Op);
  unsigned NewOpc = selectUmullSmull(N0, N1, DAG, DL, isMLA);

  if (!NewOpc) {
    if (VT.getVectorElementType() == MVT::i64) {
      // If SVE is available then i64 vector multiplications can also be made
      // legal.
      if (Subtarget->hasSVE())
        return LowerToPredicatedOp(Op, DAG, AArch64ISD::MUL_PRED);
      // Fall through to expand this.  It is not legal.
      return SDValue();
    } else
      // Other vector multiplications are legal.
      return Op;
  }

  // Legalize to a S/UMULL instruction
  SDValue Op0;
  SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
  if (!isMLA) {
    Op0 = skipExtensionForVectorMULL(N0, DAG);
    assert(Op0.getValueType().is64BitVector() &&
           Op1.getValueType().is64BitVector() &&
           "unexpected types for extended operands to VMULL");
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OVT,
                       DAG.getNode(NewOpc, DL, VT, Op0, Op1),
                       DAG.getConstant(0, DL, MVT::i64));
  }
  // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
  // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
  // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
  SDValue N00 = skipExtensionForVectorMULL(N0.getOperand(0), DAG);
  SDValue N01 = skipExtensionForVectorMULL(N0.getOperand(1), DAG);
  EVT Op1VT = Op1.getValueType();
  return DAG.getNode(
      ISD::EXTRACT_SUBVECTOR, DL, OVT,
      DAG.getNode(N0.getOpcode(), DL, VT,
                  DAG.getNode(NewOpc, DL, VT,
                              DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
                  DAG.getNode(NewOpc, DL, VT,
                              DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1)),
      DAG.getConstant(0, DL, MVT::i64));
}

static inline SDValue getPTrue(SelectionDAG &DAG, SDLoc DL, EVT VT,
                               int Pattern) {
  if (Pattern == AArch64SVEPredPattern::all)
    return DAG.getConstant(1, DL, VT);
  return DAG.getNode(AArch64ISD::PTRUE, DL, VT,
                     DAG.getTargetConstant(Pattern, DL, MVT::i32));
}

static SDValue optimizeIncrementingWhile(SDNode *N, SelectionDAG &DAG,
                                         bool IsSigned, bool IsEqual) {
  unsigned Op0 = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 1 : 0;
  unsigned Op1 = N->getOpcode() == ISD::INTRINSIC_WO_CHAIN ? 2 : 1;

  if (!N->getValueType(0).isScalableVector() ||
      !isa<ConstantSDNode>(N->getOperand(Op1)))
    return SDValue();

  SDLoc DL(N);
  APInt Y = N->getConstantOperandAPInt(Op1);

  // When the second operand is the maximum value, comparisons that include
  // equality can never fail and thus we can return an all active predicate.
  if (IsEqual)
    if (IsSigned ? Y.isMaxSignedValue() : Y.isMaxValue())
      return DAG.getConstant(1, DL, N->getValueType(0));

  if (!isa<ConstantSDNode>(N->getOperand(Op0)))
    return SDValue();

  APInt X = N->getConstantOperandAPInt(Op0);

  bool Overflow;
  APInt NumActiveElems =
      IsSigned ? Y.ssub_ov(X, Overflow) : Y.usub_ov(X, Overflow);

  if (Overflow)
    return SDValue();

  if (IsEqual) {
    APInt One(NumActiveElems.getBitWidth(), 1, IsSigned);
    NumActiveElems = IsSigned ? NumActiveElems.sadd_ov(One, Overflow)
                              : NumActiveElems.uadd_ov(One, Overflow);
    if (Overflow)
      return SDValue();
  }

  std::optional<unsigned> PredPattern =
      getSVEPredPatternFromNumElements(NumActiveElems.getZExtValue());
  unsigned MinSVEVectorSize = std::max(
      DAG.getSubtarget<AArch64Subtarget>().getMinSVEVectorSizeInBits(), 128u);
  unsigned ElementSize = 128 / N->getValueType(0).getVectorMinNumElements();
  if (PredPattern != std::nullopt &&
      NumActiveElems.getZExtValue() <= (MinSVEVectorSize / ElementSize))
    return getPTrue(DAG, DL, N->getValueType(0), *PredPattern);

  return SDValue();
}

// Returns a safe bitcast between two scalable vector predicates, where
// any newly created lanes from a widening bitcast are defined as zero.
static SDValue getSVEPredicateBitCast(EVT VT, SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  EVT InVT = Op.getValueType();

  assert(InVT.getVectorElementType() == MVT::i1 &&
         VT.getVectorElementType() == MVT::i1 &&
         "Expected a predicate-to-predicate bitcast");
  assert(VT.isScalableVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
         InVT.isScalableVector() &&
         DAG.getTargetLoweringInfo().isTypeLegal(InVT) &&
         "Only expect to cast between legal scalable predicate types!");

  // Return the operand if the cast isn't changing type,
  if (InVT == VT)
    return Op;

  // Look through casts to <vscale x 16 x i1> when their input has more lanes
  // than VT. This will increase the chances of removing casts that introduce
  // new lanes, which have to be explicitly zero'd.
  if (Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
      Op.getConstantOperandVal(0) == Intrinsic::aarch64_sve_convert_to_svbool &&
      Op.getOperand(1).getValueType().bitsGT(VT))
    Op = Op.getOperand(1);

  SDValue Reinterpret = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, VT, Op);

  // We only have to zero the lanes if new lanes are being defined, e.g. when
  // casting from <vscale x 2 x i1> to <vscale x 16 x i1>. If this is not the
  // case (e.g. when casting from <vscale x 16 x i1> -> <vscale x 2 x i1>) then
  // we can return here.
  if (InVT.bitsGT(VT))
    return Reinterpret;

  // Check if the other lanes are already known to be zeroed by
  // construction.
  if (isZeroingInactiveLanes(Op))
    return Reinterpret;

  // Zero the newly introduced lanes.
  SDValue Mask = DAG.getConstant(1, DL, InVT);
  Mask = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, VT, Mask);
  return DAG.getNode(ISD::AND, DL, VT, Reinterpret, Mask);
}

SDValue AArch64TargetLowering::getRuntimePStateSM(SelectionDAG &DAG,
                                                  SDValue Chain, SDLoc DL,
                                                  EVT VT) const {
  SDValue Callee = DAG.getExternalSymbol("__arm_sme_state",
                                         getPointerTy(DAG.getDataLayout()));
  Type *Int64Ty = Type::getInt64Ty(*DAG.getContext());
  Type *RetTy = StructType::get(Int64Ty, Int64Ty);
  TargetLowering::CallLoweringInfo CLI(DAG);
  ArgListTy Args;
  CLI.setDebugLoc(DL).setChain(Chain).setLibCallee(
      CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2,
      RetTy, Callee, std::move(Args));
  std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
  SDValue Mask = DAG.getConstant(/*PSTATE.SM*/ 1, DL, MVT::i64);
  return DAG.getNode(ISD::AND, DL, MVT::i64, CallResult.first.getOperand(0),
                     Mask);
}

// Lower an SME LDR/STR ZA intrinsic
// Case 1: If the vector number (vecnum) is an immediate in range, it gets
// folded into the instruction
//    ldr(%tileslice, %ptr, 11) -> ldr [%tileslice, 11], [%ptr, 11]
// Case 2: If the vecnum is not an immediate, then it is used to modify the base
// and tile slice registers
//    ldr(%tileslice, %ptr, %vecnum)
//    ->
//    %svl = rdsvl
//    %ptr2 = %ptr + %svl * %vecnum
//    %tileslice2 = %tileslice + %vecnum
//    ldr [%tileslice2, 0], [%ptr2, 0]
// Case 3: If the vecnum is an immediate out of range, then the same is done as
// case 2, but the base and slice registers are modified by the greatest
// multiple of 15 lower than the vecnum and the remainder is folded into the
// instruction. This means that successive loads and stores that are offset from
// each other can share the same base and slice register updates.
//    ldr(%tileslice, %ptr, 22)
//    ldr(%tileslice, %ptr, 23)
//    ->
//    %svl = rdsvl
//    %ptr2 = %ptr + %svl * 15
//    %tileslice2 = %tileslice + 15
//    ldr [%tileslice2, 7], [%ptr2, 7]
//    ldr [%tileslice2, 8], [%ptr2, 8]
// Case 4: If the vecnum is an add of an immediate, then the non-immediate
// operand and the immediate can be folded into the instruction, like case 2.
//    ldr(%tileslice, %ptr, %vecnum + 7)
//    ldr(%tileslice, %ptr, %vecnum + 8)
//    ->
//    %svl = rdsvl
//    %ptr2 = %ptr + %svl * %vecnum
//    %tileslice2 = %tileslice + %vecnum
//    ldr [%tileslice2, 7], [%ptr2, 7]
//    ldr [%tileslice2, 8], [%ptr2, 8]
// Case 5: The vecnum being an add of an immediate out of range is also handled,
// in which case the same remainder logic as case 3 is used.
SDValue LowerSMELdrStr(SDValue N, SelectionDAG &DAG, bool IsLoad) {
  SDLoc DL(N);

  SDValue TileSlice = N->getOperand(2);
  SDValue Base = N->getOperand(3);
  SDValue VecNum = N->getOperand(4);
  int32_t ConstAddend = 0;
  SDValue VarAddend = VecNum;

  // If the vnum is an add of an immediate, we can fold it into the instruction
  if (VecNum.getOpcode() == ISD::ADD &&
      isa<ConstantSDNode>(VecNum.getOperand(1))) {
    ConstAddend = cast<ConstantSDNode>(VecNum.getOperand(1))->getSExtValue();
    VarAddend = VecNum.getOperand(0);
  } else if (auto ImmNode = dyn_cast<ConstantSDNode>(VecNum)) {
    ConstAddend = ImmNode->getSExtValue();
    VarAddend = SDValue();
  }

  int32_t ImmAddend = ConstAddend % 16;
  if (int32_t C = (ConstAddend - ImmAddend)) {
    SDValue CVal = DAG.getTargetConstant(C, DL, MVT::i32);
    VarAddend = VarAddend
                    ? DAG.getNode(ISD::ADD, DL, MVT::i32, {VarAddend, CVal})
                    : CVal;
  }

  if (VarAddend) {
    // Get the vector length that will be multiplied by vnum
    auto SVL = DAG.getNode(AArch64ISD::RDSVL, DL, MVT::i64,
                           DAG.getConstant(1, DL, MVT::i32));

    // Multiply SVL and vnum then add it to the base
    SDValue Mul = DAG.getNode(
        ISD::MUL, DL, MVT::i64,
        {SVL, DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, VarAddend)});
    Base = DAG.getNode(ISD::ADD, DL, MVT::i64, {Base, Mul});
    // Just add vnum to the tileslice
    TileSlice = DAG.getNode(ISD::ADD, DL, MVT::i32, {TileSlice, VarAddend});
  }

  return DAG.getNode(IsLoad ? AArch64ISD::SME_ZA_LDR : AArch64ISD::SME_ZA_STR,
                     DL, MVT::Other,
                     {/*Chain=*/N.getOperand(0), TileSlice, Base,
                      DAG.getTargetConstant(ImmAddend, DL, MVT::i32)});
}

SDValue LowerVectorMatch(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  SDValue ID =
      DAG.getTargetConstant(Intrinsic::aarch64_sve_match, DL, MVT::i64);

  auto Op1 = Op.getOperand(1);
  auto Op2 = Op.getOperand(2);
  auto Mask = Op.getOperand(3);

  EVT Op1VT = Op1.getValueType();
  EVT Op2VT = Op2.getValueType();
  EVT ResVT = Op.getValueType();

  assert((Op1VT.getVectorElementType() == MVT::i8 ||
          Op1VT.getVectorElementType() == MVT::i16) &&
         "Expected 8-bit or 16-bit characters.");

  // Scalable vector type used to wrap operands.
  // A single container is enough for both operands because ultimately the
  // operands will have to be wrapped to the same type (nxv16i8 or nxv8i16).
  EVT OpContainerVT = Op1VT.isScalableVector()
                          ? Op1VT
                          : getContainerForFixedLengthVector(DAG, Op1VT);

  if (Op2VT.is128BitVector()) {
    // If Op2 is a full 128-bit vector, wrap it trivially in a scalable vector.
    Op2 = convertToScalableVector(DAG, OpContainerVT, Op2);
    // Further, if the result is scalable, broadcast Op2 to a full SVE register.
    if (ResVT.isScalableVector())
      Op2 = DAG.getNode(AArch64ISD::DUPLANE128, DL, OpContainerVT, Op2,
                        DAG.getTargetConstant(0, DL, MVT::i64));
  } else {
    // If Op2 is not a full 128-bit vector, we always need to broadcast it.
    unsigned Op2BitWidth = Op2VT.getFixedSizeInBits();
    MVT Op2IntVT = MVT::getIntegerVT(Op2BitWidth);
    EVT Op2PromotedVT = getPackedSVEVectorVT(Op2IntVT);
    Op2 = DAG.getBitcast(MVT::getVectorVT(Op2IntVT, 1), Op2);
    Op2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, Op2IntVT, Op2,
                      DAG.getConstant(0, DL, MVT::i64));
    Op2 = DAG.getSplatVector(Op2PromotedVT, DL, Op2);
    Op2 = DAG.getBitcast(OpContainerVT, Op2);
  }

  // If the result is scalable, we just need to carry out the MATCH.
  if (ResVT.isScalableVector())
    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ResVT, ID, Mask, Op1, Op2);

  // If the result is fixed, we can still use MATCH but we need to wrap the
  // first operand and the mask in scalable vectors before doing so.

  // Wrap the operands.
  Op1 = convertToScalableVector(DAG, OpContainerVT, Op1);
  Mask = DAG.getNode(ISD::SIGN_EXTEND, DL, Op1VT, Mask);
  Mask = convertFixedMaskToScalableVector(Mask, DAG);

  // Carry out the match.
  SDValue Match = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Mask.getValueType(),
                              ID, Mask, Op1, Op2);

  // Extract and promote the match result (nxv16i1/nxv8i1) to ResVT
  // (v16i8/v8i8).
  Match = DAG.getNode(ISD::SIGN_EXTEND, DL, OpContainerVT, Match);
  Match = convertFromScalableVector(DAG, Op1VT, Match);
  return DAG.getNode(ISD::TRUNCATE, DL, ResVT, Match);
}

SDValue AArch64TargetLowering::LowerINTRINSIC_VOID(SDValue Op,
                                                   SelectionDAG &DAG) const {
  unsigned IntNo = Op.getConstantOperandVal(1);
  SDLoc DL(Op);
  switch (IntNo) {
  default:
    return SDValue(); // Don't custom lower most intrinsics.
  case Intrinsic::aarch64_prefetch: {
    SDValue Chain = Op.getOperand(0);
    SDValue Addr = Op.getOperand(2);

    unsigned IsWrite = Op.getConstantOperandVal(3);
    unsigned Locality = Op.getConstantOperandVal(4);
    unsigned IsStream = Op.getConstantOperandVal(5);
    unsigned IsData = Op.getConstantOperandVal(6);
    unsigned PrfOp = (IsWrite << 4) |    // Load/Store bit
                     (!IsData << 3) |    // IsDataCache bit
                     (Locality << 1) |   // Cache level bits
                     (unsigned)IsStream; // Stream bit

    return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Chain,
                       DAG.getTargetConstant(PrfOp, DL, MVT::i32), Addr);
  }
  case Intrinsic::aarch64_sme_str:
  case Intrinsic::aarch64_sme_ldr: {
    return LowerSMELdrStr(Op, DAG, IntNo == Intrinsic::aarch64_sme_ldr);
  }
  case Intrinsic::aarch64_sme_za_enable:
    return DAG.getNode(
        AArch64ISD::SMSTART, DL, DAG.getVTList(MVT::Other, MVT::Glue),
        Op->getOperand(0), // Chain
        DAG.getTargetConstant((int32_t)(AArch64SVCR::SVCRZA), DL, MVT::i32));
  case Intrinsic::aarch64_sme_za_disable:
    return DAG.getNode(
        AArch64ISD::SMSTOP, DL, DAG.getVTList(MVT::Other, MVT::Glue),
        Op->getOperand(0), // Chain
        DAG.getTargetConstant((int32_t)(AArch64SVCR::SVCRZA), DL, MVT::i32));
  }
}

SDValue AArch64TargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op,
                                                      SelectionDAG &DAG) const {
  unsigned IntNo = Op.getConstantOperandVal(1);
  SDLoc DL(Op);
  switch (IntNo) {
  default:
    return SDValue(); // Don't custom lower most intrinsics.
  case Intrinsic::aarch64_mops_memset_tag: {
    auto Node = cast<MemIntrinsicSDNode>(Op.getNode());
    SDValue Chain = Node->getChain();
    SDValue Dst = Op.getOperand(2);
    SDValue Val = Op.getOperand(3);
    Val = DAG.getAnyExtOrTrunc(Val, DL, MVT::i64);
    SDValue Size = Op.getOperand(4);
    auto Alignment = Node->getMemOperand()->getAlign();
    bool IsVol = Node->isVolatile();
    auto DstPtrInfo = Node->getPointerInfo();

    const auto &SDI =
        static_cast<const AArch64SelectionDAGInfo &>(DAG.getSelectionDAGInfo());
    SDValue MS = SDI.EmitMOPS(AArch64::MOPSMemorySetTaggingPseudo, DAG, DL,
                              Chain, Dst, Val, Size, Alignment, IsVol,
                              DstPtrInfo, MachinePointerInfo{});

    // MOPS_MEMSET_TAGGING has 3 results (DstWb, SizeWb, Chain) whereas the
    // intrinsic has 2. So hide SizeWb using MERGE_VALUES. Otherwise
    // LowerOperationWrapper will complain that the number of results has
    // changed.
    return DAG.getMergeValues({MS.getValue(0), MS.getValue(2)}, DL);
  }
  }
}

SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                     SelectionDAG &DAG) const {
  unsigned IntNo = Op.getConstantOperandVal(0);
  SDLoc DL(Op);
  switch (IntNo) {
  default: return SDValue();    // Don't custom lower most intrinsics.
  case Intrinsic::thread_pointer: {
    EVT PtrVT = getPointerTy(DAG.getDataLayout());
    return DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);
  }
  case Intrinsic::aarch64_neon_abs: {
    EVT Ty = Op.getValueType();
    if (Ty == MVT::i64) {
      SDValue Result =
          DAG.getNode(ISD::BITCAST, DL, MVT::v1i64, Op.getOperand(1));
      Result = DAG.getNode(ISD::ABS, DL, MVT::v1i64, Result);
      return DAG.getNode(ISD::BITCAST, DL, MVT::i64, Result);
    } else if (Ty.isVector() && Ty.isInteger() && isTypeLegal(Ty)) {
      return DAG.getNode(ISD::ABS, DL, Ty, Op.getOperand(1));
    } else {
      report_fatal_error("Unexpected type for AArch64 NEON intrinsic");
    }
  }
  case Intrinsic::aarch64_neon_pmull64: {
    SDValue LHS = Op.getOperand(1);
    SDValue RHS = Op.getOperand(2);

    std::optional<uint64_t> LHSLane =
        getConstantLaneNumOfExtractHalfOperand(LHS);
    std::optional<uint64_t> RHSLane =
        getConstantLaneNumOfExtractHalfOperand(RHS);

    assert((!LHSLane || *LHSLane < 2) && "Expect lane to be None or 0 or 1");
    assert((!RHSLane || *RHSLane < 2) && "Expect lane to be None or 0 or 1");

    // 'aarch64_neon_pmull64' takes i64 parameters; while pmull/pmull2
    // instructions execute on SIMD registers. So canonicalize i64 to v1i64,
    // which ISel recognizes better. For example, generate a ldr into d*
    // registers as opposed to a GPR load followed by a fmov.
    auto TryVectorizeOperand = [](SDValue N, std::optional<uint64_t> NLane,
                                  std::optional<uint64_t> OtherLane,
                                  const SDLoc &DL,
                                  SelectionDAG &DAG) -> SDValue {
      // If the operand is an higher half itself, rewrite it to
      // extract_high_v2i64; this way aarch64_neon_pmull64 could
      // re-use the dag-combiner function with aarch64_neon_{pmull,smull,umull}.
      if (NLane == 1)
        return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v1i64,
                           N.getOperand(0), DAG.getConstant(1, DL, MVT::i64));

      // Operand N is not a higher half but the other operand is.
      if (OtherLane == 1) {
        // If this operand is a lower half, rewrite it to
        // extract_high_v2i64(duplane(<2 x Ty>, 0)). This saves a roundtrip to
        // align lanes of two operands. A roundtrip sequence (to move from lane
        // 1 to lane 0) is like this:
        //   mov x8, v0.d[1]
        //   fmov d0, x8
        if (NLane == 0)
          return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v1i64,
                             DAG.getNode(AArch64ISD::DUPLANE64, DL, MVT::v2i64,
                                         N.getOperand(0),
                                         DAG.getConstant(0, DL, MVT::i64)),
                             DAG.getConstant(1, DL, MVT::i64));

        // Otherwise just dup from main to all lanes.
        return DAG.getNode(AArch64ISD::DUP, DL, MVT::v1i64, N);
      }

      // Neither operand is an extract of higher half, so codegen may just use
      // the non-high version of PMULL instruction. Use v1i64 to represent i64.
      assert(N.getValueType() == MVT::i64 &&
             "Intrinsic aarch64_neon_pmull64 requires i64 parameters");
      return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v1i64, N);
    };

    LHS = TryVectorizeOperand(LHS, LHSLane, RHSLane, DL, DAG);
    RHS = TryVectorizeOperand(RHS, RHSLane, LHSLane, DL, DAG);

    return DAG.getNode(AArch64ISD::PMULL, DL, Op.getValueType(), LHS, RHS);
  }
  case Intrinsic::aarch64_neon_smax:
    return DAG.getNode(ISD::SMAX, DL, Op.getValueType(), Op.getOperand(1),
                       Op.getOperand(2));
  case Intrinsic::aarch64_neon_umax:
    return DAG.getNode(ISD::UMAX, DL, Op.getValueType(), Op.getOperand(1),
                       Op.getOperand(2));
  case Intrinsic::aarch64_neon_smin:
    return DAG.getNode(ISD::SMIN, DL, Op.getValueType(), Op.getOperand(1),
                       Op.getOperand(2));
  case Intrinsic::aarch64_neon_umin:
    return DAG.getNode(ISD::UMIN, DL, Op.getValueType(), Op.getOperand(1),
                       Op.getOperand(2));
  case Intrinsic::aarch64_neon_scalar_sqxtn:
  case Intrinsic::aarch64_neon_scalar_sqxtun:
  case Intrinsic::aarch64_neon_scalar_uqxtn: {
    assert(Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::f32);
    if (Op.getValueType() == MVT::i32)
      return DAG.getNode(ISD::BITCAST, DL, MVT::i32,
                         DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::f32,
                                     Op.getOperand(0),
                                     DAG.getNode(ISD::BITCAST, DL, MVT::f64,
                                                 Op.getOperand(1))));
    return SDValue();
  }
  case Intrinsic::aarch64_neon_sqxtn:
    return DAG.getNode(ISD::TRUNCATE_SSAT_S, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_neon_sqxtun:
    return DAG.getNode(ISD::TRUNCATE_SSAT_U, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_neon_uqxtn:
    return DAG.getNode(ISD::TRUNCATE_USAT_U, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_neon_sqshrn:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::TRUNCATE_SSAT_S, DL, Op.getValueType(),
                         DAG.getNode(AArch64ISD::VASHR, DL,
                                     Op.getOperand(1).getValueType(),
                                     Op.getOperand(1), Op.getOperand(2)));
    return SDValue();
  case Intrinsic::aarch64_neon_sqshrun:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::TRUNCATE_SSAT_U, DL, Op.getValueType(),
                         DAG.getNode(AArch64ISD::VASHR, DL,
                                     Op.getOperand(1).getValueType(),
                                     Op.getOperand(1), Op.getOperand(2)));
    return SDValue();
  case Intrinsic::aarch64_neon_uqshrn:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::TRUNCATE_USAT_U, DL, Op.getValueType(),
                         DAG.getNode(AArch64ISD::VLSHR, DL,
                                     Op.getOperand(1).getValueType(),
                                     Op.getOperand(1), Op.getOperand(2)));
    return SDValue();
  case Intrinsic::aarch64_neon_sqrshrn:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::TRUNCATE_SSAT_S, DL, Op.getValueType(),
                         DAG.getNode(AArch64ISD::SRSHR_I, DL,
                                     Op.getOperand(1).getValueType(),
                                     Op.getOperand(1), Op.getOperand(2)));
    return SDValue();
  case Intrinsic::aarch64_neon_sqrshrun:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::TRUNCATE_SSAT_U, DL, Op.getValueType(),
                         DAG.getNode(AArch64ISD::SRSHR_I, DL,
                                     Op.getOperand(1).getValueType(),
                                     Op.getOperand(1), Op.getOperand(2)));
    return SDValue();
  case Intrinsic::aarch64_neon_uqrshrn:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::TRUNCATE_USAT_U, DL, Op.getValueType(),
                         DAG.getNode(AArch64ISD::URSHR_I, DL,
                                     Op.getOperand(1).getValueType(),
                                     Op.getOperand(1), Op.getOperand(2)));
    return SDValue();
  case Intrinsic::aarch64_neon_sqadd:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::SADDSAT, DL, Op.getValueType(), Op.getOperand(1),
                         Op.getOperand(2));
    return SDValue();
  case Intrinsic::aarch64_neon_sqsub:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::SSUBSAT, DL, Op.getValueType(), Op.getOperand(1),
                         Op.getOperand(2));
    return SDValue();
  case Intrinsic::aarch64_neon_uqadd:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::UADDSAT, DL, Op.getValueType(), Op.getOperand(1),
                         Op.getOperand(2));
    return SDValue();
  case Intrinsic::aarch64_neon_uqsub:
    if (Op.getValueType().isVector())
      return DAG.getNode(ISD::USUBSAT, DL, Op.getValueType(), Op.getOperand(1),
                         Op.getOperand(2));
    return SDValue();
  case Intrinsic::aarch64_sve_whilelt:
    return optimizeIncrementingWhile(Op.getNode(), DAG, /*IsSigned=*/true,
                                     /*IsEqual=*/false);
  case Intrinsic::aarch64_sve_whilels:
    return optimizeIncrementingWhile(Op.getNode(), DAG, /*IsSigned=*/false,
                                     /*IsEqual=*/true);
  case Intrinsic::aarch64_sve_whilele:
    return optimizeIncrementingWhile(Op.getNode(), DAG, /*IsSigned=*/true,
                                     /*IsEqual=*/true);
  case Intrinsic::aarch64_sve_sunpkhi:
    return DAG.getNode(AArch64ISD::SUNPKHI, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_sunpklo:
    return DAG.getNode(AArch64ISD::SUNPKLO, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_uunpkhi:
    return DAG.getNode(AArch64ISD::UUNPKHI, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_uunpklo:
    return DAG.getNode(AArch64ISD::UUNPKLO, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_clasta_n:
    return DAG.getNode(AArch64ISD::CLASTA_N, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
  case Intrinsic::aarch64_sve_clastb_n:
    return DAG.getNode(AArch64ISD::CLASTB_N, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
  case Intrinsic::aarch64_sve_lasta:
    return DAG.getNode(AArch64ISD::LASTA, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_lastb:
    return DAG.getNode(AArch64ISD::LASTB, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_rev:
    return DAG.getNode(ISD::VECTOR_REVERSE, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_tbl:
    return DAG.getNode(AArch64ISD::TBL, DL, Op.getValueType(), Op.getOperand(1),
                       Op.getOperand(2));
  case Intrinsic::aarch64_sve_trn1:
    return DAG.getNode(AArch64ISD::TRN1, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_trn2:
    return DAG.getNode(AArch64ISD::TRN2, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_uzp1:
    return DAG.getNode(AArch64ISD::UZP1, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_uzp2:
    return DAG.getNode(AArch64ISD::UZP2, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_zip1:
    return DAG.getNode(AArch64ISD::ZIP1, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_zip2:
    return DAG.getNode(AArch64ISD::ZIP2, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_splice:
    return DAG.getNode(AArch64ISD::SPLICE, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
  case Intrinsic::aarch64_sve_ptrue:
    return getPTrue(DAG, DL, Op.getValueType(), Op.getConstantOperandVal(1));
  case Intrinsic::aarch64_sve_clz:
    return DAG.getNode(AArch64ISD::CTLZ_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sme_cntsb:
    return DAG.getNode(AArch64ISD::RDSVL, DL, Op.getValueType(),
                       DAG.getConstant(1, DL, MVT::i32));
  case Intrinsic::aarch64_sme_cntsh: {
    SDValue One = DAG.getConstant(1, DL, MVT::i32);
    SDValue Bytes = DAG.getNode(AArch64ISD::RDSVL, DL, Op.getValueType(), One);
    return DAG.getNode(ISD::SRL, DL, Op.getValueType(), Bytes, One);
  }
  case Intrinsic::aarch64_sme_cntsw: {
    SDValue Bytes = DAG.getNode(AArch64ISD::RDSVL, DL, Op.getValueType(),
                                DAG.getConstant(1, DL, MVT::i32));
    return DAG.getNode(ISD::SRL, DL, Op.getValueType(), Bytes,
                       DAG.getConstant(2, DL, MVT::i32));
  }
  case Intrinsic::aarch64_sme_cntsd: {
    SDValue Bytes = DAG.getNode(AArch64ISD::RDSVL, DL, Op.getValueType(),
                                DAG.getConstant(1, DL, MVT::i32));
    return DAG.getNode(ISD::SRL, DL, Op.getValueType(), Bytes,
                       DAG.getConstant(3, DL, MVT::i32));
  }
  case Intrinsic::aarch64_sve_cnt: {
    SDValue Data = Op.getOperand(3);
    // CTPOP only supports integer operands.
    if (Data.getValueType().isFloatingPoint())
      Data = DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Data);
    return DAG.getNode(AArch64ISD::CTPOP_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Data, Op.getOperand(1));
  }
  case Intrinsic::aarch64_sve_dupq_lane:
    return LowerDUPQLane(Op, DAG);
  case Intrinsic::aarch64_sve_convert_from_svbool:
    if (Op.getValueType() == MVT::aarch64svcount)
      return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Op.getOperand(1));
    return getSVEPredicateBitCast(Op.getValueType(), Op.getOperand(1), DAG);
  case Intrinsic::aarch64_sve_convert_to_svbool:
    if (Op.getOperand(1).getValueType() == MVT::aarch64svcount)
      return DAG.getNode(ISD::BITCAST, DL, MVT::nxv16i1, Op.getOperand(1));
    return getSVEPredicateBitCast(MVT::nxv16i1, Op.getOperand(1), DAG);
  case Intrinsic::aarch64_sve_fneg:
    return DAG.getNode(AArch64ISD::FNEG_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_frintp:
    return DAG.getNode(AArch64ISD::FCEIL_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_frintm:
    return DAG.getNode(AArch64ISD::FFLOOR_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_frinti:
    return DAG.getNode(AArch64ISD::FNEARBYINT_MERGE_PASSTHRU, DL,
                       Op.getValueType(), Op.getOperand(2), Op.getOperand(3),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_frintx:
    return DAG.getNode(AArch64ISD::FRINT_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_frinta:
    return DAG.getNode(AArch64ISD::FROUND_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_frintn:
    return DAG.getNode(AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU, DL,
                       Op.getValueType(), Op.getOperand(2), Op.getOperand(3),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_frintz:
    return DAG.getNode(AArch64ISD::FTRUNC_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_ucvtf:
    return DAG.getNode(AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU, DL,
                       Op.getValueType(), Op.getOperand(2), Op.getOperand(3),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_scvtf:
    return DAG.getNode(AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU, DL,
                       Op.getValueType(), Op.getOperand(2), Op.getOperand(3),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_fcvtzu:
    return DAG.getNode(AArch64ISD::FCVTZU_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_fcvtzs:
    return DAG.getNode(AArch64ISD::FCVTZS_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_fsqrt:
    return DAG.getNode(AArch64ISD::FSQRT_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_frecpx:
    return DAG.getNode(AArch64ISD::FRECPX_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_frecpe_x:
    return DAG.getNode(AArch64ISD::FRECPE, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_frecps_x:
    return DAG.getNode(AArch64ISD::FRECPS, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_frsqrte_x:
    return DAG.getNode(AArch64ISD::FRSQRTE, DL, Op.getValueType(),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_frsqrts_x:
    return DAG.getNode(AArch64ISD::FRSQRTS, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_sve_fabs:
    return DAG.getNode(AArch64ISD::FABS_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_abs:
    return DAG.getNode(AArch64ISD::ABS_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_neg:
    return DAG.getNode(AArch64ISD::NEG_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_insr: {
    SDValue Scalar = Op.getOperand(2);
    EVT ScalarTy = Scalar.getValueType();
    if ((ScalarTy == MVT::i8) || (ScalarTy == MVT::i16))
      Scalar = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Scalar);

    return DAG.getNode(AArch64ISD::INSR, DL, Op.getValueType(),
                       Op.getOperand(1), Scalar);
  }
  case Intrinsic::aarch64_sve_rbit:
    return DAG.getNode(AArch64ISD::BITREVERSE_MERGE_PASSTHRU, DL,
                       Op.getValueType(), Op.getOperand(2), Op.getOperand(3),
                       Op.getOperand(1));
  case Intrinsic::aarch64_sve_revb:
    return DAG.getNode(AArch64ISD::BSWAP_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_revh:
    return DAG.getNode(AArch64ISD::REVH_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_revw:
    return DAG.getNode(AArch64ISD::REVW_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_revd:
    return DAG.getNode(AArch64ISD::REVD_MERGE_PASSTHRU, DL, Op.getValueType(),
                       Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
  case Intrinsic::aarch64_sve_sxtb:
    return DAG.getNode(
        AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU, DL, Op.getValueType(),
        Op.getOperand(2), Op.getOperand(3),
        DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i8)),
        Op.getOperand(1));
  case Intrinsic::aarch64_sve_sxth:
    return DAG.getNode(
        AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU, DL, Op.getValueType(),
        Op.getOperand(2), Op.getOperand(3),
        DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i16)),
        Op.getOperand(1));
  case Intrinsic::aarch64_sve_sxtw:
    return DAG.getNode(
        AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU, DL, Op.getValueType(),
        Op.getOperand(2), Op.getOperand(3),
        DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i32)),
        Op.getOperand(1));
  case Intrinsic::aarch64_sve_uxtb:
    return DAG.getNode(
        AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU, DL, Op.getValueType(),
        Op.getOperand(2), Op.getOperand(3),
        DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i8)),
        Op.getOperand(1));
  case Intrinsic::aarch64_sve_uxth:
    return DAG.getNode(
        AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU, DL, Op.getValueType(),
        Op.getOperand(2), Op.getOperand(3),
        DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i16)),
        Op.getOperand(1));
  case Intrinsic::aarch64_sve_uxtw:
    return DAG.getNode(
        AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU, DL, Op.getValueType(),
        Op.getOperand(2), Op.getOperand(3),
        DAG.getValueType(Op.getValueType().changeVectorElementType(MVT::i32)),
        Op.getOperand(1));
  case Intrinsic::localaddress: {
    const auto &MF = DAG.getMachineFunction();
    const auto *RegInfo = Subtarget->getRegisterInfo();
    unsigned Reg = RegInfo->getLocalAddressRegister(MF);
    return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg,
                              Op.getSimpleValueType());
  }

  case Intrinsic::eh_recoverfp: {
    // FIXME: This needs to be implemented to correctly handle highly aligned
    // stack objects. For now we simply return the incoming FP. Refer D53541
    // for more details.
    SDValue FnOp = Op.getOperand(1);
    SDValue IncomingFPOp = Op.getOperand(2);
    GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
    auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
    if (!Fn)
      report_fatal_error(
          "llvm.eh.recoverfp must take a function as the first argument");
    return IncomingFPOp;
  }

  case Intrinsic::aarch64_neon_vsri:
  case Intrinsic::aarch64_neon_vsli:
  case Intrinsic::aarch64_sve_sri:
  case Intrinsic::aarch64_sve_sli: {
    EVT Ty = Op.getValueType();

    if (!Ty.isVector())
      report_fatal_error("Unexpected type for aarch64_neon_vsli");

    assert(Op.getConstantOperandVal(3) <= Ty.getScalarSizeInBits());

    bool IsShiftRight = IntNo == Intrinsic::aarch64_neon_vsri ||
                        IntNo == Intrinsic::aarch64_sve_sri;
    unsigned Opcode = IsShiftRight ? AArch64ISD::VSRI : AArch64ISD::VSLI;
    return DAG.getNode(Opcode, DL, Ty, Op.getOperand(1), Op.getOperand(2),
                       Op.getOperand(3));
  }

  case Intrinsic::aarch64_neon_srhadd:
  case Intrinsic::aarch64_neon_urhadd:
  case Intrinsic::aarch64_neon_shadd:
  case Intrinsic::aarch64_neon_uhadd: {
    bool IsSignedAdd = (IntNo == Intrinsic::aarch64_neon_srhadd ||
                        IntNo == Intrinsic::aarch64_neon_shadd);
    bool IsRoundingAdd = (IntNo == Intrinsic::aarch64_neon_srhadd ||
                          IntNo == Intrinsic::aarch64_neon_urhadd);
    unsigned Opcode = IsSignedAdd
                          ? (IsRoundingAdd ? ISD::AVGCEILS : ISD::AVGFLOORS)
                          : (IsRoundingAdd ? ISD::AVGCEILU : ISD::AVGFLOORU);
    return DAG.getNode(Opcode, DL, Op.getValueType(), Op.getOperand(1),
                       Op.getOperand(2));
  }
  case Intrinsic::aarch64_neon_saddlp:
  case Intrinsic::aarch64_neon_uaddlp: {
    unsigned Opcode = IntNo == Intrinsic::aarch64_neon_uaddlp
                          ? AArch64ISD::UADDLP
                          : AArch64ISD::SADDLP;
    return DAG.getNode(Opcode, DL, Op.getValueType(), Op.getOperand(1));
  }
  case Intrinsic::aarch64_neon_sdot:
  case Intrinsic::aarch64_neon_udot:
  case Intrinsic::aarch64_sve_sdot:
  case Intrinsic::aarch64_sve_udot: {
    unsigned Opcode = (IntNo == Intrinsic::aarch64_neon_udot ||
                       IntNo == Intrinsic::aarch64_sve_udot)
                          ? AArch64ISD::UDOT
                          : AArch64ISD::SDOT;
    return DAG.getNode(Opcode, DL, Op.getValueType(), Op.getOperand(1),
                       Op.getOperand(2), Op.getOperand(3));
  }
  case Intrinsic::aarch64_neon_usdot:
  case Intrinsic::aarch64_sve_usdot: {
    return DAG.getNode(AArch64ISD::USDOT, DL, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
  }
  case Intrinsic::aarch64_neon_saddlv:
  case Intrinsic::aarch64_neon_uaddlv: {
    EVT OpVT = Op.getOperand(1).getValueType();
    EVT ResVT = Op.getValueType();
    assert(
        ((ResVT == MVT::i32 && (OpVT == MVT::v8i8 || OpVT == MVT::v16i8 ||
                                OpVT == MVT::v8i16 || OpVT == MVT::v4i16)) ||
         (ResVT == MVT::i64 && (OpVT == MVT::v4i32 || OpVT == MVT::v2i32))) &&
        "Unexpected aarch64_neon_u/saddlv type");
    (void)OpVT;
    // In order to avoid insert_subvector, use v4i32 rather than v2i32.
    SDValue ADDLV = DAG.getNode(
        IntNo == Intrinsic::aarch64_neon_uaddlv ? AArch64ISD::UADDLV
                                                : AArch64ISD::SADDLV,
        DL, ResVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64, Op.getOperand(1));
    SDValue EXTRACT_VEC_ELT = DAG.getNode(
        ISD::EXTRACT_VECTOR_ELT, DL, ResVT == MVT::i32 ? MVT::i32 : MVT::i64,
        ADDLV, DAG.getConstant(0, DL, MVT::i64));
    return EXTRACT_VEC_ELT;
  }
  case Intrinsic::experimental_cttz_elts: {
    SDValue CttzOp = Op.getOperand(1);
    EVT VT = CttzOp.getValueType();
    assert(VT.getVectorElementType() == MVT::i1 && "Expected MVT::i1");

    if (VT.isFixedLengthVector()) {
      // We can use SVE instructions to lower this intrinsic by first creating
      // an SVE predicate register mask from the fixed-width vector.
      EVT NewVT = getTypeToTransformTo(*DAG.getContext(), VT);
      SDValue Mask = DAG.getNode(ISD::SIGN_EXTEND, DL, NewVT, CttzOp);
      CttzOp = convertFixedMaskToScalableVector(Mask, DAG);
    }

    SDValue NewCttzElts =
        DAG.getNode(AArch64ISD::CTTZ_ELTS, DL, MVT::i64, CttzOp);
    return DAG.getZExtOrTrunc(NewCttzElts, DL, Op.getValueType());
  }
  case Intrinsic::experimental_vector_match: {
    return LowerVectorMatch(Op, DAG);
  }
  }
}

bool AArch64TargetLowering::shouldExtendGSIndex(EVT VT, EVT &EltTy) const {
  if (VT.getVectorElementType() == MVT::i8 ||
      VT.getVectorElementType() == MVT::i16) {
    EltTy = MVT::i32;
    return true;
  }
  return false;
}

bool AArch64TargetLowering::shouldRemoveExtendFromGSIndex(SDValue Extend,
                                                          EVT DataVT) const {
  const EVT IndexVT = Extend.getOperand(0).getValueType();
  // SVE only supports implicit extension of 32-bit indices.
  if (!Subtarget->hasSVE() || IndexVT.getVectorElementType() != MVT::i32)
    return false;

  // Indices cannot be smaller than the main data type.
  if (IndexVT.getScalarSizeInBits() < DataVT.getScalarSizeInBits())
    return false;

  // Scalable vectors with "vscale * 2" or fewer elements sit within a 64-bit
  // element container type, which would violate the previous clause.
  return DataVT.isFixedLengthVector() || DataVT.getVectorMinNumElements() > 2;
}

bool AArch64TargetLowering::isVectorLoadExtDesirable(SDValue ExtVal) const {
  EVT ExtVT = ExtVal.getValueType();
  if (!ExtVT.isScalableVector() && !Subtarget->useSVEForFixedLengthVectors())
    return false;

  // It may be worth creating extending masked loads if there are multiple
  // masked loads using the same predicate. That way we'll end up creating
  // extending masked loads that may then get split by the legaliser. This
  // results in just one set of predicate unpacks at the start, instead of
  // multiple sets of vector unpacks after each load.
  if (auto *Ld = dyn_cast<MaskedLoadSDNode>(ExtVal->getOperand(0))) {
    if (!isLoadExtLegalOrCustom(ISD::ZEXTLOAD, ExtVT, Ld->getValueType(0))) {
      // Disable extending masked loads for fixed-width for now, since the code
      // quality doesn't look great.
      if (!ExtVT.isScalableVector())
        return false;

      unsigned NumExtMaskedLoads = 0;
      for (auto *U : Ld->getMask()->users())
        if (isa<MaskedLoadSDNode>(U))
          NumExtMaskedLoads++;

      if (NumExtMaskedLoads <= 1)
        return false;
    }
  }

  EVT PreExtScalarVT = ExtVal->getOperand(0).getValueType().getScalarType();
  return PreExtScalarVT == MVT::i8 || PreExtScalarVT == MVT::i16 ||
         PreExtScalarVT == MVT::i32 || PreExtScalarVT == MVT::i64;
}

unsigned getGatherVecOpcode(bool IsScaled, bool IsSigned, bool NeedsExtend) {
  std::map<std::tuple<bool, bool, bool>, unsigned> AddrModes = {
      {std::make_tuple(/*Scaled*/ false, /*Signed*/ false, /*Extend*/ false),
       AArch64ISD::GLD1_MERGE_ZERO},
      {std::make_tuple(/*Scaled*/ false, /*Signed*/ false, /*Extend*/ true),
       AArch64ISD::GLD1_UXTW_MERGE_ZERO},
      {std::make_tuple(/*Scaled*/ false, /*Signed*/ true, /*Extend*/ false),
       AArch64ISD::GLD1_MERGE_ZERO},
      {std::make_tuple(/*Scaled*/ false, /*Signed*/ true, /*Extend*/ true),
       AArch64ISD::GLD1_SXTW_MERGE_ZERO},
      {std::make_tuple(/*Scaled*/ true, /*Signed*/ false, /*Extend*/ false),
       AArch64ISD::GLD1_SCALED_MERGE_ZERO},
      {std::make_tuple(/*Scaled*/ true, /*Signed*/ false, /*Extend*/ true),
       AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO},
      {std::make_tuple(/*Scaled*/ true, /*Signed*/ true, /*Extend*/ false),
       AArch64ISD::GLD1_SCALED_MERGE_ZERO},
      {std::make_tuple(/*Scaled*/ true, /*Signed*/ true, /*Extend*/ true),
       AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO},
  };
  auto Key = std::make_tuple(IsScaled, IsSigned, NeedsExtend);
  return AddrModes.find(Key)->second;
}

unsigned getSignExtendedGatherOpcode(unsigned Opcode) {
  switch (Opcode) {
  default:
    llvm_unreachable("unimplemented opcode");
    return Opcode;
  case AArch64ISD::GLD1_MERGE_ZERO:
    return AArch64ISD::GLD1S_MERGE_ZERO;
  case AArch64ISD::GLD1_IMM_MERGE_ZERO:
    return AArch64ISD::GLD1S_IMM_MERGE_ZERO;
  case AArch64ISD::GLD1_UXTW_MERGE_ZERO:
    return AArch64ISD::GLD1S_UXTW_MERGE_ZERO;
  case AArch64ISD::GLD1_SXTW_MERGE_ZERO:
    return AArch64ISD::GLD1S_SXTW_MERGE_ZERO;
  case AArch64ISD::GLD1_SCALED_MERGE_ZERO:
    return AArch64ISD::GLD1S_SCALED_MERGE_ZERO;
  case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO:
    return AArch64ISD::GLD1S_UXTW_SCALED_MERGE_ZERO;
  case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO:
    return AArch64ISD::GLD1S_SXTW_SCALED_MERGE_ZERO;
  }
}

SDValue AArch64TargetLowering::LowerMGATHER(SDValue Op,
                                            SelectionDAG &DAG) const {
  MaskedGatherSDNode *MGT = cast<MaskedGatherSDNode>(Op);

  SDLoc DL(Op);
  SDValue Chain = MGT->getChain();
  SDValue PassThru = MGT->getPassThru();
  SDValue Mask = MGT->getMask();
  SDValue BasePtr = MGT->getBasePtr();
  SDValue Index = MGT->getIndex();
  SDValue Scale = MGT->getScale();
  EVT VT = Op.getValueType();
  EVT MemVT = MGT->getMemoryVT();
  ISD::LoadExtType ExtType = MGT->getExtensionType();
  ISD::MemIndexType IndexType = MGT->getIndexType();

  // SVE supports zero (and so undef) passthrough values only, everything else
  // must be handled manually by an explicit select on the load's output.
  if (!PassThru->isUndef() && !isZerosVector(PassThru.getNode())) {
    SDValue Ops[] = {Chain, DAG.getUNDEF(VT), Mask, BasePtr, Index, Scale};
    SDValue Load =
        DAG.getMaskedGather(MGT->getVTList(), MemVT, DL, Ops,
                            MGT->getMemOperand(), IndexType, ExtType);
    SDValue Select = DAG.getSelect(DL, VT, Mask, Load, PassThru);
    return DAG.getMergeValues({Select, Load.getValue(1)}, DL);
  }

  bool IsScaled = MGT->isIndexScaled();
  bool IsSigned = MGT->isIndexSigned();

  // SVE supports an index scaled by sizeof(MemVT.elt) only, everything else
  // must be calculated before hand.
  uint64_t ScaleVal = Scale->getAsZExtVal();
  if (IsScaled && ScaleVal != MemVT.getScalarStoreSize()) {
    assert(isPowerOf2_64(ScaleVal) && "Expecting power-of-two types");
    EVT IndexVT = Index.getValueType();
    Index = DAG.getNode(ISD::SHL, DL, IndexVT, Index,
                        DAG.getConstant(Log2_32(ScaleVal), DL, IndexVT));
    Scale = DAG.getTargetConstant(1, DL, Scale.getValueType());

    SDValue Ops[] = {Chain, PassThru, Mask, BasePtr, Index, Scale};
    return DAG.getMaskedGather(MGT->getVTList(), MemVT, DL, Ops,
                               MGT->getMemOperand(), IndexType, ExtType);
  }

  // Lower fixed length gather to a scalable equivalent.
  if (VT.isFixedLengthVector()) {
    assert(Subtarget->useSVEForFixedLengthVectors() &&
           "Cannot lower when not using SVE for fixed vectors!");

    // NOTE: Handle floating-point as if integer then bitcast the result.
    EVT DataVT = VT.changeVectorElementTypeToInteger();
    MemVT = MemVT.changeVectorElementTypeToInteger();

    // Find the smallest integer fixed length vector we can use for the gather.
    EVT PromotedVT = VT.changeVectorElementType(MVT::i32);
    if (DataVT.getVectorElementType() == MVT::i64 ||
        Index.getValueType().getVectorElementType() == MVT::i64 ||
        Mask.getValueType().getVectorElementType() == MVT::i64)
      PromotedVT = VT.changeVectorElementType(MVT::i64);

    // Promote vector operands except for passthrough, which we know is either
    // undef or zero, and thus best constructed directly.
    unsigned ExtOpcode = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
    Index = DAG.getNode(ExtOpcode, DL, PromotedVT, Index);
    Mask = DAG.getNode(ISD::SIGN_EXTEND, DL, PromotedVT, Mask);

    // A promoted result type forces the need for an extending load.
    if (PromotedVT != DataVT && ExtType == ISD::NON_EXTLOAD)
      ExtType = ISD::EXTLOAD;

    EVT ContainerVT = getContainerForFixedLengthVector(DAG, PromotedVT);

    // Convert fixed length vector operands to scalable.
    MemVT = ContainerVT.changeVectorElementType(MemVT.getVectorElementType());
    Index = convertToScalableVector(DAG, ContainerVT, Index);
    Mask = convertFixedMaskToScalableVector(Mask, DAG);
    PassThru = PassThru->isUndef() ? DAG.getUNDEF(ContainerVT)
                                   : DAG.getConstant(0, DL, ContainerVT);

    // Emit equivalent scalable vector gather.
    SDValue Ops[] = {Chain, PassThru, Mask, BasePtr, Index, Scale};
    SDValue Load =
        DAG.getMaskedGather(DAG.getVTList(ContainerVT, MVT::Other), MemVT, DL,
                            Ops, MGT->getMemOperand(), IndexType, ExtType);

    // Extract fixed length data then convert to the required result type.
    SDValue Result = convertFromScalableVector(DAG, PromotedVT, Load);
    Result = DAG.getNode(ISD::TRUNCATE, DL, DataVT, Result);
    if (VT.isFloatingPoint())
      Result = DAG.getNode(ISD::BITCAST, DL, VT, Result);

    return DAG.getMergeValues({Result, Load.getValue(1)}, DL);
  }

  // Everything else is legal.
  return Op;
}

SDValue AArch64TargetLowering::LowerMSCATTER(SDValue Op,
                                             SelectionDAG &DAG) const {
  MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(Op);

  SDLoc DL(Op);
  SDValue Chain = MSC->getChain();
  SDValue StoreVal = MSC->getValue();
  SDValue Mask = MSC->getMask();
  SDValue BasePtr = MSC->getBasePtr();
  SDValue Index = MSC->getIndex();
  SDValue Scale = MSC->getScale();
  EVT VT = StoreVal.getValueType();
  EVT MemVT = MSC->getMemoryVT();
  ISD::MemIndexType IndexType = MSC->getIndexType();
  bool Truncating = MSC->isTruncatingStore();

  bool IsScaled = MSC->isIndexScaled();
  bool IsSigned = MSC->isIndexSigned();

  // SVE supports an index scaled by sizeof(MemVT.elt) only, everything else
  // must be calculated before hand.
  uint64_t ScaleVal = Scale->getAsZExtVal();
  if (IsScaled && ScaleVal != MemVT.getScalarStoreSize()) {
    assert(isPowerOf2_64(ScaleVal) && "Expecting power-of-two types");
    EVT IndexVT = Index.getValueType();
    Index = DAG.getNode(ISD::SHL, DL, IndexVT, Index,
                        DAG.getConstant(Log2_32(ScaleVal), DL, IndexVT));
    Scale = DAG.getTargetConstant(1, DL, Scale.getValueType());

    SDValue Ops[] = {Chain, StoreVal, Mask, BasePtr, Index, Scale};
    return DAG.getMaskedScatter(MSC->getVTList(), MemVT, DL, Ops,
                                MSC->getMemOperand(), IndexType, Truncating);
  }

  // Lower fixed length scatter to a scalable equivalent.
  if (VT.isFixedLengthVector()) {
    assert(Subtarget->useSVEForFixedLengthVectors() &&
           "Cannot lower when not using SVE for fixed vectors!");

    // Once bitcast we treat floating-point scatters as if integer.
    if (VT.isFloatingPoint()) {
      VT = VT.changeVectorElementTypeToInteger();
      MemVT = MemVT.changeVectorElementTypeToInteger();
      StoreVal = DAG.getNode(ISD::BITCAST, DL, VT, StoreVal);
    }

    // Find the smallest integer fixed length vector we can use for the scatter.
    EVT PromotedVT = VT.changeVectorElementType(MVT::i32);
    if (VT.getVectorElementType() == MVT::i64 ||
        Index.getValueType().getVectorElementType() == MVT::i64 ||
        Mask.getValueType().getVectorElementType() == MVT::i64)
      PromotedVT = VT.changeVectorElementType(MVT::i64);

    // Promote vector operands.
    unsigned ExtOpcode = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
    Index = DAG.getNode(ExtOpcode, DL, PromotedVT, Index);
    Mask = DAG.getNode(ISD::SIGN_EXTEND, DL, PromotedVT, Mask);
    StoreVal = DAG.getNode(ISD::ANY_EXTEND, DL, PromotedVT, StoreVal);

    // A promoted value type forces the need for a truncating store.
    if (PromotedVT != VT)
      Truncating = true;

    EVT ContainerVT = getContainerForFixedLengthVector(DAG, PromotedVT);

    // Convert fixed length vector operands to scalable.
    MemVT = ContainerVT.changeVectorElementType(MemVT.getVectorElementType());
    Index = convertToScalableVector(DAG, ContainerVT, Index);
    Mask = convertFixedMaskToScalableVector(Mask, DAG);
    StoreVal = convertToScalableVector(DAG, ContainerVT, StoreVal);

    // Emit equivalent scalable vector scatter.
    SDValue Ops[] = {Chain, StoreVal, Mask, BasePtr, Index, Scale};
    return DAG.getMaskedScatter(MSC->getVTList(), MemVT, DL, Ops,
                                MSC->getMemOperand(), IndexType, Truncating);
  }

  // Everything else is legal.
  return Op;
}

SDValue AArch64TargetLowering::LowerMLOAD(SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  MaskedLoadSDNode *LoadNode = cast<MaskedLoadSDNode>(Op);
  assert(LoadNode && "Expected custom lowering of a masked load node");
  EVT VT = Op->getValueType(0);

  if (useSVEForFixedLengthVectorVT(VT, /*OverrideNEON=*/true))
    return LowerFixedLengthVectorMLoadToSVE(Op, DAG);

  SDValue PassThru = LoadNode->getPassThru();
  SDValue Mask = LoadNode->getMask();

  if (PassThru->isUndef() || isZerosVector(PassThru.getNode()))
    return Op;

  SDValue Load = DAG.getMaskedLoad(
      VT, DL, LoadNode->getChain(), LoadNode->getBasePtr(),
      LoadNode->getOffset(), Mask, DAG.getUNDEF(VT), LoadNode->getMemoryVT(),
      LoadNode->getMemOperand(), LoadNode->getAddressingMode(),
      LoadNode->getExtensionType());

  SDValue Result = DAG.getSelect(DL, VT, Mask, Load, PassThru);

  return DAG.getMergeValues({Result, Load.getValue(1)}, DL);
}

// Custom lower trunc store for v4i8 vectors, since it is promoted to v4i16.
static SDValue LowerTruncateVectorStore(SDLoc DL, StoreSDNode *ST,
                                        EVT VT, EVT MemVT,
                                        SelectionDAG &DAG) {
  assert(VT.isVector() && "VT should be a vector type");
  assert(MemVT == MVT::v4i8 && VT == MVT::v4i16);

  SDValue Value = ST->getValue();

  // It first extend the promoted v4i16 to v8i16, truncate to v8i8, and extract
  // the word lane which represent the v4i8 subvector.  It optimizes the store
  // to:
  //
  //   xtn  v0.8b, v0.8h
  //   str  s0, [x0]

  SDValue Undef = DAG.getUNDEF(MVT::i16);
  SDValue UndefVec = DAG.getBuildVector(MVT::v4i16, DL,
                                        {Undef, Undef, Undef, Undef});

  SDValue TruncExt = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i16,
                                 Value, UndefVec);
  SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, MVT::v8i8, TruncExt);

  Trunc = DAG.getNode(ISD::BITCAST, DL, MVT::v2i32, Trunc);
  SDValue ExtractTrunc = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i32,
                                     Trunc, DAG.getConstant(0, DL, MVT::i64));

  return DAG.getStore(ST->getChain(), DL, ExtractTrunc,
                      ST->getBasePtr(), ST->getMemOperand());
}

static SDValue LowerADDRSPACECAST(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  SDValue Src = Op.getOperand(0);
  MVT DestVT = Op.getSimpleValueType();
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  AddrSpaceCastSDNode *N = cast<AddrSpaceCastSDNode>(Op.getNode());

  unsigned SrcAS = N->getSrcAddressSpace();
  unsigned DestAS = N->getDestAddressSpace();
  assert(SrcAS != DestAS &&
         "addrspacecast must be between different address spaces");
  assert(TLI.getTargetMachine().getPointerSize(SrcAS) !=
             TLI.getTargetMachine().getPointerSize(DestAS) &&
         "addrspacecast must be between different ptr sizes");
  (void)TLI;

  if (SrcAS == ARM64AS::PTR32_SPTR) {
    return DAG.getNode(ISD::SIGN_EXTEND, DL, DestVT, Src,
                       DAG.getTargetConstant(0, DL, DestVT));
  } else if (SrcAS == ARM64AS::PTR32_UPTR) {
    return DAG.getNode(ISD::ZERO_EXTEND, DL, DestVT, Src,
                       DAG.getTargetConstant(0, DL, DestVT));
  } else if ((DestAS == ARM64AS::PTR32_SPTR) ||
             (DestAS == ARM64AS::PTR32_UPTR)) {
    SDValue Ext = DAG.getAnyExtOrTrunc(Src, DL, DestVT);
    SDValue Trunc = DAG.getZeroExtendInReg(Ext, DL, DestVT);
    return Trunc;
  } else {
    return Src;
  }
}

// Custom lowering for any store, vector or scalar and/or default or with
// a truncate operations.  Currently only custom lower truncate operation
// from vector v4i16 to v4i8 or volatile stores of i128.
SDValue AArch64TargetLowering::LowerSTORE(SDValue Op,
                                          SelectionDAG &DAG) const {
  SDLoc Dl(Op);
  StoreSDNode *StoreNode = cast<StoreSDNode>(Op);
  assert (StoreNode && "Can only custom lower store nodes");

  SDValue Value = StoreNode->getValue();

  EVT VT = Value.getValueType();
  EVT MemVT = StoreNode->getMemoryVT();

  if (VT.isVector()) {
    if (useSVEForFixedLengthVectorVT(
            VT,
            /*OverrideNEON=*/Subtarget->useSVEForFixedLengthVectors()))
      return LowerFixedLengthVectorStoreToSVE(Op, DAG);

    unsigned AS = StoreNode->getAddressSpace();
    Align Alignment = StoreNode->getAlign();
    if (Alignment < MemVT.getStoreSize() &&
        !allowsMisalignedMemoryAccesses(MemVT, AS, Alignment,
                                        StoreNode->getMemOperand()->getFlags(),
                                        nullptr)) {
      return scalarizeVectorStore(StoreNode, DAG);
    }

    if (StoreNode->isTruncatingStore() && VT == MVT::v4i16 &&
        MemVT == MVT::v4i8) {
      return LowerTruncateVectorStore(Dl, StoreNode, VT, MemVT, DAG);
    }
    // 256 bit non-temporal stores can be lowered to STNP. Do this as part of
    // the custom lowering, as there are no un-paired non-temporal stores and
    // legalization will break up 256 bit inputs.
    ElementCount EC = MemVT.getVectorElementCount();
    if (StoreNode->isNonTemporal() && MemVT.getSizeInBits() == 256u &&
        EC.isKnownEven() && DAG.getDataLayout().isLittleEndian() &&
        (MemVT.getScalarSizeInBits() == 8u ||
         MemVT.getScalarSizeInBits() == 16u ||
         MemVT.getScalarSizeInBits() == 32u ||
         MemVT.getScalarSizeInBits() == 64u)) {
      SDValue Lo =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, Dl,
                      MemVT.getHalfNumVectorElementsVT(*DAG.getContext()),
                      StoreNode->getValue(), DAG.getConstant(0, Dl, MVT::i64));
      SDValue Hi =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, Dl,
                      MemVT.getHalfNumVectorElementsVT(*DAG.getContext()),
                      StoreNode->getValue(),
                      DAG.getConstant(EC.getKnownMinValue() / 2, Dl, MVT::i64));
      SDValue Result = DAG.getMemIntrinsicNode(
          AArch64ISD::STNP, Dl, DAG.getVTList(MVT::Other),
          {StoreNode->getChain(), Lo, Hi, StoreNode->getBasePtr()},
          StoreNode->getMemoryVT(), StoreNode->getMemOperand());
      return Result;
    }
  } else if (MemVT == MVT::i128 && StoreNode->isVolatile()) {
    return LowerStore128(Op, DAG);
  } else if (MemVT == MVT::i64x8) {
    SDValue Value = StoreNode->getValue();
    assert(Value->getValueType(0) == MVT::i64x8);
    SDValue Chain = StoreNode->getChain();
    SDValue Base = StoreNode->getBasePtr();
    EVT PtrVT = Base.getValueType();
    for (unsigned i = 0; i < 8; i++) {
      SDValue Part = DAG.getNode(AArch64ISD::LS64_EXTRACT, Dl, MVT::i64,
                                 Value, DAG.getConstant(i, Dl, MVT::i32));
      SDValue Ptr = DAG.getNode(ISD::ADD, Dl, PtrVT, Base,
                                DAG.getConstant(i * 8, Dl, PtrVT));
      Chain = DAG.getStore(Chain, Dl, Part, Ptr, StoreNode->getPointerInfo(),
                           StoreNode->getBaseAlign());
    }
    return Chain;
  }

  return SDValue();
}

/// Lower atomic or volatile 128-bit stores to a single STP instruction.
SDValue AArch64TargetLowering::LowerStore128(SDValue Op,
                                             SelectionDAG &DAG) const {
  MemSDNode *StoreNode = cast<MemSDNode>(Op);
  assert(StoreNode->getMemoryVT() == MVT::i128);
  assert(StoreNode->isVolatile() || StoreNode->isAtomic());

  bool IsStoreRelease =
      StoreNode->getMergedOrdering() == AtomicOrdering::Release;
  if (StoreNode->isAtomic())
    assert((Subtarget->hasFeature(AArch64::FeatureLSE2) &&
            Subtarget->hasFeature(AArch64::FeatureRCPC3) && IsStoreRelease) ||
           StoreNode->getMergedOrdering() == AtomicOrdering::Unordered ||
           StoreNode->getMergedOrdering() == AtomicOrdering::Monotonic);

  SDValue Value = (StoreNode->getOpcode() == ISD::STORE ||
                   StoreNode->getOpcode() == ISD::ATOMIC_STORE)
                      ? StoreNode->getOperand(1)
                      : StoreNode->getOperand(2);
  SDLoc DL(Op);
  auto StoreValue = DAG.SplitScalar(Value, DL, MVT::i64, MVT::i64);
  unsigned Opcode = IsStoreRelease ? AArch64ISD::STILP : AArch64ISD::STP;
  if (DAG.getDataLayout().isBigEndian())
    std::swap(StoreValue.first, StoreValue.second);
  SDValue Result = DAG.getMemIntrinsicNode(
      Opcode, DL, DAG.getVTList(MVT::Other),
      {StoreNode->getChain(), StoreValue.first, StoreValue.second,
       StoreNode->getBasePtr()},
      StoreNode->getMemoryVT(), StoreNode->getMemOperand());
  return Result;
}

SDValue AArch64TargetLowering::LowerLOAD(SDValue Op,
                                         SelectionDAG &DAG) const {
  SDLoc DL(Op);
  LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
  assert(LoadNode && "Expected custom lowering of a load node");

  if (LoadNode->getMemoryVT() == MVT::i64x8) {
    SmallVector<SDValue, 8> Ops;
    SDValue Base = LoadNode->getBasePtr();
    SDValue Chain = LoadNode->getChain();
    EVT PtrVT = Base.getValueType();
    for (unsigned i = 0; i < 8; i++) {
      SDValue Ptr = DAG.getNode(ISD::ADD, DL, PtrVT, Base,
                                DAG.getConstant(i * 8, DL, PtrVT));
      SDValue Part =
          DAG.getLoad(MVT::i64, DL, Chain, Ptr, LoadNode->getPointerInfo(),
                      LoadNode->getBaseAlign());
      Ops.push_back(Part);
      Chain = SDValue(Part.getNode(), 1);
    }
    SDValue Loaded = DAG.getNode(AArch64ISD::LS64_BUILD, DL, MVT::i64x8, Ops);
    return DAG.getMergeValues({Loaded, Chain}, DL);
  }

  // Custom lowering for extending v4i8 vector loads.
  EVT VT = Op->getValueType(0);
  assert((VT == MVT::v4i16 || VT == MVT::v4i32) && "Expected v4i16 or v4i32");

  if (LoadNode->getMemoryVT() != MVT::v4i8)
    return SDValue();

  // Avoid generating unaligned loads.
  if (Subtarget->requiresStrictAlign() && LoadNode->getAlign() < Align(4))
    return SDValue();

  unsigned ExtType;
  if (LoadNode->getExtensionType() == ISD::SEXTLOAD)
    ExtType = ISD::SIGN_EXTEND;
  else if (LoadNode->getExtensionType() == ISD::ZEXTLOAD ||
           LoadNode->getExtensionType() == ISD::EXTLOAD)
    ExtType = ISD::ZERO_EXTEND;
  else
    return SDValue();

  SDValue Load = DAG.getLoad(MVT::f32, DL, LoadNode->getChain(),
                             LoadNode->getBasePtr(), MachinePointerInfo());
  SDValue Chain = Load.getValue(1);
  SDValue Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f32, Load);
  SDValue BC = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Vec);
  SDValue Ext = DAG.getNode(ExtType, DL, MVT::v8i16, BC);
  Ext = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Ext,
                    DAG.getConstant(0, DL, MVT::i64));
  if (VT == MVT::v4i32)
    Ext = DAG.getNode(ExtType, DL, MVT::v4i32, Ext);
  return DAG.getMergeValues({Ext, Chain}, DL);
}

SDValue AArch64TargetLowering::LowerVECTOR_COMPRESS(SDValue Op,
                                                    SelectionDAG &DAG) const {
  SDLoc DL(Op);
  SDValue Vec = Op.getOperand(0);
  SDValue Mask = Op.getOperand(1);
  SDValue Passthru = Op.getOperand(2);
  EVT VecVT = Vec.getValueType();
  EVT MaskVT = Mask.getValueType();
  EVT ElmtVT = VecVT.getVectorElementType();
  const bool IsFixedLength = VecVT.isFixedLengthVector();
  const bool HasPassthru = !Passthru.isUndef();
  unsigned MinElmts = VecVT.getVectorElementCount().getKnownMinValue();
  EVT FixedVecVT = MVT::getVectorVT(ElmtVT.getSimpleVT(), MinElmts);

  assert(VecVT.isVector() && "Input to VECTOR_COMPRESS must be vector.");

  if (!Subtarget->isSVEAvailable())
    return SDValue();

  if (IsFixedLength && VecVT.getSizeInBits().getFixedValue() > 128)
    return SDValue();

  // Only <vscale x {4|2} x {i32|i64}> supported for compact.
  if (MinElmts != 2 && MinElmts != 4)
    return SDValue();

  // We can use the SVE register containing the NEON vector in its lowest bits.
  if (IsFixedLength) {
    EVT ScalableVecVT =
        MVT::getScalableVectorVT(ElmtVT.getSimpleVT(), MinElmts);
    EVT ScalableMaskVT = MVT::getScalableVectorVT(
        MaskVT.getVectorElementType().getSimpleVT(), MinElmts);

    Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ScalableVecVT,
                      DAG.getUNDEF(ScalableVecVT), Vec,
                      DAG.getConstant(0, DL, MVT::i64));
    Mask = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ScalableMaskVT,
                       DAG.getUNDEF(ScalableMaskVT), Mask,
                       DAG.getConstant(0, DL, MVT::i64));
    Mask = DAG.getNode(ISD::TRUNCATE, DL,
                       ScalableMaskVT.changeVectorElementType(MVT::i1), Mask);
    Passthru = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, ScalableVecVT,
                           DAG.getUNDEF(ScalableVecVT), Passthru,
                           DAG.getConstant(0, DL, MVT::i64));

    VecVT = Vec.getValueType();
    MaskVT = Mask.getValueType();
  }

  // Get legal type for compact instruction
  EVT ContainerVT = getSVEContainerType(VecVT);
  EVT CastVT = VecVT.changeVectorElementTypeToInteger();

  // Convert to i32 or i64 for smaller types, as these are the only supported
  // sizes for compact.
  if (ContainerVT != VecVT) {
    Vec = DAG.getBitcast(CastVT, Vec);
    Vec = DAG.getNode(ISD::ANY_EXTEND, DL, ContainerVT, Vec);
  }

  SDValue Compressed = DAG.getNode(
      ISD::INTRINSIC_WO_CHAIN, DL, Vec.getValueType(),
      DAG.getConstant(Intrinsic::aarch64_sve_compact, DL, MVT::i64), Mask, Vec);

  // compact fills with 0s, so if our passthru is all 0s, do nothing here.
  if (HasPassthru && !ISD::isConstantSplatVectorAllZeros(Passthru.getNode())) {
    SDValue Offset = DAG.getNode(
        ISD::INTRINSIC_WO_CHAIN, DL, MVT::i64,
        DAG.getConstant(Intrinsic::aarch64_sve_cntp, DL, MVT::i64), Mask, Mask);

    SDValue IndexMask = DAG.getNode(
        ISD::INTRINSIC_WO_CHAIN, DL, MaskVT,
        DAG.getConstant(Intrinsic::aarch64_sve_whilelo, DL, MVT::i64),
        DAG.getConstant(0, DL, MVT::i64), Offset);

    Compressed =
        DAG.getNode(ISD::VSELECT, DL, VecVT, IndexMask, Compressed, Passthru);
  }

  // Extracting from a legal SVE type before truncating produces better code.
  if (IsFixedLength) {
    Compressed = DAG.getNode(
        ISD::EXTRACT_SUBVECTOR, DL,
        FixedVecVT.changeVectorElementType(ContainerVT.getVectorElementType()),
        Compressed, DAG.getConstant(0, DL, MVT::i64));
    CastVT = FixedVecVT.changeVectorElementTypeToInteger();
    VecVT = FixedVecVT;
  }

  // If we changed the element type before, we need to convert it back.
  if (ContainerVT != VecVT) {
    Compressed = DAG.getNode(ISD::TRUNCATE, DL, CastVT, Compressed);
    Compressed = DAG.getBitcast(VecVT, Compressed);
  }

  return Compressed;
}

// Generate SUBS and CSEL for integer abs.
SDValue AArch64TargetLowering::LowerABS(SDValue Op, SelectionDAG &DAG) const {
  MVT VT = Op.getSimpleValueType();

  if (VT.isVector())
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::ABS_MERGE_PASSTHRU);

  SDLoc DL(Op);
  SDValue Neg = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
                            Op.getOperand(0));
  // Generate SUBS & CSEL.
  SDValue Cmp =
      DAG.getNode(AArch64ISD::SUBS, DL, DAG.getVTList(VT, MVT::i32),
                  Op.getOperand(0), DAG.getConstant(0, DL, VT));
  return DAG.getNode(AArch64ISD::CSEL, DL, VT, Op.getOperand(0), Neg,
                     DAG.getConstant(AArch64CC::PL, DL, MVT::i32),
                     Cmp.getValue(1));
}

static SDValue LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
  SDValue Chain = Op.getOperand(0);
  SDValue Cond = Op.getOperand(1);
  SDValue Dest = Op.getOperand(2);

  AArch64CC::CondCode CC;
  if (SDValue Cmp = emitConjunction(DAG, Cond, CC)) {
    SDLoc DL(Op);
    SDValue CCVal = DAG.getConstant(CC, DL, MVT::i32);
    return DAG.getNode(AArch64ISD::BRCOND, DL, MVT::Other, Chain, Dest, CCVal,
                       Cmp);
  }

  return SDValue();
}

// Treat FSHR with constant shifts as legal operation, otherwise it is expanded
// FSHL is converted to FSHR before deciding what to do with it
static SDValue LowerFunnelShift(SDValue Op, SelectionDAG &DAG) {
  SDValue Shifts = Op.getOperand(2);
  // Check if the shift amount is a constant and normalise to [0, SrcBitLen)
  // If opcode is FSHL, convert it to FSHR
  if (auto *ShiftNo = dyn_cast<ConstantSDNode>(Shifts)) {
    SDLoc DL(Op);
    MVT VT = Op.getSimpleValueType();
    unsigned int NewShiftNo = ShiftNo->getZExtValue() % VT.getFixedSizeInBits();

    if (Op.getOpcode() == ISD::FSHL) {
      if (NewShiftNo == 0)
        return Op.getOperand(0);

      NewShiftNo = VT.getFixedSizeInBits() - NewShiftNo;
      return DAG.getNode(
          ISD::FSHR, DL, VT, Op.getOperand(0), Op.getOperand(1),
          DAG.getConstant(NewShiftNo, DL, Shifts.getValueType()));
    }

    if (Op.getOpcode() == ISD::FSHR) {
      if (NewShiftNo == 0)
        return Op.getOperand(1);

      if (ShiftNo->getZExtValue() == NewShiftNo)
        return Op;

      // Rewrite using the normalised shift amount.
      return DAG.getNode(
          ISD::FSHR, DL, VT, Op.getOperand(0), Op.getOperand(1),
          DAG.getConstant(NewShiftNo, DL, Shifts.getValueType()));
    }
  }

  return SDValue();
}

static SDValue LowerFLDEXP(SDValue Op, SelectionDAG &DAG) {
  SDValue X = Op.getOperand(0);
  EVT XScalarTy = X.getValueType();
  SDValue Exp = Op.getOperand(1);

  SDLoc DL(Op);
  EVT XVT, ExpVT;
  switch (Op.getSimpleValueType().SimpleTy) {
  default:
    return SDValue();
  case MVT::bf16:
  case MVT::f16:
    X = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, X);
    [[fallthrough]];
  case MVT::f32:
    XVT = MVT::nxv4f32;
    ExpVT = MVT::nxv4i32;
    break;
  case MVT::f64:
    XVT = MVT::nxv2f64;
    ExpVT = MVT::nxv2i64;
    Exp = DAG.getNode(ISD::SIGN_EXTEND, DL, MVT::i64, Exp);
    break;
  }

  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  SDValue VX =
      DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, XVT, DAG.getUNDEF(XVT), X, Zero);
  SDValue VExp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ExpVT,
                             DAG.getUNDEF(ExpVT), Exp, Zero);
  SDValue VPg = getPTrue(DAG, DL, XVT.changeVectorElementType(MVT::i1),
                         AArch64SVEPredPattern::all);
  SDValue FScale =
      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, XVT,
                  DAG.getConstant(Intrinsic::aarch64_sve_fscale, DL, MVT::i64),
                  VPg, VX, VExp);
  SDValue Final =
      DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, X.getValueType(), FScale, Zero);
  if (X.getValueType() != XScalarTy)
    Final = DAG.getNode(ISD::FP_ROUND, DL, XScalarTy, Final,
                        DAG.getIntPtrConstant(1, SDLoc(Op), /*isTarget=*/true));
  return Final;
}

SDValue AArch64TargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
                                                      SelectionDAG &DAG) const {
  return Op.getOperand(0);
}

SDValue AArch64TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
                                                    SelectionDAG &DAG) const {
  SDValue Chain = Op.getOperand(0);
  SDValue Trmp = Op.getOperand(1); // trampoline, >=32 bytes
  SDValue FPtr = Op.getOperand(2); // nested function
  SDValue Nest = Op.getOperand(3); // 'nest' parameter value

  const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();

  // ldr NestReg, .+16
  // ldr x17, .+20
  // br x17
  // .word 0
  // .nest: .qword nest
  // .fptr: .qword fptr
  SDValue OutChains[5];

  const Function *Func =
      cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
  CallingConv::ID CC = Func->getCallingConv();
  unsigned NestReg;

  switch (CC) {
  default:
    NestReg = 0x0f; // X15
    break;
  case CallingConv::ARM64EC_Thunk_X64:
    // Must be kept in sync with AArch64CallingConv.td
    NestReg = 0x04; // X4
    break;
  }

  const char FptrReg = 0x11; // X17

  SDValue Addr = Trmp;

  SDLoc DL(Op);
  OutChains[0] = DAG.getStore(
      Chain, DL, DAG.getConstant(0x58000080u | NestReg, DL, MVT::i32), Addr,
      MachinePointerInfo(TrmpAddr));

  Addr = DAG.getNode(ISD::ADD, DL, MVT::i64, Trmp,
                     DAG.getConstant(4, DL, MVT::i64));
  OutChains[1] = DAG.getStore(
      Chain, DL, DAG.getConstant(0x580000b0u | FptrReg, DL, MVT::i32), Addr,
      MachinePointerInfo(TrmpAddr, 4));

  Addr = DAG.getNode(ISD::ADD, DL, MVT::i64, Trmp,
                     DAG.getConstant(8, DL, MVT::i64));
  OutChains[2] =
      DAG.getStore(Chain, DL, DAG.getConstant(0xd61f0220u, DL, MVT::i32), Addr,
                   MachinePointerInfo(TrmpAddr, 8));

  Addr = DAG.getNode(ISD::ADD, DL, MVT::i64, Trmp,
                     DAG.getConstant(16, DL, MVT::i64));
  OutChains[3] =
      DAG.getStore(Chain, DL, Nest, Addr, MachinePointerInfo(TrmpAddr, 16));

  Addr = DAG.getNode(ISD::ADD, DL, MVT::i64, Trmp,
                     DAG.getConstant(24, DL, MVT::i64));
  OutChains[4] =
      DAG.getStore(Chain, DL, FPtr, Addr, MachinePointerInfo(TrmpAddr, 24));

  SDValue StoreToken = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, OutChains);

  SDValue EndOfTrmp = DAG.getNode(ISD::ADD, DL, MVT::i64, Trmp,
                                  DAG.getConstant(12, DL, MVT::i64));

  // Call clear cache on the trampoline instructions.
  return DAG.getNode(ISD::CLEAR_CACHE, DL, MVT::Other, StoreToken, Trmp,
                     EndOfTrmp);
}

SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
                                              SelectionDAG &DAG) const {
  LLVM_DEBUG(dbgs() << "Custom lowering: ");
  LLVM_DEBUG(Op.dump());

  switch (Op.getOpcode()) {
  default:
    llvm_unreachable("unimplemented operand");
    return SDValue();
  case ISD::BITCAST:
    return LowerBITCAST(Op, DAG);
  case ISD::GlobalAddress:
    return LowerGlobalAddress(Op, DAG);
  case ISD::GlobalTLSAddress:
    return LowerGlobalTLSAddress(Op, DAG);
  case ISD::PtrAuthGlobalAddress:
    return LowerPtrAuthGlobalAddress(Op, DAG);
  case ISD::ADJUST_TRAMPOLINE:
    return LowerADJUST_TRAMPOLINE(Op, DAG);
  case ISD::INIT_TRAMPOLINE:
    return LowerINIT_TRAMPOLINE(Op, DAG);
  case ISD::SETCC:
  case ISD::STRICT_FSETCC:
  case ISD::STRICT_FSETCCS:
    return LowerSETCC(Op, DAG);
  case ISD::SETCCCARRY:
    return LowerSETCCCARRY(Op, DAG);
  case ISD::BRCOND:
    return LowerBRCOND(Op, DAG);
  case ISD::BR_CC:
    return LowerBR_CC(Op, DAG);
  case ISD::SELECT:
    return LowerSELECT(Op, DAG);
  case ISD::SELECT_CC:
    return LowerSELECT_CC(Op, DAG);
  case ISD::JumpTable:
    return LowerJumpTable(Op, DAG);
  case ISD::BR_JT:
    return LowerBR_JT(Op, DAG);
  case ISD::BRIND:
    return LowerBRIND(Op, DAG);
  case ISD::ConstantPool:
    return LowerConstantPool(Op, DAG);
  case ISD::BlockAddress:
    return LowerBlockAddress(Op, DAG);
  case ISD::VASTART:
    return LowerVASTART(Op, DAG);
  case ISD::VACOPY:
    return LowerVACOPY(Op, DAG);
  case ISD::VAARG:
    return LowerVAARG(Op, DAG);
  case ISD::UADDO_CARRY:
    return lowerADDSUBO_CARRY(Op, DAG, AArch64ISD::ADCS, false /*unsigned*/);
  case ISD::USUBO_CARRY:
    return lowerADDSUBO_CARRY(Op, DAG, AArch64ISD::SBCS, false /*unsigned*/);
  case ISD::SADDO_CARRY:
    return lowerADDSUBO_CARRY(Op, DAG, AArch64ISD::ADCS, true /*signed*/);
  case ISD::SSUBO_CARRY:
    return lowerADDSUBO_CARRY(Op, DAG, AArch64ISD::SBCS, true /*signed*/);
  case ISD::SADDO:
  case ISD::UADDO:
  case ISD::SSUBO:
  case ISD::USUBO:
  case ISD::SMULO:
  case ISD::UMULO:
    return LowerXALUO(Op, DAG);
  case ISD::FADD:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FADD_PRED);
  case ISD::FSUB:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FSUB_PRED);
  case ISD::FMUL:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMUL_PRED);
  case ISD::FMA:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMA_PRED);
  case ISD::FDIV:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FDIV_PRED);
  case ISD::FNEG:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FNEG_MERGE_PASSTHRU);
  case ISD::FCEIL:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FCEIL_MERGE_PASSTHRU);
  case ISD::FFLOOR:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FFLOOR_MERGE_PASSTHRU);
  case ISD::FNEARBYINT:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FNEARBYINT_MERGE_PASSTHRU);
  case ISD::FRINT:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FRINT_MERGE_PASSTHRU);
  case ISD::FROUND:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FROUND_MERGE_PASSTHRU);
  case ISD::FROUNDEVEN:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU);
  case ISD::FTRUNC:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FTRUNC_MERGE_PASSTHRU);
  case ISD::FSQRT:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FSQRT_MERGE_PASSTHRU);
  case ISD::FABS:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FABS_MERGE_PASSTHRU);
  case ISD::FP_ROUND:
  case ISD::STRICT_FP_ROUND:
    return LowerFP_ROUND(Op, DAG);
  case ISD::FP_EXTEND:
  case ISD::STRICT_FP_EXTEND:
    return LowerFP_EXTEND(Op, DAG);
  case ISD::FRAMEADDR:
    return LowerFRAMEADDR(Op, DAG);
  case ISD::SPONENTRY:
    return LowerSPONENTRY(Op, DAG);
  case ISD::RETURNADDR:
    return LowerRETURNADDR(Op, DAG);
  case ISD::ADDROFRETURNADDR:
    return LowerADDROFRETURNADDR(Op, DAG);
  case ISD::CONCAT_VECTORS:
    return LowerCONCAT_VECTORS(Op, DAG);
  case ISD::INSERT_VECTOR_ELT:
    return LowerINSERT_VECTOR_ELT(Op, DAG);
  case ISD::EXTRACT_VECTOR_ELT:
    return LowerEXTRACT_VECTOR_ELT(Op, DAG);
  case ISD::BUILD_VECTOR:
    return LowerBUILD_VECTOR(Op, DAG);
  case ISD::ZERO_EXTEND_VECTOR_INREG:
    return LowerZERO_EXTEND_VECTOR_INREG(Op, DAG);
  case ISD::VECTOR_SHUFFLE:
    return LowerVECTOR_SHUFFLE(Op, DAG);
  case ISD::SPLAT_VECTOR:
    return LowerSPLAT_VECTOR(Op, DAG);
  case ISD::EXTRACT_SUBVECTOR:
    return LowerEXTRACT_SUBVECTOR(Op, DAG);
  case ISD::INSERT_SUBVECTOR:
    return LowerINSERT_SUBVECTOR(Op, DAG);
  case ISD::SDIV:
  case ISD::UDIV:
    return LowerDIV(Op, DAG);
  case ISD::SMIN:
  case ISD::UMIN:
  case ISD::SMAX:
  case ISD::UMAX:
    return LowerMinMax(Op, DAG);
  case ISD::SRA:
  case ISD::SRL:
  case ISD::SHL:
    return LowerVectorSRA_SRL_SHL(Op, DAG);
  case ISD::SHL_PARTS:
  case ISD::SRL_PARTS:
  case ISD::SRA_PARTS:
    return LowerShiftParts(Op, DAG);
  case ISD::CTPOP:
  case ISD::PARITY:
    return LowerCTPOP_PARITY(Op, DAG);
  case ISD::FCOPYSIGN:
    return LowerFCOPYSIGN(Op, DAG);
  case ISD::OR:
    return LowerVectorOR(Op, DAG);
  case ISD::XOR:
    return LowerXOR(Op, DAG);
  case ISD::PREFETCH:
    return LowerPREFETCH(Op, DAG);
  case ISD::SINT_TO_FP:
  case ISD::UINT_TO_FP:
  case ISD::STRICT_SINT_TO_FP:
  case ISD::STRICT_UINT_TO_FP:
    return LowerINT_TO_FP(Op, DAG);
  case ISD::FP_TO_SINT:
  case ISD::FP_TO_UINT:
  case ISD::STRICT_FP_TO_SINT:
  case ISD::STRICT_FP_TO_UINT:
    return LowerFP_TO_INT(Op, DAG);
  case ISD::FP_TO_SINT_SAT:
  case ISD::FP_TO_UINT_SAT:
    return LowerFP_TO_INT_SAT(Op, DAG);
  case ISD::FSINCOS:
    return LowerFSINCOS(Op, DAG);
  case ISD::GET_ROUNDING:
    return LowerGET_ROUNDING(Op, DAG);
  case ISD::SET_ROUNDING:
    return LowerSET_ROUNDING(Op, DAG);
  case ISD::GET_FPMODE:
    return LowerGET_FPMODE(Op, DAG);
  case ISD::SET_FPMODE:
    return LowerSET_FPMODE(Op, DAG);
  case ISD::RESET_FPMODE:
    return LowerRESET_FPMODE(Op, DAG);
  case ISD::MUL:
    return LowerMUL(Op, DAG);
  case ISD::MULHS:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::MULHS_PRED);
  case ISD::MULHU:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::MULHU_PRED);
  case ISD::INTRINSIC_W_CHAIN:
    return LowerINTRINSIC_W_CHAIN(Op, DAG);
  case ISD::INTRINSIC_WO_CHAIN:
    return LowerINTRINSIC_WO_CHAIN(Op, DAG);
  case ISD::INTRINSIC_VOID:
    return LowerINTRINSIC_VOID(Op, DAG);
  case ISD::ATOMIC_STORE:
    if (cast<MemSDNode>(Op)->getMemoryVT() == MVT::i128) {
      assert(Subtarget->hasLSE2() || Subtarget->hasRCPC3());
      return LowerStore128(Op, DAG);
    }
    return SDValue();
  case ISD::STORE:
    return LowerSTORE(Op, DAG);
  case ISD::MSTORE:
    return LowerFixedLengthVectorMStoreToSVE(Op, DAG);
  case ISD::MGATHER:
    return LowerMGATHER(Op, DAG);
  case ISD::MSCATTER:
    return LowerMSCATTER(Op, DAG);
  case ISD::VECREDUCE_SEQ_FADD:
    return LowerVECREDUCE_SEQ_FADD(Op, DAG);
  case ISD::VECREDUCE_ADD:
  case ISD::VECREDUCE_AND:
  case ISD::VECREDUCE_OR:
  case ISD::VECREDUCE_XOR:
  case ISD::VECREDUCE_SMAX:
  case ISD::VECREDUCE_SMIN:
  case ISD::VECREDUCE_UMAX:
  case ISD::VECREDUCE_UMIN:
  case ISD::VECREDUCE_FADD:
  case ISD::VECREDUCE_FMAX:
  case ISD::VECREDUCE_FMIN:
  case ISD::VECREDUCE_FMAXIMUM:
  case ISD::VECREDUCE_FMINIMUM:
    return LowerVECREDUCE(Op, DAG);
  case ISD::ATOMIC_LOAD_AND:
    return LowerATOMIC_LOAD_AND(Op, DAG);
  case ISD::DYNAMIC_STACKALLOC:
    return LowerDYNAMIC_STACKALLOC(Op, DAG);
  case ISD::VSCALE:
    return LowerVSCALE(Op, DAG);
  case ISD::VECTOR_COMPRESS:
    return LowerVECTOR_COMPRESS(Op, DAG);
  case ISD::ANY_EXTEND:
  case ISD::SIGN_EXTEND:
  case ISD::ZERO_EXTEND:
    return LowerFixedLengthVectorIntExtendToSVE(Op, DAG);
  case ISD::ADDRSPACECAST:
    return LowerADDRSPACECAST(Op, DAG);
  case ISD::SIGN_EXTEND_INREG: {
    // Only custom lower when ExtraVT has a legal byte based element type.
    EVT ExtraVT = cast<VTSDNode>(Op.getOperand(1))->getVT();
    EVT ExtraEltVT = ExtraVT.getVectorElementType();
    if ((ExtraEltVT != MVT::i8) && (ExtraEltVT != MVT::i16) &&
        (ExtraEltVT != MVT::i32) && (ExtraEltVT != MVT::i64))
      return SDValue();

    return LowerToPredicatedOp(Op, DAG,
                               AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU);
  }
  case ISD::TRUNCATE:
    return LowerTRUNCATE(Op, DAG);
  case ISD::MLOAD:
    return LowerMLOAD(Op, DAG);
  case ISD::LOAD:
    if (useSVEForFixedLengthVectorVT(Op.getValueType(),
                                     !Subtarget->isNeonAvailable()))
      return LowerFixedLengthVectorLoadToSVE(Op, DAG);
    return LowerLOAD(Op, DAG);
  case ISD::ADD:
  case ISD::AND:
  case ISD::SUB:
    return LowerToScalableOp(Op, DAG);
  case ISD::FMAXIMUM:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMAX_PRED);
  case ISD::FMAXNUM:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMAXNM_PRED);
  case ISD::FMINIMUM:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMIN_PRED);
  case ISD::FMINNUM:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::FMINNM_PRED);
  case ISD::VSELECT:
    return LowerFixedLengthVectorSelectToSVE(Op, DAG);
  case ISD::ABS:
    return LowerABS(Op, DAG);
  case ISD::ABDS:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::ABDS_PRED);
  case ISD::ABDU:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::ABDU_PRED);
  case ISD::AVGFLOORS:
    return LowerAVG(Op, DAG, AArch64ISD::HADDS_PRED);
  case ISD::AVGFLOORU:
    return LowerAVG(Op, DAG, AArch64ISD::HADDU_PRED);
  case ISD::AVGCEILS:
    return LowerAVG(Op, DAG, AArch64ISD::RHADDS_PRED);
  case ISD::AVGCEILU:
    return LowerAVG(Op, DAG, AArch64ISD::RHADDU_PRED);
  case ISD::BITREVERSE:
    return LowerBitreverse(Op, DAG);
  case ISD::BSWAP:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::BSWAP_MERGE_PASSTHRU);
  case ISD::CTLZ:
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::CTLZ_MERGE_PASSTHRU);
  case ISD::CTTZ:
    return LowerCTTZ(Op, DAG);
  case ISD::VECTOR_SPLICE:
    return LowerVECTOR_SPLICE(Op, DAG);
  case ISD::VECTOR_DEINTERLEAVE:
    return LowerVECTOR_DEINTERLEAVE(Op, DAG);
  case ISD::VECTOR_INTERLEAVE:
    return LowerVECTOR_INTERLEAVE(Op, DAG);
  case ISD::GET_ACTIVE_LANE_MASK:
    return LowerGET_ACTIVE_LANE_MASK(Op, DAG);
  case ISD::LRINT:
  case ISD::LLRINT:
    if (Op.getValueType().isVector())
      return LowerVectorXRINT(Op, DAG);
    [[fallthrough]];
  case ISD::LROUND:
  case ISD::LLROUND: {
    assert((Op.getOperand(0).getValueType() == MVT::f16 ||
            Op.getOperand(0).getValueType() == MVT::bf16) &&
           "Expected custom lowering of rounding operations only for f16");
    SDLoc DL(Op);
    SDValue Ext = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, Op.getOperand(0));
    return DAG.getNode(Op.getOpcode(), DL, Op.getValueType(), Ext);
  }
  case ISD::STRICT_LROUND:
  case ISD::STRICT_LLROUND:
  case ISD::STRICT_LRINT:
  case ISD::STRICT_LLRINT: {
    assert((Op.getOperand(1).getValueType() == MVT::f16 ||
            Op.getOperand(1).getValueType() == MVT::bf16) &&
           "Expected custom lowering of rounding operations only for f16");
    SDLoc DL(Op);
    SDValue Ext = DAG.getNode(ISD::STRICT_FP_EXTEND, DL, {MVT::f32, MVT::Other},
                              {Op.getOperand(0), Op.getOperand(1)});
    return DAG.getNode(Op.getOpcode(), DL, {Op.getValueType(), MVT::Other},
                       {Ext.getValue(1), Ext.getValue(0)});
  }
  case ISD::WRITE_REGISTER: {
    assert(Op.getOperand(2).getValueType() == MVT::i128 &&
           "WRITE_REGISTER custom lowering is only for 128-bit sysregs");
    SDLoc DL(Op);

    SDValue Chain = Op.getOperand(0);
    SDValue SysRegName = Op.getOperand(1);
    std::pair<SDValue, SDValue> Pair =
        DAG.SplitScalar(Op.getOperand(2), DL, MVT::i64, MVT::i64);

    // chain = MSRR(chain, sysregname, lo, hi)
    SDValue Result = DAG.getNode(AArch64ISD::MSRR, DL, MVT::Other, Chain,
                                 SysRegName, Pair.first, Pair.second);

    return Result;
  }
  case ISD::FSHL:
  case ISD::FSHR:
    return LowerFunnelShift(Op, DAG);
  case ISD::FLDEXP:
    return LowerFLDEXP(Op, DAG);
  case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM:
    return LowerVECTOR_HISTOGRAM(Op, DAG);
  case ISD::PARTIAL_REDUCE_SMLA:
  case ISD::PARTIAL_REDUCE_UMLA:
  case ISD::PARTIAL_REDUCE_SUMLA:
    return LowerPARTIAL_REDUCE_MLA(Op, DAG);
  }
}

bool AArch64TargetLowering::mergeStoresAfterLegalization(EVT VT) const {
  return !Subtarget->useSVEForFixedLengthVectors();
}

bool AArch64TargetLowering::useSVEForFixedLengthVectorVT(
    EVT VT, bool OverrideNEON) const {
  if (!VT.isFixedLengthVector() || !VT.isSimple())
    return false;

  // Don't use SVE for vectors we cannot scalarize if required.
  switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
  // Fixed length predicates should be promoted to i8.
  // NOTE: This is consistent with how NEON (and thus 64/128bit vectors) work.
  case MVT::i1:
  default:
    return false;
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
  case MVT::i64:
  case MVT::f16:
  case MVT::f32:
  case MVT::f64:
    break;
  }

  // NEON-sized vectors can be emulated using SVE instructions.
  if (OverrideNEON && (VT.is128BitVector() || VT.is64BitVector()))
    return Subtarget->isSVEorStreamingSVEAvailable();

  // Ensure NEON MVTs only belong to a single register class.
  if (VT.getFixedSizeInBits() <= 128)
    return false;

  // Ensure wider than NEON code generation is enabled.
  if (!Subtarget->useSVEForFixedLengthVectors())
    return false;

  // Don't use SVE for types that don't fit.
  if (VT.getFixedSizeInBits() > Subtarget->getMinSVEVectorSizeInBits())
    return false;

  // TODO: Perhaps an artificial restriction, but worth having whilst getting
  // the base fixed length SVE support in place.
  if (!VT.isPow2VectorType())
    return false;

  return true;
}

//===----------------------------------------------------------------------===//
//                      Calling Convention Implementation
//===----------------------------------------------------------------------===//

static unsigned getIntrinsicID(const SDNode *N) {
  unsigned Opcode = N->getOpcode();
  switch (Opcode) {
  default:
    return Intrinsic::not_intrinsic;
  case ISD::INTRINSIC_WO_CHAIN: {
    unsigned IID = N->getConstantOperandVal(0);
    if (IID < Intrinsic::num_intrinsics)
      return IID;
    return Intrinsic::not_intrinsic;
  }
  }
}

bool AArch64TargetLowering::isReassocProfitable(SelectionDAG &DAG, SDValue N0,
                                                SDValue N1) const {
  if (!N0.hasOneUse())
    return false;

  unsigned IID = getIntrinsicID(N1.getNode());
  // Avoid reassociating expressions that can be lowered to smlal/umlal.
  if (IID == Intrinsic::aarch64_neon_umull ||
      N1.getOpcode() == AArch64ISD::UMULL ||
      IID == Intrinsic::aarch64_neon_smull ||
      N1.getOpcode() == AArch64ISD::SMULL)
    return N0.getOpcode() != ISD::ADD;

  return true;
}

/// Selects the correct CCAssignFn for a given CallingConvention value.
CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
                                                     bool IsVarArg) const {
  switch (CC) {
  default:
    reportFatalUsageError("unsupported calling convention");
  case CallingConv::GHC:
    return CC_AArch64_GHC;
  case CallingConv::PreserveNone:
    // The VarArg implementation makes assumptions about register
    // argument passing that do not hold for preserve_none, so we
    // instead fall back to C argument passing.
    // The non-vararg case is handled in the CC function itself.
    if (!IsVarArg)
      return CC_AArch64_Preserve_None;
    [[fallthrough]];
  case CallingConv::C:
  case CallingConv::Fast:
  case CallingConv::PreserveMost:
  case CallingConv::PreserveAll:
  case CallingConv::CXX_FAST_TLS:
  case CallingConv::Swift:
  case CallingConv::SwiftTail:
  case CallingConv::Tail:
  case CallingConv::GRAAL:
    if (Subtarget->isTargetWindows()) {
      if (IsVarArg) {
        if (Subtarget->isWindowsArm64EC())
          return CC_AArch64_Arm64EC_VarArg;
        return CC_AArch64_Win64_VarArg;
      }
      return CC_AArch64_Win64PCS;
    }
    if (!Subtarget->isTargetDarwin())
      return CC_AArch64_AAPCS;
    if (!IsVarArg)
      return CC_AArch64_DarwinPCS;
    return Subtarget->isTargetILP32() ? CC_AArch64_DarwinPCS_ILP32_VarArg
                                      : CC_AArch64_DarwinPCS_VarArg;
  case CallingConv::Win64:
    if (IsVarArg) {
      if (Subtarget->isWindowsArm64EC())
        return CC_AArch64_Arm64EC_VarArg;
      return CC_AArch64_Win64_VarArg;
    }
    return CC_AArch64_Win64PCS;
  case CallingConv::CFGuard_Check:
    if (Subtarget->isWindowsArm64EC())
      return CC_AArch64_Arm64EC_CFGuard_Check;
    return CC_AArch64_Win64_CFGuard_Check;
  case CallingConv::AArch64_VectorCall:
  case CallingConv::AArch64_SVE_VectorCall:
  case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X0:
  case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1:
  case CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X2:
    return CC_AArch64_AAPCS;
  case CallingConv::ARM64EC_Thunk_X64:
    return CC_AArch64_Arm64EC_Thunk;
  case CallingConv::ARM64EC_Thunk_Native:
    return CC_AArch64_Arm64EC_Thunk_Native;
  }
}

CCAssignFn *
AArch64TargetLowering::CCAssignFnForReturn(CallingConv::ID CC) const {
  switch (CC) {
  default:
    return RetCC_AArch64_AAPCS;
  case CallingConv::ARM64EC_Thunk_X64:
    return RetCC_AArch64_Arm64EC_Thunk;
  case CallingConv::CFGuard_Check:
    if (Subtarget->isWindowsArm64EC())
      return RetCC_AArch64_Arm64EC_CFGuard_Check;
    return RetCC_AArch64_AAPCS;
  }
}

static bool isPassedInFPR(EVT VT) {
  return VT.isFixedLengthVector() ||
         (VT.isFloatingPoint() && !VT.isScalableVector());
}

SDValue AArch64TargetLowering::LowerFormalArguments(
    SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
    const SmallVectorImpl<ISD::InputArg> &Ins, const SDLoc &DL,
    SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals) const {
  MachineFunction &MF = DAG.getMachineFunction();
  const Function &F = MF.getFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  bool IsWin64 =
      Subtarget->isCallingConvWin64(F.getCallingConv(), F.isVarArg());
  bool StackViaX4 = CallConv == CallingConv::ARM64EC_Thunk_X64 ||
                    (isVarArg && Subtarget->isWindowsArm64EC());
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();

  SmallVector<ISD::OutputArg, 4> Outs;
  GetReturnInfo(CallConv, F.getReturnType(), F.getAttributes(), Outs,
                DAG.getTargetLoweringInfo(), MF.getDataLayout());
  if (any_of(Outs, [](ISD::OutputArg &Out){ return Out.VT.isScalableVector(); }))
    FuncInfo->setIsSVECC(true);

  // Assign locations to all of the incoming arguments.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());

  // At this point, Ins[].VT may already be promoted to i32. To correctly
  // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
  // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
  // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
  // we use a special version of AnalyzeFormalArguments to pass in ValVT and
  // LocVT.
  unsigned NumArgs = Ins.size();
  Function::const_arg_iterator CurOrigArg = F.arg_begin();
  unsigned CurArgIdx = 0;
  bool UseVarArgCC = false;
  if (IsWin64)
    UseVarArgCC = isVarArg;

  CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, UseVarArgCC);

  for (unsigned i = 0; i != NumArgs; ++i) {
    MVT ValVT = Ins[i].VT;
    if (Ins[i].isOrigArg()) {
      std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
      CurArgIdx = Ins[i].getOrigArgIndex();

      // Get type of the original argument.
      EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
                                  /*AllowUnknown*/ true);
      MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
      // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
      if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
        ValVT = MVT::i8;
      else if (ActualMVT == MVT::i16)
        ValVT = MVT::i16;
    }
    bool Res =
        AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
    assert(!Res && "Call operand has unhandled type");
    (void)Res;
  }

  SMEAttrs Attrs = FuncInfo->getSMEFnAttrs();
  bool IsLocallyStreaming =
      !Attrs.hasStreamingInterface() && Attrs.hasStreamingBody();
  assert(Chain.getOpcode() == ISD::EntryToken && "Unexpected Chain value");
  SDValue Glue = Chain.getValue(1);

  unsigned ExtraArgLocs = 0;
  for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
    CCValAssign &VA = ArgLocs[i - ExtraArgLocs];

    if (Ins[i].Flags.isByVal()) {
      // Byval is used for HFAs in the PCS, but the system should work in a
      // non-compliant manner for larger structs.
      EVT PtrVT = getPointerTy(DAG.getDataLayout());
      int Size = Ins[i].Flags.getByValSize();
      unsigned NumRegs = (Size + 7) / 8;

      // FIXME: This works on big-endian for composite byvals, which are the common
      // case. It should also work for fundamental types too.
      unsigned FrameIdx =
        MFI.CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
      SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
      InVals.push_back(FrameIdxN);

      continue;
    }

    if (Ins[i].Flags.isSwiftAsync())
      MF.getInfo<AArch64FunctionInfo>()->setHasSwiftAsyncContext(true);

    SDValue ArgValue;
    if (VA.isRegLoc()) {
      // Arguments stored in registers.
      EVT RegVT = VA.getLocVT();
      const TargetRegisterClass *RC;

      if (RegVT == MVT::i32)
        RC = &AArch64::GPR32RegClass;
      else if (RegVT == MVT::i64)
        RC = &AArch64::GPR64RegClass;
      else if (RegVT == MVT::f16 || RegVT == MVT::bf16)
        RC = &AArch64::FPR16RegClass;
      else if (RegVT == MVT::f32)
        RC = &AArch64::FPR32RegClass;
      else if (RegVT == MVT::f64 || RegVT.is64BitVector())
        RC = &AArch64::FPR64RegClass;
      else if (RegVT == MVT::f128 || RegVT.is128BitVector())
        RC = &AArch64::FPR128RegClass;
      else if (RegVT.isScalableVector() &&
               RegVT.getVectorElementType() == MVT::i1) {
        FuncInfo->setIsSVECC(true);
        RC = &AArch64::PPRRegClass;
      } else if (RegVT == MVT::aarch64svcount) {
        FuncInfo->setIsSVECC(true);
        RC = &AArch64::PPRRegClass;
      } else if (RegVT.isScalableVector()) {
        FuncInfo->setIsSVECC(true);
        RC = &AArch64::ZPRRegClass;
      } else
        llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");

      // Transform the arguments in physical registers into virtual ones.
      Register Reg = MF.addLiveIn(VA.getLocReg(), RC);

      if (IsLocallyStreaming) {
        // LocallyStreamingFunctions must insert the SMSTART in the correct
        // position, so we use Glue to ensure no instructions can be scheduled
        // between the chain of:
        //        t0: ch,glue = EntryNode
        //      t1:  res,ch,glue = CopyFromReg
        //     ...
        //   tn: res,ch,glue = CopyFromReg t(n-1), ..
        // t(n+1): ch, glue = SMSTART t0:0, ...., tn:2
        // ^^^^^^
        // This will be the new Chain/Root node.
        ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT, Glue);
        Glue = ArgValue.getValue(2);
        if (isPassedInFPR(ArgValue.getValueType())) {
          ArgValue =
              DAG.getNode(AArch64ISD::COALESCER_BARRIER, DL,
                          DAG.getVTList(ArgValue.getValueType(), MVT::Glue),
                          {ArgValue, Glue});
          Glue = ArgValue.getValue(1);
        }
      } else
        ArgValue = DAG.getCopyFromReg(Chain, DL, Reg, RegVT);

      // If this is an 8, 16 or 32-bit value, it is really passed promoted
      // to 64 bits.  Insert an assert[sz]ext to capture this, then
      // truncate to the right size.
      switch (VA.getLocInfo()) {
      default:
        llvm_unreachable("Unknown loc info!");
      case CCValAssign::Full:
        break;
      case CCValAssign::Indirect:
        assert(
            (VA.getValVT().isScalableVT() || Subtarget->isWindowsArm64EC()) &&
            "Indirect arguments should be scalable on most subtargets");
        break;
      case CCValAssign::BCvt:
        ArgValue = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), ArgValue);
        break;
      case CCValAssign::AExt:
      case CCValAssign::SExt:
      case CCValAssign::ZExt:
        break;
      case CCValAssign::AExtUpper:
        ArgValue = DAG.getNode(ISD::SRL, DL, RegVT, ArgValue,
                               DAG.getConstant(32, DL, RegVT));
        ArgValue = DAG.getZExtOrTrunc(ArgValue, DL, VA.getValVT());
        break;
      }
    } else { // VA.isRegLoc()
      assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
      unsigned ArgOffset = VA.getLocMemOffset();
      unsigned ArgSize = (VA.getLocInfo() == CCValAssign::Indirect
                              ? VA.getLocVT().getSizeInBits()
                              : VA.getValVT().getSizeInBits()) / 8;

      uint32_t BEAlign = 0;
      if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
          !Ins[i].Flags.isInConsecutiveRegs())
        BEAlign = 8 - ArgSize;

      SDValue FIN;
      MachinePointerInfo PtrInfo;
      if (StackViaX4) {
        // In both the ARM64EC varargs convention and the thunk convention,
        // arguments on the stack are accessed relative to x4, not sp. In
        // the thunk convention, there's an additional offset of 32 bytes
        // to account for the shadow store.
        unsigned ObjOffset = ArgOffset + BEAlign;
        if (CallConv == CallingConv::ARM64EC_Thunk_X64)
          ObjOffset += 32;
        Register VReg = MF.addLiveIn(AArch64::X4, &AArch64::GPR64RegClass);
        SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
        FIN = DAG.getNode(ISD::ADD, DL, MVT::i64, Val,
                          DAG.getConstant(ObjOffset, DL, MVT::i64));
        PtrInfo = MachinePointerInfo::getUnknownStack(MF);
      } else {
        int FI = MFI.CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);

        // Create load nodes to retrieve arguments from the stack.
        FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
        PtrInfo = MachinePointerInfo::getFixedStack(MF, FI);
      }

      // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
      ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
      MVT MemVT = VA.getValVT();

      switch (VA.getLocInfo()) {
      default:
        break;
      case CCValAssign::Trunc:
      case CCValAssign::BCvt:
        MemVT = VA.getLocVT();
        break;
      case CCValAssign::Indirect:
        assert((VA.getValVT().isScalableVector() ||
                Subtarget->isWindowsArm64EC()) &&
               "Indirect arguments should be scalable on most subtargets");
        MemVT = VA.getLocVT();
        break;
      case CCValAssign::SExt:
        ExtType = ISD::SEXTLOAD;
        break;
      case CCValAssign::ZExt:
        ExtType = ISD::ZEXTLOAD;
        break;
      case CCValAssign::AExt:
        ExtType = ISD::EXTLOAD;
        break;
      }

      ArgValue = DAG.getExtLoad(ExtType, DL, VA.getLocVT(), Chain, FIN, PtrInfo,
                                MemVT);
    }

    if (VA.getLocInfo() == CCValAssign::Indirect) {
      assert((VA.getValVT().isScalableVT() ||
              Subtarget->isWindowsArm64EC()) &&
             "Indirect arguments should be scalable on most subtargets");

      uint64_t PartSize = VA.getValVT().getStoreSize().getKnownMinValue();
      unsigned NumParts = 1;
      if (Ins[i].Flags.isInConsecutiveRegs()) {
        while (!Ins[i + NumParts - 1].Flags.isInConsecutiveRegsLast())
          ++NumParts;
      }

      MVT PartLoad = VA.getValVT();
      SDValue Ptr = ArgValue;

      // Ensure we generate all loads for each tuple part, whilst updating the
      // pointer after each load correctly using vscale.
      while (NumParts > 0) {
        ArgValue = DAG.getLoad(PartLoad, DL, Chain, Ptr, MachinePointerInfo());
        InVals.push_back(ArgValue);
        NumParts--;
        if (NumParts > 0) {
          SDValue BytesIncrement;
          if (PartLoad.isScalableVector()) {
            BytesIncrement = DAG.getVScale(
                DL, Ptr.getValueType(),
                APInt(Ptr.getValueSizeInBits().getFixedValue(), PartSize));
          } else {
            BytesIncrement = DAG.getConstant(
                APInt(Ptr.getValueSizeInBits().getFixedValue(), PartSize), DL,
                Ptr.getValueType());
          }
          Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
                            BytesIncrement, SDNodeFlags::NoUnsignedWrap);
          ExtraArgLocs++;
          i++;
        }
      }
    } else {
      if (Subtarget->isTargetILP32() && Ins[i].Flags.isPointer())
        ArgValue = DAG.getNode(ISD::AssertZext, DL, ArgValue.getValueType(),
                               ArgValue, DAG.getValueType(MVT::i32));

      // i1 arguments are zero-extended to i8 by the caller. Emit a
      // hint to reflect this.
      if (Ins[i].isOrigArg()) {
        Argument *OrigArg = F.getArg(Ins[i].getOrigArgIndex());
        if (OrigArg->getType()->isIntegerTy(1)) {
          if (!Ins[i].Flags.isZExt()) {
            ArgValue = DAG.getNode(AArch64ISD::ASSERT_ZEXT_BOOL, DL,
                                   ArgValue.getValueType(), ArgValue);
          }
        }
      }

      InVals.push_back(ArgValue);
    }
  }
  assert((ArgLocs.size() + ExtraArgLocs) == Ins.size());

  // Insert the SMSTART if this is a locally streaming function and
  // make sure it is Glued to the last CopyFromReg value.
  if (IsLocallyStreaming) {
    SDValue PStateSM;
    if (Attrs.hasStreamingCompatibleInterface()) {
      PStateSM = getRuntimePStateSM(DAG, Chain, DL, MVT::i64);
      Register Reg = MF.getRegInfo().createVirtualRegister(
          getRegClassFor(PStateSM.getValueType().getSimpleVT()));
      FuncInfo->setPStateSMReg(Reg);
      Chain = DAG.getCopyToReg(Chain, DL, Reg, PStateSM);
      Chain = changeStreamingMode(DAG, DL, /*Enable*/ true, Chain, Glue,
                                  AArch64SME::IfCallerIsNonStreaming, PStateSM);
    } else
      Chain = changeStreamingMode(DAG, DL, /*Enable*/ true, Chain, Glue,
                                  AArch64SME::Always);

    // Ensure that the SMSTART happens after the CopyWithChain such that its
    // chain result is used.
    for (unsigned I=0; I<InVals.size(); ++I) {
      Register Reg = MF.getRegInfo().createVirtualRegister(
          getRegClassFor(InVals[I].getValueType().getSimpleVT()));
      Chain = DAG.getCopyToReg(Chain, DL, Reg, InVals[I]);
      InVals[I] = DAG.getCopyFromReg(Chain, DL, Reg,
                                     InVals[I].getValueType());
    }
  }

  // varargs
  if (isVarArg) {
    if (DAG.getMachineFunction().getFrameInfo().hasVAStart()) {
      if (!Subtarget->isTargetDarwin() || IsWin64) {
        // The AAPCS variadic function ABI is identical to the non-variadic
        // one. As a result there may be more arguments in registers and we
        // should save them for future reference.
        // Win64 variadic functions also pass arguments in registers, but all
        // float arguments are passed in integer registers.
        saveVarArgRegisters(CCInfo, DAG, DL, Chain);
      }

      // This will point to the next argument passed via stack.
      unsigned VarArgsOffset = CCInfo.getStackSize();
      // We currently pass all varargs at 8-byte alignment, or 4 for ILP32
      VarArgsOffset =
          alignTo(VarArgsOffset, Subtarget->isTargetILP32() ? 4 : 8);
      FuncInfo->setVarArgsStackOffset(VarArgsOffset);
      FuncInfo->setVarArgsStackIndex(
          MFI.CreateFixedObject(4, VarArgsOffset, true));
    }

    if (MFI.hasMustTailInVarArgFunc()) {
      SmallVector<MVT, 2> RegParmTypes;
      RegParmTypes.push_back(MVT::i64);
      RegParmTypes.push_back(MVT::f128);
      // Compute the set of forwarded registers. The rest are scratch.
      SmallVectorImpl<ForwardedRegister> &Forwards =
                                       FuncInfo->getForwardedMustTailRegParms();
      CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes,
                                               CC_AArch64_AAPCS);

      // Conservatively forward X8, since it might be used for aggregate return.
      if (!CCInfo.isAllocated(AArch64::X8)) {
        Register X8VReg = MF.addLiveIn(AArch64::X8, &AArch64::GPR64RegClass);
        Forwards.push_back(ForwardedRegister(X8VReg, AArch64::X8, MVT::i64));
      }
    }
  }

  // On Windows, InReg pointers must be returned, so record the pointer in a
  // virtual register at the start of the function so it can be returned in the
  // epilogue.
  if (IsWin64 || F.getCallingConv() == CallingConv::ARM64EC_Thunk_X64) {
    for (unsigned I = 0, E = Ins.size(); I != E; ++I) {
      if ((F.getCallingConv() == CallingConv::ARM64EC_Thunk_X64 ||
           Ins[I].Flags.isInReg()) &&
          Ins[I].Flags.isSRet()) {
        assert(!FuncInfo->getSRetReturnReg());

        MVT PtrTy = getPointerTy(DAG.getDataLayout());
        Register Reg =
            MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
        FuncInfo->setSRetReturnReg(Reg);

        SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, Reg, InVals[I]);
        Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Copy, Chain);
        break;
      }
    }
  }

  unsigned StackArgSize = CCInfo.getStackSize();
  bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
  if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
    // This is a non-standard ABI so by fiat I say we're allowed to make full
    // use of the stack area to be popped, which must be aligned to 16 bytes in
    // any case:
    StackArgSize = alignTo(StackArgSize, 16);

    // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
    // a multiple of 16.
    FuncInfo->setArgumentStackToRestore(StackArgSize);

    // This realignment carries over to the available bytes below. Our own
    // callers will guarantee the space is free by giving an aligned value to
    // CALLSEQ_START.
  }
  // Even if we're not expected to free up the space, it's useful to know how
  // much is there while considering tail calls (because we can reuse it).
  FuncInfo->setBytesInStackArgArea(StackArgSize);

  if (Subtarget->hasCustomCallingConv())
    Subtarget->getRegisterInfo()->UpdateCustomCalleeSavedRegs(MF);

  // Create a 16 Byte TPIDR2 object. The dynamic buffer
  // will be expanded and stored in the static object later using a pseudonode.
  if (Attrs.hasZAState()) {
    TPIDR2Object &TPIDR2 = FuncInfo->getTPIDR2Obj();
    TPIDR2.FrameIndex = MFI.CreateStackObject(16, Align(16), false);
    SDValue SVL = DAG.getNode(AArch64ISD::RDSVL, DL, MVT::i64,
                              DAG.getConstant(1, DL, MVT::i32));

    SDValue Buffer;
    if (!Subtarget->isTargetWindows() && !hasInlineStackProbe(MF)) {
      Buffer = DAG.getNode(AArch64ISD::ALLOCATE_ZA_BUFFER, DL,
                           DAG.getVTList(MVT::i64, MVT::Other), {Chain, SVL});
    } else {
      SDValue Size = DAG.getNode(ISD::MUL, DL, MVT::i64, SVL, SVL);
      Buffer = DAG.getNode(ISD::DYNAMIC_STACKALLOC, DL,
                           DAG.getVTList(MVT::i64, MVT::Other),
                           {Chain, Size, DAG.getConstant(1, DL, MVT::i64)});
      MFI.CreateVariableSizedObject(Align(16), nullptr);
    }
    Chain = DAG.getNode(
        AArch64ISD::INIT_TPIDR2OBJ, DL, DAG.getVTList(MVT::Other),
        {/*Chain*/ Buffer.getValue(1), /*Buffer ptr*/ Buffer.getValue(0)});
  } else if (Attrs.hasAgnosticZAInterface()) {
    // Call __arm_sme_state_size().
    SDValue BufferSize =
        DAG.getNode(AArch64ISD::GET_SME_SAVE_SIZE, DL,
                    DAG.getVTList(MVT::i64, MVT::Other), Chain);
    Chain = BufferSize.getValue(1);

    SDValue Buffer;
    if (!Subtarget->isTargetWindows() && !hasInlineStackProbe(MF)) {
      Buffer =
          DAG.getNode(AArch64ISD::ALLOC_SME_SAVE_BUFFER, DL,
                      DAG.getVTList(MVT::i64, MVT::Other), {Chain, BufferSize});
    } else {
      // Allocate space dynamically.
      Buffer = DAG.getNode(
          ISD::DYNAMIC_STACKALLOC, DL, DAG.getVTList(MVT::i64, MVT::Other),
          {Chain, BufferSize, DAG.getConstant(1, DL, MVT::i64)});
      MFI.CreateVariableSizedObject(Align(16), nullptr);
    }

    // Copy the value to a virtual register, and save that in FuncInfo.
    Register BufferPtr =
        MF.getRegInfo().createVirtualRegister(&AArch64::GPR64RegClass);
    FuncInfo->setSMESaveBufferAddr(BufferPtr);
    Chain = DAG.getCopyToReg(Chain, DL, BufferPtr, Buffer);
  }

  if (CallConv == CallingConv::PreserveNone) {
    for (const ISD::InputArg &I : Ins) {
      if (I.Flags.isSwiftSelf() || I.Flags.isSwiftError() ||
          I.Flags.isSwiftAsync()) {
        MachineFunction &MF = DAG.getMachineFunction();
        DAG.getContext()->diagnose(DiagnosticInfoUnsupported(
            MF.getFunction(),
            "Swift attributes can't be used with preserve_none",
            DL.getDebugLoc()));
        break;
      }
    }
  }

  return Chain;
}

void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
                                                SelectionDAG &DAG,
                                                const SDLoc &DL,
                                                SDValue &Chain) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  auto PtrVT = getPointerTy(DAG.getDataLayout());
  Function &F = MF.getFunction();
  bool IsWin64 =
      Subtarget->isCallingConvWin64(F.getCallingConv(), F.isVarArg());

  SmallVector<SDValue, 8> MemOps;

  auto GPRArgRegs = AArch64::getGPRArgRegs();
  unsigned NumGPRArgRegs = GPRArgRegs.size();
  if (Subtarget->isWindowsArm64EC()) {
    // In the ARM64EC ABI, only x0-x3 are used to pass arguments to varargs
    // functions.
    NumGPRArgRegs = 4;
  }
  unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);

  unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
  int GPRIdx = 0;
  if (GPRSaveSize != 0) {
    if (IsWin64) {
      GPRIdx = MFI.CreateFixedObject(GPRSaveSize, -(int)GPRSaveSize, false);
      if (GPRSaveSize & 15)
        // The extra size here, if triggered, will always be 8.
        MFI.CreateFixedObject(16 - (GPRSaveSize & 15), -(int)alignTo(GPRSaveSize, 16), false);
    } else
      GPRIdx = MFI.CreateStackObject(GPRSaveSize, Align(8), false);

    SDValue FIN;
    if (Subtarget->isWindowsArm64EC()) {
      // With the Arm64EC ABI, we reserve the save area as usual, but we
      // compute its address relative to x4.  For a normal AArch64->AArch64
      // call, x4 == sp on entry, but calls from an entry thunk can pass in a
      // different address.
      Register VReg = MF.addLiveIn(AArch64::X4, &AArch64::GPR64RegClass);
      SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
      FIN = DAG.getNode(ISD::SUB, DL, MVT::i64, Val,
                        DAG.getConstant(GPRSaveSize, DL, MVT::i64));
    } else {
      FIN = DAG.getFrameIndex(GPRIdx, PtrVT);
    }

    for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
      Register VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
      SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
      SDValue Store =
          DAG.getStore(Val.getValue(1), DL, Val, FIN,
                       IsWin64 ? MachinePointerInfo::getFixedStack(
                                     MF, GPRIdx, (i - FirstVariadicGPR) * 8)
                               : MachinePointerInfo::getStack(MF, i * 8));
      MemOps.push_back(Store);
      FIN =
          DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
    }
  }
  FuncInfo->setVarArgsGPRIndex(GPRIdx);
  FuncInfo->setVarArgsGPRSize(GPRSaveSize);

  if (Subtarget->hasFPARMv8() && !IsWin64) {
    auto FPRArgRegs = AArch64::getFPRArgRegs();
    const unsigned NumFPRArgRegs = FPRArgRegs.size();
    unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);

    unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
    int FPRIdx = 0;
    if (FPRSaveSize != 0) {
      FPRIdx = MFI.CreateStackObject(FPRSaveSize, Align(16), false);

      SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);

      for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
        Register VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
        SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);

        SDValue Store = DAG.getStore(Val.getValue(1), DL, Val, FIN,
                                     MachinePointerInfo::getStack(MF, i * 16));
        MemOps.push_back(Store);
        FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
                          DAG.getConstant(16, DL, PtrVT));
      }
    }
    FuncInfo->setVarArgsFPRIndex(FPRIdx);
    FuncInfo->setVarArgsFPRSize(FPRSaveSize);
  }

  if (!MemOps.empty()) {
    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
  }
}

/// LowerCallResult - Lower the result values of a call into the
/// appropriate copies out of appropriate physical registers.
SDValue AArch64TargetLowering::LowerCallResult(
    SDValue Chain, SDValue InGlue, CallingConv::ID CallConv, bool isVarArg,
    const SmallVectorImpl<CCValAssign> &RVLocs, const SDLoc &DL,
    SelectionDAG &DAG, SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
    SDValue ThisVal, bool RequiresSMChange) const {
  DenseMap<unsigned, SDValue> CopiedRegs;
  // Copy all of the result registers out of their specified physreg.
  for (unsigned i = 0; i != RVLocs.size(); ++i) {
    CCValAssign VA = RVLocs[i];

    // Pass 'this' value directly from the argument to return value, to avoid
    // reg unit interference
    if (i == 0 && isThisReturn) {
      assert(!VA.needsCustom() && VA.getLocVT() == MVT::i64 &&
             "unexpected return calling convention register assignment");
      InVals.push_back(ThisVal);
      continue;
    }

    // Avoid copying a physreg twice since RegAllocFast is incompetent and only
    // allows one use of a physreg per block.
    SDValue Val = CopiedRegs.lookup(VA.getLocReg());
    if (!Val) {
      Val =
          DAG.getCopyFromReg(Chain, DL, VA.getLocReg(), VA.getLocVT(), InGlue);
      Chain = Val.getValue(1);
      InGlue = Val.getValue(2);
      CopiedRegs[VA.getLocReg()] = Val;
    }

    switch (VA.getLocInfo()) {
    default:
      llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full:
      break;
    case CCValAssign::BCvt:
      Val = DAG.getNode(ISD::BITCAST, DL, VA.getValVT(), Val);
      break;
    case CCValAssign::AExtUpper:
      Val = DAG.getNode(ISD::SRL, DL, VA.getLocVT(), Val,
                        DAG.getConstant(32, DL, VA.getLocVT()));
      [[fallthrough]];
    case CCValAssign::AExt:
      [[fallthrough]];
    case CCValAssign::ZExt:
      Val = DAG.getZExtOrTrunc(Val, DL, VA.getValVT());
      break;
    }

    if (RequiresSMChange && isPassedInFPR(VA.getValVT()))
      Val = DAG.getNode(AArch64ISD::COALESCER_BARRIER, DL,
                        DAG.getVTList(Val.getValueType(), MVT::Glue), Val);

    InVals.push_back(Val);
  }

  return Chain;
}

/// Return true if the calling convention is one that we can guarantee TCO for.
static bool canGuaranteeTCO(CallingConv::ID CC, bool GuaranteeTailCalls) {
  return (CC == CallingConv::Fast && GuaranteeTailCalls) ||
         CC == CallingConv::Tail || CC == CallingConv::SwiftTail;
}

/// Return true if we might ever do TCO for calls with this calling convention.
static bool mayTailCallThisCC(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::C:
  case CallingConv::AArch64_SVE_VectorCall:
  case CallingConv::PreserveMost:
  case CallingConv::PreserveAll:
  case CallingConv::PreserveNone:
  case CallingConv::Swift:
  case CallingConv::SwiftTail:
  case CallingConv::Tail:
  case CallingConv::Fast:
    return true;
  default:
    return false;
  }
}

/// Return true if the call convention supports varargs
/// Currently only those that pass varargs like the C
/// calling convention does are eligible
/// Calling conventions listed in this function must also
/// be properly handled in AArch64Subtarget::isCallingConvWin64
static bool callConvSupportsVarArgs(CallingConv::ID CC) {
  switch (CC) {
  case CallingConv::C:
  case CallingConv::PreserveNone:
  // SVE vector call is only partially supported, but it should
  // support named arguments being passed. Any arguments being passed
  // as varargs, are still unsupported.
  case CallingConv::AArch64_SVE_VectorCall:
    return true;
  default:
    return false;
  }
}

static void analyzeCallOperands(const AArch64TargetLowering &TLI,
                                const AArch64Subtarget *Subtarget,
                                const TargetLowering::CallLoweringInfo &CLI,
                                CCState &CCInfo) {
  const SelectionDAG &DAG = CLI.DAG;
  CallingConv::ID CalleeCC = CLI.CallConv;
  bool IsVarArg = CLI.IsVarArg;
  const SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
  bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC, IsVarArg);

  // For Arm64EC thunks, allocate 32 extra bytes at the bottom of the stack
  // for the shadow store.
  if (CalleeCC == CallingConv::ARM64EC_Thunk_X64)
    CCInfo.AllocateStack(32, Align(16));

  unsigned NumArgs = Outs.size();
  for (unsigned i = 0; i != NumArgs; ++i) {
    MVT ArgVT = Outs[i].VT;
    ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;

    bool UseVarArgCC = false;
    if (IsVarArg) {
      // On Windows, the fixed arguments in a vararg call are passed in GPRs
      // too, so use the vararg CC to force them to integer registers.
      if (IsCalleeWin64) {
        UseVarArgCC = true;
      } else {
        UseVarArgCC = !Outs[i].IsFixed;
      }
    }

    if (!UseVarArgCC) {
      // Get type of the original argument.
      EVT ActualVT =
          TLI.getValueType(DAG.getDataLayout(), CLI.Args[Outs[i].OrigArgIndex].Ty,
                       /*AllowUnknown*/ true);
      MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ArgVT;
      // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
      if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
        ArgVT = MVT::i8;
      else if (ActualMVT == MVT::i16)
        ArgVT = MVT::i16;
    }

    // FIXME: CCAssignFnForCall should be called once, for the call and not per
    // argument. This logic should exactly mirror LowerFormalArguments.
    CCAssignFn *AssignFn = TLI.CCAssignFnForCall(CalleeCC, UseVarArgCC);
    bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
    assert(!Res && "Call operand has unhandled type");
    (void)Res;
  }
}

static SMECallAttrs
getSMECallAttrs(const Function &Caller,
                const TargetLowering::CallLoweringInfo &CLI) {
  if (CLI.CB)
    return SMECallAttrs(*CLI.CB);
  if (auto *ES = dyn_cast<ExternalSymbolSDNode>(CLI.Callee))
    return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(ES->getSymbol()));
  return SMECallAttrs(SMEAttrs(Caller), SMEAttrs(SMEAttrs::Normal));
}

bool AArch64TargetLowering::isEligibleForTailCallOptimization(
    const CallLoweringInfo &CLI) const {
  CallingConv::ID CalleeCC = CLI.CallConv;
  if (!mayTailCallThisCC(CalleeCC))
    return false;

  SDValue Callee = CLI.Callee;
  bool IsVarArg = CLI.IsVarArg;
  const SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
  const SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
  const SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
  const SelectionDAG &DAG = CLI.DAG;
  MachineFunction &MF = DAG.getMachineFunction();
  const Function &CallerF = MF.getFunction();
  CallingConv::ID CallerCC = CallerF.getCallingConv();

  // SME Streaming functions are not eligible for TCO as they may require
  // the streaming mode or ZA to be restored after returning from the call.
  SMECallAttrs CallAttrs = getSMECallAttrs(CallerF, CLI);
  if (CallAttrs.requiresSMChange() || CallAttrs.requiresLazySave() ||
      CallAttrs.requiresPreservingAllZAState() ||
      CallAttrs.caller().hasStreamingBody())
    return false;

  // Functions using the C or Fast calling convention that have an SVE signature
  // preserve more registers and should assume the SVE_VectorCall CC.
  // The check for matching callee-saved regs will determine whether it is
  // eligible for TCO.
  if ((CallerCC == CallingConv::C || CallerCC == CallingConv::Fast) &&
      MF.getInfo<AArch64FunctionInfo>()->isSVECC())
    CallerCC = CallingConv::AArch64_SVE_VectorCall;

  bool CCMatch = CallerCC == CalleeCC;

  // When using the Windows calling convention on a non-windows OS, we want
  // to back up and restore X18 in such functions; we can't do a tail call
  // from those functions.
  if (CallerCC == CallingConv::Win64 && !Subtarget->isTargetWindows() &&
      CalleeCC != CallingConv::Win64)
    return false;

  // Byval parameters hand the function a pointer directly into the stack area
  // we want to reuse during a tail call. Working around this *is* possible (see
  // X86) but less efficient and uglier in LowerCall.
  for (Function::const_arg_iterator i = CallerF.arg_begin(),
                                    e = CallerF.arg_end();
       i != e; ++i) {
    if (i->hasByValAttr())
      return false;

    // On Windows, "inreg" attributes signify non-aggregate indirect returns.
    // In this case, it is necessary to save X0/X1 in the callee and return it
    // in X0. Tail call opt may interfere with this, so we disable tail call
    // opt when the caller has an "inreg" attribute -- except if the callee
    // also has that attribute on the same argument, and the same value is
    // passed.
    if (i->hasInRegAttr()) {
      unsigned ArgIdx = i - CallerF.arg_begin();
      if (!CLI.CB || CLI.CB->arg_size() <= ArgIdx)
        return false;
      AttributeSet Attrs = CLI.CB->getParamAttributes(ArgIdx);
      if (!Attrs.hasAttribute(Attribute::InReg) ||
          !Attrs.hasAttribute(Attribute::StructRet) || !i->hasStructRetAttr() ||
          CLI.CB->getArgOperand(ArgIdx) != i) {
        return false;
      }
    }
  }

  if (canGuaranteeTCO(CalleeCC, getTargetMachine().Options.GuaranteedTailCallOpt))
    return CCMatch;

  // Externally-defined functions with weak linkage should not be
  // tail-called on AArch64 when the OS does not support dynamic
  // pre-emption of symbols, as the AAELF spec requires normal calls
  // to undefined weak functions to be replaced with a NOP or jump to the
  // next instruction. The behaviour of branch instructions in this
  // situation (as used for tail calls) is implementation-defined, so we
  // cannot rely on the linker replacing the tail call with a return.
  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
    const GlobalValue *GV = G->getGlobal();
    const Triple &TT = getTargetMachine().getTargetTriple();
    if (GV->hasExternalWeakLinkage() &&
        (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
      return false;
  }

  // Now we search for cases where we can use a tail call without changing the
  // ABI. Sibcall is used in some places (particularly gcc) to refer to this
  // concept.

  // I want anyone implementing a new calling convention to think long and hard
  // about this assert.
  if (IsVarArg && !callConvSupportsVarArgs(CalleeCC))
    report_fatal_error("Unsupported variadic calling convention");

  LLVMContext &C = *DAG.getContext();
  // Check that the call results are passed in the same way.
  if (!CCState::resultsCompatible(CalleeCC, CallerCC, MF, C, Ins,
                                  CCAssignFnForCall(CalleeCC, IsVarArg),
                                  CCAssignFnForCall(CallerCC, IsVarArg)))
    return false;
  // The callee has to preserve all registers the caller needs to preserve.
  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  const uint32_t *CallerPreserved = TRI->getCallPreservedMask(MF, CallerCC);
  if (!CCMatch) {
    const uint32_t *CalleePreserved = TRI->getCallPreservedMask(MF, CalleeCC);
    if (Subtarget->hasCustomCallingConv()) {
      TRI->UpdateCustomCallPreservedMask(MF, &CallerPreserved);
      TRI->UpdateCustomCallPreservedMask(MF, &CalleePreserved);
    }
    if (!TRI->regmaskSubsetEqual(CallerPreserved, CalleePreserved))
      return false;
  }

  // Nothing more to check if the callee is taking no arguments
  if (Outs.empty())
    return true;

  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CalleeCC, IsVarArg, MF, ArgLocs, C);

  analyzeCallOperands(*this, Subtarget, CLI, CCInfo);

  if (IsVarArg && !(CLI.CB && CLI.CB->isMustTailCall())) {
    // When we are musttail, additional checks have been done and we can safely ignore this check
    // At least two cases here: if caller is fastcc then we can't have any
    // memory arguments (we'd be expected to clean up the stack afterwards). If
    // caller is C then we could potentially use its argument area.

    // FIXME: for now we take the most conservative of these in both cases:
    // disallow all variadic memory operands.
    for (const CCValAssign &ArgLoc : ArgLocs)
      if (!ArgLoc.isRegLoc())
        return false;
  }

  const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();

  // If any of the arguments is passed indirectly, it must be SVE, so the
  // 'getBytesInStackArgArea' is not sufficient to determine whether we need to
  // allocate space on the stack. That is why we determine this explicitly here
  // the call cannot be a tailcall.
  if (llvm::any_of(ArgLocs, [&](CCValAssign &A) {
        assert((A.getLocInfo() != CCValAssign::Indirect ||
                A.getValVT().isScalableVector() ||
                Subtarget->isWindowsArm64EC()) &&
               "Expected value to be scalable");
        return A.getLocInfo() == CCValAssign::Indirect;
      }))
    return false;

  // If the stack arguments for this call do not fit into our own save area then
  // the call cannot be made tail.
  if (CCInfo.getStackSize() > FuncInfo->getBytesInStackArgArea())
    return false;

  const MachineRegisterInfo &MRI = MF.getRegInfo();
  if (!parametersInCSRMatch(MRI, CallerPreserved, ArgLocs, OutVals))
    return false;

  return true;
}

SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
                                                   SelectionDAG &DAG,
                                                   MachineFrameInfo &MFI,
                                                   int ClobberedFI) const {
  SmallVector<SDValue, 8> ArgChains;
  int64_t FirstByte = MFI.getObjectOffset(ClobberedFI);
  int64_t LastByte = FirstByte + MFI.getObjectSize(ClobberedFI) - 1;

  // Include the original chain at the beginning of the list. When this is
  // used by target LowerCall hooks, this helps legalize find the
  // CALLSEQ_BEGIN node.
  ArgChains.push_back(Chain);

  // Add a chain value for each stack argument corresponding
  for (SDNode *U : DAG.getEntryNode().getNode()->users())
    if (LoadSDNode *L = dyn_cast<LoadSDNode>(U))
      if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
        if (FI->getIndex() < 0) {
          int64_t InFirstByte = MFI.getObjectOffset(FI->getIndex());
          int64_t InLastByte = InFirstByte;
          InLastByte += MFI.getObjectSize(FI->getIndex()) - 1;

          if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
              (FirstByte <= InFirstByte && InFirstByte <= LastByte))
            ArgChains.push_back(SDValue(L, 1));
        }

  // Build a tokenfactor for all the chains.
  return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
}

bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
                                                   bool TailCallOpt) const {
  return (CallCC == CallingConv::Fast && TailCallOpt) ||
         CallCC == CallingConv::Tail || CallCC == CallingConv::SwiftTail;
}

// Check if the value is zero-extended from i1 to i8
static bool checkZExtBool(SDValue Arg, const SelectionDAG &DAG) {
  unsigned SizeInBits = Arg.getValueType().getSizeInBits();
  if (SizeInBits < 8)
    return false;

  APInt RequiredZero(SizeInBits, 0xFE);
  KnownBits Bits = DAG.computeKnownBits(Arg, 4);
  bool ZExtBool = (Bits.Zero & RequiredZero) == RequiredZero;
  return ZExtBool;
}

void AArch64TargetLowering::AdjustInstrPostInstrSelection(MachineInstr &MI,
                                                          SDNode *Node) const {
  // Live-in physreg copies that are glued to SMSTART are applied as
  // implicit-def's in the InstrEmitter. Here we remove them, allowing the
  // register allocator to pass call args in callee saved regs, without extra
  // copies to avoid these fake clobbers of actually-preserved GPRs.
  if (MI.getOpcode() == AArch64::MSRpstatesvcrImm1 ||
      MI.getOpcode() == AArch64::MSRpstatePseudo) {
    for (unsigned I = MI.getNumOperands() - 1; I > 0; --I)
      if (MachineOperand &MO = MI.getOperand(I);
          MO.isReg() && MO.isImplicit() && MO.isDef() &&
          (AArch64::GPR32RegClass.contains(MO.getReg()) ||
           AArch64::GPR64RegClass.contains(MO.getReg())))
        MI.removeOperand(I);

    // The SVE vector length can change when entering/leaving streaming mode.
    // FPMR is set to 0 when entering/leaving streaming mode.
    if (MI.getOperand(0).getImm() == AArch64SVCR::SVCRSM ||
        MI.getOperand(0).getImm() == AArch64SVCR::SVCRSMZA) {
      MI.addOperand(MachineOperand::CreateReg(AArch64::VG, /*IsDef=*/false,
                                              /*IsImplicit=*/true));
      MI.addOperand(MachineOperand::CreateReg(AArch64::VG, /*IsDef=*/true,
                                              /*IsImplicit=*/true));
      MI.addOperand(MachineOperand::CreateReg(AArch64::FPMR, /*IsDef=*/true,
                                              /*IsImplicit=*/true));
    }
  }

  // Add an implicit use of 'VG' for ADDXri/SUBXri, which are instructions that
  // have nothing to do with VG, were it not that they are used to materialise a
  // frame-address. If they contain a frame-index to a scalable vector, this
  // will likely require an ADDVL instruction to materialise the address, thus
  // reading VG.
  const MachineFunction &MF = *MI.getMF();
  if (MF.getInfo<AArch64FunctionInfo>()->hasStreamingModeChanges() &&
      (MI.getOpcode() == AArch64::ADDXri ||
       MI.getOpcode() == AArch64::SUBXri)) {
    const MachineOperand &MO = MI.getOperand(1);
    if (MO.isFI() && MF.getFrameInfo().getStackID(MO.getIndex()) ==
                         TargetStackID::ScalableVector)
      MI.addOperand(MachineOperand::CreateReg(AArch64::VG, /*IsDef=*/false,
                                              /*IsImplicit=*/true));
  }
}

SDValue AArch64TargetLowering::changeStreamingMode(SelectionDAG &DAG, SDLoc DL,
                                                   bool Enable, SDValue Chain,
                                                   SDValue InGlue,
                                                   unsigned Condition,
                                                   SDValue PStateSM) const {
  MachineFunction &MF = DAG.getMachineFunction();
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  FuncInfo->setHasStreamingModeChanges(true);

  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  SDValue RegMask = DAG.getRegisterMask(TRI->getSMStartStopCallPreservedMask());
  SDValue MSROp =
      DAG.getTargetConstant((int32_t)AArch64SVCR::SVCRSM, DL, MVT::i32);
  SmallVector<SDValue> Ops = {Chain, MSROp};
  unsigned Opcode;
  if (Condition != AArch64SME::Always) {
    SDValue ConditionOp = DAG.getTargetConstant(Condition, DL, MVT::i64);
    Opcode = Enable ? AArch64ISD::COND_SMSTART : AArch64ISD::COND_SMSTOP;
    assert(PStateSM && "PStateSM should be defined");
    Ops.push_back(ConditionOp);
    Ops.push_back(PStateSM);
  } else {
    Opcode = Enable ? AArch64ISD::SMSTART : AArch64ISD::SMSTOP;
  }
  Ops.push_back(RegMask);

  if (InGlue)
    Ops.push_back(InGlue);

  return DAG.getNode(Opcode, DL, DAG.getVTList(MVT::Other, MVT::Glue), Ops);
}

// Emit a call to __arm_sme_save or __arm_sme_restore.
static SDValue emitSMEStateSaveRestore(const AArch64TargetLowering &TLI,
                                       SelectionDAG &DAG,
                                       AArch64FunctionInfo *Info, SDLoc DL,
                                       SDValue Chain, bool IsSave) {
  MachineFunction &MF = DAG.getMachineFunction();
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  FuncInfo->setSMESaveBufferUsed();

  TargetLowering::ArgListTy Args;
  TargetLowering::ArgListEntry Entry;
  Entry.Ty = PointerType::getUnqual(*DAG.getContext());
  Entry.Node =
      DAG.getCopyFromReg(Chain, DL, Info->getSMESaveBufferAddr(), MVT::i64);
  Args.push_back(Entry);

  SDValue Callee =
      DAG.getExternalSymbol(IsSave ? "__arm_sme_save" : "__arm_sme_restore",
                            TLI.getPointerTy(DAG.getDataLayout()));
  auto *RetTy = Type::getVoidTy(*DAG.getContext());
  TargetLowering::CallLoweringInfo CLI(DAG);
  CLI.setDebugLoc(DL).setChain(Chain).setLibCallee(
      CallingConv::AArch64_SME_ABI_Support_Routines_PreserveMost_From_X1, RetTy,
      Callee, std::move(Args));
  return TLI.LowerCallTo(CLI).second;
}

static AArch64SME::ToggleCondition
getSMToggleCondition(const SMECallAttrs &CallAttrs) {
  if (!CallAttrs.caller().hasStreamingCompatibleInterface() ||
      CallAttrs.caller().hasStreamingBody())
    return AArch64SME::Always;
  if (CallAttrs.callee().hasNonStreamingInterface())
    return AArch64SME::IfCallerIsStreaming;
  if (CallAttrs.callee().hasStreamingInterface())
    return AArch64SME::IfCallerIsNonStreaming;

  llvm_unreachable("Unsupported attributes");
}

/// Check whether a stack argument requires lowering in a tail call.
static bool shouldLowerTailCallStackArg(const MachineFunction &MF,
                                        const CCValAssign &VA, SDValue Arg,
                                        ISD::ArgFlagsTy Flags, int CallOffset) {
  // FIXME: We should be able to handle this case, but it's not clear how to.
  if (Flags.isZExt() || Flags.isSExt())
    return true;

  for (;;) {
    // Look through nodes that don't alter the bits of the incoming value.
    unsigned Op = Arg.getOpcode();
    if (Op == ISD::ZERO_EXTEND || Op == ISD::ANY_EXTEND || Op == ISD::BITCAST ||
        Arg->isAssert() || Op == AArch64ISD::ASSERT_ZEXT_BOOL) {
      Arg = Arg.getOperand(0);
      continue;
    }
    break;
  }

  // If the argument is a load from the same immutable stack slot, we can reuse
  // it.
  if (auto *LoadNode = dyn_cast<LoadSDNode>(Arg)) {
    if (auto *FINode = dyn_cast<FrameIndexSDNode>(LoadNode->getBasePtr())) {
      const MachineFrameInfo &MFI = MF.getFrameInfo();
      int FI = FINode->getIndex();
      if (!MFI.isImmutableObjectIndex(FI))
        return true;
      if (CallOffset != MFI.getObjectOffset(FI))
        return true;
      uint64_t SizeInBits = LoadNode->getMemoryVT().getFixedSizeInBits();
      if (SizeInBits / 8 != MFI.getObjectSize(FI))
        return true;
      return false;
    }
  }

  return true;
}

/// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
/// and add input and output parameter nodes.
SDValue
AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
                                 SmallVectorImpl<SDValue> &InVals) const {
  SelectionDAG &DAG = CLI.DAG;
  SDLoc &DL = CLI.DL;
  SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
  SmallVector<SDValue, 32> &OutVals = CLI.OutVals;
  SmallVector<ISD::InputArg, 32> &Ins = CLI.Ins;
  SDValue Chain = CLI.Chain;
  SDValue Callee = CLI.Callee;
  bool &IsTailCall = CLI.IsTailCall;
  CallingConv::ID &CallConv = CLI.CallConv;
  bool IsVarArg = CLI.IsVarArg;

  MachineFunction &MF = DAG.getMachineFunction();
  MachineFunction::CallSiteInfo CSInfo;
  bool IsThisReturn = false;

  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
  bool IsCFICall = CLI.CB && CLI.CB->isIndirectCall() && CLI.CFIType;
  bool IsSibCall = false;
  bool GuardWithBTI = false;

  if (CLI.CB && CLI.CB->hasFnAttr(Attribute::ReturnsTwice) &&
      !Subtarget->noBTIAtReturnTwice()) {
    GuardWithBTI = FuncInfo->branchTargetEnforcement();
  }

  // Analyze operands of the call, assigning locations to each operand.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CallConv, IsVarArg, MF, ArgLocs, *DAG.getContext());

  if (IsVarArg) {
    unsigned NumArgs = Outs.size();

    for (unsigned i = 0; i != NumArgs; ++i) {
      if (!Outs[i].IsFixed && Outs[i].VT.isScalableVector())
        report_fatal_error("Passing SVE types to variadic functions is "
                           "currently not supported");
    }
  }

  analyzeCallOperands(*this, Subtarget, CLI, CCInfo);

  CCAssignFn *RetCC = CCAssignFnForReturn(CallConv);
  // Assign locations to each value returned by this call.
  SmallVector<CCValAssign, 16> RVLocs;
  CCState RetCCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), RVLocs,
                    *DAG.getContext());
  RetCCInfo.AnalyzeCallResult(Ins, RetCC);

  // Check callee args/returns for SVE registers and set calling convention
  // accordingly.
  if (CallConv == CallingConv::C || CallConv == CallingConv::Fast) {
    auto HasSVERegLoc = [](CCValAssign &Loc) {
      if (!Loc.isRegLoc())
        return false;
      return AArch64::ZPRRegClass.contains(Loc.getLocReg()) ||
             AArch64::PPRRegClass.contains(Loc.getLocReg());
    };
    if (any_of(RVLocs, HasSVERegLoc) || any_of(ArgLocs, HasSVERegLoc))
      CallConv = CallingConv::AArch64_SVE_VectorCall;
  }

  if (IsTailCall) {
    // Check if it's really possible to do a tail call.
    IsTailCall = isEligibleForTailCallOptimization(CLI);

    // A sibling call is one where we're under the usual C ABI and not planning
    // to change that but can still do a tail call:
    if (!TailCallOpt && IsTailCall && CallConv != CallingConv::Tail &&
        CallConv != CallingConv::SwiftTail)
      IsSibCall = true;

    if (IsTailCall)
      ++NumTailCalls;
  }

  if (!IsTailCall && CLI.CB && CLI.CB->isMustTailCall())
    report_fatal_error("failed to perform tail call elimination on a call "
                       "site marked musttail");

  // Get a count of how many bytes are to be pushed on the stack.
  unsigned NumBytes = CCInfo.getStackSize();

  if (IsSibCall) {
    // Since we're not changing the ABI to make this a tail call, the memory
    // operands are already available in the caller's incoming argument space.
    NumBytes = 0;
  }

  // FPDiff is the byte offset of the call's argument area from the callee's.
  // Stores to callee stack arguments will be placed in FixedStackSlots offset
  // by this amount for a tail call. In a sibling call it must be 0 because the
  // caller will deallocate the entire stack and the callee still expects its
  // arguments to begin at SP+0. Completely unused for non-tail calls.
  int FPDiff = 0;

  if (IsTailCall && !IsSibCall) {
    unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();

    // Since callee will pop argument stack as a tail call, we must keep the
    // popped size 16-byte aligned.
    NumBytes = alignTo(NumBytes, 16);

    // FPDiff will be negative if this tail call requires more space than we
    // would automatically have in our incoming argument space. Positive if we
    // can actually shrink the stack.
    FPDiff = NumReusableBytes - NumBytes;

    // Update the required reserved area if this is the tail call requiring the
    // most argument stack space.
    if (FPDiff < 0 && FuncInfo->getTailCallReservedStack() < (unsigned)-FPDiff)
      FuncInfo->setTailCallReservedStack(-FPDiff);

    // The stack pointer must be 16-byte aligned at all times it's used for a
    // memory operation, which in practice means at *all* times and in
    // particular across call boundaries. Therefore our own arguments started at
    // a 16-byte aligned SP and the delta applied for the tail call should
    // satisfy the same constraint.
    assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
  }

  // Determine whether we need any streaming mode changes.
  SMECallAttrs CallAttrs = getSMECallAttrs(MF.getFunction(), CLI);

  auto DescribeCallsite =
      [&](OptimizationRemarkAnalysis &R) -> OptimizationRemarkAnalysis & {
    R << "call from '" << ore::NV("Caller", MF.getName()) << "' to '";
    if (auto *ES = dyn_cast<ExternalSymbolSDNode>(CLI.Callee))
      R << ore::NV("Callee", ES->getSymbol());
    else if (CLI.CB && CLI.CB->getCalledFunction())
      R << ore::NV("Callee", CLI.CB->getCalledFunction()->getName());
    else
      R << "unknown callee";
    R << "'";
    return R;
  };

  bool RequiresLazySave = CallAttrs.requiresLazySave();
  bool RequiresSaveAllZA = CallAttrs.requiresPreservingAllZAState();
  if (RequiresLazySave) {
    const TPIDR2Object &TPIDR2 = FuncInfo->getTPIDR2Obj();
    MachinePointerInfo MPI =
        MachinePointerInfo::getStack(MF, TPIDR2.FrameIndex);
    SDValue TPIDR2ObjAddr = DAG.getFrameIndex(
        TPIDR2.FrameIndex,
        DAG.getTargetLoweringInfo().getFrameIndexTy(DAG.getDataLayout()));
    SDValue NumZaSaveSlicesAddr =
        DAG.getNode(ISD::ADD, DL, TPIDR2ObjAddr.getValueType(), TPIDR2ObjAddr,
                    DAG.getConstant(8, DL, TPIDR2ObjAddr.getValueType()));
    SDValue NumZaSaveSlices = DAG.getNode(AArch64ISD::RDSVL, DL, MVT::i64,
                                          DAG.getConstant(1, DL, MVT::i32));
    Chain = DAG.getTruncStore(Chain, DL, NumZaSaveSlices, NumZaSaveSlicesAddr,
                              MPI, MVT::i16);
    Chain = DAG.getNode(
        ISD::INTRINSIC_VOID, DL, MVT::Other, Chain,
        DAG.getConstant(Intrinsic::aarch64_sme_set_tpidr2, DL, MVT::i32),
        TPIDR2ObjAddr);
    OptimizationRemarkEmitter ORE(&MF.getFunction());
    ORE.emit([&]() {
      auto R = CLI.CB ? OptimizationRemarkAnalysis("sme", "SMELazySaveZA",
                                                   CLI.CB)
                      : OptimizationRemarkAnalysis("sme", "SMELazySaveZA",
                                                   &MF.getFunction());
      return DescribeCallsite(R) << " sets up a lazy save for ZA";
    });
  } else if (RequiresSaveAllZA) {
    assert(!CallAttrs.callee().hasSharedZAInterface() &&
           "Cannot share state that may not exist");
    Chain = emitSMEStateSaveRestore(*this, DAG, FuncInfo, DL, Chain,
                                    /*IsSave=*/true);
  }

  SDValue PStateSM;
  bool RequiresSMChange = CallAttrs.requiresSMChange();
  if (RequiresSMChange) {
    if (CallAttrs.caller().hasStreamingInterfaceOrBody())
      PStateSM = DAG.getConstant(1, DL, MVT::i64);
    else if (CallAttrs.caller().hasNonStreamingInterface())
      PStateSM = DAG.getConstant(0, DL, MVT::i64);
    else
      PStateSM = getRuntimePStateSM(DAG, Chain, DL, MVT::i64);
    OptimizationRemarkEmitter ORE(&MF.getFunction());
    ORE.emit([&]() {
      auto R = CLI.CB ? OptimizationRemarkAnalysis("sme", "SMETransition",
                                                   CLI.CB)
                      : OptimizationRemarkAnalysis("sme", "SMETransition",
                                                   &MF.getFunction());
      DescribeCallsite(R) << " requires a streaming mode transition";
      return R;
    });
  }

  SDValue ZTFrameIdx;
  MachineFrameInfo &MFI = MF.getFrameInfo();
  bool ShouldPreserveZT0 = CallAttrs.requiresPreservingZT0();

  // If the caller has ZT0 state which will not be preserved by the callee,
  // spill ZT0 before the call.
  if (ShouldPreserveZT0) {
    unsigned ZTObj = MFI.CreateSpillStackObject(64, Align(16));
    ZTFrameIdx = DAG.getFrameIndex(
        ZTObj,
        DAG.getTargetLoweringInfo().getFrameIndexTy(DAG.getDataLayout()));

    Chain = DAG.getNode(AArch64ISD::SAVE_ZT, DL, DAG.getVTList(MVT::Other),
                        {Chain, DAG.getConstant(0, DL, MVT::i32), ZTFrameIdx});
  }

  // If caller shares ZT0 but the callee is not shared ZA, we need to stop
  // PSTATE.ZA before the call if there is no lazy-save active.
  bool DisableZA = CallAttrs.requiresDisablingZABeforeCall();
  assert((!DisableZA || !RequiresLazySave) &&
         "Lazy-save should have PSTATE.SM=1 on entry to the function");

  if (DisableZA)
    Chain = DAG.getNode(
        AArch64ISD::SMSTOP, DL, DAG.getVTList(MVT::Other, MVT::Glue), Chain,
        DAG.getTargetConstant((int32_t)(AArch64SVCR::SVCRZA), DL, MVT::i32));

  // Adjust the stack pointer for the new arguments...
  // These operations are automatically eliminated by the prolog/epilog pass
  if (!IsSibCall)
    Chain = DAG.getCALLSEQ_START(Chain, IsTailCall ? 0 : NumBytes, 0, DL);

  SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
                                        getPointerTy(DAG.getDataLayout()));

  SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
  SmallSet<unsigned, 8> RegsUsed;
  SmallVector<SDValue, 8> MemOpChains;
  auto PtrVT = getPointerTy(DAG.getDataLayout());

  if (IsVarArg && CLI.CB && CLI.CB->isMustTailCall()) {
    const auto &Forwards = FuncInfo->getForwardedMustTailRegParms();
    for (const auto &F : Forwards) {
      SDValue Val = DAG.getCopyFromReg(Chain, DL, F.VReg, F.VT);
       RegsToPass.emplace_back(F.PReg, Val);
    }
  }

  // Walk the register/memloc assignments, inserting copies/loads.
  unsigned ExtraArgLocs = 0;
  for (unsigned i = 0, e = Outs.size(); i != e; ++i) {
    CCValAssign &VA = ArgLocs[i - ExtraArgLocs];
    SDValue Arg = OutVals[i];
    ISD::ArgFlagsTy Flags = Outs[i].Flags;

    // Promote the value if needed.
    switch (VA.getLocInfo()) {
    default:
      llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full:
      break;
    case CCValAssign::SExt:
      Arg = DAG.getNode(ISD::SIGN_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::ZExt:
      Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::AExt:
      if (Outs[i].ArgVT == MVT::i1) {
        // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
        //
        // Check if we actually have to do this, because the value may
        // already be zero-extended.
        //
        // We cannot just emit a (zext i8 (trunc (assert-zext i8)))
        // and rely on DAGCombiner to fold this, because the following
        // (anyext i32) is combined with (zext i8) in DAG.getNode:
        //
        //   (ext (zext x)) -> (zext x)
        //
        // This will give us (zext i32), which we cannot remove, so
        // try to check this beforehand.
        if (!checkZExtBool(Arg, DAG)) {
          Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
          Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
        }
      }
      Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::AExtUpper:
      assert(VA.getValVT() == MVT::i32 && "only expect 32 -> 64 upper bits");
      Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
      Arg = DAG.getNode(ISD::SHL, DL, VA.getLocVT(), Arg,
                        DAG.getConstant(32, DL, VA.getLocVT()));
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getBitcast(VA.getLocVT(), Arg);
      break;
    case CCValAssign::Trunc:
      Arg = DAG.getZExtOrTrunc(Arg, DL, VA.getLocVT());
      break;
    case CCValAssign::FPExt:
      Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::Indirect:
      bool isScalable = VA.getValVT().isScalableVT();
      assert((isScalable || Subtarget->isWindowsArm64EC()) &&
             "Indirect arguments should be scalable on most subtargets");

      uint64_t StoreSize = VA.getValVT().getStoreSize().getKnownMinValue();
      uint64_t PartSize = StoreSize;
      unsigned NumParts = 1;
      if (Outs[i].Flags.isInConsecutiveRegs()) {
        while (!Outs[i + NumParts - 1].Flags.isInConsecutiveRegsLast())
          ++NumParts;
        StoreSize *= NumParts;
      }

      Type *Ty = EVT(VA.getValVT()).getTypeForEVT(*DAG.getContext());
      Align Alignment = DAG.getDataLayout().getPrefTypeAlign(Ty);
      MachineFrameInfo &MFI = MF.getFrameInfo();
      int FI = MFI.CreateStackObject(StoreSize, Alignment, false);
      if (isScalable)
        MFI.setStackID(FI, TargetStackID::ScalableVector);

      MachinePointerInfo MPI = MachinePointerInfo::getFixedStack(MF, FI);
      SDValue Ptr = DAG.getFrameIndex(
          FI, DAG.getTargetLoweringInfo().getFrameIndexTy(DAG.getDataLayout()));
      SDValue SpillSlot = Ptr;

      // Ensure we generate all stores for each tuple part, whilst updating the
      // pointer after each store correctly using vscale.
      while (NumParts) {
        SDValue Store = DAG.getStore(Chain, DL, OutVals[i], Ptr, MPI);
        MemOpChains.push_back(Store);

        NumParts--;
        if (NumParts > 0) {
          SDValue BytesIncrement;
          if (isScalable) {
            BytesIncrement = DAG.getVScale(
                DL, Ptr.getValueType(),
                APInt(Ptr.getValueSizeInBits().getFixedValue(), PartSize));
          } else {
            BytesIncrement = DAG.getConstant(
                APInt(Ptr.getValueSizeInBits().getFixedValue(), PartSize), DL,
                Ptr.getValueType());
          }
          MPI = MachinePointerInfo(MPI.getAddrSpace());
          Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
                            BytesIncrement, SDNodeFlags::NoUnsignedWrap);
          ExtraArgLocs++;
          i++;
        }
      }

      Arg = SpillSlot;
      break;
    }

    if (VA.isRegLoc()) {
      if (i == 0 && Flags.isReturned() && !Flags.isSwiftSelf() &&
          Outs[0].VT == MVT::i64) {
        assert(VA.getLocVT() == MVT::i64 &&
               "unexpected calling convention register assignment");
        assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
               "unexpected use of 'returned'");
        IsThisReturn = true;
      }
      if (RegsUsed.count(VA.getLocReg())) {
        // If this register has already been used then we're trying to pack
        // parts of an [N x i32] into an X-register. The extension type will
        // take care of putting the two halves in the right place but we have to
        // combine them.
        SDValue &Bits =
            llvm::find_if(RegsToPass,
                          [=](const std::pair<unsigned, SDValue> &Elt) {
                            return Elt.first == VA.getLocReg();
                          })
                ->second;
        Bits = DAG.getNode(ISD::OR, DL, Bits.getValueType(), Bits, Arg);
        // Call site info is used for function's parameter entry value
        // tracking. For now we track only simple cases when parameter
        // is transferred through whole register.
        llvm::erase_if(CSInfo.ArgRegPairs,
                       [&VA](MachineFunction::ArgRegPair ArgReg) {
                         return ArgReg.Reg == VA.getLocReg();
                       });
      } else {
        // Add an extra level of indirection for streaming mode changes by
        // using a pseudo copy node that cannot be rematerialised between a
        // smstart/smstop and the call by the simple register coalescer.
        if (RequiresSMChange && isPassedInFPR(Arg.getValueType()))
          Arg = DAG.getNode(AArch64ISD::COALESCER_BARRIER, DL,
                            DAG.getVTList(Arg.getValueType(), MVT::Glue), Arg);
        RegsToPass.emplace_back(VA.getLocReg(), Arg);
        RegsUsed.insert(VA.getLocReg());
        const TargetOptions &Options = DAG.getTarget().Options;
        if (Options.EmitCallSiteInfo)
          CSInfo.ArgRegPairs.emplace_back(VA.getLocReg(), i);
      }
    } else {
      assert(VA.isMemLoc());

      SDValue DstAddr;
      MachinePointerInfo DstInfo;

      // FIXME: This works on big-endian for composite byvals, which are the
      // common case. It should also work for fundamental types too.
      uint32_t BEAlign = 0;
      unsigned OpSize;
      if (VA.getLocInfo() == CCValAssign::Indirect ||
          VA.getLocInfo() == CCValAssign::Trunc)
        OpSize = VA.getLocVT().getFixedSizeInBits();
      else
        OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
                                 : VA.getValVT().getSizeInBits();
      OpSize = (OpSize + 7) / 8;
      if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
          !Flags.isInConsecutiveRegs()) {
        if (OpSize < 8)
          BEAlign = 8 - OpSize;
      }
      unsigned LocMemOffset = VA.getLocMemOffset();
      int32_t Offset = LocMemOffset + BEAlign;

      if (IsTailCall) {
        // When the frame pointer is perfectly aligned for the tail call and the
        // same stack argument is passed down intact, we can reuse it.
        if (!FPDiff && !shouldLowerTailCallStackArg(MF, VA, Arg, Flags, Offset))
          continue;

        Offset = Offset + FPDiff;
        int FI = MF.getFrameInfo().CreateFixedObject(OpSize, Offset, true);

        DstAddr = DAG.getFrameIndex(FI, PtrVT);
        DstInfo = MachinePointerInfo::getFixedStack(MF, FI);

        // Make sure any stack arguments overlapping with where we're storing
        // are loaded before this eventual operation. Otherwise they'll be
        // clobbered.
        Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
      } else {
        SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);

        DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
        DstInfo = MachinePointerInfo::getStack(MF, LocMemOffset);
      }

      if (Outs[i].Flags.isByVal()) {
        SDValue SizeNode =
            DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
        SDValue Cpy = DAG.getMemcpy(
            Chain, DL, DstAddr, Arg, SizeNode,
            Outs[i].Flags.getNonZeroByValAlign(),
            /*isVol = */ false, /*AlwaysInline = */ false,
            /*CI=*/nullptr, std::nullopt, DstInfo, MachinePointerInfo());

        MemOpChains.push_back(Cpy);
      } else {
        // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
        // promoted to a legal register type i32, we should truncate Arg back to
        // i1/i8/i16.
        if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
            VA.getValVT() == MVT::i16)
          Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);

        SDValue Store = DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo);
        MemOpChains.push_back(Store);
      }
    }
  }

  if (IsVarArg && Subtarget->isWindowsArm64EC() &&
      !(CLI.CB && CLI.CB->isMustTailCall())) {
    SDValue ParamPtr = StackPtr;
    if (IsTailCall) {
      // Create a dummy object at the top of the stack that can be used to get
      // the SP after the epilogue
      int FI = MF.getFrameInfo().CreateFixedObject(1, FPDiff, true);
      ParamPtr = DAG.getFrameIndex(FI, PtrVT);
    }

    // For vararg calls, the Arm64EC ABI requires values in x4 and x5
    // describing the argument list.  x4 contains the address of the
    // first stack parameter. x5 contains the size in bytes of all parameters
    // passed on the stack.
    RegsToPass.emplace_back(AArch64::X4, ParamPtr);
    RegsToPass.emplace_back(AArch64::X5,
                            DAG.getConstant(NumBytes, DL, MVT::i64));
  }

  if (!MemOpChains.empty())
    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);

  SDValue InGlue;
  if (RequiresSMChange) {
    if (!Subtarget->isTargetDarwin() || Subtarget->hasSVE()) {
      Chain = DAG.getNode(AArch64ISD::VG_SAVE, DL,
                          DAG.getVTList(MVT::Other, MVT::Glue), Chain);
      InGlue = Chain.getValue(1);
    }

    SDValue NewChain = changeStreamingMode(
        DAG, DL, CallAttrs.callee().hasStreamingInterface(), Chain, InGlue,
        getSMToggleCondition(CallAttrs), PStateSM);
    Chain = NewChain.getValue(0);
    InGlue = NewChain.getValue(1);
  }

  // Build a sequence of copy-to-reg nodes chained together with token chain
  // and flag operands which copy the outgoing args into the appropriate regs.
  for (auto &RegToPass : RegsToPass) {
    Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
                             RegToPass.second, InGlue);
    InGlue = Chain.getValue(1);
  }

  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
  // node so that legalize doesn't hack it.
  const GlobalValue *CalledGlobal = nullptr;
  unsigned OpFlags = 0;
  if (auto *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
    CalledGlobal = G->getGlobal();
    OpFlags = Subtarget->classifyGlobalFunctionReference(CalledGlobal,
                                                         getTargetMachine());
    if (OpFlags & AArch64II::MO_GOT) {
      Callee = DAG.getTargetGlobalAddress(CalledGlobal, DL, PtrVT, 0, OpFlags);
      Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
    } else {
      const GlobalValue *GV = G->getGlobal();
      Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
    }
  } else if (auto *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
    bool UseGot = (getTargetMachine().getCodeModel() == CodeModel::Large &&
                   Subtarget->isTargetMachO()) ||
                  MF.getFunction().getParent()->getRtLibUseGOT();
    const char *Sym = S->getSymbol();
    if (UseGot) {
      Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
      Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
    } else {
      Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
    }
  }

  // We don't usually want to end the call-sequence here because we would tidy
  // the frame up *after* the call, however in the ABI-changing tail-call case
  // we've carefully laid out the parameters so that when sp is reset they'll be
  // in the correct location.
  if (IsTailCall && !IsSibCall) {
    Chain = DAG.getCALLSEQ_END(Chain, 0, 0, InGlue, DL);
    InGlue = Chain.getValue(1);
  }

  unsigned Opc = IsTailCall ? AArch64ISD::TC_RETURN : AArch64ISD::CALL;

  std::vector<SDValue> Ops;
  Ops.push_back(Chain);
  Ops.push_back(Callee);

  // Calls with operand bundle "clang.arc.attachedcall" are special. They should
  // be expanded to the call, directly followed by a special marker sequence and
  // a call to an ObjC library function.  Use CALL_RVMARKER to do that.
  if (CLI.CB && objcarc::hasAttachedCallOpBundle(CLI.CB)) {
    assert(!IsTailCall &&
           "tail calls cannot be marked with clang.arc.attachedcall");
    Opc = AArch64ISD::CALL_RVMARKER;

    // Add a target global address for the retainRV/claimRV runtime function
    // just before the call target.
    Function *ARCFn = *objcarc::getAttachedARCFunction(CLI.CB);
    auto GA = DAG.getTargetGlobalAddress(ARCFn, DL, PtrVT);
    Ops.insert(Ops.begin() + 1, GA);

    // We may or may not need to emit both the marker and the retain/claim call.
    // Tell the pseudo expansion using an additional boolean op.
    bool ShouldEmitMarker = objcarc::attachedCallOpBundleNeedsMarker(CLI.CB);
    SDValue DoEmitMarker =
        DAG.getTargetConstant(ShouldEmitMarker, DL, MVT::i32);
    Ops.insert(Ops.begin() + 2, DoEmitMarker);
  } else if (CallConv == CallingConv::ARM64EC_Thunk_X64) {
    Opc = AArch64ISD::CALL_ARM64EC_TO_X64;
  } else if (GuardWithBTI) {
    Opc = AArch64ISD::CALL_BTI;
  }

  if (IsTailCall) {
    // Each tail call may have to adjust the stack by a different amount, so
    // this information must travel along with the operation for eventual
    // consumption by emitEpilogue.
    Ops.push_back(DAG.getSignedTargetConstant(FPDiff, DL, MVT::i32));
  }

  if (CLI.PAI) {
    const uint64_t Key = CLI.PAI->Key;
    assert((Key == AArch64PACKey::IA || Key == AArch64PACKey::IB) &&
           "Invalid auth call key");

    // Split the discriminator into address/integer components.
    SDValue AddrDisc, IntDisc;
    std::tie(IntDisc, AddrDisc) =
        extractPtrauthBlendDiscriminators(CLI.PAI->Discriminator, &DAG);

    if (Opc == AArch64ISD::CALL_RVMARKER)
      Opc = AArch64ISD::AUTH_CALL_RVMARKER;
    else
      Opc = IsTailCall ? AArch64ISD::AUTH_TC_RETURN : AArch64ISD::AUTH_CALL;
    Ops.push_back(DAG.getTargetConstant(Key, DL, MVT::i32));
    Ops.push_back(IntDisc);
    Ops.push_back(AddrDisc);
  }

  // Add argument registers to the end of the list so that they are known live
  // into the call.
  for (auto &RegToPass : RegsToPass)
    Ops.push_back(DAG.getRegister(RegToPass.first,
                                  RegToPass.second.getValueType()));

  // Add a register mask operand representing the call-preserved registers.
  const uint32_t *Mask;
  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  if (IsThisReturn) {
    // For 'this' returns, use the X0-preserving mask if applicable
    Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
    if (!Mask) {
      IsThisReturn = false;
      Mask = TRI->getCallPreservedMask(MF, CallConv);
    }
  } else
    Mask = TRI->getCallPreservedMask(MF, CallConv);

  if (Subtarget->hasCustomCallingConv())
    TRI->UpdateCustomCallPreservedMask(MF, &Mask);

  if (TRI->isAnyArgRegReserved(MF))
    TRI->emitReservedArgRegCallError(MF);

  assert(Mask && "Missing call preserved mask for calling convention");
  Ops.push_back(DAG.getRegisterMask(Mask));

  if (InGlue.getNode())
    Ops.push_back(InGlue);

  // If we're doing a tall call, use a TC_RETURN here rather than an
  // actual call instruction.
  if (IsTailCall) {
    MF.getFrameInfo().setHasTailCall();
    SDValue Ret = DAG.getNode(Opc, DL, MVT::Other, Ops);
    if (IsCFICall)
      Ret.getNode()->setCFIType(CLI.CFIType->getZExtValue());

    DAG.addNoMergeSiteInfo(Ret.getNode(), CLI.NoMerge);
    DAG.addCallSiteInfo(Ret.getNode(), std::move(CSInfo));
    if (CalledGlobal &&
        MF.getFunction().getParent()->getModuleFlag("import-call-optimization"))
      DAG.addCalledGlobal(Ret.getNode(), CalledGlobal, OpFlags);
    return Ret;
  }

  // Returns a chain and a flag for retval copy to use.
  Chain = DAG.getNode(Opc, DL, {MVT::Other, MVT::Glue}, Ops);
  if (IsCFICall)
    Chain.getNode()->setCFIType(CLI.CFIType->getZExtValue());

  DAG.addNoMergeSiteInfo(Chain.getNode(), CLI.NoMerge);
  InGlue = Chain.getValue(1);
  DAG.addCallSiteInfo(Chain.getNode(), std::move(CSInfo));
  if (CalledGlobal &&
      MF.getFunction().getParent()->getModuleFlag("import-call-optimization"))
    DAG.addCalledGlobal(Chain.getNode(), CalledGlobal, OpFlags);

  uint64_t CalleePopBytes =
      DoesCalleeRestoreStack(CallConv, TailCallOpt) ? alignTo(NumBytes, 16) : 0;

  Chain = DAG.getCALLSEQ_END(Chain, NumBytes, CalleePopBytes, InGlue, DL);
  InGlue = Chain.getValue(1);

  // Handle result values, copying them out of physregs into vregs that we
  // return.
  SDValue Result = LowerCallResult(
      Chain, InGlue, CallConv, IsVarArg, RVLocs, DL, DAG, InVals, IsThisReturn,
      IsThisReturn ? OutVals[0] : SDValue(), RequiresSMChange);

  if (!Ins.empty())
    InGlue = Result.getValue(Result->getNumValues() - 1);

  if (RequiresSMChange) {
    assert(PStateSM && "Expected a PStateSM to be set");
    Result = changeStreamingMode(
        DAG, DL, !CallAttrs.callee().hasStreamingInterface(), Result, InGlue,
        getSMToggleCondition(CallAttrs), PStateSM);

    if (!Subtarget->isTargetDarwin() || Subtarget->hasSVE()) {
      InGlue = Result.getValue(1);
      Result =
          DAG.getNode(AArch64ISD::VG_RESTORE, DL,
                      DAG.getVTList(MVT::Other, MVT::Glue), {Result, InGlue});
    }
  }

  if (CallAttrs.requiresEnablingZAAfterCall())
    // Unconditionally resume ZA.
    Result = DAG.getNode(
        AArch64ISD::SMSTART, DL, DAG.getVTList(MVT::Other, MVT::Glue), Result,
        DAG.getTargetConstant((int32_t)(AArch64SVCR::SVCRZA), DL, MVT::i32));

  if (ShouldPreserveZT0)
    Result =
        DAG.getNode(AArch64ISD::RESTORE_ZT, DL, DAG.getVTList(MVT::Other),
                    {Result, DAG.getConstant(0, DL, MVT::i32), ZTFrameIdx});

  if (RequiresLazySave) {
    // Conditionally restore the lazy save using a pseudo node.
    TPIDR2Object &TPIDR2 = FuncInfo->getTPIDR2Obj();
    SDValue RegMask = DAG.getRegisterMask(
        TRI->SMEABISupportRoutinesCallPreservedMaskFromX0());
    SDValue RestoreRoutine = DAG.getTargetExternalSymbol(
        "__arm_tpidr2_restore", getPointerTy(DAG.getDataLayout()));
    SDValue TPIDR2_EL0 = DAG.getNode(
        ISD::INTRINSIC_W_CHAIN, DL, MVT::i64, Result,
        DAG.getConstant(Intrinsic::aarch64_sme_get_tpidr2, DL, MVT::i32));

    // Copy the address of the TPIDR2 block into X0 before 'calling' the
    // RESTORE_ZA pseudo.
    SDValue Glue;
    SDValue TPIDR2Block = DAG.getFrameIndex(
        TPIDR2.FrameIndex,
        DAG.getTargetLoweringInfo().getFrameIndexTy(DAG.getDataLayout()));
    Result = DAG.getCopyToReg(Result, DL, AArch64::X0, TPIDR2Block, Glue);
    Result =
        DAG.getNode(AArch64ISD::RESTORE_ZA, DL, MVT::Other,
                    {Result, TPIDR2_EL0, DAG.getRegister(AArch64::X0, MVT::i64),
                     RestoreRoutine, RegMask, Result.getValue(1)});

    // Finally reset the TPIDR2_EL0 register to 0.
    Result = DAG.getNode(
        ISD::INTRINSIC_VOID, DL, MVT::Other, Result,
        DAG.getConstant(Intrinsic::aarch64_sme_set_tpidr2, DL, MVT::i32),
        DAG.getConstant(0, DL, MVT::i64));
    TPIDR2.Uses++;
  } else if (RequiresSaveAllZA) {
    Result = emitSMEStateSaveRestore(*this, DAG, FuncInfo, DL, Result,
                                     /*IsSave=*/false);
  }

  if (RequiresSMChange || RequiresLazySave || ShouldPreserveZT0 ||
      RequiresSaveAllZA) {
    for (unsigned I = 0; I < InVals.size(); ++I) {
      // The smstart/smstop is chained as part of the call, but when the
      // resulting chain is discarded (which happens when the call is not part
      // of a chain, e.g. a call to @llvm.cos()), we need to ensure the
      // smstart/smstop is chained to the result value. We can do that by doing
      // a vreg -> vreg copy.
      Register Reg = MF.getRegInfo().createVirtualRegister(
          getRegClassFor(InVals[I].getValueType().getSimpleVT()));
      SDValue X = DAG.getCopyToReg(Result, DL, Reg, InVals[I]);
      InVals[I] = DAG.getCopyFromReg(X, DL, Reg,
                                     InVals[I].getValueType());
    }
  }

  if (CallConv == CallingConv::PreserveNone) {
    for (const ISD::OutputArg &O : Outs) {
      if (O.Flags.isSwiftSelf() || O.Flags.isSwiftError() ||
          O.Flags.isSwiftAsync()) {
        MachineFunction &MF = DAG.getMachineFunction();
        DAG.getContext()->diagnose(DiagnosticInfoUnsupported(
            MF.getFunction(),
            "Swift attributes can't be used with preserve_none",
            DL.getDebugLoc()));
        break;
      }
    }
  }

  return Result;
}

bool AArch64TargetLowering::CanLowerReturn(
    CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
    const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context,
    const Type *RetTy) const {
  CCAssignFn *RetCC = CCAssignFnForReturn(CallConv);
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
  return CCInfo.CheckReturn(Outs, RetCC);
}

SDValue
AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
                                   bool isVarArg,
                                   const SmallVectorImpl<ISD::OutputArg> &Outs,
                                   const SmallVectorImpl<SDValue> &OutVals,
                                   const SDLoc &DL, SelectionDAG &DAG) const {
  auto &MF = DAG.getMachineFunction();
  auto *FuncInfo = MF.getInfo<AArch64FunctionInfo>();

  CCAssignFn *RetCC = CCAssignFnForReturn(CallConv);
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
  CCInfo.AnalyzeReturn(Outs, RetCC);

  // Copy the result values into the output registers.
  SDValue Glue;
  SmallVector<std::pair<unsigned, SDValue>, 4> RetVals;
  SmallSet<unsigned, 4> RegsUsed;
  for (unsigned i = 0, realRVLocIdx = 0; i != RVLocs.size();
       ++i, ++realRVLocIdx) {
    CCValAssign &VA = RVLocs[i];
    assert(VA.isRegLoc() && "Can only return in registers!");
    SDValue Arg = OutVals[realRVLocIdx];

    switch (VA.getLocInfo()) {
    default:
      llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full:
      if (Outs[i].ArgVT == MVT::i1) {
        // AAPCS requires i1 to be zero-extended to i8 by the producer of the
        // value. This is strictly redundant on Darwin (which uses "zeroext
        // i1"), but will be optimised out before ISel.
        Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
        Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
      }
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::AExt:
    case CCValAssign::ZExt:
      Arg = DAG.getZExtOrTrunc(Arg, DL, VA.getLocVT());
      break;
    case CCValAssign::AExtUpper:
      assert(VA.getValVT() == MVT::i32 && "only expect 32 -> 64 upper bits");
      Arg = DAG.getZExtOrTrunc(Arg, DL, VA.getLocVT());
      Arg = DAG.getNode(ISD::SHL, DL, VA.getLocVT(), Arg,
                        DAG.getConstant(32, DL, VA.getLocVT()));
      break;
    }

    if (RegsUsed.count(VA.getLocReg())) {
      SDValue &Bits =
          llvm::find_if(RetVals, [=](const std::pair<unsigned, SDValue> &Elt) {
            return Elt.first == VA.getLocReg();
          })->second;
      Bits = DAG.getNode(ISD::OR, DL, Bits.getValueType(), Bits, Arg);
    } else {
      RetVals.emplace_back(VA.getLocReg(), Arg);
      RegsUsed.insert(VA.getLocReg());
    }
  }

  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();

  // Emit SMSTOP before returning from a locally streaming function
  SMEAttrs FuncAttrs = FuncInfo->getSMEFnAttrs();
  if (FuncAttrs.hasStreamingBody() && !FuncAttrs.hasStreamingInterface()) {
    if (FuncAttrs.hasStreamingCompatibleInterface()) {
      Register Reg = FuncInfo->getPStateSMReg();
      assert(Reg.isValid() && "PStateSM Register is invalid");
      SDValue PStateSM = DAG.getCopyFromReg(Chain, DL, Reg, MVT::i64);
      Chain = changeStreamingMode(DAG, DL, /*Enable*/ false, Chain,
                                  /*Glue*/ SDValue(),
                                  AArch64SME::IfCallerIsNonStreaming, PStateSM);
    } else
      Chain = changeStreamingMode(DAG, DL, /*Enable*/ false, Chain,
                                  /*Glue*/ SDValue(), AArch64SME::Always);
    Glue = Chain.getValue(1);
  }

  SmallVector<SDValue, 4> RetOps(1, Chain);
  for (auto &RetVal : RetVals) {
    if (FuncAttrs.hasStreamingBody() && !FuncAttrs.hasStreamingInterface() &&
        isPassedInFPR(RetVal.second.getValueType()))
      RetVal.second =
          DAG.getNode(AArch64ISD::COALESCER_BARRIER, DL,
                      DAG.getVTList(RetVal.second.getValueType(), MVT::Glue),
                      RetVal.second);
    Chain = DAG.getCopyToReg(Chain, DL, RetVal.first, RetVal.second, Glue);
    Glue = Chain.getValue(1);
    RetOps.push_back(
        DAG.getRegister(RetVal.first, RetVal.second.getValueType()));
  }

  // Windows AArch64 ABIs require that for returning structs by value we copy
  // the sret argument into X0 for the return.
  // We saved the argument into a virtual register in the entry block,
  // so now we copy the value out and into X0.
  if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
    SDValue Val = DAG.getCopyFromReg(RetOps[0], DL, SRetReg,
                                     getPointerTy(MF.getDataLayout()));

    unsigned RetValReg = AArch64::X0;
    if (CallConv == CallingConv::ARM64EC_Thunk_X64)
      RetValReg = AArch64::X8;
    Chain = DAG.getCopyToReg(Chain, DL, RetValReg, Val, Glue);
    Glue = Chain.getValue(1);

    RetOps.push_back(
      DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
  }

  const MCPhysReg *I = TRI->getCalleeSavedRegsViaCopy(&MF);
  if (I) {
    for (; *I; ++I) {
      if (AArch64::GPR64RegClass.contains(*I))
        RetOps.push_back(DAG.getRegister(*I, MVT::i64));
      else if (AArch64::FPR64RegClass.contains(*I))
        RetOps.push_back(DAG.getRegister(*I, MVT::getFloatingPointVT(64)));
      else
        llvm_unreachable("Unexpected register class in CSRsViaCopy!");
    }
  }

  RetOps[0] = Chain; // Update chain.

  // Add the glue if we have it.
  if (Glue.getNode())
    RetOps.push_back(Glue);

  if (CallConv == CallingConv::ARM64EC_Thunk_X64) {
    // ARM64EC entry thunks use a special return sequence: instead of a regular
    // "ret" instruction, they need to explicitly call the emulator.
    EVT PtrVT = getPointerTy(DAG.getDataLayout());
    SDValue Arm64ECRetDest =
        DAG.getExternalSymbol("__os_arm64x_dispatch_ret", PtrVT);
    Arm64ECRetDest =
        getAddr(cast<ExternalSymbolSDNode>(Arm64ECRetDest), DAG, 0);
    Arm64ECRetDest = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Arm64ECRetDest,
                                 MachinePointerInfo());
    RetOps.insert(RetOps.begin() + 1, Arm64ECRetDest);
    RetOps.insert(RetOps.begin() + 2, DAG.getTargetConstant(0, DL, MVT::i32));
    return DAG.getNode(AArch64ISD::TC_RETURN, DL, MVT::Other, RetOps);
  }

  return DAG.getNode(AArch64ISD::RET_GLUE, DL, MVT::Other, RetOps);
}

//===----------------------------------------------------------------------===//
//  Other Lowering Code
//===----------------------------------------------------------------------===//

SDValue AArch64TargetLowering::getTargetNode(GlobalAddressSDNode *N, EVT Ty,
                                             SelectionDAG &DAG,
                                             unsigned Flag) const {
  return DAG.getTargetGlobalAddress(N->getGlobal(), SDLoc(N), Ty,
                                    N->getOffset(), Flag);
}

SDValue AArch64TargetLowering::getTargetNode(JumpTableSDNode *N, EVT Ty,
                                             SelectionDAG &DAG,
                                             unsigned Flag) const {
  return DAG.getTargetJumpTable(N->getIndex(), Ty, Flag);
}

SDValue AArch64TargetLowering::getTargetNode(ConstantPoolSDNode *N, EVT Ty,
                                             SelectionDAG &DAG,
                                             unsigned Flag) const {
  return DAG.getTargetConstantPool(N->getConstVal(), Ty, N->getAlign(),
                                   N->getOffset(), Flag);
}

SDValue AArch64TargetLowering::getTargetNode(BlockAddressSDNode* N, EVT Ty,
                                             SelectionDAG &DAG,
                                             unsigned Flag) const {
  return DAG.getTargetBlockAddress(N->getBlockAddress(), Ty, 0, Flag);
}

SDValue AArch64TargetLowering::getTargetNode(ExternalSymbolSDNode *N, EVT Ty,
                                             SelectionDAG &DAG,
                                             unsigned Flag) const {
  return DAG.getTargetExternalSymbol(N->getSymbol(), Ty, Flag);
}

// (loadGOT sym)
template <class NodeTy>
SDValue AArch64TargetLowering::getGOT(NodeTy *N, SelectionDAG &DAG,
                                      unsigned Flags) const {
  LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getGOT\n");
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  SDValue GotAddr = getTargetNode(N, Ty, DAG, AArch64II::MO_GOT | Flags);
  // FIXME: Once remat is capable of dealing with instructions with register
  // operands, expand this into two nodes instead of using a wrapper node.
  if (DAG.getMachineFunction()
          .getInfo<AArch64FunctionInfo>()
          ->hasELFSignedGOT())
    return SDValue(DAG.getMachineNode(AArch64::LOADgotAUTH, DL, Ty, GotAddr),
                   0);
  return DAG.getNode(AArch64ISD::LOADgot, DL, Ty, GotAddr);
}

// (wrapper %highest(sym), %higher(sym), %hi(sym), %lo(sym))
template <class NodeTy>
SDValue AArch64TargetLowering::getAddrLarge(NodeTy *N, SelectionDAG &DAG,
                                            unsigned Flags) const {
  LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getAddrLarge\n");
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  const unsigned char MO_NC = AArch64II::MO_NC;
  return DAG.getNode(
      AArch64ISD::WrapperLarge, DL, Ty,
      getTargetNode(N, Ty, DAG, AArch64II::MO_G3 | Flags),
      getTargetNode(N, Ty, DAG, AArch64II::MO_G2 | MO_NC | Flags),
      getTargetNode(N, Ty, DAG, AArch64II::MO_G1 | MO_NC | Flags),
      getTargetNode(N, Ty, DAG, AArch64II::MO_G0 | MO_NC | Flags));
}

// (addlow (adrp %hi(sym)) %lo(sym))
template <class NodeTy>
SDValue AArch64TargetLowering::getAddr(NodeTy *N, SelectionDAG &DAG,
                                       unsigned Flags) const {
  LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getAddr\n");
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  SDValue Hi = getTargetNode(N, Ty, DAG, AArch64II::MO_PAGE | Flags);
  SDValue Lo = getTargetNode(N, Ty, DAG,
                             AArch64II::MO_PAGEOFF | AArch64II::MO_NC | Flags);
  SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, Ty, Hi);
  return DAG.getNode(AArch64ISD::ADDlow, DL, Ty, ADRP, Lo);
}

// (adr sym)
template <class NodeTy>
SDValue AArch64TargetLowering::getAddrTiny(NodeTy *N, SelectionDAG &DAG,
                                           unsigned Flags) const {
  LLVM_DEBUG(dbgs() << "AArch64TargetLowering::getAddrTiny\n");
  SDLoc DL(N);
  EVT Ty = getPointerTy(DAG.getDataLayout());
  SDValue Sym = getTargetNode(N, Ty, DAG, Flags);
  return DAG.getNode(AArch64ISD::ADR, DL, Ty, Sym);
}

SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
                                                  SelectionDAG &DAG) const {
  GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
  const GlobalValue *GV = GN->getGlobal();
  unsigned OpFlags = Subtarget->ClassifyGlobalReference(GV, getTargetMachine());

  if (OpFlags != AArch64II::MO_NO_FLAG)
    assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
           "unexpected offset in global node");

  // This also catches the large code model case for Darwin, and tiny code
  // model with got relocations.
  if ((OpFlags & AArch64II::MO_GOT) != 0) {
    return getGOT(GN, DAG, OpFlags);
  }

  SDValue Result;
  if (getTargetMachine().getCodeModel() == CodeModel::Large &&
      !getTargetMachine().isPositionIndependent()) {
    Result = getAddrLarge(GN, DAG, OpFlags);
  } else if (getTargetMachine().getCodeModel() == CodeModel::Tiny) {
    Result = getAddrTiny(GN, DAG, OpFlags);
  } else {
    Result = getAddr(GN, DAG, OpFlags);
  }
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(GN);
  if (OpFlags & (AArch64II::MO_DLLIMPORT | AArch64II::MO_COFFSTUB))
    Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
                         MachinePointerInfo::getGOT(DAG.getMachineFunction()));
  return Result;
}

/// Convert a TLS address reference into the correct sequence of loads
/// and calls to compute the variable's address (for Darwin, currently) and
/// return an SDValue containing the final node.

/// Darwin only has one TLS scheme which must be capable of dealing with the
/// fully general situation, in the worst case. This means:
///     + "extern __thread" declaration.
///     + Defined in a possibly unknown dynamic library.
///
/// The general system is that each __thread variable has a [3 x i64] descriptor
/// which contains information used by the runtime to calculate the address. The
/// only part of this the compiler needs to know about is the first xword, which
/// contains a function pointer that must be called with the address of the
/// entire descriptor in "x0".
///
/// Since this descriptor may be in a different unit, in general even the
/// descriptor must be accessed via an indirect load. The "ideal" code sequence
/// is:
///     adrp x0, _var@TLVPPAGE
///     ldr x0, [x0, _var@TLVPPAGEOFF]   ; x0 now contains address of descriptor
///     ldr x1, [x0]                     ; x1 contains 1st entry of descriptor,
///                                      ; the function pointer
///     blr x1                           ; Uses descriptor address in x0
///     ; Address of _var is now in x0.
///
/// If the address of _var's descriptor *is* known to the linker, then it can
/// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
/// a slight efficiency gain.
SDValue
AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
                                                   SelectionDAG &DAG) const {
  assert(Subtarget->isTargetDarwin() &&
         "This function expects a Darwin target");

  SDLoc DL(Op);
  MVT PtrVT = getPointerTy(DAG.getDataLayout());
  MVT PtrMemVT = getPointerMemTy(DAG.getDataLayout());
  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();

  SDValue TLVPAddr =
      DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
  SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);

  // The first entry in the descriptor is a function pointer that we must call
  // to obtain the address of the variable.
  SDValue Chain = DAG.getEntryNode();
  SDValue FuncTLVGet = DAG.getLoad(
      PtrMemVT, DL, Chain, DescAddr,
      MachinePointerInfo::getGOT(DAG.getMachineFunction()),
      Align(PtrMemVT.getSizeInBits() / 8),
      MachineMemOperand::MOInvariant | MachineMemOperand::MODereferenceable);
  Chain = FuncTLVGet.getValue(1);

  // Extend loaded pointer if necessary (i.e. if ILP32) to DAG pointer.
  FuncTLVGet = DAG.getZExtOrTrunc(FuncTLVGet, DL, PtrVT);

  MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
  MFI.setAdjustsStack(true);

  // TLS calls preserve all registers except those that absolutely must be
  // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
  // silly).
  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  const uint32_t *Mask = TRI->getTLSCallPreservedMask();
  if (Subtarget->hasCustomCallingConv())
    TRI->UpdateCustomCallPreservedMask(DAG.getMachineFunction(), &Mask);

  // Finally, we can make the call. This is just a degenerate version of a
  // normal AArch64 call node: x0 takes the address of the descriptor, and
  // returns the address of the variable in this thread.
  Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());

  unsigned Opcode = AArch64ISD::CALL;
  SmallVector<SDValue, 8> Ops;
  Ops.push_back(Chain);
  Ops.push_back(FuncTLVGet);

  // With ptrauth-calls, the tlv access thunk pointer is authenticated (IA, 0).
  if (DAG.getMachineFunction().getFunction().hasFnAttribute("ptrauth-calls")) {
    Opcode = AArch64ISD::AUTH_CALL;
    Ops.push_back(DAG.getTargetConstant(AArch64PACKey::IA, DL, MVT::i32));
    Ops.push_back(DAG.getTargetConstant(0, DL, MVT::i64)); // Integer Disc.
    Ops.push_back(DAG.getRegister(AArch64::NoRegister, MVT::i64)); // Addr Disc.
  }

  Ops.push_back(DAG.getRegister(AArch64::X0, MVT::i64));
  Ops.push_back(DAG.getRegisterMask(Mask));
  Ops.push_back(Chain.getValue(1));
  Chain = DAG.getNode(Opcode, DL, DAG.getVTList(MVT::Other, MVT::Glue), Ops);
  return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
}

/// Convert a thread-local variable reference into a sequence of instructions to
/// compute the variable's address for the local exec TLS model of ELF targets.
/// The sequence depends on the maximum TLS area size.
SDValue AArch64TargetLowering::LowerELFTLSLocalExec(const GlobalValue *GV,
                                                    SDValue ThreadBase,
                                                    const SDLoc &DL,
                                                    SelectionDAG &DAG) const {
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDValue TPOff, Addr;

  switch (DAG.getTarget().Options.TLSSize) {
  default:
    llvm_unreachable("Unexpected TLS size");

  case 12: {
    // mrs   x0, TPIDR_EL0
    // add   x0, x0, :tprel_lo12:a
    SDValue Var = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_PAGEOFF);
    return SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
                                      Var,
                                      DAG.getTargetConstant(0, DL, MVT::i32)),
                   0);
  }

  case 24: {
    // mrs   x0, TPIDR_EL0
    // add   x0, x0, :tprel_hi12:a
    // add   x0, x0, :tprel_lo12_nc:a
    SDValue HiVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
    SDValue LoVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0,
        AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
    Addr = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
                                      HiVar,
                                      DAG.getTargetConstant(0, DL, MVT::i32)),
                   0);
    return SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, Addr,
                                      LoVar,
                                      DAG.getTargetConstant(0, DL, MVT::i32)),
                   0);
  }

  case 32: {
    // mrs   x1, TPIDR_EL0
    // movz  x0, #:tprel_g1:a
    // movk  x0, #:tprel_g0_nc:a
    // add   x0, x1, x0
    SDValue HiVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G1);
    SDValue LoVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0,
        AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
    TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
                                       DAG.getTargetConstant(16, DL, MVT::i32)),
                    0);
    TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar,
                                       DAG.getTargetConstant(0, DL, MVT::i32)),
                    0);
    return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
  }

  case 48: {
    // mrs   x1, TPIDR_EL0
    // movz  x0, #:tprel_g2:a
    // movk  x0, #:tprel_g1_nc:a
    // movk  x0, #:tprel_g0_nc:a
    // add   x0, x1, x0
    SDValue HiVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_G2);
    SDValue MiVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0,
        AArch64II::MO_TLS | AArch64II::MO_G1 | AArch64II::MO_NC);
    SDValue LoVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0,
        AArch64II::MO_TLS | AArch64II::MO_G0 | AArch64II::MO_NC);
    TPOff = SDValue(DAG.getMachineNode(AArch64::MOVZXi, DL, PtrVT, HiVar,
                                       DAG.getTargetConstant(32, DL, MVT::i32)),
                    0);
    TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, MiVar,
                                       DAG.getTargetConstant(16, DL, MVT::i32)),
                    0);
    TPOff = SDValue(DAG.getMachineNode(AArch64::MOVKXi, DL, PtrVT, TPOff, LoVar,
                                       DAG.getTargetConstant(0, DL, MVT::i32)),
                    0);
    return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
  }
  }
}

/// When accessing thread-local variables under either the general-dynamic or
/// local-dynamic system, we make a "TLS-descriptor" call. The variable will
/// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
/// is a function pointer to carry out the resolution.
///
/// The sequence is:
///    adrp  x0, :tlsdesc:var
///    ldr   x1, [x0, #:tlsdesc_lo12:var]
///    add   x0, x0, #:tlsdesc_lo12:var
///    .tlsdesccall var
///    blr   x1
///    (TPIDR_EL0 offset now in x0)
///
///  The above sequence must be produced unscheduled, to enable the linker to
///  optimize/relax this sequence.
///  Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
///  above sequence, and expanded really late in the compilation flow, to ensure
///  the sequence is produced as per above.
SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr,
                                                      const SDLoc &DL,
                                                      SelectionDAG &DAG) const {
  EVT PtrVT = getPointerTy(DAG.getDataLayout());

  SDValue Chain = DAG.getEntryNode();
  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);

  unsigned Opcode =
      DAG.getMachineFunction().getInfo<AArch64FunctionInfo>()->hasELFSignedGOT()
          ? AArch64ISD::TLSDESC_AUTH_CALLSEQ
          : AArch64ISD::TLSDESC_CALLSEQ;
  Chain = DAG.getNode(Opcode, DL, NodeTys, {Chain, SymAddr});
  SDValue Glue = Chain.getValue(1);

  return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
}

SDValue
AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
                                                SelectionDAG &DAG) const {
  assert(Subtarget->isTargetELF() && "This function expects an ELF target");

  const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
  AArch64FunctionInfo *MFI =
      DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();

  TLSModel::Model Model = MFI->hasELFSignedGOT()
                              ? TLSModel::GeneralDynamic
                              : getTargetMachine().getTLSModel(GA->getGlobal());

  if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
    if (Model == TLSModel::LocalDynamic)
      Model = TLSModel::GeneralDynamic;
  }

  if (getTargetMachine().getCodeModel() == CodeModel::Large &&
      Model != TLSModel::LocalExec)
    report_fatal_error("ELF TLS only supported in small memory model or "
                       "in local exec TLS model");
  // Different choices can be made for the maximum size of the TLS area for a
  // module. For the small address model, the default TLS size is 16MiB and the
  // maximum TLS size is 4GiB.
  // FIXME: add tiny and large code model support for TLS access models other
  // than local exec. We currently generate the same code as small for tiny,
  // which may be larger than needed.

  SDValue TPOff;
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);
  const GlobalValue *GV = GA->getGlobal();

  SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);

  if (Model == TLSModel::LocalExec) {
    return LowerELFTLSLocalExec(GV, ThreadBase, DL, DAG);
  } else if (Model == TLSModel::InitialExec) {
    TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
    TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
  } else if (Model == TLSModel::LocalDynamic) {
    // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
    // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
    // the beginning of the module's TLS region, followed by a DTPREL offset
    // calculation.

    // These accesses will need deduplicating if there's more than one.
    MFI->incNumLocalDynamicTLSAccesses();

    // The call needs a relocation too for linker relaxation. It doesn't make
    // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
    // the address.
    SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
                                                  AArch64II::MO_TLS);

    // Now we can calculate the offset from TPIDR_EL0 to this module's
    // thread-local area.
    TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);

    // Now use :dtprel_whatever: operations to calculate this variable's offset
    // in its thread-storage area.
    SDValue HiVar = DAG.getTargetGlobalAddress(
        GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
    SDValue LoVar = DAG.getTargetGlobalAddress(
        GV, DL, MVT::i64, 0,
        AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);

    TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
                                       DAG.getTargetConstant(0, DL, MVT::i32)),
                    0);
    TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
                                       DAG.getTargetConstant(0, DL, MVT::i32)),
                    0);
  } else if (Model == TLSModel::GeneralDynamic) {
    // The call needs a relocation too for linker relaxation. It doesn't make
    // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
    // the address.
    SDValue SymAddr =
        DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);

    // Finally we can make a call to calculate the offset from tpidr_el0.
    TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
  } else
    llvm_unreachable("Unsupported ELF TLS access model");

  return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
}

SDValue
AArch64TargetLowering::LowerWindowsGlobalTLSAddress(SDValue Op,
                                                    SelectionDAG &DAG) const {
  assert(Subtarget->isTargetWindows() && "Windows specific TLS lowering");

  SDValue Chain = DAG.getEntryNode();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);

  SDValue TEB = DAG.getRegister(AArch64::X18, MVT::i64);

  // Load the ThreadLocalStoragePointer from the TEB
  // A pointer to the TLS array is located at offset 0x58 from the TEB.
  SDValue TLSArray =
      DAG.getNode(ISD::ADD, DL, PtrVT, TEB, DAG.getIntPtrConstant(0x58, DL));
  TLSArray = DAG.getLoad(PtrVT, DL, Chain, TLSArray, MachinePointerInfo());
  Chain = TLSArray.getValue(1);

  // Load the TLS index from the C runtime;
  // This does the same as getAddr(), but without having a GlobalAddressSDNode.
  // This also does the same as LOADgot, but using a generic i32 load,
  // while LOADgot only loads i64.
  SDValue TLSIndexHi =
      DAG.getTargetExternalSymbol("_tls_index", PtrVT, AArch64II::MO_PAGE);
  SDValue TLSIndexLo = DAG.getTargetExternalSymbol(
      "_tls_index", PtrVT, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
  SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, TLSIndexHi);
  SDValue TLSIndex =
      DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, TLSIndexLo);
  TLSIndex = DAG.getLoad(MVT::i32, DL, Chain, TLSIndex, MachinePointerInfo());
  Chain = TLSIndex.getValue(1);

  // The pointer to the thread's TLS data area is at the TLS Index scaled by 8
  // offset into the TLSArray.
  TLSIndex = DAG.getNode(ISD::ZERO_EXTEND, DL, PtrVT, TLSIndex);
  SDValue Slot = DAG.getNode(ISD::SHL, DL, PtrVT, TLSIndex,
                             DAG.getConstant(3, DL, PtrVT));
  SDValue TLS = DAG.getLoad(PtrVT, DL, Chain,
                            DAG.getNode(ISD::ADD, DL, PtrVT, TLSArray, Slot),
                            MachinePointerInfo());
  Chain = TLS.getValue(1);

  const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
  const GlobalValue *GV = GA->getGlobal();
  SDValue TGAHi = DAG.getTargetGlobalAddress(
      GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
  SDValue TGALo = DAG.getTargetGlobalAddress(
      GV, DL, PtrVT, 0,
      AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);

  // Add the offset from the start of the .tls section (section base).
  SDValue Addr =
      SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TLS, TGAHi,
                                 DAG.getTargetConstant(0, DL, MVT::i32)),
              0);
  Addr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, Addr, TGALo);
  return Addr;
}

SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
                                                     SelectionDAG &DAG) const {
  const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
  if (DAG.getTarget().useEmulatedTLS())
    return LowerToTLSEmulatedModel(GA, DAG);

  if (Subtarget->isTargetDarwin())
    return LowerDarwinGlobalTLSAddress(Op, DAG);
  if (Subtarget->isTargetELF())
    return LowerELFGlobalTLSAddress(Op, DAG);
  if (Subtarget->isTargetWindows())
    return LowerWindowsGlobalTLSAddress(Op, DAG);

  llvm_unreachable("Unexpected platform trying to use TLS");
}

//===----------------------------------------------------------------------===//
//                      PtrAuthGlobalAddress lowering
//
// We have 3 lowering alternatives to choose from:
// - MOVaddrPAC: similar to MOVaddr, with added PAC.
//   If the GV doesn't need a GOT load (i.e., is locally defined)
//   materialize the pointer using adrp+add+pac. See LowerMOVaddrPAC.
//
// - LOADgotPAC: similar to LOADgot, with added PAC.
//   If the GV needs a GOT load, materialize the pointer using the usual
//   GOT adrp+ldr, +pac. Pointers in GOT are assumed to be not signed, the GOT
//   section is assumed to be read-only (for example, via relro mechanism). See
//   LowerMOVaddrPAC.
//
// - LOADauthptrstatic: similar to LOADgot, but use a
//   special stub slot instead of a GOT slot.
//   Load a signed pointer for symbol 'sym' from a stub slot named
//   'sym$auth_ptr$key$disc' filled by dynamic linker during relocation
//   resolving. This usually lowers to adrp+ldr, but also emits an entry into
//   .data with an @AUTH relocation. See LowerLOADauthptrstatic.
//
// All 3 are pseudos that are expand late to longer sequences: this lets us
// provide integrity guarantees on the to-be-signed intermediate values.
//
// LOADauthptrstatic is undesirable because it requires a large section filled
// with often similarly-signed pointers, making it a good harvesting target.
// Thus, it's only used for ptrauth references to extern_weak to avoid null
// checks.

static SDValue LowerPtrAuthGlobalAddressStatically(
    SDValue TGA, SDLoc DL, EVT VT, AArch64PACKey::ID KeyC,
    SDValue Discriminator, SDValue AddrDiscriminator, SelectionDAG &DAG) {
  const auto *TGN = cast<GlobalAddressSDNode>(TGA.getNode());
  assert(TGN->getGlobal()->hasExternalWeakLinkage());

  // Offsets and extern_weak don't mix well: ptrauth aside, you'd get the
  // offset alone as a pointer if the symbol wasn't available, which would
  // probably break null checks in users. Ptrauth complicates things further:
  // error out.
  if (TGN->getOffset() != 0)
    report_fatal_error(
        "unsupported non-zero offset in weak ptrauth global reference");

  if (!isNullConstant(AddrDiscriminator))
    report_fatal_error("unsupported weak addr-div ptrauth global");

  SDValue Key = DAG.getTargetConstant(KeyC, DL, MVT::i32);
  return SDValue(DAG.getMachineNode(AArch64::LOADauthptrstatic, DL, MVT::i64,
                                    {TGA, Key, Discriminator}),
                 0);
}

SDValue
AArch64TargetLowering::LowerPtrAuthGlobalAddress(SDValue Op,
                                                 SelectionDAG &DAG) const {
  SDValue Ptr = Op.getOperand(0);
  uint64_t KeyC = Op.getConstantOperandVal(1);
  SDValue AddrDiscriminator = Op.getOperand(2);
  uint64_t DiscriminatorC = Op.getConstantOperandVal(3);
  EVT VT = Op.getValueType();
  SDLoc DL(Op);

  if (KeyC > AArch64PACKey::LAST)
    report_fatal_error("key in ptrauth global out of range [0, " +
                       Twine((int)AArch64PACKey::LAST) + "]");

  // Blend only works if the integer discriminator is 16-bit wide.
  if (!isUInt<16>(DiscriminatorC))
    report_fatal_error(
        "constant discriminator in ptrauth global out of range [0, 0xffff]");

  // Choosing between 3 lowering alternatives is target-specific.
  if (!Subtarget->isTargetELF() && !Subtarget->isTargetMachO())
    report_fatal_error("ptrauth global lowering only supported on MachO/ELF");

  int64_t PtrOffsetC = 0;
  if (Ptr.getOpcode() == ISD::ADD) {
    PtrOffsetC = Ptr.getConstantOperandVal(1);
    Ptr = Ptr.getOperand(0);
  }
  const auto *PtrN = cast<GlobalAddressSDNode>(Ptr.getNode());
  const GlobalValue *PtrGV = PtrN->getGlobal();

  // Classify the reference to determine whether it needs a GOT load.
  const unsigned OpFlags =
      Subtarget->ClassifyGlobalReference(PtrGV, getTargetMachine());
  const bool NeedsGOTLoad = ((OpFlags & AArch64II::MO_GOT) != 0);
  assert(((OpFlags & (~AArch64II::MO_GOT)) == 0) &&
         "unsupported non-GOT op flags on ptrauth global reference");

  // Fold any offset into the GV; our pseudos expect it there.
  PtrOffsetC += PtrN->getOffset();
  SDValue TPtr = DAG.getTargetGlobalAddress(PtrGV, DL, VT, PtrOffsetC,
                                            /*TargetFlags=*/0);
  assert(PtrN->getTargetFlags() == 0 &&
         "unsupported target flags on ptrauth global");

  SDValue Key = DAG.getTargetConstant(KeyC, DL, MVT::i32);
  SDValue Discriminator = DAG.getTargetConstant(DiscriminatorC, DL, MVT::i64);
  SDValue TAddrDiscriminator = !isNullConstant(AddrDiscriminator)
                                   ? AddrDiscriminator
                                   : DAG.getRegister(AArch64::XZR, MVT::i64);

  // No GOT load needed -> MOVaddrPAC
  if (!NeedsGOTLoad) {
    assert(!PtrGV->hasExternalWeakLinkage() && "extern_weak should use GOT");
    return SDValue(
        DAG.getMachineNode(AArch64::MOVaddrPAC, DL, MVT::i64,
                           {TPtr, Key, TAddrDiscriminator, Discriminator}),
        0);
  }

  // GOT load -> LOADgotPAC
  // Note that we disallow extern_weak refs to avoid null checks later.
  if (!PtrGV->hasExternalWeakLinkage())
    return SDValue(
        DAG.getMachineNode(AArch64::LOADgotPAC, DL, MVT::i64,
                           {TPtr, Key, TAddrDiscriminator, Discriminator}),
        0);

  // extern_weak ref -> LOADauthptrstatic
  return LowerPtrAuthGlobalAddressStatically(
      TPtr, DL, VT, (AArch64PACKey::ID)KeyC, Discriminator, AddrDiscriminator,
      DAG);
}

// Looks through \param Val to determine the bit that can be used to
// check the sign of the value. It returns the unextended value and
// the sign bit position.
std::pair<SDValue, uint64_t> lookThroughSignExtension(SDValue Val) {
  if (Val.getOpcode() == ISD::SIGN_EXTEND_INREG)
    return {Val.getOperand(0),
            cast<VTSDNode>(Val.getOperand(1))->getVT().getFixedSizeInBits() -
                1};

  if (Val.getOpcode() == ISD::SIGN_EXTEND)
    return {Val.getOperand(0),
            Val.getOperand(0)->getValueType(0).getFixedSizeInBits() - 1};

  return {Val, Val.getValueSizeInBits() - 1};
}

SDValue AArch64TargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
  SDValue Chain = Op.getOperand(0);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
  SDValue LHS = Op.getOperand(2);
  SDValue RHS = Op.getOperand(3);
  SDValue Dest = Op.getOperand(4);
  SDLoc DL(Op);

  MachineFunction &MF = DAG.getMachineFunction();
  // Speculation tracking/SLH assumes that optimized TB(N)Z/CB(N)Z instructions
  // will not be produced, as they are conditional branch instructions that do
  // not set flags.
  bool ProduceNonFlagSettingCondBr =
      !MF.getFunction().hasFnAttribute(Attribute::SpeculativeLoadHardening);

  // Handle f128 first, since lowering it will result in comparing the return
  // value of a libcall against zero, which is just what the rest of LowerBR_CC
  // is expecting to deal with.
  if (LHS.getValueType() == MVT::f128) {
    softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, DL, LHS, RHS);

    // If softenSetCCOperands returned a scalar, we need to compare the result
    // against zero to select between true and false values.
    if (!RHS.getNode()) {
      RHS = DAG.getConstant(0, DL, LHS.getValueType());
      CC = ISD::SETNE;
    }
  }

  // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
  // instruction.
  if (ISD::isOverflowIntrOpRes(LHS) && isOneConstant(RHS) &&
      (CC == ISD::SETEQ || CC == ISD::SETNE)) {
    // Only lower legal XALUO ops.
    if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
      return SDValue();

    // The actual operation with overflow check.
    AArch64CC::CondCode OFCC;
    SDValue Value, Overflow;
    std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);

    if (CC == ISD::SETNE)
      OFCC = getInvertedCondCode(OFCC);
    SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);

    return DAG.getNode(AArch64ISD::BRCOND, DL, MVT::Other, Chain, Dest, CCVal,
                       Overflow);
  }

  if (LHS.getValueType().isInteger()) {
    assert((LHS.getValueType() == RHS.getValueType()) &&
           (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));

    // If the RHS of the comparison is zero, we can potentially fold this
    // to a specialized branch.
    const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
    if (RHSC && RHSC->getZExtValue() == 0 && ProduceNonFlagSettingCondBr) {
      if (CC == ISD::SETEQ) {
        // See if we can use a TBZ to fold in an AND as well.
        // TBZ has a smaller branch displacement than CBZ.  If the offset is
        // out of bounds, a late MI-layer pass rewrites branches.
        // 403.gcc is an example that hits this case.
        if (LHS.getOpcode() == ISD::AND &&
            isa<ConstantSDNode>(LHS.getOperand(1)) &&
            isPowerOf2_64(LHS.getConstantOperandVal(1))) {
          SDValue Test = LHS.getOperand(0);
          uint64_t Mask = LHS.getConstantOperandVal(1);
          return DAG.getNode(AArch64ISD::TBZ, DL, MVT::Other, Chain, Test,
                             DAG.getConstant(Log2_64(Mask), DL, MVT::i64),
                             Dest);
        }

        return DAG.getNode(AArch64ISD::CBZ, DL, MVT::Other, Chain, LHS, Dest);
      } else if (CC == ISD::SETNE) {
        // See if we can use a TBZ to fold in an AND as well.
        // TBZ has a smaller branch displacement than CBZ.  If the offset is
        // out of bounds, a late MI-layer pass rewrites branches.
        // 403.gcc is an example that hits this case.
        if (LHS.getOpcode() == ISD::AND &&
            isa<ConstantSDNode>(LHS.getOperand(1)) &&
            isPowerOf2_64(LHS.getConstantOperandVal(1))) {
          SDValue Test = LHS.getOperand(0);
          uint64_t Mask = LHS.getConstantOperandVal(1);
          return DAG.getNode(AArch64ISD::TBNZ, DL, MVT::Other, Chain, Test,
                             DAG.getConstant(Log2_64(Mask), DL, MVT::i64),
                             Dest);
        }

        return DAG.getNode(AArch64ISD::CBNZ, DL, MVT::Other, Chain, LHS, Dest);
      } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
        // Don't combine AND since emitComparison converts the AND to an ANDS
        // (a.k.a. TST) and the test in the test bit and branch instruction
        // becomes redundant.  This would also increase register pressure.
        uint64_t SignBitPos;
        std::tie(LHS, SignBitPos) = lookThroughSignExtension(LHS);
        return DAG.getNode(AArch64ISD::TBNZ, DL, MVT::Other, Chain, LHS,
                           DAG.getConstant(SignBitPos, DL, MVT::i64), Dest);
      }
    }
    if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
        LHS.getOpcode() != ISD::AND && ProduceNonFlagSettingCondBr) {
      // Don't combine AND since emitComparison converts the AND to an ANDS
      // (a.k.a. TST) and the test in the test bit and branch instruction
      // becomes redundant.  This would also increase register pressure.
      uint64_t SignBitPos;
      std::tie(LHS, SignBitPos) = lookThroughSignExtension(LHS);
      return DAG.getNode(AArch64ISD::TBZ, DL, MVT::Other, Chain, LHS,
                         DAG.getConstant(SignBitPos, DL, MVT::i64), Dest);
    }

    // Try to emit Armv9.6 CB instructions. We prefer tb{n}z/cb{n}z due to their
    // larger branch displacement but do prefer CB over cmp + br.
    if (Subtarget->hasCMPBR() &&
        AArch64CC::isValidCBCond(changeIntCCToAArch64CC(CC)) &&
        ProduceNonFlagSettingCondBr) {
      SDValue Cond =
          DAG.getTargetConstant(changeIntCCToAArch64CC(CC), DL, MVT::i32);
      return DAG.getNode(AArch64ISD::CB, DL, MVT::Other, Chain, Cond, LHS, RHS,
                         Dest);
    }

    SDValue CCVal;
    SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, DL);
    return DAG.getNode(AArch64ISD::BRCOND, DL, MVT::Other, Chain, Dest, CCVal,
                       Cmp);
  }

  assert(LHS.getValueType() == MVT::f16 || LHS.getValueType() == MVT::bf16 ||
         LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);

  // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
  // clean.  Some of them require two branches to implement.
  SDValue Cmp = emitComparison(LHS, RHS, CC, DL, DAG);
  AArch64CC::CondCode CC1, CC2;
  changeFPCCToAArch64CC(CC, CC1, CC2);
  SDValue CC1Val = DAG.getConstant(CC1, DL, MVT::i32);
  SDValue BR1 =
      DAG.getNode(AArch64ISD::BRCOND, DL, MVT::Other, Chain, Dest, CC1Val, Cmp);
  if (CC2 != AArch64CC::AL) {
    SDValue CC2Val = DAG.getConstant(CC2, DL, MVT::i32);
    return DAG.getNode(AArch64ISD::BRCOND, DL, MVT::Other, BR1, Dest, CC2Val,
                       Cmp);
  }

  return BR1;
}

SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
                                              SelectionDAG &DAG) const {
  if (!Subtarget->isNeonAvailable() &&
      !Subtarget->useSVEForFixedLengthVectors())
    return SDValue();

  EVT VT = Op.getValueType();
  EVT IntVT = VT.changeTypeToInteger();
  SDLoc DL(Op);

  SDValue In1 = Op.getOperand(0);
  SDValue In2 = Op.getOperand(1);
  EVT SrcVT = In2.getValueType();

  if (!SrcVT.bitsEq(VT))
    In2 = DAG.getFPExtendOrRound(In2, DL, VT);

  if (VT.isScalableVector())
    IntVT =
        getPackedSVEVectorVT(VT.getVectorElementType().changeTypeToInteger());

  if (VT.isFixedLengthVector() &&
      useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable())) {
    EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);

    In1 = convertToScalableVector(DAG, ContainerVT, In1);
    In2 = convertToScalableVector(DAG, ContainerVT, In2);

    SDValue Res = DAG.getNode(ISD::FCOPYSIGN, DL, ContainerVT, In1, In2);
    return convertFromScalableVector(DAG, VT, Res);
  }

  // With SVE, but without Neon, extend the scalars to scalable vectors and use
  // a SVE FCOPYSIGN.
  if (!VT.isVector() && !Subtarget->isNeonAvailable() &&
      Subtarget->isSVEorStreamingSVEAvailable()) {
    if (VT != MVT::f16 && VT != MVT::f32 && VT != MVT::f64 && VT != MVT::bf16)
      return SDValue();
    EVT SVT = getPackedSVEVectorVT(VT);

    SDValue Ins1 =
        DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, SVT, DAG.getUNDEF(SVT), In1,
                    DAG.getConstant(0, DL, MVT::i64));
    SDValue Ins2 =
        DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, SVT, DAG.getUNDEF(SVT), In2,
                    DAG.getConstant(0, DL, MVT::i64));
    SDValue FCS = DAG.getNode(ISD::FCOPYSIGN, DL, SVT, Ins1, Ins2);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, FCS,
                       DAG.getConstant(0, DL, MVT::i64));
  }

  auto BitCast = [this](EVT VT, SDValue Op, SelectionDAG &DAG) {
    if (VT.isScalableVector())
      return getSVESafeBitCast(VT, Op, DAG);

    return DAG.getBitcast(VT, Op);
  };

  SDValue VecVal1, VecVal2;
  EVT VecVT;
  auto SetVecVal = [&](int Idx = -1) {
    if (!VT.isVector()) {
      VecVal1 =
          DAG.getTargetInsertSubreg(Idx, DL, VecVT, DAG.getUNDEF(VecVT), In1);
      VecVal2 =
          DAG.getTargetInsertSubreg(Idx, DL, VecVT, DAG.getUNDEF(VecVT), In2);
    } else {
      VecVal1 = BitCast(VecVT, In1, DAG);
      VecVal2 = BitCast(VecVT, In2, DAG);
    }
  };
  if (VT.isVector()) {
    VecVT = IntVT;
    SetVecVal();
  } else if (VT == MVT::f64) {
    VecVT = MVT::v2i64;
    SetVecVal(AArch64::dsub);
  } else if (VT == MVT::f32) {
    VecVT = MVT::v4i32;
    SetVecVal(AArch64::ssub);
  } else if (VT == MVT::f16 || VT == MVT::bf16) {
    VecVT = MVT::v8i16;
    SetVecVal(AArch64::hsub);
  } else {
    llvm_unreachable("Invalid type for copysign!");
  }

  unsigned BitWidth = In1.getScalarValueSizeInBits();
  SDValue SignMaskV = DAG.getConstant(~APInt::getSignMask(BitWidth), DL, VecVT);

  // We want to materialize a mask with every bit but the high bit set, but the
  // AdvSIMD immediate moves cannot materialize that in a single instruction for
  // 64-bit elements. Instead, materialize all bits set and then negate that.
  if (VT == MVT::f64 || VT == MVT::v2f64) {
    SignMaskV = DAG.getConstant(APInt::getAllOnes(BitWidth), DL, VecVT);
    SignMaskV = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, SignMaskV);
    SignMaskV = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, SignMaskV);
    SignMaskV = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, SignMaskV);
  }

  SDValue BSP =
      DAG.getNode(AArch64ISD::BSP, DL, VecVT, SignMaskV, VecVal1, VecVal2);
  if (VT == MVT::f16 || VT == MVT::bf16)
    return DAG.getTargetExtractSubreg(AArch64::hsub, DL, VT, BSP);
  if (VT == MVT::f32)
    return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, BSP);
  if (VT == MVT::f64)
    return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, BSP);

  return BitCast(VT, BSP, DAG);
}

SDValue AArch64TargetLowering::LowerCTPOP_PARITY(SDValue Op,
                                                 SelectionDAG &DAG) const {
  if (DAG.getMachineFunction().getFunction().hasFnAttribute(
          Attribute::NoImplicitFloat))
    return SDValue();

  EVT VT = Op.getValueType();
  if (VT.isScalableVector() ||
      useSVEForFixedLengthVectorVT(VT, /*OverrideNEON=*/true))
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::CTPOP_MERGE_PASSTHRU);

  bool IsParity = Op.getOpcode() == ISD::PARITY;
  SDValue Val = Op.getOperand(0);
  SDLoc DL(Op);

  // for i32, general parity function using EORs is more efficient compared to
  // using floating point
  if (VT == MVT::i32 && IsParity)
    return SDValue();

  if (Subtarget->isSVEorStreamingSVEAvailable()) {
    if (VT == MVT::i32 || VT == MVT::i64) {
      EVT ContainerVT = VT == MVT::i32 ? MVT::nxv4i32 : MVT::nxv2i64;
      Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ContainerVT,
                        DAG.getUNDEF(ContainerVT), Val,
                        DAG.getVectorIdxConstant(0, DL));
      Val = DAG.getNode(ISD::CTPOP, DL, ContainerVT, Val);
      Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Val,
                        DAG.getVectorIdxConstant(0, DL));
      if (IsParity)
        Val = DAG.getNode(ISD::AND, DL, VT, Val, DAG.getConstant(1, DL, VT));
      return Val;
    }

    if (VT == MVT::i128) {
      Val = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Val);
      Val = convertToScalableVector(DAG, MVT::nxv2i64, Val);
      Val = DAG.getNode(ISD::CTPOP, DL, MVT::nxv2i64, Val);
      Val = convertFromScalableVector(DAG, MVT::v2i64, Val);
      Val = DAG.getNode(ISD::VECREDUCE_ADD, DL, MVT::i64, Val);
      Val = DAG.getZExtOrTrunc(Val, DL, VT);
      if (IsParity)
        Val = DAG.getNode(ISD::AND, DL, VT, Val, DAG.getConstant(1, DL, VT));
      return Val;
    }
  }

  if (!Subtarget->isNeonAvailable())
    return SDValue();

  // If there is no CNT instruction available, GPR popcount 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
  //  FMOV    X0, D0        // copy result back to integer reg
  if (VT == MVT::i32 || VT == MVT::i64) {
    if (VT == MVT::i32)
      Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
    Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);

    SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
    SDValue AddV = DAG.getNode(AArch64ISD::UADDV, DL, MVT::v8i8, CtPop);
    AddV = DAG.getNode(AArch64ISD::NVCAST, DL,
                       VT == MVT::i32 ? MVT::v2i32 : MVT::v1i64, AddV);
    AddV = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, AddV,
                       DAG.getConstant(0, DL, MVT::i64));
    if (IsParity)
      AddV = DAG.getNode(ISD::AND, DL, VT, AddV, DAG.getConstant(1, DL, VT));
    return AddV;
  } else if (VT == MVT::i128) {
    Val = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Val);

    SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v16i8, Val);
    SDValue AddV = DAG.getNode(AArch64ISD::UADDV, DL, MVT::v16i8, CtPop);
    AddV = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64,
                       DAG.getNode(AArch64ISD::NVCAST, DL, MVT::v2i64, AddV),
                       DAG.getConstant(0, DL, MVT::i64));
    AddV = DAG.getZExtOrTrunc(AddV, DL, VT);
    if (IsParity)
      AddV = DAG.getNode(ISD::AND, DL, VT, AddV, DAG.getConstant(1, DL, VT));
    return AddV;
  }

  assert(!IsParity && "ISD::PARITY of vector types not supported");

  assert((VT == MVT::v1i64 || VT == MVT::v2i64 || VT == MVT::v2i32 ||
          VT == MVT::v4i32 || VT == MVT::v4i16 || VT == MVT::v8i16) &&
         "Unexpected type for custom ctpop lowering");

  EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
  Val = DAG.getBitcast(VT8Bit, Val);
  Val = DAG.getNode(ISD::CTPOP, DL, VT8Bit, Val);

  if (Subtarget->hasDotProd() && VT.getScalarSizeInBits() != 16 &&
      VT.getVectorNumElements() >= 2) {
    EVT DT = VT == MVT::v2i64 ? MVT::v4i32 : VT;
    SDValue Zeros = DAG.getConstant(0, DL, DT);
    SDValue Ones = DAG.getConstant(1, DL, VT8Bit);

    if (VT == MVT::v2i64) {
      Val = DAG.getNode(AArch64ISD::UDOT, DL, DT, Zeros, Ones, Val);
      Val = DAG.getNode(AArch64ISD::UADDLP, DL, VT, Val);
    } else if (VT == MVT::v2i32) {
      Val = DAG.getNode(AArch64ISD::UDOT, DL, DT, Zeros, Ones, Val);
    } else if (VT == MVT::v4i32) {
      Val = DAG.getNode(AArch64ISD::UDOT, DL, DT, Zeros, Ones, Val);
    } else {
      llvm_unreachable("Unexpected type for custom ctpop lowering");
    }

    return Val;
  }

  // Widen v8i8/v16i8 CTPOP result to VT by repeatedly widening pairwise adds.
  unsigned EltSize = 8;
  unsigned NumElts = VT.is64BitVector() ? 8 : 16;
  while (EltSize != VT.getScalarSizeInBits()) {
    EltSize *= 2;
    NumElts /= 2;
    MVT WidenVT = MVT::getVectorVT(MVT::getIntegerVT(EltSize), NumElts);
    Val = DAG.getNode(AArch64ISD::UADDLP, DL, WidenVT, Val);
  }

  return Val;
}

SDValue AArch64TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isScalableVector() ||
         useSVEForFixedLengthVectorVT(
             VT, /*OverrideNEON=*/Subtarget->useSVEForFixedLengthVectors()));

  SDLoc DL(Op);
  SDValue RBIT = DAG.getNode(ISD::BITREVERSE, DL, VT, Op.getOperand(0));
  return DAG.getNode(ISD::CTLZ, DL, VT, RBIT);
}

SDValue AArch64TargetLowering::LowerMinMax(SDValue Op,
                                           SelectionDAG &DAG) const {

  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  unsigned Opcode = Op.getOpcode();
  ISD::CondCode CC;
  switch (Opcode) {
  default:
    llvm_unreachable("Wrong instruction");
  case ISD::SMAX:
    CC = ISD::SETGT;
    break;
  case ISD::SMIN:
    CC = ISD::SETLT;
    break;
  case ISD::UMAX:
    CC = ISD::SETUGT;
    break;
  case ISD::UMIN:
    CC = ISD::SETULT;
    break;
  }

  if (VT.isScalableVector() ||
      useSVEForFixedLengthVectorVT(
          VT, /*OverrideNEON=*/Subtarget->useSVEForFixedLengthVectors())) {
    switch (Opcode) {
    default:
      llvm_unreachable("Wrong instruction");
    case ISD::SMAX:
      return LowerToPredicatedOp(Op, DAG, AArch64ISD::SMAX_PRED);
    case ISD::SMIN:
      return LowerToPredicatedOp(Op, DAG, AArch64ISD::SMIN_PRED);
    case ISD::UMAX:
      return LowerToPredicatedOp(Op, DAG, AArch64ISD::UMAX_PRED);
    case ISD::UMIN:
      return LowerToPredicatedOp(Op, DAG, AArch64ISD::UMIN_PRED);
    }
  }

  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  SDValue Cond = DAG.getSetCC(DL, VT, Op0, Op1, CC);
  return DAG.getSelect(DL, VT, Cond, Op0, Op1);
}

SDValue AArch64TargetLowering::LowerBitreverse(SDValue Op,
                                               SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();

  if (VT.isScalableVector() ||
      useSVEForFixedLengthVectorVT(
          VT, /*OverrideNEON=*/Subtarget->useSVEForFixedLengthVectors()))
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::BITREVERSE_MERGE_PASSTHRU);

  SDLoc DL(Op);
  SDValue REVB;
  MVT VST;

  switch (VT.getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("Invalid type for bitreverse!");

  case MVT::v2i32: {
    VST = MVT::v8i8;
    REVB = DAG.getNode(AArch64ISD::REV32, DL, VST, Op.getOperand(0));

    break;
  }

  case MVT::v4i32: {
    VST = MVT::v16i8;
    REVB = DAG.getNode(AArch64ISD::REV32, DL, VST, Op.getOperand(0));

    break;
  }

  case MVT::v1i64: {
    VST = MVT::v8i8;
    REVB = DAG.getNode(AArch64ISD::REV64, DL, VST, Op.getOperand(0));

    break;
  }

  case MVT::v2i64: {
    VST = MVT::v16i8;
    REVB = DAG.getNode(AArch64ISD::REV64, DL, VST, Op.getOperand(0));

    break;
  }
  }

  return DAG.getNode(AArch64ISD::NVCAST, DL, VT,
                     DAG.getNode(ISD::BITREVERSE, DL, VST, REVB));
}

// Check whether the continuous comparison sequence.
static bool
isOrXorChain(SDValue N, unsigned &Num,
             SmallVector<std::pair<SDValue, SDValue>, 16> &WorkList) {
  if (Num == MaxXors)
    return false;

  // Skip the one-use zext
  if (N->getOpcode() == ISD::ZERO_EXTEND && N->hasOneUse())
    N = N->getOperand(0);

  // The leaf node must be XOR
  if (N->getOpcode() == ISD::XOR) {
    WorkList.push_back(std::make_pair(N->getOperand(0), N->getOperand(1)));
    Num++;
    return true;
  }

  // All the non-leaf nodes must be OR.
  if (N->getOpcode() != ISD::OR || !N->hasOneUse())
    return false;

  if (isOrXorChain(N->getOperand(0), Num, WorkList) &&
      isOrXorChain(N->getOperand(1), Num, WorkList))
    return true;
  return false;
}

// Transform chains of ORs and XORs, which usually outlined by memcmp/bmp.
static SDValue performOrXorChainCombine(SDNode *N, SelectionDAG &DAG) {
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SmallVector<std::pair<SDValue, SDValue>, 16> WorkList;

  // Only handle integer compares.
  if (N->getOpcode() != ISD::SETCC)
    return SDValue();

  ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(2))->get();
  // Try to express conjunction "cmp 0 (or (xor A0 A1) (xor B0 B1))" as:
  // sub A0, A1; ccmp B0, B1, 0, eq; cmp inv(Cond) flag
  unsigned NumXors = 0;
  if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && isNullConstant(RHS) &&
      LHS->getOpcode() == ISD::OR && LHS->hasOneUse() &&
      isOrXorChain(LHS, NumXors, WorkList)) {
    SDValue XOR0, XOR1;
    std::tie(XOR0, XOR1) = WorkList[0];
    unsigned LogicOp = (Cond == ISD::SETEQ) ? ISD::AND : ISD::OR;
    SDValue Cmp = DAG.getSetCC(DL, VT, XOR0, XOR1, Cond);
    for (unsigned I = 1; I < WorkList.size(); I++) {
      std::tie(XOR0, XOR1) = WorkList[I];
      SDValue CmpChain = DAG.getSetCC(DL, VT, XOR0, XOR1, Cond);
      Cmp = DAG.getNode(LogicOp, DL, VT, Cmp, CmpChain);
    }

    // Exit early by inverting the condition, which help reduce indentations.
    return Cmp;
  }

  return SDValue();
}

SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {

  if (Op.getValueType().isVector())
    return LowerVSETCC(Op, DAG);

  bool IsStrict = Op->isStrictFPOpcode();
  bool IsSignaling = Op.getOpcode() == ISD::STRICT_FSETCCS;
  unsigned OpNo = IsStrict ? 1 : 0;
  SDValue Chain;
  if (IsStrict)
    Chain = Op.getOperand(0);
  SDValue LHS = Op.getOperand(OpNo + 0);
  SDValue RHS = Op.getOperand(OpNo + 1);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(OpNo + 2))->get();
  SDLoc DL(Op);

  // We chose ZeroOrOneBooleanContents, so use zero and one.
  EVT VT = Op.getValueType();
  SDValue TVal = DAG.getConstant(1, DL, VT);
  SDValue FVal = DAG.getConstant(0, DL, VT);

  // Handle f128 first, since one possible outcome is a normal integer
  // comparison which gets picked up by the next if statement.
  if (LHS.getValueType() == MVT::f128) {
    softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, DL, LHS, RHS, Chain,
                        IsSignaling);

    // If softenSetCCOperands returned a scalar, use it.
    if (!RHS.getNode()) {
      assert(LHS.getValueType() == Op.getValueType() &&
             "Unexpected setcc expansion!");
      return IsStrict ? DAG.getMergeValues({LHS, Chain}, DL) : LHS;
    }
  }

  if (LHS.getValueType().isInteger()) {

    simplifySetCCIntoEq(CC, LHS, RHS, DAG, DL);

    SDValue CCVal;
    SDValue Cmp = getAArch64Cmp(
        LHS, RHS, ISD::getSetCCInverse(CC, LHS.getValueType()), CCVal, DAG, DL);

    // Note that we inverted the condition above, so we reverse the order of
    // the true and false operands here.  This will allow the setcc to be
    // matched to a single CSINC instruction.
    SDValue Res = DAG.getNode(AArch64ISD::CSEL, DL, VT, FVal, TVal, CCVal, Cmp);
    return IsStrict ? DAG.getMergeValues({Res, Chain}, DL) : Res;
  }

  // Now we know we're dealing with FP values.
  assert(LHS.getValueType() == MVT::bf16 || LHS.getValueType() == MVT::f16 ||
         LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);

  // If that fails, we'll need to perform an FCMP + CSEL sequence.  Go ahead
  // and do the comparison.
  SDValue Cmp;
  if (IsStrict)
    Cmp = emitStrictFPComparison(LHS, RHS, DL, DAG, Chain, IsSignaling);
  else
    Cmp = emitComparison(LHS, RHS, CC, DL, DAG);

  AArch64CC::CondCode CC1, CC2;
  changeFPCCToAArch64CC(CC, CC1, CC2);
  SDValue Res;
  if (CC2 == AArch64CC::AL) {
    changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, LHS.getValueType()), CC1,
                          CC2);
    SDValue CC1Val = DAG.getConstant(CC1, DL, MVT::i32);

    // Note that we inverted the condition above, so we reverse the order of
    // the true and false operands here.  This will allow the setcc to be
    // matched to a single CSINC instruction.
    Res = DAG.getNode(AArch64ISD::CSEL, DL, VT, FVal, TVal, CC1Val, Cmp);
  } else {
    // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
    // totally clean.  Some of them require two CSELs to implement.  As is in
    // this case, we emit the first CSEL and then emit a second using the output
    // of the first as the RHS.  We're effectively OR'ing the two CC's together.

    // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
    SDValue CC1Val = DAG.getConstant(CC1, DL, MVT::i32);
    SDValue CS1 =
        DAG.getNode(AArch64ISD::CSEL, DL, VT, TVal, FVal, CC1Val, Cmp);

    SDValue CC2Val = DAG.getConstant(CC2, DL, MVT::i32);
    Res = DAG.getNode(AArch64ISD::CSEL, DL, VT, TVal, CS1, CC2Val, Cmp);
  }
  return IsStrict ? DAG.getMergeValues({Res, Cmp.getValue(1)}, DL) : Res;
}

SDValue AArch64TargetLowering::LowerSETCCCARRY(SDValue Op,
                                               SelectionDAG &DAG) const {

  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  EVT VT = LHS.getValueType();
  if (VT != MVT::i32 && VT != MVT::i64)
    return SDValue();

  SDLoc DL(Op);
  SDValue Carry = Op.getOperand(2);
  // SBCS uses a carry not a borrow so the carry flag should be inverted first.
  SDValue InvCarry = valueToCarryFlag(Carry, DAG, true);
  SDValue Cmp = DAG.getNode(AArch64ISD::SBCS, DL, DAG.getVTList(VT, MVT::Glue),
                            LHS, RHS, InvCarry);

  EVT OpVT = Op.getValueType();
  SDValue TVal = DAG.getConstant(1, DL, OpVT);
  SDValue FVal = DAG.getConstant(0, DL, OpVT);

  ISD::CondCode Cond = cast<CondCodeSDNode>(Op.getOperand(3))->get();
  ISD::CondCode CondInv = ISD::getSetCCInverse(Cond, VT);
  SDValue CCVal =
      DAG.getConstant(changeIntCCToAArch64CC(CondInv), DL, MVT::i32);
  // Inputs are swapped because the condition is inverted. This will allow
  // matching with a single CSINC instruction.
  return DAG.getNode(AArch64ISD::CSEL, DL, OpVT, FVal, TVal, CCVal,
                     Cmp.getValue(1));
}

/// Emit vector comparison for floating-point values, producing a mask.
static SDValue emitVectorComparison(SDValue LHS, SDValue RHS,
                                    AArch64CC::CondCode CC, bool NoNans, EVT VT,
                                    const SDLoc &DL, SelectionDAG &DAG) {
  assert(VT.getSizeInBits() == LHS.getValueType().getSizeInBits() &&
         "function only supposed to emit natural comparisons");

  switch (CC) {
  default:
    return SDValue();
  case AArch64CC::NE: {
    SDValue Fcmeq = DAG.getNode(AArch64ISD::FCMEQ, DL, VT, LHS, RHS);
    // Use vector semantics for the inversion to potentially save a copy between
    // SIMD and regular registers.
    if (!LHS.getValueType().isVector()) {
      EVT VecVT =
          EVT::getVectorVT(*DAG.getContext(), VT, 128 / VT.getSizeInBits());
      SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
      SDValue MaskVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecVT,
                                    DAG.getUNDEF(VecVT), Fcmeq, Zero);
      SDValue InvertedMask = DAG.getNOT(DL, MaskVec, VecVT);
      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, InvertedMask, Zero);
    }
    return DAG.getNOT(DL, Fcmeq, VT);
  }
  case AArch64CC::EQ:
    return DAG.getNode(AArch64ISD::FCMEQ, DL, VT, LHS, RHS);
  case AArch64CC::GE:
    return DAG.getNode(AArch64ISD::FCMGE, DL, VT, LHS, RHS);
  case AArch64CC::GT:
    return DAG.getNode(AArch64ISD::FCMGT, DL, VT, LHS, RHS);
  case AArch64CC::LE:
    if (!NoNans)
      return SDValue();
    // If we ignore NaNs then we can use to the LS implementation.
    [[fallthrough]];
  case AArch64CC::LS:
    return DAG.getNode(AArch64ISD::FCMGE, DL, VT, RHS, LHS);
  case AArch64CC::LT:
    if (!NoNans)
      return SDValue();
    // If we ignore NaNs then we can use to the MI implementation.
    [[fallthrough]];
  case AArch64CC::MI:
    return DAG.getNode(AArch64ISD::FCMGT, DL, VT, RHS, LHS);
  }
}

/// For SELECT_CC, when the true/false values are (-1, 0) and the compared
/// values are scalars, try to emit a mask generating vector instruction.
static SDValue emitFloatCompareMask(SDValue LHS, SDValue RHS, SDValue TVal,
                                    SDValue FVal, ISD::CondCode CC, bool NoNaNs,
                                    const SDLoc &DL, SelectionDAG &DAG) {
  assert(!LHS.getValueType().isVector());
  assert(!RHS.getValueType().isVector());

  auto *CTVal = dyn_cast<ConstantSDNode>(TVal);
  auto *CFVal = dyn_cast<ConstantSDNode>(FVal);
  if (!CTVal || !CFVal)
    return {};
  if (!(CTVal->isAllOnes() && CFVal->isZero()) &&
      !(CTVal->isZero() && CFVal->isAllOnes()))
    return {};

  if (CTVal->isZero())
    CC = ISD::getSetCCInverse(CC, LHS.getValueType());

  EVT VT = TVal.getValueType();
  if (VT.getSizeInBits() != LHS.getValueType().getSizeInBits())
    return {};

  if (!NoNaNs && (CC == ISD::SETUO || CC == ISD::SETO)) {
    bool OneNaN = false;
    if (LHS == RHS) {
      OneNaN = true;
    } else if (DAG.isKnownNeverNaN(RHS)) {
      OneNaN = true;
      RHS = LHS;
    } else if (DAG.isKnownNeverNaN(LHS)) {
      OneNaN = true;
      LHS = RHS;
    }
    if (OneNaN)
      CC = (CC == ISD::SETUO) ? ISD::SETUNE : ISD::SETOEQ;
  }

  AArch64CC::CondCode CC1;
  AArch64CC::CondCode CC2;
  bool ShouldInvert = false;
  changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);
  SDValue Cmp = emitVectorComparison(LHS, RHS, CC1, NoNaNs, VT, DL, DAG);
  SDValue Cmp2;
  if (CC2 != AArch64CC::AL) {
    Cmp2 = emitVectorComparison(LHS, RHS, CC2, NoNaNs, VT, DL, DAG);
    if (!Cmp2)
      return {};
  }
  if (!Cmp2 && !ShouldInvert)
    return Cmp;

  EVT VecVT = EVT::getVectorVT(*DAG.getContext(), VT, 128 / VT.getSizeInBits());
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  Cmp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecVT, DAG.getUNDEF(VecVT), Cmp,
                    Zero);
  if (Cmp2) {
    Cmp2 = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VecVT, DAG.getUNDEF(VecVT),
                       Cmp2, Zero);
    Cmp = DAG.getNode(ISD::OR, DL, VecVT, Cmp, Cmp2);
  }
  if (ShouldInvert)
    Cmp = DAG.getNOT(DL, Cmp, VecVT);
  Cmp = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Cmp, Zero);
  return Cmp;
}

SDValue AArch64TargetLowering::LowerSELECT_CC(
    ISD::CondCode CC, SDValue LHS, SDValue RHS, SDValue TVal, SDValue FVal,
    iterator_range<SDNode::user_iterator> Users, bool HasNoNaNs,
    const SDLoc &DL, SelectionDAG &DAG) const {
  // Handle f128 first, because it will result in a comparison of some RTLIB
  // call result against zero.
  if (LHS.getValueType() == MVT::f128) {
    softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, DL, LHS, RHS);

    // If softenSetCCOperands returned a scalar, we need to compare the result
    // against zero to select between true and false values.
    if (!RHS.getNode()) {
      RHS = DAG.getConstant(0, DL, LHS.getValueType());
      CC = ISD::SETNE;
    }
  }

  // Also handle f16, for which we need to do a f32 comparison.
  if ((LHS.getValueType() == MVT::f16 && !Subtarget->hasFullFP16()) ||
      LHS.getValueType() == MVT::bf16) {
    LHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, LHS);
    RHS = DAG.getNode(ISD::FP_EXTEND, DL, MVT::f32, RHS);
  }

  // Next, handle integers.
  if (LHS.getValueType().isInteger()) {
    assert((LHS.getValueType() == RHS.getValueType()) &&
           (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));

    ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
    ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);
    ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
    // Check for sign pattern (SELECT_CC setgt, iN lhs, -1, 1, -1) and transform
    // into (OR (ASR lhs, N-1), 1), which requires less instructions for the
    // supported types.
    if (CC == ISD::SETGT && RHSC && RHSC->isAllOnes() && CTVal && CFVal &&
        CTVal->isOne() && CFVal->isAllOnes() &&
        LHS.getValueType() == TVal.getValueType()) {
      EVT VT = LHS.getValueType();
      SDValue Shift =
          DAG.getNode(ISD::SRA, DL, VT, LHS,
                      DAG.getConstant(VT.getSizeInBits() - 1, DL, VT));
      return DAG.getNode(ISD::OR, DL, VT, Shift, DAG.getConstant(1, DL, VT));
    }

    // Check for SMAX(lhs, 0) and SMIN(lhs, 0) patterns.
    // (SELECT_CC setgt, lhs, 0, lhs, 0) -> (BIC lhs, (SRA lhs, typesize-1))
    // (SELECT_CC setlt, lhs, 0, lhs, 0) -> (AND lhs, (SRA lhs, typesize-1))
    // Both require less instructions than compare and conditional select.
    if ((CC == ISD::SETGT || CC == ISD::SETLT) && LHS == TVal &&
        RHSC && RHSC->isZero() && CFVal && CFVal->isZero() &&
        LHS.getValueType() == RHS.getValueType()) {
      EVT VT = LHS.getValueType();
      SDValue Shift =
          DAG.getNode(ISD::SRA, DL, VT, LHS,
                      DAG.getConstant(VT.getSizeInBits() - 1, DL, VT));

      if (CC == ISD::SETGT)
        Shift = DAG.getNOT(DL, Shift, VT);

      return DAG.getNode(ISD::AND, DL, VT, LHS, Shift);
    }

    unsigned Opcode = AArch64ISD::CSEL;

    // If both the TVal and the FVal are constants, see if we can swap them in
    // order to for a CSINV or CSINC out of them.
    if (CTVal && CFVal && CTVal->isAllOnes() && CFVal->isZero()) {
      std::swap(TVal, FVal);
      std::swap(CTVal, CFVal);
      CC = ISD::getSetCCInverse(CC, LHS.getValueType());
    } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isZero()) {
      std::swap(TVal, FVal);
      std::swap(CTVal, CFVal);
      CC = ISD::getSetCCInverse(CC, LHS.getValueType());
    } else if (TVal.getOpcode() == ISD::XOR) {
      // If TVal is a NOT we want to swap TVal and FVal so that we can match
      // with a CSINV rather than a CSEL.
      if (isAllOnesConstant(TVal.getOperand(1))) {
        std::swap(TVal, FVal);
        std::swap(CTVal, CFVal);
        CC = ISD::getSetCCInverse(CC, LHS.getValueType());
      }
    } else if (TVal.getOpcode() == ISD::SUB) {
      // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
      // that we can match with a CSNEG rather than a CSEL.
      if (isNullConstant(TVal.getOperand(0))) {
        std::swap(TVal, FVal);
        std::swap(CTVal, CFVal);
        CC = ISD::getSetCCInverse(CC, LHS.getValueType());
      }
    } else if (CTVal && CFVal) {
      const int64_t TrueVal = CTVal->getSExtValue();
      const int64_t FalseVal = CFVal->getSExtValue();
      bool Swap = false;

      // If both TVal and FVal are constants, see if FVal is the
      // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
      // instead of a CSEL in that case.
      if (TrueVal == ~FalseVal) {
        Opcode = AArch64ISD::CSINV;
      } else if (FalseVal > std::numeric_limits<int64_t>::min() &&
                 TrueVal == -FalseVal) {
        Opcode = AArch64ISD::CSNEG;
      } else if (TVal.getValueType() == MVT::i32) {
        // If our operands are only 32-bit wide, make sure we use 32-bit
        // arithmetic for the check whether we can use CSINC. This ensures that
        // the addition in the check will wrap around properly in case there is
        // an overflow (which would not be the case if we do the check with
        // 64-bit arithmetic).
        const uint32_t TrueVal32 = CTVal->getZExtValue();
        const uint32_t FalseVal32 = CFVal->getZExtValue();

        if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
          Opcode = AArch64ISD::CSINC;

          if (TrueVal32 > FalseVal32) {
            Swap = true;
          }
        }
      } else {
        // 64-bit check whether we can use CSINC.
        const uint64_t TrueVal64 = TrueVal;
        const uint64_t FalseVal64 = FalseVal;

        if ((TrueVal64 == FalseVal64 + 1) || (TrueVal64 + 1 == FalseVal64)) {
          Opcode = AArch64ISD::CSINC;

          if (TrueVal > FalseVal) {
            Swap = true;
          }
        }
      }

      // Swap TVal and FVal if necessary.
      if (Swap) {
        std::swap(TVal, FVal);
        std::swap(CTVal, CFVal);
        CC = ISD::getSetCCInverse(CC, LHS.getValueType());
      }

      if (Opcode != AArch64ISD::CSEL) {
        // Drop FVal since we can get its value by simply inverting/negating
        // TVal.
        FVal = TVal;
      }
    }

    // Avoid materializing a constant when possible by reusing a known value in
    // a register.  However, don't perform this optimization if the known value
    // is one, zero or negative one in the case of a CSEL.  We can always
    // materialize these values using CSINC, CSEL and CSINV with wzr/xzr as the
    // FVal, respectively.
    ConstantSDNode *RHSVal = dyn_cast<ConstantSDNode>(RHS);
    if (Opcode == AArch64ISD::CSEL && RHSVal && !RHSVal->isOne() &&
        !RHSVal->isZero() && !RHSVal->isAllOnes()) {
      AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
      // Transform "a == C ? C : x" to "a == C ? a : x" and "a != C ? x : C" to
      // "a != C ? x : a" to avoid materializing C.
      if (CTVal && CTVal == RHSVal && AArch64CC == AArch64CC::EQ)
        TVal = LHS;
      else if (CFVal && CFVal == RHSVal && AArch64CC == AArch64CC::NE)
        FVal = LHS;
    } else if (Opcode == AArch64ISD::CSNEG && RHSVal && RHSVal->isOne()) {
      assert (CTVal && CFVal && "Expected constant operands for CSNEG.");
      // Use a CSINV to transform "a == C ? 1 : -1" to "a == C ? a : -1" to
      // avoid materializing C.
      AArch64CC::CondCode AArch64CC = changeIntCCToAArch64CC(CC);
      if (CTVal == RHSVal && AArch64CC == AArch64CC::EQ) {
        Opcode = AArch64ISD::CSINV;
        TVal = LHS;
        FVal = DAG.getConstant(0, DL, FVal.getValueType());
      }
    }

    SDValue CCVal;
    SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, DL);
    EVT VT = TVal.getValueType();
    return DAG.getNode(Opcode, DL, VT, TVal, FVal, CCVal, Cmp);
  }

  // Now we know we're dealing with FP values.
  assert(LHS.getValueType() == MVT::f16 || LHS.getValueType() == MVT::f32 ||
         LHS.getValueType() == MVT::f64);
  assert(LHS.getValueType() == RHS.getValueType());
  EVT VT = TVal.getValueType();

  // If the purpose of the comparison is to select between all ones
  // or all zeros, try to use a vector comparison because the operands are
  // already stored in SIMD registers.
  if (Subtarget->isNeonAvailable() && all_of(Users, [](const SDNode *U) {
        switch (U->getOpcode()) {
        default:
          return false;
        case ISD::INSERT_VECTOR_ELT:
        case ISD::SCALAR_TO_VECTOR:
        case AArch64ISD::DUP:
          return true;
        }
      })) {
    bool NoNaNs = getTargetMachine().Options.NoNaNsFPMath || HasNoNaNs;
    SDValue VectorCmp =
        emitFloatCompareMask(LHS, RHS, TVal, FVal, CC, NoNaNs, DL, DAG);
    if (VectorCmp)
      return VectorCmp;
  }

  SDValue Cmp = emitComparison(LHS, RHS, CC, DL, DAG);

  // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
  // clean.  Some of them require two CSELs to implement.
  AArch64CC::CondCode CC1, CC2;
  changeFPCCToAArch64CC(CC, CC1, CC2);

  if (DAG.getTarget().Options.UnsafeFPMath) {
    // Transform "a == 0.0 ? 0.0 : x" to "a == 0.0 ? a : x" and
    // "a != 0.0 ? x : 0.0" to "a != 0.0 ? x : a" to avoid materializing 0.0.
    ConstantFPSDNode *RHSVal = dyn_cast<ConstantFPSDNode>(RHS);
    if (RHSVal && RHSVal->isZero()) {
      ConstantFPSDNode *CFVal = dyn_cast<ConstantFPSDNode>(FVal);
      ConstantFPSDNode *CTVal = dyn_cast<ConstantFPSDNode>(TVal);

      if ((CC == ISD::SETEQ || CC == ISD::SETOEQ || CC == ISD::SETUEQ) &&
          CTVal && CTVal->isZero() && TVal.getValueType() == LHS.getValueType())
        TVal = LHS;
      else if ((CC == ISD::SETNE || CC == ISD::SETONE || CC == ISD::SETUNE) &&
               CFVal && CFVal->isZero() &&
               FVal.getValueType() == LHS.getValueType())
        FVal = LHS;
    }
  }

  // Emit first, and possibly only, CSEL.
  SDValue CC1Val = DAG.getConstant(CC1, DL, MVT::i32);
  SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, DL, VT, TVal, FVal, CC1Val, Cmp);

  // If we need a second CSEL, emit it, using the output of the first as the
  // RHS.  We're effectively OR'ing the two CC's together.
  if (CC2 != AArch64CC::AL) {
    SDValue CC2Val = DAG.getConstant(CC2, DL, MVT::i32);
    return DAG.getNode(AArch64ISD::CSEL, DL, VT, TVal, CS1, CC2Val, Cmp);
  }

  // Otherwise, return the output of the first CSEL.
  return CS1;
}

SDValue AArch64TargetLowering::LowerVECTOR_SPLICE(SDValue Op,
                                                  SelectionDAG &DAG) const {
  EVT Ty = Op.getValueType();
  auto Idx = Op.getConstantOperandAPInt(2);
  int64_t IdxVal = Idx.getSExtValue();
  assert(Ty.isScalableVector() &&
         "Only expect scalable vectors for custom lowering of VECTOR_SPLICE");

  // We can use the splice instruction for certain index values where we are
  // able to efficiently generate the correct predicate. The index will be
  // inverted and used directly as the input to the ptrue instruction, i.e.
  // -1 -> vl1, -2 -> vl2, etc. The predicate will then be reversed to get the
  // splice predicate. However, we can only do this if we can guarantee that
  // there are enough elements in the vector, hence we check the index <= min
  // number of elements.
  std::optional<unsigned> PredPattern;
  if (Ty.isScalableVector() && IdxVal < 0 &&
      (PredPattern = getSVEPredPatternFromNumElements(std::abs(IdxVal))) !=
          std::nullopt) {
    SDLoc DL(Op);

    // Create a predicate where all but the last -IdxVal elements are false.
    EVT PredVT = Ty.changeVectorElementType(MVT::i1);
    SDValue Pred = getPTrue(DAG, DL, PredVT, *PredPattern);
    Pred = DAG.getNode(ISD::VECTOR_REVERSE, DL, PredVT, Pred);

    // Now splice the two inputs together using the predicate.
    return DAG.getNode(AArch64ISD::SPLICE, DL, Ty, Pred, Op.getOperand(0),
                       Op.getOperand(1));
  }

  // We can select to an EXT instruction when indexing the first 256 bytes.
  unsigned BlockSize = AArch64::SVEBitsPerBlock / Ty.getVectorMinNumElements();
  if (IdxVal >= 0 && (IdxVal * BlockSize / 8) < 256)
    return Op;

  return SDValue();
}

SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
                                              SelectionDAG &DAG) const {
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  SDValue TVal = Op.getOperand(2);
  SDValue FVal = Op.getOperand(3);
  bool HasNoNans = Op->getFlags().hasNoNaNs();
  SDLoc DL(Op);
  return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, Op->users(), HasNoNans, DL,
                        DAG);
}

SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
                                           SelectionDAG &DAG) const {
  SDValue CCVal = Op->getOperand(0);
  SDValue TVal = Op->getOperand(1);
  SDValue FVal = Op->getOperand(2);
  bool HasNoNans = Op->getFlags().hasNoNaNs();
  SDLoc DL(Op);

  EVT Ty = Op.getValueType();
  if (Ty == MVT::aarch64svcount) {
    TVal = DAG.getNode(ISD::BITCAST, DL, MVT::nxv16i1, TVal);
    FVal = DAG.getNode(ISD::BITCAST, DL, MVT::nxv16i1, FVal);
    SDValue Sel =
        DAG.getNode(ISD::SELECT, DL, MVT::nxv16i1, CCVal, TVal, FVal);
    return DAG.getNode(ISD::BITCAST, DL, Ty, Sel);
  }

  if (Ty.isScalableVector()) {
    MVT PredVT = MVT::getVectorVT(MVT::i1, Ty.getVectorElementCount());
    SDValue SplatPred = DAG.getNode(ISD::SPLAT_VECTOR, DL, PredVT, CCVal);
    return DAG.getNode(ISD::VSELECT, DL, Ty, SplatPred, TVal, FVal);
  }

  if (useSVEForFixedLengthVectorVT(Ty, !Subtarget->isNeonAvailable())) {
    // FIXME: Ideally this would be the same as above using i1 types, however
    // for the moment we can't deal with fixed i1 vector types properly, so
    // instead extend the predicate to a result type sized integer vector.
    MVT SplatValVT = MVT::getIntegerVT(Ty.getScalarSizeInBits());
    MVT PredVT = MVT::getVectorVT(SplatValVT, Ty.getVectorElementCount());
    SDValue SplatVal = DAG.getSExtOrTrunc(CCVal, DL, SplatValVT);
    SDValue SplatPred = DAG.getNode(ISD::SPLAT_VECTOR, DL, PredVT, SplatVal);
    return DAG.getNode(ISD::VSELECT, DL, Ty, SplatPred, TVal, FVal);
  }

  // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
  // instruction.
  if (ISD::isOverflowIntrOpRes(CCVal)) {
    // Only lower legal XALUO ops.
    if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
      return SDValue();

    AArch64CC::CondCode OFCC;
    SDValue Value, Overflow;
    std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
    SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);

    return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
                       CCVal, Overflow);
  }

  // Lower it the same way as we would lower a SELECT_CC node.
  ISD::CondCode CC;
  SDValue LHS, RHS;
  if (CCVal.getOpcode() == ISD::SETCC) {
    LHS = CCVal.getOperand(0);
    RHS = CCVal.getOperand(1);
    CC = cast<CondCodeSDNode>(CCVal.getOperand(2))->get();
  } else {
    LHS = CCVal;
    RHS = DAG.getConstant(0, DL, CCVal.getValueType());
    CC = ISD::SETNE;
  }

  // If we are lowering a f16 and we do not have fullf16, convert to a f32 in
  // order to use FCSELSrrr
  if ((Ty == MVT::f16 || Ty == MVT::bf16) && !Subtarget->hasFullFP16()) {
    TVal = DAG.getTargetInsertSubreg(AArch64::hsub, DL, MVT::f32,
                                     DAG.getUNDEF(MVT::f32), TVal);
    FVal = DAG.getTargetInsertSubreg(AArch64::hsub, DL, MVT::f32,
                                     DAG.getUNDEF(MVT::f32), FVal);
  }

  SDValue Res =
      LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, Op->users(), HasNoNans, DL, DAG);

  if ((Ty == MVT::f16 || Ty == MVT::bf16) && !Subtarget->hasFullFP16()) {
    return DAG.getTargetExtractSubreg(AArch64::hsub, DL, Ty, Res);
  }

  return Res;
}

SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
                                              SelectionDAG &DAG) const {
  // Jump table entries as PC relative offsets. No additional tweaking
  // is necessary here. Just get the address of the jump table.
  JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);

  CodeModel::Model CM = getTargetMachine().getCodeModel();
  if (CM == CodeModel::Large && !getTargetMachine().isPositionIndependent() &&
      !Subtarget->isTargetMachO())
    return getAddrLarge(JT, DAG);
  if (CM == CodeModel::Tiny)
    return getAddrTiny(JT, DAG);
  return getAddr(JT, DAG);
}

SDValue AArch64TargetLowering::LowerBR_JT(SDValue Op,
                                          SelectionDAG &DAG) const {
  // Jump table entries as PC relative offsets. No additional tweaking
  // is necessary here. Just get the address of the jump table.
  SDLoc DL(Op);
  SDValue JT = Op.getOperand(1);
  SDValue Entry = Op.getOperand(2);
  int JTI = cast<JumpTableSDNode>(JT.getNode())->getIndex();

  auto *AFI = DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
  AFI->setJumpTableEntryInfo(JTI, 4, nullptr);

  // With aarch64-jump-table-hardening, we only expand the jump table dispatch
  // sequence later, to guarantee the integrity of the intermediate values.
  if (DAG.getMachineFunction().getFunction().hasFnAttribute(
          "aarch64-jump-table-hardening")) {
    CodeModel::Model CM = getTargetMachine().getCodeModel();
    if (Subtarget->isTargetMachO()) {
      if (CM != CodeModel::Small && CM != CodeModel::Large)
        report_fatal_error("Unsupported code-model for hardened jump-table");
    } else {
      // Note that COFF support would likely also need JUMP_TABLE_DEBUG_INFO.
      assert(Subtarget->isTargetELF() &&
             "jump table hardening only supported on MachO/ELF");
      if (CM != CodeModel::Small)
        report_fatal_error("Unsupported code-model for hardened jump-table");
    }

    SDValue X16Copy = DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::X16,
                                       Entry, SDValue());
    SDNode *B = DAG.getMachineNode(AArch64::BR_JumpTable, DL, MVT::Other,
                                   DAG.getTargetJumpTable(JTI, MVT::i32),
                                   X16Copy.getValue(0), X16Copy.getValue(1));
    return SDValue(B, 0);
  }

  SDNode *Dest =
      DAG.getMachineNode(AArch64::JumpTableDest32, DL, MVT::i64, MVT::i64, JT,
                         Entry, DAG.getTargetJumpTable(JTI, MVT::i32));
  SDValue JTInfo = DAG.getJumpTableDebugInfo(JTI, Op.getOperand(0), DL);
  return DAG.getNode(ISD::BRIND, DL, MVT::Other, JTInfo, SDValue(Dest, 0));
}

SDValue AArch64TargetLowering::LowerBRIND(SDValue Op, SelectionDAG &DAG) const {
  SDValue Chain = Op.getOperand(0);
  SDValue Dest = Op.getOperand(1);

  // BR_JT is lowered to BRIND, but the later lowering is specific to indirectbr
  // Skip over the jump-table BRINDs, where the destination is JumpTableDest32.
  if (Dest->isMachineOpcode() &&
      Dest->getMachineOpcode() == AArch64::JumpTableDest32)
    return SDValue();

  const MachineFunction &MF = DAG.getMachineFunction();
  std::optional<uint16_t> BADisc =
      Subtarget->getPtrAuthBlockAddressDiscriminatorIfEnabled(MF.getFunction());
  if (!BADisc)
    return SDValue();

  SDLoc DL(Op);

  SDValue Disc = DAG.getTargetConstant(*BADisc, DL, MVT::i64);
  SDValue Key = DAG.getTargetConstant(AArch64PACKey::IA, DL, MVT::i32);
  SDValue AddrDisc = DAG.getRegister(AArch64::XZR, MVT::i64);

  SDNode *BrA = DAG.getMachineNode(AArch64::BRA, DL, MVT::Other,
                                   {Dest, Key, Disc, AddrDisc, Chain});
  return SDValue(BrA, 0);
}

SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
                                                 SelectionDAG &DAG) const {
  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
  CodeModel::Model CM = getTargetMachine().getCodeModel();
  if (CM == CodeModel::Large) {
    // Use the GOT for the large code model on iOS.
    if (Subtarget->isTargetMachO()) {
      return getGOT(CP, DAG);
    }
    if (!getTargetMachine().isPositionIndependent())
      return getAddrLarge(CP, DAG);
  } else if (CM == CodeModel::Tiny) {
    return getAddrTiny(CP, DAG);
  }
  return getAddr(CP, DAG);
}

SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
                                               SelectionDAG &DAG) const {
  BlockAddressSDNode *BAN = cast<BlockAddressSDNode>(Op);
  const BlockAddress *BA = BAN->getBlockAddress();

  if (std::optional<uint16_t> BADisc =
          Subtarget->getPtrAuthBlockAddressDiscriminatorIfEnabled(
              *BA->getFunction())) {
    SDLoc DL(Op);

    // This isn't cheap, but BRIND is rare.
    SDValue TargetBA = DAG.getTargetBlockAddress(BA, BAN->getValueType(0));

    SDValue Disc = DAG.getTargetConstant(*BADisc, DL, MVT::i64);

    SDValue Key = DAG.getTargetConstant(AArch64PACKey::IA, DL, MVT::i32);
    SDValue AddrDisc = DAG.getRegister(AArch64::XZR, MVT::i64);

    SDNode *MOV =
        DAG.getMachineNode(AArch64::MOVaddrPAC, DL, {MVT::Other, MVT::Glue},
                           {TargetBA, Key, AddrDisc, Disc});
    return DAG.getCopyFromReg(SDValue(MOV, 0), DL, AArch64::X16, MVT::i64,
                              SDValue(MOV, 1));
  }

  CodeModel::Model CM = getTargetMachine().getCodeModel();
  if (CM == CodeModel::Large && !Subtarget->isTargetMachO()) {
    if (!getTargetMachine().isPositionIndependent())
      return getAddrLarge(BAN, DAG);
  } else if (CM == CodeModel::Tiny) {
    return getAddrTiny(BAN, DAG);
  }
  return getAddr(BAN, DAG);
}

SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
                                                 SelectionDAG &DAG) const {
  AArch64FunctionInfo *FuncInfo =
      DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();

  SDLoc DL(Op);
  SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
                                 getPointerTy(DAG.getDataLayout()));
  FR = DAG.getZExtOrTrunc(FR, DL, getPointerMemTy(DAG.getDataLayout()));
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
                      MachinePointerInfo(SV));
}

SDValue AArch64TargetLowering::LowerWin64_VASTART(SDValue Op,
                                                  SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();

  SDLoc DL(Op);
  SDValue FR;
  if (Subtarget->isWindowsArm64EC()) {
    // With the Arm64EC ABI, we compute the address of the varargs save area
    // relative to x4. For a normal AArch64->AArch64 call, x4 == sp on entry,
    // but calls from an entry thunk can pass in a different address.
    Register VReg = MF.addLiveIn(AArch64::X4, &AArch64::GPR64RegClass);
    SDValue Val = DAG.getCopyFromReg(DAG.getEntryNode(), DL, VReg, MVT::i64);
    uint64_t StackOffset;
    if (FuncInfo->getVarArgsGPRSize() > 0)
      StackOffset = -(uint64_t)FuncInfo->getVarArgsGPRSize();
    else
      StackOffset = FuncInfo->getVarArgsStackOffset();
    FR = DAG.getNode(ISD::ADD, DL, MVT::i64, Val,
                     DAG.getConstant(StackOffset, DL, MVT::i64));
  } else {
    FR = DAG.getFrameIndex(FuncInfo->getVarArgsGPRSize() > 0
                               ? FuncInfo->getVarArgsGPRIndex()
                               : FuncInfo->getVarArgsStackIndex(),
                           getPointerTy(DAG.getDataLayout()));
  }
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
                      MachinePointerInfo(SV));
}

SDValue AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
                                                  SelectionDAG &DAG) const {
  // The layout of the va_list struct is specified in the AArch64 Procedure Call
  // Standard, section B.3.
  MachineFunction &MF = DAG.getMachineFunction();
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  unsigned PtrSize = Subtarget->isTargetILP32() ? 4 : 8;
  auto PtrMemVT = getPointerMemTy(DAG.getDataLayout());
  auto PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);

  SDValue Chain = Op.getOperand(0);
  SDValue VAList = Op.getOperand(1);
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  SmallVector<SDValue, 4> MemOps;

  // void *__stack at offset 0
  unsigned Offset = 0;
  SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
  Stack = DAG.getZExtOrTrunc(Stack, DL, PtrMemVT);
  MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
                                MachinePointerInfo(SV), Align(PtrSize)));

  // void *__gr_top at offset 8 (4 on ILP32)
  Offset += PtrSize;
  int GPRSize = FuncInfo->getVarArgsGPRSize();
  if (GPRSize > 0) {
    SDValue GRTop, GRTopAddr;

    GRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                            DAG.getConstant(Offset, DL, PtrVT));

    GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
    GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
                        DAG.getSignedConstant(GPRSize, DL, PtrVT));
    GRTop = DAG.getZExtOrTrunc(GRTop, DL, PtrMemVT);

    MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
                                  MachinePointerInfo(SV, Offset),
                                  Align(PtrSize)));
  }

  // void *__vr_top at offset 16 (8 on ILP32)
  Offset += PtrSize;
  int FPRSize = FuncInfo->getVarArgsFPRSize();
  if (FPRSize > 0) {
    SDValue VRTop, VRTopAddr;
    VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                            DAG.getConstant(Offset, DL, PtrVT));

    VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
    VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
                        DAG.getSignedConstant(FPRSize, DL, PtrVT));
    VRTop = DAG.getZExtOrTrunc(VRTop, DL, PtrMemVT);

    MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
                                  MachinePointerInfo(SV, Offset),
                                  Align(PtrSize)));
  }

  // int __gr_offs at offset 24 (12 on ILP32)
  Offset += PtrSize;
  SDValue GROffsAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                                   DAG.getConstant(Offset, DL, PtrVT));
  MemOps.push_back(
      DAG.getStore(Chain, DL, DAG.getSignedConstant(-GPRSize, DL, MVT::i32),
                   GROffsAddr, MachinePointerInfo(SV, Offset), Align(4)));

  // int __vr_offs at offset 28 (16 on ILP32)
  Offset += 4;
  SDValue VROffsAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                                   DAG.getConstant(Offset, DL, PtrVT));
  MemOps.push_back(
      DAG.getStore(Chain, DL, DAG.getSignedConstant(-FPRSize, DL, MVT::i32),
                   VROffsAddr, MachinePointerInfo(SV, Offset), Align(4)));

  return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
}

SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
                                            SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  Function &F = MF.getFunction();

  if (Subtarget->isCallingConvWin64(F.getCallingConv(), F.isVarArg()))
    return LowerWin64_VASTART(Op, DAG);
  else if (Subtarget->isTargetDarwin())
    return LowerDarwin_VASTART(Op, DAG);
  else
    return LowerAAPCS_VASTART(Op, DAG);
}

SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
                                           SelectionDAG &DAG) const {
  // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
  // pointer.
  SDLoc DL(Op);
  unsigned PtrSize = Subtarget->isTargetILP32() ? 4 : 8;
  unsigned VaListSize =
      (Subtarget->isTargetDarwin() || Subtarget->isTargetWindows())
          ? PtrSize
          : Subtarget->isTargetILP32() ? 20 : 32;
  const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
  const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();

  return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1), Op.getOperand(2),
                       DAG.getConstant(VaListSize, DL, MVT::i32),
                       Align(PtrSize), false, false, /*CI=*/nullptr,
                       std::nullopt, MachinePointerInfo(DestSV),
                       MachinePointerInfo(SrcSV));
}

SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
  assert(Subtarget->isTargetDarwin() &&
         "automatic va_arg instruction only works on Darwin");

  const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue Chain = Op.getOperand(0);
  SDValue Addr = Op.getOperand(1);
  MaybeAlign Align(Op.getConstantOperandVal(3));
  unsigned MinSlotSize = Subtarget->isTargetILP32() ? 4 : 8;
  auto PtrVT = getPointerTy(DAG.getDataLayout());
  auto PtrMemVT = getPointerMemTy(DAG.getDataLayout());
  SDValue VAList =
      DAG.getLoad(PtrMemVT, DL, Chain, Addr, MachinePointerInfo(V));
  Chain = VAList.getValue(1);
  VAList = DAG.getZExtOrTrunc(VAList, DL, PtrVT);

  if (VT.isScalableVector())
    report_fatal_error("Passing SVE types to variadic functions is "
                       "currently not supported");

  if (Align && *Align > MinSlotSize) {
    VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                         DAG.getConstant(Align->value() - 1, DL, PtrVT));
    VAList =
        DAG.getNode(ISD::AND, DL, PtrVT, VAList,
                    DAG.getSignedConstant(-(int64_t)Align->value(), DL, PtrVT));
  }

  Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
  unsigned ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);

  // Scalar integer and FP values smaller than 64 bits are implicitly extended
  // up to 64 bits.  At the very least, we have to increase the striding of the
  // vaargs list to match this, and for FP values we need to introduce
  // FP_ROUND nodes as well.
  if (VT.isInteger() && !VT.isVector())
    ArgSize = std::max(ArgSize, MinSlotSize);
  bool NeedFPTrunc = false;
  if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
    ArgSize = 8;
    NeedFPTrunc = true;
  }

  // Increment the pointer, VAList, to the next vaarg
  SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                               DAG.getConstant(ArgSize, DL, PtrVT));
  VANext = DAG.getZExtOrTrunc(VANext, DL, PtrMemVT);

  // Store the incremented VAList to the legalized pointer
  SDValue APStore =
      DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V));

  // Load the actual argument out of the pointer VAList
  if (NeedFPTrunc) {
    // Load the value as an f64.
    SDValue WideFP =
        DAG.getLoad(MVT::f64, DL, APStore, VAList, MachinePointerInfo());
    // Round the value down to an f32.
    SDValue NarrowFP =
        DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
                    DAG.getIntPtrConstant(1, DL, /*isTarget=*/true));
    SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
    // Merge the rounded value with the chain output of the load.
    return DAG.getMergeValues(Ops, DL);
  }

  return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo());
}

SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
                                              SelectionDAG &DAG) const {
  MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
  MFI.setFrameAddressIsTaken(true);

  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  unsigned Depth = Op.getConstantOperandVal(0);
  SDValue FrameAddr =
      DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, MVT::i64);
  while (Depth--)
    FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
                            MachinePointerInfo());

  if (Subtarget->isTargetILP32())
    FrameAddr = DAG.getNode(ISD::AssertZext, DL, MVT::i64, FrameAddr,
                            DAG.getValueType(VT));

  return FrameAddr;
}

SDValue AArch64TargetLowering::LowerSPONENTRY(SDValue Op,
                                              SelectionDAG &DAG) const {
  MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();

  EVT VT = getPointerTy(DAG.getDataLayout());
  int FI = MFI.CreateFixedObject(4, 0, false);
  return DAG.getFrameIndex(FI, VT);
}

#define GET_REGISTER_MATCHER
#include "AArch64GenAsmMatcher.inc"

// FIXME? Maybe this could be a TableGen attribute on some registers and
// this table could be generated automatically from RegInfo.
Register AArch64TargetLowering::
getRegisterByName(const char* RegName, LLT VT, const MachineFunction &MF) const {
  Register Reg = MatchRegisterName(RegName);
  if (AArch64::X1 <= Reg && Reg <= AArch64::X28) {
    const AArch64RegisterInfo *MRI = Subtarget->getRegisterInfo();
    unsigned DwarfRegNum = MRI->getDwarfRegNum(Reg, false);
    if (!Subtarget->isXRegisterReserved(DwarfRegNum) &&
        !MRI->isReservedReg(MF, Reg))
      Reg = Register();
  }
  return Reg;
}

SDValue AArch64TargetLowering::LowerADDROFRETURNADDR(SDValue Op,
                                                     SelectionDAG &DAG) const {
  DAG.getMachineFunction().getFrameInfo().setFrameAddressIsTaken(true);

  EVT VT = Op.getValueType();
  SDLoc DL(Op);

  SDValue FrameAddr =
      DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
  SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));

  return DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset);
}

SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
                                               SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo &MFI = MF.getFrameInfo();
  MFI.setReturnAddressIsTaken(true);

  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  unsigned Depth = Op.getConstantOperandVal(0);
  SDValue ReturnAddress;
  if (Depth) {
    SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
    SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
    ReturnAddress = DAG.getLoad(
        VT, DL, DAG.getEntryNode(),
        DAG.getNode(ISD::ADD, DL, VT, FrameAddr, Offset), MachinePointerInfo());
  } else {
    // Return LR, which contains the return address. Mark it an implicit
    // live-in.
    Register Reg = MF.addLiveIn(AArch64::LR, &AArch64::GPR64RegClass);
    ReturnAddress = DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
  }

  // The XPACLRI instruction assembles to a hint-space instruction before
  // Armv8.3-A therefore this instruction can be safely used for any pre
  // Armv8.3-A architectures. On Armv8.3-A and onwards XPACI is available so use
  // that instead.
  SDNode *St;
  if (Subtarget->hasPAuth()) {
    St = DAG.getMachineNode(AArch64::XPACI, DL, VT, ReturnAddress);
  } else {
    // XPACLRI operates on LR therefore we must move the operand accordingly.
    SDValue Chain =
        DAG.getCopyToReg(DAG.getEntryNode(), DL, AArch64::LR, ReturnAddress);
    St = DAG.getMachineNode(AArch64::XPACLRI, DL, VT, Chain);
  }
  return SDValue(St, 0);
}

/// LowerShiftParts - Lower SHL_PARTS/SRA_PARTS/SRL_PARTS, which returns two
/// i32 values and take a 2 x i32 value to shift plus a shift amount.
SDValue AArch64TargetLowering::LowerShiftParts(SDValue Op,
                                               SelectionDAG &DAG) const {
  SDValue Lo, Hi;
  expandShiftParts(Op.getNode(), Lo, Hi, DAG);
  return DAG.getMergeValues({Lo, Hi}, SDLoc(Op));
}

bool AArch64TargetLowering::isOffsetFoldingLegal(
    const GlobalAddressSDNode *GA) const {
  // Offsets are folded in the DAG combine rather than here so that we can
  // intelligently choose an offset based on the uses.
  return false;
}

bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
                                         bool OptForSize) const {
  bool IsLegal = false;
  // We can materialize #0.0 as fmov $Rd, XZR for 64-bit, 32-bit cases, and
  // 16-bit case when target has full fp16 support.
  // We encode bf16 bit patterns as if they were fp16. This results in very
  // strange looking assembly but should populate the register with appropriate
  // values. Let's say we wanted to encode 0xR3FC0 which is 1.5 in BF16. We will
  // end up encoding this as the imm8 0x7f. This imm8 will be expanded to the
  // FP16 1.9375 which shares the same bit pattern as BF16 1.5.
  // FIXME: We should be able to handle f128 as well with a clever lowering.
  const APInt ImmInt = Imm.bitcastToAPInt();
  if (VT == MVT::f64)
    IsLegal = AArch64_AM::getFP64Imm(ImmInt) != -1 || Imm.isPosZero();
  else if (VT == MVT::f32)
    IsLegal = AArch64_AM::getFP32Imm(ImmInt) != -1 || Imm.isPosZero();
  else if (VT == MVT::f16 || VT == MVT::bf16)
    IsLegal =
        (Subtarget->hasFullFP16() && AArch64_AM::getFP16Imm(ImmInt) != -1) ||
        Imm.isPosZero();

  // If we can not materialize in immediate field for fmov, check if the
  // value can be encoded as the immediate operand of a logical instruction.
  // The immediate value will be created with either MOVZ, MOVN, or ORR.
  // TODO: fmov h0, w0 is also legal, however we don't have an isel pattern to
  //       generate that fmov.
  if (!IsLegal && (VT == MVT::f64 || VT == MVT::f32)) {
    // The cost is actually exactly the same for mov+fmov vs. adrp+ldr;
    // however the mov+fmov sequence is always better because of the reduced
    // cache pressure. The timings are still the same if you consider
    // movw+movk+fmov vs. adrp+ldr (it's one instruction longer, but the
    // movw+movk is fused). So we limit up to 2 instrdduction at most.
    SmallVector<AArch64_IMM::ImmInsnModel, 4> Insn;
    AArch64_IMM::expandMOVImm(ImmInt.getZExtValue(), VT.getSizeInBits(), Insn);
    assert(Insn.size() <= 4 &&
           "Should be able to build any value with at most 4 moves");
    unsigned Limit = (OptForSize ? 1 : (Subtarget->hasFuseLiterals() ? 4 : 2));
    IsLegal = Insn.size() <= Limit;
  }

  LLVM_DEBUG(dbgs() << (IsLegal ? "Legal " : "Illegal ") << VT
                    << " imm value: "; Imm.dump(););
  return IsLegal;
}

//===----------------------------------------------------------------------===//
//                          AArch64 Optimization Hooks
//===----------------------------------------------------------------------===//

static SDValue getEstimate(const AArch64Subtarget *ST, unsigned Opcode,
                           SDValue Operand, SelectionDAG &DAG,
                           int &ExtraSteps) {
  EVT VT = Operand.getValueType();
  if ((ST->hasNEON() &&
       (VT == MVT::f64 || VT == MVT::v1f64 || VT == MVT::v2f64 ||
        VT == MVT::f32 || VT == MVT::v1f32 || VT == MVT::v2f32 ||
        VT == MVT::v4f32)) ||
      (ST->hasSVE() &&
       (VT == MVT::nxv8f16 || VT == MVT::nxv4f32 || VT == MVT::nxv2f64))) {
    if (ExtraSteps == TargetLoweringBase::ReciprocalEstimate::Unspecified) {
      // For the reciprocal estimates, convergence is quadratic, so the number
      // of digits is doubled after each iteration.  In ARMv8, the accuracy of
      // the initial estimate is 2^-8.  Thus the number of extra steps to refine
      // the result for float (23 mantissa bits) is 2 and for double (52
      // mantissa bits) is 3.
      constexpr unsigned AccurateBits = 8;
      unsigned DesiredBits = APFloat::semanticsPrecision(VT.getFltSemantics());
      ExtraSteps = DesiredBits <= AccurateBits
                       ? 0
                       : Log2_64_Ceil(DesiredBits) - Log2_64_Ceil(AccurateBits);
    }

    return DAG.getNode(Opcode, SDLoc(Operand), VT, Operand);
  }

  return SDValue();
}

SDValue
AArch64TargetLowering::getSqrtInputTest(SDValue Op, SelectionDAG &DAG,
                                        const DenormalMode &Mode) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  EVT CCVT = getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT);
  SDValue FPZero = DAG.getConstantFP(0.0, DL, VT);
  return DAG.getSetCC(DL, CCVT, Op, FPZero, ISD::SETEQ);
}

SDValue
AArch64TargetLowering::getSqrtResultForDenormInput(SDValue Op,
                                                   SelectionDAG &DAG) const {
  return Op;
}

SDValue AArch64TargetLowering::getSqrtEstimate(SDValue Operand,
                                               SelectionDAG &DAG, int Enabled,
                                               int &ExtraSteps,
                                               bool &UseOneConst,
                                               bool Reciprocal) const {
  if (Enabled == ReciprocalEstimate::Enabled ||
      (Enabled == ReciprocalEstimate::Unspecified && Subtarget->useRSqrt()))
    if (SDValue Estimate = getEstimate(Subtarget, AArch64ISD::FRSQRTE, Operand,
                                       DAG, ExtraSteps)) {
      SDLoc DL(Operand);
      EVT VT = Operand.getValueType();

      SDNodeFlags Flags = SDNodeFlags::AllowReassociation;

      // Newton reciprocal square root iteration: E * 0.5 * (3 - X * E^2)
      // AArch64 reciprocal square root iteration instruction: 0.5 * (3 - M * N)
      for (int i = ExtraSteps; i > 0; --i) {
        SDValue Step = DAG.getNode(ISD::FMUL, DL, VT, Estimate, Estimate,
                                   Flags);
        Step = DAG.getNode(AArch64ISD::FRSQRTS, DL, VT, Operand, Step, Flags);
        Estimate = DAG.getNode(ISD::FMUL, DL, VT, Estimate, Step, Flags);
      }
      if (!Reciprocal)
        Estimate = DAG.getNode(ISD::FMUL, DL, VT, Operand, Estimate, Flags);

      ExtraSteps = 0;
      return Estimate;
    }

  return SDValue();
}

SDValue AArch64TargetLowering::getRecipEstimate(SDValue Operand,
                                                SelectionDAG &DAG, int Enabled,
                                                int &ExtraSteps) const {
  if (Enabled == ReciprocalEstimate::Enabled)
    if (SDValue Estimate = getEstimate(Subtarget, AArch64ISD::FRECPE, Operand,
                                       DAG, ExtraSteps)) {
      SDLoc DL(Operand);
      EVT VT = Operand.getValueType();

      SDNodeFlags Flags = SDNodeFlags::AllowReassociation;

      // Newton reciprocal iteration: E * (2 - X * E)
      // AArch64 reciprocal iteration instruction: (2 - M * N)
      for (int i = ExtraSteps; i > 0; --i) {
        SDValue Step = DAG.getNode(AArch64ISD::FRECPS, DL, VT, Operand,
                                   Estimate, Flags);
        Estimate = DAG.getNode(ISD::FMUL, DL, VT, Estimate, Step, Flags);
      }

      ExtraSteps = 0;
      return Estimate;
    }

  return SDValue();
}

//===----------------------------------------------------------------------===//
//                          AArch64 Inline Assembly Support
//===----------------------------------------------------------------------===//

// Table of Constraints
// TODO: This is the current set of constraints supported by ARM for the
// compiler, not all of them may make sense.
//
// r - A general register
// w - An FP/SIMD register of some size in the range v0-v31
// x - An FP/SIMD register of some size in the range v0-v15
// I - Constant that can be used with an ADD instruction
// J - Constant that can be used with a SUB instruction
// K - Constant that can be used with a 32-bit logical instruction
// L - Constant that can be used with a 64-bit logical instruction
// M - Constant that can be used as a 32-bit MOV immediate
// N - Constant that can be used as a 64-bit MOV immediate
// Q - A memory reference with base register and no offset
// S - A symbolic address
// Y - Floating point constant zero
// Z - Integer constant zero
//
//   Note that general register operands will be output using their 64-bit x
// register name, whatever the size of the variable, unless the asm operand
// is prefixed by the %w modifier. Floating-point and SIMD register operands
// will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
// %q modifier.
const char *AArch64TargetLowering::LowerXConstraint(EVT ConstraintVT) const {
  // At this point, we have to lower this constraint to something else, so we
  // lower it to an "r" or "w". However, by doing this we will force the result
  // to be in register, while the X constraint is much more permissive.
  //
  // Although we are correct (we are free to emit anything, without
  // constraints), we might break use cases that would expect us to be more
  // efficient and emit something else.
  if (!Subtarget->hasFPARMv8())
    return "r";

  if (ConstraintVT.isFloatingPoint())
    return "w";

  if (ConstraintVT.isVector() &&
     (ConstraintVT.getSizeInBits() == 64 ||
      ConstraintVT.getSizeInBits() == 128))
    return "w";

  return "r";
}

enum class PredicateConstraint { Uph, Upl, Upa };

// Returns a {Reg, RegisterClass} tuple if the constraint is
// a specific predicate register.
//
// For some constraint like "{pn3}" the default path in
// TargetLowering::getRegForInlineAsmConstraint() leads it to determine that a
// suitable register class for this register is "PPRorPNR", after which it
// determines that nxv16i1 is an appropriate type for the constraint, which is
// not what we want. The code here pre-empts this by matching the register
// explicitly.
static std::optional<std::pair<unsigned, const TargetRegisterClass *>>
parseSVERegAsConstraint(StringRef Constraint) {
  if (!Constraint.starts_with('{') || !Constraint.ends_with('}') ||
      (Constraint[1] != 'p' && Constraint[1] != 'z'))
    return std::nullopt;

  bool IsPredicate = Constraint[1] == 'p';
  Constraint = Constraint.substr(2, Constraint.size() - 3);
  bool IsPredicateAsCount = IsPredicate && Constraint.starts_with("n");
  if (IsPredicateAsCount)
    Constraint = Constraint.drop_front(1);

  unsigned V;
  if (Constraint.getAsInteger(10, V) || V > 31)
    return std::nullopt;

  if (IsPredicateAsCount)
    return std::make_pair(AArch64::PN0 + V, &AArch64::PNRRegClass);
  if (IsPredicate)
    return std::make_pair(AArch64::P0 + V, &AArch64::PPRRegClass);
  return std::make_pair(AArch64::Z0 + V, &AArch64::ZPRRegClass);
}

static std::optional<PredicateConstraint>
parsePredicateConstraint(StringRef Constraint) {
  return StringSwitch<std::optional<PredicateConstraint>>(Constraint)
      .Case("Uph", PredicateConstraint::Uph)
      .Case("Upl", PredicateConstraint::Upl)
      .Case("Upa", PredicateConstraint::Upa)
      .Default(std::nullopt);
}

static const TargetRegisterClass *
getPredicateRegisterClass(PredicateConstraint Constraint, EVT VT) {
  if (VT != MVT::aarch64svcount &&
      (!VT.isScalableVector() || VT.getVectorElementType() != MVT::i1))
    return nullptr;

  switch (Constraint) {
  case PredicateConstraint::Uph:
    return VT == MVT::aarch64svcount ? &AArch64::PNR_p8to15RegClass
                                     : &AArch64::PPR_p8to15RegClass;
  case PredicateConstraint::Upl:
    return VT == MVT::aarch64svcount ? &AArch64::PNR_3bRegClass
                                     : &AArch64::PPR_3bRegClass;
  case PredicateConstraint::Upa:
    return VT == MVT::aarch64svcount ? &AArch64::PNRRegClass
                                     : &AArch64::PPRRegClass;
  }

  llvm_unreachable("Missing PredicateConstraint!");
}

enum class ReducedGprConstraint { Uci, Ucj };

static std::optional<ReducedGprConstraint>
parseReducedGprConstraint(StringRef Constraint) {
  return StringSwitch<std::optional<ReducedGprConstraint>>(Constraint)
      .Case("Uci", ReducedGprConstraint::Uci)
      .Case("Ucj", ReducedGprConstraint::Ucj)
      .Default(std::nullopt);
}

static const TargetRegisterClass *
getReducedGprRegisterClass(ReducedGprConstraint Constraint, EVT VT) {
  if (!VT.isScalarInteger() || VT.getFixedSizeInBits() > 64)
    return nullptr;

  switch (Constraint) {
  case ReducedGprConstraint::Uci:
    return &AArch64::MatrixIndexGPR32_8_11RegClass;
  case ReducedGprConstraint::Ucj:
    return &AArch64::MatrixIndexGPR32_12_15RegClass;
  }

  llvm_unreachable("Missing ReducedGprConstraint!");
}

// The set of cc code supported is from
// https://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html#Flag-Output-Operands
static AArch64CC::CondCode parseConstraintCode(llvm::StringRef Constraint) {
  AArch64CC::CondCode Cond = StringSwitch<AArch64CC::CondCode>(Constraint)
                                 .Case("{@cchi}", AArch64CC::HI)
                                 .Case("{@cccs}", AArch64CC::HS)
                                 .Case("{@cclo}", AArch64CC::LO)
                                 .Case("{@ccls}", AArch64CC::LS)
                                 .Case("{@cccc}", AArch64CC::LO)
                                 .Case("{@cceq}", AArch64CC::EQ)
                                 .Case("{@ccgt}", AArch64CC::GT)
                                 .Case("{@ccge}", AArch64CC::GE)
                                 .Case("{@cclt}", AArch64CC::LT)
                                 .Case("{@ccle}", AArch64CC::LE)
                                 .Case("{@cchs}", AArch64CC::HS)
                                 .Case("{@ccne}", AArch64CC::NE)
                                 .Case("{@ccvc}", AArch64CC::VC)
                                 .Case("{@ccpl}", AArch64CC::PL)
                                 .Case("{@ccvs}", AArch64CC::VS)
                                 .Case("{@ccmi}", AArch64CC::MI)
                                 .Default(AArch64CC::Invalid);
  return Cond;
}

/// Helper function to create 'CSET', which is equivalent to 'CSINC <Wd>, WZR,
/// WZR, invert(<cond>)'.
static SDValue getSETCC(AArch64CC::CondCode CC, SDValue NZCV, const SDLoc &DL,
                        SelectionDAG &DAG) {
  return DAG.getNode(
      AArch64ISD::CSINC, DL, MVT::i32, DAG.getConstant(0, DL, MVT::i32),
      DAG.getConstant(0, DL, MVT::i32),
      DAG.getConstant(getInvertedCondCode(CC), DL, MVT::i32), NZCV);
}

// Lower @cc flag output via getSETCC.
SDValue AArch64TargetLowering::LowerAsmOutputForConstraint(
    SDValue &Chain, SDValue &Glue, const SDLoc &DL,
    const AsmOperandInfo &OpInfo, SelectionDAG &DAG) const {
  AArch64CC::CondCode Cond = parseConstraintCode(OpInfo.ConstraintCode);
  if (Cond == AArch64CC::Invalid)
    return SDValue();
  // The output variable should be a scalar integer.
  if (OpInfo.ConstraintVT.isVector() || !OpInfo.ConstraintVT.isInteger() ||
      OpInfo.ConstraintVT.getSizeInBits() < 8)
    report_fatal_error("Flag output operand is of invalid type");

  // Get NZCV register. Only update chain when copyfrom is glued.
  if (Glue.getNode()) {
    Glue = DAG.getCopyFromReg(Chain, DL, AArch64::NZCV, MVT::i32, Glue);
    Chain = Glue.getValue(1);
  } else
    Glue = DAG.getCopyFromReg(Chain, DL, AArch64::NZCV, MVT::i32);
  // Extract CC code.
  SDValue CC = getSETCC(Cond, Glue, DL, DAG);

  SDValue Result;

  // Truncate or ZERO_EXTEND based on value types.
  if (OpInfo.ConstraintVT.getSizeInBits() <= 32)
    Result = DAG.getNode(ISD::TRUNCATE, DL, OpInfo.ConstraintVT, CC);
  else
    Result = DAG.getNode(ISD::ZERO_EXTEND, DL, OpInfo.ConstraintVT, CC);

  return Result;
}

/// getConstraintType - Given a constraint letter, return the type of
/// constraint it is for this target.
AArch64TargetLowering::ConstraintType
AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    default:
      break;
    case 'x':
    case 'w':
    case 'y':
      return C_RegisterClass;
    // An address with a single base register. Due to the way we
    // currently handle addresses it is the same as 'r'.
    case 'Q':
      return C_Memory;
    case 'I':
    case 'J':
    case 'K':
    case 'L':
    case 'M':
    case 'N':
    case 'Y':
    case 'Z':
      return C_Immediate;
    case 'z':
    case 'S': // A symbol or label reference with a constant offset
      return C_Other;
    }
  } else if (parsePredicateConstraint(Constraint))
    return C_RegisterClass;
  else if (parseReducedGprConstraint(Constraint))
    return C_RegisterClass;
  else if (parseConstraintCode(Constraint) != AArch64CC::Invalid)
    return C_Other;
  return TargetLowering::getConstraintType(Constraint);
}

/// Examine constraint type and operand type and determine a weight value.
/// This object must already have been set up with the operand type
/// and the current alternative constraint selected.
TargetLowering::ConstraintWeight
AArch64TargetLowering::getSingleConstraintMatchWeight(
    AsmOperandInfo &info, const char *constraint) const {
  ConstraintWeight weight = CW_Invalid;
  Value *CallOperandVal = info.CallOperandVal;
  // If we don't have a value, we can't do a match,
  // but allow it at the lowest weight.
  if (!CallOperandVal)
    return CW_Default;
  Type *type = CallOperandVal->getType();
  // Look at the constraint type.
  switch (*constraint) {
  default:
    weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
    break;
  case 'x':
  case 'w':
  case 'y':
    if (type->isFloatingPointTy() || type->isVectorTy())
      weight = CW_Register;
    break;
  case 'z':
    weight = CW_Constant;
    break;
  case 'U':
    if (parsePredicateConstraint(constraint) ||
        parseReducedGprConstraint(constraint))
      weight = CW_Register;
    break;
  }
  return weight;
}

std::pair<unsigned, const TargetRegisterClass *>
AArch64TargetLowering::getRegForInlineAsmConstraint(
    const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    case 'r':
      if (VT.isScalableVector())
        return std::make_pair(0U, nullptr);
      if (Subtarget->hasLS64() && VT.getSizeInBits() == 512)
        return std::make_pair(0U, &AArch64::GPR64x8ClassRegClass);
      if (VT.getFixedSizeInBits() == 64)
        return std::make_pair(0U, &AArch64::GPR64commonRegClass);
      return std::make_pair(0U, &AArch64::GPR32commonRegClass);
    case 'w': {
      if (!Subtarget->hasFPARMv8())
        break;
      if (VT.isScalableVector()) {
        if (VT.getVectorElementType() != MVT::i1)
          return std::make_pair(0U, &AArch64::ZPRRegClass);
        return std::make_pair(0U, nullptr);
      }
      if (VT == MVT::Other)
        break;
      uint64_t VTSize = VT.getFixedSizeInBits();
      if (VTSize == 16)
        return std::make_pair(0U, &AArch64::FPR16RegClass);
      if (VTSize == 32)
        return std::make_pair(0U, &AArch64::FPR32RegClass);
      if (VTSize == 64)
        return std::make_pair(0U, &AArch64::FPR64RegClass);
      if (VTSize == 128)
        return std::make_pair(0U, &AArch64::FPR128RegClass);
      break;
    }
    // The instructions that this constraint is designed for can
    // only take 128-bit registers so just use that regclass.
    case 'x':
      if (!Subtarget->hasFPARMv8())
        break;
      if (VT.isScalableVector())
        return std::make_pair(0U, &AArch64::ZPR_4bRegClass);
      if (VT.getSizeInBits() == 128)
        return std::make_pair(0U, &AArch64::FPR128_loRegClass);
      break;
    case 'y':
      if (!Subtarget->hasFPARMv8())
        break;
      if (VT.isScalableVector())
        return std::make_pair(0U, &AArch64::ZPR_3bRegClass);
      break;
    }
  } else {
    if (const auto P = parseSVERegAsConstraint(Constraint)) {
      // SME functions that are not in streaming mode, should
      // still observe clobbers of Z-registers by clobbering
      // the lower 128bits of those registers.
      if (AArch64::ZPRRegClass.hasSubClassEq(P->second) &&
          !Subtarget->isSVEorStreamingSVEAvailable())
        return std::make_pair(TRI->getSubReg(P->first, AArch64::zsub),
                              &AArch64::FPR128RegClass);
      return *P;
    }
    if (const auto PC = parsePredicateConstraint(Constraint))
      if (const auto *RegClass = getPredicateRegisterClass(*PC, VT))
        return std::make_pair(0U, RegClass);

    if (const auto RGC = parseReducedGprConstraint(Constraint))
      if (const auto *RegClass = getReducedGprRegisterClass(*RGC, VT))
        return std::make_pair(0U, RegClass);
  }
  if (StringRef("{cc}").equals_insensitive(Constraint) ||
      parseConstraintCode(Constraint) != AArch64CC::Invalid)
    return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);

  if (Constraint == "{za}") {
    return std::make_pair(unsigned(AArch64::ZA), &AArch64::MPRRegClass);
  }

  if (Constraint == "{zt0}") {
    return std::make_pair(unsigned(AArch64::ZT0), &AArch64::ZTRRegClass);
  }

  // Use the default implementation in TargetLowering to convert the register
  // constraint into a member of a register class.
  std::pair<unsigned, const TargetRegisterClass *> Res;
  Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);

  // Not found as a standard register?
  if (!Res.second) {
    unsigned Size = Constraint.size();
    if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
        tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
      int RegNo;
      bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
      if (!Failed && RegNo >= 0 && RegNo <= 31) {
        // v0 - v31 are aliases of q0 - q31 or d0 - d31 depending on size.
        // By default we'll emit v0-v31 for this unless there's a modifier where
        // we'll emit the correct register as well.
        if (VT != MVT::Other && VT.getSizeInBits() == 64) {
          Res.first = AArch64::FPR64RegClass.getRegister(RegNo);
          Res.second = &AArch64::FPR64RegClass;
        } else {
          Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
          Res.second = &AArch64::FPR128RegClass;
        }
      }
    }
  }

  if (Res.second && !Subtarget->hasFPARMv8() &&
      !AArch64::GPR32allRegClass.hasSubClassEq(Res.second) &&
      !AArch64::GPR64allRegClass.hasSubClassEq(Res.second))
    return std::make_pair(0U, nullptr);

  return Res;
}

EVT AArch64TargetLowering::getAsmOperandValueType(const DataLayout &DL,
                                                  llvm::Type *Ty,
                                                  bool AllowUnknown) const {
  if (Subtarget->hasLS64() && Ty->isIntegerTy(512))
    return EVT(MVT::i64x8);

  return TargetLowering::getAsmOperandValueType(DL, Ty, AllowUnknown);
}

/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
/// vector.  If it is invalid, don't add anything to Ops.
void AArch64TargetLowering::LowerAsmOperandForConstraint(
    SDValue Op, StringRef Constraint, std::vector<SDValue> &Ops,
    SelectionDAG &DAG) const {
  SDValue Result;

  // Currently only support length 1 constraints.
  if (Constraint.size() != 1)
    return;

  char ConstraintLetter = Constraint[0];
  switch (ConstraintLetter) {
  default:
    break;

  // This set of constraints deal with valid constants for various instructions.
  // Validate and return a target constant for them if we can.
  case 'z': {
    // 'z' maps to xzr or wzr so it needs an input of 0.
    if (!isNullConstant(Op))
      return;

    if (Op.getValueType() == MVT::i64)
      Result = DAG.getRegister(AArch64::XZR, MVT::i64);
    else
      Result = DAG.getRegister(AArch64::WZR, MVT::i32);
    break;
  }
  case 'S':
    // Use the generic code path for "s". In GCC's aarch64 port, "S" is
    // supported for PIC while "s" isn't, making "s" less useful. We implement
    // "S" but not "s".
    TargetLowering::LowerAsmOperandForConstraint(Op, "s", Ops, DAG);
    break;

  case 'I':
  case 'J':
  case 'K':
  case 'L':
  case 'M':
  case 'N':
    ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
    if (!C)
      return;

    // Grab the value and do some validation.
    uint64_t CVal = C->getZExtValue();
    switch (ConstraintLetter) {
    // The I constraint applies only to simple ADD or SUB immediate operands:
    // i.e. 0 to 4095 with optional shift by 12
    // The J constraint applies only to ADD or SUB immediates that would be
    // valid when negated, i.e. if [an add pattern] were to be output as a SUB
    // instruction [or vice versa], in other words -1 to -4095 with optional
    // left shift by 12.
    case 'I':
      if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
        break;
      return;
    case 'J': {
      uint64_t NVal = -C->getSExtValue();
      if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
        CVal = C->getSExtValue();
        break;
      }
      return;
    }
    // The K and L constraints apply *only* to logical immediates, including
    // what used to be the MOVI alias for ORR (though the MOVI alias has now
    // been removed and MOV should be used). So these constraints have to
    // distinguish between bit patterns that are valid 32-bit or 64-bit
    // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
    // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
    // versa.
    case 'K':
      if (AArch64_AM::isLogicalImmediate(CVal, 32))
        break;
      return;
    case 'L':
      if (AArch64_AM::isLogicalImmediate(CVal, 64))
        break;
      return;
    // The M and N constraints are a superset of K and L respectively, for use
    // with the MOV (immediate) alias. As well as the logical immediates they
    // also match 32 or 64-bit immediates that can be loaded either using a
    // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
    // (M) or 64-bit 0x1234000000000000 (N) etc.
    // As a note some of this code is liberally stolen from the asm parser.
    case 'M': {
      if (!isUInt<32>(CVal))
        return;
      if (AArch64_AM::isLogicalImmediate(CVal, 32))
        break;
      if ((CVal & 0xFFFF) == CVal)
        break;
      if ((CVal & 0xFFFF0000ULL) == CVal)
        break;
      uint64_t NCVal = ~(uint32_t)CVal;
      if ((NCVal & 0xFFFFULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF0000ULL) == NCVal)
        break;
      return;
    }
    case 'N': {
      if (AArch64_AM::isLogicalImmediate(CVal, 64))
        break;
      if ((CVal & 0xFFFFULL) == CVal)
        break;
      if ((CVal & 0xFFFF0000ULL) == CVal)
        break;
      if ((CVal & 0xFFFF00000000ULL) == CVal)
        break;
      if ((CVal & 0xFFFF000000000000ULL) == CVal)
        break;
      uint64_t NCVal = ~CVal;
      if ((NCVal & 0xFFFFULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF0000ULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF00000000ULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
        break;
      return;
    }
    default:
      return;
    }

    // All assembler immediates are 64-bit integers.
    Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
    break;
  }

  if (Result.getNode()) {
    Ops.push_back(Result);
    return;
  }

  return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
}

//===----------------------------------------------------------------------===//
//                     AArch64 Advanced SIMD Support
//===----------------------------------------------------------------------===//

/// WidenVector - Given a value in the V64 register class, produce the
/// equivalent value in the V128 register class.
static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
  EVT VT = V64Reg.getValueType();
  unsigned NarrowSize = VT.getVectorNumElements();
  MVT EltTy = VT.getVectorElementType().getSimpleVT();
  MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
  SDLoc DL(V64Reg);

  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
                     V64Reg, DAG.getConstant(0, DL, MVT::i64));
}

/// getExtFactor - Determine the adjustment factor for the position when
/// generating an "extract from vector registers" instruction.
static unsigned getExtFactor(SDValue &V) {
  EVT EltType = V.getValueType().getVectorElementType();
  return EltType.getSizeInBits() / 8;
}

// Check if a vector is built from one vector via extracted elements of
// another together with an AND mask, ensuring that all elements fit
// within range. This can be reconstructed using AND and NEON's TBL1.
SDValue ReconstructShuffleWithRuntimeMask(SDValue Op, SelectionDAG &DAG) {
  assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  assert(!VT.isScalableVector() &&
         "Scalable vectors cannot be used with ISD::BUILD_VECTOR");

  // Can only recreate a shuffle with 16xi8 or 8xi8 elements, as they map
  // directly to TBL1.
  if (VT != MVT::v16i8 && VT != MVT::v8i8)
    return SDValue();

  unsigned NumElts = VT.getVectorNumElements();
  assert((NumElts == 8 || NumElts == 16) &&
         "Need to have exactly 8 or 16 elements in vector.");

  SDValue SourceVec;
  SDValue MaskSourceVec;
  SmallVector<SDValue, 16> AndMaskConstants;

  for (unsigned i = 0; i < NumElts; ++i) {
    SDValue V = Op.getOperand(i);
    if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      return SDValue();

    SDValue OperandSourceVec = V.getOperand(0);
    if (!SourceVec)
      SourceVec = OperandSourceVec;
    else if (SourceVec != OperandSourceVec)
      return SDValue();

    // This only looks at shuffles with elements that are
    // a) truncated by a constant AND mask extracted from a mask vector, or
    // b) extracted directly from a mask vector.
    SDValue MaskSource = V.getOperand(1);
    if (MaskSource.getOpcode() == ISD::AND) {
      if (!isa<ConstantSDNode>(MaskSource.getOperand(1)))
        return SDValue();

      AndMaskConstants.push_back(MaskSource.getOperand(1));
      MaskSource = MaskSource->getOperand(0);
    } else if (!AndMaskConstants.empty()) {
      // Either all or no operands should have an AND mask.
      return SDValue();
    }

    // An ANY_EXTEND may be inserted between the AND and the source vector
    // extraction. We don't care about that, so we can just skip it.
    if (MaskSource.getOpcode() == ISD::ANY_EXTEND)
      MaskSource = MaskSource.getOperand(0);

    if (MaskSource.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      return SDValue();

    SDValue MaskIdx = MaskSource.getOperand(1);
    if (!isa<ConstantSDNode>(MaskIdx) ||
        !cast<ConstantSDNode>(MaskIdx)->getConstantIntValue()->equalsInt(i))
      return SDValue();

    // We only apply this if all elements come from the same vector with the
    // same vector type.
    if (!MaskSourceVec) {
      MaskSourceVec = MaskSource->getOperand(0);
      if (MaskSourceVec.getValueType() != VT)
        return SDValue();
    } else if (MaskSourceVec != MaskSource->getOperand(0)) {
      return SDValue();
    }
  }

  // We need a v16i8 for TBL, so we extend the source with a placeholder vector
  // for v8i8 to get a v16i8. As the pattern we are replacing is extract +
  // insert, we know that the index in the mask must be smaller than the number
  // of elements in the source, or we would have an out-of-bounds access.
  if (NumElts == 8)
    SourceVec = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, SourceVec,
                            DAG.getUNDEF(VT));

  // Preconditions met, so we can use a vector (AND +) TBL to build this vector.
  if (!AndMaskConstants.empty())
    MaskSourceVec = DAG.getNode(ISD::AND, DL, VT, MaskSourceVec,
                                DAG.getBuildVector(VT, DL, AndMaskConstants));

  return DAG.getNode(
      ISD::INTRINSIC_WO_CHAIN, DL, VT,
      DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), SourceVec,
      MaskSourceVec);
}

// Gather data to see if the operation can be modelled as a
// shuffle in combination with VEXTs.
SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
                                                  SelectionDAG &DAG) const {
  assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
  LLVM_DEBUG(dbgs() << "AArch64TargetLowering::ReconstructShuffle\n");
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  assert(!VT.isScalableVector() &&
         "Scalable vectors cannot be used with ISD::BUILD_VECTOR");
  unsigned NumElts = VT.getVectorNumElements();

  struct ShuffleSourceInfo {
    SDValue Vec;
    unsigned MinElt;
    unsigned MaxElt;

    // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
    // be compatible with the shuffle we intend to construct. As a result
    // ShuffleVec will be some sliding window into the original Vec.
    SDValue ShuffleVec;

    // Code should guarantee that element i in Vec starts at element "WindowBase
    // + i * WindowScale in ShuffleVec".
    int WindowBase;
    int WindowScale;

    ShuffleSourceInfo(SDValue Vec)
      : Vec(Vec), MinElt(std::numeric_limits<unsigned>::max()), MaxElt(0),
          ShuffleVec(Vec), WindowBase(0), WindowScale(1) {}

    bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
  };

  // First gather all vectors used as an immediate source for this BUILD_VECTOR
  // node.
  SmallVector<ShuffleSourceInfo, 2> Sources;
  for (unsigned i = 0; i < NumElts; ++i) {
    SDValue V = Op.getOperand(i);
    if (V.isUndef())
      continue;
    else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
             !isa<ConstantSDNode>(V.getOperand(1)) ||
             V.getOperand(0).getValueType().isScalableVector()) {
      LLVM_DEBUG(
          dbgs() << "Reshuffle failed: "
                    "a shuffle can only come from building a vector from "
                    "various elements of other fixed-width vectors, provided "
                    "their indices are constant\n");
      return SDValue();
    }

    // Add this element source to the list if it's not already there.
    SDValue SourceVec = V.getOperand(0);
    auto Source = find(Sources, SourceVec);
    if (Source == Sources.end())
      Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));

    // Update the minimum and maximum lane number seen.
    unsigned EltNo = V.getConstantOperandVal(1);
    Source->MinElt = std::min(Source->MinElt, EltNo);
    Source->MaxElt = std::max(Source->MaxElt, EltNo);
  }

  // If we have 3 or 4 sources, try to generate a TBL, which will at least be
  // better than moving to/from gpr registers for larger vectors.
  if ((Sources.size() == 3 || Sources.size() == 4) && NumElts > 4) {
    // Construct a mask for the tbl. We may need to adjust the index for types
    // larger than i8.
    SmallVector<unsigned, 16> Mask;
    unsigned OutputFactor = VT.getScalarSizeInBits() / 8;
    for (unsigned I = 0; I < NumElts; ++I) {
      SDValue V = Op.getOperand(I);
      if (V.isUndef()) {
        for (unsigned OF = 0; OF < OutputFactor; OF++)
          Mask.push_back(-1);
        continue;
      }
      // Set the Mask lanes adjusted for the size of the input and output
      // lanes. The Mask is always i8, so it will set OutputFactor lanes per
      // output element, adjusted in their positions per input and output types.
      unsigned Lane = V.getConstantOperandVal(1);
      for (unsigned S = 0; S < Sources.size(); S++) {
        if (V.getOperand(0) == Sources[S].Vec) {
          unsigned InputSize = Sources[S].Vec.getScalarValueSizeInBits();
          unsigned InputBase = 16 * S + Lane * InputSize / 8;
          for (unsigned OF = 0; OF < OutputFactor; OF++)
            Mask.push_back(InputBase + OF);
          break;
        }
      }
    }

    // Construct the tbl3/tbl4 out of an intrinsic, the sources converted to
    // v16i8, and the TBLMask
    SmallVector<SDValue, 16> TBLOperands;
    TBLOperands.push_back(DAG.getConstant(Sources.size() == 3
                                              ? Intrinsic::aarch64_neon_tbl3
                                              : Intrinsic::aarch64_neon_tbl4,
                                          DL, MVT::i32));
    for (unsigned i = 0; i < Sources.size(); i++) {
      SDValue Src = Sources[i].Vec;
      EVT SrcVT = Src.getValueType();
      Src = DAG.getBitcast(SrcVT.is64BitVector() ? MVT::v8i8 : MVT::v16i8, Src);
      assert((SrcVT.is64BitVector() || SrcVT.is128BitVector()) &&
             "Expected a legally typed vector");
      if (SrcVT.is64BitVector())
        Src = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, Src,
                          DAG.getUNDEF(MVT::v8i8));
      TBLOperands.push_back(Src);
    }

    SmallVector<SDValue, 16> TBLMask;
    for (unsigned i = 0; i < Mask.size(); i++)
      TBLMask.push_back(DAG.getConstant(Mask[i], DL, MVT::i32));
    assert((Mask.size() == 8 || Mask.size() == 16) &&
           "Expected a v8i8 or v16i8 Mask");
    TBLOperands.push_back(DAG.getBuildVector(
        Mask.size() == 8 ? MVT::v8i8 : MVT::v16i8, DL, TBLMask));

    SDValue Shuffle =
        DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL,
                    Mask.size() == 8 ? MVT::v8i8 : MVT::v16i8, TBLOperands);
    return DAG.getBitcast(VT, Shuffle);
  }

  if (Sources.size() > 2) {
    LLVM_DEBUG(dbgs() << "Reshuffle failed: currently only do something "
                      << "sensible when at most two source vectors are "
                      << "involved\n");
    return SDValue();
  }

  // Find out the smallest element size among result and two sources, and use
  // it as element size to build the shuffle_vector.
  EVT SmallestEltTy = VT.getVectorElementType();
  for (auto &Source : Sources) {
    EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
    if (SrcEltTy.bitsLT(SmallestEltTy)) {
      SmallestEltTy = SrcEltTy;
    }
  }
  unsigned ResMultiplier =
      VT.getScalarSizeInBits() / SmallestEltTy.getFixedSizeInBits();
  uint64_t VTSize = VT.getFixedSizeInBits();
  NumElts = VTSize / SmallestEltTy.getFixedSizeInBits();
  EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);

  // If the source vector is too wide or too narrow, we may nevertheless be able
  // to construct a compatible shuffle either by concatenating it with UNDEF or
  // extracting a suitable range of elements.
  for (auto &Src : Sources) {
    EVT SrcVT = Src.ShuffleVec.getValueType();

    TypeSize SrcVTSize = SrcVT.getSizeInBits();
    if (SrcVTSize == TypeSize::getFixed(VTSize))
      continue;

    // This stage of the search produces a source with the same element type as
    // the original, but with a total width matching the BUILD_VECTOR output.
    EVT EltVT = SrcVT.getVectorElementType();
    unsigned NumSrcElts = VTSize / EltVT.getFixedSizeInBits();
    EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);

    if (SrcVTSize.getFixedValue() < VTSize) {
      assert(2 * SrcVTSize == VTSize);
      // We can pad out the smaller vector for free, so if it's part of a
      // shuffle...
      Src.ShuffleVec =
          DAG.getNode(ISD::CONCAT_VECTORS, DL, DestVT, Src.ShuffleVec,
                      DAG.getUNDEF(Src.ShuffleVec.getValueType()));
      continue;
    }

    if (SrcVTSize.getFixedValue() != 2 * VTSize) {
      LLVM_DEBUG(
          dbgs() << "Reshuffle failed: result vector too small to extract\n");
      return SDValue();
    }

    if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
      LLVM_DEBUG(
          dbgs() << "Reshuffle failed: span too large for a VEXT to cope\n");
      return SDValue();
    }

    if (Src.MinElt >= NumSrcElts) {
      // The extraction can just take the second half
      Src.ShuffleVec =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, DestVT, Src.ShuffleVec,
                      DAG.getConstant(NumSrcElts, DL, MVT::i64));
      Src.WindowBase = -NumSrcElts;
    } else if (Src.MaxElt < NumSrcElts) {
      // The extraction can just take the first half
      Src.ShuffleVec =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, DestVT, Src.ShuffleVec,
                      DAG.getConstant(0, DL, MVT::i64));
    } else {
      // An actual VEXT is needed
      SDValue VEXTSrc1 =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, DestVT, Src.ShuffleVec,
                      DAG.getConstant(0, DL, MVT::i64));
      SDValue VEXTSrc2 =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, DestVT, Src.ShuffleVec,
                      DAG.getConstant(NumSrcElts, DL, MVT::i64));
      unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);

      if (!SrcVT.is64BitVector()) {
        LLVM_DEBUG(
          dbgs() << "Reshuffle failed: don't know how to lower AArch64ISD::EXT "
                    "for SVE vectors.");
        return SDValue();
      }

      Src.ShuffleVec =
          DAG.getNode(AArch64ISD::EXT, DL, DestVT, VEXTSrc1, VEXTSrc2,
                      DAG.getConstant(Imm, DL, MVT::i32));
      Src.WindowBase = -Src.MinElt;
    }
  }

  // Another possible incompatibility occurs from the vector element types. We
  // can fix this by bitcasting the source vectors to the same type we intend
  // for the shuffle.
  for (auto &Src : Sources) {
    EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
    if (SrcEltTy == SmallestEltTy)
      continue;
    assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
    if (DAG.getDataLayout().isBigEndian()) {
      Src.ShuffleVec =
          DAG.getNode(AArch64ISD::NVCAST, DL, ShuffleVT, Src.ShuffleVec);
    } else {
      Src.ShuffleVec = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Src.ShuffleVec);
    }
    Src.WindowScale =
        SrcEltTy.getFixedSizeInBits() / SmallestEltTy.getFixedSizeInBits();
    Src.WindowBase *= Src.WindowScale;
  }

  // Final check before we try to actually produce a shuffle.
  LLVM_DEBUG({
    for (auto Src : Sources)
      assert(Src.ShuffleVec.getValueType() == ShuffleVT);
  });

  // The stars all align, our next step is to produce the mask for the shuffle.
  SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
  int BitsPerShuffleLane = ShuffleVT.getScalarSizeInBits();
  for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
    SDValue Entry = Op.getOperand(i);
    if (Entry.isUndef())
      continue;

    auto Src = find(Sources, Entry.getOperand(0));
    int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();

    // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
    // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
    // segment.
    EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
    int BitsDefined = std::min(OrigEltTy.getScalarSizeInBits(),
                               VT.getScalarSizeInBits());
    int LanesDefined = BitsDefined / BitsPerShuffleLane;

    // This source is expected to fill ResMultiplier lanes of the final shuffle,
    // starting at the appropriate offset.
    int *LaneMask = &Mask[i * ResMultiplier];

    int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
    ExtractBase += NumElts * (Src - Sources.begin());
    for (int j = 0; j < LanesDefined; ++j)
      LaneMask[j] = ExtractBase + j;
  }

  // Final check before we try to produce nonsense...
  if (!isShuffleMaskLegal(Mask, ShuffleVT)) {
    LLVM_DEBUG(dbgs() << "Reshuffle failed: illegal shuffle mask\n");
    return SDValue();
  }

  SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
  for (unsigned i = 0; i < Sources.size(); ++i)
    ShuffleOps[i] = Sources[i].ShuffleVec;

  SDValue Shuffle =
      DAG.getVectorShuffle(ShuffleVT, DL, ShuffleOps[0], ShuffleOps[1], Mask);
  SDValue V;
  if (DAG.getDataLayout().isBigEndian()) {
    V = DAG.getNode(AArch64ISD::NVCAST, DL, VT, Shuffle);
  } else {
    V = DAG.getNode(ISD::BITCAST, DL, VT, Shuffle);
  }

  LLVM_DEBUG(dbgs() << "Reshuffle, creating node: "; Shuffle.dump();
             dbgs() << "Reshuffle, creating node: "; V.dump(););

  return V;
}

// check if an EXT instruction can handle the shuffle mask when the
// vector sources of the shuffle are the same.
static bool isSingletonEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
  unsigned NumElts = VT.getVectorNumElements();

  // Assume that the first shuffle index is not UNDEF.  Fail if it is.
  if (M[0] < 0)
    return false;

  Imm = M[0];

  // If this is a VEXT shuffle, the immediate value is the index of the first
  // element.  The other shuffle indices must be the successive elements after
  // the first one.
  unsigned ExpectedElt = Imm;
  for (unsigned i = 1; i < NumElts; ++i) {
    // Increment the expected index.  If it wraps around, just follow it
    // back to index zero and keep going.
    ++ExpectedElt;
    if (ExpectedElt == NumElts)
      ExpectedElt = 0;

    if (M[i] < 0)
      continue; // ignore UNDEF indices
    if (ExpectedElt != static_cast<unsigned>(M[i]))
      return false;
  }

  return true;
}

// Detect patterns of a0,a1,a2,a3,b0,b1,b2,b3,c0,c1,c2,c3,d0,d1,d2,d3 from
// v4i32s. This is really a truncate, which we can construct out of (legal)
// concats and truncate nodes.
static SDValue ReconstructTruncateFromBuildVector(SDValue V, SelectionDAG &DAG) {
  if (V.getValueType() != MVT::v16i8)
    return SDValue();
  assert(V.getNumOperands() == 16 && "Expected 16 operands on the BUILDVECTOR");

  for (unsigned X = 0; X < 4; X++) {
    // Check the first item in each group is an extract from lane 0 of a v4i32
    // or v4i16.
    SDValue BaseExt = V.getOperand(X * 4);
    if (BaseExt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
        (BaseExt.getOperand(0).getValueType() != MVT::v4i16 &&
         BaseExt.getOperand(0).getValueType() != MVT::v4i32) ||
        !isa<ConstantSDNode>(BaseExt.getOperand(1)) ||
        BaseExt.getConstantOperandVal(1) != 0)
      return SDValue();
    SDValue Base = BaseExt.getOperand(0);
    // And check the other items are extracts from the same vector.
    for (unsigned Y = 1; Y < 4; Y++) {
      SDValue Ext = V.getOperand(X * 4 + Y);
      if (Ext.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
          Ext.getOperand(0) != Base ||
          !isa<ConstantSDNode>(Ext.getOperand(1)) ||
          Ext.getConstantOperandVal(1) != Y)
        return SDValue();
    }
  }

  // Turn the buildvector into a series of truncates and concates, which will
  // become uzip1's. Any v4i32s we found get truncated to v4i16, which are
  // concat together to produce 2 v8i16. These are both truncated and concat
  // together.
  SDLoc DL(V);
  SDValue Trunc[4] = {
      V.getOperand(0).getOperand(0), V.getOperand(4).getOperand(0),
      V.getOperand(8).getOperand(0), V.getOperand(12).getOperand(0)};
  for (SDValue &V : Trunc)
    if (V.getValueType() == MVT::v4i32)
      V = DAG.getNode(ISD::TRUNCATE, DL, MVT::v4i16, V);
  SDValue Concat0 =
      DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i16, Trunc[0], Trunc[1]);
  SDValue Concat1 =
      DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i16, Trunc[2], Trunc[3]);
  SDValue Trunc0 = DAG.getNode(ISD::TRUNCATE, DL, MVT::v8i8, Concat0);
  SDValue Trunc1 = DAG.getNode(ISD::TRUNCATE, DL, MVT::v8i8, Concat1);
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, Trunc0, Trunc1);
}

/// Check if a vector shuffle corresponds to a DUP instructions with a larger
/// element width than the vector lane type. If that is the case the function
/// returns true and writes the value of the DUP instruction lane operand into
/// DupLaneOp
static bool isWideDUPMask(ArrayRef<int> M, EVT VT, unsigned BlockSize,
                          unsigned &DupLaneOp) {
  assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
         "Only possible block sizes for wide DUP are: 16, 32, 64");

  if (BlockSize <= VT.getScalarSizeInBits())
    return false;
  if (BlockSize % VT.getScalarSizeInBits() != 0)
    return false;
  if (VT.getSizeInBits() % BlockSize != 0)
    return false;

  size_t SingleVecNumElements = VT.getVectorNumElements();
  size_t NumEltsPerBlock = BlockSize / VT.getScalarSizeInBits();
  size_t NumBlocks = VT.getSizeInBits() / BlockSize;

  // We are looking for masks like
  // [0, 1, 0, 1] or [2, 3, 2, 3] or [4, 5, 6, 7, 4, 5, 6, 7] where any element
  // might be replaced by 'undefined'. BlockIndices will eventually contain
  // lane indices of the duplicated block (i.e. [0, 1], [2, 3] and [4, 5, 6, 7]
  // for the above examples)
  SmallVector<int, 8> BlockElts(NumEltsPerBlock, -1);
  for (size_t BlockIndex = 0; BlockIndex < NumBlocks; BlockIndex++)
    for (size_t I = 0; I < NumEltsPerBlock; I++) {
      int Elt = M[BlockIndex * NumEltsPerBlock + I];
      if (Elt < 0)
        continue;
      // For now we don't support shuffles that use the second operand
      if ((unsigned)Elt >= SingleVecNumElements)
        return false;
      if (BlockElts[I] < 0)
        BlockElts[I] = Elt;
      else if (BlockElts[I] != Elt)
        return false;
    }

  // We found a candidate block (possibly with some undefs). It must be a
  // sequence of consecutive integers starting with a value divisible by
  // NumEltsPerBlock with some values possibly replaced by undef-s.

  // Find first non-undef element
  auto FirstRealEltIter = find_if(BlockElts, [](int Elt) { return Elt >= 0; });
  assert(FirstRealEltIter != BlockElts.end() &&
         "Shuffle with all-undefs must have been caught by previous cases, "
         "e.g. isSplat()");
  if (FirstRealEltIter == BlockElts.end()) {
    DupLaneOp = 0;
    return true;
  }

  // Index of FirstRealElt in BlockElts
  size_t FirstRealIndex = FirstRealEltIter - BlockElts.begin();

  if ((unsigned)*FirstRealEltIter < FirstRealIndex)
    return false;
  // BlockElts[0] must have the following value if it isn't undef:
  size_t Elt0 = *FirstRealEltIter - FirstRealIndex;

  // Check the first element
  if (Elt0 % NumEltsPerBlock != 0)
    return false;
  // Check that the sequence indeed consists of consecutive integers (modulo
  // undefs)
  for (size_t I = 0; I < NumEltsPerBlock; I++)
    if (BlockElts[I] >= 0 && (unsigned)BlockElts[I] != Elt0 + I)
      return false;

  DupLaneOp = Elt0 / NumEltsPerBlock;
  return true;
}

// check if an EXT instruction can handle the shuffle mask when the
// vector sources of the shuffle are different.
static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
                      unsigned &Imm) {
  // Look for the first non-undef element.
  const int *FirstRealElt = find_if(M, [](int Elt) { return Elt >= 0; });

  // Benefit form APInt to handle overflow when calculating expected element.
  unsigned NumElts = VT.getVectorNumElements();
  unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
  APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1, /*isSigned=*/false,
                            /*implicitTrunc=*/true);
  // The following shuffle indices must be the successive elements after the
  // first real element.
  bool FoundWrongElt = std::any_of(FirstRealElt + 1, M.end(), [&](int Elt) {
    return Elt != ExpectedElt++ && Elt != -1;
  });
  if (FoundWrongElt)
    return false;

  // The index of an EXT is the first element if it is not UNDEF.
  // Watch out for the beginning UNDEFs. The EXT index should be the expected
  // value of the first element.  E.g.
  // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
  // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
  // ExpectedElt is the last mask index plus 1.
  Imm = ExpectedElt.getZExtValue();

  // There are two difference cases requiring to reverse input vectors.
  // For example, for vector <4 x i32> we have the following cases,
  // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
  // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
  // For both cases, we finally use mask <5, 6, 7, 0>, which requires
  // to reverse two input vectors.
  if (Imm < NumElts)
    ReverseEXT = true;
  else
    Imm -= NumElts;

  return true;
}

/// isZIP_v_undef_Mask - Special case of isZIPMask for canonical form of
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
static bool isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  unsigned NumElts = VT.getVectorNumElements();
  if (NumElts % 2 != 0)
    return false;
  WhichResult = (M[0] == 0 ? 0 : 1);
  unsigned Idx = WhichResult * NumElts / 2;
  for (unsigned i = 0; i != NumElts; i += 2) {
    if ((M[i] >= 0 && (unsigned)M[i] != Idx) ||
        (M[i + 1] >= 0 && (unsigned)M[i + 1] != Idx))
      return false;
    Idx += 1;
  }

  return true;
}

/// isUZP_v_undef_Mask - Special case of isUZPMask for canonical form of
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
static bool isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  unsigned Half = VT.getVectorNumElements() / 2;
  WhichResult = (M[0] == 0 ? 0 : 1);
  for (unsigned j = 0; j != 2; ++j) {
    unsigned Idx = WhichResult;
    for (unsigned i = 0; i != Half; ++i) {
      int MIdx = M[i + j * Half];
      if (MIdx >= 0 && (unsigned)MIdx != Idx)
        return false;
      Idx += 2;
    }
  }

  return true;
}

/// isTRN_v_undef_Mask - Special case of isTRNMask for canonical form of
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
static bool isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  unsigned NumElts = VT.getVectorNumElements();
  if (NumElts % 2 != 0)
    return false;
  WhichResult = (M[0] == 0 ? 0 : 1);
  for (unsigned i = 0; i < NumElts; i += 2) {
    if ((M[i] >= 0 && (unsigned)M[i] != i + WhichResult) ||
        (M[i + 1] >= 0 && (unsigned)M[i + 1] != i + WhichResult))
      return false;
  }
  return true;
}

static bool isINSMask(ArrayRef<int> M, int NumInputElements,
                      bool &DstIsLeft, int &Anomaly) {
  if (M.size() != static_cast<size_t>(NumInputElements))
    return false;

  int NumLHSMatch = 0, NumRHSMatch = 0;
  int LastLHSMismatch = -1, LastRHSMismatch = -1;

  for (int i = 0; i < NumInputElements; ++i) {
    if (M[i] == -1) {
      ++NumLHSMatch;
      ++NumRHSMatch;
      continue;
    }

    if (M[i] == i)
      ++NumLHSMatch;
    else
      LastLHSMismatch = i;

    if (M[i] == i + NumInputElements)
      ++NumRHSMatch;
    else
      LastRHSMismatch = i;
  }

  if (NumLHSMatch == NumInputElements - 1) {
    DstIsLeft = true;
    Anomaly = LastLHSMismatch;
    return true;
  } else if (NumRHSMatch == NumInputElements - 1) {
    DstIsLeft = false;
    Anomaly = LastRHSMismatch;
    return true;
  }

  return false;
}

static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
  if (VT.getSizeInBits() != 128)
    return false;

  unsigned NumElts = VT.getVectorNumElements();

  for (int I = 0, E = NumElts / 2; I != E; I++) {
    if (Mask[I] != I)
      return false;
  }

  int Offset = NumElts / 2;
  for (int I = NumElts / 2, E = NumElts; I != E; I++) {
    if (Mask[I] != I + SplitLHS * Offset)
      return false;
  }

  return true;
}

static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue V0 = Op.getOperand(0);
  SDValue V1 = Op.getOperand(1);
  ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();

  if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
      VT.getVectorElementType() != V1.getValueType().getVectorElementType())
    return SDValue();

  bool SplitV0 = V0.getValueSizeInBits() == 128;

  if (!isConcatMask(Mask, VT, SplitV0))
    return SDValue();

  EVT CastVT = VT.getHalfNumVectorElementsVT(*DAG.getContext());
  if (SplitV0) {
    V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
                     DAG.getConstant(0, DL, MVT::i64));
  }
  if (V1.getValueSizeInBits() == 128) {
    V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
                     DAG.getConstant(0, DL, MVT::i64));
  }
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
}

/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
/// the specified operations to build the shuffle. ID is the perfect-shuffle
//ID, V1 and V2 are the original shuffle inputs. PFEntry is the Perfect shuffle
//table entry and LHS/RHS are the immediate inputs for this stage of the
//shuffle.
static SDValue GeneratePerfectShuffle(unsigned ID, SDValue V1, SDValue V2,
                                      unsigned PFEntry, SDValue LHS,
                                      SDValue RHS, SelectionDAG &DAG,
                                      const SDLoc &DL) {
  unsigned OpNum = (PFEntry >> 26) & 0x0F;
  unsigned LHSID = (PFEntry >> 13) & ((1 << 13) - 1);
  unsigned RHSID = (PFEntry >> 0) & ((1 << 13) - 1);

  enum {
    OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
    OP_VREV,
    OP_VDUP0,
    OP_VDUP1,
    OP_VDUP2,
    OP_VDUP3,
    OP_VEXT1,
    OP_VEXT2,
    OP_VEXT3,
    OP_VUZPL,  // VUZP, left result
    OP_VUZPR,  // VUZP, right result
    OP_VZIPL,  // VZIP, left result
    OP_VZIPR,  // VZIP, right result
    OP_VTRNL,  // VTRN, left result
    OP_VTRNR,  // VTRN, right result
    OP_MOVLANE // Move lane. RHSID is the lane to move into
  };

  if (OpNum == OP_COPY) {
    if (LHSID == (1 * 9 + 2) * 9 + 3)
      return LHS;
    assert(LHSID == ((4 * 9 + 5) * 9 + 6) * 9 + 7 && "Illegal OP_COPY!");
    return RHS;
  }

  if (OpNum == OP_MOVLANE) {
    // Decompose a PerfectShuffle ID to get the Mask for lane Elt
    auto getPFIDLane = [](unsigned ID, int Elt) -> int {
      assert(Elt < 4 && "Expected Perfect Lanes to be less than 4");
      Elt = 3 - Elt;
      while (Elt > 0) {
        ID /= 9;
        Elt--;
      }
      return (ID % 9 == 8) ? -1 : ID % 9;
    };

    // For OP_MOVLANE shuffles, the RHSID represents the lane to move into. We
    // get the lane to move from the PFID, which is always from the
    // original vectors (V1 or V2).
    SDValue OpLHS = GeneratePerfectShuffle(
        LHSID, V1, V2, PerfectShuffleTable[LHSID], LHS, RHS, DAG, DL);
    EVT VT = OpLHS.getValueType();
    assert(RHSID < 8 && "Expected a lane index for RHSID!");
    unsigned ExtLane = 0;
    SDValue Input;

    // OP_MOVLANE are either D movs (if bit 0x4 is set) or S movs. D movs
    // convert into a higher type.
    if (RHSID & 0x4) {
      int MaskElt = getPFIDLane(ID, (RHSID & 0x01) << 1) >> 1;
      if (MaskElt == -1)
        MaskElt = (getPFIDLane(ID, ((RHSID & 0x01) << 1) + 1) - 1) >> 1;
      assert(MaskElt >= 0 && "Didn't expect an undef movlane index!");
      ExtLane = MaskElt < 2 ? MaskElt : (MaskElt - 2);
      Input = MaskElt < 2 ? V1 : V2;
      if (VT.getScalarSizeInBits() == 16) {
        Input = DAG.getBitcast(MVT::v2f32, Input);
        OpLHS = DAG.getBitcast(MVT::v2f32, OpLHS);
      } else {
        assert(VT.getScalarSizeInBits() == 32 &&
               "Expected 16 or 32 bit shuffle elements");
        Input = DAG.getBitcast(MVT::v2f64, Input);
        OpLHS = DAG.getBitcast(MVT::v2f64, OpLHS);
      }
    } else {
      int MaskElt = getPFIDLane(ID, RHSID);
      assert(MaskElt >= 0 && "Didn't expect an undef movlane index!");
      ExtLane = MaskElt < 4 ? MaskElt : (MaskElt - 4);
      Input = MaskElt < 4 ? V1 : V2;
      // Be careful about creating illegal types. Use f16 instead of i16.
      if (VT == MVT::v4i16) {
        Input = DAG.getBitcast(MVT::v4f16, Input);
        OpLHS = DAG.getBitcast(MVT::v4f16, OpLHS);
      }
    }
    SDValue Ext = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
                              Input.getValueType().getVectorElementType(),
                              Input, DAG.getVectorIdxConstant(ExtLane, DL));
    SDValue Ins =
        DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, Input.getValueType(), OpLHS,
                    Ext, DAG.getVectorIdxConstant(RHSID & 0x3, DL));
    return DAG.getBitcast(VT, Ins);
  }

  SDValue OpLHS, OpRHS;
  OpLHS = GeneratePerfectShuffle(LHSID, V1, V2, PerfectShuffleTable[LHSID], LHS,
                                 RHS, DAG, DL);
  OpRHS = GeneratePerfectShuffle(RHSID, V1, V2, PerfectShuffleTable[RHSID], LHS,
                                 RHS, DAG, DL);
  EVT VT = OpLHS.getValueType();

  switch (OpNum) {
  default:
    llvm_unreachable("Unknown shuffle opcode!");
  case OP_VREV:
    // VREV divides the vector in half and swaps within the half.
    if (VT.getVectorElementType() == MVT::i32 ||
        VT.getVectorElementType() == MVT::f32)
      return DAG.getNode(AArch64ISD::REV64, DL, VT, OpLHS);
    // vrev <4 x i16> -> REV32
    if (VT.getVectorElementType() == MVT::i16 ||
        VT.getVectorElementType() == MVT::f16 ||
        VT.getVectorElementType() == MVT::bf16)
      return DAG.getNode(AArch64ISD::REV32, DL, VT, OpLHS);
    // vrev <4 x i8> -> REV16
    assert(VT.getVectorElementType() == MVT::i8);
    return DAG.getNode(AArch64ISD::REV16, DL, VT, OpLHS);
  case OP_VDUP0:
  case OP_VDUP1:
  case OP_VDUP2:
  case OP_VDUP3: {
    EVT EltTy = VT.getVectorElementType();
    unsigned Opcode;
    if (EltTy == MVT::i8)
      Opcode = AArch64ISD::DUPLANE8;
    else if (EltTy == MVT::i16 || EltTy == MVT::f16 || EltTy == MVT::bf16)
      Opcode = AArch64ISD::DUPLANE16;
    else if (EltTy == MVT::i32 || EltTy == MVT::f32)
      Opcode = AArch64ISD::DUPLANE32;
    else if (EltTy == MVT::i64 || EltTy == MVT::f64)
      Opcode = AArch64ISD::DUPLANE64;
    else
      llvm_unreachable("Invalid vector element type?");

    if (VT.getSizeInBits() == 64)
      OpLHS = WidenVector(OpLHS, DAG);
    SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, DL, MVT::i64);
    return DAG.getNode(Opcode, DL, VT, OpLHS, Lane);
  }
  case OP_VEXT1:
  case OP_VEXT2:
  case OP_VEXT3: {
    unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
    return DAG.getNode(AArch64ISD::EXT, DL, VT, OpLHS, OpRHS,
                       DAG.getConstant(Imm, DL, MVT::i32));
  }
  case OP_VUZPL:
    return DAG.getNode(AArch64ISD::UZP1, DL, VT, OpLHS, OpRHS);
  case OP_VUZPR:
    return DAG.getNode(AArch64ISD::UZP2, DL, VT, OpLHS, OpRHS);
  case OP_VZIPL:
    return DAG.getNode(AArch64ISD::ZIP1, DL, VT, OpLHS, OpRHS);
  case OP_VZIPR:
    return DAG.getNode(AArch64ISD::ZIP2, DL, VT, OpLHS, OpRHS);
  case OP_VTRNL:
    return DAG.getNode(AArch64ISD::TRN1, DL, VT, OpLHS, OpRHS);
  case OP_VTRNR:
    return DAG.getNode(AArch64ISD::TRN2, DL, VT, OpLHS, OpRHS);
  }
}

static SDValue GenerateTBL(SDValue Op, ArrayRef<int> ShuffleMask,
                           SelectionDAG &DAG) {
  // Check to see if we can use the TBL instruction.
  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);
  SDLoc DL(Op);

  EVT EltVT = Op.getValueType().getVectorElementType();
  unsigned BytesPerElt = EltVT.getSizeInBits() / 8;

  bool Swap = false;
  if (V1.isUndef() || isZerosVector(V1.getNode())) {
    std::swap(V1, V2);
    Swap = true;
  }

  // If the V2 source is undef or zero then we can use a tbl1, as tbl1 will fill
  // out of range values with 0s. We do need to make sure that any out-of-range
  // values are really out-of-range for a v16i8 vector.
  bool IsUndefOrZero = V2.isUndef() || isZerosVector(V2.getNode());
  MVT IndexVT = MVT::v8i8;
  unsigned IndexLen = 8;
  if (Op.getValueSizeInBits() == 128) {
    IndexVT = MVT::v16i8;
    IndexLen = 16;
  }

  SmallVector<SDValue, 8> TBLMask;
  for (int Val : ShuffleMask) {
    for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
      unsigned Offset = Byte + Val * BytesPerElt;
      if (Swap)
        Offset = Offset < IndexLen ? Offset + IndexLen : Offset - IndexLen;
      if (IsUndefOrZero && Offset >= IndexLen)
        Offset = 255;
      TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
    }
  }

  SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
  SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);

  SDValue Shuffle;
  if (IsUndefOrZero) {
    if (IndexLen == 8)
      V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
    Shuffle = DAG.getNode(
        ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
        DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
        DAG.getBuildVector(IndexVT, DL, ArrayRef(TBLMask.data(), IndexLen)));
  } else {
    if (IndexLen == 8) {
      V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
      Shuffle = DAG.getNode(
          ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
          DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
          DAG.getBuildVector(IndexVT, DL, ArrayRef(TBLMask.data(), IndexLen)));
    } else {
      // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
      // cannot currently represent the register constraints on the input
      // table registers.
      //  Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
      //                   DAG.getBuildVector(IndexVT, DL, &TBLMask[0],
      //                   IndexLen));
      Shuffle = DAG.getNode(
          ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
          DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32), V1Cst,
          V2Cst,
          DAG.getBuildVector(IndexVT, DL, ArrayRef(TBLMask.data(), IndexLen)));
    }
  }
  return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
}

static unsigned getDUPLANEOp(EVT EltType) {
  if (EltType == MVT::i8)
    return AArch64ISD::DUPLANE8;
  if (EltType == MVT::i16 || EltType == MVT::f16 || EltType == MVT::bf16)
    return AArch64ISD::DUPLANE16;
  if (EltType == MVT::i32 || EltType == MVT::f32)
    return AArch64ISD::DUPLANE32;
  if (EltType == MVT::i64 || EltType == MVT::f64)
    return AArch64ISD::DUPLANE64;

  llvm_unreachable("Invalid vector element type?");
}

static SDValue constructDup(SDValue V, int Lane, SDLoc DL, EVT VT,
                            unsigned Opcode, SelectionDAG &DAG) {
  // Try to eliminate a bitcasted extract subvector before a DUPLANE.
  auto getScaledOffsetDup = [](SDValue BitCast, int &LaneC, MVT &CastVT) {
    // Match: dup (bitcast (extract_subv X, C)), LaneC
    if (BitCast.getOpcode() != ISD::BITCAST ||
        BitCast.getOperand(0).getOpcode() != ISD::EXTRACT_SUBVECTOR)
      return false;

    // The extract index must align in the destination type. That may not
    // happen if the bitcast is from narrow to wide type.
    SDValue Extract = BitCast.getOperand(0);
    unsigned ExtIdx = Extract.getConstantOperandVal(1);
    unsigned SrcEltBitWidth = Extract.getScalarValueSizeInBits();
    unsigned ExtIdxInBits = ExtIdx * SrcEltBitWidth;
    unsigned CastedEltBitWidth = BitCast.getScalarValueSizeInBits();
    if (ExtIdxInBits % CastedEltBitWidth != 0)
      return false;

    // Can't handle cases where vector size is not 128-bit
    if (!Extract.getOperand(0).getValueType().is128BitVector())
      return false;

    // Update the lane value by offsetting with the scaled extract index.
    LaneC += ExtIdxInBits / CastedEltBitWidth;

    // Determine the casted vector type of the wide vector input.
    // dup (bitcast (extract_subv X, C)), LaneC --> dup (bitcast X), LaneC'
    // Examples:
    // dup (bitcast (extract_subv v2f64 X, 1) to v2f32), 1 --> dup v4f32 X, 3
    // dup (bitcast (extract_subv v16i8 X, 8) to v4i16), 1 --> dup v8i16 X, 5
    unsigned SrcVecNumElts =
        Extract.getOperand(0).getValueSizeInBits() / CastedEltBitWidth;
    CastVT = MVT::getVectorVT(BitCast.getSimpleValueType().getScalarType(),
                              SrcVecNumElts);
    return true;
  };
  MVT CastVT;
  if (getScaledOffsetDup(V, Lane, CastVT)) {
    V = DAG.getBitcast(CastVT, V.getOperand(0).getOperand(0));
  } else if (V.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
             V.getOperand(0).getValueType().is128BitVector()) {
    // The lane is incremented by the index of the extract.
    // Example: dup v2f32 (extract v4f32 X, 2), 1 --> dup v4f32 X, 3
    Lane += V.getConstantOperandVal(1);
    V = V.getOperand(0);
  } else if (V.getOpcode() == ISD::CONCAT_VECTORS) {
    // The lane is decremented if we are splatting from the 2nd operand.
    // Example: dup v4i32 (concat v2i32 X, v2i32 Y), 3 --> dup v4i32 Y, 1
    unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
    Lane -= Idx * VT.getVectorNumElements() / 2;
    V = WidenVector(V.getOperand(Idx), DAG);
  } else if (VT.getSizeInBits() == 64) {
    // Widen the operand to 128-bit register with undef.
    V = WidenVector(V, DAG);
  }
  return DAG.getNode(Opcode, DL, VT, V, DAG.getConstant(Lane, DL, MVT::i64));
}

// Try to widen element type to get a new mask value for a better permutation
// sequence, so that we can use NEON shuffle instructions, such as zip1/2,
// UZP1/2, TRN1/2, REV, INS, etc.
// For example:
//  shufflevector <4 x i32> %a, <4 x i32> %b,
//                <4 x i32> <i32 6, i32 7, i32 2, i32 3>
// is equivalent to:
//  shufflevector <2 x i64> %a, <2 x i64> %b, <2 x i32> <i32 3, i32 1>
// Finally, we can get:
//  mov     v0.d[0], v1.d[1]
static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  EVT ScalarVT = VT.getVectorElementType();
  unsigned ElementSize = ScalarVT.getFixedSizeInBits();
  SDValue V0 = Op.getOperand(0);
  SDValue V1 = Op.getOperand(1);
  ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();

  // If combining adjacent elements, like two i16's -> i32, two i32's -> i64 ...
  // We need to make sure the wider element type is legal. Thus, ElementSize
  // should be not larger than 32 bits, and i1 type should also be excluded.
  if (ElementSize > 32 || ElementSize == 1)
    return SDValue();

  SmallVector<int, 8> NewMask;
  if (widenShuffleMaskElts(Mask, NewMask)) {
    MVT NewEltVT = VT.isFloatingPoint()
                       ? MVT::getFloatingPointVT(ElementSize * 2)
                       : MVT::getIntegerVT(ElementSize * 2);
    MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
    if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
      V0 = DAG.getBitcast(NewVT, V0);
      V1 = DAG.getBitcast(NewVT, V1);
      return DAG.getBitcast(VT,
                            DAG.getVectorShuffle(NewVT, DL, V0, V1, NewMask));
    }
  }

  return SDValue();
}

// Try to fold shuffle (tbl2, tbl2) into a single tbl4.
static SDValue tryToConvertShuffleOfTbl2ToTbl4(SDValue Op,
                                               ArrayRef<int> ShuffleMask,
                                               SelectionDAG &DAG) {
  SDValue Tbl1 = Op->getOperand(0);
  SDValue Tbl2 = Op->getOperand(1);
  SDLoc DL(Op);
  SDValue Tbl2ID =
      DAG.getTargetConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i64);

  EVT VT = Op.getValueType();
  if (Tbl1.getOpcode() != ISD::INTRINSIC_WO_CHAIN ||
      Tbl1.getOperand(0) != Tbl2ID ||
      Tbl2.getOpcode() != ISD::INTRINSIC_WO_CHAIN ||
      Tbl2.getOperand(0) != Tbl2ID)
    return SDValue();

  if (Tbl1.getValueType() != MVT::v16i8 || Tbl2.getValueType() != MVT::v16i8)
    return SDValue();

  SDValue Mask1 = Tbl1.getOperand(3);
  SDValue Mask2 = Tbl2.getOperand(3);
  if (Mask1.getOpcode() != ISD::BUILD_VECTOR ||
      Mask2.getOpcode() != ISD::BUILD_VECTOR)
    return SDValue();

  SmallVector<SDValue, 16> TBLMaskParts(16, SDValue());
  for (unsigned I = 0; I < 16; I++) {
    if (ShuffleMask[I] < 16)
      TBLMaskParts[I] = Mask1.getOperand(ShuffleMask[I]);
    else {
      auto *C = dyn_cast<ConstantSDNode>(Mask2.getOperand(ShuffleMask[I] - 16));
      if (!C)
        return SDValue();
      TBLMaskParts[I] = DAG.getConstant(C->getSExtValue() + 32, DL, MVT::i32);
    }
  }

  SDValue TBLMask = DAG.getBuildVector(VT, DL, TBLMaskParts);
  SDValue ID =
      DAG.getTargetConstant(Intrinsic::aarch64_neon_tbl4, DL, MVT::i64);

  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::v16i8,
                     {ID, Tbl1->getOperand(1), Tbl1->getOperand(2),
                      Tbl2->getOperand(1), Tbl2->getOperand(2), TBLMask});
}

// Baseline legalization for ZERO_EXTEND_VECTOR_INREG will blend-in zeros,
// but we don't have an appropriate instruction,
// so custom-lower it as ZIP1-with-zeros.
SDValue
AArch64TargetLowering::LowerZERO_EXTEND_VECTOR_INREG(SDValue Op,
                                                     SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue SrcOp = Op.getOperand(0);
  EVT SrcVT = SrcOp.getValueType();
  assert(VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits() == 0 &&
         "Unexpected extension factor.");
  unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits();
  // FIXME: support multi-step zipping?
  if (Scale != 2)
    return SDValue();
  SDValue Zeros = DAG.getConstant(0, DL, SrcVT);
  return DAG.getBitcast(VT,
                        DAG.getNode(AArch64ISD::ZIP1, DL, SrcVT, SrcOp, Zeros));
}

SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
                                                   SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();

  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());

  if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()))
    return LowerFixedLengthVECTOR_SHUFFLEToSVE(Op, DAG);

  // Convert shuffles that are directly supported on NEON to target-specific
  // DAG nodes, instead of keeping them as shuffles and matching them again
  // during code selection.  This is more efficient and avoids the possibility
  // of inconsistencies between legalization and selection.
  ArrayRef<int> ShuffleMask = SVN->getMask();

  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);

  assert(V1.getValueType() == VT && "Unexpected VECTOR_SHUFFLE type!");
  assert(ShuffleMask.size() == VT.getVectorNumElements() &&
         "Unexpected VECTOR_SHUFFLE mask size!");

  if (SDValue Res = tryToConvertShuffleOfTbl2ToTbl4(Op, ShuffleMask, DAG))
    return Res;

  if (SVN->isSplat()) {
    int Lane = SVN->getSplatIndex();
    // If this is undef splat, generate it via "just" vdup, if possible.
    if (Lane == -1)
      Lane = 0;

    if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
      return DAG.getNode(AArch64ISD::DUP, DL, V1.getValueType(),
                         V1.getOperand(0));
    // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
    // constant. If so, we can just reference the lane's definition directly.
    if (V1.getOpcode() == ISD::BUILD_VECTOR &&
        !isa<ConstantSDNode>(V1.getOperand(Lane)))
      return DAG.getNode(AArch64ISD::DUP, DL, VT, V1.getOperand(Lane));

    // Otherwise, duplicate from the lane of the input vector.
    unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
    return constructDup(V1, Lane, DL, VT, Opcode, DAG);
  }

  // Check if the mask matches a DUP for a wider element
  for (unsigned LaneSize : {64U, 32U, 16U}) {
    unsigned Lane = 0;
    if (isWideDUPMask(ShuffleMask, VT, LaneSize, Lane)) {
      unsigned Opcode = LaneSize == 64 ? AArch64ISD::DUPLANE64
                                       : LaneSize == 32 ? AArch64ISD::DUPLANE32
                                                        : AArch64ISD::DUPLANE16;
      // Cast V1 to an integer vector with required lane size
      MVT NewEltTy = MVT::getIntegerVT(LaneSize);
      unsigned NewEltCount = VT.getSizeInBits() / LaneSize;
      MVT NewVecTy = MVT::getVectorVT(NewEltTy, NewEltCount);
      V1 = DAG.getBitcast(NewVecTy, V1);
      // Construct the DUP instruction
      V1 = constructDup(V1, Lane, DL, NewVecTy, Opcode, DAG);
      // Cast back to the original type
      return DAG.getBitcast(VT, V1);
    }
  }

  unsigned NumElts = VT.getVectorNumElements();
  unsigned EltSize = VT.getScalarSizeInBits();
  if (isREVMask(ShuffleMask, EltSize, NumElts, 64))
    return DAG.getNode(AArch64ISD::REV64, DL, V1.getValueType(), V1);
  if (isREVMask(ShuffleMask, EltSize, NumElts, 32))
    return DAG.getNode(AArch64ISD::REV32, DL, V1.getValueType(), V1);
  if (isREVMask(ShuffleMask, EltSize, NumElts, 16))
    return DAG.getNode(AArch64ISD::REV16, DL, V1.getValueType(), V1);

  if (((NumElts == 8 && EltSize == 16) || (NumElts == 16 && EltSize == 8)) &&
      ShuffleVectorInst::isReverseMask(ShuffleMask, ShuffleMask.size())) {
    SDValue Rev = DAG.getNode(AArch64ISD::REV64, DL, VT, V1);
    return DAG.getNode(AArch64ISD::EXT, DL, VT, Rev, Rev,
                       DAG.getConstant(8, DL, MVT::i32));
  }

  bool ReverseEXT = false;
  unsigned Imm;
  if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
    if (ReverseEXT)
      std::swap(V1, V2);
    Imm *= getExtFactor(V1);
    return DAG.getNode(AArch64ISD::EXT, DL, V1.getValueType(), V1, V2,
                       DAG.getConstant(Imm, DL, MVT::i32));
  } else if (V2->isUndef() && isSingletonEXTMask(ShuffleMask, VT, Imm)) {
    Imm *= getExtFactor(V1);
    return DAG.getNode(AArch64ISD::EXT, DL, V1.getValueType(), V1, V1,
                       DAG.getConstant(Imm, DL, MVT::i32));
  }

  unsigned WhichResult;
  if (isZIPMask(ShuffleMask, NumElts, WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
    return DAG.getNode(Opc, DL, V1.getValueType(), V1, V2);
  }
  if (isUZPMask(ShuffleMask, NumElts, WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
    return DAG.getNode(Opc, DL, V1.getValueType(), V1, V2);
  }
  if (isTRNMask(ShuffleMask, NumElts, WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
    return DAG.getNode(Opc, DL, V1.getValueType(), V1, V2);
  }

  if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
    return DAG.getNode(Opc, DL, V1.getValueType(), V1, V1);
  }
  if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
    return DAG.getNode(Opc, DL, V1.getValueType(), V1, V1);
  }
  if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
    return DAG.getNode(Opc, DL, V1.getValueType(), V1, V1);
  }

  if (SDValue Concat = tryFormConcatFromShuffle(Op, DAG))
    return Concat;

  bool DstIsLeft;
  int Anomaly;
  int NumInputElements = V1.getValueType().getVectorNumElements();
  if (isINSMask(ShuffleMask, NumInputElements, DstIsLeft, Anomaly)) {
    SDValue DstVec = DstIsLeft ? V1 : V2;
    SDValue DstLaneV = DAG.getConstant(Anomaly, DL, MVT::i64);

    SDValue SrcVec = V1;
    int SrcLane = ShuffleMask[Anomaly];
    if (SrcLane >= NumInputElements) {
      SrcVec = V2;
      SrcLane -= NumElts;
    }
    SDValue SrcLaneV = DAG.getConstant(SrcLane, DL, MVT::i64);

    EVT ScalarVT = VT.getVectorElementType();

    if (ScalarVT.getFixedSizeInBits() < 32 && ScalarVT.isInteger())
      ScalarVT = MVT::i32;

    return DAG.getNode(
        ISD::INSERT_VECTOR_ELT, DL, VT, DstVec,
        DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarVT, SrcVec, SrcLaneV),
        DstLaneV);
  }

  if (SDValue NewSD = tryWidenMaskForShuffle(Op, DAG))
    return NewSD;

  // If the shuffle is not directly supported and it has 4 elements, use
  // the PerfectShuffle-generated table to synthesize it from other shuffles.
  if (NumElts == 4) {
    unsigned PFIndexes[4];
    for (unsigned i = 0; i != 4; ++i) {
      if (ShuffleMask[i] < 0)
        PFIndexes[i] = 8;
      else
        PFIndexes[i] = ShuffleMask[i];
    }

    // Compute the index in the perfect shuffle table.
    unsigned PFTableIndex = PFIndexes[0] * 9 * 9 * 9 + PFIndexes[1] * 9 * 9 +
                            PFIndexes[2] * 9 + PFIndexes[3];
    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
    return GeneratePerfectShuffle(PFTableIndex, V1, V2, PFEntry, V1, V2, DAG,
                                  DL);
  }

  // Check for a "select shuffle", generating a BSL to pick between lanes in
  // V1/V2.
  if (ShuffleVectorInst::isSelectMask(ShuffleMask, NumElts)) {
    assert(VT.getScalarSizeInBits() <= 32 &&
           "Expected larger vector element sizes to be handled already");
    SmallVector<SDValue> MaskElts;
    for (int M : ShuffleMask)
      MaskElts.push_back(DAG.getConstant(
          M >= static_cast<int>(NumElts) ? 0 : 0xffffffff, DL, MVT::i32));
    EVT IVT = VT.changeVectorElementTypeToInteger();
    SDValue MaskConst = DAG.getBuildVector(IVT, DL, MaskElts);
    return DAG.getBitcast(VT, DAG.getNode(AArch64ISD::BSP, DL, IVT, MaskConst,
                                          DAG.getBitcast(IVT, V1),
                                          DAG.getBitcast(IVT, V2)));
  }

  // Fall back to generating a TBL
  return GenerateTBL(Op, ShuffleMask, DAG);
}

SDValue AArch64TargetLowering::LowerSPLAT_VECTOR(SDValue Op,
                                                 SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();

  if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()))
    return LowerToScalableOp(Op, DAG);

  assert(VT.isScalableVector() && VT.getVectorElementType() == MVT::i1 &&
         "Unexpected vector type!");

  // We can handle the constant cases during isel.
  if (isa<ConstantSDNode>(Op.getOperand(0)))
    return Op;

  // There isn't a natural way to handle the general i1 case, so we use some
  // trickery with whilelo.
  SDLoc DL(Op);
  SDValue SplatVal = DAG.getAnyExtOrTrunc(Op.getOperand(0), DL, MVT::i64);
  SplatVal = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, MVT::i64, SplatVal,
                         DAG.getValueType(MVT::i1));
  SDValue ID =
      DAG.getTargetConstant(Intrinsic::aarch64_sve_whilelo, DL, MVT::i64);
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  if (VT == MVT::nxv1i1)
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::nxv1i1,
                       DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::nxv2i1, ID,
                                   Zero, SplatVal),
                       Zero);
  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT, ID, Zero, SplatVal);
}

SDValue AArch64TargetLowering::LowerDUPQLane(SDValue Op,
                                             SelectionDAG &DAG) const {
  SDLoc DL(Op);

  EVT VT = Op.getValueType();
  if (!isTypeLegal(VT) || !VT.isScalableVector())
    return SDValue();

  // Current lowering only supports the SVE-ACLE types.
  if (VT.getSizeInBits().getKnownMinValue() != AArch64::SVEBitsPerBlock)
    return SDValue();

  // The DUPQ operation is independent of element type so normalise to i64s.
  SDValue Idx128 = Op.getOperand(2);

  // DUPQ can be used when idx is in range.
  auto *CIdx = dyn_cast<ConstantSDNode>(Idx128);
  if (CIdx && (CIdx->getZExtValue() <= 3)) {
    SDValue CI = DAG.getTargetConstant(CIdx->getZExtValue(), DL, MVT::i64);
    return DAG.getNode(AArch64ISD::DUPLANE128, DL, VT, Op.getOperand(1), CI);
  }

  SDValue V = DAG.getNode(ISD::BITCAST, DL, MVT::nxv2i64, Op.getOperand(1));

  // The ACLE says this must produce the same result as:
  //   svtbl(data, svadd_x(svptrue_b64(),
  //                       svand_x(svptrue_b64(), svindex_u64(0, 1), 1),
  //                       index * 2))
  SDValue One = DAG.getConstant(1, DL, MVT::i64);
  SDValue SplatOne = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, One);

  // create the vector 0,1,0,1,...
  SDValue SV = DAG.getStepVector(DL, MVT::nxv2i64);
  SV = DAG.getNode(ISD::AND, DL, MVT::nxv2i64, SV, SplatOne);

  // create the vector idx64,idx64+1,idx64,idx64+1,...
  SDValue Idx64 = DAG.getNode(ISD::ADD, DL, MVT::i64, Idx128, Idx128);
  SDValue SplatIdx64 = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, Idx64);
  SDValue ShuffleMask = DAG.getNode(ISD::ADD, DL, MVT::nxv2i64, SV, SplatIdx64);

  // create the vector Val[idx64],Val[idx64+1],Val[idx64],Val[idx64+1],...
  SDValue TBL = DAG.getNode(AArch64ISD::TBL, DL, MVT::nxv2i64, V, ShuffleMask);
  return DAG.getNode(ISD::BITCAST, DL, VT, TBL);
}


static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
                               APInt &UndefBits) {
  EVT VT = BVN->getValueType(0);
  APInt SplatBits, SplatUndef;
  unsigned SplatBitSize;
  bool HasAnyUndefs;
  if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
    unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;

    for (unsigned i = 0; i < NumSplats; ++i) {
      CnstBits <<= SplatBitSize;
      UndefBits <<= SplatBitSize;
      CnstBits |= SplatBits.zextOrTrunc(VT.getSizeInBits());
      UndefBits |= (SplatBits ^ SplatUndef).zextOrTrunc(VT.getSizeInBits());
    }

    return true;
  }

  return false;
}

// Try 64-bit splatted SIMD immediate.
static SDValue tryAdvSIMDModImm64(unsigned NewOp, SDValue Op, SelectionDAG &DAG,
                                 const APInt &Bits) {
  if (Bits.getHiBits(64) == Bits.getLoBits(64)) {
    uint64_t Value = Bits.zextOrTrunc(64).getZExtValue();
    EVT VT = Op.getValueType();
    MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v2i64 : MVT::f64;

    if (AArch64_AM::isAdvSIMDModImmType10(Value)) {
      Value = AArch64_AM::encodeAdvSIMDModImmType10(Value);

      SDLoc DL(Op);
      SDValue Mov =
          DAG.getNode(NewOp, DL, MovTy, DAG.getConstant(Value, DL, MVT::i32));
      return DAG.getNode(AArch64ISD::NVCAST, DL, VT, Mov);
    }
  }

  return SDValue();
}

// Try 32-bit splatted SIMD immediate.
static SDValue tryAdvSIMDModImm32(unsigned NewOp, SDValue Op, SelectionDAG &DAG,
                                  const APInt &Bits,
                                  const SDValue *LHS = nullptr) {
  EVT VT = Op.getValueType();
  if (VT.isFixedLengthVector() &&
      !DAG.getSubtarget<AArch64Subtarget>().isNeonAvailable())
    return SDValue();

  if (Bits.getHiBits(64) == Bits.getLoBits(64)) {
    uint64_t Value = Bits.zextOrTrunc(64).getZExtValue();
    MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
    bool isAdvSIMDModImm = false;
    uint64_t Shift;

    if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType1(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType1(Value);
      Shift = 0;
    }
    else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType2(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType2(Value);
      Shift = 8;
    }
    else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType3(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType3(Value);
      Shift = 16;
    }
    else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType4(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType4(Value);
      Shift = 24;
    }

    if (isAdvSIMDModImm) {
      SDLoc DL(Op);
      SDValue Mov;

      if (LHS)
        Mov = DAG.getNode(NewOp, DL, MovTy,
                          DAG.getNode(AArch64ISD::NVCAST, DL, MovTy, *LHS),
                          DAG.getConstant(Value, DL, MVT::i32),
                          DAG.getConstant(Shift, DL, MVT::i32));
      else
        Mov =
            DAG.getNode(NewOp, DL, MovTy, DAG.getConstant(Value, DL, MVT::i32),
                        DAG.getConstant(Shift, DL, MVT::i32));

      return DAG.getNode(AArch64ISD::NVCAST, DL, VT, Mov);
    }
  }

  return SDValue();
}

// Try 16-bit splatted SIMD immediate.
static SDValue tryAdvSIMDModImm16(unsigned NewOp, SDValue Op, SelectionDAG &DAG,
                                  const APInt &Bits,
                                  const SDValue *LHS = nullptr) {
  EVT VT = Op.getValueType();
  if (VT.isFixedLengthVector() &&
      !DAG.getSubtarget<AArch64Subtarget>().isNeonAvailable())
    return SDValue();

  if (Bits.getHiBits(64) == Bits.getLoBits(64)) {
    uint64_t Value = Bits.zextOrTrunc(64).getZExtValue();
    MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v8i16 : MVT::v4i16;
    bool isAdvSIMDModImm = false;
    uint64_t Shift;

    if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType5(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType5(Value);
      Shift = 0;
    }
    else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType6(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType6(Value);
      Shift = 8;
    }

    if (isAdvSIMDModImm) {
      SDLoc DL(Op);
      SDValue Mov;

      if (LHS)
        Mov = DAG.getNode(NewOp, DL, MovTy,
                          DAG.getNode(AArch64ISD::NVCAST, DL, MovTy, *LHS),
                          DAG.getConstant(Value, DL, MVT::i32),
                          DAG.getConstant(Shift, DL, MVT::i32));
      else
        Mov =
            DAG.getNode(NewOp, DL, MovTy, DAG.getConstant(Value, DL, MVT::i32),
                        DAG.getConstant(Shift, DL, MVT::i32));

      return DAG.getNode(AArch64ISD::NVCAST, DL, VT, Mov);
    }
  }

  return SDValue();
}

// Try 32-bit splatted SIMD immediate with shifted ones.
static SDValue tryAdvSIMDModImm321s(unsigned NewOp, SDValue Op,
                                    SelectionDAG &DAG, const APInt &Bits) {
  if (Bits.getHiBits(64) == Bits.getLoBits(64)) {
    uint64_t Value = Bits.zextOrTrunc(64).getZExtValue();
    EVT VT = Op.getValueType();
    MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v4i32 : MVT::v2i32;
    bool isAdvSIMDModImm = false;
    uint64_t Shift;

    if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType7(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType7(Value);
      Shift = 264;
    }
    else if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType8(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType8(Value);
      Shift = 272;
    }

    if (isAdvSIMDModImm) {
      SDLoc DL(Op);
      SDValue Mov =
          DAG.getNode(NewOp, DL, MovTy, DAG.getConstant(Value, DL, MVT::i32),
                      DAG.getConstant(Shift, DL, MVT::i32));
      return DAG.getNode(AArch64ISD::NVCAST, DL, VT, Mov);
    }
  }

  return SDValue();
}

// Try 8-bit splatted SIMD immediate.
static SDValue tryAdvSIMDModImm8(unsigned NewOp, SDValue Op, SelectionDAG &DAG,
                                 const APInt &Bits) {
  if (Bits.getHiBits(64) == Bits.getLoBits(64)) {
    uint64_t Value = Bits.zextOrTrunc(64).getZExtValue();
    EVT VT = Op.getValueType();
    MVT MovTy = (VT.getSizeInBits() == 128) ? MVT::v16i8 : MVT::v8i8;

    if (AArch64_AM::isAdvSIMDModImmType9(Value)) {
      Value = AArch64_AM::encodeAdvSIMDModImmType9(Value);

      SDLoc DL(Op);
      SDValue Mov =
          DAG.getNode(NewOp, DL, MovTy, DAG.getConstant(Value, DL, MVT::i32));
      return DAG.getNode(AArch64ISD::NVCAST, DL, VT, Mov);
    }
  }

  return SDValue();
}

// Try FP splatted SIMD immediate.
static SDValue tryAdvSIMDModImmFP(unsigned NewOp, SDValue Op, SelectionDAG &DAG,
                                  const APInt &Bits) {
  if (Bits.getHiBits(64) == Bits.getLoBits(64)) {
    uint64_t Value = Bits.zextOrTrunc(64).getZExtValue();
    EVT VT = Op.getValueType();
    bool isWide = (VT.getSizeInBits() == 128);
    MVT MovTy;
    bool isAdvSIMDModImm = false;

    if ((isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType11(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType11(Value);
      MovTy = isWide ? MVT::v4f32 : MVT::v2f32;
    }
    else if (isWide &&
             (isAdvSIMDModImm = AArch64_AM::isAdvSIMDModImmType12(Value))) {
      Value = AArch64_AM::encodeAdvSIMDModImmType12(Value);
      MovTy = MVT::v2f64;
    }

    if (isAdvSIMDModImm) {
      SDLoc DL(Op);
      SDValue Mov =
          DAG.getNode(NewOp, DL, MovTy, DAG.getConstant(Value, DL, MVT::i32));
      return DAG.getNode(AArch64ISD::NVCAST, DL, VT, Mov);
    }
  }

  return SDValue();
}

// Specialized code to quickly find if PotentialBVec is a BuildVector that
// consists of only the same constant int value, returned in reference arg
// ConstVal
static bool isAllConstantBuildVector(const SDValue &PotentialBVec,
                                     uint64_t &ConstVal) {
  BuildVectorSDNode *Bvec = dyn_cast<BuildVectorSDNode>(PotentialBVec);
  if (!Bvec)
    return false;
  ConstantSDNode *FirstElt = dyn_cast<ConstantSDNode>(Bvec->getOperand(0));
  if (!FirstElt)
    return false;
  EVT VT = Bvec->getValueType(0);
  unsigned NumElts = VT.getVectorNumElements();
  for (unsigned i = 1; i < NumElts; ++i)
    if (dyn_cast<ConstantSDNode>(Bvec->getOperand(i)) != FirstElt)
      return false;
  ConstVal = FirstElt->getZExtValue();
  return true;
}

static bool isAllInactivePredicate(SDValue N) {
  // Look through cast.
  while (N.getOpcode() == AArch64ISD::REINTERPRET_CAST)
    N = N.getOperand(0);

  return ISD::isConstantSplatVectorAllZeros(N.getNode());
}

static bool isAllActivePredicate(SelectionDAG &DAG, SDValue N) {
  unsigned NumElts = N.getValueType().getVectorMinNumElements();

  // Look through cast.
  while (N.getOpcode() == AArch64ISD::REINTERPRET_CAST) {
    N = N.getOperand(0);
    // When reinterpreting from a type with fewer elements the "new" elements
    // are not active, so bail if they're likely to be used.
    if (N.getValueType().getVectorMinNumElements() < NumElts)
      return false;
  }

  if (ISD::isConstantSplatVectorAllOnes(N.getNode()))
    return true;

  // "ptrue p.<ty>, all" can be considered all active when <ty> is the same size
  // or smaller than the implicit element type represented by N.
  // NOTE: A larger element count implies a smaller element type.
  if (N.getOpcode() == AArch64ISD::PTRUE &&
      N.getConstantOperandVal(0) == AArch64SVEPredPattern::all)
    return N.getValueType().getVectorMinNumElements() >= NumElts;

  // If we're compiling for a specific vector-length, we can check if the
  // pattern's VL equals that of the scalable vector at runtime.
  if (N.getOpcode() == AArch64ISD::PTRUE) {
    const auto &Subtarget = DAG.getSubtarget<AArch64Subtarget>();
    unsigned MinSVESize = Subtarget.getMinSVEVectorSizeInBits();
    unsigned MaxSVESize = Subtarget.getMaxSVEVectorSizeInBits();
    if (MaxSVESize && MinSVESize == MaxSVESize) {
      unsigned VScale = MaxSVESize / AArch64::SVEBitsPerBlock;
      unsigned PatNumElts =
          getNumElementsFromSVEPredPattern(N.getConstantOperandVal(0));
      return PatNumElts == (NumElts * VScale);
    }
  }

  return false;
}

// Attempt to form a vector S[LR]I from (or (and X, BvecC1), (lsl Y, C2)),
// to (SLI X, Y, C2), where X and Y have matching vector types, BvecC1 is a
// BUILD_VECTORs with constant element C1, C2 is a constant, and:
//   - for the SLI case: C1 == ~(Ones(ElemSizeInBits) << C2)
//   - for the SRI case: C1 == ~(Ones(ElemSizeInBits) >> C2)
// The (or (lsl Y, C2), (and X, BvecC1)) case is also handled.
static SDValue tryLowerToSLI(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);

  if (!VT.isVector())
    return SDValue();

  SDLoc DL(N);

  SDValue And;
  SDValue Shift;

  SDValue FirstOp = N->getOperand(0);
  unsigned FirstOpc = FirstOp.getOpcode();
  SDValue SecondOp = N->getOperand(1);
  unsigned SecondOpc = SecondOp.getOpcode();

  // Is one of the operands an AND or a BICi? The AND may have been optimised to
  // a BICi in order to use an immediate instead of a register.
  // Is the other operand an shl or lshr? This will have been turned into:
  // AArch64ISD::VSHL vector, #shift or AArch64ISD::VLSHR vector, #shift
  // or (AArch64ISD::SHL_PRED || AArch64ISD::SRL_PRED) mask, vector, #shiftVec.
  if ((FirstOpc == ISD::AND || FirstOpc == AArch64ISD::BICi) &&
      (SecondOpc == AArch64ISD::VSHL || SecondOpc == AArch64ISD::VLSHR ||
       SecondOpc == AArch64ISD::SHL_PRED ||
       SecondOpc == AArch64ISD::SRL_PRED)) {
    And = FirstOp;
    Shift = SecondOp;

  } else if ((SecondOpc == ISD::AND || SecondOpc == AArch64ISD::BICi) &&
             (FirstOpc == AArch64ISD::VSHL || FirstOpc == AArch64ISD::VLSHR ||
              FirstOpc == AArch64ISD::SHL_PRED ||
              FirstOpc == AArch64ISD::SRL_PRED)) {
    And = SecondOp;
    Shift = FirstOp;
  } else
    return SDValue();

  bool IsAnd = And.getOpcode() == ISD::AND;
  bool IsShiftRight = Shift.getOpcode() == AArch64ISD::VLSHR ||
                      Shift.getOpcode() == AArch64ISD::SRL_PRED;
  bool ShiftHasPredOp = Shift.getOpcode() == AArch64ISD::SHL_PRED ||
                        Shift.getOpcode() == AArch64ISD::SRL_PRED;

  // Is the shift amount constant and are all lanes active?
  uint64_t C2;
  if (ShiftHasPredOp) {
    if (!isAllActivePredicate(DAG, Shift.getOperand(0)))
      return SDValue();
    APInt C;
    if (!ISD::isConstantSplatVector(Shift.getOperand(2).getNode(), C))
      return SDValue();
    C2 = C.getZExtValue();
  } else if (ConstantSDNode *C2node =
                 dyn_cast<ConstantSDNode>(Shift.getOperand(1)))
    C2 = C2node->getZExtValue();
  else
    return SDValue();

  APInt C1AsAPInt;
  unsigned ElemSizeInBits = VT.getScalarSizeInBits();
  if (IsAnd) {
    // Is the and mask vector all constant?
    if (!ISD::isConstantSplatVector(And.getOperand(1).getNode(), C1AsAPInt))
      return SDValue();
  } else {
    // Reconstruct the corresponding AND immediate from the two BICi immediates.
    ConstantSDNode *C1nodeImm = dyn_cast<ConstantSDNode>(And.getOperand(1));
    ConstantSDNode *C1nodeShift = dyn_cast<ConstantSDNode>(And.getOperand(2));
    assert(C1nodeImm && C1nodeShift);
    C1AsAPInt = ~(C1nodeImm->getAPIntValue() << C1nodeShift->getAPIntValue());
    C1AsAPInt = C1AsAPInt.zextOrTrunc(ElemSizeInBits);
  }

  // Is C1 == ~(Ones(ElemSizeInBits) << C2) or
  // C1 == ~(Ones(ElemSizeInBits) >> C2), taking into account
  // how much one can shift elements of a particular size?
  if (C2 > ElemSizeInBits)
    return SDValue();

  APInt RequiredC1 = IsShiftRight ? APInt::getHighBitsSet(ElemSizeInBits, C2)
                                  : APInt::getLowBitsSet(ElemSizeInBits, C2);
  if (C1AsAPInt != RequiredC1)
    return SDValue();

  SDValue X = And.getOperand(0);
  SDValue Y = ShiftHasPredOp ? Shift.getOperand(1) : Shift.getOperand(0);
  SDValue Imm = ShiftHasPredOp ? DAG.getTargetConstant(C2, DL, MVT::i32)
                               : Shift.getOperand(1);

  unsigned Inst = IsShiftRight ? AArch64ISD::VSRI : AArch64ISD::VSLI;
  SDValue ResultSLI = DAG.getNode(Inst, DL, VT, X, Y, Imm);

  return ResultSLI;
}

SDValue AArch64TargetLowering::LowerVectorOR(SDValue Op,
                                             SelectionDAG &DAG) const {
  if (useSVEForFixedLengthVectorVT(Op.getValueType(),
                                   !Subtarget->isNeonAvailable()))
    return LowerToScalableOp(Op, DAG);

  // Attempt to form a vector S[LR]I from (or (and X, C1), (lsl Y, C2))
  if (SDValue Res = tryLowerToSLI(Op.getNode(), DAG))
    return Res;

  EVT VT = Op.getValueType();
  if (VT.isScalableVector())
    return Op;

  SDValue LHS = Op.getOperand(0);
  BuildVectorSDNode *BVN =
      dyn_cast<BuildVectorSDNode>(Op.getOperand(1).getNode());
  if (!BVN) {
    // OR commutes, so try swapping the operands.
    LHS = Op.getOperand(1);
    BVN = dyn_cast<BuildVectorSDNode>(Op.getOperand(0).getNode());
  }
  if (!BVN)
    return Op;

  APInt DefBits(VT.getSizeInBits(), 0);
  APInt UndefBits(VT.getSizeInBits(), 0);
  if (resolveBuildVector(BVN, DefBits, UndefBits)) {
    SDValue NewOp;

    if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::ORRi, Op, DAG,
                                    DefBits, &LHS)) ||
        (NewOp = tryAdvSIMDModImm16(AArch64ISD::ORRi, Op, DAG,
                                    DefBits, &LHS)))
      return NewOp;

    if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::ORRi, Op, DAG,
                                    UndefBits, &LHS)) ||
        (NewOp = tryAdvSIMDModImm16(AArch64ISD::ORRi, Op, DAG,
                                    UndefBits, &LHS)))
      return NewOp;
  }

  // We can always fall back to a non-immediate OR.
  return Op;
}

// Normalize the operands of BUILD_VECTOR. The value of constant operands will
// be truncated to fit element width.
static SDValue NormalizeBuildVector(SDValue Op,
                                    SelectionDAG &DAG) {
  assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  EVT EltTy= VT.getVectorElementType();

  if (EltTy.isFloatingPoint() || EltTy.getSizeInBits() > 16)
    return Op;

  SmallVector<SDValue, 16> Ops;
  for (SDValue Lane : Op->ops()) {
    // For integer vectors, type legalization would have promoted the
    // operands already. Otherwise, if Op is a floating-point splat
    // (with operands cast to integers), then the only possibilities
    // are constants and UNDEFs.
    if (auto *CstLane = dyn_cast<ConstantSDNode>(Lane)) {
      Lane = DAG.getConstant(
          CstLane->getAPIntValue().trunc(EltTy.getSizeInBits()).getZExtValue(),
          DL, MVT::i32);
    } else if (Lane.getNode()->isUndef()) {
      Lane = DAG.getUNDEF(MVT::i32);
    } else {
      assert(Lane.getValueType() == MVT::i32 &&
             "Unexpected BUILD_VECTOR operand type");
    }
    Ops.push_back(Lane);
  }
  return DAG.getBuildVector(VT, DL, Ops);
}

static SDValue ConstantBuildVector(SDValue Op, SelectionDAG &DAG,
                                   const AArch64Subtarget *ST) {
  EVT VT = Op.getValueType();
  assert((VT.getSizeInBits() == 64 || VT.getSizeInBits() == 128) &&
         "Expected a legal NEON vector");

  APInt DefBits(VT.getSizeInBits(), 0);
  APInt UndefBits(VT.getSizeInBits(), 0);
  BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
  if (resolveBuildVector(BVN, DefBits, UndefBits)) {
    auto TryMOVIWithBits = [&](APInt DefBits) {
      SDValue NewOp;
      if ((NewOp =
               tryAdvSIMDModImm64(AArch64ISD::MOVIedit, Op, DAG, DefBits)) ||
          (NewOp =
               tryAdvSIMDModImm32(AArch64ISD::MOVIshift, Op, DAG, DefBits)) ||
          (NewOp =
               tryAdvSIMDModImm321s(AArch64ISD::MOVImsl, Op, DAG, DefBits)) ||
          (NewOp =
               tryAdvSIMDModImm16(AArch64ISD::MOVIshift, Op, DAG, DefBits)) ||
          (NewOp = tryAdvSIMDModImm8(AArch64ISD::MOVI, Op, DAG, DefBits)) ||
          (NewOp = tryAdvSIMDModImmFP(AArch64ISD::FMOV, Op, DAG, DefBits)))
        return NewOp;

      APInt NotDefBits = ~DefBits;
      if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::MVNIshift, Op, DAG,
                                      NotDefBits)) ||
          (NewOp = tryAdvSIMDModImm321s(AArch64ISD::MVNImsl, Op, DAG,
                                        NotDefBits)) ||
          (NewOp =
               tryAdvSIMDModImm16(AArch64ISD::MVNIshift, Op, DAG, NotDefBits)))
        return NewOp;
      return SDValue();
    };
    if (SDValue R = TryMOVIWithBits(DefBits))
      return R;
    if (SDValue R = TryMOVIWithBits(UndefBits))
      return R;

    // See if a fneg of the constant can be materialized with a MOVI, etc
    auto TryWithFNeg = [&](APInt DefBits, MVT FVT) {
      // FNegate each sub-element of the constant
      assert(VT.getSizeInBits() % FVT.getScalarSizeInBits() == 0);
      APInt Neg = APInt::getHighBitsSet(FVT.getSizeInBits(), 1)
                      .zext(VT.getSizeInBits());
      APInt NegBits(VT.getSizeInBits(), 0);
      unsigned NumElts = VT.getSizeInBits() / FVT.getScalarSizeInBits();
      for (unsigned i = 0; i < NumElts; i++)
        NegBits |= Neg << (FVT.getScalarSizeInBits() * i);
      NegBits = DefBits ^ NegBits;

      // Try to create the new constants with MOVI, and if so generate a fneg
      // for it.
      if (SDValue NewOp = TryMOVIWithBits(NegBits)) {
        SDLoc DL(Op);
        MVT VFVT = NumElts == 1 ? FVT : MVT::getVectorVT(FVT, NumElts);
        return DAG.getNode(
            AArch64ISD::NVCAST, DL, VT,
            DAG.getNode(ISD::FNEG, DL, VFVT,
                        DAG.getNode(AArch64ISD::NVCAST, DL, VFVT, NewOp)));
      }
      return SDValue();
    };
    SDValue R;
    if ((R = TryWithFNeg(DefBits, MVT::f32)) ||
        (R = TryWithFNeg(DefBits, MVT::f64)) ||
        (ST->hasFullFP16() && (R = TryWithFNeg(DefBits, MVT::f16))))
      return R;
  }

  return SDValue();
}

SDValue AArch64TargetLowering::LowerFixedLengthBuildVectorToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
  auto *BVN = cast<BuildVectorSDNode>(Op);

  if (auto SeqInfo = BVN->isConstantSequence()) {
    SDValue Start = DAG.getConstant(SeqInfo->first, DL, ContainerVT);
    SDValue Steps = DAG.getStepVector(DL, ContainerVT, SeqInfo->second);
    SDValue Seq = DAG.getNode(ISD::ADD, DL, ContainerVT, Start, Steps);
    return convertFromScalableVector(DAG, VT, Seq);
  }

  unsigned NumElems = VT.getVectorNumElements();
  if (!VT.isPow2VectorType() || VT.getFixedSizeInBits() > 128 ||
      NumElems <= 1 || BVN->isConstant())
    return SDValue();

  auto IsExtractElt = [](SDValue Op) {
    return Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT;
  };

  // For integer types that are not already in vectors limit to at most four
  // elements. This is an arbitrary restriction to avoid many fmovs from GPRs.
  if (VT.getScalarType().isInteger() &&
      NumElems - count_if(Op->op_values(), IsExtractElt) > 4)
    return SDValue();

  // Lower (pow2) BUILD_VECTORS that are <= 128-bit to a sequence of ZIP1s.
  SDValue ZeroI64 = DAG.getConstant(0, DL, MVT::i64);
  SmallVector<SDValue, 16> Intermediates = map_to_vector<16>(
      Op->op_values(), [&, Undef = DAG.getUNDEF(ContainerVT)](SDValue Op) {
        return Op.isUndef() ? Undef
                            : DAG.getNode(ISD::INSERT_VECTOR_ELT, DL,
                                          ContainerVT, Undef, Op, ZeroI64);
      });

  ElementCount ZipEC = ContainerVT.getVectorElementCount();
  while (Intermediates.size() > 1) {
    EVT ZipVT = getPackedSVEVectorVT(ZipEC);

    for (unsigned I = 0; I < Intermediates.size(); I += 2) {
      SDValue Op0 = DAG.getBitcast(ZipVT, Intermediates[I + 0]);
      SDValue Op1 = DAG.getBitcast(ZipVT, Intermediates[I + 1]);
      Intermediates[I / 2] =
          Op1.isUndef() ? Op0
                        : DAG.getNode(AArch64ISD::ZIP1, DL, ZipVT, Op0, Op1);
    }

    Intermediates.resize(Intermediates.size() / 2);
    ZipEC = ZipEC.divideCoefficientBy(2);
  }

  assert(Intermediates.size() == 1);
  SDValue Vec = DAG.getBitcast(ContainerVT, Intermediates[0]);
  return convertFromScalableVector(DAG, VT, Vec);
}

SDValue AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op,
                                                 SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();

  bool OverrideNEON = !Subtarget->isNeonAvailable() ||
                      cast<BuildVectorSDNode>(Op)->isConstantSequence();
  if (useSVEForFixedLengthVectorVT(VT, OverrideNEON))
    return LowerFixedLengthBuildVectorToSVE(Op, DAG);

  // Try to build a simple constant vector.
  Op = NormalizeBuildVector(Op, DAG);
  // Thought this might return a non-BUILD_VECTOR (e.g. CONCAT_VECTORS), if so,
  // abort.
  if (Op.getOpcode() != ISD::BUILD_VECTOR)
    return SDValue();

  // Certain vector constants, used to express things like logical NOT and
  // arithmetic NEG, are passed through unmodified.  This allows special
  // patterns for these operations to match, which will lower these constants
  // to whatever is proven necessary.
  BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
  if (BVN->isConstant()) {
    if (ConstantSDNode *Const = BVN->getConstantSplatNode()) {
      unsigned BitSize = VT.getVectorElementType().getSizeInBits();
      APInt Val(BitSize,
                Const->getAPIntValue().zextOrTrunc(BitSize).getZExtValue());
      if (Val.isZero() || (VT.isInteger() && Val.isAllOnes()))
        return Op;
    }
    if (ConstantFPSDNode *Const = BVN->getConstantFPSplatNode())
      if (Const->isZero() && !Const->isNegative())
        return Op;
  }

  if (SDValue V = ConstantBuildVector(Op, DAG, Subtarget))
    return V;

  // Scan through the operands to find some interesting properties we can
  // exploit:
  //   1) If only one value is used, we can use a DUP, or
  //   2) if only the low element is not undef, we can just insert that, or
  //   3) if only one constant value is used (w/ some non-constant lanes),
  //      we can splat the constant value into the whole vector then fill
  //      in the non-constant lanes.
  //   4) FIXME: If different constant values are used, but we can intelligently
  //             select the values we'll be overwriting for the non-constant
  //             lanes such that we can directly materialize the vector
  //             some other way (MOVI, e.g.), we can be sneaky.
  //   5) if all operands are EXTRACT_VECTOR_ELT, check for VUZP.
  SDLoc DL(Op);
  unsigned NumElts = VT.getVectorNumElements();
  bool isOnlyLowElement = true;
  bool usesOnlyOneValue = true;
  bool usesOnlyOneConstantValue = true;
  bool isConstant = true;
  bool AllLanesExtractElt = true;
  unsigned NumConstantLanes = 0;
  unsigned NumDifferentLanes = 0;
  unsigned NumUndefLanes = 0;
  SDValue Value;
  SDValue ConstantValue;
  SmallMapVector<SDValue, unsigned, 16> DifferentValueMap;
  unsigned ConsecutiveValCount = 0;
  SDValue PrevVal;
  for (unsigned i = 0; i < NumElts; ++i) {
    SDValue V = Op.getOperand(i);
    if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      AllLanesExtractElt = false;
    if (V.isUndef()) {
      ++NumUndefLanes;
      continue;
    }
    if (i > 0)
      isOnlyLowElement = false;
    if (!isIntOrFPConstant(V))
      isConstant = false;

    if (isIntOrFPConstant(V)) {
      ++NumConstantLanes;
      if (!ConstantValue.getNode())
        ConstantValue = V;
      else if (ConstantValue != V)
        usesOnlyOneConstantValue = false;
    }

    if (!Value.getNode())
      Value = V;
    else if (V != Value) {
      usesOnlyOneValue = false;
      ++NumDifferentLanes;
    }

    if (PrevVal != V) {
      ConsecutiveValCount = 0;
      PrevVal = V;
    }

    // Keep different values and its last consecutive count. For example,
    //
    //  t22: v16i8 = build_vector t23, t23, t23, t23, t23, t23, t23, t23,
    //                            t24, t24, t24, t24, t24, t24, t24, t24
    //  t23 = consecutive count 8
    //  t24 = consecutive count 8
    // ------------------------------------------------------------------
    //  t22: v16i8 = build_vector t24, t24, t23, t23, t23, t23, t23, t24,
    //                            t24, t24, t24, t24, t24, t24, t24, t24
    //  t23 = consecutive count 5
    //  t24 = consecutive count 9
    DifferentValueMap[V] = ++ConsecutiveValCount;
  }

  if (!Value.getNode()) {
    LLVM_DEBUG(
        dbgs() << "LowerBUILD_VECTOR: value undefined, creating undef node\n");
    return DAG.getUNDEF(VT);
  }

  // Convert BUILD_VECTOR where all elements but the lowest are undef into
  // SCALAR_TO_VECTOR, except for when we have a single-element constant vector
  // as SimplifyDemandedBits will just turn that back into BUILD_VECTOR.
  if (isOnlyLowElement && !(NumElts == 1 && isIntOrFPConstant(Value))) {
    LLVM_DEBUG(dbgs() << "LowerBUILD_VECTOR: only low element used, creating 1 "
                         "SCALAR_TO_VECTOR node\n");
    return DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Value);
  }

  if (AllLanesExtractElt) {
    SDNode *Vector = nullptr;
    bool Even = false;
    bool Odd = false;
    // Check whether the extract elements match the Even pattern <0,2,4,...> or
    // the Odd pattern <1,3,5,...>.
    for (unsigned i = 0; i < NumElts; ++i) {
      SDValue V = Op.getOperand(i);
      const SDNode *N = V.getNode();
      if (!isa<ConstantSDNode>(N->getOperand(1))) {
        Even = false;
        Odd = false;
        break;
      }
      SDValue N0 = N->getOperand(0);

      // All elements are extracted from the same vector.
      if (!Vector) {
        Vector = N0.getNode();
        // Check that the type of EXTRACT_VECTOR_ELT matches the type of
        // BUILD_VECTOR.
        if (VT.getVectorElementType() !=
            N0.getValueType().getVectorElementType())
          break;
      } else if (Vector != N0.getNode()) {
        Odd = false;
        Even = false;
        break;
      }

      // Extracted values are either at Even indices <0,2,4,...> or at Odd
      // indices <1,3,5,...>.
      uint64_t Val = N->getConstantOperandVal(1);
      if (Val == 2 * i) {
        Even = true;
        continue;
      }
      if (Val - 1 == 2 * i) {
        Odd = true;
        continue;
      }

      // Something does not match: abort.
      Odd = false;
      Even = false;
      break;
    }
    if (Even || Odd) {
      SDValue LHS =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SDValue(Vector, 0),
                      DAG.getConstant(0, DL, MVT::i64));
      SDValue RHS =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SDValue(Vector, 0),
                      DAG.getConstant(NumElts, DL, MVT::i64));

      if (Even && !Odd)
        return DAG.getNode(AArch64ISD::UZP1, DL, VT, LHS, RHS);
      if (Odd && !Even)
        return DAG.getNode(AArch64ISD::UZP2, DL, VT, LHS, RHS);
    }
  }

  // Use DUP for non-constant splats. For f32 constant splats, reduce to
  // i32 and try again.
  if (usesOnlyOneValue) {
    if (!isConstant) {
      if (Value.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
          Value.getValueType() != VT) {
        LLVM_DEBUG(
            dbgs() << "LowerBUILD_VECTOR: use DUP for non-constant splats\n");
        return DAG.getNode(AArch64ISD::DUP, DL, VT, Value);
      }

      // This is actually a DUPLANExx operation, which keeps everything vectory.

      SDValue Lane = Value.getOperand(1);
      Value = Value.getOperand(0);
      if (Value.getValueSizeInBits() == 64) {
        LLVM_DEBUG(
            dbgs() << "LowerBUILD_VECTOR: DUPLANE works on 128-bit vectors, "
                      "widening it\n");
        Value = WidenVector(Value, DAG);
      }

      unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
      return DAG.getNode(Opcode, DL, VT, Value, Lane);
    }

    if (VT.getVectorElementType().isFloatingPoint()) {
      SmallVector<SDValue, 8> Ops;
      EVT EltTy = VT.getVectorElementType();
      assert ((EltTy == MVT::f16 || EltTy == MVT::bf16 || EltTy == MVT::f32 ||
               EltTy == MVT::f64) && "Unsupported floating-point vector type");
      LLVM_DEBUG(
          dbgs() << "LowerBUILD_VECTOR: float constant splats, creating int "
                    "BITCASTS, and try again\n");
      MVT NewType = MVT::getIntegerVT(EltTy.getSizeInBits());
      for (unsigned i = 0; i < NumElts; ++i)
        Ops.push_back(DAG.getNode(ISD::BITCAST, DL, NewType, Op.getOperand(i)));
      EVT VecVT = EVT::getVectorVT(*DAG.getContext(), NewType, NumElts);
      SDValue Val = DAG.getBuildVector(VecVT, DL, Ops);
      LLVM_DEBUG(dbgs() << "LowerBUILD_VECTOR: trying to lower new vector: ";
                 Val.dump(););
      Val = LowerBUILD_VECTOR(Val, DAG);
      if (Val.getNode())
        return DAG.getNode(ISD::BITCAST, DL, VT, Val);
    }
  }

  // If we need to insert a small number of different non-constant elements and
  // the vector width is sufficiently large, prefer using DUP with the common
  // value and INSERT_VECTOR_ELT for the different lanes. If DUP is preferred,
  // skip the constant lane handling below.
  bool PreferDUPAndInsert =
      !isConstant && NumDifferentLanes >= 1 &&
      NumDifferentLanes < ((NumElts - NumUndefLanes) / 2) &&
      NumDifferentLanes >= NumConstantLanes;

  // If there was only one constant value used and for more than one lane,
  // start by splatting that value, then replace the non-constant lanes. This
  // is better than the default, which will perform a separate initialization
  // for each lane.
  if (!PreferDUPAndInsert && NumConstantLanes > 0 && usesOnlyOneConstantValue) {
    // Firstly, try to materialize the splat constant.
    SDValue Val = DAG.getSplatBuildVector(VT, DL, ConstantValue);
    unsigned BitSize = VT.getScalarSizeInBits();
    APInt ConstantValueAPInt(1, 0);
    if (auto *C = dyn_cast<ConstantSDNode>(ConstantValue))
      ConstantValueAPInt = C->getAPIntValue().zextOrTrunc(BitSize);
    if (!isNullConstant(ConstantValue) && !isNullFPConstant(ConstantValue) &&
        !ConstantValueAPInt.isAllOnes()) {
      Val = ConstantBuildVector(Val, DAG, Subtarget);
      if (!Val)
        // Otherwise, materialize the constant and splat it.
        Val = DAG.getNode(AArch64ISD::DUP, DL, VT, ConstantValue);
    }

    // Now insert the non-constant lanes.
    for (unsigned i = 0; i < NumElts; ++i) {
      SDValue V = Op.getOperand(i);
      SDValue LaneIdx = DAG.getConstant(i, DL, MVT::i64);
      if (!isIntOrFPConstant(V))
        // Note that type legalization likely mucked about with the VT of the
        // source operand, so we may have to convert it here before inserting.
        Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Val, V, LaneIdx);
    }
    return Val;
  }

  // This will generate a load from the constant pool.
  if (isConstant) {
    LLVM_DEBUG(
        dbgs() << "LowerBUILD_VECTOR: all elements are constant, use default "
                  "expansion\n");
    return SDValue();
  }

  // Detect patterns of a0,a1,a2,a3,b0,b1,b2,b3,c0,c1,c2,c3,d0,d1,d2,d3 from
  // v4i32s. This is really a truncate, which we can construct out of (legal)
  // concats and truncate nodes.
  if (SDValue M = ReconstructTruncateFromBuildVector(Op, DAG))
    return M;

  // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
  if (NumElts >= 4) {
    if (SDValue Shuffle = ReconstructShuffle(Op, DAG))
      return Shuffle;

    if (SDValue Shuffle = ReconstructShuffleWithRuntimeMask(Op, DAG))
      return Shuffle;
  }

  if (PreferDUPAndInsert) {
    // First, build a constant vector with the common element.
    SmallVector<SDValue, 8> Ops(NumElts, Value);
    SDValue NewVector = LowerBUILD_VECTOR(DAG.getBuildVector(VT, DL, Ops), DAG);
    // Next, insert the elements that do not match the common value.
    for (unsigned I = 0; I < NumElts; ++I)
      if (Op.getOperand(I) != Value)
        NewVector =
            DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NewVector,
                        Op.getOperand(I), DAG.getConstant(I, DL, MVT::i64));

    return NewVector;
  }

  // If vector consists of two different values, try to generate two DUPs and
  // (CONCAT_VECTORS or VECTOR_SHUFFLE).
  if (DifferentValueMap.size() == 2 && NumUndefLanes == 0) {
    SmallVector<SDValue, 2> Vals;
    // Check the consecutive count of the value is the half number of vector
    // elements. In this case, we can use CONCAT_VECTORS. For example,
    //
    // canUseVECTOR_CONCAT = true;
    //  t22: v16i8 = build_vector t23, t23, t23, t23, t23, t23, t23, t23,
    //                            t24, t24, t24, t24, t24, t24, t24, t24
    //
    // canUseVECTOR_CONCAT = false;
    //  t22: v16i8 = build_vector t23, t23, t23, t23, t23, t24, t24, t24,
    //                            t24, t24, t24, t24, t24, t24, t24, t24
    bool canUseVECTOR_CONCAT = true;
    for (auto Pair : DifferentValueMap) {
      // Check different values have same length which is NumElts / 2.
      if (Pair.second != NumElts / 2)
        canUseVECTOR_CONCAT = false;
      Vals.push_back(Pair.first);
    }

    // If canUseVECTOR_CONCAT is true, we can generate two DUPs and
    // CONCAT_VECTORs. For example,
    //
    //  t22: v16i8 = BUILD_VECTOR t23, t23, t23, t23, t23, t23, t23, t23,
    //                            t24, t24, t24, t24, t24, t24, t24, t24
    // ==>
    //    t26: v8i8 = AArch64ISD::DUP t23
    //    t28: v8i8 = AArch64ISD::DUP t24
    //  t29: v16i8 = concat_vectors t26, t28
    if (canUseVECTOR_CONCAT) {
      EVT SubVT = VT.getHalfNumVectorElementsVT(*DAG.getContext());
      if (isTypeLegal(SubVT) && SubVT.isVector() &&
          SubVT.getVectorNumElements() >= 2) {
        SmallVector<SDValue, 8> Ops1(NumElts / 2, Vals[0]);
        SmallVector<SDValue, 8> Ops2(NumElts / 2, Vals[1]);
        SDValue DUP1 =
            LowerBUILD_VECTOR(DAG.getBuildVector(SubVT, DL, Ops1), DAG);
        SDValue DUP2 =
            LowerBUILD_VECTOR(DAG.getBuildVector(SubVT, DL, Ops2), DAG);
        SDValue CONCAT_VECTORS =
            DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, DUP1, DUP2);
        return CONCAT_VECTORS;
      }
    }

    // Let's try to generate VECTOR_SHUFFLE. For example,
    //
    //  t24: v8i8 = BUILD_VECTOR t25, t25, t25, t25, t26, t26, t26, t26
    //  ==>
    //    t27: v8i8 = BUILD_VECTOR t26, t26, t26, t26, t26, t26, t26, t26
    //    t28: v8i8 = BUILD_VECTOR t25, t25, t25, t25, t25, t25, t25, t25
    //  t29: v8i8 = vector_shuffle<0,1,2,3,12,13,14,15> t27, t28
    if (NumElts >= 8) {
      SmallVector<int, 16> MaskVec;
      // Build mask for VECTOR_SHUFLLE.
      SDValue FirstLaneVal = Op.getOperand(0);
      for (unsigned i = 0; i < NumElts; ++i) {
        SDValue Val = Op.getOperand(i);
        if (FirstLaneVal == Val)
          MaskVec.push_back(i);
        else
          MaskVec.push_back(i + NumElts);
      }

      SmallVector<SDValue, 8> Ops1(NumElts, Vals[0]);
      SmallVector<SDValue, 8> Ops2(NumElts, Vals[1]);
      SDValue VEC1 = DAG.getBuildVector(VT, DL, Ops1);
      SDValue VEC2 = DAG.getBuildVector(VT, DL, Ops2);
      SDValue VECTOR_SHUFFLE =
          DAG.getVectorShuffle(VT, DL, VEC1, VEC2, MaskVec);
      return VECTOR_SHUFFLE;
    }
  }

  // If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
  // know the default expansion would otherwise fall back on something even
  // worse. For a vector with one or two non-undef values, that's
  // scalar_to_vector for the elements followed by a shuffle (provided the
  // shuffle is valid for the target) and materialization element by element
  // on the stack followed by a load for everything else.
  if (!isConstant && !usesOnlyOneValue) {
    LLVM_DEBUG(
        dbgs() << "LowerBUILD_VECTOR: alternatives failed, creating sequence "
                  "of INSERT_VECTOR_ELT\n");

    SDValue Vec = DAG.getUNDEF(VT);
    SDValue Op0 = Op.getOperand(0);
    unsigned i = 0;

    // Use SCALAR_TO_VECTOR for lane zero to
    // a) Avoid a RMW dependency on the full vector register, and
    // b) Allow the register coalescer to fold away the copy if the
    //    value is already in an S or D register, and we're forced to emit an
    //    INSERT_SUBREG that we can't fold anywhere.
    //
    // We also allow types like i8 and i16 which are illegal scalar but legal
    // vector element types. After type-legalization the inserted value is
    // extended (i32) and it is safe to cast them to the vector type by ignoring
    // the upper bits of the lowest lane (e.g. v8i8, v4i16).
    if (!Op0.isUndef()) {
      LLVM_DEBUG(dbgs() << "Creating node for op0, it is not undefined:\n");
      Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Op0);
      ++i;
    }
    LLVM_DEBUG({
      if (i < NumElts)
        dbgs() << "Creating nodes for the other vector elements:\n";
    });
    for (; i < NumElts; ++i) {
      SDValue V = Op.getOperand(i);
      if (V.isUndef())
        continue;
      SDValue LaneIdx = DAG.getConstant(i, DL, MVT::i64);
      Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, Vec, V, LaneIdx);
    }
    return Vec;
  }

  LLVM_DEBUG(
      dbgs() << "LowerBUILD_VECTOR: use default expansion, failed to find "
                "better alternative\n");
  return SDValue();
}

SDValue AArch64TargetLowering::LowerCONCAT_VECTORS(SDValue Op,
                                                   SelectionDAG &DAG) const {
  if (useSVEForFixedLengthVectorVT(Op.getValueType(),
                                   !Subtarget->isNeonAvailable()))
    return LowerFixedLengthConcatVectorsToSVE(Op, DAG);

  assert(Op.getValueType().isScalableVector() &&
         isTypeLegal(Op.getValueType()) &&
         "Expected legal scalable vector type!");

  if (isTypeLegal(Op.getOperand(0).getValueType())) {
    unsigned NumOperands = Op->getNumOperands();
    assert(NumOperands > 1 && isPowerOf2_32(NumOperands) &&
           "Unexpected number of operands in CONCAT_VECTORS");

    if (NumOperands == 2)
      return Op;

    // Concat each pair of subvectors and pack into the lower half of the array.
    SmallVector<SDValue> ConcatOps(Op->ops());
    while (ConcatOps.size() > 1) {
      for (unsigned I = 0, E = ConcatOps.size(); I != E; I += 2) {
        SDValue V1 = ConcatOps[I];
        SDValue V2 = ConcatOps[I + 1];
        EVT SubVT = V1.getValueType();
        EVT PairVT = SubVT.getDoubleNumVectorElementsVT(*DAG.getContext());
        ConcatOps[I / 2] =
            DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(Op), PairVT, V1, V2);
      }
      ConcatOps.resize(ConcatOps.size() / 2);
    }
    return ConcatOps[0];
  }

  return SDValue();
}

SDValue AArch64TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
                                                      SelectionDAG &DAG) const {
  assert(Op.getOpcode() == ISD::INSERT_VECTOR_ELT && "Unknown opcode!");

  if (useSVEForFixedLengthVectorVT(Op.getValueType(),
                                   !Subtarget->isNeonAvailable()))
    return LowerFixedLengthInsertVectorElt(Op, DAG);

  EVT VT = Op.getOperand(0).getValueType();

  if (VT.getScalarType() == MVT::i1) {
    EVT VectorVT = getPromotedVTForPredicate(VT);
    SDLoc DL(Op);
    SDValue ExtendedVector =
        DAG.getAnyExtOrTrunc(Op.getOperand(0), DL, VectorVT);
    SDValue ExtendedValue =
        DAG.getAnyExtOrTrunc(Op.getOperand(1), DL,
                             VectorVT.getScalarType().getSizeInBits() < 32
                                 ? MVT::i32
                                 : VectorVT.getScalarType());
    ExtendedVector =
        DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VectorVT, ExtendedVector,
                    ExtendedValue, Op.getOperand(2));
    return DAG.getAnyExtOrTrunc(ExtendedVector, DL, VT);
  }

  // Check for non-constant or out of range lane.
  ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(2));
  if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
    return SDValue();

  return Op;
}

SDValue
AArch64TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
                                               SelectionDAG &DAG) const {
  assert(Op.getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Unknown opcode!");
  EVT VT = Op.getOperand(0).getValueType();

  if (VT.getScalarType() == MVT::i1) {
    // We can't directly extract from an SVE predicate; extend it first.
    // (This isn't the only possible lowering, but it's straightforward.)
    EVT VectorVT = getPromotedVTForPredicate(VT);
    SDLoc DL(Op);
    SDValue Extend =
        DAG.getNode(ISD::ANY_EXTEND, DL, VectorVT, Op.getOperand(0));
    MVT ExtractTy = VectorVT == MVT::nxv2i64 ? MVT::i64 : MVT::i32;
    SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtractTy,
                                  Extend, Op.getOperand(1));
    return DAG.getAnyExtOrTrunc(Extract, DL, Op.getValueType());
  }

  if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()))
    return LowerFixedLengthExtractVectorElt(Op, DAG);

  // Check for non-constant or out of range lane.
  ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Op.getOperand(1));
  if (!CI || CI->getZExtValue() >= VT.getVectorNumElements())
    return SDValue();

  // Insertion/extraction are legal for V128 types.
  if (VT == MVT::v16i8 || VT == MVT::v8i16 || VT == MVT::v4i32 ||
      VT == MVT::v2i64 || VT == MVT::v4f32 || VT == MVT::v2f64 ||
      VT == MVT::v8f16 || VT == MVT::v8bf16)
    return Op;

  if (VT != MVT::v8i8 && VT != MVT::v4i16 && VT != MVT::v2i32 &&
      VT != MVT::v1i64 && VT != MVT::v2f32 && VT != MVT::v4f16 &&
      VT != MVT::v4bf16)
    return SDValue();

  // For V64 types, we perform extraction by expanding the value
  // to a V128 type and perform the extraction on that.
  SDLoc DL(Op);
  SDValue WideVec = WidenVector(Op.getOperand(0), DAG);
  EVT WideTy = WideVec.getValueType();

  EVT ExtrTy = WideTy.getVectorElementType();
  if (ExtrTy == MVT::i16 || ExtrTy == MVT::i8)
    ExtrTy = MVT::i32;

  // For extractions, we just return the result directly.
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ExtrTy, WideVec,
                     Op.getOperand(1));
}

SDValue AArch64TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
                                                      SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() &&
         "Only cases that extract a fixed length vector are supported!");
  EVT InVT = Op.getOperand(0).getValueType();

  // If we don't have legal types yet, do nothing
  if (!isTypeLegal(InVT))
    return SDValue();

  if (InVT.is128BitVector()) {
    assert(VT.is64BitVector() && "Extracting unexpected vector type!");
    unsigned Idx = Op.getConstantOperandVal(1);

    // This will get lowered to an appropriate EXTRACT_SUBREG in ISel.
    if (Idx == 0)
      return Op;

    // If this is extracting the upper 64-bits of a 128-bit vector, we match
    // that directly.
    if (Idx * InVT.getScalarSizeInBits() == 64 && Subtarget->isNeonAvailable())
      return Op;
  }

  if (InVT.isScalableVector() ||
      useSVEForFixedLengthVectorVT(InVT, !Subtarget->isNeonAvailable())) {
    SDLoc DL(Op);
    SDValue Vec = Op.getOperand(0);
    SDValue Idx = Op.getOperand(1);

    EVT PackedVT = getPackedSVEVectorVT(InVT.getVectorElementType());
    if (PackedVT != InVT) {
      // Pack input into the bottom part of an SVE register and try again.
      SDValue Container = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, PackedVT,
                                      DAG.getUNDEF(PackedVT), Vec,
                                      DAG.getVectorIdxConstant(0, DL));
      return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Container, Idx);
    }

    // This will get matched by custom code during ISelDAGToDAG.
    if (isNullConstant(Idx))
      return Op;

    assert(InVT.isScalableVector() && "Unexpected vector type!");
    // Move requested subvector to the start of the vector and try again.
    SDValue Splice = DAG.getNode(ISD::VECTOR_SPLICE, DL, InVT, Vec, Vec, Idx);
    return convertFromScalableVector(DAG, VT, Splice);
  }

  return SDValue();
}

SDValue AArch64TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op,
                                                     SelectionDAG &DAG) const {
  assert(Op.getValueType().isScalableVector() &&
         "Only expect to lower inserts into scalable vectors!");

  EVT InVT = Op.getOperand(1).getValueType();
  unsigned Idx = Op.getConstantOperandVal(2);

  SDValue Vec0 = Op.getOperand(0);
  SDValue Vec1 = Op.getOperand(1);
  SDLoc DL(Op);
  EVT VT = Op.getValueType();

  if (InVT.isScalableVector()) {
    if (!isTypeLegal(VT))
      return SDValue();

    // Break down insert_subvector into simpler parts.
    if (VT.getVectorElementType() == MVT::i1) {
      unsigned NumElts = VT.getVectorMinNumElements();
      EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext());

      SDValue Lo, Hi;
      Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, Vec0,
                       DAG.getVectorIdxConstant(0, DL));
      Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, Vec0,
                       DAG.getVectorIdxConstant(NumElts / 2, DL));
      if (Idx < (NumElts / 2))
        Lo = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, HalfVT, Lo, Vec1,
                         DAG.getVectorIdxConstant(Idx, DL));
      else
        Hi = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, HalfVT, Hi, Vec1,
                         DAG.getVectorIdxConstant(Idx - (NumElts / 2), DL));

      return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
    }

    // We can select these directly.
    if (isTypeLegal(InVT) && Vec0.isUndef())
      return Op;

    // Ensure the subvector is half the size of the main vector.
    if (VT.getVectorElementCount() != (InVT.getVectorElementCount() * 2))
      return SDValue();

    // Here narrow and wide refers to the vector element types. After "casting"
    // both vectors must have the same bit length and so because the subvector
    // has fewer elements, those elements need to be bigger.
    EVT NarrowVT = getPackedSVEVectorVT(VT.getVectorElementCount());
    EVT WideVT = getPackedSVEVectorVT(InVT.getVectorElementCount());

    // NOP cast operands to the largest legal vector of the same element count.
    if (VT.isFloatingPoint()) {
      Vec0 = getSVESafeBitCast(NarrowVT, Vec0, DAG);
      Vec1 = getSVESafeBitCast(NarrowVT, Vec1, DAG);
    } else {
      // Legal integer vectors are already their largest so Vec0 is fine as is.
      Vec1 = DAG.getNode(ISD::ANY_EXTEND, DL, WideVT, Vec1);
      Vec1 = DAG.getNode(AArch64ISD::NVCAST, DL, NarrowVT, Vec1);
    }

    // To replace the top/bottom half of vector V with vector SubV we widen the
    // preserved half of V, concatenate this to SubV (the order depending on the
    // half being replaced) and then narrow the result.
    SDValue Narrow;
    if (Idx == 0) {
      SDValue HiVec0 = DAG.getNode(AArch64ISD::UUNPKHI, DL, WideVT, Vec0);
      HiVec0 = DAG.getNode(AArch64ISD::NVCAST, DL, NarrowVT, HiVec0);
      Narrow = DAG.getNode(AArch64ISD::UZP1, DL, NarrowVT, Vec1, HiVec0);
    } else {
      assert(Idx == InVT.getVectorMinNumElements() &&
             "Invalid subvector index!");
      SDValue LoVec0 = DAG.getNode(AArch64ISD::UUNPKLO, DL, WideVT, Vec0);
      LoVec0 = DAG.getNode(AArch64ISD::NVCAST, DL, NarrowVT, LoVec0);
      Narrow = DAG.getNode(AArch64ISD::UZP1, DL, NarrowVT, LoVec0, Vec1);
    }

    return getSVESafeBitCast(VT, Narrow, DAG);
  }

  if (Idx == 0 && isPackedVectorType(VT, DAG)) {
    // This will be matched by custom code during ISelDAGToDAG.
    if (Vec0.isUndef())
      return Op;

    std::optional<unsigned> PredPattern =
        getSVEPredPatternFromNumElements(InVT.getVectorNumElements());
    auto PredTy = VT.changeVectorElementType(MVT::i1);
    SDValue PTrue = getPTrue(DAG, DL, PredTy, *PredPattern);
    SDValue ScalableVec1 = convertToScalableVector(DAG, VT, Vec1);
    return DAG.getNode(ISD::VSELECT, DL, VT, PTrue, ScalableVec1, Vec0);
  }

  return SDValue();
}

static bool isPow2Splat(SDValue Op, uint64_t &SplatVal, bool &Negated) {
  if (Op.getOpcode() != AArch64ISD::DUP &&
      Op.getOpcode() != ISD::SPLAT_VECTOR &&
      Op.getOpcode() != ISD::BUILD_VECTOR)
    return false;

  if (Op.getOpcode() == ISD::BUILD_VECTOR &&
      !isAllConstantBuildVector(Op, SplatVal))
    return false;

  if (Op.getOpcode() != ISD::BUILD_VECTOR &&
      !isa<ConstantSDNode>(Op->getOperand(0)))
    return false;

  SplatVal = Op->getConstantOperandVal(0);
  if (Op.getValueType().getVectorElementType() != MVT::i64)
    SplatVal = (int32_t)SplatVal;

  Negated = false;
  if (isPowerOf2_64(SplatVal))
    return true;

  Negated = true;
  if (isPowerOf2_64(-SplatVal)) {
    SplatVal = -SplatVal;
    return true;
  }

  return false;
}

SDValue AArch64TargetLowering::LowerDIV(SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);

  if (useSVEForFixedLengthVectorVT(VT, /*OverrideNEON=*/true))
    return LowerFixedLengthVectorIntDivideToSVE(Op, DAG);

  assert(VT.isScalableVector() && "Expected a scalable vector.");

  bool Signed = Op.getOpcode() == ISD::SDIV;
  unsigned PredOpcode = Signed ? AArch64ISD::SDIV_PRED : AArch64ISD::UDIV_PRED;

  bool Negated;
  uint64_t SplatVal;
  if (Signed && isPow2Splat(Op.getOperand(1), SplatVal, Negated)) {
    SDValue Pg = getPredicateForScalableVector(DAG, DL, VT);
    SDValue Res =
        DAG.getNode(AArch64ISD::SRAD_MERGE_OP1, DL, VT, Pg, Op->getOperand(0),
                    DAG.getTargetConstant(Log2_64(SplatVal), DL, MVT::i32));
    if (Negated)
      Res = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT), Res);

    return Res;
  }

  if (VT == MVT::nxv4i32 || VT == MVT::nxv2i64)
    return LowerToPredicatedOp(Op, DAG, PredOpcode);

  // SVE doesn't have i8 and i16 DIV operations; widen them to 32-bit
  // operations, and truncate the result.
  EVT WidenedVT;
  if (VT == MVT::nxv16i8)
    WidenedVT = MVT::nxv8i16;
  else if (VT == MVT::nxv8i16)
    WidenedVT = MVT::nxv4i32;
  else
    llvm_unreachable("Unexpected Custom DIV operation");

  unsigned UnpkLo = Signed ? AArch64ISD::SUNPKLO : AArch64ISD::UUNPKLO;
  unsigned UnpkHi = Signed ? AArch64ISD::SUNPKHI : AArch64ISD::UUNPKHI;
  SDValue Op0Lo = DAG.getNode(UnpkLo, DL, WidenedVT, Op.getOperand(0));
  SDValue Op1Lo = DAG.getNode(UnpkLo, DL, WidenedVT, Op.getOperand(1));
  SDValue Op0Hi = DAG.getNode(UnpkHi, DL, WidenedVT, Op.getOperand(0));
  SDValue Op1Hi = DAG.getNode(UnpkHi, DL, WidenedVT, Op.getOperand(1));
  SDValue ResultLo = DAG.getNode(Op.getOpcode(), DL, WidenedVT, Op0Lo, Op1Lo);
  SDValue ResultHi = DAG.getNode(Op.getOpcode(), DL, WidenedVT, Op0Hi, Op1Hi);
  SDValue ResultLoCast = DAG.getNode(AArch64ISD::NVCAST, DL, VT, ResultLo);
  SDValue ResultHiCast = DAG.getNode(AArch64ISD::NVCAST, DL, VT, ResultHi);
  return DAG.getNode(AArch64ISD::UZP1, DL, VT, ResultLoCast, ResultHiCast);
}

bool AArch64TargetLowering::shouldExpandBuildVectorWithShuffles(
    EVT VT, unsigned DefinedValues) const {
  if (!Subtarget->isNeonAvailable())
    return false;
  return TargetLowering::shouldExpandBuildVectorWithShuffles(VT, DefinedValues);
}

bool AArch64TargetLowering::isShuffleMaskLegal(ArrayRef<int> M, EVT VT) const {
  // Currently no fixed length shuffles that require SVE are legal.
  if (useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()))
    return false;

  if (VT.getVectorNumElements() == 4 &&
      (VT.is128BitVector() || VT.is64BitVector())) {
    unsigned Cost = getPerfectShuffleCost(M);
    if (Cost <= 1)
      return true;
  }

  bool DummyBool;
  int DummyInt;
  unsigned DummyUnsigned;

  unsigned EltSize = VT.getScalarSizeInBits();
  unsigned NumElts = VT.getVectorNumElements();
  return (ShuffleVectorSDNode::isSplatMask(M) ||
          isREVMask(M, EltSize, NumElts, 64) ||
          isREVMask(M, EltSize, NumElts, 32) ||
          isREVMask(M, EltSize, NumElts, 16) ||
          isEXTMask(M, VT, DummyBool, DummyUnsigned) ||
          isTRNMask(M, NumElts, DummyUnsigned) ||
          isUZPMask(M, NumElts, DummyUnsigned) ||
          isZIPMask(M, NumElts, DummyUnsigned) ||
          isTRN_v_undef_Mask(M, VT, DummyUnsigned) ||
          isUZP_v_undef_Mask(M, VT, DummyUnsigned) ||
          isZIP_v_undef_Mask(M, VT, DummyUnsigned) ||
          isINSMask(M, NumElts, DummyBool, DummyInt) ||
          isConcatMask(M, VT, VT.getSizeInBits() == 128));
}

bool AArch64TargetLowering::isVectorClearMaskLegal(ArrayRef<int> M,
                                                   EVT VT) const {
  // Just delegate to the generic legality, clear masks aren't special.
  return isShuffleMaskLegal(M, VT);
}

/// getVShiftImm - Check if this is a valid build_vector for the immediate
/// operand of a vector shift operation, where all the elements of the
/// build_vector must have the same constant integer value.
static bool getVShiftImm(SDValue Op, unsigned ElementBits, int64_t &Cnt) {
  // Ignore bit_converts.
  while (Op.getOpcode() == ISD::BITCAST)
    Op = Op.getOperand(0);
  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
  APInt SplatBits, SplatUndef;
  unsigned SplatBitSize;
  bool HasAnyUndefs;
  if (!BVN || !BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize,
                                    HasAnyUndefs, ElementBits) ||
      SplatBitSize > ElementBits)
    return false;
  Cnt = SplatBits.getSExtValue();
  return true;
}

/// isVShiftLImm - Check if this is a valid build_vector for the immediate
/// operand of a vector shift left operation.  That value must be in the range:
///   0 <= Value < ElementBits for a left shift; or
///   0 <= Value <= ElementBits for a long left shift.
static bool isVShiftLImm(SDValue Op, EVT VT, bool isLong, int64_t &Cnt) {
  assert(VT.isVector() && "vector shift count is not a vector type");
  int64_t ElementBits = VT.getScalarSizeInBits();
  if (!getVShiftImm(Op, ElementBits, Cnt))
    return false;
  return (Cnt >= 0 && (isLong ? Cnt - 1 : Cnt) < ElementBits);
}

/// isVShiftRImm - Check if this is a valid build_vector for the immediate
/// operand of a vector shift right operation. The value must be in the range:
///   1 <= Value <= ElementBits for a right shift; or
static bool isVShiftRImm(SDValue Op, EVT VT, bool isNarrow, int64_t &Cnt) {
  assert(VT.isVector() && "vector shift count is not a vector type");
  int64_t ElementBits = VT.getScalarSizeInBits();
  if (!getVShiftImm(Op, ElementBits, Cnt))
    return false;
  return (Cnt >= 1 && Cnt <= (isNarrow ? ElementBits / 2 : ElementBits));
}

SDValue AArch64TargetLowering::LowerTRUNCATE(SDValue Op,
                                             SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();

  if (VT.getScalarType() == MVT::i1) {
    // Lower i1 truncate to `(x & 1) != 0`.
    SDLoc DL(Op);
    EVT OpVT = Op.getOperand(0).getValueType();
    SDValue Zero = DAG.getConstant(0, DL, OpVT);
    SDValue One = DAG.getConstant(1, DL, OpVT);
    SDValue And = DAG.getNode(ISD::AND, DL, OpVT, Op.getOperand(0), One);
    return DAG.getSetCC(DL, VT, And, Zero, ISD::SETNE);
  }

  if (!VT.isVector() || VT.isScalableVector())
    return SDValue();

  if (useSVEForFixedLengthVectorVT(Op.getOperand(0).getValueType(),
                                   !Subtarget->isNeonAvailable()))
    return LowerFixedLengthVectorTruncateToSVE(Op, DAG);

  return SDValue();
}

// Check if we can we lower this SRL to a rounding shift instruction. ResVT is
// possibly a truncated type, it tells how many bits of the value are to be
// used.
static bool canLowerSRLToRoundingShiftForVT(SDValue Shift, EVT ResVT,
                                            SelectionDAG &DAG,
                                            unsigned &ShiftValue,
                                            SDValue &RShOperand) {
  if (Shift->getOpcode() != ISD::SRL)
    return false;

  EVT VT = Shift.getValueType();
  assert(VT.isScalableVT());

  auto ShiftOp1 =
      dyn_cast_or_null<ConstantSDNode>(DAG.getSplatValue(Shift->getOperand(1)));
  if (!ShiftOp1)
    return false;

  ShiftValue = ShiftOp1->getZExtValue();
  if (ShiftValue < 1 || ShiftValue > ResVT.getScalarSizeInBits())
    return false;

  SDValue Add = Shift->getOperand(0);
  if (Add->getOpcode() != ISD::ADD || !Add->hasOneUse())
    return false;

  assert(ResVT.getScalarSizeInBits() <= VT.getScalarSizeInBits() &&
         "ResVT must be truncated or same type as the shift.");
  // Check if an overflow can lead to incorrect results.
  uint64_t ExtraBits = VT.getScalarSizeInBits() - ResVT.getScalarSizeInBits();
  if (ShiftValue > ExtraBits && !Add->getFlags().hasNoUnsignedWrap())
    return false;

  auto AddOp1 =
      dyn_cast_or_null<ConstantSDNode>(DAG.getSplatValue(Add->getOperand(1)));
  if (!AddOp1)
    return false;
  uint64_t AddValue = AddOp1->getZExtValue();
  if (AddValue != 1ULL << (ShiftValue - 1))
    return false;

  RShOperand = Add->getOperand(0);
  return true;
}

SDValue AArch64TargetLowering::LowerVectorSRA_SRL_SHL(SDValue Op,
                                                      SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  int64_t Cnt;

  if (!Op.getOperand(1).getValueType().isVector())
    return Op;
  unsigned EltSize = VT.getScalarSizeInBits();

  switch (Op.getOpcode()) {
  case ISD::SHL:
    if (VT.isScalableVector() ||
        useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable()))
      return LowerToPredicatedOp(Op, DAG, AArch64ISD::SHL_PRED);

    if (isVShiftLImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize)
      return DAG.getNode(AArch64ISD::VSHL, DL, VT, Op.getOperand(0),
                         DAG.getConstant(Cnt, DL, MVT::i32));
    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
                       DAG.getConstant(Intrinsic::aarch64_neon_ushl, DL,
                                       MVT::i32),
                       Op.getOperand(0), Op.getOperand(1));
  case ISD::SRA:
  case ISD::SRL:
    if (VT.isScalableVector() &&
        (Subtarget->hasSVE2() ||
         (Subtarget->hasSME() && Subtarget->isStreaming()))) {
      SDValue RShOperand;
      unsigned ShiftValue;
      if (canLowerSRLToRoundingShiftForVT(Op, VT, DAG, ShiftValue, RShOperand))
        return DAG.getNode(AArch64ISD::URSHR_I_PRED, DL, VT,
                           getPredicateForVector(DAG, DL, VT), RShOperand,
                           DAG.getTargetConstant(ShiftValue, DL, MVT::i32));
    }

    if (VT.isScalableVector() ||
        useSVEForFixedLengthVectorVT(VT, !Subtarget->isNeonAvailable())) {
      unsigned Opc = Op.getOpcode() == ISD::SRA ? AArch64ISD::SRA_PRED
                                                : AArch64ISD::SRL_PRED;
      return LowerToPredicatedOp(Op, DAG, Opc);
    }

    // Right shift immediate
    if (isVShiftRImm(Op.getOperand(1), VT, false, Cnt) && Cnt < EltSize) {
      unsigned Opc =
          (Op.getOpcode() == ISD::SRA) ? AArch64ISD::VASHR : AArch64ISD::VLSHR;
      return DAG.getNode(Opc, DL, VT, Op.getOperand(0),
                         DAG.getConstant(Cnt, DL, MVT::i32), Op->getFlags());
    }

    // Right shift register.  Note, there is not a shift right register
    // instruction, but the shift left register instruction takes a signed
    // value, where negative numbers specify a right shift.
    unsigned Opc = (Op.getOpcode() == ISD::SRA) ? Intrinsic::aarch64_neon_sshl
                                                : Intrinsic::aarch64_neon_ushl;
    // negate the shift amount
    SDValue NegShift = DAG.getNode(ISD::SUB, DL, VT, DAG.getConstant(0, DL, VT),
                                   Op.getOperand(1));
    SDValue NegShiftLeft =
        DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
                    DAG.getConstant(Opc, DL, MVT::i32), Op.getOperand(0),
                    NegShift);
    return NegShiftLeft;
  }

  llvm_unreachable("unexpected shift opcode");
}

SDValue AArch64TargetLowering::LowerVSETCC(SDValue Op,
                                           SelectionDAG &DAG) const {
  if (Op.getValueType().isScalableVector())
    return LowerToPredicatedOp(Op, DAG, AArch64ISD::SETCC_MERGE_ZERO);

  if (useSVEForFixedLengthVectorVT(Op.getOperand(0).getValueType(),
                                   !Subtarget->isNeonAvailable()))
    return LowerFixedLengthVectorSetccToSVE(Op, DAG);

  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  EVT CmpVT = LHS.getValueType().changeVectorElementTypeToInteger();
  SDLoc DL(Op);

  if (LHS.getValueType().getVectorElementType().isInteger())
    return Op;

  assert(((!Subtarget->hasFullFP16() &&
           LHS.getValueType().getVectorElementType() != MVT::f16) ||
          LHS.getValueType().getVectorElementType() != MVT::bf16 ||
          LHS.getValueType().getVectorElementType() != MVT::f128) &&
         "Unexpected type!");

  // Lower isnan(x) | isnan(never-nan) to x != x.
  // Lower !isnan(x) & !isnan(never-nan) to x == x.
  if (CC == ISD::SETUO || CC == ISD::SETO) {
    bool OneNaN = false;
    if (LHS == RHS) {
      OneNaN = true;
    } else if (DAG.isKnownNeverNaN(RHS)) {
      OneNaN = true;
      RHS = LHS;
    } else if (DAG.isKnownNeverNaN(LHS)) {
      OneNaN = true;
      LHS = RHS;
    }
    if (OneNaN) {
      CC = CC == ISD::SETUO ? ISD::SETUNE : ISD::SETOEQ;
    }
  }

  // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
  // clean.  Some of them require two branches to implement.
  AArch64CC::CondCode CC1, CC2;
  bool ShouldInvert;
  changeVectorFPCCToAArch64CC(CC, CC1, CC2, ShouldInvert);

  bool NoNaNs =
      getTargetMachine().Options.NoNaNsFPMath || Op->getFlags().hasNoNaNs();
  SDValue Cmp = emitVectorComparison(LHS, RHS, CC1, NoNaNs, CmpVT, DL, DAG);
  if (!Cmp.getNode())
    return SDValue();

  if (CC2 != AArch64CC::AL) {
    SDValue Cmp2 = emitVectorComparison(LHS, RHS, CC2, NoNaNs, CmpVT, DL, DAG);
    if (!Cmp2.getNode())
      return SDValue();

    Cmp = DAG.getNode(ISD::OR, DL, CmpVT, Cmp, Cmp2);
  }

  Cmp = DAG.getSExtOrTrunc(Cmp, DL, Op.getValueType());

  if (ShouldInvert)
    Cmp = DAG.getNOT(DL, Cmp, Cmp.getValueType());

  return Cmp;
}

static SDValue getReductionSDNode(unsigned Op, SDLoc DL, SDValue ScalarOp,
                                  SelectionDAG &DAG) {
  SDValue VecOp = ScalarOp.getOperand(0);
  auto Rdx = DAG.getNode(Op, DL, VecOp.getSimpleValueType(), VecOp);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarOp.getValueType(), Rdx,
                     DAG.getConstant(0, DL, MVT::i64));
}

static SDValue getVectorBitwiseReduce(unsigned Opcode, SDValue Vec, EVT VT,
                                      SDLoc DL, SelectionDAG &DAG) {
  unsigned ScalarOpcode;
  switch (Opcode) {
  case ISD::VECREDUCE_AND:
    ScalarOpcode = ISD::AND;
    break;
  case ISD::VECREDUCE_OR:
    ScalarOpcode = ISD::OR;
    break;
  case ISD::VECREDUCE_XOR:
    ScalarOpcode = ISD::XOR;
    break;
  default:
    llvm_unreachable("Expected bitwise vector reduction");
    return SDValue();
  }

  EVT VecVT = Vec.getValueType();
  assert(VecVT.isFixedLengthVector() && VecVT.isPow2VectorType() &&
         "Expected power-of-2 length vector");

  EVT ElemVT = VecVT.getVectorElementType();

  SDValue Result;
  unsigned NumElems = VecVT.getVectorNumElements();

  // Special case for boolean reductions
  if (ElemVT == MVT::i1) {
    // Split large vectors into smaller ones
    if (NumElems > 16) {
      SDValue Lo, Hi;
      std::tie(Lo, Hi) = DAG.SplitVector(Vec, DL);
      EVT HalfVT = Lo.getValueType();
      SDValue HalfVec = DAG.getNode(ScalarOpcode, DL, HalfVT, Lo, Hi);
      return getVectorBitwiseReduce(Opcode, HalfVec, VT, DL, DAG);
    }

    // Results of setcc operations get widened to 128 bits if their input
    // operands are 128 bits wide, otherwise vectors that are less than 64 bits
    // get widened to neatly fit a 64 bit register, so e.g. <4 x i1> gets
    // lowered to either <4 x i16> or <4 x i32>. Sign extending to this element
    // size leads to the best codegen, since e.g. setcc results might need to be
    // truncated otherwise.
    unsigned ExtendedWidth = 64;
    if (Vec.getOpcode() == ISD::SETCC &&
        Vec.getOperand(0).getValueSizeInBits() >= 128) {
      ExtendedWidth = 128;
    }
    EVT ExtendedVT = MVT::getIntegerVT(std::max(ExtendedWidth / NumElems, 8u));

    // any_ext doesn't work with umin/umax, so only use it for uadd.
    unsigned ExtendOp =
        ScalarOpcode == ISD::XOR ? ISD::ANY_EXTEND : ISD::SIGN_EXTEND;
    SDValue Extended = DAG.getNode(
        ExtendOp, DL, VecVT.changeVectorElementType(ExtendedVT), Vec);
    // The uminp/uminv and umaxp/umaxv instructions don't have .2d variants, so
    // in that case we bitcast the sign extended values from v2i64 to v4i32
    // before reduction for optimal code generation.
    if ((ScalarOpcode == ISD::AND || ScalarOpcode == ISD::OR) &&
        NumElems == 2 && ExtendedWidth == 128) {
      Extended = DAG.getBitcast(MVT::v4i32, Extended);
      ExtendedVT = MVT::i32;
    }
    switch (ScalarOpcode) {
    case ISD::AND:
      Result = DAG.getNode(ISD::VECREDUCE_UMIN, DL, ExtendedVT, Extended);
      break;
    case ISD::OR:
      Result = DAG.getNode(ISD::VECREDUCE_UMAX, DL, ExtendedVT, Extended);
      break;
    case ISD::XOR:
      Result = DAG.getNode(ISD::VECREDUCE_ADD, DL, ExtendedVT, Extended);
      break;
    default:
      llvm_unreachable("Unexpected Opcode");
    }

    Result = DAG.getAnyExtOrTrunc(Result, DL, MVT::i1);
  } else {
    // Iteratively split the vector in half and combine using the bitwise
    // operation until it fits in a 64 bit register.
    while (VecVT.getSizeInBits() > 64) {
      SDValue Lo, Hi;
      std::tie(Lo, Hi) = DAG.SplitVector(Vec, DL);
      VecVT = Lo.getValueType();
      NumElems = VecVT.getVectorNumElements();
      Vec = DAG.getNode(ScalarOpcode, DL, VecVT, Lo, Hi);
    }

    EVT ScalarVT = EVT::getIntegerVT(*DAG.getContext(), VecVT.getSizeInBits());

    // Do the remaining work on a scalar since it allows the code generator to
    // combine the shift and bitwise operation into one instruction and since
    // integer instructions can have higher throughput than vector instructions.
    SDValue Scalar = DAG.getBitcast(ScalarVT, Vec);

    // Iteratively combine the lower and upper halves of the scalar using the
    // bitwise operation, halving the relevant region of the scalar in each
    // iteration, until the relevant region is just one element of the original
    // vector.
    for (unsigned Shift = NumElems / 2; Shift > 0; Shift /= 2) {
      SDValue ShiftAmount =
          DAG.getConstant(Shift * ElemVT.getSizeInBits(), DL, MVT::i64);
      SDValue Shifted =
          DAG.getNode(ISD::SRL, DL, ScalarVT, Scalar, ShiftAmount);
      Scalar = DAG.getNode(ScalarOpcode, DL, ScalarVT, Scalar, Shifted);
    }

    Result = DAG.getAnyExtOrTrunc(Scalar, DL, ElemVT);
  }

  return DAG.getAnyExtOrTrunc(Result, DL, VT);
}

SDValue AArch64TargetLowering::LowerVECREDUCE(SDValue Op,
                                              SelectionDAG &DAG) const {
  SDValue Src = Op.getOperand(0);
  EVT SrcVT = Src.getValueType();

  // Scalarize v2f16 to turn it into a faddp. This will be more efficient than
  // widening by inserting zeroes.
  if (Subtarget->hasFullFP16() && Op.getOpcode() == ISD::VECREDUCE_FADD &&
      SrcVT == MVT::v2f16) {
    SDLoc DL(Op);
    return DAG.getNode(ISD::FADD, DL, MVT::f16,
                       DAG.getExtractVectorElt(DL, MVT::f16, Src, 0),
                       DAG.getExtractVectorElt(DL, MVT::f16, Src, 1));
  }

  // Try to lower fixed length reductions to SVE.
  bool OverrideNEON = !Subtarget->isNeonAvailable() ||
                      Op.getOpcode() == ISD::VECREDUCE_AND ||
                      Op.getOpcode() == ISD::VECREDUCE_OR ||
                      Op.getOpcode() == ISD::VECREDUCE_XOR ||
                      Op.getOpcode() == ISD::VECREDUCE_FADD ||
                      (Op.getOpcode() != ISD::VECREDUCE_ADD &&
                       SrcVT.getVectorElementType() == MVT::i64);
  if (SrcVT.isScalableVector() ||
      useSVEForFixedLengthVectorVT(
          SrcVT, OverrideNEON && Subtarget->useSVEForFixedLengthVectors())) {

    if (SrcVT.getVectorElementType() == MVT::i1)
      return LowerPredReductionToSVE(Op, DAG);

    switch (Op.getOpcode()) {
    case ISD::VECREDUCE_ADD:
      return LowerReductionToSVE(AArch64ISD::UADDV_PRED, Op, DAG);
    case ISD::VECREDUCE_AND:
      return LowerReductionToSVE(AArch64ISD::ANDV_PRED, Op, DAG);
    case ISD::VECREDUCE_OR:
      return LowerReductionToSVE(AArch64ISD::ORV_PRED, Op, DAG);
    case ISD::VECREDUCE_SMAX:
      return LowerReductionToSVE(AArch64ISD::SMAXV_PRED, Op, DAG);
    case ISD::VECREDUCE_SMIN:
      return LowerReductionToSVE(AArch64ISD::SMINV_PRED, Op, DAG);
    case ISD::VECREDUCE_UMAX:
      return LowerReductionToSVE(AArch64ISD::UMAXV_PRED, Op, DAG);
    case ISD::VECREDUCE_UMIN:
      return LowerReductionToSVE(AArch64ISD::UMINV_PRED, Op, DAG);
    case ISD::VECREDUCE_XOR:
      return LowerReductionToSVE(AArch64ISD::EORV_PRED, Op, DAG);
    case ISD::VECREDUCE_FADD:
      return LowerReductionToSVE(AArch64ISD::FADDV_PRED, Op, DAG);
    case ISD::VECREDUCE_FMAX:
      return LowerReductionToSVE(AArch64ISD::FMAXNMV_PRED, Op, DAG);
    case ISD::VECREDUCE_FMIN:
      return LowerReductionToSVE(AArch64ISD::FMINNMV_PRED, Op, DAG);
    case ISD::VECREDUCE_FMAXIMUM:
      return LowerReductionToSVE(AArch64ISD::FMAXV_PRED, Op, DAG);
    case ISD::VECREDUCE_FMINIMUM:
      return LowerReductionToSVE(AArch64ISD::FMINV_PRED, Op, DAG);
    default:
      llvm_unreachable("Unhandled fixed length reduction");
    }
  }

  // Lower NEON reductions.
  SDLoc DL(Op);
  switch (Op.getOpcode()) {
  case ISD::VECREDUCE_AND:
  case ISD::VECREDUCE_OR:
  case ISD::VECREDUCE_XOR:
    return getVectorBitwiseReduce(Op.getOpcode(), Op.getOperand(0),
                                  Op.getValueType(), DL, DAG);
  case ISD::VECREDUCE_ADD:
    return getReductionSDNode(AArch64ISD::UADDV, DL, Op, DAG);
  case ISD::VECREDUCE_SMAX:
    return getReductionSDNode(AArch64ISD::SMAXV, DL, Op, DAG);
  case ISD::VECREDUCE_SMIN:
    return getReductionSDNode(AArch64ISD::SMINV, DL, Op, DAG);
  case ISD::VECREDUCE_UMAX:
    return getReductionSDNode(AArch64ISD::UMAXV, DL, Op, DAG);
  case ISD::VECREDUCE_UMIN:
    return getReductionSDNode(AArch64ISD::UMINV, DL, Op, DAG);
  default:
    llvm_unreachable("Unhandled reduction");
  }
}

SDValue AArch64TargetLowering::LowerATOMIC_LOAD_AND(SDValue Op,
                                                    SelectionDAG &DAG) const {
  auto &Subtarget = DAG.getSubtarget<AArch64Subtarget>();
  // No point replacing if we don't have the relevant instruction/libcall anyway
  if (!Subtarget.hasLSE() && !Subtarget.outlineAtomics())
    return SDValue();

  // LSE has an atomic load-clear instruction, but not a load-and.
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  assert(VT != MVT::i128 && "Handled elsewhere, code replicated.");
  SDValue RHS = Op.getOperand(2);
  AtomicSDNode *AN = cast<AtomicSDNode>(Op.getNode());
  RHS = DAG.getNode(ISD::XOR, DL, VT, DAG.getAllOnesConstant(DL, VT), RHS);
  return DAG.getAtomic(ISD::ATOMIC_LOAD_CLR, DL, AN->getMemoryVT(),
                       Op.getOperand(0), Op.getOperand(1), RHS,
                       AN->getMemOperand());
}

SDValue
AArch64TargetLowering::LowerWindowsDYNAMIC_STACKALLOC(SDValue Op,
                                                      SelectionDAG &DAG) const {

  SDLoc DL(Op);
  // Get the inputs.
  SDNode *Node = Op.getNode();
  SDValue Chain = Op.getOperand(0);
  SDValue Size = Op.getOperand(1);
  MaybeAlign Align =
      cast<ConstantSDNode>(Op.getOperand(2))->getMaybeAlignValue();
  EVT VT = Node->getValueType(0);

  if (DAG.getMachineFunction().getFunction().hasFnAttribute(
          "no-stack-arg-probe")) {
    SDValue SP = DAG.getCopyFromReg(Chain, DL, AArch64::SP, MVT::i64);
    Chain = SP.getValue(1);
    SP = DAG.getNode(ISD::SUB, DL, MVT::i64, SP, Size);
    if (Align)
      SP =
          DAG.getNode(ISD::AND, DL, VT, SP.getValue(0),
                      DAG.getSignedConstant(-(uint64_t)Align->value(), DL, VT));
    Chain = DAG.getCopyToReg(Chain, DL, AArch64::SP, SP);
    SDValue Ops[2] = {SP, Chain};
    return DAG.getMergeValues(Ops, DL);
  }

  Chain = DAG.getCALLSEQ_START(Chain, 0, 0, DL);

  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDValue Callee = DAG.getTargetExternalSymbol(Subtarget->getChkStkName(),
                                               PtrVT, 0);

  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  const uint32_t *Mask = TRI->getWindowsStackProbePreservedMask();
  if (Subtarget->hasCustomCallingConv())
    TRI->UpdateCustomCallPreservedMask(DAG.getMachineFunction(), &Mask);

  Size = DAG.getNode(ISD::SRL, DL, MVT::i64, Size,
                     DAG.getConstant(4, DL, MVT::i64));
  Chain = DAG.getCopyToReg(Chain, DL, AArch64::X15, Size, SDValue());
  Chain =
      DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
                  Chain, Callee, DAG.getRegister(AArch64::X15, MVT::i64),
                  DAG.getRegisterMask(Mask), Chain.getValue(1));
  // To match the actual intent better, we should read the output from X15 here
  // again (instead of potentially spilling it to the stack), but rereading Size
  // from X15 here doesn't work at -O0, since it thinks that X15 is undefined
  // here.

  Size = DAG.getNode(ISD::SHL, DL, MVT::i64, Size,
                     DAG.getConstant(4, DL, MVT::i64));

  SDValue SP = DAG.getCopyFromReg(Chain, DL, AArch64::SP, MVT::i64);
  Chain = SP.getValue(1);
  SP = DAG.getNode(ISD::SUB, DL, MVT::i64, SP, Size);
  if (Align)
    SP = DAG.getNode(ISD::AND, DL, VT, SP.getValue(0),
                     DAG.getSignedConstant(-(uint64_t)Align->value(), DL, VT));
  Chain = DAG.getCopyToReg(Chain, DL, AArch64::SP, SP);

  Chain = DAG.getCALLSEQ_END(Chain, 0, 0, SDValue(), DL);

  SDValue Ops[2] = {SP, Chain};
  return DAG.getMergeValues(Ops, DL);
}

SDValue
AArch64TargetLowering::LowerInlineDYNAMIC_STACKALLOC(SDValue Op,
                                                     SelectionDAG &DAG) const {
  // Get the inputs.
  SDNode *Node = Op.getNode();
  SDValue Chain = Op.getOperand(0);
  SDValue Size = Op.getOperand(1);

  MaybeAlign Align =
      cast<ConstantSDNode>(Op.getOperand(2))->getMaybeAlignValue();
  SDLoc DL(Op);
  EVT VT = Node->getValueType(0);

  // Construct the new SP value in a GPR.
  SDValue SP = DAG.getCopyFromReg(Chain, DL, AArch64::SP, MVT::i64);
  Chain = SP.getValue(1);
  SP = DAG.getNode(ISD::SUB, DL, MVT::i64, SP, Size);
  if (Align)
    SP = DAG.getNode(ISD::AND, DL, VT, SP.getValue(0),
                     DAG.getSignedConstant(-(uint64_t)Align->value(), DL, VT));

  // Set the real SP to the new value with a probing loop.
  Chain = DAG.getNode(AArch64ISD::PROBED_ALLOCA, DL, MVT::Other, Chain, SP);
  SDValue Ops[2] = {SP, Chain};
  return DAG.getMergeValues(Ops, DL);
}

SDValue
AArch64TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
                                               SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();

  if (Subtarget->isTargetWindows())
    return LowerWindowsDYNAMIC_STACKALLOC(Op, DAG);
  else if (hasInlineStackProbe(MF))
    return LowerInlineDYNAMIC_STACKALLOC(Op, DAG);
  else
    return SDValue();
}

SDValue AArch64TargetLowering::LowerAVG(SDValue Op, SelectionDAG &DAG,
                                        unsigned NewOp) const {
  if (Subtarget->hasSVE2())
    return LowerToPredicatedOp(Op, DAG, NewOp);

  // Default to expand.
  return SDValue();
}

SDValue AArch64TargetLowering::LowerVSCALE(SDValue Op,
                                           SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT != MVT::i64 && "Expected illegal VSCALE node");

  SDLoc DL(Op);
  APInt MulImm = Op.getConstantOperandAPInt(0);
  return DAG.getZExtOrTrunc(DAG.getVScale(DL, MVT::i64, MulImm.sext(64)), DL,
                            VT);
}

/// Set the IntrinsicInfo for the `aarch64_sve_st<N>` intrinsics.
template <unsigned NumVecs>
static bool
setInfoSVEStN(const AArch64TargetLowering &TLI, const DataLayout &DL,
              AArch64TargetLowering::IntrinsicInfo &Info, const CallInst &CI) {
  Info.opc = ISD::INTRINSIC_VOID;
  // Retrieve EC from first vector argument.
  const EVT VT = TLI.getMemValueType(DL, CI.getArgOperand(0)->getType());
  ElementCount EC = VT.getVectorElementCount();
#ifndef NDEBUG
  // Check the assumption that all input vectors are the same type.
  for (unsigned I = 0; I < NumVecs; ++I)
    assert(VT == TLI.getMemValueType(DL, CI.getArgOperand(I)->getType()) &&
           "Invalid type.");
#endif
  // memVT is `NumVecs * VT`.
  Info.memVT = EVT::getVectorVT(CI.getType()->getContext(), VT.getScalarType(),
                                EC * NumVecs);
  Info.ptrVal = CI.getArgOperand(CI.arg_size() - 1);
  Info.offset = 0;
  Info.align.reset();
  Info.flags = MachineMemOperand::MOStore;
  return true;
}

/// getTgtMemIntrinsic - Represent NEON load and store intrinsics as
/// MemIntrinsicNodes.  The associated MachineMemOperands record the alignment
/// specified in the intrinsic calls.
bool AArch64TargetLowering::getTgtMemIntrinsic(IntrinsicInfo &Info,
                                               const CallInst &I,
                                               MachineFunction &MF,
                                               unsigned Intrinsic) const {
  auto &DL = I.getDataLayout();
  switch (Intrinsic) {
  case Intrinsic::aarch64_sve_st2:
    return setInfoSVEStN<2>(*this, DL, Info, I);
  case Intrinsic::aarch64_sve_st3:
    return setInfoSVEStN<3>(*this, DL, Info, I);
  case Intrinsic::aarch64_sve_st4:
    return setInfoSVEStN<4>(*this, DL, Info, I);
  case Intrinsic::aarch64_neon_ld2:
  case Intrinsic::aarch64_neon_ld3:
  case Intrinsic::aarch64_neon_ld4:
  case Intrinsic::aarch64_neon_ld1x2:
  case Intrinsic::aarch64_neon_ld1x3:
  case Intrinsic::aarch64_neon_ld1x4: {
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    uint64_t NumElts = DL.getTypeSizeInBits(I.getType()) / 64;
    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
    Info.ptrVal = I.getArgOperand(I.arg_size() - 1);
    Info.offset = 0;
    Info.align.reset();
    // volatile loads with NEON intrinsics not supported
    Info.flags = MachineMemOperand::MOLoad;
    return true;
  }
  case Intrinsic::aarch64_neon_ld2lane:
  case Intrinsic::aarch64_neon_ld3lane:
  case Intrinsic::aarch64_neon_ld4lane:
  case Intrinsic::aarch64_neon_ld2r:
  case Intrinsic::aarch64_neon_ld3r:
  case Intrinsic::aarch64_neon_ld4r: {
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    // ldx return struct with the same vec type
    Type *RetTy = I.getType();
    auto *StructTy = cast<StructType>(RetTy);
    unsigned NumElts = StructTy->getNumElements();
    Type *VecTy = StructTy->getElementType(0);
    MVT EleVT = MVT::getVT(VecTy).getVectorElementType();
    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), EleVT, NumElts);
    Info.ptrVal = I.getArgOperand(I.arg_size() - 1);
    Info.offset = 0;
    Info.align.reset();
    // volatile loads with NEON intrinsics not supported
    Info.flags = MachineMemOperand::MOLoad;
    return true;
  }
  case Intrinsic::aarch64_neon_st2:
  case Intrinsic::aarch64_neon_st3:
  case Intrinsic::aarch64_neon_st4:
  case Intrinsic::aarch64_neon_st1x2:
  case Intrinsic::aarch64_neon_st1x3:
  case Intrinsic::aarch64_neon_st1x4: {
    Info.opc = ISD::INTRINSIC_VOID;
    unsigned NumElts = 0;
    for (const Value *Arg : I.args()) {
      Type *ArgTy = Arg->getType();
      if (!ArgTy->isVectorTy())
        break;
      NumElts += DL.getTypeSizeInBits(ArgTy) / 64;
    }
    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), MVT::i64, NumElts);
    Info.ptrVal = I.getArgOperand(I.arg_size() - 1);
    Info.offset = 0;
    Info.align.reset();
    // volatile stores with NEON intrinsics not supported
    Info.flags = MachineMemOperand::MOStore;
    return true;
  }
  case Intrinsic::aarch64_neon_st2lane:
  case Intrinsic::aarch64_neon_st3lane:
  case Intrinsic::aarch64_neon_st4lane: {
    Info.opc = ISD::INTRINSIC_VOID;
    unsigned NumElts = 0;
    // all the vector type is same
    Type *VecTy = I.getArgOperand(0)->getType();
    MVT EleVT = MVT::getVT(VecTy).getVectorElementType();

    for (const Value *Arg : I.args()) {
      Type *ArgTy = Arg->getType();
      if (!ArgTy->isVectorTy())
        break;
      NumElts += 1;
    }

    Info.memVT = EVT::getVectorVT(I.getType()->getContext(), EleVT, NumElts);
    Info.ptrVal = I.getArgOperand(I.arg_size() - 1);
    Info.offset = 0;
    Info.align.reset();
    // volatile stores with NEON intrinsics not supported
    Info.flags = MachineMemOperand::MOStore;
    return true;
  }
  case Intrinsic::aarch64_ldaxr:
  case Intrinsic::aarch64_ldxr: {
    Type *ValTy = I.getParamElementType(0);
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(ValTy);
    Info.ptrVal = I.getArgOperand(0);
    Info.offset = 0;
    Info.align = DL.getABITypeAlign(ValTy);
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
    return true;
  }
  case Intrinsic::aarch64_stlxr:
  case Intrinsic::aarch64_stxr: {
    Type *ValTy = I.getParamElementType(1);
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(ValTy);
    Info.ptrVal = I.getArgOperand(1);
    Info.offset = 0;
    Info.align = DL.getABITypeAlign(ValTy);
    Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile;
    return true;
  }
  case Intrinsic::aarch64_ldaxp:
  case Intrinsic::aarch64_ldxp:
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::i128;
    Info.ptrVal = I.getArgOperand(0);
    Info.offset = 0;
    Info.align = Align(16);
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile;
    return true;
  case Intrinsic::aarch64_stlxp:
  case Intrinsic::aarch64_stxp:
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::i128;
    Info.ptrVal = I.getArgOperand(2);
    Info.offset = 0;
    Info.align = Align(16);
    Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MOVolatile;
    return true;
  case Intrinsic::aarch64_sve_ldnt1: {
    Type *ElTy = cast<VectorType>(I.getType())->getElementType();
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(I.getType());
    Info.ptrVal = I.getArgOperand(1);
    Info.offset = 0;
    Info.align = DL.getABITypeAlign(ElTy);
    Info.flags = MachineMemOperand::MOLoad | MachineMemOperand::MONonTemporal;
    return true;
  }
  case Intrinsic::aarch64_sve_stnt1: {
    Type *ElTy =
        cast<VectorType>(I.getArgOperand(0)->getType())->getElementType();
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(I.getOperand(0)->getType());
    Info.ptrVal = I.getArgOperand(2);
    Info.offset = 0;
    Info.align = DL.getABITypeAlign(ElTy);
    Info.flags = MachineMemOperand::MOStore | MachineMemOperand::MONonTemporal;
    return true;
  }
  case Intrinsic::aarch64_mops_memset_tag: {
    Value *Dst = I.getArgOperand(0);
    Value *Val = I.getArgOperand(1);
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.memVT = MVT::getVT(Val->getType());
    Info.ptrVal = Dst;
    Info.offset = 0;
    Info.align = I.getParamAlign(0).valueOrOne();
    Info.flags = MachineMemOperand::MOStore;
    // The size of the memory being operated on is unknown at this point
    Info.size = MemoryLocation::UnknownSize;
    return true;
  }
  default:
    break;
  }

  return false;
}

bool AArch64TargetLowering::shouldReduceLoadWidth(
    SDNode *Load, ISD::LoadExtType ExtTy, EVT NewVT,
    std::optional<unsigned> ByteOffset) const {
  // TODO: This may be worth removing. Check regression tests for diffs.
  if (!TargetLoweringBase::shouldReduceLoadWidth(Load, ExtTy, NewVT,
                                                 ByteOffset))
    return false;

  // If we're reducing the load width in order to avoid having to use an extra
  // instruction to do extension then it's probably a good idea.
  if (ExtTy != ISD::NON_EXTLOAD)
    return true;
  // Don't reduce load width if it would prevent us from combining a shift into
  // the offset.
  MemSDNode *Mem = dyn_cast<MemSDNode>(Load);
  assert(Mem);
  const SDValue &Base = Mem->getBasePtr();
  if (Base.getOpcode() == ISD::ADD &&
      Base.getOperand(1).getOpcode() == ISD::SHL &&
      Base.getOperand(1).hasOneUse() &&
      Base.getOperand(1).getOperand(1).getOpcode() == ISD::Constant) {
    // It's unknown whether a scalable vector has a power-of-2 bitwidth.
    if (Mem->getMemoryVT().isScalableVector())
      return false;
    // The shift can be combined if it matches the size of the value being
    // loaded (and so reducing the width would make it not match).
    uint64_t ShiftAmount = Base.getOperand(1).getConstantOperandVal(1);
    uint64_t LoadBytes = Mem->getMemoryVT().getSizeInBits()/8;
    if (ShiftAmount == Log2_32(LoadBytes))
      return false;
  }
  // We have no reason to disallow reducing the load width, so allow it.
  return true;
}

// Treat a sext_inreg(extract(..)) as free if it has multiple uses.
bool AArch64TargetLowering::shouldRemoveRedundantExtend(SDValue Extend) const {
  EVT VT = Extend.getValueType();
  if ((VT == MVT::i64 || VT == MVT::i32) && Extend->use_size()) {
    SDValue Extract = Extend.getOperand(0);
    if (Extract.getOpcode() == ISD::ANY_EXTEND && Extract.hasOneUse())
      Extract = Extract.getOperand(0);
    if (Extract.getOpcode() == ISD::EXTRACT_VECTOR_ELT && Extract.hasOneUse()) {
      EVT VecVT = Extract.getOperand(0).getValueType();
      if (VecVT.getScalarType() == MVT::i8 || VecVT.getScalarType() == MVT::i16)
        return false;
    }
  }
  return true;
}

// Truncations from 64-bit GPR to 32-bit GPR is free.
bool AArch64TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
  if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
    return false;
  uint64_t NumBits1 = Ty1->getPrimitiveSizeInBits().getFixedValue();
  uint64_t NumBits2 = Ty2->getPrimitiveSizeInBits().getFixedValue();
  return NumBits1 > NumBits2;
}
bool AArch64TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
  if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
    return false;
  uint64_t NumBits1 = VT1.getFixedSizeInBits();
  uint64_t NumBits2 = VT2.getFixedSizeInBits();
  return NumBits1 > NumBits2;
}

/// Check if it is profitable to hoist instruction in then/else to if.
/// Not profitable if I and it's user can form a FMA instruction
/// because we prefer FMSUB/FMADD.
bool AArch64TargetLowering::isProfitableToHoist(Instruction *I) const {
  if (I->getOpcode() != Instruction::FMul)
    return true;

  if (!I->hasOneUse())
    return true;

  Instruction *User = I->user_back();

  if (!(User->getOpcode() == Instruction::FSub ||
        User->getOpcode() == Instruction::FAdd))
    return true;

  const TargetOptions &Options = getTargetMachine().Options;
  const Function *F = I->getFunction();
  const DataLayout &DL = F->getDataLayout();
  Type *Ty = User->getOperand(0)->getType();

  return !(isFMAFasterThanFMulAndFAdd(*F, Ty) &&
           isOperationLegalOrCustom(ISD::FMA, getValueType(DL, Ty)) &&
           (Options.AllowFPOpFusion == FPOpFusion::Fast ||
            Options.UnsafeFPMath));
}

// All 32-bit GPR operations implicitly zero the high-half of the corresponding
// 64-bit GPR.
bool AArch64TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
  if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
    return false;
  unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
  unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
  return NumBits1 == 32 && NumBits2 == 64;
}
bool AArch64TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
  if (VT1.isVector() || VT2.isVector() || !VT1.isInteger() || !VT2.isInteger())
    return false;
  unsigned NumBits1 = VT1.getSizeInBits();
  unsigned NumBits2 = VT2.getSizeInBits();
  return NumBits1 == 32 && NumBits2 == 64;
}

bool AArch64TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
  EVT VT1 = Val.getValueType();
  if (isZExtFree(VT1, VT2)) {
    return true;
  }

  if (Val.getOpcode() != ISD::LOAD)
    return false;

  // 8-, 16-, and 32-bit integer loads all implicitly zero-extend.
  return (VT1.isSimple() && !VT1.isVector() && VT1.isInteger() &&
          VT2.isSimple() && !VT2.isVector() && VT2.isInteger() &&
          VT1.getSizeInBits() <= 32);
}

bool AArch64TargetLowering::isExtFreeImpl(const Instruction *Ext) const {
  if (isa<FPExtInst>(Ext))
    return false;

  // Vector types are not free.
  if (Ext->getType()->isVectorTy())
    return false;

  for (const Use &U : Ext->uses()) {
    // The extension is free if we can fold it with a left shift in an
    // addressing mode or an arithmetic operation: add, sub, and cmp.

    // Is there a shift?
    const Instruction *Instr = cast<Instruction>(U.getUser());

    // Is this a constant shift?
    switch (Instr->getOpcode()) {
    case Instruction::Shl:
      if (!isa<ConstantInt>(Instr->getOperand(1)))
        return false;
      break;
    case Instruction::GetElementPtr: {
      gep_type_iterator GTI = gep_type_begin(Instr);
      auto &DL = Ext->getDataLayout();
      std::advance(GTI, U.getOperandNo()-1);
      Type *IdxTy = GTI.getIndexedType();
      // This extension will end up with a shift because of the scaling factor.
      // 8-bit sized types have a scaling factor of 1, thus a shift amount of 0.
      // Get the shift amount based on the scaling factor:
      // log2(sizeof(IdxTy)) - log2(8).
      if (IdxTy->isScalableTy())
        return false;
      uint64_t ShiftAmt =
          llvm::countr_zero(DL.getTypeStoreSizeInBits(IdxTy).getFixedValue()) -
          3;
      // Is the constant foldable in the shift of the addressing mode?
      // I.e., shift amount is between 1 and 4 inclusive.
      if (ShiftAmt == 0 || ShiftAmt > 4)
        return false;
      break;
    }
    case Instruction::Trunc:
      // Check if this is a noop.
      // trunc(sext ty1 to ty2) to ty1.
      if (Instr->getType() == Ext->getOperand(0)->getType())
        continue;
      [[fallthrough]];
    default:
      return false;
    }

    // At this point we can use the bfm family, so this extension is free
    // for that use.
  }
  return true;
}

static bool createTblShuffleMask(unsigned SrcWidth, unsigned DstWidth,
                                 unsigned NumElts, bool IsLittleEndian,
                                 SmallVectorImpl<int> &Mask) {
  if (DstWidth % 8 != 0 || DstWidth <= 16 || DstWidth > 64)
    return false;

  assert(DstWidth % SrcWidth == 0 &&
         "TBL lowering is not supported for a conversion instruction with this "
         "source and destination element type.");

  unsigned Factor = DstWidth / SrcWidth;
  unsigned MaskLen = NumElts * Factor;

  Mask.clear();
  Mask.resize(MaskLen, NumElts);

  unsigned SrcIndex = 0;
  for (unsigned I = IsLittleEndian ? 0 : Factor - 1; I < MaskLen; I += Factor)
    Mask[I] = SrcIndex++;

  return true;
}

static Value *createTblShuffleForZExt(IRBuilderBase &Builder, Value *Op,
                                      FixedVectorType *ZExtTy,
                                      FixedVectorType *DstTy,
                                      bool IsLittleEndian) {
  auto *SrcTy = cast<FixedVectorType>(Op->getType());
  unsigned NumElts = SrcTy->getNumElements();
  auto SrcWidth = cast<IntegerType>(SrcTy->getElementType())->getBitWidth();
  auto DstWidth = cast<IntegerType>(DstTy->getElementType())->getBitWidth();

  SmallVector<int> Mask;
  if (!createTblShuffleMask(SrcWidth, DstWidth, NumElts, IsLittleEndian, Mask))
    return nullptr;

  auto *FirstEltZero = Builder.CreateInsertElement(
      PoisonValue::get(SrcTy), Builder.getIntN(SrcWidth, 0), uint64_t(0));
  Value *Result = Builder.CreateShuffleVector(Op, FirstEltZero, Mask);
  Result = Builder.CreateBitCast(Result, DstTy);
  if (DstTy != ZExtTy)
    Result = Builder.CreateZExt(Result, ZExtTy);
  return Result;
}

static Value *createTblShuffleForSExt(IRBuilderBase &Builder, Value *Op,
                                      FixedVectorType *DstTy,
                                      bool IsLittleEndian) {
  auto *SrcTy = cast<FixedVectorType>(Op->getType());
  auto SrcWidth = cast<IntegerType>(SrcTy->getElementType())->getBitWidth();
  auto DstWidth = cast<IntegerType>(DstTy->getElementType())->getBitWidth();

  SmallVector<int> Mask;
  if (!createTblShuffleMask(SrcWidth, DstWidth, SrcTy->getNumElements(),
                            !IsLittleEndian, Mask))
    return nullptr;

  auto *FirstEltZero = Builder.CreateInsertElement(
      PoisonValue::get(SrcTy), Builder.getIntN(SrcWidth, 0), uint64_t(0));

  return Builder.CreateShuffleVector(Op, FirstEltZero, Mask);
}

static void createTblForTrunc(TruncInst *TI, bool IsLittleEndian) {
  IRBuilder<> Builder(TI);
  SmallVector<Value *> Parts;
  int NumElements = cast<FixedVectorType>(TI->getType())->getNumElements();
  auto *SrcTy = cast<FixedVectorType>(TI->getOperand(0)->getType());
  auto *DstTy = cast<FixedVectorType>(TI->getType());
  assert(SrcTy->getElementType()->isIntegerTy() &&
         "Non-integer type source vector element is not supported");
  assert(DstTy->getElementType()->isIntegerTy(8) &&
         "Unsupported destination vector element type");
  unsigned SrcElemTySz =
      cast<IntegerType>(SrcTy->getElementType())->getBitWidth();
  unsigned DstElemTySz =
      cast<IntegerType>(DstTy->getElementType())->getBitWidth();
  assert((SrcElemTySz % DstElemTySz == 0) &&
         "Cannot lower truncate to tbl instructions for a source element size "
         "that is not divisible by the destination element size");
  unsigned TruncFactor = SrcElemTySz / DstElemTySz;
  assert((SrcElemTySz == 16 || SrcElemTySz == 32 || SrcElemTySz == 64) &&
         "Unsupported source vector element type size");
  Type *VecTy = FixedVectorType::get(Builder.getInt8Ty(), 16);

  // Create a mask to choose every nth byte from the source vector table of
  // bytes to create the truncated destination vector, where 'n' is the truncate
  // ratio. For example, for a truncate from Yxi64 to Yxi8, choose
  // 0,8,16,..Y*8th bytes for the little-endian format
  SmallVector<Constant *, 16> MaskConst;
  for (int Itr = 0; Itr < 16; Itr++) {
    if (Itr < NumElements)
      MaskConst.push_back(Builder.getInt8(
          IsLittleEndian ? Itr * TruncFactor
                         : Itr * TruncFactor + (TruncFactor - 1)));
    else
      MaskConst.push_back(Builder.getInt8(255));
  }

  int MaxTblSz = 128 * 4;
  int MaxSrcSz = SrcElemTySz * NumElements;
  int ElemsPerTbl =
      (MaxTblSz > MaxSrcSz) ? NumElements : (MaxTblSz / SrcElemTySz);
  assert(ElemsPerTbl <= 16 &&
         "Maximum elements selected using TBL instruction cannot exceed 16!");

  int ShuffleCount = 128 / SrcElemTySz;
  SmallVector<int> ShuffleLanes;
  for (int i = 0; i < ShuffleCount; ++i)
    ShuffleLanes.push_back(i);

  // Create TBL's table of bytes in 1,2,3 or 4 FP/SIMD registers using shuffles
  // over the source vector. If TBL's maximum 4 FP/SIMD registers are saturated,
  // call TBL & save the result in a vector of TBL results for combining later.
  SmallVector<Value *> Results;
  while (ShuffleLanes.back() < NumElements) {
    Parts.push_back(Builder.CreateBitCast(
        Builder.CreateShuffleVector(TI->getOperand(0), ShuffleLanes), VecTy));

    if (Parts.size() == 4) {
      Parts.push_back(ConstantVector::get(MaskConst));
      Results.push_back(
          Builder.CreateIntrinsic(Intrinsic::aarch64_neon_tbl4, VecTy, Parts));
      Parts.clear();
    }

    for (int i = 0; i < ShuffleCount; ++i)
      ShuffleLanes[i] += ShuffleCount;
  }

  assert((Parts.empty() || Results.empty()) &&
         "Lowering trunc for vectors requiring different TBL instructions is "
         "not supported!");
  // Call TBL for the residual table bytes present in 1,2, or 3 FP/SIMD
  // registers
  if (!Parts.empty()) {
    Intrinsic::ID TblID;
    switch (Parts.size()) {
    case 1:
      TblID = Intrinsic::aarch64_neon_tbl1;
      break;
    case 2:
      TblID = Intrinsic::aarch64_neon_tbl2;
      break;
    case 3:
      TblID = Intrinsic::aarch64_neon_tbl3;
      break;
    }

    Parts.push_back(ConstantVector::get(MaskConst));
    Results.push_back(Builder.CreateIntrinsic(TblID, VecTy, Parts));
  }

  // Extract the destination vector from TBL result(s) after combining them
  // where applicable. Currently, at most two TBLs are supported.
  assert(Results.size() <= 2 && "Trunc lowering does not support generation of "
                                "more than 2 tbl instructions!");
  Value *FinalResult = Results[0];
  if (Results.size() == 1) {
    if (ElemsPerTbl < 16) {
      SmallVector<int> FinalMask(ElemsPerTbl);
      std::iota(FinalMask.begin(), FinalMask.end(), 0);
      FinalResult = Builder.CreateShuffleVector(Results[0], FinalMask);
    }
  } else {
    SmallVector<int> FinalMask(ElemsPerTbl * Results.size());
    if (ElemsPerTbl < 16) {
      std::iota(FinalMask.begin(), FinalMask.begin() + ElemsPerTbl, 0);
      std::iota(FinalMask.begin() + ElemsPerTbl, FinalMask.end(), 16);
    } else {
      std::iota(FinalMask.begin(), FinalMask.end(), 0);
    }
    FinalResult =
        Builder.CreateShuffleVector(Results[0], Results[1], FinalMask);
  }

  TI->replaceAllUsesWith(FinalResult);
  TI->eraseFromParent();
}

bool AArch64TargetLowering::optimizeExtendOrTruncateConversion(
    Instruction *I, Loop *L, const TargetTransformInfo &TTI) const {
  // shuffle_vector instructions are serialized when targeting SVE,
  // see LowerSPLAT_VECTOR. This peephole is not beneficial.
  if (!EnableExtToTBL || Subtarget->useSVEForFixedLengthVectors())
    return false;

  // Try to optimize conversions using tbl. This requires materializing constant
  // index vectors, which can increase code size and add loads. Skip the
  // transform unless the conversion is in a loop block guaranteed to execute
  // and we are not optimizing for size.
  Function *F = I->getParent()->getParent();
  if (!L || L->getHeader() != I->getParent() || F->hasOptSize())
    return false;

  auto *SrcTy = dyn_cast<FixedVectorType>(I->getOperand(0)->getType());
  auto *DstTy = dyn_cast<FixedVectorType>(I->getType());
  if (!SrcTy || !DstTy)
    return false;

  // Convert 'zext <Y x i8> %x to <Y x i8X>' to a shuffle that can be
  // lowered to tbl instructions to insert the original i8 elements
  // into i8x lanes. This is enabled for cases where it is beneficial.
  auto *ZExt = dyn_cast<ZExtInst>(I);
  if (ZExt && SrcTy->getElementType()->isIntegerTy(8)) {
    auto DstWidth = DstTy->getElementType()->getScalarSizeInBits();
    if (DstWidth % 8 != 0)
      return false;

    auto *TruncDstType =
        cast<FixedVectorType>(VectorType::getTruncatedElementVectorType(DstTy));
    // If the ZExt can be lowered to a single ZExt to the next power-of-2 and
    // the remaining ZExt folded into the user, don't use tbl lowering.
    auto SrcWidth = SrcTy->getElementType()->getScalarSizeInBits();
    if (TTI.getCastInstrCost(I->getOpcode(), DstTy, TruncDstType,
                             TargetTransformInfo::getCastContextHint(I),
                             TTI::TCK_SizeAndLatency, I) == TTI::TCC_Free) {
      if (SrcWidth * 2 >= TruncDstType->getElementType()->getScalarSizeInBits())
        return false;

      DstTy = TruncDstType;
    }

    // mul(zext(i8), sext) can be transformed into smull(zext, sext) which
    // performs one extend implicitly. If DstWidth is at most 4 * SrcWidth, at
    // most one extra extend step is needed and using tbl is not profitable.
    // Similarly, bail out if partial_reduce(acc, zext(i8)) can be lowered to a
    // udot instruction.
    if (SrcWidth * 4 <= DstWidth) {
      if (all_of(I->users(), [&](auto *U) {
            auto *SingleUser = cast<Instruction>(&*U);
            if (match(SingleUser, m_c_Mul(m_Specific(I), m_SExt(m_Value()))))
              return true;
            if (match(SingleUser,
                      m_Intrinsic<
                          Intrinsic::experimental_vector_partial_reduce_add>(
                          m_Value(), m_Specific(I))))
              return true;
            return false;
          }))
        return false;
    }

    if (DstTy->getScalarSizeInBits() >= 64)
      return false;

    IRBuilder<> Builder(ZExt);
    Value *Result = createTblShuffleForZExt(
        Builder, ZExt->getOperand(0), cast<FixedVectorType>(ZExt->getType()),
        DstTy, Subtarget->isLittleEndian());
    if (!Result)
      return false;
    ZExt->replaceAllUsesWith(Result);
    ZExt->eraseFromParent();
    return true;
  }

  auto *UIToFP = dyn_cast<UIToFPInst>(I);
  if (UIToFP && ((SrcTy->getElementType()->isIntegerTy(8) &&
                  DstTy->getElementType()->isFloatTy()) ||
                 (SrcTy->getElementType()->isIntegerTy(16) &&
                  DstTy->getElementType()->isDoubleTy()))) {
    IRBuilder<> Builder(I);
    Value *ZExt = createTblShuffleForZExt(
        Builder, I->getOperand(0), FixedVectorType::getInteger(DstTy),
        FixedVectorType::getInteger(DstTy), Subtarget->isLittleEndian());
    assert(ZExt && "Cannot fail for the i8 to float conversion");
    auto *UI = Builder.CreateUIToFP(ZExt, DstTy);
    I->replaceAllUsesWith(UI);
    I->eraseFromParent();
    return true;
  }

  auto *SIToFP = dyn_cast<SIToFPInst>(I);
  if (SIToFP && SrcTy->getElementType()->isIntegerTy(8) &&
      DstTy->getElementType()->isFloatTy()) {
    IRBuilder<> Builder(I);
    auto *Shuffle = createTblShuffleForSExt(Builder, I->getOperand(0),
                                            FixedVectorType::getInteger(DstTy),
                                            Subtarget->isLittleEndian());
    assert(Shuffle && "Cannot fail for the i8 to float conversion");
    auto *Cast = Builder.CreateBitCast(Shuffle, VectorType::getInteger(DstTy));
    auto *AShr = Builder.CreateAShr(Cast, 24, "", true);
    auto *SI = Builder.CreateSIToFP(AShr, DstTy);
    I->replaceAllUsesWith(SI);
    I->eraseFromParent();
    return true;
  }

  // Convert 'fptoui <(8|16) x float> to <(8|16) x i8>' to a wide fptoui
  // followed by a truncate lowered to using tbl.4.
  auto *FPToUI = dyn_cast<FPToUIInst>(I);
  if (FPToUI &&
      (SrcTy->getNumElements() == 8 || SrcTy->getNumElements() == 16) &&
      SrcTy->getElementType()->isFloatTy() &&
      DstTy->getElementType()->isIntegerTy(8)) {
    IRBuilder<> Builder(I);
    auto *WideConv = Builder.CreateFPToUI(FPToUI->getOperand(0),
                                          VectorType::getInteger(SrcTy));
    auto *TruncI = Builder.CreateTrunc(WideConv, DstTy);
    I->replaceAllUsesWith(TruncI);
    I->eraseFromParent();
    createTblForTrunc(cast<TruncInst>(TruncI), Subtarget->isLittleEndian());
    return true;
  }

  // Convert 'trunc <(8|16) x (i32|i64)> %x to <(8|16) x i8>' to an appropriate
  // tbl instruction selecting the lowest/highest (little/big endian) 8 bits
  // per lane of the input that is represented using 1,2,3 or 4 128-bit table
  // registers
  auto *TI = dyn_cast<TruncInst>(I);
  if (TI && DstTy->getElementType()->isIntegerTy(8) &&
      ((SrcTy->getElementType()->isIntegerTy(32) ||
        SrcTy->getElementType()->isIntegerTy(64)) &&
       (SrcTy->getNumElements() == 16 || SrcTy->getNumElements() == 8))) {
    createTblForTrunc(TI, Subtarget->isLittleEndian());
    return true;
  }

  return false;
}

bool AArch64TargetLowering::hasPairedLoad(EVT LoadedType,
                                          Align &RequiredAlignment) const {
  if (!LoadedType.isSimple() ||
      (!LoadedType.isInteger() && !LoadedType.isFloatingPoint()))
    return false;
  // Cyclone supports unaligned accesses.
  RequiredAlignment = Align(1);
  unsigned NumBits = LoadedType.getSizeInBits();
  return NumBits == 32 || NumBits == 64;
}

/// A helper function for determining the number of interleaved accesses we
/// will generate when lowering accesses of the given type.
unsigned AArch64TargetLowering::getNumInterleavedAccesses(
    VectorType *VecTy, const DataLayout &DL, bool UseScalable) const {
  unsigned VecSize = 128;
  unsigned ElSize = DL.getTypeSizeInBits(VecTy->getElementType());
  unsigned MinElts = VecTy->getElementCount().getKnownMinValue();
  if (UseScalable && isa<FixedVectorType>(VecTy))
    VecSize = std::max(Subtarget->getMinSVEVectorSizeInBits(), 128u);
  return std::max<unsigned>(1, (MinElts * ElSize + 127) / VecSize);
}

MachineMemOperand::Flags
AArch64TargetLowering::getTargetMMOFlags(const Instruction &I) const {
  if (Subtarget->getProcFamily() == AArch64Subtarget::Falkor &&
      I.hasMetadata(FALKOR_STRIDED_ACCESS_MD))
    return MOStridedAccess;
  return MachineMemOperand::MONone;
}

bool AArch64TargetLowering::isLegalInterleavedAccessType(
    VectorType *VecTy, const DataLayout &DL, bool &UseScalable) const {
  unsigned ElSize = DL.getTypeSizeInBits(VecTy->getElementType());
  auto EC = VecTy->getElementCount();
  unsigned MinElts = EC.getKnownMinValue();

  UseScalable = false;

  if (isa<FixedVectorType>(VecTy) && !Subtarget->isNeonAvailable() &&
      (!Subtarget->useSVEForFixedLengthVectors() ||
       !getSVEPredPatternFromNumElements(MinElts)))
    return false;

  if (isa<ScalableVectorType>(VecTy) &&
      !Subtarget->isSVEorStreamingSVEAvailable())
    return false;

  // Ensure the number of vector elements is greater than 1.
  if (MinElts < 2)
    return false;

  // Ensure the element type is legal.
  if (ElSize != 8 && ElSize != 16 && ElSize != 32 && ElSize != 64)
    return false;

  if (EC.isScalable()) {
    UseScalable = true;
    return isPowerOf2_32(MinElts) && (MinElts * ElSize) % 128 == 0;
  }

  unsigned VecSize = DL.getTypeSizeInBits(VecTy);
  if (Subtarget->useSVEForFixedLengthVectors()) {
    unsigned MinSVEVectorSize =
        std::max(Subtarget->getMinSVEVectorSizeInBits(), 128u);
    if (VecSize % MinSVEVectorSize == 0 ||
        (VecSize < MinSVEVectorSize && isPowerOf2_32(MinElts) &&
         (!Subtarget->isNeonAvailable() || VecSize > 128))) {
      UseScalable = true;
      return true;
    }
  }

  // Ensure the total vector size is 64 or a multiple of 128. Types larger than
  // 128 will be split into multiple interleaved accesses.
  return Subtarget->isNeonAvailable() && (VecSize == 64 || VecSize % 128 == 0);
}

static ScalableVectorType *getSVEContainerIRType(FixedVectorType *VTy) {
  if (VTy->getElementType() == Type::getDoubleTy(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 2);

  if (VTy->getElementType() == Type::getFloatTy(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 4);

  if (VTy->getElementType() == Type::getBFloatTy(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 8);

  if (VTy->getElementType() == Type::getHalfTy(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 8);

  if (VTy->getElementType() == Type::getInt64Ty(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 2);

  if (VTy->getElementType() == Type::getInt32Ty(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 4);

  if (VTy->getElementType() == Type::getInt16Ty(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 8);

  if (VTy->getElementType() == Type::getInt8Ty(VTy->getContext()))
    return ScalableVectorType::get(VTy->getElementType(), 16);

  llvm_unreachable("Cannot handle input vector type");
}

static Function *getStructuredLoadFunction(Module *M, unsigned Factor,
                                           bool Scalable, Type *LDVTy,
                                           Type *PtrTy) {
  assert(Factor >= 2 && Factor <= 4 && "Invalid interleave factor");
  static const Intrinsic::ID SVELoads[3] = {Intrinsic::aarch64_sve_ld2_sret,
                                            Intrinsic::aarch64_sve_ld3_sret,
                                            Intrinsic::aarch64_sve_ld4_sret};
  static const Intrinsic::ID NEONLoads[3] = {Intrinsic::aarch64_neon_ld2,
                                             Intrinsic::aarch64_neon_ld3,
                                             Intrinsic::aarch64_neon_ld4};
  if (Scalable)
    return Intrinsic::getOrInsertDeclaration(M, SVELoads[Factor - 2], {LDVTy});

  return Intrinsic::getOrInsertDeclaration(M, NEONLoads[Factor - 2],
                                           {LDVTy, PtrTy});
}

static Function *getStructuredStoreFunction(Module *M, unsigned Factor,
                                            bool Scalable, Type *STVTy,
                                            Type *PtrTy) {
  assert(Factor >= 2 && Factor <= 4 && "Invalid interleave factor");
  static const Intrinsic::ID SVEStores[3] = {Intrinsic::aarch64_sve_st2,
                                             Intrinsic::aarch64_sve_st3,
                                             Intrinsic::aarch64_sve_st4};
  static const Intrinsic::ID NEONStores[3] = {Intrinsic::aarch64_neon_st2,
                                              Intrinsic::aarch64_neon_st3,
                                              Intrinsic::aarch64_neon_st4};
  if (Scalable)
    return Intrinsic::getOrInsertDeclaration(M, SVEStores[Factor - 2], {STVTy});

  return Intrinsic::getOrInsertDeclaration(M, NEONStores[Factor - 2],
                                           {STVTy, PtrTy});
}

/// Lower an interleaved load into a ldN intrinsic.
///
/// E.g. Lower an interleaved load (Factor = 2):
///        %wide.vec = load <8 x i32>, <8 x i32>* %ptr
///        %v0 = shuffle %wide.vec, undef, <0, 2, 4, 6>  ; Extract even elements
///        %v1 = shuffle %wide.vec, undef, <1, 3, 5, 7>  ; Extract odd elements
///
///      Into:
///        %ld2 = { <4 x i32>, <4 x i32> } call llvm.aarch64.neon.ld2(%ptr)
///        %vec0 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 0
///        %vec1 = extractelement { <4 x i32>, <4 x i32> } %ld2, i32 1
bool AArch64TargetLowering::lowerInterleavedLoad(
    Instruction *Load, Value *Mask, ArrayRef<ShuffleVectorInst *> Shuffles,
    ArrayRef<unsigned> Indices, unsigned Factor) const {
  assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
         "Invalid interleave factor");
  assert(!Shuffles.empty() && "Empty shufflevector input");
  assert(Shuffles.size() == Indices.size() &&
         "Unmatched number of shufflevectors and indices");

  auto *LI = dyn_cast<LoadInst>(Load);
  if (!LI)
    return false;
  assert(!Mask && "Unexpected mask on a load");

  const DataLayout &DL = LI->getDataLayout();

  VectorType *VTy = Shuffles[0]->getType();

  // Skip if we do not have NEON and skip illegal vector types. We can
  // "legalize" wide vector types into multiple interleaved accesses as long as
  // the vector types are divisible by 128.
  bool UseScalable;
  if (!isLegalInterleavedAccessType(VTy, DL, UseScalable))
    return false;

  // Check if the interleave is a zext(shuffle), that can be better optimized
  // into shift / and masks. For the moment we do this just for uitofp (not
  // zext) to avoid issues with widening instructions.
  if (Shuffles.size() == 4 && all_of(Shuffles, [](ShuffleVectorInst *SI) {
        return SI->hasOneUse() && match(SI->user_back(), m_UIToFP(m_Value())) &&
               SI->getType()->getScalarSizeInBits() * 4 ==
                   SI->user_back()->getType()->getScalarSizeInBits();
      }))
    return false;

  unsigned NumLoads = getNumInterleavedAccesses(VTy, DL, UseScalable);

  auto *FVTy = cast<FixedVectorType>(VTy);

  // A pointer vector can not be the return type of the ldN intrinsics. Need to
  // load integer vectors first and then convert to pointer vectors.
  Type *EltTy = FVTy->getElementType();
  if (EltTy->isPointerTy())
    FVTy =
        FixedVectorType::get(DL.getIntPtrType(EltTy), FVTy->getNumElements());

  // If we're going to generate more than one load, reset the sub-vector type
  // to something legal.
  FVTy = FixedVectorType::get(FVTy->getElementType(),
                              FVTy->getNumElements() / NumLoads);

  auto *LDVTy =
      UseScalable ? cast<VectorType>(getSVEContainerIRType(FVTy)) : FVTy;

  IRBuilder<> Builder(LI);

  // The base address of the load.
  Value *BaseAddr = LI->getPointerOperand();

  Type *PtrTy = LI->getPointerOperandType();
  Type *PredTy = VectorType::get(Type::getInt1Ty(LDVTy->getContext()),
                                 LDVTy->getElementCount());

  Function *LdNFunc = getStructuredLoadFunction(LI->getModule(), Factor,
                                                UseScalable, LDVTy, PtrTy);

  // Holds sub-vectors extracted from the load intrinsic return values. The
  // sub-vectors are associated with the shufflevector instructions they will
  // replace.
  DenseMap<ShuffleVectorInst *, SmallVector<Value *, 4>> SubVecs;

  Value *PTrue = nullptr;
  if (UseScalable) {
    std::optional<unsigned> PgPattern =
        getSVEPredPatternFromNumElements(FVTy->getNumElements());
    if (Subtarget->getMinSVEVectorSizeInBits() ==
            Subtarget->getMaxSVEVectorSizeInBits() &&
        Subtarget->getMinSVEVectorSizeInBits() == DL.getTypeSizeInBits(FVTy))
      PgPattern = AArch64SVEPredPattern::all;

    auto *PTruePat =
        ConstantInt::get(Type::getInt32Ty(LDVTy->getContext()), *PgPattern);
    PTrue = Builder.CreateIntrinsic(Intrinsic::aarch64_sve_ptrue, {PredTy},
                                    {PTruePat});
  }

  for (unsigned LoadCount = 0; LoadCount < NumLoads; ++LoadCount) {

    // If we're generating more than one load, compute the base address of
    // subsequent loads as an offset from the previous.
    if (LoadCount > 0)
      BaseAddr = Builder.CreateConstGEP1_32(LDVTy->getElementType(), BaseAddr,
                                            FVTy->getNumElements() * Factor);

    CallInst *LdN;
    if (UseScalable)
      LdN = Builder.CreateCall(LdNFunc, {PTrue, BaseAddr}, "ldN");
    else
      LdN = Builder.CreateCall(LdNFunc, BaseAddr, "ldN");

    // Extract and store the sub-vectors returned by the load intrinsic.
    for (unsigned i = 0; i < Shuffles.size(); i++) {
      ShuffleVectorInst *SVI = Shuffles[i];
      unsigned Index = Indices[i];

      Value *SubVec = Builder.CreateExtractValue(LdN, Index);

      if (UseScalable)
        SubVec = Builder.CreateExtractVector(FVTy, SubVec, uint64_t(0));

      // Convert the integer vector to pointer vector if the element is pointer.
      if (EltTy->isPointerTy())
        SubVec = Builder.CreateIntToPtr(
            SubVec, FixedVectorType::get(SVI->getType()->getElementType(),
                                         FVTy->getNumElements()));

      SubVecs[SVI].push_back(SubVec);
    }
  }

  // Replace uses of the shufflevector instructions with the sub-vectors
  // returned by the load intrinsic. If a shufflevector instruction is
  // associated with more than one sub-vector, those sub-vectors will be
  // concatenated into a single wide vector.
  for (ShuffleVectorInst *SVI : Shuffles) {
    auto &SubVec = SubVecs[SVI];
    auto *WideVec =
        SubVec.size() > 1 ? concatenateVectors(Builder, SubVec) : SubVec[0];
    SVI->replaceAllUsesWith(WideVec);
  }

  return true;
}

template <typename Iter>
bool hasNearbyPairedStore(Iter It, Iter End, Value *Ptr, const DataLayout &DL) {
  int MaxLookupDist = 20;
  unsigned IdxWidth = DL.getIndexSizeInBits(0);
  APInt OffsetA(IdxWidth, 0), OffsetB(IdxWidth, 0);
  const Value *PtrA1 =
      Ptr->stripAndAccumulateInBoundsConstantOffsets(DL, OffsetA);

  while (++It != End) {
    if (It->isDebugOrPseudoInst())
      continue;
    if (MaxLookupDist-- == 0)
      break;
    if (const auto *SI = dyn_cast<StoreInst>(&*It)) {
      const Value *PtrB1 =
          SI->getPointerOperand()->stripAndAccumulateInBoundsConstantOffsets(
              DL, OffsetB);
      if (PtrA1 == PtrB1 &&
          (OffsetA.sextOrTrunc(IdxWidth) - OffsetB.sextOrTrunc(IdxWidth))
                  .abs() == 16)
        return true;
    }
  }

  return false;
}

/// Lower an interleaved store into a stN intrinsic.
///
/// E.g. Lower an interleaved store (Factor = 3):
///        %i.vec = shuffle <8 x i32> %v0, <8 x i32> %v1,
///                 <0, 4, 8, 1, 5, 9, 2, 6, 10, 3, 7, 11>
///        store <12 x i32> %i.vec, <12 x i32>* %ptr
///
///      Into:
///        %sub.v0 = shuffle <8 x i32> %v0, <8 x i32> v1, <0, 1, 2, 3>
///        %sub.v1 = shuffle <8 x i32> %v0, <8 x i32> v1, <4, 5, 6, 7>
///        %sub.v2 = shuffle <8 x i32> %v0, <8 x i32> v1, <8, 9, 10, 11>
///        call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
///
/// Note that the new shufflevectors will be removed and we'll only generate one
/// st3 instruction in CodeGen.
///
/// Example for a more general valid mask (Factor 3). Lower:
///        %i.vec = shuffle <32 x i32> %v0, <32 x i32> %v1,
///                 <4, 32, 16, 5, 33, 17, 6, 34, 18, 7, 35, 19>
///        store <12 x i32> %i.vec, <12 x i32>* %ptr
///
///      Into:
///        %sub.v0 = shuffle <32 x i32> %v0, <32 x i32> v1, <4, 5, 6, 7>
///        %sub.v1 = shuffle <32 x i32> %v0, <32 x i32> v1, <32, 33, 34, 35>
///        %sub.v2 = shuffle <32 x i32> %v0, <32 x i32> v1, <16, 17, 18, 19>
///        call void llvm.aarch64.neon.st3(%sub.v0, %sub.v1, %sub.v2, %ptr)
bool AArch64TargetLowering::lowerInterleavedStore(Instruction *Store,
                                                  Value *LaneMask,
                                                  ShuffleVectorInst *SVI,
                                                  unsigned Factor) const {

  assert(Factor >= 2 && Factor <= getMaxSupportedInterleaveFactor() &&
         "Invalid interleave factor");
  auto *SI = dyn_cast<StoreInst>(Store);
  if (!SI)
    return false;
  assert(!LaneMask && "Unexpected mask on store");

  auto *VecTy = cast<FixedVectorType>(SVI->getType());
  assert(VecTy->getNumElements() % Factor == 0 && "Invalid interleaved store");

  unsigned LaneLen = VecTy->getNumElements() / Factor;
  Type *EltTy = VecTy->getElementType();
  auto *SubVecTy = FixedVectorType::get(EltTy, LaneLen);

  const DataLayout &DL = SI->getDataLayout();
  bool UseScalable;

  // Skip if we do not have NEON and skip illegal vector types. We can
  // "legalize" wide vector types into multiple interleaved accesses as long as
  // the vector types are divisible by 128.
  if (!isLegalInterleavedAccessType(SubVecTy, DL, UseScalable))
    return false;

  unsigned NumStores = getNumInterleavedAccesses(SubVecTy, DL, UseScalable);

  Value *Op0 = SVI->getOperand(0);
  Value *Op1 = SVI->getOperand(1);
  IRBuilder<> Builder(SI);

  // StN intrinsics don't support pointer vectors as arguments. Convert pointer
  // vectors to integer vectors.
  if (EltTy->isPointerTy()) {
    Type *IntTy = DL.getIntPtrType(EltTy);
    unsigned NumOpElts =
        cast<FixedVectorType>(Op0->getType())->getNumElements();

    // Convert to the corresponding integer vector.
    auto *IntVecTy = FixedVectorType::get(IntTy, NumOpElts);
    Op0 = Builder.CreatePtrToInt(Op0, IntVecTy);
    Op1 = Builder.CreatePtrToInt(Op1, IntVecTy);

    SubVecTy = FixedVectorType::get(IntTy, LaneLen);
  }

  // If we're going to generate more than one store, reset the lane length
  // and sub-vector type to something legal.
  LaneLen /= NumStores;
  SubVecTy = FixedVectorType::get(SubVecTy->getElementType(), LaneLen);

  auto *STVTy = UseScalable ? cast<VectorType>(getSVEContainerIRType(SubVecTy))
                            : SubVecTy;

  // The base address of the store.
  Value *BaseAddr = SI->getPointerOperand();

  auto Mask = SVI->getShuffleMask();

  // Sanity check if all the indices are NOT in range.
  // If mask is `poison`, `Mask` may be a vector of -1s.
  // If all of them are `poison`, OOB read will happen later.
  if (llvm::all_of(Mask, [](int Idx) { return Idx == PoisonMaskElem; })) {
    return false;
  }
  // A 64bit st2 which does not start at element 0 will involved adding extra
  // ext elements making the st2 unprofitable, and if there is a nearby store
  // that points to BaseAddr+16 or BaseAddr-16 then it can be better left as a
  // zip;ldp pair which has higher throughput.
  if (Factor == 2 && SubVecTy->getPrimitiveSizeInBits() == 64 &&
      (Mask[0] != 0 ||
       hasNearbyPairedStore(SI->getIterator(), SI->getParent()->end(), BaseAddr,
                            DL) ||
       hasNearbyPairedStore(SI->getReverseIterator(), SI->getParent()->rend(),
                            BaseAddr, DL)))
    return false;

  Type *PtrTy = SI->getPointerOperandType();
  Type *PredTy = VectorType::get(Type::getInt1Ty(STVTy->getContext()),
                                 STVTy->getElementCount());

  Function *StNFunc = getStructuredStoreFunction(SI->getModule(), Factor,
                                                 UseScalable, STVTy, PtrTy);

  Value *PTrue = nullptr;
  if (UseScalable) {
    std::optional<unsigned> PgPattern =
        getSVEPredPatternFromNumElements(SubVecTy->getNumElements());
    if (Subtarget->getMinSVEVectorSizeInBits() ==
            Subtarget->getMaxSVEVectorSizeInBits() &&
        Subtarget->getMinSVEVectorSizeInBits() ==
            DL.getTypeSizeInBits(SubVecTy))
      PgPattern = AArch64SVEPredPattern::all;

    auto *PTruePat =
        ConstantInt::get(Type::getInt32Ty(STVTy->getContext()), *PgPattern);
    PTrue = Builder.CreateIntrinsic(Intrinsic::aarch64_sve_ptrue, {PredTy},
                                    {PTruePat});
  }

  for (unsigned StoreCount = 0; StoreCount < NumStores; ++StoreCount) {

    SmallVector<Value *, 5> Ops;

    // Split the shufflevector operands into sub vectors for the new stN call.
    for (unsigned i = 0; i < Factor; i++) {
      Value *Shuffle;
      unsigned IdxI = StoreCount * LaneLen * Factor + i;
      if (Mask[IdxI] >= 0) {
        Shuffle = Builder.CreateShuffleVector(
            Op0, Op1, createSequentialMask(Mask[IdxI], LaneLen, 0));
      } else {
        unsigned StartMask = 0;
        for (unsigned j = 1; j < LaneLen; j++) {
          unsigned IdxJ = StoreCount * LaneLen * Factor + j * Factor + i;
          if (Mask[IdxJ] >= 0) {
            StartMask = Mask[IdxJ] - j;
            break;
          }
        }
        // Note: Filling undef gaps with random elements is ok, since
        // those elements were being written anyway (with undefs).
        // In the case of all undefs we're defaulting to using elems from 0
        // Note: StartMask cannot be negative, it's checked in
        // isReInterleaveMask
        Shuffle = Builder.CreateShuffleVector(
            Op0, Op1, createSequentialMask(StartMask, LaneLen, 0));
      }

      if (UseScalable)
        Shuffle = Builder.CreateInsertVector(STVTy, PoisonValue::get(STVTy),
                                             Shuffle, uint64_t(0));

      Ops.push_back(Shuffle);
    }

    if (UseScalable)
      Ops.push_back(PTrue);

    // If we generating more than one store, we compute the base address of
    // subsequent stores as an offset from the previous.
    if (StoreCount > 0)
      BaseAddr = Builder.CreateConstGEP1_32(SubVecTy->getElementType(),
                                            BaseAddr, LaneLen * Factor);

    Ops.push_back(BaseAddr);
    Builder.CreateCall(StNFunc, Ops);
  }
  return true;
}

bool AArch64TargetLowering::lowerDeinterleaveIntrinsicToLoad(
    Instruction *Load, Value *Mask, IntrinsicInst *DI) const {
  const unsigned Factor = getDeinterleaveIntrinsicFactor(DI->getIntrinsicID());
  if (Factor != 2 && Factor != 4) {
    LLVM_DEBUG(dbgs() << "Matching ld2 and ld4 patterns failed\n");
    return false;
  }
  auto *LI = dyn_cast<LoadInst>(Load);
  if (!LI)
    return false;
  assert(!Mask && "Unexpected mask on a load\n");

  VectorType *VTy = getDeinterleavedVectorType(DI);

  const DataLayout &DL = LI->getModule()->getDataLayout();
  bool UseScalable;
  if (!isLegalInterleavedAccessType(VTy, DL, UseScalable))
    return false;

  // TODO: Add support for using SVE instructions with fixed types later, using
  // the code from lowerInterleavedLoad to obtain the correct container type.
  if (UseScalable && !VTy->isScalableTy())
    return false;

  unsigned NumLoads = getNumInterleavedAccesses(VTy, DL, UseScalable);
  VectorType *LdTy =
      VectorType::get(VTy->getElementType(),
                      VTy->getElementCount().divideCoefficientBy(NumLoads));

  Type *PtrTy = LI->getPointerOperandType();
  Function *LdNFunc = getStructuredLoadFunction(LI->getModule(), Factor,
                                                UseScalable, LdTy, PtrTy);

  IRBuilder<> Builder(LI);
  Value *Pred = nullptr;
  if (UseScalable)
    Pred =
        Builder.CreateVectorSplat(LdTy->getElementCount(), Builder.getTrue());

  Value *BaseAddr = LI->getPointerOperand();
  Value *Result = nullptr;
  if (NumLoads > 1) {
    // Create multiple legal small ldN.
    SmallVector<Value *, 4> ExtractedLdValues(Factor, PoisonValue::get(VTy));
    for (unsigned I = 0; I < NumLoads; ++I) {
      Value *Offset = Builder.getInt64(I * Factor);

      Value *Address = Builder.CreateGEP(LdTy, BaseAddr, {Offset});
      Value *LdN = nullptr;
      if (UseScalable)
        LdN = Builder.CreateCall(LdNFunc, {Pred, Address}, "ldN");
      else
        LdN = Builder.CreateCall(LdNFunc, Address, "ldN");
      Value *Idx =
          Builder.getInt64(I * LdTy->getElementCount().getKnownMinValue());
      for (unsigned J = 0; J < Factor; ++J) {
        ExtractedLdValues[J] = Builder.CreateInsertVector(
            VTy, ExtractedLdValues[J], Builder.CreateExtractValue(LdN, J), Idx);
      }
      LLVM_DEBUG(dbgs() << "LdN4 res: "; LdN->dump());
    }

    // Merge the values from different factors.
    Result = PoisonValue::get(DI->getType());
    for (unsigned J = 0; J < Factor; ++J)
      Result = Builder.CreateInsertValue(Result, ExtractedLdValues[J], J);
  } else {
    if (UseScalable)
      Result = Builder.CreateCall(LdNFunc, {Pred, BaseAddr}, "ldN");
    else
      Result = Builder.CreateCall(LdNFunc, BaseAddr, "ldN");
  }

  // Replace output of deinterleave2 intrinsic by output of ldN2/ldN4
  DI->replaceAllUsesWith(Result);
  return true;
}

bool AArch64TargetLowering::lowerInterleaveIntrinsicToStore(
    Instruction *Store, Value *Mask,
    ArrayRef<Value *> InterleavedValues) const {
  unsigned Factor = InterleavedValues.size();
  if (Factor != 2 && Factor != 4) {
    LLVM_DEBUG(dbgs() << "Matching st2 and st4 patterns failed\n");
    return false;
  }
  StoreInst *SI = dyn_cast<StoreInst>(Store);
  if (!SI)
    return false;
  assert(!Mask && "Unexpected mask on plain store");

  VectorType *VTy = cast<VectorType>(InterleavedValues[0]->getType());
  const DataLayout &DL = SI->getModule()->getDataLayout();

  bool UseScalable;
  if (!isLegalInterleavedAccessType(VTy, DL, UseScalable))
    return false;

  // TODO: Add support for using SVE instructions with fixed types later, using
  // the code from lowerInterleavedStore to obtain the correct container type.
  if (UseScalable && !VTy->isScalableTy())
    return false;

  unsigned NumStores = getNumInterleavedAccesses(VTy, DL, UseScalable);

  VectorType *StTy =
      VectorType::get(VTy->getElementType(),
                      VTy->getElementCount().divideCoefficientBy(NumStores));

  Type *PtrTy = SI->getPointerOperandType();
  Function *StNFunc = getStructuredStoreFunction(SI->getModule(), Factor,
                                                 UseScalable, StTy, PtrTy);

  IRBuilder<> Builder(SI);

  Value *BaseAddr = SI->getPointerOperand();
  Value *Pred = nullptr;

  if (UseScalable)
    Pred =
        Builder.CreateVectorSplat(StTy->getElementCount(), Builder.getTrue());

  auto ExtractedValues = InterleavedValues;
  SmallVector<Value *, 4> StoreOperands(InterleavedValues);
  if (UseScalable)
    StoreOperands.push_back(Pred);
  StoreOperands.push_back(BaseAddr);
  for (unsigned I = 0; I < NumStores; ++I) {
    Value *Address = BaseAddr;
    if (NumStores > 1) {
      Value *Offset = Builder.getInt64(I * Factor);
      Address = Builder.CreateGEP(StTy, BaseAddr, {Offset});
      Value *Idx =
          Builder.getInt64(I * StTy->getElementCount().getKnownMinValue());
      for (unsigned J = 0; J < Factor; J++) {
        StoreOperands[J] =
            Builder.CreateExtractVector(StTy, ExtractedValues[J], Idx);
      }
      // update the address
      StoreOperands[StoreOperands.size() - 1] = Address;
    }
    Builder.CreateCall(StNFunc, StoreOperands);
  }
  return true;
}

EVT AArch64TargetLowering::getOptimalMemOpType(
    LLVMContext &Context, const MemOp &Op,
    const AttributeList &FuncAttributes) const {
  bool CanImplicitFloat = !FuncAttributes.hasFnAttr(Attribute::NoImplicitFloat);
  bool CanUseNEON = Subtarget->hasNEON() && CanImplicitFloat;
  bool CanUseFP = Subtarget->hasFPARMv8() && CanImplicitFloat;
  // Only use AdvSIMD to implement memset of 32-byte and above. It would have
  // taken one instruction to materialize the v2i64 zero and one store (with
  // restrictive addressing mode). Just do i64 stores.
  bool IsSmallMemset = Op.isMemset() && Op.size() < 32;
  auto AlignmentIsAcceptable = [&](EVT VT, Align AlignCheck) {
    if (Op.isAligned(AlignCheck))
      return true;
    unsigned Fast;
    return allowsMisalignedMemoryAccesses(VT, 0, Align(1),
                                          MachineMemOperand::MONone, &Fast) &&
           Fast;
  };

  if (CanUseNEON && Op.isMemset() && !IsSmallMemset &&
      AlignmentIsAcceptable(MVT::v16i8, Align(16)))
    return MVT::v16i8;
  if (CanUseFP && !IsSmallMemset && AlignmentIsAcceptable(MVT::f128, Align(16)))
    return MVT::f128;
  if (Op.size() >= 8 && AlignmentIsAcceptable(MVT::i64, Align(8)))
    return MVT::i64;
  if (Op.size() >= 4 && AlignmentIsAcceptable(MVT::i32, Align(4)))
    return MVT::i32;
  return MVT::Other;
}

LLT AArch64TargetLowering::getOptimalMemOpLLT(
    const MemOp &Op, const AttributeList &FuncAttributes) const {
  bool CanImplicitFloat = !FuncAttributes.hasFnAttr(Attribute::NoImplicitFloat);
  bool CanUseNEON = Subtarget->hasNEON() && CanImplicitFloat;
  bool CanUseFP = Subtarget->hasFPARMv8() && CanImplicitFloat;
  // Only use AdvSIMD to implement memset of 32-byte and above. It would have
  // taken one instruction to materialize the v2i64 zero and one store (with
  // restrictive addressing mode). Just do i64 stores.
  bool IsSmallMemset = Op.isMemset() && Op.size() < 32;
  auto AlignmentIsAcceptable = [&](EVT VT, Align AlignCheck) {
    if (Op.isAligned(AlignCheck))
      return true;
    unsigned Fast;
    return allowsMisalignedMemoryAccesses(VT, 0, Align(1),
                                          MachineMemOperand::MONone, &Fast) &&
           Fast;
  };

  if (CanUseNEON && Op.isMemset() && !IsSmallMemset &&
      AlignmentIsAcceptable(MVT::v2i64, Align(16)))
    return LLT::fixed_vector(2, 64);
  if (CanUseFP && !IsSmallMemset && AlignmentIsAcceptable(MVT::f128, Align(16)))
    return LLT::scalar(128);
  if (Op.size() >= 8 && AlignmentIsAcceptable(MVT::i64, Align(8)))
    return LLT::scalar(64);
  if (Op.size() >= 4 && AlignmentIsAcceptable(MVT::i32, Align(4)))
    return LLT::scalar(32);
  return LLT();
}

// 12-bit optionally shifted immediates are legal for adds.
bool AArch64TargetLowering::isLegalAddImmediate(int64_t Immed) const {
  if (Immed == std::numeric_limits<int64_t>::min()) {
    return false;
  }
  // Same encoding for add/sub, just flip the sign.
  return isLegalArithImmed((uint64_t)std::abs(Immed));
}

bool AArch64TargetLowering::isLegalAddScalableImmediate(int64_t Imm) const {
  // We will only emit addvl/inc* instructions for SVE2
  if (!Subtarget->hasSVE2())
    return false;

  // addvl's immediates are in terms of the number of bytes in a register.
  // Since there are 16 in the base supported size (128bits), we need to
  // divide the immediate by that much to give us a useful immediate to
  // multiply by vscale. We can't have a remainder as a result of this.
  if (Imm % 16 == 0)
    return isInt<6>(Imm / 16);

  // Inc[b|h|w|d] instructions take a pattern and a positive immediate
  // multiplier. For now, assume a pattern of 'all'. Incb would be a subset
  // of addvl as a result, so only take h|w|d into account.
  // Dec[h|w|d] will cover subtractions.
  // Immediates are in the range [1,16], so we can't do a 2's complement check.
  // FIXME: Can we make use of other patterns to cover other immediates?

  // inch|dech
  if (Imm % 8 == 0)
    return std::abs(Imm / 8) <= 16;
  // incw|decw
  if (Imm % 4 == 0)
    return std::abs(Imm / 4) <= 16;
  // incd|decd
  if (Imm % 2 == 0)
    return std::abs(Imm / 2) <= 16;

  return false;
}

// Return false to prevent folding
// (mul (add x, c1), c2) -> (add (mul x, c2), c2*c1) in DAGCombine,
// if the folding leads to worse code.
bool AArch64TargetLowering::isMulAddWithConstProfitable(
    SDValue AddNode, SDValue ConstNode) const {
  // Let the DAGCombiner decide for vector types and large types.
  const EVT VT = AddNode.getValueType();
  if (VT.isVector() || VT.getScalarSizeInBits() > 64)
    return true;

  // It is worse if c1 is legal add immediate, while c1*c2 is not
  // and has to be composed by at least two instructions.
  const ConstantSDNode *C1Node = cast<ConstantSDNode>(AddNode.getOperand(1));
  const ConstantSDNode *C2Node = cast<ConstantSDNode>(ConstNode);
  const int64_t C1 = C1Node->getSExtValue();
  const APInt C1C2 = C1Node->getAPIntValue() * C2Node->getAPIntValue();
  if (!isLegalAddImmediate(C1) || isLegalAddImmediate(C1C2.getSExtValue()))
    return true;
  SmallVector<AArch64_IMM::ImmInsnModel, 4> Insn;
  // Adapt to the width of a register.
  unsigned BitSize = VT.getSizeInBits() <= 32 ? 32 : 64;
  AArch64_IMM::expandMOVImm(C1C2.getZExtValue(), BitSize, Insn);
  if (Insn.size() > 1)
    return false;

  // Default to true and let the DAGCombiner decide.
  return true;
}

// Integer comparisons are implemented with ADDS/SUBS, so the range of valid
// immediates is the same as for an add or a sub.
bool AArch64TargetLowering::isLegalICmpImmediate(int64_t Immed) const {
  return isLegalAddImmediate(Immed);
}

/// isLegalAddressingMode - Return true if the addressing mode represented
/// by AM is legal for this target, for a load/store of the specified type.
bool AArch64TargetLowering::isLegalAddressingMode(const DataLayout &DL,
                                                  const AddrMode &AMode, Type *Ty,
                                                  unsigned AS, Instruction *I) const {
  // AArch64 has five basic addressing modes:
  //  reg
  //  reg + 9-bit signed offset
  //  reg + SIZE_IN_BYTES * 12-bit unsigned offset
  //  reg1 + reg2
  //  reg + SIZE_IN_BYTES * reg

  // No global is ever allowed as a base.
  if (AMode.BaseGV)
    return false;

  // No reg+reg+imm addressing.
  if (AMode.HasBaseReg && AMode.BaseOffs && AMode.Scale)
    return false;

  // Canonicalise `1*ScaledReg + imm` into `BaseReg + imm` and
  // `2*ScaledReg` into `BaseReg + ScaledReg`
  AddrMode AM = AMode;
  if (AM.Scale && !AM.HasBaseReg) {
    if (AM.Scale == 1) {
      AM.HasBaseReg = true;
      AM.Scale = 0;
    } else if (AM.Scale == 2) {
      AM.HasBaseReg = true;
      AM.Scale = 1;
    } else {
      return false;
    }
  }

  // A base register is required in all addressing modes.
  if (!AM.HasBaseReg)
    return false;

  if (Ty->isScalableTy()) {
    if (isa<ScalableVectorType>(Ty)) {
      // See if we have a foldable vscale-based offset, for vector types which
      // are either legal or smaller than the minimum; more work will be
      // required if we need to consider addressing for types which need
      // legalization by splitting.
      uint64_t VecNumBytes = DL.getTypeSizeInBits(Ty).getKnownMinValue() / 8;
      if (AM.HasBaseReg && !AM.BaseOffs && AM.ScalableOffset && !AM.Scale &&
          (AM.ScalableOffset % VecNumBytes == 0) && VecNumBytes <= 16 &&
          isPowerOf2_64(VecNumBytes))
        return isInt<4>(AM.ScalableOffset / (int64_t)VecNumBytes);

      uint64_t VecElemNumBytes =
          DL.getTypeSizeInBits(cast<VectorType>(Ty)->getElementType()) / 8;
      return AM.HasBaseReg && !AM.BaseOffs && !AM.ScalableOffset &&
             (AM.Scale == 0 || (uint64_t)AM.Scale == VecElemNumBytes);
    }

    return AM.HasBaseReg && !AM.BaseOffs && !AM.ScalableOffset && !AM.Scale;
  }

  // No scalable offsets allowed for non-scalable types.
  if (AM.ScalableOffset)
    return false;

  // check reg + imm case:
  // i.e., reg + 0, reg + imm9, reg + SIZE_IN_BYTES * uimm12
  uint64_t NumBytes = 0;
  if (Ty->isSized()) {
    uint64_t NumBits = DL.getTypeSizeInBits(Ty);
    NumBytes = NumBits / 8;
    if (!isPowerOf2_64(NumBits))
      NumBytes = 0;
  }

  return Subtarget->getInstrInfo()->isLegalAddressingMode(NumBytes, AM.BaseOffs,
                                                          AM.Scale);
}

// Check whether the 2 offsets belong to the same imm24 range, and their high
// 12bits are same, then their high part can be decoded with the offset of add.
int64_t
AArch64TargetLowering::getPreferredLargeGEPBaseOffset(int64_t MinOffset,
                                                      int64_t MaxOffset) const {
  int64_t HighPart = MinOffset & ~0xfffULL;
  if (MinOffset >> 12 == MaxOffset >> 12 && isLegalAddImmediate(HighPart)) {
    // Rebase the value to an integer multiple of imm12.
    return HighPart;
  }

  return 0;
}

bool AArch64TargetLowering::shouldConsiderGEPOffsetSplit() const {
  // Consider splitting large offset of struct or array.
  return true;
}

bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(
    const MachineFunction &MF, EVT VT) const {
  EVT ScalarVT = VT.getScalarType();

  if (!ScalarVT.isSimple())
    return false;

  switch (ScalarVT.getSimpleVT().SimpleTy) {
  case MVT::f16:
    return Subtarget->hasFullFP16();
  case MVT::f32:
  case MVT::f64:
    return true;
  case MVT::bf16:
    return VT.isScalableVector() && Subtarget->hasSVEB16B16();
  default:
    break;
  }

  return false;
}

bool AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(const Function &F,
                                                       Type *Ty) const {
  switch (Ty->getScalarType()->getTypeID()) {
  case Type::FloatTyID:
  case Type::DoubleTyID:
    return true;
  default:
    return false;
  }
}

bool AArch64TargetLowering::generateFMAsInMachineCombiner(
    EVT VT, CodeGenOptLevel OptLevel) const {
  return (OptLevel >= CodeGenOptLevel::Aggressive) && !VT.isScalableVector() &&
         !useSVEForFixedLengthVectorVT(VT);
}

const MCPhysReg *
AArch64TargetLowering::getScratchRegisters(CallingConv::ID) const {
  // LR is a callee-save register, but we must treat it as clobbered by any call
  // site. Hence we include LR in the scratch registers, which are in turn added
  // as implicit-defs for stackmaps and patchpoints.
  static const MCPhysReg ScratchRegs[] = {
    AArch64::X16, AArch64::X17, AArch64::LR, 0
  };
  return ScratchRegs;
}

ArrayRef<MCPhysReg> AArch64TargetLowering::getRoundingControlRegisters() const {
  static const MCPhysReg RCRegs[] = {AArch64::FPCR};
  return RCRegs;
}

bool
AArch64TargetLowering::isDesirableToCommuteWithShift(const SDNode *N,
                                                     CombineLevel Level) const {
  assert((N->getOpcode() == ISD::SHL || N->getOpcode() == ISD::SRA ||
          N->getOpcode() == ISD::SRL) &&
         "Expected shift op");

  SDValue ShiftLHS = N->getOperand(0);
  EVT VT = N->getValueType(0);

  if (!ShiftLHS->hasOneUse())
    return false;

  if (ShiftLHS.getOpcode() == ISD::SIGN_EXTEND &&
      !ShiftLHS.getOperand(0)->hasOneUse())
    return false;

  // If ShiftLHS is unsigned bit extraction: ((x >> C) & mask), then do not
  // combine it with shift 'N' to let it be lowered to UBFX except:
  // ((x >> C) & mask) << C.
  if (ShiftLHS.getOpcode() == ISD::AND && (VT == MVT::i32 || VT == MVT::i64) &&
      isa<ConstantSDNode>(ShiftLHS.getOperand(1))) {
    uint64_t TruncMask = ShiftLHS.getConstantOperandVal(1);
    if (isMask_64(TruncMask)) {
      SDValue AndLHS = ShiftLHS.getOperand(0);
      if (AndLHS.getOpcode() == ISD::SRL) {
        if (auto *SRLC = dyn_cast<ConstantSDNode>(AndLHS.getOperand(1))) {
          if (N->getOpcode() == ISD::SHL)
            if (auto *SHLC = dyn_cast<ConstantSDNode>(N->getOperand(1)))
              return SRLC->getZExtValue() == SHLC->getZExtValue();
          return false;
        }
      }
    }
  }
  return true;
}

bool AArch64TargetLowering::isDesirableToCommuteXorWithShift(
    const SDNode *N) const {
  assert(N->getOpcode() == ISD::XOR &&
         (N->getOperand(0).getOpcode() == ISD::SHL ||
          N->getOperand(0).getOpcode() == ISD::SRL) &&
         "Expected XOR(SHIFT) pattern");

  // Only commute if the entire NOT mask is a hidden shifted mask.
  auto *XorC = dyn_cast<ConstantSDNode>(N->getOperand(1));
  auto *ShiftC = dyn_cast<ConstantSDNode>(N->getOperand(0).getOperand(1));
  if (XorC && ShiftC) {
    unsigned MaskIdx, MaskLen;
    if (XorC->getAPIntValue().isShiftedMask(MaskIdx, MaskLen)) {
      unsigned ShiftAmt = ShiftC->getZExtValue();
      unsigned BitWidth = N->getValueType(0).getScalarSizeInBits();
      if (N->getOperand(0).getOpcode() == ISD::SHL)
        return MaskIdx == ShiftAmt && MaskLen == (BitWidth - ShiftAmt);
      return MaskIdx == 0 && MaskLen == (BitWidth - ShiftAmt);
    }
  }

  return false;
}

bool AArch64TargetLowering::shouldFoldConstantShiftPairToMask(
    const SDNode *N, CombineLevel Level) const {
  assert(((N->getOpcode() == ISD::SHL &&
           N->getOperand(0).getOpcode() == ISD::SRL) ||
          (N->getOpcode() == ISD::SRL &&
           N->getOperand(0).getOpcode() == ISD::SHL)) &&
         "Expected shift-shift mask");
  // Don't allow multiuse shift folding with the same shift amount.
  if (!N->getOperand(0)->hasOneUse())
    return false;

  // Only fold srl(shl(x,c1),c2) iff C1 >= C2 to prevent loss of UBFX patterns.
  EVT VT = N->getValueType(0);
  if (N->getOpcode() == ISD::SRL && (VT == MVT::i32 || VT == MVT::i64)) {
    auto *C1 = dyn_cast<ConstantSDNode>(N->getOperand(0).getOperand(1));
    auto *C2 = dyn_cast<ConstantSDNode>(N->getOperand(1));
    return (!C1 || !C2 || C1->getZExtValue() >= C2->getZExtValue());
  }

  // We do not need to fold when this shifting used in specific load case:
  // (ldr x, (add x, (shl (srl x, c1) 2)))
  if (N->getOpcode() == ISD::SHL && N->hasOneUse()) {
    if (auto C2 = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
      unsigned ShlAmt = C2->getZExtValue();
      if (auto ShouldADD = *N->user_begin();
          ShouldADD->getOpcode() == ISD::ADD && ShouldADD->hasOneUse()) {
        if (auto ShouldLOAD = dyn_cast<LoadSDNode>(*ShouldADD->user_begin())) {
          unsigned ByteVT = ShouldLOAD->getMemoryVT().getSizeInBits() / 8;
          if ((1ULL << ShlAmt) == ByteVT &&
              isIndexedLoadLegal(ISD::PRE_INC, ShouldLOAD->getMemoryVT()))
            return false;
        }
      }
    }
  }

  return true;
}

bool AArch64TargetLowering::shouldFoldSelectWithIdentityConstant(
    unsigned BinOpcode, EVT VT, unsigned SelectOpcode, SDValue X,
    SDValue Y) const {
  return VT.isScalableVector() && isTypeLegal(VT) &&
         SelectOpcode == ISD::VSELECT;
}

bool AArch64TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                                              Type *Ty) const {
  assert(Ty->isIntegerTy());

  unsigned BitSize = Ty->getPrimitiveSizeInBits();
  if (BitSize == 0)
    return false;

  int64_t Val = Imm.getSExtValue();
  if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, BitSize))
    return true;

  if ((int64_t)Val < 0)
    Val = ~Val;
  if (BitSize == 32)
    Val &= (1LL << 32) - 1;

  unsigned Shift = llvm::Log2_64((uint64_t)Val) / 16;
  // MOVZ is free so return true for one or fewer MOVK.
  return Shift < 3;
}

bool AArch64TargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
                                                    unsigned Index) const {
  if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
    return false;

  return (Index == 0 || Index == ResVT.getVectorMinNumElements());
}

/// Turn vector tests of the signbit in the form of:
///   xor (sra X, elt_size(X)-1), -1
/// into:
///   cmge X, X, #0
static SDValue foldVectorXorShiftIntoCmp(SDNode *N, SelectionDAG &DAG,
                                         const AArch64Subtarget *Subtarget) {
  EVT VT = N->getValueType(0);
  if (!Subtarget->hasNEON() || !VT.isVector())
    return SDValue();

  // There must be a shift right algebraic before the xor, and the xor must be a
  // 'not' operation.
  SDValue Shift = N->getOperand(0);
  SDValue Ones = N->getOperand(1);
  if (Shift.getOpcode() != AArch64ISD::VASHR || !Shift.hasOneUse() ||
      !ISD::isBuildVectorAllOnes(Ones.getNode()))
    return SDValue();

  // The shift should be smearing the sign bit across each vector element.
  auto *ShiftAmt = dyn_cast<ConstantSDNode>(Shift.getOperand(1));
  EVT ShiftEltTy = Shift.getValueType().getVectorElementType();
  if (!ShiftAmt || ShiftAmt->getZExtValue() != ShiftEltTy.getSizeInBits() - 1)
    return SDValue();

  SDLoc DL(N);
  SDValue Zero = DAG.getConstant(0, DL, Shift.getValueType());
  return DAG.getSetCC(DL, VT, Shift.getOperand(0), Zero, ISD::SETGE);
}

// Given a vecreduce_add node, detect the below pattern and convert it to the
// node sequence with UABDL, [S|U]ADB and UADDLP.
//
// i32 vecreduce_add(
//  v16i32 abs(
//    v16i32 sub(
//     v16i32 [sign|zero]_extend(v16i8 a), v16i32 [sign|zero]_extend(v16i8 b))))
//
//  or
//
// i32 vecreduce_add(
//  v16i32 zext(
//   v16i16 abs(
//    v16i16 sub(
//     v16i16 [sign|zero]_extend(v16i8 a), v16i16 [sign|zero]_extend(v16i8 b))))
//
// =================>
// i32 vecreduce_add(
//   v4i32 UADDLP(
//     v8i16 add(
//       v8i16 zext(
//         v8i8 [S|U]ABD low8:v16i8 a, low8:v16i8 b
//       v8i16 zext(
//         v8i8 [S|U]ABD high8:v16i8 a, high8:v16i8 b
static SDValue performVecReduceAddCombineWithUADDLP(SDNode *N,
                                                    SelectionDAG &DAG) {
  // Assumed i32 vecreduce_add
  if (N->getValueType(0) != MVT::i32)
    return SDValue();

  SDValue VecReduceOp0 = N->getOperand(0);
  bool SawTrailingZext = false;
  // Look through an optional post-ABS ZEXT from v16i16 -> v16i32.
  if (VecReduceOp0.getOpcode() == ISD::ZERO_EXTEND &&
      VecReduceOp0->getValueType(0) == MVT::v16i32 &&
      VecReduceOp0->getOperand(0)->getOpcode() == ISD::ABS &&
      VecReduceOp0->getOperand(0)->getValueType(0) == MVT::v16i16) {
    SawTrailingZext = true;
    VecReduceOp0 = VecReduceOp0.getOperand(0);
  }

  // Peel off an optional post-ABS extend (v16i16 -> v16i32).
  MVT AbsInputVT = SawTrailingZext ? MVT::v16i16 : MVT::v16i32;
  // Assumed v16i16 or v16i32 abs input
  unsigned Opcode = VecReduceOp0.getOpcode();
  if (Opcode != ISD::ABS || VecReduceOp0->getValueType(0) != AbsInputVT)
    return SDValue();

  SDValue ABS = VecReduceOp0;
  // Assumed v16i16 or v16i32 sub
  if (ABS->getOperand(0)->getOpcode() != ISD::SUB ||
      ABS->getOperand(0)->getValueType(0) != AbsInputVT)
    return SDValue();

  SDValue SUB = ABS->getOperand(0);
  unsigned Opcode0 = SUB->getOperand(0).getOpcode();
  unsigned Opcode1 = SUB->getOperand(1).getOpcode();
  // Assumed v16i16 or v16i32 type
  if (SUB->getOperand(0)->getValueType(0) != AbsInputVT ||
      SUB->getOperand(1)->getValueType(0) != AbsInputVT)
    return SDValue();

  // Assumed zext or sext
  bool IsZExt = false;
  if (Opcode0 == ISD::ZERO_EXTEND && Opcode1 == ISD::ZERO_EXTEND) {
    IsZExt = true;
  } else if (Opcode0 == ISD::SIGN_EXTEND && Opcode1 == ISD::SIGN_EXTEND) {
    IsZExt = false;
  } else
    return SDValue();

  SDValue EXT0 = SUB->getOperand(0);
  SDValue EXT1 = SUB->getOperand(1);
  // Assumed zext's operand has v16i8 type
  if (EXT0->getOperand(0)->getValueType(0) != MVT::v16i8 ||
      EXT1->getOperand(0)->getValueType(0) != MVT::v16i8)
    return SDValue();

  // Pattern is detected. Let's convert it to sequence of nodes.
  SDLoc DL(N);

  // First, create the node pattern of UABD/SABD.
  SDValue UABDHigh8Op0 =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, EXT0->getOperand(0),
                  DAG.getConstant(8, DL, MVT::i64));
  SDValue UABDHigh8Op1 =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, EXT1->getOperand(0),
                  DAG.getConstant(8, DL, MVT::i64));
  SDValue UABDHigh8 = DAG.getNode(IsZExt ? ISD::ABDU : ISD::ABDS, DL, MVT::v8i8,
                                  UABDHigh8Op0, UABDHigh8Op1);
  SDValue UABDL = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, UABDHigh8);

  // Second, create the node pattern of UABAL.
  SDValue UABDLo8Op0 =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, EXT0->getOperand(0),
                  DAG.getConstant(0, DL, MVT::i64));
  SDValue UABDLo8Op1 =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, EXT1->getOperand(0),
                  DAG.getConstant(0, DL, MVT::i64));
  SDValue UABDLo8 = DAG.getNode(IsZExt ? ISD::ABDU : ISD::ABDS, DL, MVT::v8i8,
                                UABDLo8Op0, UABDLo8Op1);
  SDValue ZExtUABD = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, UABDLo8);
  SDValue UABAL = DAG.getNode(ISD::ADD, DL, MVT::v8i16, UABDL, ZExtUABD);

  // Third, create the node of UADDLP.
  SDValue UADDLP = DAG.getNode(AArch64ISD::UADDLP, DL, MVT::v4i32, UABAL);

  // Fourth, create the node of VECREDUCE_ADD.
  return DAG.getNode(ISD::VECREDUCE_ADD, DL, MVT::i32, UADDLP);
}

static SDValue
performActiveLaneMaskCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                             const AArch64Subtarget *ST) {
  if (DCI.isBeforeLegalize())
    return SDValue();

  if (SDValue While = optimizeIncrementingWhile(N, DCI.DAG, /*IsSigned=*/false,
                                                /*IsEqual=*/false))
    return While;

  if (!N->getValueType(0).isScalableVector() ||
      (!ST->hasSVE2p1() && !(ST->hasSME2() && ST->isStreaming())))
    return SDValue();

  unsigned NumUses = N->use_size();
  auto MaskEC = N->getValueType(0).getVectorElementCount();
  if (!MaskEC.isKnownMultipleOf(NumUses))
    return SDValue();

  ElementCount ExtMinEC = MaskEC.divideCoefficientBy(NumUses);
  if (ExtMinEC.getKnownMinValue() < 2)
    return SDValue();

  SmallVector<SDNode *> Extracts(NumUses, nullptr);
  for (SDNode *Use : N->users()) {
    if (Use->getOpcode() != ISD::EXTRACT_SUBVECTOR)
      return SDValue();

    // Ensure the extract type is correct (e.g. if NumUses is 4 and
    // the mask return type is nxv8i1, each extract should be nxv2i1.
    if (Use->getValueType(0).getVectorElementCount() != ExtMinEC)
      return SDValue();

    // There should be exactly one extract for each part of the mask.
    unsigned Offset = Use->getConstantOperandVal(1);
    unsigned Part = Offset / ExtMinEC.getKnownMinValue();
    if (Extracts[Part] != nullptr)
      return SDValue();

    Extracts[Part] = Use;
  }

  SelectionDAG &DAG = DCI.DAG;
  SDLoc DL(N);
  SDValue ID =
      DAG.getTargetConstant(Intrinsic::aarch64_sve_whilelo_x2, DL, MVT::i64);

  SDValue Idx = N->getOperand(0);
  SDValue TC = N->getOperand(1);
  EVT OpVT = Idx.getValueType();
  if (OpVT != MVT::i64) {
    Idx = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Idx);
    TC = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, TC);
  }

  // Create the whilelo_x2 intrinsics from each pair of extracts
  EVT ExtVT = Extracts[0]->getValueType(0);
  auto R =
      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, {ExtVT, ExtVT}, {ID, Idx, TC});
  DCI.CombineTo(Extracts[0], R.getValue(0));
  DCI.CombineTo(Extracts[1], R.getValue(1));

  if (NumUses == 2)
    return SDValue(N, 0);

  auto Elts = DAG.getElementCount(DL, OpVT, ExtVT.getVectorElementCount() * 2);
  for (unsigned I = 2; I < NumUses; I += 2) {
    // After the first whilelo_x2, we need to increment the starting value.
    Idx = DAG.getNode(ISD::UADDSAT, DL, OpVT, Idx, Elts);
    R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, {ExtVT, ExtVT}, {ID, Idx, TC});
    DCI.CombineTo(Extracts[I], R.getValue(0));
    DCI.CombineTo(Extracts[I + 1], R.getValue(1));
  }

  return SDValue(N, 0);
}

// Turn a v8i8/v16i8 extended vecreduce into a udot/sdot and vecreduce
//   vecreduce.add(ext(A)) to vecreduce.add(DOT(zero, A, one))
//   vecreduce.add(mul(ext(A), ext(B))) to vecreduce.add(DOT(zero, A, B))
// If we have vectors larger than v16i8 we extract v16i8 vectors,
// Follow the same steps above to get DOT instructions concatenate them
// and generate vecreduce.add(concat_vector(DOT, DOT2, ..)).
static SDValue performVecReduceAddCombine(SDNode *N, SelectionDAG &DAG,
                                          const AArch64Subtarget *ST) {
  if (!ST->isNeonAvailable())
    return SDValue();

  if (!ST->hasDotProd())
    return performVecReduceAddCombineWithUADDLP(N, DAG);

  SDValue Op0 = N->getOperand(0);
  if (N->getValueType(0) != MVT::i32 || Op0.getValueType().isScalableVT() ||
      Op0.getValueType().getVectorElementType() != MVT::i32)
    return SDValue();

  unsigned ExtOpcode = Op0.getOpcode();
  SDValue A = Op0;
  SDValue B;
  unsigned DotOpcode;
  if (ExtOpcode == ISD::MUL) {
    A = Op0.getOperand(0);
    B = Op0.getOperand(1);
    if (A.getOperand(0).getValueType() != B.getOperand(0).getValueType())
      return SDValue();
    auto OpCodeA = A.getOpcode();
    if (OpCodeA != ISD::ZERO_EXTEND && OpCodeA != ISD::SIGN_EXTEND)
      return SDValue();

    auto OpCodeB = B.getOpcode();
    if (OpCodeB != ISD::ZERO_EXTEND && OpCodeB != ISD::SIGN_EXTEND)
      return SDValue();

    if (OpCodeA == OpCodeB) {
      DotOpcode =
          OpCodeA == ISD::ZERO_EXTEND ? AArch64ISD::UDOT : AArch64ISD::SDOT;
    } else {
      // Check USDOT support support
      if (!ST->hasMatMulInt8())
        return SDValue();
      DotOpcode = AArch64ISD::USDOT;
      if (OpCodeA == ISD::SIGN_EXTEND)
        std::swap(A, B);
    }
  } else if (ExtOpcode == ISD::ZERO_EXTEND) {
    DotOpcode = AArch64ISD::UDOT;
  } else if (ExtOpcode == ISD::SIGN_EXTEND) {
    DotOpcode = AArch64ISD::SDOT;
  } else {
    return SDValue();
  }

  EVT Op0VT = A.getOperand(0).getValueType();
  bool IsValidElementCount = Op0VT.getVectorNumElements() % 8 == 0;
  bool IsValidSize = Op0VT.getScalarSizeInBits() == 8;
  if (!IsValidElementCount || !IsValidSize)
    return SDValue();

  SDLoc DL(Op0);
  // For non-mla reductions B can be set to 1. For MLA we take the operand of
  // the extend B.
  if (!B)
    B = DAG.getConstant(1, DL, Op0VT);
  else
    B = B.getOperand(0);

  unsigned IsMultipleOf16 = Op0VT.getVectorNumElements() % 16 == 0;
  unsigned NumOfVecReduce;
  EVT TargetType;
  if (IsMultipleOf16) {
    NumOfVecReduce = Op0VT.getVectorNumElements() / 16;
    TargetType = MVT::v4i32;
  } else {
    NumOfVecReduce = Op0VT.getVectorNumElements() / 8;
    TargetType = MVT::v2i32;
  }
  // Handle the case where we need to generate only one Dot operation.
  if (NumOfVecReduce == 1) {
    SDValue Zeros = DAG.getConstant(0, DL, TargetType);
    SDValue Dot = DAG.getNode(DotOpcode, DL, Zeros.getValueType(), Zeros,
                              A.getOperand(0), B);
    return DAG.getNode(ISD::VECREDUCE_ADD, DL, N->getValueType(0), Dot);
  }
  // Generate Dot instructions that are multiple of 16.
  unsigned VecReduce16Num = Op0VT.getVectorNumElements() / 16;
  SmallVector<SDValue, 4> SDotVec16;
  unsigned I = 0;
  for (; I < VecReduce16Num; I += 1) {
    SDValue Zeros = DAG.getConstant(0, DL, MVT::v4i32);
    SDValue Op0 =
        DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v16i8, A.getOperand(0),
                    DAG.getConstant(I * 16, DL, MVT::i64));
    SDValue Op1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v16i8, B,
                              DAG.getConstant(I * 16, DL, MVT::i64));
    SDValue Dot =
        DAG.getNode(DotOpcode, DL, Zeros.getValueType(), Zeros, Op0, Op1);
    SDotVec16.push_back(Dot);
  }
  // Concatenate dot operations.
  EVT SDot16EVT =
      EVT::getVectorVT(*DAG.getContext(), MVT::i32, 4 * VecReduce16Num);
  SDValue ConcatSDot16 =
      DAG.getNode(ISD::CONCAT_VECTORS, DL, SDot16EVT, SDotVec16);
  SDValue VecReduceAdd16 =
      DAG.getNode(ISD::VECREDUCE_ADD, DL, N->getValueType(0), ConcatSDot16);
  unsigned VecReduce8Num = (Op0VT.getVectorNumElements() % 16) / 8;
  if (VecReduce8Num == 0)
    return VecReduceAdd16;

  // Generate the remainder Dot operation that is multiple of 8.
  SDValue Zeros = DAG.getConstant(0, DL, MVT::v2i32);
  SDValue Vec8Op0 =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, A.getOperand(0),
                  DAG.getConstant(I * 16, DL, MVT::i64));
  SDValue Vec8Op1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8, B,
                                DAG.getConstant(I * 16, DL, MVT::i64));
  SDValue Dot =
      DAG.getNode(DotOpcode, DL, Zeros.getValueType(), Zeros, Vec8Op0, Vec8Op1);
  SDValue VecReduceAdd8 =
      DAG.getNode(ISD::VECREDUCE_ADD, DL, N->getValueType(0), Dot);
  return DAG.getNode(ISD::ADD, DL, N->getValueType(0), VecReduceAdd16,
                     VecReduceAdd8);
}

// Given an (integer) vecreduce, we know the order of the inputs does not
// matter. We can convert UADDV(add(zext(extract_lo(x)), zext(extract_hi(x))))
// into UADDV(UADDLP(x)). This can also happen through an extra add, where we
// transform UADDV(add(y, add(zext(extract_lo(x)), zext(extract_hi(x))))).
static SDValue performUADDVAddCombine(SDValue A, SelectionDAG &DAG) {
  auto DetectAddExtract = [&](SDValue A) {
    // Look for add(zext(extract_lo(x)), zext(extract_hi(x))), returning
    // UADDLP(x) if found.
    assert(A.getOpcode() == ISD::ADD);
    EVT VT = A.getValueType();
    SDValue Op0 = A.getOperand(0);
    SDValue Op1 = A.getOperand(1);
    if (Op0.getOpcode() != Op1.getOpcode() ||
        (Op0.getOpcode() != ISD::ZERO_EXTEND &&
         Op0.getOpcode() != ISD::SIGN_EXTEND))
      return SDValue();
    SDValue Ext0 = Op0.getOperand(0);
    SDValue Ext1 = Op1.getOperand(0);
    if (Ext0.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
        Ext1.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
        Ext0.getOperand(0) != Ext1.getOperand(0))
      return SDValue();
    // Check that the type is twice the add types, and the extract are from
    // upper/lower parts of the same source.
    if (Ext0.getOperand(0).getValueType().getVectorNumElements() !=
        VT.getVectorNumElements() * 2)
      return SDValue();
    if ((Ext0.getConstantOperandVal(1) != 0 ||
         Ext1.getConstantOperandVal(1) != VT.getVectorNumElements()) &&
        (Ext1.getConstantOperandVal(1) != 0 ||
         Ext0.getConstantOperandVal(1) != VT.getVectorNumElements()))
      return SDValue();
    unsigned Opcode = Op0.getOpcode() == ISD::ZERO_EXTEND ? AArch64ISD::UADDLP
                                                          : AArch64ISD::SADDLP;
    return DAG.getNode(Opcode, SDLoc(A), VT, Ext0.getOperand(0));
  };

  if (SDValue R = DetectAddExtract(A))
    return R;

  if (A.getOperand(0).getOpcode() == ISD::ADD && A.getOperand(0).hasOneUse())
    if (SDValue R = performUADDVAddCombine(A.getOperand(0), DAG))
      return DAG.getNode(ISD::ADD, SDLoc(A), A.getValueType(), R,
                         A.getOperand(1));
  if (A.getOperand(1).getOpcode() == ISD::ADD && A.getOperand(1).hasOneUse())
    if (SDValue R = performUADDVAddCombine(A.getOperand(1), DAG))
      return DAG.getNode(ISD::ADD, SDLoc(A), A.getValueType(), R,
                         A.getOperand(0));
  return SDValue();
}

// We can convert a UADDV(add(zext(64-bit source), zext(64-bit source))) into
// UADDLV(concat), where the concat represents the 64-bit zext sources.
static SDValue performUADDVZextCombine(SDValue A, SelectionDAG &DAG) {
  // Look for add(zext(64-bit source), zext(64-bit source)), returning
  // UADDLV(concat(zext, zext)) if found.
  assert(A.getOpcode() == ISD::ADD);
  EVT VT = A.getValueType();
  if (VT != MVT::v8i16 && VT != MVT::v4i32 && VT != MVT::v2i64)
    return SDValue();
  SDValue Op0 = A.getOperand(0);
  SDValue Op1 = A.getOperand(1);
  if (Op0.getOpcode() != ISD::ZERO_EXTEND || Op0.getOpcode() != Op1.getOpcode())
    return SDValue();
  SDValue Ext0 = Op0.getOperand(0);
  SDValue Ext1 = Op1.getOperand(0);
  EVT ExtVT0 = Ext0.getValueType();
  EVT ExtVT1 = Ext1.getValueType();
  // Check zext VTs are the same and 64-bit length.
  if (ExtVT0 != ExtVT1 ||
      VT.getScalarSizeInBits() != (2 * ExtVT0.getScalarSizeInBits()))
    return SDValue();
  // Get VT for concat of zext sources.
  EVT PairVT = ExtVT0.getDoubleNumVectorElementsVT(*DAG.getContext());
  SDValue Concat =
      DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(A), PairVT, Ext0, Ext1);

  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::v2i64:
  case MVT::v4i32:
    return DAG.getNode(AArch64ISD::UADDLV, SDLoc(A), VT, Concat);
  case MVT::v8i16: {
    SDValue Uaddlv =
        DAG.getNode(AArch64ISD::UADDLV, SDLoc(A), MVT::v4i32, Concat);
    return DAG.getNode(AArch64ISD::NVCAST, SDLoc(A), MVT::v8i16, Uaddlv);
  }
  default:
    llvm_unreachable("Unhandled vector type");
  }
}

static SDValue performUADDVCombine(SDNode *N, SelectionDAG &DAG) {
  SDValue A = N->getOperand(0);
  if (A.getOpcode() == ISD::ADD) {
    if (SDValue R = performUADDVAddCombine(A, DAG))
      return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0), R);
    else if (SDValue R = performUADDVZextCombine(A, DAG))
      return R;
  }
  return SDValue();
}

static SDValue performXorCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const AArch64Subtarget *Subtarget) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  return foldVectorXorShiftIntoCmp(N, DAG, Subtarget);
}

SDValue
AArch64TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor,
                                     SelectionDAG &DAG,
                                     SmallVectorImpl<SDNode *> &Created) const {
  AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
  if (isIntDivCheap(N->getValueType(0), Attr))
    return SDValue(N, 0); // Lower SDIV as SDIV

  EVT VT = N->getValueType(0);

  // If SVE is available, we can generate
  //  sdiv(x,y) -> ptrue + asrd          , where 'y' is positive pow-2 divisor.
  //  sdiv(x,y) -> ptrue + asrd + subr   , where 'y' is negative pow-2 divisor.
  if (VT.isVector() && Subtarget->isSVEorStreamingSVEAvailable())
    return SDValue(N, 0);

  // fold (sdiv X, pow2)
  if ((VT != MVT::i32 && VT != MVT::i64) ||
      !(Divisor.isPowerOf2() || Divisor.isNegatedPowerOf2()))
    return SDValue();

  // If the divisor is 2 or -2, the default expansion is better. It will add
  // (N->getValueType(0) >> (BitWidth - 1)) to it before shifting right.
  if (Divisor == 2 ||
      Divisor == APInt(Divisor.getBitWidth(), -2, /*isSigned*/ true))
    return SDValue();

  return TargetLowering::buildSDIVPow2WithCMov(N, Divisor, DAG, Created);
}

SDValue
AArch64TargetLowering::BuildSREMPow2(SDNode *N, const APInt &Divisor,
                                     SelectionDAG &DAG,
                                     SmallVectorImpl<SDNode *> &Created) const {
  AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes();
  if (isIntDivCheap(N->getValueType(0), Attr))
    return SDValue(N, 0); // Lower SREM as SREM

  EVT VT = N->getValueType(0);

  // For scalable and fixed types, mark them as cheap so we can handle it much
  // later. This allows us to handle larger than legal types.
  if (VT.isScalableVector() || Subtarget->useSVEForFixedLengthVectors())
    return SDValue(N, 0);

  // fold (srem X, pow2)
  if ((VT != MVT::i32 && VT != MVT::i64) ||
      !(Divisor.isPowerOf2() || Divisor.isNegatedPowerOf2()))
    return SDValue();

  unsigned Lg2 = Divisor.countr_zero();
  if (Lg2 == 0)
    return SDValue();

  SDLoc DL(N);
  SDValue N0 = N->getOperand(0);
  SDValue Pow2MinusOne = DAG.getConstant((1ULL << Lg2) - 1, DL, VT);
  SDValue Zero = DAG.getConstant(0, DL, VT);
  SDValue CCVal, CSNeg;
  if (Lg2 == 1) {
    SDValue Cmp = getAArch64Cmp(N0, Zero, ISD::SETGE, CCVal, DAG, DL);
    SDValue And = DAG.getNode(ISD::AND, DL, VT, N0, Pow2MinusOne);
    CSNeg = DAG.getNode(AArch64ISD::CSNEG, DL, VT, And, And, CCVal, Cmp);

    Created.push_back(Cmp.getNode());
    Created.push_back(And.getNode());
  } else {
    SDValue CCVal = DAG.getConstant(AArch64CC::MI, DL, MVT_CC);
    SDVTList VTs = DAG.getVTList(VT, MVT::i32);

    SDValue Negs = DAG.getNode(AArch64ISD::SUBS, DL, VTs, Zero, N0);
    SDValue AndPos = DAG.getNode(ISD::AND, DL, VT, N0, Pow2MinusOne);
    SDValue AndNeg = DAG.getNode(ISD::AND, DL, VT, Negs, Pow2MinusOne);
    CSNeg = DAG.getNode(AArch64ISD::CSNEG, DL, VT, AndPos, AndNeg, CCVal,
                        Negs.getValue(1));

    Created.push_back(Negs.getNode());
    Created.push_back(AndPos.getNode());
    Created.push_back(AndNeg.getNode());
  }

  return CSNeg;
}

static std::optional<unsigned> IsSVECntIntrinsic(SDValue S) {
  switch(getIntrinsicID(S.getNode())) {
  default:
    break;
  case Intrinsic::aarch64_sve_cntb:
    return 8;
  case Intrinsic::aarch64_sve_cnth:
    return 16;
  case Intrinsic::aarch64_sve_cntw:
    return 32;
  case Intrinsic::aarch64_sve_cntd:
    return 64;
  }
  return {};
}

/// Calculates what the pre-extend type is, based on the extension
/// operation node provided by \p Extend.
///
/// In the case that \p Extend is a SIGN_EXTEND or a ZERO_EXTEND, the
/// pre-extend type is pulled directly from the operand, while other extend
/// operations need a bit more inspection to get this information.
///
/// \param Extend The SDNode from the DAG that represents the extend operation
///
/// \returns The type representing the \p Extend source type, or \p MVT::Other
/// if no valid type can be determined
static EVT calculatePreExtendType(SDValue Extend) {
  switch (Extend.getOpcode()) {
  case ISD::SIGN_EXTEND:
  case ISD::ZERO_EXTEND:
  case ISD::ANY_EXTEND:
    return Extend.getOperand(0).getValueType();
  case ISD::AssertSext:
  case ISD::AssertZext:
  case ISD::SIGN_EXTEND_INREG: {
    VTSDNode *TypeNode = dyn_cast<VTSDNode>(Extend.getOperand(1));
    if (!TypeNode)
      return MVT::Other;
    return TypeNode->getVT();
  }
  case ISD::AND: {
    ConstantSDNode *Constant =
        dyn_cast<ConstantSDNode>(Extend.getOperand(1).getNode());
    if (!Constant)
      return MVT::Other;

    uint32_t Mask = Constant->getZExtValue();

    if (Mask == UCHAR_MAX)
      return MVT::i8;
    else if (Mask == USHRT_MAX)
      return MVT::i16;
    else if (Mask == UINT_MAX)
      return MVT::i32;

    return MVT::Other;
  }
  default:
    return MVT::Other;
  }
}

/// Combines a buildvector(sext/zext) or shuffle(sext/zext, undef) node pattern
/// into sext/zext(buildvector) or sext/zext(shuffle) making use of the vector
/// SExt/ZExt rather than the scalar SExt/ZExt
static SDValue performBuildShuffleExtendCombine(SDValue BV, SelectionDAG &DAG) {
  EVT VT = BV.getValueType();
  if (BV.getOpcode() != ISD::BUILD_VECTOR &&
      BV.getOpcode() != ISD::VECTOR_SHUFFLE)
    return SDValue();

  // Use the first item in the buildvector/shuffle to get the size of the
  // extend, and make sure it looks valid.
  SDValue Extend = BV->getOperand(0);
  unsigned ExtendOpcode = Extend.getOpcode();
  bool IsAnyExt = ExtendOpcode == ISD::ANY_EXTEND;
  bool IsSExt = ExtendOpcode == ISD::SIGN_EXTEND ||
                ExtendOpcode == ISD::SIGN_EXTEND_INREG ||
                ExtendOpcode == ISD::AssertSext;
  if (!IsAnyExt && !IsSExt && ExtendOpcode != ISD::ZERO_EXTEND &&
      ExtendOpcode != ISD::AssertZext && ExtendOpcode != ISD::AND)
    return SDValue();
  // Shuffle inputs are vector, limit to SIGN_EXTEND/ZERO_EXTEND/ANY_EXTEND to
  // ensure calculatePreExtendType will work without issue.
  if (BV.getOpcode() == ISD::VECTOR_SHUFFLE &&
      ExtendOpcode != ISD::SIGN_EXTEND && ExtendOpcode != ISD::ZERO_EXTEND)
    return SDValue();

  // Restrict valid pre-extend data type
  EVT PreExtendType = calculatePreExtendType(Extend);
  if (PreExtendType == MVT::Other ||
      PreExtendType.getScalarSizeInBits() != VT.getScalarSizeInBits() / 2)
    return SDValue();

  // Make sure all other operands are equally extended.
  bool SeenZExtOrSExt = !IsAnyExt;
  for (SDValue Op : drop_begin(BV->ops())) {
    if (Op.isUndef())
      continue;

    if (calculatePreExtendType(Op) != PreExtendType)
      return SDValue();

    unsigned Opc = Op.getOpcode();
    if (Opc == ISD::ANY_EXTEND)
      continue;

    bool OpcIsSExt = Opc == ISD::SIGN_EXTEND || Opc == ISD::SIGN_EXTEND_INREG ||
                     Opc == ISD::AssertSext;

    if (SeenZExtOrSExt && OpcIsSExt != IsSExt)
      return SDValue();

    IsSExt = OpcIsSExt;
    SeenZExtOrSExt = true;
  }

  SDValue NBV;
  SDLoc DL(BV);
  if (BV.getOpcode() == ISD::BUILD_VECTOR) {
    EVT PreExtendVT = VT.changeVectorElementType(PreExtendType);
    EVT PreExtendLegalType =
        PreExtendType.getScalarSizeInBits() < 32 ? MVT::i32 : PreExtendType;
    SmallVector<SDValue, 8> NewOps;
    for (SDValue Op : BV->ops())
      NewOps.push_back(Op.isUndef() ? DAG.getUNDEF(PreExtendLegalType)
                                    : DAG.getAnyExtOrTrunc(Op.getOperand(0), DL,
                                                           PreExtendLegalType));
    NBV = DAG.getNode(ISD::BUILD_VECTOR, DL, PreExtendVT, NewOps);
  } else { // BV.getOpcode() == ISD::VECTOR_SHUFFLE
    EVT PreExtendVT = VT.changeVectorElementType(PreExtendType.getScalarType());
    NBV = DAG.getVectorShuffle(PreExtendVT, DL, BV.getOperand(0).getOperand(0),
                               BV.getOperand(1).isUndef()
                                   ? DAG.getUNDEF(PreExtendVT)
                                   : BV.getOperand(1).getOperand(0),
                               cast<ShuffleVectorSDNode>(BV)->getMask());
  }
  unsigned ExtOpc = !SeenZExtOrSExt
                        ? ISD::ANY_EXTEND
                        : (IsSExt ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND);
  return DAG.getNode(ExtOpc, DL, VT, NBV);
}

/// Combines a mul(dup(sext/zext)) node pattern into mul(sext/zext(dup))
/// making use of the vector SExt/ZExt rather than the scalar SExt/ZExt
static SDValue performMulVectorExtendCombine(SDNode *Mul, SelectionDAG &DAG) {
  // If the value type isn't a vector, none of the operands are going to be dups
  EVT VT = Mul->getValueType(0);
  if (VT != MVT::v8i16 && VT != MVT::v4i32 && VT != MVT::v2i64)
    return SDValue();

  SDValue Op0 = performBuildShuffleExtendCombine(Mul->getOperand(0), DAG);
  SDValue Op1 = performBuildShuffleExtendCombine(Mul->getOperand(1), DAG);

  // Neither operands have been changed, don't make any further changes
  if (!Op0 && !Op1)
    return SDValue();

  SDLoc DL(Mul);
  return DAG.getNode(Mul->getOpcode(), DL, VT, Op0 ? Op0 : Mul->getOperand(0),
                     Op1 ? Op1 : Mul->getOperand(1));
}

// Combine v4i32 Mul(And(Srl(X, 15), 0x10001), 0xffff) -> v8i16 CMLTz
// Same for other types with equivalent constants.
static SDValue performMulVectorCmpZeroCombine(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if (VT != MVT::v2i64 && VT != MVT::v1i64 && VT != MVT::v2i32 &&
      VT != MVT::v4i32 && VT != MVT::v4i16 && VT != MVT::v8i16)
    return SDValue();
  if (N->getOperand(0).getOpcode() != ISD::AND ||
      N->getOperand(0).getOperand(0).getOpcode() != ISD::SRL)
    return SDValue();

  SDValue And = N->getOperand(0);
  SDValue Srl = And.getOperand(0);

  APInt V1, V2, V3;
  if (!ISD::isConstantSplatVector(N->getOperand(1).getNode(), V1) ||
      !ISD::isConstantSplatVector(And.getOperand(1).getNode(), V2) ||
      !ISD::isConstantSplatVector(Srl.getOperand(1).getNode(), V3))
    return SDValue();

  unsigned HalfSize = VT.getScalarSizeInBits() / 2;
  if (!V1.isMask(HalfSize) || V2 != (1ULL | 1ULL << HalfSize) ||
      V3 != (HalfSize - 1))
    return SDValue();

  EVT HalfVT = EVT::getVectorVT(*DAG.getContext(),
                                EVT::getIntegerVT(*DAG.getContext(), HalfSize),
                                VT.getVectorElementCount() * 2);

  SDLoc DL(N);
  SDValue In = DAG.getNode(AArch64ISD::NVCAST, DL, HalfVT, Srl.getOperand(0));
  SDValue Zero = DAG.getConstant(0, DL, In.getValueType());
  SDValue CM = DAG.getSetCC(DL, HalfVT, Zero, In, ISD::SETGT);
  return DAG.getNode(AArch64ISD::NVCAST, DL, VT, CM);
}

// Transform vector add(zext i8 to i32, zext i8 to i32)
//  into sext(add(zext(i8 to i16), zext(i8 to i16)) to i32)
// This allows extra uses of saddl/uaddl at the lower vector widths, and less
// extends.
static SDValue performVectorExtCombine(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if (!VT.isFixedLengthVector() || VT.getSizeInBits() <= 128 ||
      (N->getOperand(0).getOpcode() != ISD::ZERO_EXTEND &&
       N->getOperand(0).getOpcode() != ISD::SIGN_EXTEND) ||
      (N->getOperand(1).getOpcode() != ISD::ZERO_EXTEND &&
       N->getOperand(1).getOpcode() != ISD::SIGN_EXTEND) ||
      N->getOperand(0).getOperand(0).getValueType() !=
          N->getOperand(1).getOperand(0).getValueType())
    return SDValue();

  if (N->getOpcode() == ISD::MUL &&
      N->getOperand(0).getOpcode() != N->getOperand(1).getOpcode())
    return SDValue();

  SDValue N0 = N->getOperand(0).getOperand(0);
  SDValue N1 = N->getOperand(1).getOperand(0);
  EVT InVT = N0.getValueType();

  EVT S1 = InVT.getScalarType();
  EVT S2 = VT.getScalarType();
  if ((S2 == MVT::i32 && S1 == MVT::i8) ||
      (S2 == MVT::i64 && (S1 == MVT::i8 || S1 == MVT::i16))) {
    SDLoc DL(N);
    EVT HalfVT = EVT::getVectorVT(*DAG.getContext(),
                                  S2.getHalfSizedIntegerVT(*DAG.getContext()),
                                  VT.getVectorElementCount());
    SDValue NewN0 = DAG.getNode(N->getOperand(0).getOpcode(), DL, HalfVT, N0);
    SDValue NewN1 = DAG.getNode(N->getOperand(1).getOpcode(), DL, HalfVT, N1);
    SDValue NewOp = DAG.getNode(N->getOpcode(), DL, HalfVT, NewN0, NewN1);
    return DAG.getNode(N->getOpcode() == ISD::MUL ? N->getOperand(0).getOpcode()
                                                  : (unsigned)ISD::SIGN_EXTEND,
                       DL, VT, NewOp);
  }
  return SDValue();
}

static SDValue performMulCombine(SDNode *N, SelectionDAG &DAG,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const AArch64Subtarget *Subtarget) {

  if (SDValue Ext = performMulVectorExtendCombine(N, DAG))
    return Ext;
  if (SDValue Ext = performMulVectorCmpZeroCombine(N, DAG))
    return Ext;
  if (SDValue Ext = performVectorExtCombine(N, DAG))
    return Ext;

  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  // Canonicalize X*(Y+1) -> X*Y+X and (X+1)*Y -> X*Y+Y,
  // and in MachineCombiner pass, add+mul will be combined into madd.
  // Similarly, X*(1-Y) -> X - X*Y and (1-Y)*X -> X - Y*X.
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue N0 = N->getOperand(0);
  SDValue N1 = N->getOperand(1);
  SDValue MulOper;
  unsigned AddSubOpc;

  auto IsAddSubWith1 = [&](SDValue V) -> bool {
    AddSubOpc = V->getOpcode();
    if ((AddSubOpc == ISD::ADD || AddSubOpc == ISD::SUB) && V->hasOneUse()) {
      SDValue Opnd = V->getOperand(1);
      MulOper = V->getOperand(0);
      if (AddSubOpc == ISD::SUB)
        std::swap(Opnd, MulOper);
      if (auto C = dyn_cast<ConstantSDNode>(Opnd))
        return C->isOne();
    }
    return false;
  };

  if (IsAddSubWith1(N0)) {
    SDValue MulVal = DAG.getNode(ISD::MUL, DL, VT, N1, MulOper);
    return DAG.getNode(AddSubOpc, DL, VT, N1, MulVal);
  }

  if (IsAddSubWith1(N1)) {
    SDValue MulVal = DAG.getNode(ISD::MUL, DL, VT, N0, MulOper);
    return DAG.getNode(AddSubOpc, DL, VT, N0, MulVal);
  }

  // The below optimizations require a constant RHS.
  if (!isa<ConstantSDNode>(N1))
    return SDValue();

  ConstantSDNode *C = cast<ConstantSDNode>(N1);
  const APInt &ConstValue = C->getAPIntValue();

  // Allow the scaling to be folded into the `cnt` instruction by preventing
  // the scaling to be obscured here. This makes it easier to pattern match.
  if (IsSVECntIntrinsic(N0) ||
     (N0->getOpcode() == ISD::TRUNCATE &&
      (IsSVECntIntrinsic(N0->getOperand(0)))))
       if (ConstValue.sge(1) && ConstValue.sle(16))
         return SDValue();

  // Multiplication of a power of two plus/minus one can be done more
  // cheaply as shift+add/sub. For now, this is true unilaterally. If
  // future CPUs have a cheaper MADD instruction, this may need to be
  // gated on a subtarget feature. For Cyclone, 32-bit MADD is 4 cycles and
  // 64-bit is 5 cycles, so this is always a win.
  // More aggressively, some multiplications N0 * C can be lowered to
  // shift+add+shift if the constant C = A * B where A = 2^N + 1 and B = 2^M,
  // e.g. 6=3*2=(2+1)*2, 45=(1+4)*(1+8)
  // TODO: lower more cases.

  // TrailingZeroes is used to test if the mul can be lowered to
  // shift+add+shift.
  unsigned TrailingZeroes = ConstValue.countr_zero();
  if (TrailingZeroes) {
    // Conservatively do not lower to shift+add+shift if the mul might be
    // folded into smul or umul.
    if (N0->hasOneUse() && (isSignExtended(N0, DAG) ||
                            isZeroExtended(N0, DAG)))
      return SDValue();
    // Conservatively do not lower to shift+add+shift if the mul might be
    // folded into madd or msub.
    if (N->hasOneUse() && (N->user_begin()->getOpcode() == ISD::ADD ||
                           N->user_begin()->getOpcode() == ISD::SUB))
      return SDValue();
  }
  // Use ShiftedConstValue instead of ConstValue to support both shift+add/sub
  // and shift+add+shift.
  APInt ShiftedConstValue = ConstValue.ashr(TrailingZeroes);
  unsigned ShiftAmt;

  auto Shl = [&](SDValue N0, unsigned N1) {
    if (!N0.getNode())
      return SDValue();
    // If shift causes overflow, ignore this combine.
    if (N1 >= N0.getValueSizeInBits())
      return SDValue();
    SDValue RHS = DAG.getConstant(N1, DL, MVT::i64);
    return DAG.getNode(ISD::SHL, DL, VT, N0, RHS);
  };
  auto Add = [&](SDValue N0, SDValue N1) {
    if (!N0.getNode() || !N1.getNode())
      return SDValue();
    return DAG.getNode(ISD::ADD, DL, VT, N0, N1);
  };
  auto Sub = [&](SDValue N0, SDValue N1) {
    if (!N0.getNode() || !N1.getNode())
      return SDValue();
    return DAG.getNode(ISD::SUB, DL, VT, N0, N1);
  };
  auto Negate = [&](SDValue N) {
    if (!N0.getNode())
      return SDValue();
    SDValue Zero = DAG.getConstant(0, DL, VT);
    return DAG.getNode(ISD::SUB, DL, VT, Zero, N);
  };

  // Can the const C be decomposed into (1+2^M1)*(1+2^N1), eg:
  // C = 45 is equal to (1+4)*(1+8), we don't decompose it into (1+2)*(16-1) as
  // the (2^N - 1) can't be execused via a single instruction.
  auto isPowPlusPlusConst = [](APInt C, APInt &M, APInt &N) {
    unsigned BitWidth = C.getBitWidth();
    for (unsigned i = 1; i < BitWidth / 2; i++) {
      APInt Rem;
      APInt X(BitWidth, (1 << i) + 1);
      APInt::sdivrem(C, X, N, Rem);
      APInt NVMinus1 = N - 1;
      if (Rem == 0 && NVMinus1.isPowerOf2()) {
        M = X;
        return true;
      }
    }
    return false;
  };

  // Can the const C be decomposed into (2^M + 1) * 2^N + 1), eg:
  // C = 11 is equal to (1+4)*2+1, we don't decompose it into (1+2)*4-1 as
  // the (2^N - 1) can't be execused via a single instruction.
  auto isPowPlusPlusOneConst = [](APInt C, APInt &M, APInt &N) {
    APInt CVMinus1 = C - 1;
    if (CVMinus1.isNegative())
      return false;
    unsigned TrailingZeroes = CVMinus1.countr_zero();
    APInt SCVMinus1 = CVMinus1.ashr(TrailingZeroes) - 1;
    if (SCVMinus1.isPowerOf2()) {
      unsigned BitWidth = SCVMinus1.getBitWidth();
      M = APInt(BitWidth, SCVMinus1.logBase2());
      N = APInt(BitWidth, TrailingZeroes);
      return true;
    }
    return false;
  };

  // Can the const C be decomposed into (1 - (1 - 2^M) * 2^N), eg:
  // C = 29 is equal to 1 - (1 - 2^3) * 2^2.
  auto isPowMinusMinusOneConst = [](APInt C, APInt &M, APInt &N) {
    APInt CVMinus1 = C - 1;
    if (CVMinus1.isNegative())
      return false;
    unsigned TrailingZeroes = CVMinus1.countr_zero();
    APInt CVPlus1 = CVMinus1.ashr(TrailingZeroes) + 1;
    if (CVPlus1.isPowerOf2()) {
      unsigned BitWidth = CVPlus1.getBitWidth();
      M = APInt(BitWidth, CVPlus1.logBase2());
      N = APInt(BitWidth, TrailingZeroes);
      return true;
    }
    return false;
  };

  if (ConstValue.isNonNegative()) {
    // (mul x, (2^N + 1) * 2^M) => (shl (add (shl x, N), x), M)
    // (mul x, 2^N - 1) => (sub (shl x, N), x)
    // (mul x, (2^(N-M) - 1) * 2^M) => (sub (shl x, N), (shl x, M))
    // (mul x, (2^M + 1) * (2^N + 1))
    //     => MV = (add (shl x, M), x); (add (shl MV, N), MV)
    // (mul x, (2^M + 1) * 2^N + 1))
    //     =>  MV = add (shl x, M), x); add (shl MV, N), x)
    // (mul x, 1 - (1 - 2^M) * 2^N))
    //     =>  MV = sub (x - (shl x, M)); sub (x - (shl MV, N))
    APInt SCVMinus1 = ShiftedConstValue - 1;
    APInt SCVPlus1 = ShiftedConstValue + 1;
    APInt CVPlus1 = ConstValue + 1;
    APInt CVM, CVN;
    if (SCVMinus1.isPowerOf2()) {
      ShiftAmt = SCVMinus1.logBase2();
      return Shl(Add(Shl(N0, ShiftAmt), N0), TrailingZeroes);
    } else if (CVPlus1.isPowerOf2()) {
      ShiftAmt = CVPlus1.logBase2();
      return Sub(Shl(N0, ShiftAmt), N0);
    } else if (SCVPlus1.isPowerOf2()) {
      ShiftAmt = SCVPlus1.logBase2() + TrailingZeroes;
      return Sub(Shl(N0, ShiftAmt), Shl(N0, TrailingZeroes));
    }
    if (Subtarget->hasALULSLFast() &&
        isPowPlusPlusConst(ConstValue, CVM, CVN)) {
      APInt CVMMinus1 = CVM - 1;
      APInt CVNMinus1 = CVN - 1;
      unsigned ShiftM1 = CVMMinus1.logBase2();
      unsigned ShiftN1 = CVNMinus1.logBase2();
      // ALULSLFast implicate that Shifts <= 4 places are fast
      if (ShiftM1 <= 4 && ShiftN1 <= 4) {
        SDValue MVal = Add(Shl(N0, ShiftM1), N0);
        return Add(Shl(MVal, ShiftN1), MVal);
      }
    }
    if (Subtarget->hasALULSLFast() &&
        isPowPlusPlusOneConst(ConstValue, CVM, CVN)) {
      unsigned ShiftM = CVM.getZExtValue();
      unsigned ShiftN = CVN.getZExtValue();
      // ALULSLFast implicate that Shifts <= 4 places are fast
      if (ShiftM <= 4 && ShiftN <= 4) {
        SDValue MVal = Add(Shl(N0, CVM.getZExtValue()), N0);
        return Add(Shl(MVal, CVN.getZExtValue()), N0);
      }
    }

    if (Subtarget->hasALULSLFast() &&
        isPowMinusMinusOneConst(ConstValue, CVM, CVN)) {
      unsigned ShiftM = CVM.getZExtValue();
      unsigned ShiftN = CVN.getZExtValue();
      // ALULSLFast implicate that Shifts <= 4 places are fast
      if (ShiftM <= 4 && ShiftN <= 4) {
        SDValue MVal = Sub(N0, Shl(N0, CVM.getZExtValue()));
        return Sub(N0, Shl(MVal, CVN.getZExtValue()));
      }
    }
  } else {
    // (mul x, -(2^N - 1)) => (sub x, (shl x, N))
    // (mul x, -(2^N + 1)) => - (add (shl x, N), x)
    // (mul x, -(2^(N-M) - 1) * 2^M) => (sub (shl x, M), (shl x, N))
    APInt SCVPlus1 = -ShiftedConstValue + 1;
    APInt CVNegPlus1 = -ConstValue + 1;
    APInt CVNegMinus1 = -ConstValue - 1;
    if (CVNegPlus1.isPowerOf2()) {
      ShiftAmt = CVNegPlus1.logBase2();
      return Sub(N0, Shl(N0, ShiftAmt));
    } else if (CVNegMinus1.isPowerOf2()) {
      ShiftAmt = CVNegMinus1.logBase2();
      return Negate(Add(Shl(N0, ShiftAmt), N0));
    } else if (SCVPlus1.isPowerOf2()) {
      ShiftAmt = SCVPlus1.logBase2() + TrailingZeroes;
      return Sub(Shl(N0, TrailingZeroes), Shl(N0, ShiftAmt));
    }
  }

  return SDValue();
}

static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
                                                         SelectionDAG &DAG) {
  // Take advantage of vector comparisons producing 0 or -1 in each lane to
  // optimize away operation when it's from a constant.
  //
  // The general transformation is:
  //    UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
  //       AND(VECTOR_CMP(x,y), constant2)
  //    constant2 = UNARYOP(constant)

  // Early exit if this isn't a vector operation, the operand of the
  // unary operation isn't a bitwise AND, or if the sizes of the operations
  // aren't the same.
  EVT VT = N->getValueType(0);
  if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
      N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
      VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
    return SDValue();

  // Now check that the other operand of the AND is a constant. We could
  // make the transformation for non-constant splats as well, but it's unclear
  // that would be a benefit as it would not eliminate any operations, just
  // perform one more step in scalar code before moving to the vector unit.
  if (BuildVectorSDNode *BV =
          dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
    // Bail out if the vector isn't a constant.
    if (!BV->isConstant())
      return SDValue();

    // Everything checks out. Build up the new and improved node.
    SDLoc DL(N);
    EVT IntVT = BV->getValueType(0);
    // Create a new constant of the appropriate type for the transformed
    // DAG.
    SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
    // The AND node needs bitcasts to/from an integer vector type around it.
    SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
    SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
                                 N->getOperand(0)->getOperand(0), MaskConst);
    SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
    return Res;
  }

  return SDValue();
}

/// Tries to replace scalar FP <-> INT conversions with SVE in streaming
/// functions, this can help to reduce the number of fmovs to/from GPRs.
static SDValue
tryToReplaceScalarFPConversionWithSVE(SDNode *N, SelectionDAG &DAG,
                                      TargetLowering::DAGCombinerInfo &DCI,
                                      const AArch64Subtarget *Subtarget) {
  if (N->isStrictFPOpcode())
    return SDValue();

  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  if (!Subtarget->isSVEorStreamingSVEAvailable() ||
      (!Subtarget->isStreaming() && !Subtarget->isStreamingCompatible()))
    return SDValue();

  auto isSupportedType = [](EVT VT) {
    return !VT.isVector() && VT != MVT::bf16 && VT != MVT::f128;
  };

  SDValue SrcVal = N->getOperand(0);
  EVT SrcTy = SrcVal.getValueType();
  EVT DestTy = N->getValueType(0);

  if (!isSupportedType(SrcTy) || !isSupportedType(DestTy))
    return SDValue();

  EVT SrcVecTy;
  EVT DestVecTy;
  if (DestTy.bitsGT(SrcTy)) {
    DestVecTy = getPackedSVEVectorVT(DestTy);
    SrcVecTy = DestVecTy.changeVectorElementType(SrcTy);
  } else {
    SrcVecTy = getPackedSVEVectorVT(SrcTy);
    DestVecTy = SrcVecTy.changeVectorElementType(DestTy);
  }

  // Ensure the resulting src/dest vector type is legal.
  if (SrcVecTy == MVT::nxv2i32 || DestVecTy == MVT::nxv2i32)
    return SDValue();

  SDLoc DL(N);
  SDValue ZeroIdx = DAG.getVectorIdxConstant(0, DL);
  SDValue Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, SrcVecTy,
                            DAG.getUNDEF(SrcVecTy), SrcVal, ZeroIdx);
  SDValue Convert = DAG.getNode(N->getOpcode(), DL, DestVecTy, Vec);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, DestTy, Convert, ZeroIdx);
}

static SDValue performIntToFpCombine(SDNode *N, SelectionDAG &DAG,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     const AArch64Subtarget *Subtarget) {
  // First try to optimize away the conversion when it's conditionally from
  // a constant. Vectors only.
  if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
    return Res;

  if (SDValue Res =
          tryToReplaceScalarFPConversionWithSVE(N, DAG, DCI, Subtarget))
    return Res;

  EVT VT = N->getValueType(0);
  if (VT != MVT::f32 && VT != MVT::f64)
    return SDValue();

  // Only optimize when the source and destination types have the same width.
  if (VT.getSizeInBits() != N->getOperand(0).getValueSizeInBits())
    return SDValue();

  // If the result of an integer load is only used by an integer-to-float
  // conversion, use a fp load instead and a AdvSIMD scalar {S|U}CVTF instead.
  // This eliminates an "integer-to-vector-move" UOP and improves throughput.
  SDValue N0 = N->getOperand(0);
  if (Subtarget->isNeonAvailable() && ISD::isNormalLoad(N0.getNode()) &&
      N0.hasOneUse() &&
      // Do not change the width of a volatile load.
      !cast<LoadSDNode>(N0)->isVolatile()) {
    LoadSDNode *LN0 = cast<LoadSDNode>(N0);
    SDValue Load = DAG.getLoad(VT, SDLoc(N), LN0->getChain(), LN0->getBasePtr(),
                               LN0->getPointerInfo(), LN0->getAlign(),
                               LN0->getMemOperand()->getFlags());

    // Make sure successors of the original load stay after it by updating them
    // to use the new Chain.
    DAG.ReplaceAllUsesOfValueWith(SDValue(LN0, 1), Load.getValue(1));

    unsigned Opcode =
        (N->getOpcode() == ISD::SINT_TO_FP) ? AArch64ISD::SITOF : AArch64ISD::UITOF;
    return DAG.getNode(Opcode, SDLoc(N), VT, Load);
  }

  return SDValue();
}

/// Fold a floating-point multiply by power of two into floating-point to
/// fixed-point conversion.
static SDValue performFpToIntCombine(SDNode *N, SelectionDAG &DAG,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     const AArch64Subtarget *Subtarget) {
  if (SDValue Res =
          tryToReplaceScalarFPConversionWithSVE(N, DAG, DCI, Subtarget))
    return Res;

  if (!Subtarget->isNeonAvailable())
    return SDValue();

  if (!N->getValueType(0).isSimple())
    return SDValue();

  SDValue Op = N->getOperand(0);
  if (!Op.getValueType().isSimple() || Op.getOpcode() != ISD::FMUL)
    return SDValue();

  if (!Op.getValueType().is64BitVector() && !Op.getValueType().is128BitVector())
    return SDValue();

  SDValue ConstVec = Op->getOperand(1);
  if (!isa<BuildVectorSDNode>(ConstVec))
    return SDValue();

  MVT FloatTy = Op.getSimpleValueType().getVectorElementType();
  uint32_t FloatBits = FloatTy.getSizeInBits();
  if (FloatBits != 32 && FloatBits != 64 &&
      (FloatBits != 16 || !Subtarget->hasFullFP16()))
    return SDValue();

  MVT IntTy = N->getSimpleValueType(0).getVectorElementType();
  uint32_t IntBits = IntTy.getSizeInBits();
  if (IntBits != 16 && IntBits != 32 && IntBits != 64)
    return SDValue();

  // Avoid conversions where iN is larger than the float (e.g., float -> i64).
  if (IntBits > FloatBits)
    return SDValue();

  BitVector UndefElements;
  BuildVectorSDNode *BV = cast<BuildVectorSDNode>(ConstVec);
  int32_t Bits = IntBits == 64 ? 64 : 32;
  int32_t C = BV->getConstantFPSplatPow2ToLog2Int(&UndefElements, Bits + 1);
  if (C == -1 || C == 0 || C > Bits)
    return SDValue();

  EVT ResTy = Op.getValueType().changeVectorElementTypeToInteger();
  if (!DAG.getTargetLoweringInfo().isTypeLegal(ResTy))
    return SDValue();

  if (N->getOpcode() == ISD::FP_TO_SINT_SAT ||
      N->getOpcode() == ISD::FP_TO_UINT_SAT) {
    EVT SatVT = cast<VTSDNode>(N->getOperand(1))->getVT();
    if (SatVT.getScalarSizeInBits() != IntBits || IntBits != FloatBits)
      return SDValue();
  }

  SDLoc DL(N);
  bool IsSigned = (N->getOpcode() == ISD::FP_TO_SINT ||
                   N->getOpcode() == ISD::FP_TO_SINT_SAT);
  unsigned IntrinsicOpcode = IsSigned ? Intrinsic::aarch64_neon_vcvtfp2fxs
                                      : Intrinsic::aarch64_neon_vcvtfp2fxu;
  SDValue FixConv =
      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ResTy,
                  DAG.getConstant(IntrinsicOpcode, DL, MVT::i32),
                  Op->getOperand(0), DAG.getConstant(C, DL, MVT::i32));
  // We can handle smaller integers by generating an extra trunc.
  if (IntBits < FloatBits)
    FixConv = DAG.getNode(ISD::TRUNCATE, DL, N->getValueType(0), FixConv);

  return FixConv;
}

static SDValue tryCombineToBSL(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                               const AArch64TargetLowering &TLI) {
  EVT VT = N->getValueType(0);
  SelectionDAG &DAG = DCI.DAG;
  SDLoc DL(N);
  const auto &Subtarget = DAG.getSubtarget<AArch64Subtarget>();

  if (!VT.isVector())
    return SDValue();

  if (VT.isScalableVector() && !Subtarget.hasSVE2())
    return SDValue();

  if (VT.isFixedLengthVector() &&
      (!Subtarget.isNeonAvailable() || TLI.useSVEForFixedLengthVectorVT(VT)))
    return SDValue();

  SDValue N0 = N->getOperand(0);
  if (N0.getOpcode() != ISD::AND)
    return SDValue();

  SDValue N1 = N->getOperand(1);
  if (N1.getOpcode() != ISD::AND)
    return SDValue();

  // InstCombine does (not (neg a)) => (add a -1).
  // Try: (or (and (neg a) b) (and (add a -1) c)) => (bsl (neg a) b c)
  // Loop over all combinations of AND operands.
  for (int i = 1; i >= 0; --i) {
    for (int j = 1; j >= 0; --j) {
      SDValue O0 = N0->getOperand(i);
      SDValue O1 = N1->getOperand(j);
      SDValue Sub, Add, SubSibling, AddSibling;

      // Find a SUB and an ADD operand, one from each AND.
      if (O0.getOpcode() == ISD::SUB && O1.getOpcode() == ISD::ADD) {
        Sub = O0;
        Add = O1;
        SubSibling = N0->getOperand(1 - i);
        AddSibling = N1->getOperand(1 - j);
      } else if (O0.getOpcode() == ISD::ADD && O1.getOpcode() == ISD::SUB) {
        Add = O0;
        Sub = O1;
        AddSibling = N0->getOperand(1 - i);
        SubSibling = N1->getOperand(1 - j);
      } else
        continue;

      if (!ISD::isConstantSplatVectorAllZeros(Sub.getOperand(0).getNode()))
        continue;

      // Constant ones is always righthand operand of the Add.
      if (!ISD::isConstantSplatVectorAllOnes(Add.getOperand(1).getNode()))
        continue;

      if (Sub.getOperand(1) != Add.getOperand(0))
        continue;

      return DAG.getNode(AArch64ISD::BSP, DL, VT, Sub, SubSibling, AddSibling);
    }
  }

  // (or (and a b) (and (not a) c)) => (bsl a b c)
  // We only have to look for constant vectors here since the general, variable
  // case can be handled in TableGen.
  unsigned Bits = VT.getScalarSizeInBits();
  uint64_t BitMask = Bits == 64 ? -1ULL : ((1ULL << Bits) - 1);
  for (int i = 1; i >= 0; --i)
    for (int j = 1; j >= 0; --j) {
      APInt Val1, Val2;

      if (ISD::isConstantSplatVector(N0->getOperand(i).getNode(), Val1) &&
          ISD::isConstantSplatVector(N1->getOperand(j).getNode(), Val2) &&
          (BitMask & ~Val1.getZExtValue()) == Val2.getZExtValue()) {
        return DAG.getNode(AArch64ISD::BSP, DL, VT, N0->getOperand(i),
                           N0->getOperand(1 - i), N1->getOperand(1 - j));
      }
      BuildVectorSDNode *BVN0 = dyn_cast<BuildVectorSDNode>(N0->getOperand(i));
      BuildVectorSDNode *BVN1 = dyn_cast<BuildVectorSDNode>(N1->getOperand(j));
      if (!BVN0 || !BVN1)
        continue;

      bool FoundMatch = true;
      for (unsigned k = 0; k < VT.getVectorNumElements(); ++k) {
        ConstantSDNode *CN0 = dyn_cast<ConstantSDNode>(BVN0->getOperand(k));
        ConstantSDNode *CN1 = dyn_cast<ConstantSDNode>(BVN1->getOperand(k));
        if (!CN0 || !CN1 ||
            CN0->getZExtValue() != (BitMask & ~CN1->getZExtValue())) {
          FoundMatch = false;
          break;
        }
      }
      if (FoundMatch)
        return DAG.getNode(AArch64ISD::BSP, DL, VT, N0->getOperand(i),
                           N0->getOperand(1 - i), N1->getOperand(1 - j));
    }

  return SDValue();
}

// Given a tree of and/or(csel(0, 1, cc0), csel(0, 1, cc1)), we may be able to
// convert to csel(ccmp(.., cc0)), depending on cc1:

// (AND (CSET cc0 cmp0) (CSET cc1 (CMP x1 y1)))
// =>
// (CSET cc1 (CCMP x1 y1 !cc1 cc0 cmp0))
//
// (OR (CSET cc0 cmp0) (CSET cc1 (CMP x1 y1)))
// =>
// (CSET cc1 (CCMP x1 y1 cc1 !cc0 cmp0))
static SDValue performANDORCSELCombine(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  SDValue CSel0 = N->getOperand(0);
  SDValue CSel1 = N->getOperand(1);

  if (CSel0.getOpcode() != AArch64ISD::CSEL ||
      CSel1.getOpcode() != AArch64ISD::CSEL)
    return SDValue();

  if (!CSel0->hasOneUse() || !CSel1->hasOneUse())
    return SDValue();

  if (!isNullConstant(CSel0.getOperand(0)) ||
      !isOneConstant(CSel0.getOperand(1)) ||
      !isNullConstant(CSel1.getOperand(0)) ||
      !isOneConstant(CSel1.getOperand(1)))
    return SDValue();

  SDValue Cmp0 = CSel0.getOperand(3);
  SDValue Cmp1 = CSel1.getOperand(3);
  AArch64CC::CondCode CC0 = (AArch64CC::CondCode)CSel0.getConstantOperandVal(2);
  AArch64CC::CondCode CC1 = (AArch64CC::CondCode)CSel1.getConstantOperandVal(2);
  if (!Cmp0->hasOneUse() || !Cmp1->hasOneUse())
    return SDValue();
  if (Cmp1.getOpcode() != AArch64ISD::SUBS &&
      Cmp0.getOpcode() == AArch64ISD::SUBS) {
    std::swap(Cmp0, Cmp1);
    std::swap(CC0, CC1);
  }

  if (Cmp1.getOpcode() != AArch64ISD::SUBS)
    return SDValue();

  SDLoc DL(N);
  SDValue CCmp, Condition;
  unsigned NZCV;

  if (N->getOpcode() == ISD::AND) {
    AArch64CC::CondCode InvCC0 = AArch64CC::getInvertedCondCode(CC0);
    Condition = DAG.getConstant(InvCC0, DL, MVT_CC);
    NZCV = AArch64CC::getNZCVToSatisfyCondCode(CC1);
  } else {
    AArch64CC::CondCode InvCC1 = AArch64CC::getInvertedCondCode(CC1);
    Condition = DAG.getConstant(CC0, DL, MVT_CC);
    NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvCC1);
  }

  SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);

  auto *Op1 = dyn_cast<ConstantSDNode>(Cmp1.getOperand(1));
  if (Op1 && Op1->getAPIntValue().isNegative() &&
      Op1->getAPIntValue().sgt(-32)) {
    // CCMP accept the constant int the range [0, 31]
    // if the Op1 is a constant in the range [-31, -1], we
    // can select to CCMN to avoid the extra mov
    SDValue AbsOp1 =
        DAG.getConstant(Op1->getAPIntValue().abs(), DL, Op1->getValueType(0));
    CCmp = DAG.getNode(AArch64ISD::CCMN, DL, MVT_CC, Cmp1.getOperand(0), AbsOp1,
                       NZCVOp, Condition, Cmp0);
  } else {
    CCmp = DAG.getNode(AArch64ISD::CCMP, DL, MVT_CC, Cmp1.getOperand(0),
                       Cmp1.getOperand(1), NZCVOp, Condition, Cmp0);
  }
  return DAG.getNode(AArch64ISD::CSEL, DL, VT, CSel0.getOperand(0),
                     CSel0.getOperand(1), DAG.getConstant(CC1, DL, MVT::i32),
                     CCmp);
}

static SDValue performORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                                const AArch64Subtarget *Subtarget,
                                const AArch64TargetLowering &TLI) {
  SelectionDAG &DAG = DCI.DAG;
  EVT VT = N->getValueType(0);

  if (SDValue R = performANDORCSELCombine(N, DAG))
    return R;

  if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
    return SDValue();

  if (SDValue Res = tryCombineToBSL(N, DCI, TLI))
    return Res;

  return SDValue();
}

static bool isConstantSplatVectorMaskForType(SDNode *N, EVT MemVT) {
  if (!MemVT.getVectorElementType().isSimple())
    return false;

  uint64_t MaskForTy = 0ull;
  switch (MemVT.getVectorElementType().getSimpleVT().SimpleTy) {
  case MVT::i8:
    MaskForTy = 0xffull;
    break;
  case MVT::i16:
    MaskForTy = 0xffffull;
    break;
  case MVT::i32:
    MaskForTy = 0xffffffffull;
    break;
  default:
    return false;
    break;
  }

  if (N->getOpcode() == AArch64ISD::DUP || N->getOpcode() == ISD::SPLAT_VECTOR)
    if (auto *Op0 = dyn_cast<ConstantSDNode>(N->getOperand(0)))
      return Op0->getAPIntValue().getLimitedValue() == MaskForTy;

  return false;
}

static SDValue performReinterpretCastCombine(SDNode *N) {
  SDValue LeafOp = SDValue(N, 0);
  SDValue Op = N->getOperand(0);
  while (Op.getOpcode() == AArch64ISD::REINTERPRET_CAST &&
         LeafOp.getValueType() != Op.getValueType())
    Op = Op->getOperand(0);
  if (LeafOp.getValueType() == Op.getValueType())
    return Op;
  return SDValue();
}

static SDValue performSVEAndCombine(SDNode *N,
                                    TargetLowering::DAGCombinerInfo &DCI) {
  SelectionDAG &DAG = DCI.DAG;
  SDValue Src = N->getOperand(0);
  unsigned Opc = Src->getOpcode();

  // Zero/any extend of an unsigned unpack
  if (Opc == AArch64ISD::UUNPKHI || Opc == AArch64ISD::UUNPKLO) {
    SDValue UnpkOp = Src->getOperand(0);
    SDValue Dup = N->getOperand(1);

    if (Dup.getOpcode() != ISD::SPLAT_VECTOR)
      return SDValue();

    SDLoc DL(N);
    ConstantSDNode *C = dyn_cast<ConstantSDNode>(Dup->getOperand(0));
    if (!C)
      return SDValue();

    uint64_t ExtVal = C->getZExtValue();

    auto MaskAndTypeMatch = [ExtVal](EVT VT) -> bool {
      return ((ExtVal == 0xFF && VT == MVT::i8) ||
              (ExtVal == 0xFFFF && VT == MVT::i16) ||
              (ExtVal == 0xFFFFFFFF && VT == MVT::i32));
    };

    // If the mask is fully covered by the unpack, we don't need to push
    // a new AND onto the operand
    EVT EltTy = UnpkOp->getValueType(0).getVectorElementType();
    if (MaskAndTypeMatch(EltTy))
      return Src;

    // If this is 'and (uunpklo/hi (extload MemTy -> ExtTy)), mask', then check
    // to see if the mask is all-ones of size MemTy.
    auto MaskedLoadOp = dyn_cast<MaskedLoadSDNode>(UnpkOp);
    if (MaskedLoadOp && (MaskedLoadOp->getExtensionType() == ISD::ZEXTLOAD ||
                         MaskedLoadOp->getExtensionType() == ISD::EXTLOAD)) {
      EVT EltTy = MaskedLoadOp->getMemoryVT().getVectorElementType();
      if (MaskAndTypeMatch(EltTy))
        return Src;
    }

    // Truncate to prevent a DUP with an over wide constant
    APInt Mask = C->getAPIntValue().trunc(EltTy.getSizeInBits());

    // Otherwise, make sure we propagate the AND to the operand
    // of the unpack
    Dup = DAG.getNode(ISD::SPLAT_VECTOR, DL, UnpkOp->getValueType(0),
                      DAG.getConstant(Mask.zextOrTrunc(32), DL, MVT::i32));

    SDValue And = DAG.getNode(ISD::AND, DL,
                              UnpkOp->getValueType(0), UnpkOp, Dup);

    return DAG.getNode(Opc, DL, N->getValueType(0), And);
  }

  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  // If both sides of AND operations are i1 splat_vectors then
  // we can produce just i1 splat_vector as the result.
  if (isAllActivePredicate(DAG, N->getOperand(0)))
    return N->getOperand(1);
  if (isAllActivePredicate(DAG, N->getOperand(1)))
    return N->getOperand(0);

  if (!EnableCombineMGatherIntrinsics)
    return SDValue();

  SDValue Mask = N->getOperand(1);

  if (!Src.hasOneUse())
    return SDValue();

  EVT MemVT;

  // SVE load instructions perform an implicit zero-extend, which makes them
  // perfect candidates for combining.
  switch (Opc) {
  case AArch64ISD::LD1_MERGE_ZERO:
  case AArch64ISD::LDNF1_MERGE_ZERO:
  case AArch64ISD::LDFF1_MERGE_ZERO:
    MemVT = cast<VTSDNode>(Src->getOperand(3))->getVT();
    break;
  case AArch64ISD::GLD1_MERGE_ZERO:
  case AArch64ISD::GLD1_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1_SXTW_MERGE_ZERO:
  case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1_UXTW_MERGE_ZERO:
  case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1_IMM_MERGE_ZERO:
  case AArch64ISD::GLDFF1_MERGE_ZERO:
  case AArch64ISD::GLDFF1_SCALED_MERGE_ZERO:
  case AArch64ISD::GLDFF1_SXTW_MERGE_ZERO:
  case AArch64ISD::GLDFF1_SXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLDFF1_UXTW_MERGE_ZERO:
  case AArch64ISD::GLDFF1_UXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLDFF1_IMM_MERGE_ZERO:
  case AArch64ISD::GLDNT1_MERGE_ZERO:
    MemVT = cast<VTSDNode>(Src->getOperand(4))->getVT();
    break;
  default:
    return SDValue();
  }

  if (isConstantSplatVectorMaskForType(Mask.getNode(), MemVT))
    return Src;

  return SDValue();
}

// Transform and(fcmp(a, b), fcmp(c, d)) into fccmp(fcmp(a, b), c, d)
static SDValue performANDSETCCCombine(SDNode *N,
                                      TargetLowering::DAGCombinerInfo &DCI) {

  // This function performs an optimization on a specific pattern involving
  // an AND operation and SETCC (Set Condition Code) node.

  SDValue SetCC = N->getOperand(0);
  EVT VT = N->getValueType(0);
  SelectionDAG &DAG = DCI.DAG;

  // Checks if the current node (N) is used by any SELECT instruction and
  // returns an empty SDValue to avoid applying the optimization to prevent
  // incorrect results
  for (auto U : N->users())
    if (U->getOpcode() == ISD::SELECT)
      return SDValue();

  // Check if the operand is a SETCC node with floating-point comparison
  if (SetCC.getOpcode() == ISD::SETCC &&
      SetCC.getOperand(0).getValueType() == MVT::f32) {

    SDValue Cmp;
    AArch64CC::CondCode CC;

    // Check if the DAG is after legalization and if we can emit the conjunction
    if (!DCI.isBeforeLegalize() &&
        (Cmp = emitConjunction(DAG, SDValue(N, 0), CC))) {

      AArch64CC::CondCode InvertedCC = AArch64CC::getInvertedCondCode(CC);

      SDLoc DL(N);
      return DAG.getNode(AArch64ISD::CSINC, DL, VT, DAG.getConstant(0, DL, VT),
                         DAG.getConstant(0, DL, VT),
                         DAG.getConstant(InvertedCC, DL, MVT::i32), Cmp);
    }
  }
  return SDValue();
}

static SDValue performANDCombine(SDNode *N,
                                 TargetLowering::DAGCombinerInfo &DCI) {
  SelectionDAG &DAG = DCI.DAG;
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  EVT VT = N->getValueType(0);

  if (SDValue R = performANDORCSELCombine(N, DAG))
    return R;

  if (SDValue R = performANDSETCCCombine(N,DCI))
    return R;

  if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
    return SDValue();

  if (VT.isScalableVector())
    return performSVEAndCombine(N, DCI);

  // The combining code below works only for NEON vectors. In particular, it
  // does not work for SVE when dealing with vectors wider than 128 bits.
  if (!VT.is64BitVector() && !VT.is128BitVector())
    return SDValue();

  BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(RHS.getNode());
  if (!BVN)
    return SDValue();

  // AND does not accept an immediate, so check if we can use a BIC immediate
  // instruction instead. We do this here instead of using a (and x, (mvni imm))
  // pattern in isel, because some immediates may be lowered to the preferred
  // (and x, (movi imm)) form, even though an mvni representation also exists.
  APInt DefBits(VT.getSizeInBits(), 0);
  APInt UndefBits(VT.getSizeInBits(), 0);
  if (resolveBuildVector(BVN, DefBits, UndefBits)) {
    SDValue NewOp;

    // Any bits known to already be 0 need not be cleared again, which can help
    // reduce the size of the immediate to one supported by the instruction.
    KnownBits Known = DAG.computeKnownBits(LHS);
    APInt ZeroSplat(VT.getSizeInBits(), 0);
    for (unsigned I = 0; I < VT.getSizeInBits() / Known.Zero.getBitWidth(); I++)
      ZeroSplat |= Known.Zero.zext(VT.getSizeInBits())
                   << (Known.Zero.getBitWidth() * I);

    DefBits = ~(DefBits | ZeroSplat);
    if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::BICi, SDValue(N, 0), DAG,
                                    DefBits, &LHS)) ||
        (NewOp = tryAdvSIMDModImm16(AArch64ISD::BICi, SDValue(N, 0), DAG,
                                    DefBits, &LHS)))
      return NewOp;

    UndefBits = ~(UndefBits | ZeroSplat);
    if ((NewOp = tryAdvSIMDModImm32(AArch64ISD::BICi, SDValue(N, 0), DAG,
                                    UndefBits, &LHS)) ||
        (NewOp = tryAdvSIMDModImm16(AArch64ISD::BICi, SDValue(N, 0), DAG,
                                    UndefBits, &LHS)))
      return NewOp;
  }

  return SDValue();
}

static SDValue performFADDCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI) {
  SelectionDAG &DAG = DCI.DAG;
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  EVT VT = N->getValueType(0);
  SDLoc DL(N);

  if (!N->getFlags().hasAllowReassociation())
    return SDValue();

  // Combine fadd(a, vcmla(b, c, d)) -> vcmla(fadd(a, b), b, c)
  auto ReassocComplex = [&](SDValue A, SDValue B) {
    if (A.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
      return SDValue();
    unsigned Opc = A.getConstantOperandVal(0);
    if (Opc != Intrinsic::aarch64_neon_vcmla_rot0 &&
        Opc != Intrinsic::aarch64_neon_vcmla_rot90 &&
        Opc != Intrinsic::aarch64_neon_vcmla_rot180 &&
        Opc != Intrinsic::aarch64_neon_vcmla_rot270)
      return SDValue();
    SDValue VCMLA = DAG.getNode(
        ISD::INTRINSIC_WO_CHAIN, DL, VT, A.getOperand(0),
        DAG.getNode(ISD::FADD, DL, VT, A.getOperand(1), B, N->getFlags()),
        A.getOperand(2), A.getOperand(3));
    VCMLA->setFlags(A->getFlags());
    return VCMLA;
  };
  if (SDValue R = ReassocComplex(LHS, RHS))
    return R;
  if (SDValue R = ReassocComplex(RHS, LHS))
    return R;

  return SDValue();
}

static bool hasPairwiseAdd(unsigned Opcode, EVT VT, bool FullFP16) {
  switch (Opcode) {
  case ISD::STRICT_FADD:
  case ISD::FADD:
    return (FullFP16 && VT == MVT::f16) || VT == MVT::f32 || VT == MVT::f64;
  case ISD::ADD:
    return VT == MVT::i64;
  default:
    return false;
  }
}

static SDValue getPTest(SelectionDAG &DAG, EVT VT, SDValue Pg, SDValue Op,
                        AArch64CC::CondCode Cond);

static bool isPredicateCCSettingOp(SDValue N) {
  if ((N.getOpcode() == ISD::SETCC) ||
      // get_active_lane_mask is lowered to a whilelo instruction.
      (N.getOpcode() == ISD::GET_ACTIVE_LANE_MASK) ||
      (N.getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
       (N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilege ||
        N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilegt ||
        N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilehi ||
        N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilehs ||
        N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilele ||
        N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilelo ||
        N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilels ||
        N.getConstantOperandVal(0) == Intrinsic::aarch64_sve_whilelt)))
    return true;

  return false;
}

// Materialize : i1 = extract_vector_elt t37, Constant:i64<0>
// ... into: "ptrue p, all" + PTEST
static SDValue
performFirstTrueTestVectorCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  const AArch64Subtarget *Subtarget) {
  assert(N->getOpcode() == ISD::EXTRACT_VECTOR_ELT);
  // Make sure PTEST can be legalised with illegal types.
  if (!Subtarget->hasSVE() || DCI.isBeforeLegalize())
    return SDValue();

  SDValue N0 = N->getOperand(0);
  EVT VT = N0.getValueType();

  if (!VT.isScalableVector() || VT.getVectorElementType() != MVT::i1 ||
      !isNullConstant(N->getOperand(1)))
    return SDValue();

  // Restricted the DAG combine to only cases where we're extracting from a
  // flag-setting operation.
  if (!isPredicateCCSettingOp(N0))
    return SDValue();

  // Extracts of lane 0 for SVE can be expressed as PTEST(Op, FIRST) ? 1 : 0
  SelectionDAG &DAG = DCI.DAG;
  SDValue Pg = getPTrue(DAG, SDLoc(N), VT, AArch64SVEPredPattern::all);
  return getPTest(DAG, N->getValueType(0), Pg, N0, AArch64CC::FIRST_ACTIVE);
}

// Materialize : Idx = (add (mul vscale, NumEls), -1)
//               i1 = extract_vector_elt t37, Constant:i64<Idx>
//     ... into: "ptrue p, all" + PTEST
static SDValue
performLastTrueTestVectorCombine(SDNode *N,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 const AArch64Subtarget *Subtarget) {
  assert(N->getOpcode() == ISD::EXTRACT_VECTOR_ELT);
  // Make sure PTEST is legal types.
  if (!Subtarget->hasSVE() || DCI.isBeforeLegalize())
    return SDValue();

  SDValue N0 = N->getOperand(0);
  EVT OpVT = N0.getValueType();

  if (!OpVT.isScalableVector() || OpVT.getVectorElementType() != MVT::i1)
    return SDValue();

  // Idx == (add (mul vscale, NumEls), -1)
  SDValue Idx = N->getOperand(1);
  if (Idx.getOpcode() != ISD::ADD || !isAllOnesConstant(Idx.getOperand(1)))
    return SDValue();

  SDValue VS = Idx.getOperand(0);
  if (VS.getOpcode() != ISD::VSCALE)
    return SDValue();

  unsigned NumEls = OpVT.getVectorElementCount().getKnownMinValue();
  if (VS.getConstantOperandVal(0) != NumEls)
    return SDValue();

  // Extracts of lane EC-1 for SVE can be expressed as PTEST(Op, LAST) ? 1 : 0
  SelectionDAG &DAG = DCI.DAG;
  SDValue Pg = getPTrue(DAG, SDLoc(N), OpVT, AArch64SVEPredPattern::all);
  return getPTest(DAG, N->getValueType(0), Pg, N0, AArch64CC::LAST_ACTIVE);
}

static SDValue
performExtractLastActiveCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                                const AArch64Subtarget *Subtarget) {
  assert(N->getOpcode() == ISD::EXTRACT_VECTOR_ELT);
  SelectionDAG &DAG = DCI.DAG;
  SDValue Vec = N->getOperand(0);
  SDValue Idx = N->getOperand(1);

  if (DCI.isBeforeLegalize() || Idx.getOpcode() != ISD::VECTOR_FIND_LAST_ACTIVE)
    return SDValue();

  // Only legal for 8, 16, 32, and 64 bit element types.
  EVT EltVT = Vec.getValueType().getVectorElementType();
  if (!is_contained(ArrayRef({MVT::i8, MVT::i16, MVT::i32, MVT::i64, MVT::f16,
                              MVT::bf16, MVT::f32, MVT::f64}),
                    EltVT.getSimpleVT().SimpleTy))
    return SDValue();

  SDValue Mask = Idx.getOperand(0);
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  if (!TLI.isOperationLegal(ISD::VECTOR_FIND_LAST_ACTIVE, Mask.getValueType()))
    return SDValue();

  return DAG.getNode(AArch64ISD::LASTB, SDLoc(N), N->getValueType(0), Mask,
                     Vec);
}

static SDValue
performExtractVectorEltCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                               const AArch64Subtarget *Subtarget) {
  assert(N->getOpcode() == ISD::EXTRACT_VECTOR_ELT);
  if (SDValue Res = performFirstTrueTestVectorCombine(N, DCI, Subtarget))
    return Res;
  if (SDValue Res = performLastTrueTestVectorCombine(N, DCI, Subtarget))
    return Res;
  if (SDValue Res = performExtractLastActiveCombine(N, DCI, Subtarget))
    return Res;

  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);

  EVT VT = N->getValueType(0);
  const bool FullFP16 = Subtarget->hasFullFP16();
  bool IsStrict = N0->isStrictFPOpcode();

  // extract(dup x) -> x
  if (N0.getOpcode() == AArch64ISD::DUP)
    return VT.isInteger() ? DAG.getZExtOrTrunc(N0.getOperand(0), SDLoc(N), VT)
                          : N0.getOperand(0);

  // Rewrite for pairwise fadd pattern
  //   (f32 (extract_vector_elt
  //           (fadd (vXf32 Other)
  //                 (vector_shuffle (vXf32 Other) undef <1,X,...> )) 0))
  // ->
  //   (f32 (fadd (extract_vector_elt (vXf32 Other) 0)
  //              (extract_vector_elt (vXf32 Other) 1))
  // For strict_fadd we need to make sure the old strict_fadd can be deleted, so
  // we can only do this when it's used only by the extract_vector_elt.
  if (isNullConstant(N1) && hasPairwiseAdd(N0->getOpcode(), VT, FullFP16) &&
      (!IsStrict || N0.hasOneUse())) {
    SDLoc DL(N0);
    SDValue N00 = N0->getOperand(IsStrict ? 1 : 0);
    SDValue N01 = N0->getOperand(IsStrict ? 2 : 1);

    ShuffleVectorSDNode *Shuffle = dyn_cast<ShuffleVectorSDNode>(N01);
    SDValue Other = N00;

    // And handle the commutative case.
    if (!Shuffle) {
      Shuffle = dyn_cast<ShuffleVectorSDNode>(N00);
      Other = N01;
    }

    if (Shuffle && Shuffle->getMaskElt(0) == 1 &&
        Other == Shuffle->getOperand(0)) {
      SDValue Extract1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Other,
                                     DAG.getConstant(0, DL, MVT::i64));
      SDValue Extract2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Other,
                                     DAG.getConstant(1, DL, MVT::i64));
      if (!IsStrict)
        return DAG.getNode(N0->getOpcode(), DL, VT, Extract1, Extract2);

      // For strict_fadd we need uses of the final extract_vector to be replaced
      // with the strict_fadd, but we also need uses of the chain output of the
      // original strict_fadd to use the chain output of the new strict_fadd as
      // otherwise it may not be deleted.
      SDValue Ret = DAG.getNode(N0->getOpcode(), DL,
                                {VT, MVT::Other},
                                {N0->getOperand(0), Extract1, Extract2});
      DAG.ReplaceAllUsesOfValueWith(SDValue(N, 0), Ret);
      DAG.ReplaceAllUsesOfValueWith(N0.getValue(1), Ret.getValue(1));
      return SDValue(N, 0);
    }
  }

  return SDValue();
}

static SDValue performConcatVectorsCombine(SDNode *N,
                                           TargetLowering::DAGCombinerInfo &DCI,
                                           SelectionDAG &DAG) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue N0 = N->getOperand(0), N1 = N->getOperand(1);
  unsigned N0Opc = N0->getOpcode(), N1Opc = N1->getOpcode();

  if (VT.isScalableVector())
    return SDValue();

  if (N->getNumOperands() == 2 && N0Opc == ISD::TRUNCATE &&
      N1Opc == ISD::TRUNCATE) {
    SDValue N00 = N0->getOperand(0);
    SDValue N10 = N1->getOperand(0);
    EVT N00VT = N00.getValueType();
    unsigned N00Opc = N00.getOpcode(), N10Opc = N10.getOpcode();

    // Optimize concat_vectors of truncated vectors, where the intermediate
    // type is illegal, to avoid said illegality,  e.g.,
    //   (v4i16 (concat_vectors (v2i16 (truncate (v2i64))),
    //                          (v2i16 (truncate (v2i64)))))
    // ->
    //   (v4i16 (truncate (vector_shuffle (v4i32 (bitcast (v2i64))),
    //                                    (v4i32 (bitcast (v2i64))),
    //                                    <0, 2, 4, 6>)))
    // This isn't really target-specific, but ISD::TRUNCATE legality isn't keyed
    // on both input and result type, so we might generate worse code.
    // On AArch64 we know it's fine for v2i64->v4i16 and v4i32->v8i8.
    if (N00VT == N10.getValueType() &&
        (N00VT == MVT::v2i64 || N00VT == MVT::v4i32) &&
        N00VT.getScalarSizeInBits() == 4 * VT.getScalarSizeInBits()) {
      MVT MidVT = (N00VT == MVT::v2i64 ? MVT::v4i32 : MVT::v8i16);
      SmallVector<int, 8> Mask(MidVT.getVectorNumElements());
      for (size_t i = 0; i < Mask.size(); ++i)
        Mask[i] = i * 2;
      return DAG.getNode(ISD::TRUNCATE, DL, VT,
                         DAG.getVectorShuffle(
                             MidVT, DL,
                             DAG.getNode(ISD::BITCAST, DL, MidVT, N00),
                             DAG.getNode(ISD::BITCAST, DL, MidVT, N10), Mask));
    }

    // Optimize two large shifts and a combine into a single combine and shift
    // For AArch64 architectures, sequences like the following:
    //
    //     ushr    v0.4s, v0.4s, #20
    //     ushr    v1.4s, v1.4s, #20
    //     uzp1    v0.8h, v0.8h, v1.8h
    //
    // Can be optimized to:
    //
    //     uzp2    v0.8h, v0.8h, v1.8h
    //     ushr    v0.8h, v0.8h, #4
    //
    // This optimization reduces instruction count.
    if (N00Opc == AArch64ISD::VLSHR && N10Opc == AArch64ISD::VLSHR &&
        N00->getOperand(1) == N10->getOperand(1)) {
      SDValue N000 = N00->getOperand(0);
      SDValue N100 = N10->getOperand(0);
      uint64_t N001ConstVal = N00->getConstantOperandVal(1),
               N101ConstVal = N10->getConstantOperandVal(1),
               NScalarSize = N->getValueType(0).getScalarSizeInBits();

      if (N001ConstVal == N101ConstVal && N001ConstVal > NScalarSize) {
        N000 = DAG.getNode(AArch64ISD::NVCAST, DL, VT, N000);
        N100 = DAG.getNode(AArch64ISD::NVCAST, DL, VT, N100);
        SDValue Uzp = DAG.getNode(AArch64ISD::UZP2, DL, VT, N000, N100);
        SDValue NewShiftConstant =
            DAG.getConstant(N001ConstVal - NScalarSize, DL, MVT::i32);

        return DAG.getNode(AArch64ISD::VLSHR, DL, VT, Uzp, NewShiftConstant);
      }
    }
  }

  if (N->getOperand(0).getValueType() == MVT::v4i8 ||
      N->getOperand(0).getValueType() == MVT::v2i16 ||
      N->getOperand(0).getValueType() == MVT::v2i8) {
    EVT SrcVT = N->getOperand(0).getValueType();
    // If we have a concat of v4i8 loads, convert them to a buildvector of f32
    // loads to prevent having to go through the v4i8 load legalization that
    // needs to extend each element into a larger type.
    if (N->getNumOperands() % 2 == 0 &&
        all_of(N->op_values(), [SrcVT](SDValue V) {
          if (V.getValueType() != SrcVT)
            return false;
          if (V.isUndef())
            return true;
          LoadSDNode *LD = dyn_cast<LoadSDNode>(V);
          return LD && V.hasOneUse() && LD->isSimple() && !LD->isIndexed() &&
                 LD->getExtensionType() == ISD::NON_EXTLOAD;
        })) {
      EVT FVT = SrcVT == MVT::v2i8 ? MVT::f16 : MVT::f32;
      EVT NVT = EVT::getVectorVT(*DAG.getContext(), FVT, N->getNumOperands());
      SmallVector<SDValue> Ops;

      for (unsigned i = 0; i < N->getNumOperands(); i++) {
        SDValue V = N->getOperand(i);
        if (V.isUndef())
          Ops.push_back(DAG.getUNDEF(FVT));
        else {
          LoadSDNode *LD = cast<LoadSDNode>(V);
          SDValue NewLoad = DAG.getLoad(FVT, DL, LD->getChain(),
                                        LD->getBasePtr(), LD->getMemOperand());
          DAG.ReplaceAllUsesOfValueWith(SDValue(LD, 1), NewLoad.getValue(1));
          Ops.push_back(NewLoad);
        }
      }
      return DAG.getBitcast(N->getValueType(0),
                            DAG.getBuildVector(NVT, DL, Ops));
    }
  }

  // Canonicalise concat_vectors to replace concatenations of truncated nots
  // with nots of concatenated truncates. This in some cases allows for multiple
  // redundant negations to be eliminated.
  //  (concat_vectors (v4i16 (truncate (not (v4i32)))),
  //                  (v4i16 (truncate (not (v4i32)))))
  // ->
  //  (not (concat_vectors (v4i16 (truncate (v4i32))),
  //                       (v4i16 (truncate (v4i32)))))
  if (N->getNumOperands() == 2 && N0Opc == ISD::TRUNCATE &&
      N1Opc == ISD::TRUNCATE && N->isOnlyUserOf(N0.getNode()) &&
      N->isOnlyUserOf(N1.getNode())) {
    auto isBitwiseVectorNegate = [](SDValue V) {
      return V->getOpcode() == ISD::XOR &&
             ISD::isConstantSplatVectorAllOnes(V.getOperand(1).getNode());
    };
    SDValue N00 = N0->getOperand(0);
    SDValue N10 = N1->getOperand(0);
    if (isBitwiseVectorNegate(N00) && N0->isOnlyUserOf(N00.getNode()) &&
        isBitwiseVectorNegate(N10) && N1->isOnlyUserOf(N10.getNode())) {
      return DAG.getNOT(
          DL,
          DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
                      DAG.getNode(ISD::TRUNCATE, DL, N0.getValueType(),
                                  N00->getOperand(0)),
                      DAG.getNode(ISD::TRUNCATE, DL, N1.getValueType(),
                                  N10->getOperand(0))),
          VT);
    }
  }

  // Wait till after everything is legalized to try this. That way we have
  // legal vector types and such.
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  // Optimise concat_vectors of two identical binops with a 128-bit destination
  // size, combine into an binop of two contacts of the source vectors. eg:
  // concat(uhadd(a,b), uhadd(c, d)) -> uhadd(concat(a, c), concat(b, d))
  if (N->getNumOperands() == 2 && N0Opc == N1Opc && VT.is128BitVector() &&
      (DAG.getTargetLoweringInfo().isBinOp(N0Opc) ||
       isVectorizedBinOp(N0Opc)) &&
      N0->hasOneUse() && N1->hasOneUse()) {
    SDValue N00 = N0->getOperand(0);
    SDValue N01 = N0->getOperand(1);
    SDValue N10 = N1->getOperand(0);
    SDValue N11 = N1->getOperand(1);

    if (!N00.isUndef() && !N01.isUndef() && !N10.isUndef() && !N11.isUndef()) {
      SDValue Concat0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, N00, N10);
      SDValue Concat1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, N01, N11);
      return DAG.getNode(N0Opc, DL, VT, Concat0, Concat1);
    }
  }

  auto IsRSHRN = [](SDValue Shr) {
    if (Shr.getOpcode() != AArch64ISD::VLSHR)
      return false;
    SDValue Op = Shr.getOperand(0);
    EVT VT = Op.getValueType();
    unsigned ShtAmt = Shr.getConstantOperandVal(1);
    if (ShtAmt > VT.getScalarSizeInBits() / 2 || Op.getOpcode() != ISD::ADD)
      return false;

    APInt Imm;
    if (Op.getOperand(1).getOpcode() == AArch64ISD::MOVIshift)
      Imm = APInt(VT.getScalarSizeInBits(),
                  Op.getOperand(1).getConstantOperandVal(0)
                      << Op.getOperand(1).getConstantOperandVal(1));
    else if (Op.getOperand(1).getOpcode() == AArch64ISD::DUP &&
             isa<ConstantSDNode>(Op.getOperand(1).getOperand(0)))
      Imm = APInt(VT.getScalarSizeInBits(),
                  Op.getOperand(1).getConstantOperandVal(0));
    else
      return false;

    if (Imm != 1ULL << (ShtAmt - 1))
      return false;
    return true;
  };

  // concat(rshrn(x), rshrn(y)) -> rshrn(concat(x, y))
  if (N->getNumOperands() == 2 && IsRSHRN(N0) &&
      ((IsRSHRN(N1) &&
        N0.getConstantOperandVal(1) == N1.getConstantOperandVal(1)) ||
       N1.isUndef())) {
    SDValue X = N0.getOperand(0).getOperand(0);
    SDValue Y = N1.isUndef() ? DAG.getUNDEF(X.getValueType())
                             : N1.getOperand(0).getOperand(0);
    EVT BVT =
        X.getValueType().getDoubleNumVectorElementsVT(*DCI.DAG.getContext());
    SDValue CC = DAG.getNode(ISD::CONCAT_VECTORS, DL, BVT, X, Y);
    SDValue Add = DAG.getNode(
        ISD::ADD, DL, BVT, CC,
        DAG.getConstant(1ULL << (N0.getConstantOperandVal(1) - 1), DL, BVT));
    SDValue Shr =
        DAG.getNode(AArch64ISD::VLSHR, DL, BVT, Add, N0.getOperand(1));
    return Shr;
  }

  // concat(zip1(a, b), zip2(a, b)) is zip1(a, b)
  if (N->getNumOperands() == 2 && N0Opc == AArch64ISD::ZIP1 &&
      N1Opc == AArch64ISD::ZIP2 && N0.getOperand(0) == N1.getOperand(0) &&
      N0.getOperand(1) == N1.getOperand(1)) {
    SDValue E0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, N0.getOperand(0),
                             DAG.getUNDEF(N0.getValueType()));
    SDValue E1 = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, N0.getOperand(1),
                             DAG.getUNDEF(N0.getValueType()));
    return DAG.getNode(AArch64ISD::ZIP1, DL, VT, E0, E1);
  }

  // If we see a (concat_vectors (v1x64 A), (v1x64 A)) it's really a vector
  // splat. The indexed instructions are going to be expecting a DUPLANE64, so
  // canonicalise to that.
  if (N->getNumOperands() == 2 && N0 == N1 && VT.getVectorNumElements() == 2) {
    assert(VT.getScalarSizeInBits() == 64);
    return DAG.getNode(AArch64ISD::DUPLANE64, DL, VT, WidenVector(N0, DAG),
                       DAG.getConstant(0, DL, MVT::i64));
  }

  // Canonicalise concat_vectors so that the right-hand vector has as few
  // bit-casts as possible before its real operation. The primary matching
  // destination for these operations will be the narrowing "2" instructions,
  // which depend on the operation being performed on this right-hand vector.
  // For example,
  //    (concat_vectors LHS,  (v1i64 (bitconvert (v4i16 RHS))))
  // becomes
  //    (bitconvert (concat_vectors (v4i16 (bitconvert LHS)), RHS))

  if (N->getNumOperands() != 2 || N1Opc != ISD::BITCAST)
    return SDValue();
  SDValue RHS = N1->getOperand(0);
  MVT RHSTy = RHS.getValueType().getSimpleVT();
  // If the RHS is not a vector, this is not the pattern we're looking for.
  if (!RHSTy.isVector())
    return SDValue();

  LLVM_DEBUG(
      dbgs() << "aarch64-lower: concat_vectors bitcast simplification\n");

  MVT ConcatTy = MVT::getVectorVT(RHSTy.getVectorElementType(),
                                  RHSTy.getVectorNumElements() * 2);
  return DAG.getNode(ISD::BITCAST, DL, VT,
                     DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatTy,
                                 DAG.getNode(ISD::BITCAST, DL, RHSTy, N0),
                                 RHS));
}

static SDValue
performExtractSubvectorCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                               SelectionDAG &DAG) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  EVT VT = N->getValueType(0);
  if (!VT.isScalableVector() || VT.getVectorElementType() != MVT::i1)
    return SDValue();

  SDValue V = N->getOperand(0);

  // NOTE: This combine exists in DAGCombiner, but that version's legality check
  // blocks this combine because the non-const case requires custom lowering.
  //
  // ty1 extract_vector(ty2 splat(const))) -> ty1 splat(const)
  if (V.getOpcode() == ISD::SPLAT_VECTOR)
    if (isa<ConstantSDNode>(V.getOperand(0)))
      return DAG.getNode(ISD::SPLAT_VECTOR, SDLoc(N), VT, V.getOperand(0));

  return SDValue();
}

static SDValue
performInsertSubvectorCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                              SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue Vec = N->getOperand(0);
  SDValue SubVec = N->getOperand(1);
  uint64_t IdxVal = N->getConstantOperandVal(2);
  EVT VecVT = Vec.getValueType();
  EVT SubVT = SubVec.getValueType();

  // Promote fixed length vector zeros.
  if (VecVT.isScalableVector() && SubVT.isFixedLengthVector() &&
      Vec.isUndef() && isZerosVector(SubVec.getNode()))
    return VecVT.isInteger() ? DAG.getConstant(0, DL, VecVT)
                             : DAG.getConstantFP(0, DL, VecVT);

  // Only do this for legal fixed vector types.
  if (!VecVT.isFixedLengthVector() ||
      !DAG.getTargetLoweringInfo().isTypeLegal(VecVT) ||
      !DAG.getTargetLoweringInfo().isTypeLegal(SubVT))
    return SDValue();

  // Ignore widening patterns.
  if (IdxVal == 0 && Vec.isUndef())
    return SDValue();

  // Subvector must be half the width and an "aligned" insertion.
  unsigned NumSubElts = SubVT.getVectorNumElements();
  if ((SubVT.getSizeInBits() * 2) != VecVT.getSizeInBits() ||
      (IdxVal != 0 && IdxVal != NumSubElts))
    return SDValue();

  // Fold insert_subvector -> concat_vectors
  // insert_subvector(Vec,Sub,lo) -> concat_vectors(Sub,extract(Vec,hi))
  // insert_subvector(Vec,Sub,hi) -> concat_vectors(extract(Vec,lo),Sub)
  SDValue Lo, Hi;
  if (IdxVal == 0) {
    Lo = SubVec;
    Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, Vec,
                     DAG.getVectorIdxConstant(NumSubElts, DL));
  } else {
    Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, Vec,
                     DAG.getVectorIdxConstant(0, DL));
    Hi = SubVec;
  }
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VecVT, Lo, Hi);
}

static SDValue tryCombineFixedPointConvert(SDNode *N,
                                           TargetLowering::DAGCombinerInfo &DCI,
                                           SelectionDAG &DAG) {
  // Wait until after everything is legalized to try this. That way we have
  // legal vector types and such.
  if (DCI.isBeforeLegalizeOps())
    return SDValue();
  // Transform a scalar conversion of a value from a lane extract into a
  // lane extract of a vector conversion. E.g., from foo1 to foo2:
  // double foo1(int64x2_t a) { return vcvtd_n_f64_s64(a[1], 9); }
  // double foo2(int64x2_t a) { return vcvtq_n_f64_s64(a, 9)[1]; }
  //
  // The second form interacts better with instruction selection and the
  // register allocator to avoid cross-class register copies that aren't
  // coalescable due to a lane reference.

  // Check the operand and see if it originates from a lane extract.
  SDValue Op1 = N->getOperand(1);
  if (Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
    return SDValue();

  // Yep, no additional predication needed. Perform the transform.
  SDValue IID = N->getOperand(0);
  SDValue Shift = N->getOperand(2);
  SDValue Vec = Op1.getOperand(0);
  SDValue Lane = Op1.getOperand(1);
  EVT ResTy = N->getValueType(0);
  EVT VecResTy;
  SDLoc DL(N);

  // The vector width should be 128 bits by the time we get here, even
  // if it started as 64 bits (the extract_vector handling will have
  // done so). Bail if it is not.
  if (Vec.getValueSizeInBits() != 128)
    return SDValue();

  if (Vec.getValueType() == MVT::v4i32)
    VecResTy = MVT::v4f32;
  else if (Vec.getValueType() == MVT::v2i64)
    VecResTy = MVT::v2f64;
  else
    return SDValue();

  SDValue Convert =
      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VecResTy, IID, Vec, Shift);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResTy, Convert, Lane);
}

// AArch64 high-vector "long" operations are formed by performing the non-high
// version on an extract_subvector of each operand which gets the high half:
//
//  (longop2 LHS, RHS) == (longop (extract_high LHS), (extract_high RHS))
//
// However, there are cases which don't have an extract_high explicitly, but
// have another operation that can be made compatible with one for free. For
// example:
//
//  (dupv64 scalar) --> (extract_high (dup128 scalar))
//
// This routine does the actual conversion of such DUPs, once outer routines
// have determined that everything else is in order.
// It also supports immediate DUP-like nodes (MOVI/MVNi), which we can fold
// similarly here.
static SDValue tryExtendDUPToExtractHigh(SDValue N, SelectionDAG &DAG) {
  MVT VT = N.getSimpleValueType();
  if (N.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
      N.getConstantOperandVal(1) == 0)
    N = N.getOperand(0);

  switch (N.getOpcode()) {
  case AArch64ISD::DUP:
  case AArch64ISD::DUPLANE8:
  case AArch64ISD::DUPLANE16:
  case AArch64ISD::DUPLANE32:
  case AArch64ISD::DUPLANE64:
  case AArch64ISD::MOVI:
  case AArch64ISD::MOVIshift:
  case AArch64ISD::MOVIedit:
  case AArch64ISD::MOVImsl:
  case AArch64ISD::MVNIshift:
  case AArch64ISD::MVNImsl:
    break;
  default:
    // FMOV could be supported, but isn't very useful, as it would only occur
    // if you passed a bitcast' floating point immediate to an eligible long
    // integer op (addl, smull, ...).
    return SDValue();
  }

  if (!VT.is64BitVector())
    return SDValue();

  SDLoc DL(N);
  unsigned NumElems = VT.getVectorNumElements();
  if (N.getValueType().is64BitVector()) {
    MVT ElementTy = VT.getVectorElementType();
    MVT NewVT = MVT::getVectorVT(ElementTy, NumElems * 2);
    N = DAG.getNode(N->getOpcode(), DL, NewVT, N->ops());
  }

  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, N,
                     DAG.getConstant(NumElems, DL, MVT::i64));
}

static bool isEssentiallyExtractHighSubvector(SDValue N) {
  if (N.getOpcode() == ISD::BITCAST)
    N = N.getOperand(0);
  if (N.getOpcode() != ISD::EXTRACT_SUBVECTOR)
    return false;
  if (N.getOperand(0).getValueType().isScalableVector())
    return false;
  return N.getConstantOperandAPInt(1) ==
         N.getOperand(0).getValueType().getVectorNumElements() / 2;
}

/// Helper structure to keep track of ISD::SET_CC operands.
struct GenericSetCCInfo {
  const SDValue *Opnd0;
  const SDValue *Opnd1;
  ISD::CondCode CC;
};

/// Helper structure to keep track of a SET_CC lowered into AArch64 code.
struct AArch64SetCCInfo {
  const SDValue *Cmp;
  AArch64CC::CondCode CC;
};

/// Helper structure to keep track of SetCC information.
union SetCCInfo {
  GenericSetCCInfo Generic;
  AArch64SetCCInfo AArch64;
};

/// Helper structure to be able to read SetCC information.  If set to
/// true, IsAArch64 field, Info is a AArch64SetCCInfo, otherwise Info is a
/// GenericSetCCInfo.
struct SetCCInfoAndKind {
  SetCCInfo Info;
  bool IsAArch64;
};

/// Check whether or not \p Op is a SET_CC operation, either a generic or
/// an
/// AArch64 lowered one.
/// \p SetCCInfo is filled accordingly.
/// \post SetCCInfo is meanginfull only when this function returns true.
/// \return True when Op is a kind of SET_CC operation.
static bool isSetCC(SDValue Op, SetCCInfoAndKind &SetCCInfo) {
  // If this is a setcc, this is straight forward.
  if (Op.getOpcode() == ISD::SETCC) {
    SetCCInfo.Info.Generic.Opnd0 = &Op.getOperand(0);
    SetCCInfo.Info.Generic.Opnd1 = &Op.getOperand(1);
    SetCCInfo.Info.Generic.CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
    SetCCInfo.IsAArch64 = false;
    return true;
  }
  // Otherwise, check if this is a matching csel instruction.
  // In other words:
  // - csel 1, 0, cc
  // - csel 0, 1, !cc
  if (Op.getOpcode() != AArch64ISD::CSEL)
    return false;
  // Set the information about the operands.
  // TODO: we want the operands of the Cmp not the csel
  SetCCInfo.Info.AArch64.Cmp = &Op.getOperand(3);
  SetCCInfo.IsAArch64 = true;
  SetCCInfo.Info.AArch64.CC =
      static_cast<AArch64CC::CondCode>(Op.getConstantOperandVal(2));

  // Check that the operands matches the constraints:
  // (1) Both operands must be constants.
  // (2) One must be 1 and the other must be 0.
  ConstantSDNode *TValue = dyn_cast<ConstantSDNode>(Op.getOperand(0));
  ConstantSDNode *FValue = dyn_cast<ConstantSDNode>(Op.getOperand(1));

  // Check (1).
  if (!TValue || !FValue)
    return false;

  // Check (2).
  if (!TValue->isOne()) {
    // Update the comparison when we are interested in !cc.
    std::swap(TValue, FValue);
    SetCCInfo.Info.AArch64.CC =
        AArch64CC::getInvertedCondCode(SetCCInfo.Info.AArch64.CC);
  }
  return TValue->isOne() && FValue->isZero();
}

// Returns true if Op is setcc or zext of setcc.
static bool isSetCCOrZExtSetCC(const SDValue& Op, SetCCInfoAndKind &Info) {
  if (isSetCC(Op, Info))
    return true;
  return ((Op.getOpcode() == ISD::ZERO_EXTEND) &&
    isSetCC(Op->getOperand(0), Info));
}

// The folding we want to perform is:
// (add x, [zext] (setcc cc ...) )
//   -->
// (csel x, (add x, 1), !cc ...)
//
// The latter will get matched to a CSINC instruction.
static SDValue performSetccAddFolding(SDNode *Op, SelectionDAG &DAG) {
  assert(Op && Op->getOpcode() == ISD::ADD && "Unexpected operation!");
  SDValue LHS = Op->getOperand(0);
  SDValue RHS = Op->getOperand(1);
  SetCCInfoAndKind InfoAndKind;

  // If both operands are a SET_CC, then we don't want to perform this
  // folding and create another csel as this results in more instructions
  // (and higher register usage).
  if (isSetCCOrZExtSetCC(LHS, InfoAndKind) &&
      isSetCCOrZExtSetCC(RHS, InfoAndKind))
    return SDValue();

  // If neither operand is a SET_CC, give up.
  if (!isSetCCOrZExtSetCC(LHS, InfoAndKind)) {
    std::swap(LHS, RHS);
    if (!isSetCCOrZExtSetCC(LHS, InfoAndKind))
      return SDValue();
  }

  // FIXME: This could be generatized to work for FP comparisons.
  EVT CmpVT = InfoAndKind.IsAArch64
                  ? InfoAndKind.Info.AArch64.Cmp->getOperand(0).getValueType()
                  : InfoAndKind.Info.Generic.Opnd0->getValueType();
  if (CmpVT != MVT::i32 && CmpVT != MVT::i64)
    return SDValue();

  SDValue CCVal;
  SDValue Cmp;
  SDLoc DL(Op);
  if (InfoAndKind.IsAArch64) {
    CCVal = DAG.getConstant(
        AArch64CC::getInvertedCondCode(InfoAndKind.Info.AArch64.CC), DL,
        MVT::i32);
    Cmp = *InfoAndKind.Info.AArch64.Cmp;
  } else
    Cmp = getAArch64Cmp(
        *InfoAndKind.Info.Generic.Opnd0, *InfoAndKind.Info.Generic.Opnd1,
        ISD::getSetCCInverse(InfoAndKind.Info.Generic.CC, CmpVT), CCVal, DAG,
        DL);

  EVT VT = Op->getValueType(0);
  LHS = DAG.getNode(ISD::ADD, DL, VT, RHS, DAG.getConstant(1, DL, VT));
  return DAG.getNode(AArch64ISD::CSEL, DL, VT, RHS, LHS, CCVal, Cmp);
}

// ADD(UADDV a, UADDV b) -->  UADDV(ADD a, b)
static SDValue performAddUADDVCombine(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  // Only scalar integer and vector types.
  if (N->getOpcode() != ISD::ADD || !VT.isScalarInteger())
    return SDValue();

  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  if (LHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
      RHS.getOpcode() != ISD::EXTRACT_VECTOR_ELT || LHS.getValueType() != VT)
    return SDValue();

  auto *LHSN1 = dyn_cast<ConstantSDNode>(LHS->getOperand(1));
  auto *RHSN1 = dyn_cast<ConstantSDNode>(RHS->getOperand(1));
  if (!LHSN1 || LHSN1 != RHSN1 || !RHSN1->isZero())
    return SDValue();

  SDValue Op1 = LHS->getOperand(0);
  SDValue Op2 = RHS->getOperand(0);
  EVT OpVT1 = Op1.getValueType();
  EVT OpVT2 = Op2.getValueType();
  if (Op1.getOpcode() != AArch64ISD::UADDV || OpVT1 != OpVT2 ||
      Op2.getOpcode() != AArch64ISD::UADDV ||
      OpVT1.getVectorElementType() != VT)
    return SDValue();

  SDValue Val1 = Op1.getOperand(0);
  SDValue Val2 = Op2.getOperand(0);
  EVT ValVT = Val1->getValueType(0);
  SDLoc DL(N);
  SDValue AddVal = DAG.getNode(ISD::ADD, DL, ValVT, Val1, Val2);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
                     DAG.getNode(AArch64ISD::UADDV, DL, ValVT, AddVal),
                     DAG.getConstant(0, DL, MVT::i64));
}

/// Perform the scalar expression combine in the form of:
///   CSEL(c, 1, cc) + b => CSINC(b+c, b, cc)
///   CSNEG(c, -1, cc) + b => CSINC(b+c, b, cc)
static SDValue performAddCSelIntoCSinc(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if (!VT.isScalarInteger() || N->getOpcode() != ISD::ADD)
    return SDValue();

  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);

  // Handle commutivity.
  if (LHS.getOpcode() != AArch64ISD::CSEL &&
      LHS.getOpcode() != AArch64ISD::CSNEG) {
    std::swap(LHS, RHS);
    if (LHS.getOpcode() != AArch64ISD::CSEL &&
        LHS.getOpcode() != AArch64ISD::CSNEG) {
      return SDValue();
    }
  }

  if (!LHS.hasOneUse())
    return SDValue();

  AArch64CC::CondCode AArch64CC =
      static_cast<AArch64CC::CondCode>(LHS.getConstantOperandVal(2));

  // The CSEL should include a const one operand, and the CSNEG should include
  // One or NegOne operand.
  ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(LHS.getOperand(0));
  ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
  if (!CTVal || !CFVal)
    return SDValue();

  if (!(LHS.getOpcode() == AArch64ISD::CSEL &&
        (CTVal->isOne() || CFVal->isOne())) &&
      !(LHS.getOpcode() == AArch64ISD::CSNEG &&
        (CTVal->isOne() || CFVal->isAllOnes())))
    return SDValue();

  // Switch CSEL(1, c, cc) to CSEL(c, 1, !cc)
  if (LHS.getOpcode() == AArch64ISD::CSEL && CTVal->isOne() &&
      !CFVal->isOne()) {
    std::swap(CTVal, CFVal);
    AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
  }

  SDLoc DL(N);
  // Switch CSNEG(1, c, cc) to CSNEG(-c, -1, !cc)
  if (LHS.getOpcode() == AArch64ISD::CSNEG && CTVal->isOne() &&
      !CFVal->isAllOnes()) {
    APInt C = -1 * CFVal->getAPIntValue();
    CTVal = cast<ConstantSDNode>(DAG.getConstant(C, DL, VT));
    CFVal = cast<ConstantSDNode>(DAG.getAllOnesConstant(DL, VT));
    AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
  }

  // It might be neutral for larger constants, as the immediate need to be
  // materialized in a register.
  APInt ADDC = CTVal->getAPIntValue();
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  if (!TLI.isLegalAddImmediate(ADDC.getSExtValue()))
    return SDValue();

  assert(((LHS.getOpcode() == AArch64ISD::CSEL && CFVal->isOne()) ||
          (LHS.getOpcode() == AArch64ISD::CSNEG && CFVal->isAllOnes())) &&
         "Unexpected constant value");

  SDValue NewNode = DAG.getNode(ISD::ADD, DL, VT, RHS, SDValue(CTVal, 0));
  SDValue CCVal = DAG.getConstant(AArch64CC, DL, MVT::i32);
  SDValue Cmp = LHS.getOperand(3);

  return DAG.getNode(AArch64ISD::CSINC, DL, VT, NewNode, RHS, CCVal, Cmp);
}

// ADD(UDOT(zero, x, y), A) -->  UDOT(A, x, y)
static SDValue performAddDotCombine(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if (N->getOpcode() != ISD::ADD)
    return SDValue();

  SDValue Dot = N->getOperand(0);
  SDValue A = N->getOperand(1);
  // Handle commutivity
  auto isZeroDot = [](SDValue Dot) {
    return (Dot.getOpcode() == AArch64ISD::UDOT ||
            Dot.getOpcode() == AArch64ISD::SDOT) &&
           isZerosVector(Dot.getOperand(0).getNode());
  };
  if (!isZeroDot(Dot))
    std::swap(Dot, A);
  if (!isZeroDot(Dot))
    return SDValue();

  return DAG.getNode(Dot.getOpcode(), SDLoc(N), VT, A, Dot.getOperand(1),
                     Dot.getOperand(2));
}

static bool isNegatedInteger(SDValue Op) {
  return Op.getOpcode() == ISD::SUB && isNullConstant(Op.getOperand(0));
}

static SDValue getNegatedInteger(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue Zero = DAG.getConstant(0, DL, VT);
  return DAG.getNode(ISD::SUB, DL, VT, Zero, Op);
}

// Try to fold
//
// (neg (csel X, Y)) -> (csel (neg X), (neg Y))
//
// The folding helps csel to be matched with csneg without generating
// redundant neg instruction, which includes negation of the csel expansion
// of abs node lowered by lowerABS.
static SDValue performNegCSelCombine(SDNode *N, SelectionDAG &DAG) {
  if (!isNegatedInteger(SDValue(N, 0)))
    return SDValue();

  SDValue CSel = N->getOperand(1);
  if (CSel.getOpcode() != AArch64ISD::CSEL || !CSel->hasOneUse())
    return SDValue();

  SDValue N0 = CSel.getOperand(0);
  SDValue N1 = CSel.getOperand(1);

  // If both of them is not negations, it's not worth the folding as it
  // introduces two additional negations while reducing one negation.
  if (!isNegatedInteger(N0) && !isNegatedInteger(N1))
    return SDValue();

  SDValue N0N = getNegatedInteger(N0, DAG);
  SDValue N1N = getNegatedInteger(N1, DAG);

  SDLoc DL(N);
  EVT VT = CSel.getValueType();
  return DAG.getNode(AArch64ISD::CSEL, DL, VT, N0N, N1N, CSel.getOperand(2),
                     CSel.getOperand(3));
}

// The basic add/sub long vector instructions have variants with "2" on the end
// which act on the high-half of their inputs. They are normally matched by
// patterns like:
//
// (add (zeroext (extract_high LHS)),
//      (zeroext (extract_high RHS)))
// -> uaddl2 vD, vN, vM
//
// However, if one of the extracts is something like a duplicate, this
// instruction can still be used profitably. This function puts the DAG into a
// more appropriate form for those patterns to trigger.
static SDValue performAddSubLongCombine(SDNode *N,
                                        TargetLowering::DAGCombinerInfo &DCI) {
  SelectionDAG &DAG = DCI.DAG;
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  MVT VT = N->getSimpleValueType(0);
  if (!VT.is128BitVector()) {
    if (N->getOpcode() == ISD::ADD)
      return performSetccAddFolding(N, DAG);
    return SDValue();
  }

  // Make sure both branches are extended in the same way.
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  if ((LHS.getOpcode() != ISD::ZERO_EXTEND &&
       LHS.getOpcode() != ISD::SIGN_EXTEND) ||
      LHS.getOpcode() != RHS.getOpcode())
    return SDValue();

  unsigned ExtType = LHS.getOpcode();

  // It's not worth doing if at least one of the inputs isn't already an
  // extract, but we don't know which it'll be so we have to try both.
  if (isEssentiallyExtractHighSubvector(LHS.getOperand(0))) {
    RHS = tryExtendDUPToExtractHigh(RHS.getOperand(0), DAG);
    if (!RHS.getNode())
      return SDValue();

    RHS = DAG.getNode(ExtType, SDLoc(N), VT, RHS);
  } else if (isEssentiallyExtractHighSubvector(RHS.getOperand(0))) {
    LHS = tryExtendDUPToExtractHigh(LHS.getOperand(0), DAG);
    if (!LHS.getNode())
      return SDValue();

    LHS = DAG.getNode(ExtType, SDLoc(N), VT, LHS);
  }

  return DAG.getNode(N->getOpcode(), SDLoc(N), VT, LHS, RHS);
}

static bool isCMP(SDValue Op) {
  return Op.getOpcode() == AArch64ISD::SUBS &&
         !Op.getNode()->hasAnyUseOfValue(0);
}

// (CSEL 1 0 CC Cond) => CC
// (CSEL 0 1 CC Cond) => !CC
static std::optional<AArch64CC::CondCode> getCSETCondCode(SDValue Op) {
  if (Op.getOpcode() != AArch64ISD::CSEL)
    return std::nullopt;
  auto CC = static_cast<AArch64CC::CondCode>(Op.getConstantOperandVal(2));
  if (CC == AArch64CC::AL || CC == AArch64CC::NV)
    return std::nullopt;
  SDValue OpLHS = Op.getOperand(0);
  SDValue OpRHS = Op.getOperand(1);
  if (isOneConstant(OpLHS) && isNullConstant(OpRHS))
    return CC;
  if (isNullConstant(OpLHS) && isOneConstant(OpRHS))
    return getInvertedCondCode(CC);

  return std::nullopt;
}

// (ADC{S} l r (CMP (CSET HS carry) 1)) => (ADC{S} l r carry)
// (SBC{S} l r (CMP 0 (CSET LO carry))) => (SBC{S} l r carry)
static SDValue foldOverflowCheck(SDNode *Op, SelectionDAG &DAG, bool IsAdd) {
  SDValue CmpOp = Op->getOperand(2);
  if (!isCMP(CmpOp))
    return SDValue();

  if (IsAdd) {
    if (!isOneConstant(CmpOp.getOperand(1)))
      return SDValue();
  } else {
    if (!isNullConstant(CmpOp.getOperand(0)))
      return SDValue();
  }

  SDValue CsetOp = CmpOp->getOperand(IsAdd ? 0 : 1);
  auto CC = getCSETCondCode(CsetOp);
  if (CC != (IsAdd ? AArch64CC::HS : AArch64CC::LO))
    return SDValue();

  return DAG.getNode(Op->getOpcode(), SDLoc(Op), Op->getVTList(),
                     Op->getOperand(0), Op->getOperand(1),
                     CsetOp.getOperand(3));
}

// (ADC x 0 cond) => (CINC x HS cond)
static SDValue foldADCToCINC(SDNode *N, SelectionDAG &DAG) {
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  SDValue Cond = N->getOperand(2);

  if (!isNullConstant(RHS))
    return SDValue();

  EVT VT = N->getValueType(0);
  SDLoc DL(N);

  // (CINC x cc cond) <=> (CSINC x x !cc cond)
  SDValue CC = DAG.getConstant(AArch64CC::LO, DL, MVT::i32);
  return DAG.getNode(AArch64ISD::CSINC, DL, VT, LHS, LHS, CC, Cond);
}

static SDValue performBuildVectorCombine(SDNode *N,
                                         TargetLowering::DAGCombinerInfo &DCI,
                                         SelectionDAG &DAG) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);

  if (DAG.getSubtarget<AArch64Subtarget>().isNeonAvailable() &&
      (VT == MVT::v4f16 || VT == MVT::v4bf16)) {
    SDValue Elt0 = N->getOperand(0), Elt1 = N->getOperand(1),
            Elt2 = N->getOperand(2), Elt3 = N->getOperand(3);
    if (Elt0->getOpcode() == ISD::FP_ROUND &&
        Elt1->getOpcode() == ISD::FP_ROUND &&
        isa<ConstantSDNode>(Elt0->getOperand(1)) &&
        isa<ConstantSDNode>(Elt1->getOperand(1)) &&
        Elt0->getConstantOperandVal(1) == Elt1->getConstantOperandVal(1) &&
        Elt0->getOperand(0)->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
        Elt1->getOperand(0)->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
        // Constant index.
        isa<ConstantSDNode>(Elt0->getOperand(0)->getOperand(1)) &&
        isa<ConstantSDNode>(Elt1->getOperand(0)->getOperand(1)) &&
        Elt0->getOperand(0)->getOperand(0) ==
            Elt1->getOperand(0)->getOperand(0) &&
        Elt0->getOperand(0)->getConstantOperandVal(1) == 0 &&
        Elt1->getOperand(0)->getConstantOperandVal(1) == 1) {
      SDValue LowLanesSrcVec = Elt0->getOperand(0)->getOperand(0);
      if (LowLanesSrcVec.getValueType() == MVT::v2f64) {
        SDValue HighLanes;
        if (Elt2->getOpcode() == ISD::UNDEF &&
            Elt3->getOpcode() == ISD::UNDEF) {
          HighLanes = DAG.getUNDEF(MVT::v2f32);
        } else if (Elt2->getOpcode() == ISD::FP_ROUND &&
                   Elt3->getOpcode() == ISD::FP_ROUND &&
                   isa<ConstantSDNode>(Elt2->getOperand(1)) &&
                   isa<ConstantSDNode>(Elt3->getOperand(1)) &&
                   Elt2->getConstantOperandVal(1) ==
                       Elt3->getConstantOperandVal(1) &&
                   Elt2->getOperand(0)->getOpcode() ==
                       ISD::EXTRACT_VECTOR_ELT &&
                   Elt3->getOperand(0)->getOpcode() ==
                       ISD::EXTRACT_VECTOR_ELT &&
                   // Constant index.
                   isa<ConstantSDNode>(Elt2->getOperand(0)->getOperand(1)) &&
                   isa<ConstantSDNode>(Elt3->getOperand(0)->getOperand(1)) &&
                   Elt2->getOperand(0)->getOperand(0) ==
                       Elt3->getOperand(0)->getOperand(0) &&
                   Elt2->getOperand(0)->getConstantOperandVal(1) == 0 &&
                   Elt3->getOperand(0)->getConstantOperandVal(1) == 1) {
          SDValue HighLanesSrcVec = Elt2->getOperand(0)->getOperand(0);
          HighLanes =
              DAG.getNode(AArch64ISD::FCVTXN, DL, MVT::v2f32, HighLanesSrcVec);
        }
        if (HighLanes) {
          SDValue DoubleToSingleSticky =
              DAG.getNode(AArch64ISD::FCVTXN, DL, MVT::v2f32, LowLanesSrcVec);
          SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
                                       DoubleToSingleSticky, HighLanes);
          return DAG.getNode(ISD::FP_ROUND, DL, VT, Concat,
                             Elt0->getOperand(1));
        }
      }
    }
  }

  if (VT == MVT::v2f64) {
    SDValue Elt0 = N->getOperand(0), Elt1 = N->getOperand(1);
    if (Elt0->getOpcode() == ISD::FP_EXTEND &&
        Elt1->getOpcode() == ISD::FP_EXTEND &&
        Elt0->getOperand(0)->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
        Elt1->getOperand(0)->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
        Elt0->getOperand(0)->getOperand(0) ==
            Elt1->getOperand(0)->getOperand(0) &&
        // Constant index.
        isa<ConstantSDNode>(Elt0->getOperand(0)->getOperand(1)) &&
        isa<ConstantSDNode>(Elt1->getOperand(0)->getOperand(1)) &&
        Elt0->getOperand(0)->getConstantOperandVal(1) + 1 ==
            Elt1->getOperand(0)->getConstantOperandVal(1) &&
        // EXTRACT_SUBVECTOR requires that Idx be a constant multiple of
        // ResultType's known minimum vector length.
        Elt0->getOperand(0)->getConstantOperandVal(1) %
                VT.getVectorMinNumElements() ==
            0) {
      SDValue SrcVec = Elt0->getOperand(0)->getOperand(0);
      if (SrcVec.getValueType() == MVT::v4f16 ||
          SrcVec.getValueType() == MVT::v4bf16) {
        SDValue HalfToSingle =
            DAG.getNode(ISD::FP_EXTEND, DL, MVT::v4f32, SrcVec);
        SDValue SubvectorIdx = Elt0->getOperand(0)->getOperand(1);
        SDValue Extract = DAG.getNode(
            ISD::EXTRACT_SUBVECTOR, DL, VT.changeVectorElementType(MVT::f32),
            HalfToSingle, SubvectorIdx);
        return DAG.getNode(ISD::FP_EXTEND, DL, VT, Extract);
      }
    }
  }

  // A build vector of two extracted elements is equivalent to an
  // extract subvector where the inner vector is any-extended to the
  // extract_vector_elt VT.
  //    (build_vector (extract_elt_iXX_to_i32 vec Idx+0)
  //                  (extract_elt_iXX_to_i32 vec Idx+1))
  // => (extract_subvector (anyext_iXX_to_i32 vec) Idx)

  // For now, only consider the v2i32 case, which arises as a result of
  // legalization.
  if (VT != MVT::v2i32)
    return SDValue();

  SDValue Elt0 = N->getOperand(0), Elt1 = N->getOperand(1);
  // Reminder, EXTRACT_VECTOR_ELT has the effect of any-extending to its VT.
  if (Elt0->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
      Elt1->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
      // Constant index.
      isa<ConstantSDNode>(Elt0->getOperand(1)) &&
      isa<ConstantSDNode>(Elt1->getOperand(1)) &&
      // Both EXTRACT_VECTOR_ELT from same vector...
      Elt0->getOperand(0) == Elt1->getOperand(0) &&
      // ... and contiguous. First element's index +1 == second element's index.
      Elt0->getConstantOperandVal(1) + 1 == Elt1->getConstantOperandVal(1) &&
      // EXTRACT_SUBVECTOR requires that Idx be a constant multiple of
      // ResultType's known minimum vector length.
      Elt0->getConstantOperandVal(1) % VT.getVectorMinNumElements() == 0) {
    SDValue VecToExtend = Elt0->getOperand(0);
    EVT ExtVT = VecToExtend.getValueType().changeVectorElementType(MVT::i32);
    if (!DAG.getTargetLoweringInfo().isTypeLegal(ExtVT))
      return SDValue();

    SDValue SubvectorIdx = DAG.getVectorIdxConstant(Elt0->getConstantOperandVal(1), DL);

    SDValue Ext = DAG.getNode(ISD::ANY_EXTEND, DL, ExtVT, VecToExtend);
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Ext,
                       SubvectorIdx);
  }

  return SDValue();
}

// A special combine for the sqdmulh family of instructions.
// smin( sra ( mul( sext v0, sext v1 ) ), SHIFT_AMOUNT ),
// SATURATING_VAL ) can be reduced to sqdmulh(...)
static SDValue trySQDMULHCombine(SDNode *N, SelectionDAG &DAG) {

  if (N->getOpcode() != ISD::SMIN)
    return SDValue();

  EVT DestVT = N->getValueType(0);

  if (!DestVT.isVector() || DestVT.getScalarSizeInBits() > 64 ||
      DestVT.isScalableVector())
    return SDValue();

  ConstantSDNode *Clamp = isConstOrConstSplat(N->getOperand(1));

  if (!Clamp)
    return SDValue();

  MVT ScalarType;
  unsigned ShiftAmt = 0;
  switch (Clamp->getSExtValue()) {
  case (1ULL << 15) - 1:
    ScalarType = MVT::i16;
    ShiftAmt = 16;
    break;
  case (1ULL << 31) - 1:
    ScalarType = MVT::i32;
    ShiftAmt = 32;
    break;
  default:
    return SDValue();
  }

  SDValue Sra = N->getOperand(0);
  if (Sra.getOpcode() != ISD::SRA || !Sra.hasOneUse())
    return SDValue();

  ConstantSDNode *RightShiftVec = isConstOrConstSplat(Sra.getOperand(1));
  if (!RightShiftVec)
    return SDValue();
  unsigned SExtValue = RightShiftVec->getSExtValue();

  if (SExtValue != (ShiftAmt - 1))
    return SDValue();

  SDValue Mul = Sra.getOperand(0);
  if (Mul.getOpcode() != ISD::MUL)
    return SDValue();

  SDValue SExt0 = Mul.getOperand(0);
  SDValue SExt1 = Mul.getOperand(1);

  if (SExt0.getOpcode() != ISD::SIGN_EXTEND ||
      SExt1.getOpcode() != ISD::SIGN_EXTEND)
    return SDValue();

  EVT SExt0Type = SExt0.getOperand(0).getValueType();
  EVT SExt1Type = SExt1.getOperand(0).getValueType();

  if (SExt0Type != SExt1Type || SExt0Type.getScalarType() != ScalarType ||
      SExt0Type.getFixedSizeInBits() > 128 || !SExt0Type.isPow2VectorType() ||
      SExt0Type.getVectorNumElements() == 1)
    return SDValue();

  SDLoc DL(N);
  SDValue V0 = SExt0.getOperand(0);
  SDValue V1 = SExt1.getOperand(0);

  // Ensure input vectors are extended to legal types
  if (SExt0Type.getFixedSizeInBits() < 64) {
    unsigned VecNumElements = SExt0Type.getVectorNumElements();
    EVT ExtVecVT = MVT::getVectorVT(MVT::getIntegerVT(64 / VecNumElements),
                                    VecNumElements);
    V0 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVecVT, V0);
    V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVecVT, V1);
  }

  SDValue SQDMULH =
      DAG.getNode(AArch64ISD::SQDMULH, DL, V0.getValueType(), V0, V1);

  return DAG.getNode(ISD::SIGN_EXTEND, DL, DestVT, SQDMULH);
}

static SDValue performSMINCombine(SDNode *N, SelectionDAG &DAG) {
  if (SDValue V = trySQDMULHCombine(N, DAG)) {
    return V;
  }

  return SDValue();
}

static SDValue performTruncateCombine(SDNode *N, SelectionDAG &DAG,
                                      TargetLowering::DAGCombinerInfo &DCI) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue N0 = N->getOperand(0);
  if (VT.isFixedLengthVector() && VT.is64BitVector() && N0.hasOneUse() &&
      N0.getOpcode() == AArch64ISD::DUP) {
    SDValue Op = N0.getOperand(0);
    if (VT.getScalarType() == MVT::i32 &&
        N0.getOperand(0).getValueType().getScalarType() == MVT::i64)
      Op = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Op);
    return DAG.getNode(N0.getOpcode(), DL, VT, Op);
  }

  // Performing the following combine produces a preferable form for ISEL.
  // i32 (trunc (extract Vi64, idx)) -> i32 (extract (nvcast Vi32), idx*2))
  if (DCI.isAfterLegalizeDAG() && N0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
      N0.hasOneUse()) {
    SDValue Op = N0.getOperand(0);
    SDValue ExtractIndexNode = N0.getOperand(1);
    if (!isa<ConstantSDNode>(ExtractIndexNode))
      return SDValue();

    // For a legal DAG, EXTRACT_VECTOR_ELT can only have produced an i32 or i64.
    // So we can only expect: i32 (trunc (i64 (extract Vi64, idx))).
    assert((VT == MVT::i32 && N0.getValueType() == MVT::i64) &&
           "Unexpected legalisation result!");

    EVT SrcVectorType = Op.getValueType();
    // We also assume that SrcVectorType cannot be a V64 (see
    // LowerEXTRACT_VECTOR_ELT).
    assert((SrcVectorType == MVT::v2i64 || SrcVectorType == MVT::nxv2i64) &&
           "Unexpected legalisation result!");

    unsigned ExtractIndex =
        cast<ConstantSDNode>(ExtractIndexNode)->getZExtValue();
    MVT CastVT = SrcVectorType.isScalableVector() ? MVT::nxv4i32 : MVT::v4i32;

    Op = DAG.getNode(AArch64ISD::NVCAST, DL, CastVT, Op);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op,
                       DAG.getVectorIdxConstant(ExtractIndex * 2, DL));
  }

  return SDValue();
}

// Check an node is an extend or shift operand
static bool isExtendOrShiftOperand(SDValue N) {
  unsigned Opcode = N.getOpcode();
  if (ISD::isExtOpcode(Opcode) || Opcode == ISD::SIGN_EXTEND_INREG) {
    EVT SrcVT;
    if (Opcode == ISD::SIGN_EXTEND_INREG)
      SrcVT = cast<VTSDNode>(N.getOperand(1))->getVT();
    else
      SrcVT = N.getOperand(0).getValueType();

    return SrcVT == MVT::i32 || SrcVT == MVT::i16 || SrcVT == MVT::i8;
  } else if (Opcode == ISD::AND) {
    ConstantSDNode *CSD = dyn_cast<ConstantSDNode>(N.getOperand(1));
    if (!CSD)
      return false;
    uint64_t AndMask = CSD->getZExtValue();
    return AndMask == 0xff || AndMask == 0xffff || AndMask == 0xffffffff;
  } else if (Opcode == ISD::SHL || Opcode == ISD::SRL || Opcode == ISD::SRA) {
    return isa<ConstantSDNode>(N.getOperand(1));
  }

  return false;
}

// (N - Y) + Z --> (Z - Y) + N
// when N is an extend or shift operand
static SDValue performAddCombineSubShift(SDNode *N, SDValue SUB, SDValue Z,
                                         SelectionDAG &DAG) {
  auto IsOneUseExtend = [](SDValue N) {
    return N.hasOneUse() && isExtendOrShiftOperand(N);
  };

  // DAGCombiner will revert the combination when Z is constant cause
  // dead loop. So don't enable the combination when Z is constant.
  // If Z is one use shift C, we also can't do the optimization.
  // It will falling to self infinite loop.
  if (isa<ConstantSDNode>(Z) || IsOneUseExtend(Z))
    return SDValue();

  if (SUB.getOpcode() != ISD::SUB || !SUB.hasOneUse())
    return SDValue();

  SDValue Shift = SUB.getOperand(0);
  if (!IsOneUseExtend(Shift))
    return SDValue();

  SDLoc DL(N);
  EVT VT = N->getValueType(0);

  SDValue Y = SUB.getOperand(1);
  SDValue NewSub = DAG.getNode(ISD::SUB, DL, VT, Z, Y);
  return DAG.getNode(ISD::ADD, DL, VT, NewSub, Shift);
}

static SDValue performAddCombineForShiftedOperands(SDNode *N,
                                                   SelectionDAG &DAG) {
  // NOTE: Swapping LHS and RHS is not done for SUB, since SUB is not
  // commutative.
  if (N->getOpcode() != ISD::ADD)
    return SDValue();

  // Bail out when value type is not one of {i32, i64}, since AArch64 ADD with
  // shifted register is only available for i32 and i64.
  EVT VT = N->getValueType(0);
  if (VT != MVT::i32 && VT != MVT::i64)
    return SDValue();

  SDLoc DL(N);
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);

  if (SDValue Val = performAddCombineSubShift(N, LHS, RHS, DAG))
    return Val;
  if (SDValue Val = performAddCombineSubShift(N, RHS, LHS, DAG))
    return Val;

  uint64_t LHSImm = 0, RHSImm = 0;
  // If both operand are shifted by imm and shift amount is not greater than 4
  // for one operand, swap LHS and RHS to put operand with smaller shift amount
  // on RHS.
  //
  // On many AArch64 processors (Cortex A78, Neoverse N1/N2/V1, etc), ADD with
  // LSL shift (shift <= 4) has smaller latency and larger throughput than ADD
  // with LSL (shift > 4). For the rest of processors, this is no-op for
  // performance or correctness.
  if (isOpcWithIntImmediate(LHS.getNode(), ISD::SHL, LHSImm) &&
      isOpcWithIntImmediate(RHS.getNode(), ISD::SHL, RHSImm) && LHSImm <= 4 &&
      RHSImm > 4 && LHS.hasOneUse())
    return DAG.getNode(ISD::ADD, DL, VT, RHS, LHS);

  return SDValue();
}

// The mid end will reassociate sub(sub(x, m1), m2) to sub(x, add(m1, m2))
// This reassociates it back to allow the creation of more mls instructions.
static SDValue performSubAddMULCombine(SDNode *N, SelectionDAG &DAG) {
  if (N->getOpcode() != ISD::SUB)
    return SDValue();

  SDValue Add = N->getOperand(1);
  SDValue X = N->getOperand(0);
  if (Add.getOpcode() != ISD::ADD)
    return SDValue();

  if (!Add.hasOneUse())
    return SDValue();
  if (DAG.isConstantIntBuildVectorOrConstantInt(X))
    return SDValue();

  SDValue M1 = Add.getOperand(0);
  SDValue M2 = Add.getOperand(1);
  if (M1.getOpcode() != ISD::MUL && M1.getOpcode() != AArch64ISD::SMULL &&
      M1.getOpcode() != AArch64ISD::UMULL)
    return SDValue();
  if (M2.getOpcode() != ISD::MUL && M2.getOpcode() != AArch64ISD::SMULL &&
      M2.getOpcode() != AArch64ISD::UMULL)
    return SDValue();

  EVT VT = N->getValueType(0);
  SDValue Sub = DAG.getNode(ISD::SUB, SDLoc(N), VT, X, M1);
  return DAG.getNode(ISD::SUB, SDLoc(N), VT, Sub, M2);
}

// Combine into mla/mls.
// This works on the patterns of:
//   add v1, (mul v2, v3)
//   sub v1, (mul v2, v3)
// for vectors of type <1 x i64> and <2 x i64> when SVE is available.
// It will transform the add/sub to a scalable version, so that we can
// make use of SVE's MLA/MLS that will be generated for that pattern
static SDValue
performSVEMulAddSubCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
  SelectionDAG &DAG = DCI.DAG;
  // Make sure that the types are legal
  if (!DCI.isAfterLegalizeDAG())
    return SDValue();
  // Before using SVE's features, check first if it's available.
  if (!DAG.getSubtarget<AArch64Subtarget>().hasSVE())
    return SDValue();

  if (N->getOpcode() != ISD::ADD && N->getOpcode() != ISD::SUB)
    return SDValue();

  if (!N->getValueType(0).isFixedLengthVector())
    return SDValue();

  auto performOpt = [&DAG, &N](SDValue Op0, SDValue Op1) -> SDValue {
    if (Op1.getOpcode() != ISD::EXTRACT_SUBVECTOR)
      return SDValue();

    if (!cast<ConstantSDNode>(Op1->getOperand(1))->isZero())
      return SDValue();

    SDValue MulValue = Op1->getOperand(0);
    if (MulValue.getOpcode() != AArch64ISD::MUL_PRED)
      return SDValue();

    if (!Op1.hasOneUse() || !MulValue.hasOneUse())
      return SDValue();

    EVT ScalableVT = MulValue.getValueType();
    if (!ScalableVT.isScalableVector())
      return SDValue();

    SDValue ScaledOp = convertToScalableVector(DAG, ScalableVT, Op0);
    SDValue NewValue =
        DAG.getNode(N->getOpcode(), SDLoc(N), ScalableVT, {ScaledOp, MulValue});
    return convertFromScalableVector(DAG, N->getValueType(0), NewValue);
  };

  if (SDValue res = performOpt(N->getOperand(0), N->getOperand(1)))
    return res;
  else if (N->getOpcode() == ISD::ADD)
    return performOpt(N->getOperand(1), N->getOperand(0));

  return SDValue();
}

// Given a i64 add from a v1i64 extract, convert to a neon v1i64 add. This can
// help, for example, to produce ssra from sshr+add.
static SDValue performAddSubIntoVectorOp(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if (VT != MVT::i64 ||
      DAG.getTargetLoweringInfo().isOperationExpand(N->getOpcode(), MVT::v1i64))
    return SDValue();
  SDValue Op0 = N->getOperand(0);
  SDValue Op1 = N->getOperand(1);

  // At least one of the operands should be an extract, and the other should be
  // something that is easy to convert to v1i64 type (in this case a load).
  if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT &&
      Op0.getOpcode() != ISD::LOAD)
    return SDValue();
  if (Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT &&
      Op1.getOpcode() != ISD::LOAD)
    return SDValue();

  SDLoc DL(N);
  if (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
      Op0.getOperand(0).getValueType() == MVT::v1i64) {
    Op0 = Op0.getOperand(0);
    Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v1i64, Op1);
  } else if (Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
             Op1.getOperand(0).getValueType() == MVT::v1i64) {
    Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v1i64, Op0);
    Op1 = Op1.getOperand(0);
  } else
    return SDValue();

  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64,
                     DAG.getNode(N->getOpcode(), DL, MVT::v1i64, Op0, Op1),
                     DAG.getConstant(0, DL, MVT::i64));
}

static bool isLoadOrMultipleLoads(SDValue B, SmallVector<LoadSDNode *> &Loads) {
  SDValue BV = peekThroughOneUseBitcasts(B);
  if (!BV->hasOneUse())
    return false;
  if (auto *Ld = dyn_cast<LoadSDNode>(BV)) {
    if (!Ld || !Ld->isSimple())
      return false;
    Loads.push_back(Ld);
    return true;
  } else if (BV.getOpcode() == ISD::BUILD_VECTOR ||
             BV.getOpcode() == ISD::CONCAT_VECTORS) {
    for (unsigned Op = 0; Op < BV.getNumOperands(); Op++) {
      auto *Ld = dyn_cast<LoadSDNode>(BV.getOperand(Op));
      if (!Ld || !Ld->isSimple() || !BV.getOperand(Op).hasOneUse())
        return false;
      Loads.push_back(Ld);
    }
    return true;
  } else if (B.getOpcode() == ISD::VECTOR_SHUFFLE) {
    // Try to find a tree of shuffles and concats from how IR shuffles of loads
    // are lowered. Note that this only comes up because we do not always visit
    // operands before uses. After that is fixed this can be removed and in the
    // meantime this is fairly specific to the lowering we expect from IR.
    // t46: v16i8 = vector_shuffle<0,1,2,3,4,5,6,7,8,9,10,11,16,17,18,19> t44, t45
    //   t44: v16i8 = vector_shuffle<0,1,2,3,4,5,6,7,16,17,18,19,u,u,u,u> t42, t43
    //     t42: v16i8 = concat_vectors t40, t36, undef:v4i8, undef:v4i8
    //       t40: v4i8,ch = load<(load (s32) from %ir.17)> t0, t22, undef:i64
    //       t36: v4i8,ch = load<(load (s32) from %ir.13)> t0, t18, undef:i64
    //     t43: v16i8 = concat_vectors t32, undef:v4i8, undef:v4i8, undef:v4i8
    //       t32: v4i8,ch = load<(load (s32) from %ir.9)> t0, t14, undef:i64
    //   t45: v16i8 = concat_vectors t28, undef:v4i8, undef:v4i8, undef:v4i8
    //     t28: v4i8,ch = load<(load (s32) from %ir.0)> t0, t2, undef:i64
    if (B.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE ||
        B.getOperand(0).getOperand(0).getOpcode() != ISD::CONCAT_VECTORS ||
        B.getOperand(0).getOperand(1).getOpcode() != ISD::CONCAT_VECTORS ||
        B.getOperand(1).getOpcode() != ISD::CONCAT_VECTORS ||
        B.getOperand(1).getNumOperands() != 4)
      return false;
    auto SV1 = cast<ShuffleVectorSDNode>(B);
    auto SV2 = cast<ShuffleVectorSDNode>(B.getOperand(0));
    int NumElts = B.getValueType().getVectorNumElements();
    int NumSubElts = NumElts / 4;
    for (int I = 0; I < NumSubElts; I++) {
      // <0,1,2,3,4,5,6,7,8,9,10,11,16,17,18,19>
      if (SV1->getMaskElt(I) != I ||
          SV1->getMaskElt(I + NumSubElts) != I + NumSubElts ||
          SV1->getMaskElt(I + NumSubElts * 2) != I + NumSubElts * 2 ||
          SV1->getMaskElt(I + NumSubElts * 3) != I + NumElts)
        return false;
      // <0,1,2,3,4,5,6,7,16,17,18,19,u,u,u,u>
      if (SV2->getMaskElt(I) != I ||
          SV2->getMaskElt(I + NumSubElts) != I + NumSubElts ||
          SV2->getMaskElt(I + NumSubElts * 2) != I + NumElts)
        return false;
    }
    auto *Ld0 = dyn_cast<LoadSDNode>(SV2->getOperand(0).getOperand(0));
    auto *Ld1 = dyn_cast<LoadSDNode>(SV2->getOperand(0).getOperand(1));
    auto *Ld2 = dyn_cast<LoadSDNode>(SV2->getOperand(1).getOperand(0));
    auto *Ld3 = dyn_cast<LoadSDNode>(B.getOperand(1).getOperand(0));
    if (!Ld0 || !Ld1 || !Ld2 || !Ld3 || !Ld0->isSimple() || !Ld1->isSimple() ||
        !Ld2->isSimple() || !Ld3->isSimple())
      return false;
    Loads.push_back(Ld0);
    Loads.push_back(Ld1);
    Loads.push_back(Ld2);
    Loads.push_back(Ld3);
    return true;
  }
  return false;
}

static bool areLoadedOffsetButOtherwiseSame(SDValue Op0, SDValue Op1,
                                            SelectionDAG &DAG,
                                            unsigned &NumSubLoads) {
  if (!Op0.hasOneUse() || !Op1.hasOneUse())
    return false;

  SmallVector<LoadSDNode *> Loads0, Loads1;
  if (isLoadOrMultipleLoads(Op0, Loads0) &&
      isLoadOrMultipleLoads(Op1, Loads1)) {
    if (NumSubLoads && Loads0.size() != NumSubLoads)
      return false;
    NumSubLoads = Loads0.size();
    return Loads0.size() == Loads1.size() &&
           all_of(zip(Loads0, Loads1), [&DAG](auto L) {
             unsigned Size = get<0>(L)->getValueType(0).getSizeInBits();
             return Size == get<1>(L)->getValueType(0).getSizeInBits() &&
                    DAG.areNonVolatileConsecutiveLoads(get<1>(L), get<0>(L),
                                                       Size / 8, 1);
           });
  }

  if (Op0.getOpcode() != Op1.getOpcode())
    return false;

  switch (Op0.getOpcode()) {
  case ISD::ADD:
  case ISD::SUB:
    return areLoadedOffsetButOtherwiseSame(Op0.getOperand(0), Op1.getOperand(0),
                                           DAG, NumSubLoads) &&
           areLoadedOffsetButOtherwiseSame(Op0.getOperand(1), Op1.getOperand(1),
                                           DAG, NumSubLoads);
  case ISD::SIGN_EXTEND:
  case ISD::ANY_EXTEND:
  case ISD::ZERO_EXTEND:
    EVT XVT = Op0.getOperand(0).getValueType();
    if (XVT.getScalarSizeInBits() != 8 && XVT.getScalarSizeInBits() != 16 &&
        XVT.getScalarSizeInBits() != 32)
      return false;
    return areLoadedOffsetButOtherwiseSame(Op0.getOperand(0), Op1.getOperand(0),
                                           DAG, NumSubLoads);
  }
  return false;
}

// This method attempts to fold trees of add(ext(load p), shl(ext(load p+4))
// into a single load of twice the size, that we extract the bottom part and top
// part so that the shl can use a shll2 instruction. The two loads in that
// example can also be larger trees of instructions, which are identical except
// for the leaves which are all loads offset from the LHS, including
// buildvectors of multiple loads. For example the RHS tree could be
// sub(zext(buildvec(load p+4, load q+4)), zext(buildvec(load r+4, load s+4)))
// Whilst it can be common for the larger loads to replace LDP instructions
// (which doesn't gain anything on it's own), the larger loads can help create
// more efficient code, and in buildvectors prevent the need for ld1 lane
// inserts which can be slower than normal loads.
static SDValue performExtBinopLoadFold(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if (!VT.isFixedLengthVector() ||
      (VT.getScalarSizeInBits() != 16 && VT.getScalarSizeInBits() != 32 &&
       VT.getScalarSizeInBits() != 64))
    return SDValue();

  SDValue Other = N->getOperand(0);
  SDValue Shift = N->getOperand(1);
  if (Shift.getOpcode() != ISD::SHL && N->getOpcode() != ISD::SUB)
    std::swap(Shift, Other);
  APInt ShiftAmt;
  if (Shift.getOpcode() != ISD::SHL || !Shift.hasOneUse() ||
      !ISD::isConstantSplatVector(Shift.getOperand(1).getNode(), ShiftAmt))
    return SDValue();

  if (!ISD::isExtOpcode(Shift.getOperand(0).getOpcode()) ||
      !ISD::isExtOpcode(Other.getOpcode()) ||
      Shift.getOperand(0).getOperand(0).getValueType() !=
          Other.getOperand(0).getValueType() ||
      !Other.hasOneUse() || !Shift.getOperand(0).hasOneUse())
    return SDValue();

  SDValue Op0 = Other.getOperand(0);
  SDValue Op1 = Shift.getOperand(0).getOperand(0);

  unsigned NumSubLoads = 0;
  if (!areLoadedOffsetButOtherwiseSame(Op0, Op1, DAG, NumSubLoads))
    return SDValue();

  // Attempt to rule out some unprofitable cases using heuristics (some working
  // around suboptimal code generation), notably if the extend not be able to
  // use ushll2 instructions as the types are not large enough. Otherwise zip's
  // will need to be created which can increase the instruction count.
  unsigned NumElts = Op0.getValueType().getVectorNumElements();
  unsigned NumSubElts = NumElts / NumSubLoads;
  if (NumSubElts * VT.getScalarSizeInBits() < 128 ||
      (Other.getOpcode() != Shift.getOperand(0).getOpcode() &&
       Op0.getValueType().getSizeInBits() < 128 &&
       !DAG.getTargetLoweringInfo().isTypeLegal(Op0.getValueType())))
    return SDValue();

  // Recreate the tree with the new combined loads.
  std::function<SDValue(SDValue, SDValue, SelectionDAG &)> GenCombinedTree =
      [&GenCombinedTree](SDValue Op0, SDValue Op1, SelectionDAG &DAG) {
        EVT DVT =
            Op0.getValueType().getDoubleNumVectorElementsVT(*DAG.getContext());

        SmallVector<LoadSDNode *> Loads0, Loads1;
        if (isLoadOrMultipleLoads(Op0, Loads0) &&
            isLoadOrMultipleLoads(Op1, Loads1)) {
          EVT LoadVT = EVT::getVectorVT(
              *DAG.getContext(), Op0.getValueType().getScalarType(),
              Op0.getValueType().getVectorNumElements() / Loads0.size());
          EVT DLoadVT = LoadVT.getDoubleNumVectorElementsVT(*DAG.getContext());

          SmallVector<SDValue> NewLoads;
          for (const auto &[L0, L1] : zip(Loads0, Loads1)) {
            SDValue Load = DAG.getLoad(DLoadVT, SDLoc(L0), L0->getChain(),
                                       L0->getBasePtr(), L0->getPointerInfo(),
                                       L0->getBaseAlign());
            DAG.makeEquivalentMemoryOrdering(L0, Load.getValue(1));
            DAG.makeEquivalentMemoryOrdering(L1, Load.getValue(1));
            NewLoads.push_back(Load);
          }
          return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(Op0), DVT, NewLoads);
        }

        SmallVector<SDValue> Ops;
        for (const auto &[O0, O1] : zip(Op0->op_values(), Op1->op_values()))
          Ops.push_back(GenCombinedTree(O0, O1, DAG));
        return DAG.getNode(Op0.getOpcode(), SDLoc(Op0), DVT, Ops);
      };
  SDValue NewOp = GenCombinedTree(Op0, Op1, DAG);

  SmallVector<int> LowMask(NumElts, 0), HighMask(NumElts, 0);
  int Hi = NumSubElts, Lo = 0;
  for (unsigned i = 0; i < NumSubLoads; i++) {
    for (unsigned j = 0; j < NumSubElts; j++) {
      LowMask[i * NumSubElts + j] = Lo++;
      HighMask[i * NumSubElts + j] = Hi++;
    }
    Lo += NumSubElts;
    Hi += NumSubElts;
  }
  SDLoc DL(N);
  SDValue Ext0, Ext1;
  // Extract the top and bottom lanes, then extend the result. Possibly extend
  // the result then extract the lanes if the two operands match as it produces
  // slightly smaller code.
  if (Other.getOpcode() != Shift.getOperand(0).getOpcode()) {
    SDValue SubL = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, Op0.getValueType(),
                               NewOp, DAG.getConstant(0, DL, MVT::i64));
    SDValue SubH =
        DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, Op0.getValueType(), NewOp,
                    DAG.getConstant(NumSubElts * NumSubLoads, DL, MVT::i64));
    SDValue Extr0 =
        DAG.getVectorShuffle(Op0.getValueType(), DL, SubL, SubH, LowMask);
    SDValue Extr1 =
        DAG.getVectorShuffle(Op0.getValueType(), DL, SubL, SubH, HighMask);
    Ext0 = DAG.getNode(Other.getOpcode(), DL, VT, Extr0);
    Ext1 = DAG.getNode(Shift.getOperand(0).getOpcode(), DL, VT, Extr1);
  } else {
    EVT DVT = VT.getDoubleNumVectorElementsVT(*DAG.getContext());
    SDValue Ext = DAG.getNode(Other.getOpcode(), DL, DVT, NewOp);
    SDValue SubL = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Ext,
                               DAG.getConstant(0, DL, MVT::i64));
    SDValue SubH =
        DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Ext,
                    DAG.getConstant(NumSubElts * NumSubLoads, DL, MVT::i64));
    Ext0 = DAG.getVectorShuffle(VT, DL, SubL, SubH, LowMask);
    Ext1 = DAG.getVectorShuffle(VT, DL, SubL, SubH, HighMask);
  }
  SDValue NShift =
      DAG.getNode(Shift.getOpcode(), DL, VT, Ext1, Shift.getOperand(1));
  return DAG.getNode(N->getOpcode(), DL, VT, Ext0, NShift);
}

static SDValue performAddSubCombine(SDNode *N,
                                    TargetLowering::DAGCombinerInfo &DCI) {
  // Try to change sum of two reductions.
  if (SDValue Val = performAddUADDVCombine(N, DCI.DAG))
    return Val;
  if (SDValue Val = performAddDotCombine(N, DCI.DAG))
    return Val;
  if (SDValue Val = performAddCSelIntoCSinc(N, DCI.DAG))
    return Val;
  if (SDValue Val = performNegCSelCombine(N, DCI.DAG))
    return Val;
  if (SDValue Val = performVectorExtCombine(N, DCI.DAG))
    return Val;
  if (SDValue Val = performAddCombineForShiftedOperands(N, DCI.DAG))
    return Val;
  if (SDValue Val = performSubAddMULCombine(N, DCI.DAG))
    return Val;
  if (SDValue Val = performSVEMulAddSubCombine(N, DCI))
    return Val;
  if (SDValue Val = performAddSubIntoVectorOp(N, DCI.DAG))
    return Val;

  if (SDValue Val = performExtBinopLoadFold(N, DCI.DAG))
    return Val;

  return performAddSubLongCombine(N, DCI);
}

// Massage DAGs which we can use the high-half "long" operations on into
// something isel will recognize better. E.g.
//
// (aarch64_neon_umull (extract_high vec) (dupv64 scalar)) -->
//   (aarch64_neon_umull (extract_high (v2i64 vec)))
//                     (extract_high (v2i64 (dup128 scalar)))))
//
static SDValue tryCombineLongOpWithDup(unsigned IID, SDNode *N,
                                       TargetLowering::DAGCombinerInfo &DCI,
                                       SelectionDAG &DAG) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  SDValue LHS = N->getOperand((IID == Intrinsic::not_intrinsic) ? 0 : 1);
  SDValue RHS = N->getOperand((IID == Intrinsic::not_intrinsic) ? 1 : 2);
  assert(LHS.getValueType().is64BitVector() &&
         RHS.getValueType().is64BitVector() &&
         "unexpected shape for long operation");

  // Either node could be a DUP, but it's not worth doing both of them (you'd
  // just as well use the non-high version) so look for a corresponding extract
  // operation on the other "wing".
  if (isEssentiallyExtractHighSubvector(LHS)) {
    RHS = tryExtendDUPToExtractHigh(RHS, DAG);
    if (!RHS.getNode())
      return SDValue();
  } else if (isEssentiallyExtractHighSubvector(RHS)) {
    LHS = tryExtendDUPToExtractHigh(LHS, DAG);
    if (!LHS.getNode())
      return SDValue();
  } else
    return SDValue();

  if (IID == Intrinsic::not_intrinsic)
    return DAG.getNode(N->getOpcode(), SDLoc(N), N->getValueType(0), LHS, RHS);

  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), N->getValueType(0),
                     N->getOperand(0), LHS, RHS);
}

static SDValue tryCombineShiftImm(unsigned IID, SDNode *N, SelectionDAG &DAG) {
  MVT ElemTy = N->getSimpleValueType(0).getScalarType();
  unsigned ElemBits = ElemTy.getSizeInBits();

  int64_t ShiftAmount;
  if (BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(N->getOperand(2))) {
    APInt SplatValue, SplatUndef;
    unsigned SplatBitSize;
    bool HasAnyUndefs;
    if (!BVN->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
                              HasAnyUndefs, ElemBits) ||
        SplatBitSize != ElemBits)
      return SDValue();

    ShiftAmount = SplatValue.getSExtValue();
  } else if (ConstantSDNode *CVN = dyn_cast<ConstantSDNode>(N->getOperand(2))) {
    ShiftAmount = CVN->getSExtValue();
  } else
    return SDValue();

  // If the shift amount is zero, remove the shift intrinsic.
  if (ShiftAmount == 0 && IID != Intrinsic::aarch64_neon_sqshlu)
    return N->getOperand(1);

  unsigned Opcode;
  bool IsRightShift;
  switch (IID) {
  default:
    llvm_unreachable("Unknown shift intrinsic");
  case Intrinsic::aarch64_neon_sqshl:
    Opcode = AArch64ISD::SQSHL_I;
    IsRightShift = false;
    break;
  case Intrinsic::aarch64_neon_uqshl:
    Opcode = AArch64ISD::UQSHL_I;
    IsRightShift = false;
    break;
  case Intrinsic::aarch64_neon_srshl:
    Opcode = AArch64ISD::SRSHR_I;
    IsRightShift = true;
    break;
  case Intrinsic::aarch64_neon_urshl:
    Opcode = AArch64ISD::URSHR_I;
    IsRightShift = true;
    break;
  case Intrinsic::aarch64_neon_sqshlu:
    Opcode = AArch64ISD::SQSHLU_I;
    IsRightShift = false;
    break;
  case Intrinsic::aarch64_neon_sshl:
  case Intrinsic::aarch64_neon_ushl:
    // For positive shift amounts we can use SHL, as ushl/sshl perform a regular
    // left shift for positive shift amounts. For negative shifts we can use a
    // VASHR/VLSHR as appropriate.
    if (ShiftAmount < 0) {
      Opcode = IID == Intrinsic::aarch64_neon_sshl ? AArch64ISD::VASHR
                                                   : AArch64ISD::VLSHR;
      ShiftAmount = -ShiftAmount;
    } else
      Opcode = AArch64ISD::VSHL;
    IsRightShift = false;
    break;
  }

  EVT VT = N->getValueType(0);
  SDValue Op = N->getOperand(1);
  SDLoc DL(N);
  if (VT == MVT::i64) {
    Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v1i64, Op);
    VT = MVT::v1i64;
  }

  if (IsRightShift && ShiftAmount <= -1 && ShiftAmount >= -(int)ElemBits) {
    Op = DAG.getNode(Opcode, DL, VT, Op,
                     DAG.getSignedConstant(-ShiftAmount, DL, MVT::i32));
    if (N->getValueType(0) == MVT::i64)
      Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64, Op,
                       DAG.getConstant(0, DL, MVT::i64));
    return Op;
  } else if (!IsRightShift && ShiftAmount >= 0 && ShiftAmount < ElemBits) {
    Op = DAG.getNode(Opcode, DL, VT, Op,
                     DAG.getConstant(ShiftAmount, DL, MVT::i32));
    if (N->getValueType(0) == MVT::i64)
      Op = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64, Op,
                       DAG.getConstant(0, DL, MVT::i64));
    return Op;
  }

  return SDValue();
}

// The CRC32[BH] instructions ignore the high bits of their data operand. Since
// the intrinsics must be legal and take an i32, this means there's almost
// certainly going to be a zext in the DAG which we can eliminate.
static SDValue tryCombineCRC32(unsigned Mask, SDNode *N, SelectionDAG &DAG) {
  SDValue AndN = N->getOperand(2);
  if (AndN.getOpcode() != ISD::AND)
    return SDValue();

  ConstantSDNode *CMask = dyn_cast<ConstantSDNode>(AndN.getOperand(1));
  if (!CMask || CMask->getZExtValue() != Mask)
    return SDValue();

  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SDLoc(N), MVT::i32,
                     N->getOperand(0), N->getOperand(1), AndN.getOperand(0));
}

static SDValue combineAcrossLanesIntrinsic(unsigned Opc, SDNode *N,
                                           SelectionDAG &DAG) {
  SDLoc DL(N);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0),
                     DAG.getNode(Opc, DL, N->getOperand(1).getSimpleValueType(),
                                 N->getOperand(1)),
                     DAG.getConstant(0, DL, MVT::i64));
}

static SDValue LowerSVEIntrinsicIndex(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue Op1 = N->getOperand(1);
  SDValue Op2 = N->getOperand(2);
  EVT ScalarTy = Op2.getValueType();
  if ((ScalarTy == MVT::i8) || (ScalarTy == MVT::i16))
    ScalarTy = MVT::i32;

  // Lower index_vector(base, step) to mul(step step_vector(1)) + splat(base).
  SDValue StepVector = DAG.getStepVector(DL, N->getValueType(0));
  SDValue Step = DAG.getNode(ISD::SPLAT_VECTOR, DL, N->getValueType(0), Op2);
  SDValue Mul = DAG.getNode(ISD::MUL, DL, N->getValueType(0), StepVector, Step);
  SDValue Base = DAG.getNode(ISD::SPLAT_VECTOR, DL, N->getValueType(0), Op1);
  return DAG.getNode(ISD::ADD, DL, N->getValueType(0), Mul, Base);
}

static SDValue LowerSVEIntrinsicDUP(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue Scalar = N->getOperand(3);
  EVT ScalarTy = Scalar.getValueType();

  if ((ScalarTy == MVT::i8) || (ScalarTy == MVT::i16))
    Scalar = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Scalar);

  SDValue Passthru = N->getOperand(1);
  SDValue Pred = N->getOperand(2);
  return DAG.getNode(AArch64ISD::DUP_MERGE_PASSTHRU, DL, N->getValueType(0),
                     Pred, Scalar, Passthru);
}

static SDValue LowerSVEIntrinsicEXT(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  LLVMContext &Ctx = *DAG.getContext();
  EVT VT = N->getValueType(0);

  assert(VT.isScalableVector() && "Expected a scalable vector.");

  // Current lowering only supports the SVE-ACLE types.
  if (VT.getSizeInBits().getKnownMinValue() != AArch64::SVEBitsPerBlock)
    return SDValue();

  unsigned ElemSize = VT.getVectorElementType().getSizeInBits() / 8;
  unsigned ByteSize = VT.getSizeInBits().getKnownMinValue() / 8;
  EVT ByteVT =
      EVT::getVectorVT(Ctx, MVT::i8, ElementCount::getScalable(ByteSize));

  // Convert everything to the domain of EXT (i.e bytes).
  SDValue Op0 = DAG.getNode(ISD::BITCAST, DL, ByteVT, N->getOperand(1));
  SDValue Op1 = DAG.getNode(ISD::BITCAST, DL, ByteVT, N->getOperand(2));
  SDValue Op2 = DAG.getNode(ISD::MUL, DL, MVT::i32, N->getOperand(3),
                            DAG.getConstant(ElemSize, DL, MVT::i32));

  SDValue EXT = DAG.getNode(AArch64ISD::EXT, DL, ByteVT, Op0, Op1, Op2);
  return DAG.getNode(ISD::BITCAST, DL, VT, EXT);
}

static SDValue tryConvertSVEWideCompare(SDNode *N, ISD::CondCode CC,
                                        TargetLowering::DAGCombinerInfo &DCI,
                                        SelectionDAG &DAG) {
  if (DCI.isBeforeLegalize())
    return SDValue();

  SDValue Comparator = N->getOperand(3);
  if (Comparator.getOpcode() == AArch64ISD::DUP ||
      Comparator.getOpcode() == ISD::SPLAT_VECTOR) {
    unsigned IID = getIntrinsicID(N);
    EVT VT = N->getValueType(0);
    EVT CmpVT = N->getOperand(2).getValueType();
    SDValue Pred = N->getOperand(1);
    SDValue Imm;
    SDLoc DL(N);

    switch (IID) {
    default:
      llvm_unreachable("Called with wrong intrinsic!");
      break;

    // Signed comparisons
    case Intrinsic::aarch64_sve_cmpeq_wide:
    case Intrinsic::aarch64_sve_cmpne_wide:
    case Intrinsic::aarch64_sve_cmpge_wide:
    case Intrinsic::aarch64_sve_cmpgt_wide:
    case Intrinsic::aarch64_sve_cmplt_wide:
    case Intrinsic::aarch64_sve_cmple_wide: {
      if (auto *CN = dyn_cast<ConstantSDNode>(Comparator.getOperand(0))) {
        int64_t ImmVal = CN->getSExtValue();
        if (ImmVal >= -16 && ImmVal <= 15)
          Imm = DAG.getSignedConstant(ImmVal, DL, MVT::i32);
        else
          return SDValue();
      }
      break;
    }
    // Unsigned comparisons
    case Intrinsic::aarch64_sve_cmphs_wide:
    case Intrinsic::aarch64_sve_cmphi_wide:
    case Intrinsic::aarch64_sve_cmplo_wide:
    case Intrinsic::aarch64_sve_cmpls_wide:  {
      if (auto *CN = dyn_cast<ConstantSDNode>(Comparator.getOperand(0))) {
        uint64_t ImmVal = CN->getZExtValue();
        if (ImmVal <= 127)
          Imm = DAG.getConstant(ImmVal, DL, MVT::i32);
        else
          return SDValue();
      }
      break;
    }
    }

    if (!Imm)
      return SDValue();

    SDValue Splat = DAG.getNode(ISD::SPLAT_VECTOR, DL, CmpVT, Imm);
    return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, DL, VT, Pred,
                       N->getOperand(2), Splat, DAG.getCondCode(CC));
  }

  return SDValue();
}

static SDValue getPTest(SelectionDAG &DAG, EVT VT, SDValue Pg, SDValue Op,
                        AArch64CC::CondCode Cond) {
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();

  SDLoc DL(Op);
  assert(Op.getValueType().isScalableVector() &&
         TLI.isTypeLegal(Op.getValueType()) &&
         "Expected legal scalable vector type!");
  assert(Op.getValueType() == Pg.getValueType() &&
         "Expected same type for PTEST operands");

  // Ensure target specific opcodes are using legal type.
  EVT OutVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
  SDValue TVal = DAG.getConstant(1, DL, OutVT);
  SDValue FVal = DAG.getConstant(0, DL, OutVT);

  // Ensure operands have type nxv16i1.
  if (Op.getValueType() != MVT::nxv16i1) {
    if ((Cond == AArch64CC::ANY_ACTIVE || Cond == AArch64CC::NONE_ACTIVE) &&
        isZeroingInactiveLanes(Op))
      Pg = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, MVT::nxv16i1, Pg);
    else
      Pg = getSVEPredicateBitCast(MVT::nxv16i1, Pg, DAG);
    Op = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, MVT::nxv16i1, Op);
  }

  // Set condition code (CC) flags.
  SDValue Test = DAG.getNode(
      Cond == AArch64CC::ANY_ACTIVE ? AArch64ISD::PTEST_ANY : AArch64ISD::PTEST,
      DL, MVT::i32, Pg, Op);

  // Convert CC to integer based on requested condition.
  // NOTE: Cond is inverted to promote CSEL's removal when it feeds a compare.
  SDValue CC = DAG.getConstant(getInvertedCondCode(Cond), DL, MVT::i32);
  SDValue Res = DAG.getNode(AArch64ISD::CSEL, DL, OutVT, FVal, TVal, CC, Test);
  return DAG.getZExtOrTrunc(Res, DL, VT);
}

static SDValue combineSVEReductionInt(SDNode *N, unsigned Opc,
                                      SelectionDAG &DAG) {
  SDLoc DL(N);

  SDValue Pred = N->getOperand(1);
  SDValue VecToReduce = N->getOperand(2);

  // NOTE: The integer reduction's result type is not always linked to the
  // operand's element type so we construct it from the intrinsic's result type.
  EVT ReduceVT = getPackedSVEVectorVT(N->getValueType(0));
  SDValue Reduce = DAG.getNode(Opc, DL, ReduceVT, Pred, VecToReduce);

  // SVE reductions set the whole vector register with the first element
  // containing the reduction result, which we'll now extract.
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0), Reduce,
                     Zero);
}

static SDValue combineSVEReductionFP(SDNode *N, unsigned Opc,
                                     SelectionDAG &DAG) {
  SDLoc DL(N);

  SDValue Pred = N->getOperand(1);
  SDValue VecToReduce = N->getOperand(2);

  EVT ReduceVT = VecToReduce.getValueType();
  SDValue Reduce = DAG.getNode(Opc, DL, ReduceVT, Pred, VecToReduce);

  // SVE reductions set the whole vector register with the first element
  // containing the reduction result, which we'll now extract.
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0), Reduce,
                     Zero);
}

static SDValue combineSVEReductionOrderedFP(SDNode *N, unsigned Opc,
                                            SelectionDAG &DAG) {
  SDLoc DL(N);

  SDValue Pred = N->getOperand(1);
  SDValue InitVal = N->getOperand(2);
  SDValue VecToReduce = N->getOperand(3);
  EVT ReduceVT = VecToReduce.getValueType();

  // Ordered reductions use the first lane of the result vector as the
  // reduction's initial value.
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  InitVal = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ReduceVT,
                        DAG.getUNDEF(ReduceVT), InitVal, Zero);

  SDValue Reduce = DAG.getNode(Opc, DL, ReduceVT, Pred, InitVal, VecToReduce);

  // SVE reductions set the whole vector register with the first element
  // containing the reduction result, which we'll now extract.
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, N->getValueType(0), Reduce,
                     Zero);
}

// If a merged operation has no inactive lanes we can relax it to a predicated
// or unpredicated operation, which potentially allows better isel (perhaps
// using immediate forms) or relaxing register reuse requirements.
static SDValue convertMergedOpToPredOp(SDNode *N, unsigned Opc,
                                       SelectionDAG &DAG, bool UnpredOp = false,
                                       bool SwapOperands = false) {
  assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN && "Expected intrinsic!");
  assert(N->getNumOperands() == 4 && "Expected 3 operand intrinsic!");
  SDValue Pg = N->getOperand(1);
  SDValue Op1 = N->getOperand(SwapOperands ? 3 : 2);
  SDValue Op2 = N->getOperand(SwapOperands ? 2 : 3);

  // ISD way to specify an all active predicate.
  if (isAllActivePredicate(DAG, Pg)) {
    if (UnpredOp)
      return DAG.getNode(Opc, SDLoc(N), N->getValueType(0), Op1, Op2);

    return DAG.getNode(Opc, SDLoc(N), N->getValueType(0), Pg, Op1, Op2);
  }

  // FUTURE: SplatVector(true)
  return SDValue();
}

SDValue tryLowerPartialReductionToDot(SDNode *N,
                                      const AArch64Subtarget *Subtarget,
                                      SelectionDAG &DAG) {

  assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
         getIntrinsicID(N) ==
             Intrinsic::experimental_vector_partial_reduce_add &&
         "Expected a partial reduction node");

  bool Scalable = N->getValueType(0).isScalableVector();
  if (Scalable && !Subtarget->isSVEorStreamingSVEAvailable())
    return SDValue();
  if (!Scalable && (!Subtarget->isNeonAvailable() || !Subtarget->hasDotProd()))
    return SDValue();

  SDLoc DL(N);

  SDValue Op2 = N->getOperand(2);
  unsigned Op2Opcode = Op2->getOpcode();
  SDValue MulOpLHS, MulOpRHS;
  bool MulOpLHSIsSigned, MulOpRHSIsSigned;
  if (ISD::isExtOpcode(Op2Opcode)) {
    MulOpLHSIsSigned = MulOpRHSIsSigned = (Op2Opcode == ISD::SIGN_EXTEND);
    MulOpLHS = Op2->getOperand(0);
    MulOpRHS = DAG.getConstant(1, DL, MulOpLHS.getValueType());
  } else if (Op2Opcode == ISD::MUL) {
    SDValue ExtMulOpLHS = Op2->getOperand(0);
    SDValue ExtMulOpRHS = Op2->getOperand(1);

    unsigned ExtMulOpLHSOpcode = ExtMulOpLHS->getOpcode();
    unsigned ExtMulOpRHSOpcode = ExtMulOpRHS->getOpcode();
    if (!ISD::isExtOpcode(ExtMulOpLHSOpcode) ||
        !ISD::isExtOpcode(ExtMulOpRHSOpcode))
      return SDValue();

    MulOpLHSIsSigned = ExtMulOpLHSOpcode == ISD::SIGN_EXTEND;
    MulOpRHSIsSigned = ExtMulOpRHSOpcode == ISD::SIGN_EXTEND;

    MulOpLHS = ExtMulOpLHS->getOperand(0);
    MulOpRHS = ExtMulOpRHS->getOperand(0);

    if (MulOpLHS.getValueType() != MulOpRHS.getValueType())
      return SDValue();
  } else
    return SDValue();

  SDValue Acc = N->getOperand(1);
  EVT ReducedVT = N->getValueType(0);
  EVT MulSrcVT = MulOpLHS.getValueType();

  // Dot products operate on chunks of four elements so there must be four times
  // as many elements in the wide type
  if (!(ReducedVT == MVT::nxv4i64 && MulSrcVT == MVT::nxv16i8) &&
      !(ReducedVT == MVT::nxv4i32 && MulSrcVT == MVT::nxv16i8) &&
      !(ReducedVT == MVT::nxv2i64 && MulSrcVT == MVT::nxv8i16) &&
      !(ReducedVT == MVT::v4i64 && MulSrcVT == MVT::v16i8) &&
      !(ReducedVT == MVT::v4i32 && MulSrcVT == MVT::v16i8) &&
      !(ReducedVT == MVT::v2i32 && MulSrcVT == MVT::v8i8))
    return SDValue();

  // If the extensions are mixed, we should lower it to a usdot instead
  unsigned Opcode = 0;
  if (MulOpLHSIsSigned != MulOpRHSIsSigned) {
    if (!Subtarget->hasMatMulInt8())
      return SDValue();

    bool Scalable = N->getValueType(0).isScalableVT();
    // There's no nxv2i64 version of usdot
    if (Scalable && ReducedVT != MVT::nxv4i32 && ReducedVT != MVT::nxv4i64)
      return SDValue();

    Opcode = AArch64ISD::USDOT;
    // USDOT expects the signed operand to be last
    if (!MulOpRHSIsSigned)
      std::swap(MulOpLHS, MulOpRHS);
  } else
    Opcode = MulOpLHSIsSigned ? AArch64ISD::SDOT : AArch64ISD::UDOT;

  // Partial reduction lowering for (nx)v16i8 to (nx)v4i64 requires an i32 dot
  // product followed by a zero / sign extension
  if ((ReducedVT == MVT::nxv4i64 && MulSrcVT == MVT::nxv16i8) ||
      (ReducedVT == MVT::v4i64 && MulSrcVT == MVT::v16i8)) {
    EVT ReducedVTI32 =
        (ReducedVT.isScalableVector()) ? MVT::nxv4i32 : MVT::v4i32;

    SDValue DotI32 =
        DAG.getNode(Opcode, DL, ReducedVTI32,
                    DAG.getConstant(0, DL, ReducedVTI32), MulOpLHS, MulOpRHS);
    SDValue Extended = DAG.getSExtOrTrunc(DotI32, DL, ReducedVT);
    return DAG.getNode(ISD::ADD, DL, ReducedVT, Acc, Extended);
  }

  return DAG.getNode(Opcode, DL, ReducedVT, Acc, MulOpLHS, MulOpRHS);
}

SDValue tryLowerPartialReductionToWideAdd(SDNode *N,
                                          const AArch64Subtarget *Subtarget,
                                          SelectionDAG &DAG) {

  assert(N->getOpcode() == ISD::INTRINSIC_WO_CHAIN &&
         getIntrinsicID(N) ==
             Intrinsic::experimental_vector_partial_reduce_add &&
         "Expected a partial reduction node");

  if (!Subtarget->hasSVE2() && !Subtarget->isStreamingSVEAvailable())
    return SDValue();

  SDLoc DL(N);

  if (!ISD::isExtOpcode(N->getOperand(2).getOpcode()))
    return SDValue();
  SDValue Acc = N->getOperand(1);
  SDValue Ext = N->getOperand(2);
  EVT AccVT = Acc.getValueType();
  EVT ExtVT = Ext.getValueType();
  if (ExtVT.getVectorElementType() != AccVT.getVectorElementType())
    return SDValue();

  SDValue ExtOp = Ext->getOperand(0);
  EVT ExtOpVT = ExtOp.getValueType();

  if (!(ExtOpVT == MVT::nxv4i32 && AccVT == MVT::nxv2i64) &&
      !(ExtOpVT == MVT::nxv8i16 && AccVT == MVT::nxv4i32) &&
      !(ExtOpVT == MVT::nxv16i8 && AccVT == MVT::nxv8i16))
    return SDValue();

  bool ExtOpIsSigned = Ext.getOpcode() == ISD::SIGN_EXTEND;
  unsigned BottomOpcode =
      ExtOpIsSigned ? AArch64ISD::SADDWB : AArch64ISD::UADDWB;
  unsigned TopOpcode = ExtOpIsSigned ? AArch64ISD::SADDWT : AArch64ISD::UADDWT;
  SDValue BottomNode = DAG.getNode(BottomOpcode, DL, AccVT, Acc, ExtOp);
  return DAG.getNode(TopOpcode, DL, AccVT, BottomNode, ExtOp);
}

static SDValue combineSVEBitSel(unsigned IID, SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  SDValue Op1 = N->getOperand(1);
  SDValue Op2 = N->getOperand(2);
  SDValue Op3 = N->getOperand(3);

  switch (IID) {
  default:
    llvm_unreachable("Called with wrong intrinsic!");
  case Intrinsic::aarch64_sve_bsl:
    return DAG.getNode(AArch64ISD::BSP, DL, VT, Op3, Op1, Op2);
  case Intrinsic::aarch64_sve_bsl1n:
    return DAG.getNode(AArch64ISD::BSP, DL, VT, Op3, DAG.getNOT(DL, Op1, VT),
                       Op2);
  case Intrinsic::aarch64_sve_bsl2n:
    return DAG.getNode(AArch64ISD::BSP, DL, VT, Op3, Op1,
                       DAG.getNOT(DL, Op2, VT));
  case Intrinsic::aarch64_sve_nbsl:
    return DAG.getNOT(DL, DAG.getNode(AArch64ISD::BSP, DL, VT, Op3, Op1, Op2),
                      VT);
  }
}

static SDValue performIntrinsicCombine(SDNode *N,
                                       TargetLowering::DAGCombinerInfo &DCI,
                                       const AArch64Subtarget *Subtarget) {
  SelectionDAG &DAG = DCI.DAG;
  unsigned IID = getIntrinsicID(N);
  switch (IID) {
  default:
    break;
  case Intrinsic::experimental_vector_partial_reduce_add: {
    if (SDValue Dot = tryLowerPartialReductionToDot(N, Subtarget, DAG))
      return Dot;
    if (SDValue WideAdd = tryLowerPartialReductionToWideAdd(N, Subtarget, DAG))
      return WideAdd;
    SDLoc DL(N);
    SDValue Input = N->getOperand(2);
    return DAG.getNode(ISD::PARTIAL_REDUCE_UMLA, DL, N->getValueType(0),
                       N->getOperand(1), Input,
                       DAG.getConstant(1, DL, Input.getValueType()));
  }
  case Intrinsic::aarch64_neon_vcvtfxs2fp:
  case Intrinsic::aarch64_neon_vcvtfxu2fp:
    return tryCombineFixedPointConvert(N, DCI, DAG);
  case Intrinsic::aarch64_neon_saddv:
    return combineAcrossLanesIntrinsic(AArch64ISD::SADDV, N, DAG);
  case Intrinsic::aarch64_neon_uaddv:
    return combineAcrossLanesIntrinsic(AArch64ISD::UADDV, N, DAG);
  case Intrinsic::aarch64_neon_sminv:
    return combineAcrossLanesIntrinsic(AArch64ISD::SMINV, N, DAG);
  case Intrinsic::aarch64_neon_uminv:
    return combineAcrossLanesIntrinsic(AArch64ISD::UMINV, N, DAG);
  case Intrinsic::aarch64_neon_smaxv:
    return combineAcrossLanesIntrinsic(AArch64ISD::SMAXV, N, DAG);
  case Intrinsic::aarch64_neon_umaxv:
    return combineAcrossLanesIntrinsic(AArch64ISD::UMAXV, N, DAG);
  case Intrinsic::aarch64_neon_fmax:
    return DAG.getNode(ISD::FMAXIMUM, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_fmin:
    return DAG.getNode(ISD::FMINIMUM, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_fmaxnm:
    return DAG.getNode(ISD::FMAXNUM, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_fminnm:
    return DAG.getNode(ISD::FMINNUM, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_smull:
    return DAG.getNode(AArch64ISD::SMULL, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_umull:
    return DAG.getNode(AArch64ISD::UMULL, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_pmull:
    return DAG.getNode(AArch64ISD::PMULL, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_sqdmull:
    return tryCombineLongOpWithDup(IID, N, DCI, DAG);
  case Intrinsic::aarch64_neon_sqshl:
  case Intrinsic::aarch64_neon_uqshl:
  case Intrinsic::aarch64_neon_sqshlu:
  case Intrinsic::aarch64_neon_srshl:
  case Intrinsic::aarch64_neon_urshl:
  case Intrinsic::aarch64_neon_sshl:
  case Intrinsic::aarch64_neon_ushl:
    return tryCombineShiftImm(IID, N, DAG);
  case Intrinsic::aarch64_neon_sabd:
    return DAG.getNode(ISD::ABDS, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_neon_uabd:
    return DAG.getNode(ISD::ABDU, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_crc32b:
  case Intrinsic::aarch64_crc32cb:
    return tryCombineCRC32(0xff, N, DAG);
  case Intrinsic::aarch64_crc32h:
  case Intrinsic::aarch64_crc32ch:
    return tryCombineCRC32(0xffff, N, DAG);
  case Intrinsic::aarch64_sve_saddv:
    // There is no i64 version of SADDV because the sign is irrelevant.
    if (N->getOperand(2)->getValueType(0).getVectorElementType() == MVT::i64)
      return combineSVEReductionInt(N, AArch64ISD::UADDV_PRED, DAG);
    else
      return combineSVEReductionInt(N, AArch64ISD::SADDV_PRED, DAG);
  case Intrinsic::aarch64_sve_uaddv:
    return combineSVEReductionInt(N, AArch64ISD::UADDV_PRED, DAG);
  case Intrinsic::aarch64_sve_smaxv:
    return combineSVEReductionInt(N, AArch64ISD::SMAXV_PRED, DAG);
  case Intrinsic::aarch64_sve_umaxv:
    return combineSVEReductionInt(N, AArch64ISD::UMAXV_PRED, DAG);
  case Intrinsic::aarch64_sve_sminv:
    return combineSVEReductionInt(N, AArch64ISD::SMINV_PRED, DAG);
  case Intrinsic::aarch64_sve_uminv:
    return combineSVEReductionInt(N, AArch64ISD::UMINV_PRED, DAG);
  case Intrinsic::aarch64_sve_orv:
    return combineSVEReductionInt(N, AArch64ISD::ORV_PRED, DAG);
  case Intrinsic::aarch64_sve_eorv:
    return combineSVEReductionInt(N, AArch64ISD::EORV_PRED, DAG);
  case Intrinsic::aarch64_sve_andv:
    return combineSVEReductionInt(N, AArch64ISD::ANDV_PRED, DAG);
  case Intrinsic::aarch64_sve_index:
    return LowerSVEIntrinsicIndex(N, DAG);
  case Intrinsic::aarch64_sve_dup:
    return LowerSVEIntrinsicDUP(N, DAG);
  case Intrinsic::aarch64_sve_dup_x:
    return DAG.getNode(ISD::SPLAT_VECTOR, SDLoc(N), N->getValueType(0),
                       N->getOperand(1));
  case Intrinsic::aarch64_sve_ext:
    return LowerSVEIntrinsicEXT(N, DAG);
  case Intrinsic::aarch64_sve_mul_u:
    return DAG.getNode(AArch64ISD::MUL_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_smulh_u:
    return DAG.getNode(AArch64ISD::MULHS_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_umulh_u:
    return DAG.getNode(AArch64ISD::MULHU_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_smin_u:
    return DAG.getNode(AArch64ISD::SMIN_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_umin_u:
    return DAG.getNode(AArch64ISD::UMIN_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_smax_u:
    return DAG.getNode(AArch64ISD::SMAX_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_umax_u:
    return DAG.getNode(AArch64ISD::UMAX_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_lsl_u:
    return DAG.getNode(AArch64ISD::SHL_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_lsr_u:
    return DAG.getNode(AArch64ISD::SRL_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_asr_u:
    return DAG.getNode(AArch64ISD::SRA_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fadd_u:
    return DAG.getNode(AArch64ISD::FADD_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fdiv_u:
    return DAG.getNode(AArch64ISD::FDIV_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fmax_u:
    return DAG.getNode(AArch64ISD::FMAX_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fmaxnm_u:
    return DAG.getNode(AArch64ISD::FMAXNM_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fmla_u:
    return DAG.getNode(AArch64ISD::FMA_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(3), N->getOperand(4),
                       N->getOperand(2));
  case Intrinsic::aarch64_sve_fmin_u:
    return DAG.getNode(AArch64ISD::FMIN_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fminnm_u:
    return DAG.getNode(AArch64ISD::FMINNM_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fmul_u:
    return DAG.getNode(AArch64ISD::FMUL_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_fsub_u:
    return DAG.getNode(AArch64ISD::FSUB_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_add_u:
    return DAG.getNode(ISD::ADD, SDLoc(N), N->getValueType(0), N->getOperand(2),
                       N->getOperand(3));
  case Intrinsic::aarch64_sve_sub_u:
    return DAG.getNode(ISD::SUB, SDLoc(N), N->getValueType(0), N->getOperand(2),
                       N->getOperand(3));
  case Intrinsic::aarch64_sve_subr:
    return convertMergedOpToPredOp(N, ISD::SUB, DAG, true, true);
  case Intrinsic::aarch64_sve_and_u:
    return DAG.getNode(ISD::AND, SDLoc(N), N->getValueType(0), N->getOperand(2),
                       N->getOperand(3));
  case Intrinsic::aarch64_sve_bic_u:
    return DAG.getNode(AArch64ISD::BIC, SDLoc(N), N->getValueType(0),
                       N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_saddwb:
    return DAG.getNode(AArch64ISD::SADDWB, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_saddwt:
    return DAG.getNode(AArch64ISD::SADDWT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_uaddwb:
    return DAG.getNode(AArch64ISD::UADDWB, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_uaddwt:
    return DAG.getNode(AArch64ISD::UADDWT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_eor_u:
    return DAG.getNode(ISD::XOR, SDLoc(N), N->getValueType(0), N->getOperand(2),
                       N->getOperand(3));
  case Intrinsic::aarch64_sve_orr_u:
    return DAG.getNode(ISD::OR, SDLoc(N), N->getValueType(0), N->getOperand(2),
                       N->getOperand(3));
  case Intrinsic::aarch64_sve_sabd_u:
    return DAG.getNode(ISD::ABDS, SDLoc(N), N->getValueType(0),
                       N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_uabd_u:
    return DAG.getNode(ISD::ABDU, SDLoc(N), N->getValueType(0),
                       N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_sdiv_u:
    return DAG.getNode(AArch64ISD::SDIV_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_udiv_u:
    return DAG.getNode(AArch64ISD::UDIV_PRED, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_sqadd:
    return convertMergedOpToPredOp(N, ISD::SADDSAT, DAG, true);
  case Intrinsic::aarch64_sve_sqsub_u:
    return DAG.getNode(ISD::SSUBSAT, SDLoc(N), N->getValueType(0),
                       N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_uqadd:
    return convertMergedOpToPredOp(N, ISD::UADDSAT, DAG, true);
  case Intrinsic::aarch64_sve_uqsub_u:
    return DAG.getNode(ISD::USUBSAT, SDLoc(N), N->getValueType(0),
                       N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_sqadd_x:
    return DAG.getNode(ISD::SADDSAT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_sqsub_x:
    return DAG.getNode(ISD::SSUBSAT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_uqadd_x:
    return DAG.getNode(ISD::UADDSAT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_uqsub_x:
    return DAG.getNode(ISD::USUBSAT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_asrd:
    return DAG.getNode(AArch64ISD::SRAD_MERGE_OP1, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_cmphs:
    if (!N->getOperand(2).getValueType().isFloatingPoint())
      return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N),
                         N->getValueType(0), N->getOperand(1), N->getOperand(2),
                         N->getOperand(3), DAG.getCondCode(ISD::SETUGE));
    break;
  case Intrinsic::aarch64_sve_cmphi:
    if (!N->getOperand(2).getValueType().isFloatingPoint())
      return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N),
                         N->getValueType(0), N->getOperand(1), N->getOperand(2),
                         N->getOperand(3), DAG.getCondCode(ISD::SETUGT));
    break;
  case Intrinsic::aarch64_sve_fcmpge:
  case Intrinsic::aarch64_sve_cmpge:
    return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N),
                       N->getValueType(0), N->getOperand(1), N->getOperand(2),
                       N->getOperand(3), DAG.getCondCode(ISD::SETGE));
    break;
  case Intrinsic::aarch64_sve_fcmpgt:
  case Intrinsic::aarch64_sve_cmpgt:
    return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N),
                       N->getValueType(0), N->getOperand(1), N->getOperand(2),
                       N->getOperand(3), DAG.getCondCode(ISD::SETGT));
    break;
  case Intrinsic::aarch64_sve_fcmpeq:
  case Intrinsic::aarch64_sve_cmpeq:
    return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N),
                       N->getValueType(0), N->getOperand(1), N->getOperand(2),
                       N->getOperand(3), DAG.getCondCode(ISD::SETEQ));
    break;
  case Intrinsic::aarch64_sve_fcmpne:
  case Intrinsic::aarch64_sve_cmpne:
    return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N),
                       N->getValueType(0), N->getOperand(1), N->getOperand(2),
                       N->getOperand(3), DAG.getCondCode(ISD::SETNE));
    break;
  case Intrinsic::aarch64_sve_fcmpuo:
    return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, SDLoc(N),
                       N->getValueType(0), N->getOperand(1), N->getOperand(2),
                       N->getOperand(3), DAG.getCondCode(ISD::SETUO));
    break;
  case Intrinsic::aarch64_sve_fadda:
    return combineSVEReductionOrderedFP(N, AArch64ISD::FADDA_PRED, DAG);
  case Intrinsic::aarch64_sve_faddv:
    return combineSVEReductionFP(N, AArch64ISD::FADDV_PRED, DAG);
  case Intrinsic::aarch64_sve_fmaxnmv:
    return combineSVEReductionFP(N, AArch64ISD::FMAXNMV_PRED, DAG);
  case Intrinsic::aarch64_sve_fmaxv:
    return combineSVEReductionFP(N, AArch64ISD::FMAXV_PRED, DAG);
  case Intrinsic::aarch64_sve_fminnmv:
    return combineSVEReductionFP(N, AArch64ISD::FMINNMV_PRED, DAG);
  case Intrinsic::aarch64_sve_fminv:
    return combineSVEReductionFP(N, AArch64ISD::FMINV_PRED, DAG);
  case Intrinsic::aarch64_sve_sel:
    return DAG.getNode(ISD::VSELECT, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2), N->getOperand(3));
  case Intrinsic::aarch64_sve_cmpeq_wide:
    return tryConvertSVEWideCompare(N, ISD::SETEQ, DCI, DAG);
  case Intrinsic::aarch64_sve_cmpne_wide:
    return tryConvertSVEWideCompare(N, ISD::SETNE, DCI, DAG);
  case Intrinsic::aarch64_sve_cmpge_wide:
    return tryConvertSVEWideCompare(N, ISD::SETGE, DCI, DAG);
  case Intrinsic::aarch64_sve_cmpgt_wide:
    return tryConvertSVEWideCompare(N, ISD::SETGT, DCI, DAG);
  case Intrinsic::aarch64_sve_cmplt_wide:
    return tryConvertSVEWideCompare(N, ISD::SETLT, DCI, DAG);
  case Intrinsic::aarch64_sve_cmple_wide:
    return tryConvertSVEWideCompare(N, ISD::SETLE, DCI, DAG);
  case Intrinsic::aarch64_sve_cmphs_wide:
    return tryConvertSVEWideCompare(N, ISD::SETUGE, DCI, DAG);
  case Intrinsic::aarch64_sve_cmphi_wide:
    return tryConvertSVEWideCompare(N, ISD::SETUGT, DCI, DAG);
  case Intrinsic::aarch64_sve_cmplo_wide:
    return tryConvertSVEWideCompare(N, ISD::SETULT, DCI, DAG);
  case Intrinsic::aarch64_sve_cmpls_wide:
    return tryConvertSVEWideCompare(N, ISD::SETULE, DCI, DAG);
  case Intrinsic::aarch64_sve_ptest_any:
    return getPTest(DAG, N->getValueType(0), N->getOperand(1), N->getOperand(2),
                    AArch64CC::ANY_ACTIVE);
  case Intrinsic::aarch64_sve_ptest_first:
    return getPTest(DAG, N->getValueType(0), N->getOperand(1), N->getOperand(2),
                    AArch64CC::FIRST_ACTIVE);
  case Intrinsic::aarch64_sve_ptest_last:
    return getPTest(DAG, N->getValueType(0), N->getOperand(1), N->getOperand(2),
                    AArch64CC::LAST_ACTIVE);
  case Intrinsic::aarch64_sve_whilelo:
    return DAG.getNode(ISD::GET_ACTIVE_LANE_MASK, SDLoc(N), N->getValueType(0),
                       N->getOperand(1), N->getOperand(2));
  case Intrinsic::aarch64_sve_bsl:
  case Intrinsic::aarch64_sve_bsl1n:
  case Intrinsic::aarch64_sve_bsl2n:
  case Intrinsic::aarch64_sve_nbsl:
    return combineSVEBitSel(IID, N, DAG);
  }
  return SDValue();
}

static bool isCheapToExtend(const SDValue &N) {
  unsigned OC = N->getOpcode();
  return OC == ISD::LOAD || OC == ISD::MLOAD ||
         ISD::isConstantSplatVectorAllZeros(N.getNode());
}

static SDValue
performSignExtendSetCCCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                              SelectionDAG &DAG) {
  // If we have (sext (setcc A B)) and A and B are cheap to extend,
  // we can move the sext into the arguments and have the same result. For
  // example, if A and B are both loads, we can make those extending loads and
  // avoid an extra instruction. This pattern appears often in VLS code
  // generation where the inputs to the setcc have a different size to the
  // instruction that wants to use the result of the setcc.
  assert(N->getOpcode() == ISD::SIGN_EXTEND &&
         N->getOperand(0)->getOpcode() == ISD::SETCC);
  const SDValue SetCC = N->getOperand(0);

  const SDValue CCOp0 = SetCC.getOperand(0);
  const SDValue CCOp1 = SetCC.getOperand(1);
  if (!CCOp0->getValueType(0).isInteger() ||
      !CCOp1->getValueType(0).isInteger())
    return SDValue();

  ISD::CondCode Code =
      cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get();

  ISD::NodeType ExtType =
      isSignedIntSetCC(Code) ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;

  if (isCheapToExtend(SetCC.getOperand(0)) &&
      isCheapToExtend(SetCC.getOperand(1))) {
    const SDValue Ext1 =
        DAG.getNode(ExtType, SDLoc(N), N->getValueType(0), CCOp0);
    const SDValue Ext2 =
        DAG.getNode(ExtType, SDLoc(N), N->getValueType(0), CCOp1);

    return DAG.getSetCC(
        SDLoc(SetCC), N->getValueType(0), Ext1, Ext2,
        cast<CondCodeSDNode>(SetCC->getOperand(2).getNode())->get());
  }

  return SDValue();
}

// Convert zext(extract(shuffle a, b, [0,4,8,12])) -> and(uzp1(a, b), 255)
// This comes from interleaved vectorization. It is performed late to capture
// uitofp converts too.
static SDValue performZExtDeinterleaveShuffleCombine(SDNode *N,
                                                     SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if ((VT != MVT::v4i32 && VT != MVT::v8i16) ||
      N->getOpcode() != ISD::ZERO_EXTEND ||
      N->getOperand(0).getOpcode() != ISD::EXTRACT_SUBVECTOR)
    return SDValue();

  unsigned ExtOffset = N->getOperand(0).getConstantOperandVal(1);
  if (ExtOffset != 0 && ExtOffset != VT.getVectorNumElements())
    return SDValue();

  EVT InVT = N->getOperand(0).getOperand(0).getValueType();
  auto *Shuffle = dyn_cast<ShuffleVectorSDNode>(N->getOperand(0).getOperand(0));
  if (!Shuffle ||
      InVT.getVectorNumElements() != VT.getVectorNumElements() * 2 ||
      InVT.getScalarSizeInBits() * 2 != VT.getScalarSizeInBits())
    return SDValue();

  unsigned Idx;
  bool IsDeInterleave = ShuffleVectorInst::isDeInterleaveMaskOfFactor(
      Shuffle->getMask().slice(ExtOffset, VT.getVectorNumElements()), 4, Idx);
  // An undef interleave shuffle can come up after other canonicalizations,
  // where the shuffle has been converted to
  //   zext(extract(shuffle b, undef, [u,u,0,4]))
  bool IsUndefDeInterleave = false;
  if (!IsDeInterleave)
    IsUndefDeInterleave =
        Shuffle->getOperand(1).isUndef() &&
        all_of(
            Shuffle->getMask().slice(ExtOffset, VT.getVectorNumElements() / 2),
            [](int M) { return M < 0; }) &&
        ShuffleVectorInst::isDeInterleaveMaskOfFactor(
            Shuffle->getMask().slice(ExtOffset + VT.getVectorNumElements() / 2,
                                     VT.getVectorNumElements() / 2),
            4, Idx);
  if ((!IsDeInterleave && !IsUndefDeInterleave) || Idx >= 4)
    return SDValue();
  SDLoc DL(N);
  SDValue BC1 = DAG.getNode(AArch64ISD::NVCAST, DL, VT,
                            Shuffle->getOperand(IsUndefDeInterleave ? 1 : 0));
  SDValue BC2 = DAG.getNode(AArch64ISD::NVCAST, DL, VT,
                            Shuffle->getOperand(IsUndefDeInterleave ? 0 : 1));
  SDValue UZP = DAG.getNode(Idx < 2 ? AArch64ISD::UZP1 : AArch64ISD::UZP2, DL,
                            VT, BC1, BC2);
  if ((Idx & 1) == 1)
    UZP = DAG.getNode(ISD::SRL, DL, VT, UZP,
                      DAG.getConstant(InVT.getScalarSizeInBits(), DL, VT));
  return DAG.getNode(
      ISD::AND, DL, VT, UZP,
      DAG.getConstant((1 << InVT.getScalarSizeInBits()) - 1, DL, VT));
}

// This comes up similar to the above when lowering deinterleaving shuffles from
// zexts. We have legalized the operations in the generally case to
// zext(extract_subvector(uzp(a, b))), which can be converted to and(a, mask) if
// the extract is to the low half and the uzp is uzp1. There would be an extra
// shift if the uzp was uzp2 to grab the upper half. Due to the combine above
// there could also be an existing and / shift that can be combined in, either
// before of after the extract.
static SDValue performZExtUZPCombine(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  if (N->getOpcode() != ISD::ZERO_EXTEND ||
      (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16))
    return SDValue();

  SDValue Op = N->getOperand(0);
  unsigned ExtOffset = (unsigned)-1;
  if (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
    ExtOffset = Op.getConstantOperandVal(1);
    Op = Op.getOperand(0);
  }

  unsigned Shift = 0;
  APInt Mask = APInt::getLowBitsSet(VT.getScalarSizeInBits(),
                                    Op.getValueType().getScalarSizeInBits());

  if (Op.getOpcode() == AArch64ISD::VLSHR) {
    Shift = Op.getConstantOperandVal(1);
    Op = Op.getOperand(0);
    Mask = Mask.lshr(Shift);
  }
  if (Op.getOpcode() == ISD::AND &&
      ISD::isConstantSplatVector(Op.getOperand(1).getNode(), Mask)) {
    Op = Op.getOperand(0);
    Mask = Mask.zext(VT.getScalarSizeInBits());
  } else if (Op.getOpcode() == AArch64ISD::BICi) {
    Mask = ~APInt(Op.getValueType().getScalarSizeInBits(),
                  Op.getConstantOperandVal(1) << Op.getConstantOperandVal(2));
    Mask = Mask.zext(VT.getScalarSizeInBits());
    Op = Op.getOperand(0);
  }

  if (ExtOffset == (unsigned)-1) {
    if (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
      ExtOffset = Op.getConstantOperandVal(1);
      Op = Op.getOperand(0);
    } else
      return SDValue();
  }
  if (ExtOffset != 0 && ExtOffset != VT.getVectorNumElements())
    return SDValue();

  if (Op.getOpcode() != AArch64ISD::UZP1 && Op.getOpcode() != AArch64ISD::UZP2)
    return SDValue();
  if (Op.getOpcode() == AArch64ISD::UZP2)
    Shift += VT.getScalarSizeInBits() / 2;

  SDLoc DL(N);
  SDValue BC = DAG.getNode(AArch64ISD::NVCAST, DL, VT,
                           Op.getOperand(ExtOffset == 0 ? 0 : 1));
  if (Shift != 0)
    BC = DAG.getNode(AArch64ISD::VLSHR, DL, VT, BC,
                     DAG.getConstant(Shift, DL, MVT::i32));
  return DAG.getNode(ISD::AND, DL, VT, BC, DAG.getConstant(Mask, DL, VT));
}

static SDValue performExtendCombine(SDNode *N,
                                    TargetLowering::DAGCombinerInfo &DCI,
                                    SelectionDAG &DAG) {
  // If we see something like (zext (sabd (extract_high ...), (DUP ...))) then
  // we can convert that DUP into another extract_high (of a bigger DUP), which
  // helps the backend to decide that an sabdl2 would be useful, saving a real
  // extract_high operation.
  if (!DCI.isBeforeLegalizeOps() && N->getOpcode() == ISD::ZERO_EXTEND &&
      N->getOperand(0).getValueType().is64BitVector() &&
      (N->getOperand(0).getOpcode() == ISD::ABDU ||
       N->getOperand(0).getOpcode() == ISD::ABDS)) {
    SDNode *ABDNode = N->getOperand(0).getNode();
    SDValue NewABD =
        tryCombineLongOpWithDup(Intrinsic::not_intrinsic, ABDNode, DCI, DAG);
    if (!NewABD.getNode())
      return SDValue();

    return DAG.getNode(ISD::ZERO_EXTEND, SDLoc(N), N->getValueType(0), NewABD);
  }

  if (SDValue R = performZExtDeinterleaveShuffleCombine(N, DAG))
    return R;
  if (SDValue R = performZExtUZPCombine(N, DAG))
    return R;

  if (N->getValueType(0).isFixedLengthVector() &&
      N->getOpcode() == ISD::SIGN_EXTEND &&
      N->getOperand(0)->getOpcode() == ISD::SETCC)
    return performSignExtendSetCCCombine(N, DCI, DAG);

  // If we see (any_extend (bswap ...)) with bswap returning an i16, we know
  // that the top half of the result register must be unused, due to the
  // any_extend. This means that we can replace this pattern with (rev16
  // (any_extend ...)). This saves a machine instruction compared to (lsr (rev
  // ...)), which is what this pattern would otherwise be lowered to.
  // Only apply this optimisation if any_extend in original pattern to i32 or
  // i64, because this type will become the input type to REV16 in the new
  // pattern, so must be a legitimate REV16 input type.
  SDValue Bswap = N->getOperand(0);
  if (N->getOpcode() == ISD::ANY_EXTEND && Bswap.getOpcode() == ISD::BSWAP &&
      Bswap.getValueType() == MVT::i16 &&
      (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64)) {
    SDLoc DL(N);
    SDValue NewAnyExtend = DAG.getNode(ISD::ANY_EXTEND, DL, N->getValueType(0),
                                       Bswap->getOperand(0));
    return DAG.getNode(AArch64ISD::REV16, SDLoc(N), N->getValueType(0),
                       NewAnyExtend);
  }

  return SDValue();
}

static SDValue splitStoreSplat(SelectionDAG &DAG, StoreSDNode &St,
                               SDValue SplatVal, unsigned NumVecElts) {
  assert(!St.isTruncatingStore() && "cannot split truncating vector store");
  Align OrigAlignment = St.getAlign();
  unsigned EltOffset = SplatVal.getValueType().getSizeInBits() / 8;

  // Create scalar stores. This is at least as good as the code sequence for a
  // split unaligned store which is a dup.s, ext.b, and two stores.
  // Most of the time the three stores should be replaced by store pair
  // instructions (stp).
  SDLoc DL(&St);
  SDValue BasePtr = St.getBasePtr();
  uint64_t BaseOffset = 0;

  const MachinePointerInfo &PtrInfo = St.getPointerInfo();
  SDValue NewST1 =
      DAG.getStore(St.getChain(), DL, SplatVal, BasePtr, PtrInfo,
                   OrigAlignment, St.getMemOperand()->getFlags());

  // As this in ISel, we will not merge this add which may degrade results.
  if (BasePtr->getOpcode() == ISD::ADD &&
      isa<ConstantSDNode>(BasePtr->getOperand(1))) {
    BaseOffset = cast<ConstantSDNode>(BasePtr->getOperand(1))->getSExtValue();
    BasePtr = BasePtr->getOperand(0);
  }

  unsigned Offset = EltOffset;
  while (--NumVecElts) {
    Align Alignment = commonAlignment(OrigAlignment, Offset);
    SDValue OffsetPtr =
        DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
                    DAG.getConstant(BaseOffset + Offset, DL, MVT::i64));
    NewST1 = DAG.getStore(NewST1.getValue(0), DL, SplatVal, OffsetPtr,
                          PtrInfo.getWithOffset(Offset), Alignment,
                          St.getMemOperand()->getFlags());
    Offset += EltOffset;
  }
  return NewST1;
}

// Returns an SVE type that ContentTy can be trivially sign or zero extended
// into.
static MVT getSVEContainerType(EVT ContentTy) {
  assert(ContentTy.isSimple() && "No SVE containers for extended types");

  switch (ContentTy.getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("No known SVE container for this MVT type");
  case MVT::nxv2i8:
  case MVT::nxv2i16:
  case MVT::nxv2i32:
  case MVT::nxv2i64:
  case MVT::nxv2f32:
  case MVT::nxv2f64:
    return MVT::nxv2i64;
  case MVT::nxv4i8:
  case MVT::nxv4i16:
  case MVT::nxv4i32:
  case MVT::nxv4f32:
    return MVT::nxv4i32;
  case MVT::nxv8i8:
  case MVT::nxv8i16:
  case MVT::nxv8f16:
  case MVT::nxv8bf16:
    return MVT::nxv8i16;
  case MVT::nxv16i8:
    return MVT::nxv16i8;
  }
}

static SDValue performLD1Combine(SDNode *N, SelectionDAG &DAG, unsigned Opc) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);

  if (VT.getSizeInBits().getKnownMinValue() > AArch64::SVEBitsPerBlock)
    return SDValue();

  EVT ContainerVT = VT;
  if (ContainerVT.isInteger())
    ContainerVT = getSVEContainerType(ContainerVT);

  SDVTList VTs = DAG.getVTList(ContainerVT, MVT::Other);
  SDValue Ops[] = { N->getOperand(0), // Chain
                    N->getOperand(2), // Pg
                    N->getOperand(3), // Base
                    DAG.getValueType(VT) };

  SDValue Load = DAG.getNode(Opc, DL, VTs, Ops);
  SDValue LoadChain = SDValue(Load.getNode(), 1);

  if (ContainerVT.isInteger() && (VT != ContainerVT))
    Load = DAG.getNode(ISD::TRUNCATE, DL, VT, Load.getValue(0));

  return DAG.getMergeValues({ Load, LoadChain }, DL);
}

static SDValue performLDNT1Combine(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  EVT VT = N->getValueType(0);
  EVT PtrTy = N->getOperand(3).getValueType();

  EVT LoadVT = VT;
  if (VT.isFloatingPoint())
    LoadVT = VT.changeTypeToInteger();

  auto *MINode = cast<MemIntrinsicSDNode>(N);
  SDValue PassThru = DAG.getConstant(0, DL, LoadVT);
  SDValue L = DAG.getMaskedLoad(LoadVT, DL, MINode->getChain(),
                                MINode->getOperand(3), DAG.getUNDEF(PtrTy),
                                MINode->getOperand(2), PassThru,
                                MINode->getMemoryVT(), MINode->getMemOperand(),
                                ISD::UNINDEXED, ISD::NON_EXTLOAD, false);

   if (VT.isFloatingPoint()) {
     SDValue Ops[] = { DAG.getNode(ISD::BITCAST, DL, VT, L), L.getValue(1) };
     return DAG.getMergeValues(Ops, DL);
   }

  return L;
}

template <unsigned Opcode>
static SDValue performLD1ReplicateCombine(SDNode *N, SelectionDAG &DAG) {
  static_assert(Opcode == AArch64ISD::LD1RQ_MERGE_ZERO ||
                    Opcode == AArch64ISD::LD1RO_MERGE_ZERO,
                "Unsupported opcode.");
  SDLoc DL(N);
  EVT VT = N->getValueType(0);

  EVT LoadVT = VT;
  if (VT.isFloatingPoint())
    LoadVT = VT.changeTypeToInteger();

  SDValue Ops[] = {N->getOperand(0), N->getOperand(2), N->getOperand(3)};
  SDValue Load = DAG.getNode(Opcode, DL, {LoadVT, MVT::Other}, Ops);
  SDValue LoadChain = SDValue(Load.getNode(), 1);

  if (VT.isFloatingPoint())
    Load = DAG.getNode(ISD::BITCAST, DL, VT, Load.getValue(0));

  return DAG.getMergeValues({Load, LoadChain}, DL);
}

static SDValue performST1Combine(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue Data = N->getOperand(2);
  EVT DataVT = Data.getValueType();
  EVT HwSrcVt = getSVEContainerType(DataVT);
  SDValue InputVT = DAG.getValueType(DataVT);

  if (DataVT.isFloatingPoint())
    InputVT = DAG.getValueType(HwSrcVt);

  SDValue SrcNew;
  if (Data.getValueType().isFloatingPoint())
    SrcNew = DAG.getNode(ISD::BITCAST, DL, HwSrcVt, Data);
  else
    SrcNew = DAG.getNode(ISD::ANY_EXTEND, DL, HwSrcVt, Data);

  SDValue Ops[] = { N->getOperand(0), // Chain
                    SrcNew,
                    N->getOperand(4), // Base
                    N->getOperand(3), // Pg
                    InputVT
                  };

  return DAG.getNode(AArch64ISD::ST1_PRED, DL, N->getValueType(0), Ops);
}

static SDValue performSTNT1Combine(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);

  SDValue Data = N->getOperand(2);
  EVT DataVT = Data.getValueType();
  EVT PtrTy = N->getOperand(4).getValueType();

  if (DataVT.isFloatingPoint())
    Data = DAG.getNode(ISD::BITCAST, DL, DataVT.changeTypeToInteger(), Data);

  auto *MINode = cast<MemIntrinsicSDNode>(N);
  return DAG.getMaskedStore(MINode->getChain(), DL, Data, MINode->getOperand(4),
                            DAG.getUNDEF(PtrTy), MINode->getOperand(3),
                            MINode->getMemoryVT(), MINode->getMemOperand(),
                            ISD::UNINDEXED, false, false);
}

/// Replace a splat of zeros to a vector store by scalar stores of WZR/XZR.  The
/// load store optimizer pass will merge them to store pair stores.  This should
/// be better than a movi to create the vector zero followed by a vector store
/// if the zero constant is not re-used, since one instructions and one register
/// live range will be removed.
///
/// For example, the final generated code should be:
///
///   stp xzr, xzr, [x0]
///
/// instead of:
///
///   movi v0.2d, #0
///   str q0, [x0]
///
static SDValue replaceZeroVectorStore(SelectionDAG &DAG, StoreSDNode &St) {
  SDValue StVal = St.getValue();
  EVT VT = StVal.getValueType();

  // Avoid scalarizing zero splat stores for scalable vectors.
  if (VT.isScalableVector())
    return SDValue();

  // It is beneficial to scalarize a zero splat store for 2 or 3 i64 elements or
  // 2, 3 or 4 i32 elements.
  int NumVecElts = VT.getVectorNumElements();
  if (!(((NumVecElts == 2 || NumVecElts == 3) &&
         VT.getVectorElementType().getSizeInBits() == 64) ||
        ((NumVecElts == 2 || NumVecElts == 3 || NumVecElts == 4) &&
         VT.getVectorElementType().getSizeInBits() == 32)))
    return SDValue();

  if (StVal.getOpcode() != ISD::BUILD_VECTOR)
    return SDValue();

  // If the zero constant has more than one use then the vector store could be
  // better since the constant mov will be amortized and stp q instructions
  // should be able to be formed.
  if (!StVal.hasOneUse())
    return SDValue();

  // If the store is truncating then it's going down to i16 or smaller, which
  // means it can be implemented in a single store anyway.
  if (St.isTruncatingStore())
    return SDValue();

  // If the immediate offset of the address operand is too large for the stp
  // instruction, then bail out.
  if (DAG.isBaseWithConstantOffset(St.getBasePtr())) {
    int64_t Offset = St.getBasePtr()->getConstantOperandVal(1);
    if (Offset < -512 || Offset > 504)
      return SDValue();
  }

  for (int I = 0; I < NumVecElts; ++I) {
    SDValue EltVal = StVal.getOperand(I);
    if (!isNullConstant(EltVal) && !isNullFPConstant(EltVal))
      return SDValue();
  }

  // Use a CopyFromReg WZR/XZR here to prevent
  // DAGCombiner::MergeConsecutiveStores from undoing this transformation.
  SDLoc DL(&St);
  unsigned ZeroReg;
  EVT ZeroVT;
  if (VT.getVectorElementType().getSizeInBits() == 32) {
    ZeroReg = AArch64::WZR;
    ZeroVT = MVT::i32;
  } else {
    ZeroReg = AArch64::XZR;
    ZeroVT = MVT::i64;
  }
  SDValue SplatVal =
      DAG.getCopyFromReg(DAG.getEntryNode(), DL, ZeroReg, ZeroVT);
  return splitStoreSplat(DAG, St, SplatVal, NumVecElts);
}

/// Replace a splat of a scalar to a vector store by scalar stores of the scalar
/// value. The load store optimizer pass will merge them to store pair stores.
/// This has better performance than a splat of the scalar followed by a split
/// vector store. Even if the stores are not merged it is four stores vs a dup,
/// followed by an ext.b and two stores.
static SDValue replaceSplatVectorStore(SelectionDAG &DAG, StoreSDNode &St) {
  SDValue StVal = St.getValue();
  EVT VT = StVal.getValueType();

  // Don't replace floating point stores, they possibly won't be transformed to
  // stp because of the store pair suppress pass.
  if (VT.isFloatingPoint())
    return SDValue();

  // We can express a splat as store pair(s) for 2 or 4 elements.
  unsigned NumVecElts = VT.getVectorNumElements();
  if (NumVecElts != 4 && NumVecElts != 2)
    return SDValue();

  // If the store is truncating then it's going down to i16 or smaller, which
  // means it can be implemented in a single store anyway.
  if (St.isTruncatingStore())
    return SDValue();

  // Check that this is a splat.
  // Make sure that each of the relevant vector element locations are inserted
  // to, i.e. 0 and 1 for v2i64 and 0, 1, 2, 3 for v4i32.
  std::bitset<4> IndexNotInserted((1 << NumVecElts) - 1);
  SDValue SplatVal;
  for (unsigned I = 0; I < NumVecElts; ++I) {
    // Check for insert vector elements.
    if (StVal.getOpcode() != ISD::INSERT_VECTOR_ELT)
      return SDValue();

    // Check that same value is inserted at each vector element.
    if (I == 0)
      SplatVal = StVal.getOperand(1);
    else if (StVal.getOperand(1) != SplatVal)
      return SDValue();

    // Check insert element index.
    ConstantSDNode *CIndex = dyn_cast<ConstantSDNode>(StVal.getOperand(2));
    if (!CIndex)
      return SDValue();
    uint64_t IndexVal = CIndex->getZExtValue();
    if (IndexVal >= NumVecElts)
      return SDValue();
    IndexNotInserted.reset(IndexVal);

    StVal = StVal.getOperand(0);
  }
  // Check that all vector element locations were inserted to.
  if (IndexNotInserted.any())
      return SDValue();

  return splitStoreSplat(DAG, St, SplatVal, NumVecElts);
}

static SDValue splitStores(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                           SelectionDAG &DAG,
                           const AArch64Subtarget *Subtarget) {

  StoreSDNode *S = cast<StoreSDNode>(N);
  if (S->isVolatile() || S->isIndexed())
    return SDValue();

  SDValue StVal = S->getValue();
  EVT VT = StVal.getValueType();

  if (!VT.isFixedLengthVector())
    return SDValue();

  // If we get a splat of zeros, convert this vector store to a store of
  // scalars. They will be merged into store pairs of xzr thereby removing one
  // instruction and one register.
  if (SDValue ReplacedZeroSplat = replaceZeroVectorStore(DAG, *S))
    return ReplacedZeroSplat;

  // FIXME: The logic for deciding if an unaligned store should be split should
  // be included in TLI.allowsMisalignedMemoryAccesses(), and there should be
  // a call to that function here.

  if (!Subtarget->isMisaligned128StoreSlow())
    return SDValue();

  // Don't split at -Oz.
  if (DAG.getMachineFunction().getFunction().hasMinSize())
    return SDValue();

  // Don't split v2i64 vectors. Memcpy lowering produces those and splitting
  // those up regresses performance on micro-benchmarks and olden/bh.
  if (VT.getVectorNumElements() < 2 || VT == MVT::v2i64)
    return SDValue();

  // Split unaligned 16B stores. They are terrible for performance.
  // Don't split stores with alignment of 1 or 2. Code that uses clang vector
  // extensions can use this to mark that it does not want splitting to happen
  // (by underspecifying alignment to be 1 or 2). Furthermore, the chance of
  // eliminating alignment hazards is only 1 in 8 for alignment of 2.
  if (VT.getSizeInBits() != 128 || S->getAlign() >= Align(16) ||
      S->getAlign() <= Align(2))
    return SDValue();

  // If we get a splat of a scalar convert this vector store to a store of
  // scalars. They will be merged into store pairs thereby removing two
  // instructions.
  if (SDValue ReplacedSplat = replaceSplatVectorStore(DAG, *S))
    return ReplacedSplat;

  SDLoc DL(S);

  // Split VT into two.
  EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext());
  unsigned NumElts = HalfVT.getVectorNumElements();
  SDValue SubVector0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
                                   DAG.getConstant(0, DL, MVT::i64));
  SDValue SubVector1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, StVal,
                                   DAG.getConstant(NumElts, DL, MVT::i64));
  SDValue BasePtr = S->getBasePtr();
  SDValue NewST1 =
      DAG.getStore(S->getChain(), DL, SubVector0, BasePtr, S->getPointerInfo(),
                   S->getAlign(), S->getMemOperand()->getFlags());
  SDValue OffsetPtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr,
                                  DAG.getConstant(8, DL, MVT::i64));
  return DAG.getStore(NewST1.getValue(0), DL, SubVector1, OffsetPtr,
                      S->getPointerInfo(), S->getAlign(),
                      S->getMemOperand()->getFlags());
}

static SDValue performSpliceCombine(SDNode *N, SelectionDAG &DAG) {
  assert(N->getOpcode() == AArch64ISD::SPLICE && "Unexpected Opcode!");

  // splice(pg, op1, undef) -> op1
  if (N->getOperand(2).isUndef())
    return N->getOperand(1);

  return SDValue();
}

static SDValue performUnpackCombine(SDNode *N, SelectionDAG &DAG,
                                    const AArch64Subtarget *Subtarget) {
  assert((N->getOpcode() == AArch64ISD::UUNPKHI ||
          N->getOpcode() == AArch64ISD::UUNPKLO) &&
         "Unexpected Opcode!");

  // uunpklo/hi undef -> undef
  if (N->getOperand(0).isUndef())
    return DAG.getUNDEF(N->getValueType(0));

  // If this is a masked load followed by an UUNPKLO, fold this into a masked
  // extending load.  We can do this even if this is already a masked
  // {z,}extload.
  if (N->getOperand(0).getOpcode() == ISD::MLOAD &&
      N->getOpcode() == AArch64ISD::UUNPKLO) {
    MaskedLoadSDNode *MLD = cast<MaskedLoadSDNode>(N->getOperand(0));
    SDValue Mask = MLD->getMask();
    SDLoc DL(N);

    if (MLD->isUnindexed() && MLD->getExtensionType() != ISD::SEXTLOAD &&
        SDValue(MLD, 0).hasOneUse() && Mask->getOpcode() == AArch64ISD::PTRUE &&
        (MLD->getPassThru()->isUndef() ||
         isZerosVector(MLD->getPassThru().getNode()))) {
      unsigned MinSVESize = Subtarget->getMinSVEVectorSizeInBits();
      unsigned PgPattern = Mask->getConstantOperandVal(0);
      EVT VT = N->getValueType(0);

      // Ensure we can double the size of the predicate pattern
      unsigned NumElts = getNumElementsFromSVEPredPattern(PgPattern);
      if (NumElts &&
          NumElts * VT.getVectorElementType().getSizeInBits() <= MinSVESize) {
        Mask =
            getPTrue(DAG, DL, VT.changeVectorElementType(MVT::i1), PgPattern);
        SDValue PassThru = DAG.getConstant(0, DL, VT);
        SDValue NewLoad = DAG.getMaskedLoad(
            VT, DL, MLD->getChain(), MLD->getBasePtr(), MLD->getOffset(), Mask,
            PassThru, MLD->getMemoryVT(), MLD->getMemOperand(),
            MLD->getAddressingMode(), ISD::ZEXTLOAD);

        DAG.ReplaceAllUsesOfValueWith(SDValue(MLD, 1), NewLoad.getValue(1));

        return NewLoad;
      }
    }
  }

  return SDValue();
}

static bool isHalvingTruncateAndConcatOfLegalIntScalableType(SDNode *N) {
  if (N->getOpcode() != AArch64ISD::UZP1)
    return false;
  SDValue Op0 = N->getOperand(0);
  EVT SrcVT = Op0->getValueType(0);
  EVT DstVT = N->getValueType(0);
  return (SrcVT == MVT::nxv8i16 && DstVT == MVT::nxv16i8) ||
         (SrcVT == MVT::nxv4i32 && DstVT == MVT::nxv8i16) ||
         (SrcVT == MVT::nxv2i64 && DstVT == MVT::nxv4i32);
}

// Try to combine rounding shifts where the operands come from an extend, and
// the result is truncated and combined into one vector.
//   uzp1(rshrnb(uunpklo(X),C), rshrnb(uunpkhi(X), C)) -> urshr(X, C)
static SDValue tryCombineExtendRShTrunc(SDNode *N, SelectionDAG &DAG) {
  assert(N->getOpcode() == AArch64ISD::UZP1 && "Only UZP1 expected.");
  SDValue Op0 = N->getOperand(0);
  SDValue Op1 = N->getOperand(1);
  EVT ResVT = N->getValueType(0);

  unsigned RshOpc = Op0.getOpcode();
  if (RshOpc != AArch64ISD::RSHRNB_I)
    return SDValue();

  // Same op code and imm value?
  SDValue ShiftValue = Op0.getOperand(1);
  if (RshOpc != Op1.getOpcode() || ShiftValue != Op1.getOperand(1))
    return SDValue();

  // Same unextended operand value?
  SDValue Lo = Op0.getOperand(0);
  SDValue Hi = Op1.getOperand(0);
  if (Lo.getOpcode() != AArch64ISD::UUNPKLO &&
      Hi.getOpcode() != AArch64ISD::UUNPKHI)
    return SDValue();
  SDValue OrigArg = Lo.getOperand(0);
  if (OrigArg != Hi.getOperand(0))
    return SDValue();

  SDLoc DL(N);
  return DAG.getNode(AArch64ISD::URSHR_I_PRED, DL, ResVT,
                     getPredicateForVector(DAG, DL, ResVT), OrigArg,
                     ShiftValue);
}

// Try to simplify:
//    t1 = nxv8i16 add(X, 1 << (ShiftValue - 1))
//    t2 = nxv8i16 srl(t1, ShiftValue)
// to
//    t1 = nxv8i16 rshrnb(X, shiftvalue).
// rshrnb will zero the top half bits of each element. Therefore, this combine
// should only be performed when a following instruction with the rshrnb
// as an operand does not care about the top half of each element. For example,
// a uzp1 or a truncating store.
static SDValue trySimplifySrlAddToRshrnb(SDValue Srl, SelectionDAG &DAG,
                                         const AArch64Subtarget *Subtarget) {
  EVT VT = Srl->getValueType(0);
  if (!VT.isScalableVector() || !Subtarget->hasSVE2())
    return SDValue();

  EVT ResVT;
  if (VT == MVT::nxv8i16)
    ResVT = MVT::nxv16i8;
  else if (VT == MVT::nxv4i32)
    ResVT = MVT::nxv8i16;
  else if (VT == MVT::nxv2i64)
    ResVT = MVT::nxv4i32;
  else
    return SDValue();

  SDLoc DL(Srl);
  unsigned ShiftValue;
  SDValue RShOperand;
  if (!canLowerSRLToRoundingShiftForVT(Srl, ResVT, DAG, ShiftValue, RShOperand))
    return SDValue();
  SDValue Rshrnb = DAG.getNode(
      AArch64ISD::RSHRNB_I, DL, ResVT,
      {RShOperand, DAG.getTargetConstant(ShiftValue, DL, MVT::i32)});
  return DAG.getNode(AArch64ISD::NVCAST, DL, VT, Rshrnb);
}

static SDValue isNVCastToHalfWidthElements(SDValue V) {
  if (V.getOpcode() != AArch64ISD::NVCAST)
    return SDValue();

  SDValue Op = V.getOperand(0);
  if (!Op.getValueType().isVector() ||
      V.getValueType().getVectorElementCount() !=
          Op.getValueType().getVectorElementCount() * 2)
    return SDValue();

  return Op;
}

static SDValue performUzpCombine(SDNode *N, SelectionDAG &DAG,
                                 const AArch64Subtarget *Subtarget) {
  SDLoc DL(N);
  SDValue Op0 = N->getOperand(0);
  SDValue Op1 = N->getOperand(1);
  EVT ResVT = N->getValueType(0);

  // uzp(extract_lo(x), extract_hi(x)) -> extract_lo(uzp x, x)
  if (Op0.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
      Op1.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
      Op0.getOperand(0) == Op1.getOperand(0)) {

    SDValue SourceVec = Op0.getOperand(0);
    uint64_t ExtIdx0 = Op0.getConstantOperandVal(1);
    uint64_t ExtIdx1 = Op1.getConstantOperandVal(1);
    uint64_t NumElements = SourceVec.getValueType().getVectorMinNumElements();
    if (ExtIdx0 == 0 && ExtIdx1 == NumElements / 2) {
      EVT OpVT = Op0.getOperand(1).getValueType();
      EVT WidenedResVT = ResVT.getDoubleNumVectorElementsVT(*DAG.getContext());
      SDValue Uzp = DAG.getNode(N->getOpcode(), DL, WidenedResVT, SourceVec,
                                DAG.getUNDEF(WidenedResVT));
      return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ResVT, Uzp,
                         DAG.getConstant(0, DL, OpVT));
    }
  }

  // Following optimizations only work with uzp1.
  if (N->getOpcode() == AArch64ISD::UZP2)
    return SDValue();

  // uzp1(x, undef) -> concat(truncate(x), undef)
  if (Op1.getOpcode() == ISD::UNDEF) {
    EVT BCVT = MVT::Other, HalfVT = MVT::Other;
    switch (ResVT.getSimpleVT().SimpleTy) {
    default:
      break;
    case MVT::v16i8:
      BCVT = MVT::v8i16;
      HalfVT = MVT::v8i8;
      break;
    case MVT::v8i16:
      BCVT = MVT::v4i32;
      HalfVT = MVT::v4i16;
      break;
    case MVT::v4i32:
      BCVT = MVT::v2i64;
      HalfVT = MVT::v2i32;
      break;
    }
    if (BCVT != MVT::Other) {
      SDValue BC = DAG.getBitcast(BCVT, Op0);
      SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, BC);
      return DAG.getNode(ISD::CONCAT_VECTORS, DL, ResVT, Trunc,
                         DAG.getUNDEF(HalfVT));
    }
  }

  if (SDValue Urshr = tryCombineExtendRShTrunc(N, DAG))
    return Urshr;

  if (SDValue PreCast = isNVCastToHalfWidthElements(Op0)) {
    if (SDValue Rshrnb = trySimplifySrlAddToRshrnb(PreCast, DAG, Subtarget)) {
      Rshrnb = DAG.getNode(AArch64ISD::NVCAST, DL, ResVT, Rshrnb);
      return DAG.getNode(AArch64ISD::UZP1, DL, ResVT, Rshrnb, Op1);
    }
  }

  if (SDValue PreCast = isNVCastToHalfWidthElements(Op1)) {
    if (SDValue Rshrnb = trySimplifySrlAddToRshrnb(PreCast, DAG, Subtarget)) {
      Rshrnb = DAG.getNode(AArch64ISD::NVCAST, DL, ResVT, Rshrnb);
      return DAG.getNode(AArch64ISD::UZP1, DL, ResVT, Op0, Rshrnb);
    }
  }

  // uzp1<ty>(nvcast(unpklo(uzp1<ty>(x, y))), z) => uzp1<ty>(x, z)
  if (SDValue PreCast = isNVCastToHalfWidthElements(Op0)) {
    if (PreCast.getOpcode() == AArch64ISD::UUNPKLO) {
      if (PreCast.getOperand(0).getOpcode() == AArch64ISD::UZP1) {
        SDValue X = PreCast.getOperand(0).getOperand(0);
        return DAG.getNode(AArch64ISD::UZP1, DL, ResVT, X, Op1);
      }
    }
  }

  // uzp1<ty>(x, nvcast(unpkhi(uzp1<ty>(y, z)))) => uzp1<ty>(x, z)
  if (SDValue PreCast = isNVCastToHalfWidthElements(Op1)) {
    if (PreCast.getOpcode() == AArch64ISD::UUNPKHI) {
      if (PreCast.getOperand(0).getOpcode() == AArch64ISD::UZP1) {
        SDValue Z = PreCast.getOperand(0).getOperand(1);
        return DAG.getNode(AArch64ISD::UZP1, DL, ResVT, Op0, Z);
      }
    }
  }

  // These optimizations only work on little endian.
  if (!DAG.getDataLayout().isLittleEndian())
    return SDValue();

  // uzp1(bitcast(x), bitcast(y)) -> uzp1(x, y)
  // Example:
  // nxv4i32 = uzp1 bitcast(nxv4i32 x to nxv2i64), bitcast(nxv4i32 y to nxv2i64)
  // to
  // nxv4i32 = uzp1 nxv4i32 x, nxv4i32 y
  if (isHalvingTruncateAndConcatOfLegalIntScalableType(N) &&
      Op0.getOpcode() == ISD::BITCAST && Op1.getOpcode() == ISD::BITCAST) {
    if (Op0.getOperand(0).getValueType() == Op1.getOperand(0).getValueType()) {
      return DAG.getNode(AArch64ISD::UZP1, DL, ResVT, Op0.getOperand(0),
                         Op1.getOperand(0));
    }
  }

  if (ResVT != MVT::v2i32 && ResVT != MVT::v4i16 && ResVT != MVT::v8i8)
    return SDValue();

  SDValue SourceOp0 = peekThroughBitcasts(Op0);
  SDValue SourceOp1 = peekThroughBitcasts(Op1);

  // truncating uzp1(x, y) -> xtn(concat (x, y))
  if (SourceOp0.getValueType() == SourceOp1.getValueType()) {
    EVT Op0Ty = SourceOp0.getValueType();
    if ((ResVT == MVT::v4i16 && Op0Ty == MVT::v2i32) ||
        (ResVT == MVT::v8i8 && Op0Ty == MVT::v4i16)) {
      SDValue Concat =
          DAG.getNode(ISD::CONCAT_VECTORS, DL,
                      Op0Ty.getDoubleNumVectorElementsVT(*DAG.getContext()),
                      SourceOp0, SourceOp1);
      return DAG.getNode(ISD::TRUNCATE, DL, ResVT, Concat);
    }
  }

  // uzp1(xtn x, xtn y) -> xtn(uzp1 (x, y))
  if (SourceOp0.getOpcode() != ISD::TRUNCATE ||
      SourceOp1.getOpcode() != ISD::TRUNCATE)
    return SDValue();
  SourceOp0 = SourceOp0.getOperand(0);
  SourceOp1 = SourceOp1.getOperand(0);

  if (SourceOp0.getValueType() != SourceOp1.getValueType() ||
      !SourceOp0.getValueType().isSimple())
    return SDValue();

  EVT ResultTy;

  switch (SourceOp0.getSimpleValueType().SimpleTy) {
  case MVT::v2i64:
    ResultTy = MVT::v4i32;
    break;
  case MVT::v4i32:
    ResultTy = MVT::v8i16;
    break;
  case MVT::v8i16:
    ResultTy = MVT::v16i8;
    break;
  default:
    return SDValue();
  }

  SDValue UzpOp0 = DAG.getNode(ISD::BITCAST, DL, ResultTy, SourceOp0);
  SDValue UzpOp1 = DAG.getNode(ISD::BITCAST, DL, ResultTy, SourceOp1);
  SDValue UzpResult =
      DAG.getNode(AArch64ISD::UZP1, DL, UzpOp0.getValueType(), UzpOp0, UzpOp1);

  EVT BitcastResultTy;

  switch (ResVT.getSimpleVT().SimpleTy) {
  case MVT::v2i32:
    BitcastResultTy = MVT::v2i64;
    break;
  case MVT::v4i16:
    BitcastResultTy = MVT::v4i32;
    break;
  case MVT::v8i8:
    BitcastResultTy = MVT::v8i16;
    break;
  default:
    llvm_unreachable("Should be one of {v2i32, v4i16, v8i8}");
  }

  return DAG.getNode(ISD::TRUNCATE, DL, ResVT,
                     DAG.getNode(ISD::BITCAST, DL, BitcastResultTy, UzpResult));
}

static SDValue performGLD1Combine(SDNode *N, SelectionDAG &DAG) {
  unsigned Opc = N->getOpcode();

  const bool Scaled = Opc == AArch64ISD::GLD1_SCALED_MERGE_ZERO ||
                      Opc == AArch64ISD::GLD1S_SCALED_MERGE_ZERO;
  const bool Signed = Opc == AArch64ISD::GLD1S_MERGE_ZERO ||
                      Opc == AArch64ISD::GLD1S_SCALED_MERGE_ZERO;
  const bool Extended = Opc == AArch64ISD::GLD1_SXTW_MERGE_ZERO ||
                        Opc == AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO ||
                        Opc == AArch64ISD::GLD1_UXTW_MERGE_ZERO ||
                        Opc == AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO;

  SDLoc DL(N);
  SDValue Chain = N->getOperand(0);
  SDValue Pg = N->getOperand(1);
  SDValue Base = N->getOperand(2);
  SDValue Offset = N->getOperand(3);
  SDValue Ty = N->getOperand(4);

  EVT ResVT = N->getValueType(0);

  const auto OffsetOpc = Offset.getOpcode();
  const bool OffsetIsZExt =
      OffsetOpc == AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU;
  const bool OffsetIsSExt =
      OffsetOpc == AArch64ISD::SIGN_EXTEND_INREG_MERGE_PASSTHRU;

  // Fold sign/zero extensions of vector offsets into GLD1 nodes where possible.
  if (!Extended && (OffsetIsSExt || OffsetIsZExt)) {
    SDValue ExtPg = Offset.getOperand(0);
    VTSDNode *ExtFrom = cast<VTSDNode>(Offset.getOperand(2).getNode());
    EVT ExtFromEVT = ExtFrom->getVT().getVectorElementType();

    // If the predicate for the sign- or zero-extended offset is the
    // same as the predicate used for this load and the sign-/zero-extension
    // was from a 32-bits...
    if (ExtPg == Pg && ExtFromEVT == MVT::i32) {
      SDValue UnextendedOffset = Offset.getOperand(1);

      unsigned NewOpc = getGatherVecOpcode(Scaled, OffsetIsSExt, true);
      if (Signed)
        NewOpc = getSignExtendedGatherOpcode(NewOpc);

      return DAG.getNode(NewOpc, DL, {ResVT, MVT::Other},
                         {Chain, Pg, Base, UnextendedOffset, Ty});
    }
  }

  return SDValue();
}

/// Optimize a vector shift instruction and its operand if shifted out
/// bits are not used.
static SDValue performVectorShiftCombine(SDNode *N,
                                         const AArch64TargetLowering &TLI,
                                         TargetLowering::DAGCombinerInfo &DCI) {
  assert(N->getOpcode() == AArch64ISD::VASHR ||
         N->getOpcode() == AArch64ISD::VLSHR);

  SDValue Op = N->getOperand(0);
  unsigned OpScalarSize = Op.getScalarValueSizeInBits();

  unsigned ShiftImm = N->getConstantOperandVal(1);
  assert(OpScalarSize > ShiftImm && "Invalid shift imm");

  // Remove sign_extend_inreg (ashr(shl(x)) based on the number of sign bits.
  if (N->getOpcode() == AArch64ISD::VASHR &&
      Op.getOpcode() == AArch64ISD::VSHL &&
      N->getOperand(1) == Op.getOperand(1))
    if (DCI.DAG.ComputeNumSignBits(Op.getOperand(0)) > ShiftImm)
      return Op.getOperand(0);

  // If the shift is exact, the shifted out bits matter.
  if (N->getFlags().hasExact())
    return SDValue();

  APInt ShiftedOutBits = APInt::getLowBitsSet(OpScalarSize, ShiftImm);
  APInt DemandedMask = ~ShiftedOutBits;

  if (TLI.SimplifyDemandedBits(Op, DemandedMask, DCI))
    return SDValue(N, 0);

  return SDValue();
}

static SDValue performSunpkloCombine(SDNode *N, SelectionDAG &DAG) {
  // sunpklo(sext(pred)) -> sext(extract_low_half(pred))
  // This transform works in partnership with performSetCCPunpkCombine to
  // remove unnecessary transfer of predicates into standard registers and back
  if (N->getOperand(0).getOpcode() == ISD::SIGN_EXTEND &&
      N->getOperand(0)->getOperand(0)->getValueType(0).getScalarType() ==
          MVT::i1) {
    SDValue CC = N->getOperand(0)->getOperand(0);
    auto VT = CC->getValueType(0).getHalfNumVectorElementsVT(*DAG.getContext());
    SDValue Unpk = DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(N), VT, CC,
                               DAG.getVectorIdxConstant(0, SDLoc(N)));
    return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), N->getValueType(0), Unpk);
  }

  return SDValue();
}

/// Target-specific DAG combine function for post-increment LD1 (lane) and
/// post-increment LD1R.
static SDValue performPostLD1Combine(SDNode *N,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     bool IsLaneOp) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  SelectionDAG &DAG = DCI.DAG;
  EVT VT = N->getValueType(0);

  if (!VT.is128BitVector() && !VT.is64BitVector())
    return SDValue();

  // If it is not LOAD, can not do such combine.
  unsigned LoadIdx = IsLaneOp ? 1 : 0;
  LoadSDNode *LD = dyn_cast<LoadSDNode>(N->getOperand(LoadIdx).getNode());
  if (!LD)
    return SDValue();

  // If the Generic combiner already helped form a pre- or post-indexed load,
  // skip forming one here.
  if (LD->isIndexed())
    return SDValue();

  // The vector lane must be a constant in the LD1LANE opcode.
  SDValue Lane;
  if (IsLaneOp) {
    Lane = N->getOperand(2);
    auto *LaneC = dyn_cast<ConstantSDNode>(Lane);
    if (!LaneC || LaneC->getZExtValue() >= VT.getVectorNumElements())
      return SDValue();
    if (LaneC->getZExtValue() == 0 && isNullOrNullSplat(N->getOperand(0)))
      return SDValue();
  }

  LoadSDNode *LoadSDN = cast<LoadSDNode>(LD);
  EVT MemVT = LoadSDN->getMemoryVT();
  // Check if memory operand is the same type as the vector element.
  if (MemVT != VT.getVectorElementType())
    return SDValue();

  // Check if there are other uses. If so, do not combine as it will introduce
  // an extra load.
  for (SDUse &U : LD->uses()) {
    if (U.getResNo() == 1) // Ignore uses of the chain result.
      continue;
    if (U.getUser() != N)
      return SDValue();
  }

  // If there is one use and it can splat the value, prefer that operation.
  // TODO: This could be expanded to more operations if they reliably use the
  // index variants.
  if (N->hasOneUse()) {
    unsigned UseOpc = N->user_begin()->getOpcode();
    if (UseOpc == ISD::FMUL || UseOpc == ISD::FMA)
      return SDValue();
  }

  SDValue Addr = LD->getOperand(1);
  SDValue Vector = N->getOperand(0);
  // Search for a use of the address operand that is an increment.
  for (SDUse &Use : Addr->uses()) {
    SDNode *User = Use.getUser();
    if (User->getOpcode() != ISD::ADD || Use.getResNo() != Addr.getResNo())
      continue;

    // If the increment is a constant, it must match the memory ref size.
    SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
    if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
      uint32_t IncVal = CInc->getZExtValue();
      unsigned NumBytes = VT.getScalarSizeInBits() / 8;
      if (IncVal != NumBytes)
        continue;
      Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
    }

    // To avoid cycle construction make sure that neither the load nor the add
    // are predecessors to each other or the Vector.
    SmallPtrSet<const SDNode *, 32> Visited;
    SmallVector<const SDNode *, 16> Worklist;
    Visited.insert(Addr.getNode());
    Worklist.push_back(User);
    Worklist.push_back(LD);
    Worklist.push_back(Vector.getNode());
    if (SDNode::hasPredecessorHelper(LD, Visited, Worklist) ||
        SDNode::hasPredecessorHelper(User, Visited, Worklist))
      continue;

    SmallVector<SDValue, 8> Ops;
    Ops.push_back(LD->getOperand(0));  // Chain
    if (IsLaneOp) {
      Ops.push_back(Vector);           // The vector to be inserted
      Ops.push_back(Lane);             // The lane to be inserted in the vector
    }
    Ops.push_back(Addr);
    Ops.push_back(Inc);

    EVT Tys[3] = { VT, MVT::i64, MVT::Other };
    SDVTList SDTys = DAG.getVTList(Tys);
    unsigned NewOp = IsLaneOp ? AArch64ISD::LD1LANEpost : AArch64ISD::LD1DUPpost;
    SDValue UpdN = DAG.getMemIntrinsicNode(NewOp, SDLoc(N), SDTys, Ops,
                                           MemVT,
                                           LoadSDN->getMemOperand());

    // Update the uses.
    SDValue NewResults[] = {
        SDValue(LD, 0),            // The result of load
        SDValue(UpdN.getNode(), 2) // Chain
    };
    DCI.CombineTo(LD, NewResults);
    DCI.CombineTo(N, SDValue(UpdN.getNode(), 0));     // Dup/Inserted Result
    DCI.CombineTo(User, SDValue(UpdN.getNode(), 1));  // Write back register

    break;
  }
  return SDValue();
}

/// Simplify ``Addr`` given that the top byte of it is ignored by HW during
/// address translation.
static bool performTBISimplification(SDValue Addr,
                                     TargetLowering::DAGCombinerInfo &DCI,
                                     SelectionDAG &DAG) {
  APInt DemandedMask = APInt::getLowBitsSet(64, 56);
  KnownBits Known;
  TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
                                        !DCI.isBeforeLegalizeOps());
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  if (TLI.SimplifyDemandedBits(Addr, DemandedMask, Known, TLO)) {
    DCI.CommitTargetLoweringOpt(TLO);
    return true;
  }
  return false;
}

static SDValue foldTruncStoreOfExt(SelectionDAG &DAG, SDNode *N) {
  assert((N->getOpcode() == ISD::STORE || N->getOpcode() == ISD::MSTORE) &&
         "Expected STORE dag node in input!");

  if (auto Store = dyn_cast<StoreSDNode>(N)) {
    if (!Store->isTruncatingStore() || Store->isIndexed())
      return SDValue();
    SDValue Ext = Store->getValue();
    auto ExtOpCode = Ext.getOpcode();
    if (ExtOpCode != ISD::ZERO_EXTEND && ExtOpCode != ISD::SIGN_EXTEND &&
        ExtOpCode != ISD::ANY_EXTEND)
      return SDValue();
    SDValue Orig = Ext->getOperand(0);
    if (Store->getMemoryVT() != Orig.getValueType())
      return SDValue();
    return DAG.getStore(Store->getChain(), SDLoc(Store), Orig,
                        Store->getBasePtr(), Store->getMemOperand());
  }

  return SDValue();
}

// A custom combine to lower load <3 x i8> as the more efficient sequence
// below:
//    ldrb wX, [x0, #2]
//    ldrh wY, [x0]
//    orr wX, wY, wX, lsl #16
//    fmov s0, wX
//
// Note that an alternative sequence with even fewer (although usually more
// complex/expensive) instructions would be:
//   ld1r.4h { v0 }, [x0], #2
//   ld1.b { v0 }[2], [x0]
//
// Generating this sequence unfortunately results in noticeably worse codegen
// for code that extends the loaded v3i8, due to legalization breaking vector
// shuffle detection in a way that is very difficult to work around.
// TODO: Revisit once v3i8 legalization has been improved in general.
static SDValue combineV3I8LoadExt(LoadSDNode *LD, SelectionDAG &DAG) {
  EVT MemVT = LD->getMemoryVT();
  if (MemVT != EVT::getVectorVT(*DAG.getContext(), MVT::i8, 3) ||
      LD->getBaseAlign() >= 4)
    return SDValue();

  SDLoc DL(LD);
  MachineFunction &MF = DAG.getMachineFunction();
  SDValue Chain = LD->getChain();
  SDValue BasePtr = LD->getBasePtr();
  MachineMemOperand *MMO = LD->getMemOperand();
  assert(LD->getOffset().isUndef() && "undef offset expected");

  // Load 2 x i8, then 1 x i8.
  SDValue L16 = DAG.getLoad(MVT::i16, DL, Chain, BasePtr, MMO);
  TypeSize Offset2 = TypeSize::getFixed(2);
  SDValue L8 = DAG.getLoad(MVT::i8, DL, Chain,
                           DAG.getMemBasePlusOffset(BasePtr, Offset2, DL),
                           MF.getMachineMemOperand(MMO, 2, 1));

  // Extend to i32.
  SDValue Ext16 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, L16);
  SDValue Ext8 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, L8);

  // Pack 2 x i8 and 1 x i8 in an i32 and convert to v4i8.
  SDValue Shl = DAG.getNode(ISD::SHL, DL, MVT::i32, Ext8,
                            DAG.getConstant(16, DL, MVT::i32));
  SDValue Or = DAG.getNode(ISD::OR, DL, MVT::i32, Ext16, Shl);
  SDValue Cast = DAG.getNode(ISD::BITCAST, DL, MVT::v4i8, Or);

  // Extract v3i8 again.
  SDValue Extract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MemVT, Cast,
                                DAG.getConstant(0, DL, MVT::i64));
  SDValue TokenFactor = DAG.getNode(
      ISD::TokenFactor, DL, MVT::Other,
      {SDValue(cast<SDNode>(L16), 1), SDValue(cast<SDNode>(L8), 1)});
  return DAG.getMergeValues({Extract, TokenFactor}, DL);
}

// Perform TBI simplification if supported by the target and try to break up
// nontemporal loads larger than 256-bits loads for odd types so LDNPQ 256-bit
// load instructions can be selected.
static SDValue performLOADCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  SelectionDAG &DAG,
                                  const AArch64Subtarget *Subtarget) {
  if (Subtarget->supportsAddressTopByteIgnored())
    performTBISimplification(N->getOperand(1), DCI, DAG);

  LoadSDNode *LD = cast<LoadSDNode>(N);
  EVT RegVT = LD->getValueType(0);
  EVT MemVT = LD->getMemoryVT();
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  SDLoc DL(LD);

  // Cast ptr32 and ptr64 pointers to the default address space before a load.
  unsigned AddrSpace = LD->getAddressSpace();
  if (AddrSpace == ARM64AS::PTR64 || AddrSpace == ARM64AS::PTR32_SPTR ||
      AddrSpace == ARM64AS::PTR32_UPTR) {
    MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
    if (PtrVT != LD->getBasePtr().getSimpleValueType()) {
      SDValue Cast =
          DAG.getAddrSpaceCast(DL, PtrVT, LD->getBasePtr(), AddrSpace, 0);
      return DAG.getExtLoad(LD->getExtensionType(), DL, RegVT, LD->getChain(),
                            Cast, LD->getPointerInfo(), MemVT,
                            LD->getBaseAlign(),
                            LD->getMemOperand()->getFlags());
    }
  }

  if (LD->isVolatile() || !Subtarget->isLittleEndian())
    return SDValue(N, 0);

  if (SDValue Res = combineV3I8LoadExt(LD, DAG))
    return Res;

  if (!LD->isNonTemporal())
    return SDValue(N, 0);

  if (MemVT.isScalableVector() || MemVT.getSizeInBits() <= 256 ||
      MemVT.getSizeInBits() % 256 == 0 ||
      256 % MemVT.getScalarSizeInBits() != 0)
    return SDValue(N, 0);

  SDValue Chain = LD->getChain();
  SDValue BasePtr = LD->getBasePtr();
  SDNodeFlags Flags = LD->getFlags();
  SmallVector<SDValue, 4> LoadOps;
  SmallVector<SDValue, 4> LoadOpsChain;
  // Replace any non temporal load over 256-bit with a series of 256 bit loads
  // and a scalar/vector load less than 256. This way we can utilize 256-bit
  // loads and reduce the amount of load instructions generated.
  MVT NewVT =
      MVT::getVectorVT(MemVT.getVectorElementType().getSimpleVT(),
                       256 / MemVT.getVectorElementType().getSizeInBits());
  unsigned Num256Loads = MemVT.getSizeInBits() / 256;
  // Create all 256-bit loads starting from offset 0 and up to Num256Loads-1*32.
  for (unsigned I = 0; I < Num256Loads; I++) {
    unsigned PtrOffset = I * 32;
    SDValue NewPtr = DAG.getMemBasePlusOffset(
        BasePtr, TypeSize::getFixed(PtrOffset), DL, Flags);
    Align NewAlign = commonAlignment(LD->getAlign(), PtrOffset);
    SDValue NewLoad = DAG.getLoad(
        NewVT, DL, Chain, NewPtr, LD->getPointerInfo().getWithOffset(PtrOffset),
        NewAlign, LD->getMemOperand()->getFlags(), LD->getAAInfo());
    LoadOps.push_back(NewLoad);
    LoadOpsChain.push_back(SDValue(cast<SDNode>(NewLoad), 1));
  }

  // Process remaining bits of the load operation.
  // This is done by creating an UNDEF vector to match the size of the
  // 256-bit loads and inserting the remaining load to it. We extract the
  // original load type at the end using EXTRACT_SUBVECTOR instruction.
  unsigned BitsRemaining = MemVT.getSizeInBits() % 256;
  unsigned PtrOffset = (MemVT.getSizeInBits() - BitsRemaining) / 8;
  MVT RemainingVT = MVT::getVectorVT(
      MemVT.getVectorElementType().getSimpleVT(),
      BitsRemaining / MemVT.getVectorElementType().getSizeInBits());
  SDValue NewPtr = DAG.getMemBasePlusOffset(
      BasePtr, TypeSize::getFixed(PtrOffset), DL, Flags);
  Align NewAlign = commonAlignment(LD->getAlign(), PtrOffset);
  SDValue RemainingLoad =
      DAG.getLoad(RemainingVT, DL, Chain, NewPtr,
                  LD->getPointerInfo().getWithOffset(PtrOffset), NewAlign,
                  LD->getMemOperand()->getFlags(), LD->getAAInfo());
  SDValue UndefVector = DAG.getUNDEF(NewVT);
  SDValue InsertIdx = DAG.getVectorIdxConstant(0, DL);
  SDValue ExtendedRemainingLoad =
      DAG.getNode(ISD::INSERT_SUBVECTOR, DL, NewVT,
                  {UndefVector, RemainingLoad, InsertIdx});
  LoadOps.push_back(ExtendedRemainingLoad);
  LoadOpsChain.push_back(SDValue(cast<SDNode>(RemainingLoad), 1));
  EVT ConcatVT =
      EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
                       LoadOps.size() * NewVT.getVectorNumElements());
  SDValue ConcatVectors =
      DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatVT, LoadOps);
  // Extract the original vector type size.
  SDValue ExtractSubVector =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MemVT,
                  {ConcatVectors, DAG.getVectorIdxConstant(0, DL)});
  SDValue TokenFactor =
      DAG.getNode(ISD::TokenFactor, DL, MVT::Other, LoadOpsChain);
  return DAG.getMergeValues({ExtractSubVector, TokenFactor}, DL);
}

static EVT tryGetOriginalBoolVectorType(SDValue Op, int Depth = 0) {
  EVT VecVT = Op.getValueType();
  assert(VecVT.isVector() && VecVT.getVectorElementType() == MVT::i1 &&
         "Need boolean vector type.");

  if (Depth > 3)
    return MVT::INVALID_SIMPLE_VALUE_TYPE;

  // We can get the base type from a vector compare or truncate.
  if (Op.getOpcode() == ISD::SETCC || Op.getOpcode() == ISD::TRUNCATE)
    return Op.getOperand(0).getValueType();

  // If an operand is a bool vector, continue looking.
  EVT BaseVT = MVT::INVALID_SIMPLE_VALUE_TYPE;
  for (SDValue Operand : Op->op_values()) {
    if (Operand.getValueType() != VecVT)
      continue;

    EVT OperandVT = tryGetOriginalBoolVectorType(Operand, Depth + 1);
    if (!BaseVT.isSimple())
      BaseVT = OperandVT;
    else if (OperandVT != BaseVT)
      return MVT::INVALID_SIMPLE_VALUE_TYPE;
  }

  return BaseVT;
}

// When converting a <N x iX> vector to <N x i1> to store or use as a scalar
// iN, we can use a trick that extracts the i^th bit from the i^th element and
// then performs a vector add to get a scalar bitmask. This requires that each
// element's bits are either all 1 or all 0.
static SDValue vectorToScalarBitmask(SDNode *N, SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue ComparisonResult(N, 0);
  EVT VecVT = ComparisonResult.getValueType();
  assert(VecVT.isVector() && "Must be a vector type");

  unsigned NumElts = VecVT.getVectorNumElements();
  if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
    return SDValue();

  if (VecVT.getVectorElementType() != MVT::i1 &&
      !DAG.getTargetLoweringInfo().isTypeLegal(VecVT))
    return SDValue();

  // If we can find the original types to work on instead of a vector of i1,
  // we can avoid extend/extract conversion instructions.
  if (VecVT.getVectorElementType() == MVT::i1) {
    VecVT = tryGetOriginalBoolVectorType(ComparisonResult);
    if (!VecVT.isSimple()) {
      unsigned BitsPerElement = std::max(64 / NumElts, 8u); // >= 64-bit vector
      VecVT = MVT::getVectorVT(MVT::getIntegerVT(BitsPerElement), NumElts);
    }
  }
  VecVT = VecVT.changeVectorElementTypeToInteger();

  // Large vectors don't map directly to this conversion, so to avoid too many
  // edge cases, we don't apply it here. The conversion will likely still be
  // applied later via multiple smaller vectors, whose results are concatenated.
  if (VecVT.getSizeInBits() > 128)
    return SDValue();

  // Ensure that all elements' bits are either 0s or 1s.
  ComparisonResult = DAG.getSExtOrTrunc(ComparisonResult, DL, VecVT);

  SmallVector<SDValue, 16> MaskConstants;
  if (DAG.getSubtarget<AArch64Subtarget>().isNeonAvailable() &&
      VecVT == MVT::v16i8) {
    // v16i8 is a special case, as we have 16 entries but only 8 positional bits
    // per entry. We split it into two halves, apply the mask, zip the halves to
    // create 8x 16-bit values, and the perform the vector reduce.
    for (unsigned Half = 0; Half < 2; ++Half) {
      for (unsigned MaskBit = 1; MaskBit <= 128; MaskBit *= 2) {
        MaskConstants.push_back(DAG.getConstant(MaskBit, DL, MVT::i32));
      }
    }
    SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, MaskConstants);
    SDValue RepresentativeBits =
        DAG.getNode(ISD::AND, DL, VecVT, ComparisonResult, Mask);

    SDValue UpperRepresentativeBits =
        DAG.getNode(AArch64ISD::EXT, DL, VecVT, RepresentativeBits,
                    RepresentativeBits, DAG.getConstant(8, DL, MVT::i32));
    SDValue Zipped = DAG.getNode(AArch64ISD::ZIP1, DL, VecVT,
                                 RepresentativeBits, UpperRepresentativeBits);
    Zipped = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Zipped);
    return DAG.getNode(ISD::VECREDUCE_ADD, DL, MVT::i16, Zipped);
  }

  // All other vector sizes.
  unsigned MaxBitMask = 1u << (VecVT.getVectorNumElements() - 1);
  for (unsigned MaskBit = 1; MaskBit <= MaxBitMask; MaskBit *= 2) {
    MaskConstants.push_back(DAG.getConstant(MaskBit, DL, MVT::i64));
  }

  SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecVT, MaskConstants);
  SDValue RepresentativeBits =
      DAG.getNode(ISD::AND, DL, VecVT, ComparisonResult, Mask);
  EVT ResultVT = MVT::getIntegerVT(std::max<unsigned>(
      NumElts, VecVT.getVectorElementType().getSizeInBits()));
  return DAG.getNode(ISD::VECREDUCE_ADD, DL, ResultVT, RepresentativeBits);
}

static SDValue combineBoolVectorAndTruncateStore(SelectionDAG &DAG,
                                                 StoreSDNode *Store) {
  if (!Store->isTruncatingStore())
    return SDValue();

  SDLoc DL(Store);
  SDValue VecOp = Store->getValue();
  EVT VT = VecOp.getValueType();
  EVT MemVT = Store->getMemoryVT();

  if (!MemVT.isVector() || !VT.isVector() ||
      MemVT.getVectorElementType() != MVT::i1)
    return SDValue();

  // If we are storing a vector that we are currently building, let
  // `scalarizeVectorStore()` handle this more efficiently.
  if (VecOp.getOpcode() == ISD::BUILD_VECTOR)
    return SDValue();

  VecOp = DAG.getNode(ISD::TRUNCATE, DL, MemVT, VecOp);
  SDValue VectorBits = vectorToScalarBitmask(VecOp.getNode(), DAG);
  if (!VectorBits)
    return SDValue();

  EVT StoreVT =
      EVT::getIntegerVT(*DAG.getContext(), MemVT.getStoreSizeInBits());
  SDValue ExtendedBits = DAG.getZExtOrTrunc(VectorBits, DL, StoreVT);
  return DAG.getStore(Store->getChain(), DL, ExtendedBits, Store->getBasePtr(),
                      Store->getMemOperand());
}

bool isHalvingTruncateOfLegalScalableType(EVT SrcVT, EVT DstVT) {
  return (SrcVT == MVT::nxv8i16 && DstVT == MVT::nxv8i8) ||
         (SrcVT == MVT::nxv4i32 && DstVT == MVT::nxv4i16) ||
         (SrcVT == MVT::nxv2i64 && DstVT == MVT::nxv2i32);
}

// Combine store (trunc X to <3 x i8>) to sequence of ST1.b.
static SDValue combineI8TruncStore(StoreSDNode *ST, SelectionDAG &DAG,
                                   const AArch64Subtarget *Subtarget) {
  SDValue Value = ST->getValue();
  EVT ValueVT = Value.getValueType();

  if (ST->isVolatile() || !Subtarget->isLittleEndian() ||
      Value.getOpcode() != ISD::TRUNCATE ||
      ValueVT != EVT::getVectorVT(*DAG.getContext(), MVT::i8, 3))
    return SDValue();

  assert(ST->getOffset().isUndef() && "undef offset expected");
  SDLoc DL(ST);
  auto WideVT = EVT::getVectorVT(
      *DAG.getContext(),
      Value->getOperand(0).getValueType().getVectorElementType(), 4);
  SDValue UndefVector = DAG.getUNDEF(WideVT);
  SDValue WideTrunc = DAG.getNode(
      ISD::INSERT_SUBVECTOR, DL, WideVT,
      {UndefVector, Value->getOperand(0), DAG.getVectorIdxConstant(0, DL)});
  SDValue Cast = DAG.getNode(
      ISD::BITCAST, DL, WideVT.getSizeInBits() == 64 ? MVT::v8i8 : MVT::v16i8,
      WideTrunc);

  MachineFunction &MF = DAG.getMachineFunction();
  SDValue Chain = ST->getChain();
  MachineMemOperand *MMO = ST->getMemOperand();
  unsigned IdxScale = WideVT.getScalarSizeInBits() / 8;
  SDValue E2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8, Cast,
                           DAG.getConstant(2 * IdxScale, DL, MVT::i64));
  TypeSize Offset2 = TypeSize::getFixed(2);
  SDValue Ptr2 = DAG.getMemBasePlusOffset(ST->getBasePtr(), Offset2, DL);
  Chain = DAG.getStore(Chain, DL, E2, Ptr2, MF.getMachineMemOperand(MMO, 2, 1));

  SDValue E1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8, Cast,
                           DAG.getConstant(1 * IdxScale, DL, MVT::i64));
  TypeSize Offset1 = TypeSize::getFixed(1);
  SDValue Ptr1 = DAG.getMemBasePlusOffset(ST->getBasePtr(), Offset1, DL);
  Chain = DAG.getStore(Chain, DL, E1, Ptr1, MF.getMachineMemOperand(MMO, 1, 1));

  SDValue E0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i8, Cast,
                           DAG.getConstant(0, DL, MVT::i64));
  Chain = DAG.getStore(Chain, DL, E0, ST->getBasePtr(),
                       MF.getMachineMemOperand(MMO, 0, 1));
  return Chain;
}

static unsigned getFPSubregForVT(EVT VT) {
  assert(VT.isSimple() && "Expected simple VT");
  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::aarch64mfp8:
    return AArch64::bsub;
  case MVT::f16:
    return AArch64::hsub;
  case MVT::f32:
    return AArch64::ssub;
  case MVT::f64:
    return AArch64::dsub;
  default:
    llvm_unreachable("Unexpected VT!");
  }
}

static SDValue performSTORECombine(SDNode *N,
                                   TargetLowering::DAGCombinerInfo &DCI,
                                   SelectionDAG &DAG,
                                   const AArch64Subtarget *Subtarget) {
  StoreSDNode *ST = cast<StoreSDNode>(N);
  SDValue Chain = ST->getChain();
  SDValue Value = ST->getValue();
  SDValue Ptr = ST->getBasePtr();
  EVT ValueVT = Value.getValueType();
  EVT MemVT = ST->getMemoryVT();
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  SDLoc DL(ST);

  auto hasValidElementTypeForFPTruncStore = [](EVT VT) {
    EVT EltVT = VT.getVectorElementType();
    return EltVT == MVT::f32 || EltVT == MVT::f64;
  };

  // Cast ptr32 and ptr64 pointers to the default address space before a store.
  unsigned AddrSpace = ST->getAddressSpace();
  if (AddrSpace == ARM64AS::PTR64 || AddrSpace == ARM64AS::PTR32_SPTR ||
      AddrSpace == ARM64AS::PTR32_UPTR) {
    MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
    if (PtrVT != Ptr.getSimpleValueType()) {
      SDValue Cast = DAG.getAddrSpaceCast(DL, PtrVT, Ptr, AddrSpace, 0);
      return DAG.getStore(Chain, DL, Value, Cast, ST->getPointerInfo(),
                          ST->getBaseAlign(), ST->getMemOperand()->getFlags(),
                          ST->getAAInfo());
    }
  }

  if (SDValue Res = combineI8TruncStore(ST, DAG, Subtarget))
    return Res;

  // If this is an FP_ROUND followed by a store, fold this into a truncating
  // store. We can do this even if this is already a truncstore.
  // We purposefully don't care about legality of the nodes here as we know
  // they can be split down into something legal.
  if (DCI.isBeforeLegalizeOps() && Value.getOpcode() == ISD::FP_ROUND &&
      Value.getNode()->hasOneUse() && ST->isUnindexed() &&
      Subtarget->useSVEForFixedLengthVectors() &&
      ValueVT.isFixedLengthVector() &&
      ValueVT.getFixedSizeInBits() >= Subtarget->getMinSVEVectorSizeInBits() &&
      hasValidElementTypeForFPTruncStore(Value.getOperand(0).getValueType()))
    return DAG.getTruncStore(Chain, DL, Value.getOperand(0), Ptr, MemVT,
                             ST->getMemOperand());

  if (SDValue Split = splitStores(N, DCI, DAG, Subtarget))
    return Split;

  if (Subtarget->supportsAddressTopByteIgnored() &&
      performTBISimplification(N->getOperand(2), DCI, DAG))
    return SDValue(N, 0);

  if (SDValue Store = foldTruncStoreOfExt(DAG, N))
    return Store;

  if (SDValue Store = combineBoolVectorAndTruncateStore(DAG, ST))
    return Store;

  if (ST->isTruncatingStore() &&
      isHalvingTruncateOfLegalScalableType(ValueVT, MemVT)) {
    if (SDValue Rshrnb =
            trySimplifySrlAddToRshrnb(ST->getOperand(1), DAG, Subtarget)) {
      return DAG.getTruncStore(ST->getChain(), ST, Rshrnb, ST->getBasePtr(),
                               MemVT, ST->getMemOperand());
    }
  }

  // This is an integer vector_extract_elt followed by a (possibly truncating)
  // store. We may be able to replace this with a store of an FP subregister.
  if (DCI.isAfterLegalizeDAG() && ST->isUnindexed() &&
      Value.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {

    SDValue Vector = Value.getOperand(0);
    SDValue ExtIdx = Value.getOperand(1);
    EVT VectorVT = Vector.getValueType();
    EVT ElemVT = VectorVT.getVectorElementType();

    if (!ValueVT.isInteger())
      return SDValue();

    // Propagate zero constants (applying this fold may miss optimizations).
    if (ISD::isConstantSplatVectorAllZeros(Vector.getNode())) {
      SDValue ZeroElt = DAG.getConstant(0, DL, ValueVT);
      DAG.ReplaceAllUsesWith(Value, ZeroElt);
      return SDValue();
    }

    if (ValueVT != MemVT && !ST->isTruncatingStore())
      return SDValue();

    // This could generate an additional extract if the index is non-zero and
    // the extracted value has multiple uses.
    auto *ExtCst = dyn_cast<ConstantSDNode>(ExtIdx);
    if ((!ExtCst || !ExtCst->isZero()) && !Value.hasOneUse())
      return SDValue();

    // These can lower to st1, which is preferable if we're unlikely to fold the
    // addressing into the store.
    if (Subtarget->isNeonAvailable() && ElemVT == MemVT &&
        (VectorVT.is64BitVector() || VectorVT.is128BitVector()) && ExtCst &&
        !ExtCst->isZero() && ST->getBasePtr().getOpcode() != ISD::ADD)
      return SDValue();

    if (MemVT == MVT::i64 || MemVT == MVT::i32) {
      // Heuristic: If there are other users of w/x integer scalars extracted
      // from this vector that won't fold into the store -- abandon folding.
      // Applying this fold may disrupt paired stores.
      for (const auto &Use : Vector->uses()) {
        if (Use.getResNo() != Vector.getResNo())
          continue;
        const SDNode *User = Use.getUser();
        if (User->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
            (!User->hasOneUse() ||
             (*User->user_begin())->getOpcode() != ISD::STORE))
          return SDValue();
      }
    }

    SDValue ExtVector = Vector;
    if (!ExtCst || !ExtCst->isZero()) {
      // Handle extracting from lanes != 0.
      SDValue Ext = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
                                Value.getValueType(), Vector, ExtIdx);
      SDValue Zero = DAG.getVectorIdxConstant(0, DL);
      ExtVector = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VectorVT,
                              DAG.getUNDEF(VectorVT), Ext, Zero);
    }

    EVT FPMemVT = MemVT == MVT::i8
                      ? MVT::aarch64mfp8
                      : EVT::getFloatingPointVT(MemVT.getSizeInBits());
    SDValue FPSubreg = DAG.getTargetExtractSubreg(getFPSubregForVT(FPMemVT), DL,
                                                  FPMemVT, ExtVector);

    return DAG.getStore(ST->getChain(), DL, FPSubreg, ST->getBasePtr(),
                        ST->getMemOperand());
  }

  return SDValue();
}

static SDValue performMSTORECombine(SDNode *N,
                                    TargetLowering::DAGCombinerInfo &DCI,
                                    SelectionDAG &DAG,
                                    const AArch64Subtarget *Subtarget) {
  MaskedStoreSDNode *MST = cast<MaskedStoreSDNode>(N);
  SDValue Value = MST->getValue();
  SDValue Mask = MST->getMask();
  SDLoc DL(N);

  // If this is a UZP1 followed by a masked store, fold this into a masked
  // truncating store.  We can do this even if this is already a masked
  // truncstore.
  if (Value.getOpcode() == AArch64ISD::UZP1 && Value->hasOneUse() &&
      MST->isUnindexed() && Mask->getOpcode() == AArch64ISD::PTRUE &&
      Value.getValueType().isInteger()) {
    Value = Value.getOperand(0);
    if (Value.getOpcode() == ISD::BITCAST) {
      EVT HalfVT =
          Value.getValueType().getHalfNumVectorElementsVT(*DAG.getContext());
      EVT InVT = Value.getOperand(0).getValueType();

      if (HalfVT.widenIntegerVectorElementType(*DAG.getContext()) == InVT) {
        unsigned MinSVESize = Subtarget->getMinSVEVectorSizeInBits();
        unsigned PgPattern = Mask->getConstantOperandVal(0);

        // Ensure we can double the size of the predicate pattern
        unsigned NumElts = getNumElementsFromSVEPredPattern(PgPattern);
        if (NumElts && NumElts * InVT.getVectorElementType().getSizeInBits() <=
                           MinSVESize) {
          Mask = getPTrue(DAG, DL, InVT.changeVectorElementType(MVT::i1),
                          PgPattern);
          return DAG.getMaskedStore(MST->getChain(), DL, Value.getOperand(0),
                                    MST->getBasePtr(), MST->getOffset(), Mask,
                                    MST->getMemoryVT(), MST->getMemOperand(),
                                    MST->getAddressingMode(),
                                    /*IsTruncating=*/true);
        }
      }
    }
  }

  if (MST->isTruncatingStore()) {
    EVT ValueVT = Value->getValueType(0);
    EVT MemVT = MST->getMemoryVT();
    if (!isHalvingTruncateOfLegalScalableType(ValueVT, MemVT))
      return SDValue();
    if (SDValue Rshrnb = trySimplifySrlAddToRshrnb(Value, DAG, Subtarget)) {
      return DAG.getMaskedStore(MST->getChain(), DL, Rshrnb, MST->getBasePtr(),
                                MST->getOffset(), MST->getMask(),
                                MST->getMemoryVT(), MST->getMemOperand(),
                                MST->getAddressingMode(), true);
    }
  }

  return SDValue();
}

/// \return true if part of the index was folded into the Base.
static bool foldIndexIntoBase(SDValue &BasePtr, SDValue &Index, SDValue Scale,
                              SDLoc DL, SelectionDAG &DAG) {
  // This function assumes a vector of i64 indices.
  EVT IndexVT = Index.getValueType();
  if (!IndexVT.isVector() || IndexVT.getVectorElementType() != MVT::i64)
    return false;

  // Simplify:
  //   BasePtr = Ptr
  //   Index = X + splat(Offset)
  // ->
  //   BasePtr = Ptr + Offset * scale.
  //   Index = X
  if (Index.getOpcode() == ISD::ADD) {
    if (auto Offset = DAG.getSplatValue(Index.getOperand(1))) {
      Offset = DAG.getNode(ISD::MUL, DL, MVT::i64, Offset, Scale);
      BasePtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr, Offset);
      Index = Index.getOperand(0);
      return true;
    }
  }

  // Simplify:
  //   BasePtr = Ptr
  //   Index = (X + splat(Offset)) << splat(Shift)
  // ->
  //   BasePtr = Ptr + (Offset << Shift) * scale)
  //   Index = X << splat(shift)
  if (Index.getOpcode() == ISD::SHL &&
      Index.getOperand(0).getOpcode() == ISD::ADD) {
    SDValue Add = Index.getOperand(0);
    SDValue ShiftOp = Index.getOperand(1);
    SDValue OffsetOp = Add.getOperand(1);
    if (auto Shift = DAG.getSplatValue(ShiftOp))
      if (auto Offset = DAG.getSplatValue(OffsetOp)) {
        Offset = DAG.getNode(ISD::SHL, DL, MVT::i64, Offset, Shift);
        Offset = DAG.getNode(ISD::MUL, DL, MVT::i64, Offset, Scale);
        BasePtr = DAG.getNode(ISD::ADD, DL, MVT::i64, BasePtr, Offset);
        Index = DAG.getNode(ISD::SHL, DL, Index.getValueType(),
                            Add.getOperand(0), ShiftOp);
        return true;
      }
  }

  return false;
}

// Analyse the specified address returning true if a more optimal addressing
// mode is available. When returning true all parameters are updated to reflect
// their recommended values.
static bool findMoreOptimalIndexType(const MaskedGatherScatterSDNode *N,
                                     SDValue &BasePtr, SDValue &Index,
                                     SelectionDAG &DAG) {
  // Try to iteratively fold parts of the index into the base pointer to
  // simplify the index as much as possible.
  bool Changed = false;
  while (foldIndexIntoBase(BasePtr, Index, N->getScale(), SDLoc(N), DAG))
    Changed = true;

  // Only consider element types that are pointer sized as smaller types can
  // be easily promoted.
  EVT IndexVT = Index.getValueType();
  if (IndexVT.getVectorElementType() != MVT::i64 || IndexVT == MVT::nxv2i64)
    return Changed;

  // Can indices be trivially shrunk?
  EVT DataVT = N->getOperand(1).getValueType();
  // Don't attempt to shrink the index for fixed vectors of 64 bit data since it
  // will later be re-extended to 64 bits in legalization
  if (DataVT.isFixedLengthVector() && DataVT.getScalarSizeInBits() == 64)
    return Changed;
  if (ISD::isVectorShrinkable(Index.getNode(), 32, N->isIndexSigned())) {
    EVT NewIndexVT = IndexVT.changeVectorElementType(MVT::i32);
    Index = DAG.getNode(ISD::TRUNCATE, SDLoc(N), NewIndexVT, Index);
    return true;
  }

  // Match:
  //   Index = step(const)
  int64_t Stride = 0;
  if (Index.getOpcode() == ISD::STEP_VECTOR) {
    Stride = cast<ConstantSDNode>(Index.getOperand(0))->getSExtValue();
  }
  // Match:
  //   Index = step(const) << shift(const)
  else if (Index.getOpcode() == ISD::SHL &&
           Index.getOperand(0).getOpcode() == ISD::STEP_VECTOR) {
    SDValue RHS = Index.getOperand(1);
    if (auto *Shift =
            dyn_cast_or_null<ConstantSDNode>(DAG.getSplatValue(RHS))) {
      int64_t Step = (int64_t)Index.getOperand(0).getConstantOperandVal(1);
      Stride = Step << Shift->getZExtValue();
    }
  }

  // Return early because no supported pattern is found.
  if (Stride == 0)
    return Changed;

  if (Stride < std::numeric_limits<int32_t>::min() ||
      Stride > std::numeric_limits<int32_t>::max())
    return Changed;

  const auto &Subtarget = DAG.getSubtarget<AArch64Subtarget>();
  unsigned MaxVScale =
      Subtarget.getMaxSVEVectorSizeInBits() / AArch64::SVEBitsPerBlock;
  int64_t LastElementOffset =
      IndexVT.getVectorMinNumElements() * Stride * MaxVScale;

  if (LastElementOffset < std::numeric_limits<int32_t>::min() ||
      LastElementOffset > std::numeric_limits<int32_t>::max())
    return Changed;

  EVT NewIndexVT = IndexVT.changeVectorElementType(MVT::i32);
  // Stride does not scale explicitly by 'Scale', because it happens in
  // the gather/scatter addressing mode.
  Index = DAG.getStepVector(SDLoc(N), NewIndexVT, APInt(32, Stride, true));
  return true;
}

static SDValue performMaskedGatherScatterCombine(
    SDNode *N, TargetLowering::DAGCombinerInfo &DCI, SelectionDAG &DAG) {
  if (!DCI.isBeforeLegalize())
    return SDValue();
  MaskedGatherScatterSDNode *MGS = cast<MaskedGatherScatterSDNode>(N);

  SDLoc DL(MGS);
  SDValue Chain = MGS->getChain();
  SDValue Scale = MGS->getScale();
  SDValue Index = MGS->getIndex();
  SDValue Mask = MGS->getMask();
  SDValue BasePtr = MGS->getBasePtr();
  ISD::MemIndexType IndexType = MGS->getIndexType();

  if (!findMoreOptimalIndexType(MGS, BasePtr, Index, DAG))
    return SDValue();

  // Here we catch such cases early and change MGATHER's IndexType to allow
  // the use of an Index that's more legalisation friendly.
  if (auto *MGT = dyn_cast<MaskedGatherSDNode>(MGS)) {
    SDValue PassThru = MGT->getPassThru();
    SDValue Ops[] = {Chain, PassThru, Mask, BasePtr, Index, Scale};
    return DAG.getMaskedGather(
        DAG.getVTList(N->getValueType(0), MVT::Other), MGT->getMemoryVT(), DL,
        Ops, MGT->getMemOperand(), IndexType, MGT->getExtensionType());
  }
  if (auto *MSC = dyn_cast<MaskedScatterSDNode>(MGS)) {
    SDValue Data = MSC->getValue();
    SDValue Ops[] = {Chain, Data, Mask, BasePtr, Index, Scale};
    return DAG.getMaskedScatter(DAG.getVTList(MVT::Other), MSC->getMemoryVT(),
                                DL, Ops, MSC->getMemOperand(), IndexType,
                                MSC->isTruncatingStore());
  }
  auto *HG = cast<MaskedHistogramSDNode>(MGS);
  SDValue Ops[] = {Chain, HG->getInc(), Mask,          BasePtr,
                   Index, Scale,        HG->getIntID()};
  return DAG.getMaskedHistogram(DAG.getVTList(MVT::Other), HG->getMemoryVT(),
                                DL, Ops, HG->getMemOperand(), IndexType);
}

/// Target-specific DAG combine function for NEON load/store intrinsics
/// to merge base address updates.
static SDValue performNEONPostLDSTCombine(SDNode *N,
                                          TargetLowering::DAGCombinerInfo &DCI,
                                          SelectionDAG &DAG) {
  if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
    return SDValue();

  unsigned AddrOpIdx = N->getNumOperands() - 1;
  SDValue Addr = N->getOperand(AddrOpIdx);

  // Search for a use of the address operand that is an increment.
  for (SDUse &Use : Addr->uses()) {
    SDNode *User = Use.getUser();
    if (User->getOpcode() != ISD::ADD || Use.getResNo() != Addr.getResNo())
      continue;

    // Check that the add is independent of the load/store.  Otherwise, folding
    // it would create a cycle.
    SmallPtrSet<const SDNode *, 32> Visited;
    SmallVector<const SDNode *, 16> Worklist;
    Visited.insert(Addr.getNode());
    Worklist.push_back(N);
    Worklist.push_back(User);
    if (SDNode::hasPredecessorHelper(N, Visited, Worklist) ||
        SDNode::hasPredecessorHelper(User, Visited, Worklist))
      continue;

    // Find the new opcode for the updating load/store.
    bool IsStore = false;
    bool IsLaneOp = false;
    bool IsDupOp = false;
    unsigned NewOpc = 0;
    unsigned NumVecs = 0;
    unsigned IntNo = N->getConstantOperandVal(1);
    switch (IntNo) {
    default: llvm_unreachable("unexpected intrinsic for Neon base update");
    case Intrinsic::aarch64_neon_ld2:       NewOpc = AArch64ISD::LD2post;
      NumVecs = 2; break;
    case Intrinsic::aarch64_neon_ld3:       NewOpc = AArch64ISD::LD3post;
      NumVecs = 3; break;
    case Intrinsic::aarch64_neon_ld4:       NewOpc = AArch64ISD::LD4post;
      NumVecs = 4; break;
    case Intrinsic::aarch64_neon_st2:       NewOpc = AArch64ISD::ST2post;
      NumVecs = 2; IsStore = true; break;
    case Intrinsic::aarch64_neon_st3:       NewOpc = AArch64ISD::ST3post;
      NumVecs = 3; IsStore = true; break;
    case Intrinsic::aarch64_neon_st4:       NewOpc = AArch64ISD::ST4post;
      NumVecs = 4; IsStore = true; break;
    case Intrinsic::aarch64_neon_ld1x2:     NewOpc = AArch64ISD::LD1x2post;
      NumVecs = 2; break;
    case Intrinsic::aarch64_neon_ld1x3:     NewOpc = AArch64ISD::LD1x3post;
      NumVecs = 3; break;
    case Intrinsic::aarch64_neon_ld1x4:     NewOpc = AArch64ISD::LD1x4post;
      NumVecs = 4; break;
    case Intrinsic::aarch64_neon_st1x2:     NewOpc = AArch64ISD::ST1x2post;
      NumVecs = 2; IsStore = true; break;
    case Intrinsic::aarch64_neon_st1x3:     NewOpc = AArch64ISD::ST1x3post;
      NumVecs = 3; IsStore = true; break;
    case Intrinsic::aarch64_neon_st1x4:     NewOpc = AArch64ISD::ST1x4post;
      NumVecs = 4; IsStore = true; break;
    case Intrinsic::aarch64_neon_ld2r:      NewOpc = AArch64ISD::LD2DUPpost;
      NumVecs = 2; IsDupOp = true; break;
    case Intrinsic::aarch64_neon_ld3r:      NewOpc = AArch64ISD::LD3DUPpost;
      NumVecs = 3; IsDupOp = true; break;
    case Intrinsic::aarch64_neon_ld4r:      NewOpc = AArch64ISD::LD4DUPpost;
      NumVecs = 4; IsDupOp = true; break;
    case Intrinsic::aarch64_neon_ld2lane:   NewOpc = AArch64ISD::LD2LANEpost;
      NumVecs = 2; IsLaneOp = true; break;
    case Intrinsic::aarch64_neon_ld3lane:   NewOpc = AArch64ISD::LD3LANEpost;
      NumVecs = 3; IsLaneOp = true; break;
    case Intrinsic::aarch64_neon_ld4lane:   NewOpc = AArch64ISD::LD4LANEpost;
      NumVecs = 4; IsLaneOp = true; break;
    case Intrinsic::aarch64_neon_st2lane:   NewOpc = AArch64ISD::ST2LANEpost;
      NumVecs = 2; IsStore = true; IsLaneOp = true; break;
    case Intrinsic::aarch64_neon_st3lane:   NewOpc = AArch64ISD::ST3LANEpost;
      NumVecs = 3; IsStore = true; IsLaneOp = true; break;
    case Intrinsic::aarch64_neon_st4lane:   NewOpc = AArch64ISD::ST4LANEpost;
      NumVecs = 4; IsStore = true; IsLaneOp = true; break;
    }

    EVT VecTy;
    if (IsStore)
      VecTy = N->getOperand(2).getValueType();
    else
      VecTy = N->getValueType(0);

    // If the increment is a constant, it must match the memory ref size.
    SDValue Inc = User->getOperand(User->getOperand(0) == Addr ? 1 : 0);
    if (ConstantSDNode *CInc = dyn_cast<ConstantSDNode>(Inc.getNode())) {
      uint32_t IncVal = CInc->getZExtValue();
      unsigned NumBytes = NumVecs * VecTy.getSizeInBits() / 8;
      if (IsLaneOp || IsDupOp)
        NumBytes /= VecTy.getVectorNumElements();
      if (IncVal != NumBytes)
        continue;
      Inc = DAG.getRegister(AArch64::XZR, MVT::i64);
    }
    SmallVector<SDValue, 8> Ops;
    Ops.push_back(N->getOperand(0)); // Incoming chain
    // Load lane and store have vector list as input.
    if (IsLaneOp || IsStore)
      for (unsigned i = 2; i < AddrOpIdx; ++i)
        Ops.push_back(N->getOperand(i));
    Ops.push_back(Addr); // Base register
    Ops.push_back(Inc);

    // Return Types.
    EVT Tys[6];
    unsigned NumResultVecs = (IsStore ? 0 : NumVecs);
    unsigned n;
    for (n = 0; n < NumResultVecs; ++n)
      Tys[n] = VecTy;
    Tys[n++] = MVT::i64;  // Type of write back register
    Tys[n] = MVT::Other;  // Type of the chain
    SDVTList SDTys = DAG.getVTList(ArrayRef(Tys, NumResultVecs + 2));

    MemIntrinsicSDNode *MemInt = cast<MemIntrinsicSDNode>(N);
    SDValue UpdN = DAG.getMemIntrinsicNode(NewOpc, SDLoc(N), SDTys, Ops,
                                           MemInt->getMemoryVT(),
                                           MemInt->getMemOperand());

    // Update the uses.
    std::vector<SDValue> NewResults;
    for (unsigned i = 0; i < NumResultVecs; ++i) {
      NewResults.push_back(SDValue(UpdN.getNode(), i));
    }
    NewResults.push_back(SDValue(UpdN.getNode(), NumResultVecs + 1));
    DCI.CombineTo(N, NewResults);
    DCI.CombineTo(User, SDValue(UpdN.getNode(), NumResultVecs));

    break;
  }
  return SDValue();
}

// Checks to see if the value is the prescribed width and returns information
// about its extension mode.
static
bool checkValueWidth(SDValue V, unsigned width, ISD::LoadExtType &ExtType) {
  ExtType = ISD::NON_EXTLOAD;
  switch(V.getNode()->getOpcode()) {
  default:
    return false;
  case ISD::LOAD: {
    LoadSDNode *LoadNode = cast<LoadSDNode>(V.getNode());
    if ((LoadNode->getMemoryVT() == MVT::i8 && width == 8)
       || (LoadNode->getMemoryVT() == MVT::i16 && width == 16)) {
      ExtType = LoadNode->getExtensionType();
      return true;
    }
    return false;
  }
  case ISD::AssertSext: {
    VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
    if ((TypeNode->getVT() == MVT::i8 && width == 8)
       || (TypeNode->getVT() == MVT::i16 && width == 16)) {
      ExtType = ISD::SEXTLOAD;
      return true;
    }
    return false;
  }
  case ISD::AssertZext: {
    VTSDNode *TypeNode = cast<VTSDNode>(V.getNode()->getOperand(1));
    if ((TypeNode->getVT() == MVT::i8 && width == 8)
       || (TypeNode->getVT() == MVT::i16 && width == 16)) {
      ExtType = ISD::ZEXTLOAD;
      return true;
    }
    return false;
  }
  case ISD::Constant:
  case ISD::TargetConstant: {
    return std::abs(cast<ConstantSDNode>(V.getNode())->getSExtValue()) <
           1LL << (width - 1);
  }
  }

  return true;
}

// This function does a whole lot of voodoo to determine if the tests are
// equivalent without and with a mask. Essentially what happens is that given a
// DAG resembling:
//
//  +-------------+ +-------------+ +-------------+ +-------------+
//  |    Input    | | AddConstant | | CompConstant| |     CC      |
//  +-------------+ +-------------+ +-------------+ +-------------+
//           |           |           |               |
//           V           V           |    +----------+
//          +-------------+  +----+  |    |
//          |     ADD     |  |0xff|  |    |
//          +-------------+  +----+  |    |
//                  |           |    |    |
//                  V           V    |    |
//                 +-------------+   |    |
//                 |     AND     |   |    |
//                 +-------------+   |    |
//                      |            |    |
//                      +-----+      |    |
//                            |      |    |
//                            V      V    V
//                           +-------------+
//                           |     CMP     |
//                           +-------------+
//
// The AND node may be safely removed for some combinations of inputs. In
// particular we need to take into account the extension type of the Input,
// the exact values of AddConstant, CompConstant, and CC, along with the nominal
// width of the input (this can work for any width inputs, the above graph is
// specific to 8 bits.
//
// The specific equations were worked out by generating output tables for each
// AArch64CC value in terms of and AddConstant (w1), CompConstant(w2). The
// problem was simplified by working with 4 bit inputs, which means we only
// needed to reason about 24 distinct bit patterns: 8 patterns unique to zero
// extension (8,15), 8 patterns unique to sign extensions (-8,-1), and 8
// patterns present in both extensions (0,7). For every distinct set of
// AddConstant and CompConstants bit patterns we can consider the masked and
// unmasked versions to be equivalent if the result of this function is true for
// all 16 distinct bit patterns of for the current extension type of Input (w0).
//
//   sub      w8, w0, w1
//   and      w10, w8, #0x0f
//   cmp      w8, w2
//   cset     w9, AArch64CC
//   cmp      w10, w2
//   cset     w11, AArch64CC
//   cmp      w9, w11
//   cset     w0, eq
//   ret
//
// Since the above function shows when the outputs are equivalent it defines
// when it is safe to remove the AND. Unfortunately it only runs on AArch64 and
// would be expensive to run during compiles. The equations below were written
// in a test harness that confirmed they gave equivalent outputs to the above
// for all inputs function, so they can be used determine if the removal is
// legal instead.
//
// isEquivalentMaskless() is the code for testing if the AND can be removed
// factored out of the DAG recognition as the DAG can take several forms.

static bool isEquivalentMaskless(unsigned CC, unsigned width,
                                 ISD::LoadExtType ExtType, int AddConstant,
                                 int CompConstant) {
  // By being careful about our equations and only writing the in term
  // symbolic values and well known constants (0, 1, -1, MaxUInt) we can
  // make them generally applicable to all bit widths.
  int MaxUInt = (1 << width);

  // For the purposes of these comparisons sign extending the type is
  // equivalent to zero extending the add and displacing it by half the integer
  // width. Provided we are careful and make sure our equations are valid over
  // the whole range we can just adjust the input and avoid writing equations
  // for sign extended inputs.
  if (ExtType == ISD::SEXTLOAD)
    AddConstant -= (1 << (width-1));

  switch(CC) {
  case AArch64CC::LE:
  case AArch64CC::GT:
    if ((AddConstant == 0) ||
        (CompConstant == MaxUInt - 1 && AddConstant < 0) ||
        (AddConstant >= 0 && CompConstant < 0) ||
        (AddConstant <= 0 && CompConstant <= 0 && CompConstant < AddConstant))
      return true;
    break;
  case AArch64CC::LT:
  case AArch64CC::GE:
    if ((AddConstant == 0) ||
        (AddConstant >= 0 && CompConstant <= 0) ||
        (AddConstant <= 0 && CompConstant <= 0 && CompConstant <= AddConstant))
      return true;
    break;
  case AArch64CC::HI:
  case AArch64CC::LS:
    if ((AddConstant >= 0 && CompConstant < 0) ||
       (AddConstant <= 0 && CompConstant >= -1 &&
        CompConstant < AddConstant + MaxUInt))
      return true;
   break;
  case AArch64CC::PL:
  case AArch64CC::MI:
    if ((AddConstant == 0) ||
        (AddConstant > 0 && CompConstant <= 0) ||
        (AddConstant < 0 && CompConstant <= AddConstant))
      return true;
    break;
  case AArch64CC::LO:
  case AArch64CC::HS:
    if ((AddConstant >= 0 && CompConstant <= 0) ||
        (AddConstant <= 0 && CompConstant >= 0 &&
         CompConstant <= AddConstant + MaxUInt))
      return true;
    break;
  case AArch64CC::EQ:
  case AArch64CC::NE:
    if ((AddConstant > 0 && CompConstant < 0) ||
        (AddConstant < 0 && CompConstant >= 0 &&
         CompConstant < AddConstant + MaxUInt) ||
        (AddConstant >= 0 && CompConstant >= 0 &&
         CompConstant >= AddConstant) ||
        (AddConstant <= 0 && CompConstant < 0 && CompConstant < AddConstant))
      return true;
    break;
  case AArch64CC::VS:
  case AArch64CC::VC:
  case AArch64CC::AL:
  case AArch64CC::NV:
    return true;
  case AArch64CC::Invalid:
    break;
  }

  return false;
}

// (X & C) >u Mask --> (X & (C & (~Mask)) != 0
// (X & C) <u Pow2 --> (X & (C & ~(Pow2-1)) == 0
static SDValue performSubsToAndsCombine(SDNode *N, SDNode *SubsNode,
                                        SDNode *AndNode, SelectionDAG &DAG,
                                        unsigned CCIndex, unsigned CmpIndex,
                                        unsigned CC) {
  ConstantSDNode *SubsC = dyn_cast<ConstantSDNode>(SubsNode->getOperand(1));
  if (!SubsC)
    return SDValue();

  APInt SubsAP = SubsC->getAPIntValue();
  if (CC == AArch64CC::HI) {
    if (!SubsAP.isMask())
      return SDValue();
  } else if (CC == AArch64CC::LO) {
    if (!SubsAP.isPowerOf2())
      return SDValue();
  } else
    return SDValue();

  ConstantSDNode *AndC = dyn_cast<ConstantSDNode>(AndNode->getOperand(1));
  if (!AndC)
    return SDValue();

  APInt MaskAP = CC == AArch64CC::HI ? SubsAP : (SubsAP - 1);

  SDLoc DL(N);
  APInt AndSMask = (~MaskAP) & AndC->getAPIntValue();
  SDValue ANDS = DAG.getNode(
      AArch64ISD::ANDS, DL, SubsNode->getVTList(), AndNode->getOperand(0),
      DAG.getConstant(AndSMask, DL, SubsC->getValueType(0)));
  SDValue AArch64_CC =
      DAG.getConstant(CC == AArch64CC::HI ? AArch64CC::NE : AArch64CC::EQ, DL,
                      N->getOperand(CCIndex)->getValueType(0));

  // For now, only performCSELCombine and performBRCONDCombine call this
  // function. And both of them pass 2 for CCIndex, 3 for CmpIndex with 4
  // operands. So just init the ops direct to simplify the code. If we have some
  // other case with different CCIndex, CmpIndex, we need to use for loop to
  // rewrite the code here.
  // TODO: Do we need to assert number of operand is 4 here?
  assert((CCIndex == 2 && CmpIndex == 3) &&
         "Expected CCIndex to be 2 and CmpIndex to be 3.");
  SDValue Ops[] = {N->getOperand(0), N->getOperand(1), AArch64_CC,
                   ANDS.getValue(1)};
  return DAG.getNode(N->getOpcode(), N, N->getVTList(), Ops);
}

static
SDValue performCONDCombine(SDNode *N,
                           TargetLowering::DAGCombinerInfo &DCI,
                           SelectionDAG &DAG, unsigned CCIndex,
                           unsigned CmpIndex) {
  unsigned CC = cast<ConstantSDNode>(N->getOperand(CCIndex))->getSExtValue();
  SDNode *SubsNode = N->getOperand(CmpIndex).getNode();
  unsigned CondOpcode = SubsNode->getOpcode();

  if (CondOpcode != AArch64ISD::SUBS || SubsNode->hasAnyUseOfValue(0) ||
      !SubsNode->hasOneUse())
    return SDValue();

  // There is a SUBS feeding this condition. Is it fed by a mask we can
  // use?

  SDNode *AndNode = SubsNode->getOperand(0).getNode();
  unsigned MaskBits = 0;

  if (AndNode->getOpcode() != ISD::AND)
    return SDValue();

  if (SDValue Val = performSubsToAndsCombine(N, SubsNode, AndNode, DAG, CCIndex,
                                             CmpIndex, CC))
    return Val;

  if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(AndNode->getOperand(1))) {
    uint32_t CNV = CN->getZExtValue();
    if (CNV == 255)
      MaskBits = 8;
    else if (CNV == 65535)
      MaskBits = 16;
  }

  if (!MaskBits)
    return SDValue();

  SDValue AddValue = AndNode->getOperand(0);

  if (AddValue.getOpcode() != ISD::ADD)
    return SDValue();

  // The basic dag structure is correct, grab the inputs and validate them.

  SDValue AddInputValue1 = AddValue.getNode()->getOperand(0);
  SDValue AddInputValue2 = AddValue.getNode()->getOperand(1);
  SDValue SubsInputValue = SubsNode->getOperand(1);

  // The mask is present and the provenance of all the values is a smaller type,
  // lets see if the mask is superfluous.

  if (!isa<ConstantSDNode>(AddInputValue2.getNode()) ||
      !isa<ConstantSDNode>(SubsInputValue.getNode()))
    return SDValue();

  ISD::LoadExtType ExtType;

  if (!checkValueWidth(SubsInputValue, MaskBits, ExtType) ||
      !checkValueWidth(AddInputValue2, MaskBits, ExtType) ||
      !checkValueWidth(AddInputValue1, MaskBits, ExtType) )
    return SDValue();

  if(!isEquivalentMaskless(CC, MaskBits, ExtType,
                cast<ConstantSDNode>(AddInputValue2.getNode())->getSExtValue(),
                cast<ConstantSDNode>(SubsInputValue.getNode())->getSExtValue()))
    return SDValue();

  // The AND is not necessary, remove it.

  SDVTList VTs = DAG.getVTList(SubsNode->getValueType(0),
                               SubsNode->getValueType(1));
  SDValue Ops[] = { AddValue, SubsNode->getOperand(1) };

  SDValue NewValue = DAG.getNode(CondOpcode, SDLoc(SubsNode), VTs, Ops);
  DAG.ReplaceAllUsesWith(SubsNode, NewValue.getNode());

  return SDValue(N, 0);
}

// Optimize compare with zero and branch.
static SDValue performBRCONDCombine(SDNode *N,
                                    TargetLowering::DAGCombinerInfo &DCI,
                                    SelectionDAG &DAG) {
  MachineFunction &MF = DAG.getMachineFunction();
  // Speculation tracking/SLH assumes that optimized TB(N)Z/CB(N)Z instructions
  // will not be produced, as they are conditional branch instructions that do
  // not set flags.
  if (MF.getFunction().hasFnAttribute(Attribute::SpeculativeLoadHardening))
    return SDValue();

  if (SDValue NV = performCONDCombine(N, DCI, DAG, 2, 3))
    N = NV.getNode();
  SDValue Chain = N->getOperand(0);
  SDValue Dest = N->getOperand(1);
  SDValue CCVal = N->getOperand(2);
  SDValue Cmp = N->getOperand(3);

  assert(isa<ConstantSDNode>(CCVal) && "Expected a ConstantSDNode here!");
  unsigned CC = CCVal->getAsZExtVal();
  if (CC != AArch64CC::EQ && CC != AArch64CC::NE)
    return SDValue();

  // Fold away brcond(NE, cmp(csel(1, 0, CC, Cmp), 1)) -> brcond(~CC, Cmp)
  if (isCMP(Cmp) && CC == AArch64CC::NE && isOneConstant(Cmp.getOperand(1))) {
    SDValue CSel = Cmp.getOperand(0);
    auto CSelCC = getCSETCondCode(CSel);
    if (CSelCC) {
      SDLoc DL(N);
      return DAG.getNode(
          N->getOpcode(), DL, N->getVTList(), Chain, Dest,
          DAG.getConstant(getInvertedCondCode(*CSelCC), DL, MVT::i32),
          CSel.getOperand(3));
    }
  }

  unsigned CmpOpc = Cmp.getOpcode();
  if (CmpOpc != AArch64ISD::ADDS && CmpOpc != AArch64ISD::SUBS)
    return SDValue();

  // Only attempt folding if there is only one use of the flag and no use of the
  // value.
  if (!Cmp->hasNUsesOfValue(0, 0) || !Cmp->hasNUsesOfValue(1, 1))
    return SDValue();

  SDValue LHS = Cmp.getOperand(0);
  SDValue RHS = Cmp.getOperand(1);

  assert(LHS.getValueType() == RHS.getValueType() &&
         "Expected the value type to be the same for both operands!");
  if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
    return SDValue();

  if (isNullConstant(LHS))
    std::swap(LHS, RHS);

  if (!isNullConstant(RHS))
    return SDValue();

  if (LHS.getOpcode() == ISD::SHL || LHS.getOpcode() == ISD::SRA ||
      LHS.getOpcode() == ISD::SRL)
    return SDValue();

  // Fold the compare into the branch instruction.
  SDValue BR;
  if (CC == AArch64CC::EQ)
    BR = DAG.getNode(AArch64ISD::CBZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);
  else
    BR = DAG.getNode(AArch64ISD::CBNZ, SDLoc(N), MVT::Other, Chain, LHS, Dest);

  // Do not add new nodes to DAG combiner worklist.
  DCI.CombineTo(N, BR, false);

  return SDValue();
}

static SDValue foldCSELofCTTZ(SDNode *N, SelectionDAG &DAG) {
  unsigned CC = N->getConstantOperandVal(2);
  SDValue SUBS = N->getOperand(3);
  SDValue Zero, CTTZ;

  if (CC == AArch64CC::EQ && SUBS.getOpcode() == AArch64ISD::SUBS) {
    Zero = N->getOperand(0);
    CTTZ = N->getOperand(1);
  } else if (CC == AArch64CC::NE && SUBS.getOpcode() == AArch64ISD::SUBS) {
    Zero = N->getOperand(1);
    CTTZ = N->getOperand(0);
  } else
    return SDValue();

  if ((CTTZ.getOpcode() != ISD::CTTZ && CTTZ.getOpcode() != ISD::TRUNCATE) ||
      (CTTZ.getOpcode() == ISD::TRUNCATE &&
       CTTZ.getOperand(0).getOpcode() != ISD::CTTZ))
    return SDValue();

  assert((CTTZ.getValueType() == MVT::i32 || CTTZ.getValueType() == MVT::i64) &&
         "Illegal type in CTTZ folding");

  if (!isNullConstant(Zero) || !isNullConstant(SUBS.getOperand(1)))
    return SDValue();

  SDValue X = CTTZ.getOpcode() == ISD::TRUNCATE
                  ? CTTZ.getOperand(0).getOperand(0)
                  : CTTZ.getOperand(0);

  if (X != SUBS.getOperand(0))
    return SDValue();

  unsigned BitWidth = CTTZ.getOpcode() == ISD::TRUNCATE
                          ? CTTZ.getOperand(0).getValueSizeInBits()
                          : CTTZ.getValueSizeInBits();
  SDValue BitWidthMinusOne =
      DAG.getConstant(BitWidth - 1, SDLoc(N), CTTZ.getValueType());
  return DAG.getNode(ISD::AND, SDLoc(N), CTTZ.getValueType(), CTTZ,
                     BitWidthMinusOne);
}

// (CSEL l r EQ (CMP (CSEL x y cc2 cond) x)) => (CSEL l r cc2 cond)
// (CSEL l r EQ (CMP (CSEL x y cc2 cond) y)) => (CSEL l r !cc2 cond)
// Where x and y are constants and x != y

// (CSEL l r NE (CMP (CSEL x y cc2 cond) x)) => (CSEL l r !cc2 cond)
// (CSEL l r NE (CMP (CSEL x y cc2 cond) y)) => (CSEL l r cc2 cond)
// Where x and y are constants and x != y
static SDValue foldCSELOfCSEL(SDNode *Op, SelectionDAG &DAG) {
  SDValue L = Op->getOperand(0);
  SDValue R = Op->getOperand(1);
  AArch64CC::CondCode OpCC =
      static_cast<AArch64CC::CondCode>(Op->getConstantOperandVal(2));

  SDValue OpCmp = Op->getOperand(3);
  if (!isCMP(OpCmp))
    return SDValue();

  SDValue CmpLHS = OpCmp.getOperand(0);
  SDValue CmpRHS = OpCmp.getOperand(1);

  if (CmpRHS.getOpcode() == AArch64ISD::CSEL)
    std::swap(CmpLHS, CmpRHS);
  else if (CmpLHS.getOpcode() != AArch64ISD::CSEL)
    return SDValue();

  SDValue X = CmpLHS->getOperand(0);
  SDValue Y = CmpLHS->getOperand(1);
  if (!isa<ConstantSDNode>(X) || !isa<ConstantSDNode>(Y) || X == Y) {
    return SDValue();
  }

  // If one of the constant is opaque constant, x,y sdnode is still different
  // but the real value maybe the same. So check APInt here to make sure the
  // code is correct.
  ConstantSDNode *CX = cast<ConstantSDNode>(X);
  ConstantSDNode *CY = cast<ConstantSDNode>(Y);
  if (CX->getAPIntValue() == CY->getAPIntValue())
    return SDValue();

  AArch64CC::CondCode CC =
      static_cast<AArch64CC::CondCode>(CmpLHS->getConstantOperandVal(2));
  SDValue Cond = CmpLHS->getOperand(3);

  if (CmpRHS == Y)
    CC = AArch64CC::getInvertedCondCode(CC);
  else if (CmpRHS != X)
    return SDValue();

  if (OpCC == AArch64CC::NE)
    CC = AArch64CC::getInvertedCondCode(CC);
  else if (OpCC != AArch64CC::EQ)
    return SDValue();

  SDLoc DL(Op);
  EVT VT = Op->getValueType(0);

  SDValue CCValue = DAG.getConstant(CC, DL, MVT::i32);
  return DAG.getNode(AArch64ISD::CSEL, DL, VT, L, R, CCValue, Cond);
}

// Reassociate the true/false expressions of a CSEL instruction to obtain a
// common subexpression with the comparison instruction. For example, change
// (CSEL (ADD (ADD x y) -c) f LO (SUBS x c)) to
// (CSEL (ADD (SUBS x c) y) f LO (SUBS x c)) such that (SUBS x c) is a common
// subexpression.
static SDValue reassociateCSELOperandsForCSE(SDNode *N, SelectionDAG &DAG) {
  SDValue SubsNode = N->getOperand(3);
  if (SubsNode.getOpcode() != AArch64ISD::SUBS || !SubsNode.hasOneUse())
    return SDValue();

  SDValue CmpOpToMatch = SubsNode.getOperand(1);
  SDValue CmpOpOther = SubsNode.getOperand(0);
  EVT VT = N->getValueType(0);

  unsigned ExpectedOpcode;
  SDValue ExpectedOp;
  SDValue SubsOp;
  auto *CmpOpConst = dyn_cast<ConstantSDNode>(CmpOpToMatch);
  if (CmpOpConst) {
    ExpectedOpcode = ISD::ADD;
    ExpectedOp =
        DAG.getConstant(-CmpOpConst->getAPIntValue(), SDLoc(CmpOpConst),
                        CmpOpConst->getValueType(0));
    SubsOp = DAG.getConstant(CmpOpConst->getAPIntValue(), SDLoc(CmpOpConst),
                             CmpOpConst->getValueType(0));
  } else {
    ExpectedOpcode = ISD::SUB;
    ExpectedOp = CmpOpToMatch;
    SubsOp = CmpOpToMatch;
  }

  // Get the operand that can be reassociated with the SUBS instruction.
  auto GetReassociationOp = [&](SDValue Op, SDValue ExpectedOp) {
    if (Op.getOpcode() != ExpectedOpcode)
      return SDValue();
    if (Op.getOperand(0).getOpcode() != ISD::ADD ||
        !Op.getOperand(0).hasOneUse())
      return SDValue();
    SDValue X = Op.getOperand(0).getOperand(0);
    SDValue Y = Op.getOperand(0).getOperand(1);
    if (X != CmpOpOther)
      std::swap(X, Y);
    if (X != CmpOpOther)
      return SDValue();
    if (ExpectedOp != Op.getOperand(1))
      return SDValue();
    return Y;
  };

  // Try the reassociation using the given constant and condition code.
  auto Fold = [&](AArch64CC::CondCode NewCC, SDValue ExpectedOp,
                  SDValue SubsOp) {
    SDValue TReassocOp = GetReassociationOp(N->getOperand(0), ExpectedOp);
    SDValue FReassocOp = GetReassociationOp(N->getOperand(1), ExpectedOp);
    if (!TReassocOp && !FReassocOp)
      return SDValue();

    SDValue NewCmp = DAG.getNode(AArch64ISD::SUBS, SDLoc(SubsNode),
                                 DAG.getVTList(VT, MVT_CC), CmpOpOther, SubsOp);

    auto Reassociate = [&](SDValue ReassocOp, unsigned OpNum) {
      if (!ReassocOp)
        return N->getOperand(OpNum);
      SDValue Res = DAG.getNode(ISD::ADD, SDLoc(N->getOperand(OpNum)), VT,
                                NewCmp.getValue(0), ReassocOp);
      DAG.ReplaceAllUsesWith(N->getOperand(OpNum), Res);
      return Res;
    };

    SDValue TValReassoc = Reassociate(TReassocOp, 0);
    SDValue FValReassoc = Reassociate(FReassocOp, 1);
    return DAG.getNode(AArch64ISD::CSEL, SDLoc(N), VT, TValReassoc, FValReassoc,
                       DAG.getConstant(NewCC, SDLoc(N->getOperand(2)), MVT_CC),
                       NewCmp.getValue(1));
  };

  auto CC = static_cast<AArch64CC::CondCode>(N->getConstantOperandVal(2));

  // First, try to eliminate the compare instruction by searching for a
  // subtraction with the same constant.
  if (SDValue R = Fold(CC, ExpectedOp, SubsOp))
    return R;

  if (!CmpOpConst) {
    // Try again with the operands of the SUBS instruction and the condition
    // swapped. Due to canonicalization, this only helps for non-constant
    // operands of the SUBS instruction.
    std::swap(CmpOpToMatch, CmpOpOther);
    if (SDValue R = Fold(getSwappedCondition(CC), CmpOpToMatch, CmpOpToMatch))
      return R;
    return SDValue();
  }

  if ((CC == AArch64CC::EQ || CC == AArch64CC::NE) && !CmpOpConst->isZero())
    return SDValue();

  // Next, search for a subtraction with a slightly different constant. By
  // adjusting the condition code, we can still eliminate the compare
  // instruction. Adjusting the constant is only valid if it does not result
  // in signed/unsigned wrap for signed/unsigned comparisons, respectively.
  // Since such comparisons are trivially true/false, we should not encounter
  // them here but check for them nevertheless to be on the safe side.
  auto CheckedFold = [&](bool Check, APInt NewCmpConst,
                         AArch64CC::CondCode NewCC) {
    auto ExpectedOp = DAG.getConstant(-NewCmpConst, SDLoc(CmpOpConst),
                                      CmpOpConst->getValueType(0));
    auto SubsOp = DAG.getConstant(NewCmpConst, SDLoc(CmpOpConst),
                                  CmpOpConst->getValueType(0));
    return Check ? Fold(NewCC, ExpectedOp, SubsOp) : SDValue();
  };
  switch (CC) {
  case AArch64CC::EQ:
  case AArch64CC::LS:
    return CheckedFold(!CmpOpConst->getAPIntValue().isMaxValue(),
                       CmpOpConst->getAPIntValue() + 1, AArch64CC::LO);
  case AArch64CC::NE:
  case AArch64CC::HI:
    return CheckedFold(!CmpOpConst->getAPIntValue().isMaxValue(),
                       CmpOpConst->getAPIntValue() + 1, AArch64CC::HS);
  case AArch64CC::LO:
    return CheckedFold(!CmpOpConst->getAPIntValue().isZero(),
                       CmpOpConst->getAPIntValue() - 1, AArch64CC::LS);
  case AArch64CC::HS:
    return CheckedFold(!CmpOpConst->getAPIntValue().isZero(),
                       CmpOpConst->getAPIntValue() - 1, AArch64CC::HI);
  case AArch64CC::LT:
    return CheckedFold(!CmpOpConst->getAPIntValue().isMinSignedValue(),
                       CmpOpConst->getAPIntValue() - 1, AArch64CC::LE);
  case AArch64CC::LE:
    return CheckedFold(!CmpOpConst->getAPIntValue().isMaxSignedValue(),
                       CmpOpConst->getAPIntValue() + 1, AArch64CC::LT);
  case AArch64CC::GT:
    return CheckedFold(!CmpOpConst->getAPIntValue().isMaxSignedValue(),
                       CmpOpConst->getAPIntValue() + 1, AArch64CC::GE);
  case AArch64CC::GE:
    return CheckedFold(!CmpOpConst->getAPIntValue().isMinSignedValue(),
                       CmpOpConst->getAPIntValue() - 1, AArch64CC::GT);
  default:
    return SDValue();
  }
}

static SDValue foldCSELofLASTB(SDNode *Op, SelectionDAG &DAG) {
  AArch64CC::CondCode OpCC =
      static_cast<AArch64CC::CondCode>(Op->getConstantOperandVal(2));

  if (OpCC != AArch64CC::NE)
    return SDValue();

  SDValue PTest = Op->getOperand(3);
  if (PTest.getOpcode() != AArch64ISD::PTEST_ANY)
    return SDValue();

  SDValue TruePred = PTest.getOperand(0);
  SDValue AnyPred = PTest.getOperand(1);

  if (TruePred.getOpcode() == AArch64ISD::REINTERPRET_CAST)
    TruePred = TruePred.getOperand(0);

  if (AnyPred.getOpcode() == AArch64ISD::REINTERPRET_CAST)
    AnyPred = AnyPred.getOperand(0);

  if (TruePred != AnyPred && !isAllActivePredicate(DAG, TruePred))
    return SDValue();

  SDValue LastB = Op->getOperand(0);
  SDValue Default = Op->getOperand(1);

  if (LastB.getOpcode() != AArch64ISD::LASTB || LastB.getOperand(0) != AnyPred)
    return SDValue();

  return DAG.getNode(AArch64ISD::CLASTB_N, SDLoc(Op), Op->getValueType(0),
                     AnyPred, Default, LastB.getOperand(1));
}

// Optimize CSEL instructions
static SDValue performCSELCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  SelectionDAG &DAG) {
  // CSEL x, x, cc -> x
  if (N->getOperand(0) == N->getOperand(1))
    return N->getOperand(0);

  if (SDValue R = foldCSELOfCSEL(N, DAG))
    return R;

  // Try to reassociate the true/false expressions so that we can do CSE with
  // a SUBS instruction used to perform the comparison.
  if (SDValue R = reassociateCSELOperandsForCSE(N, DAG))
    return R;

  // CSEL 0, cttz(X), eq(X, 0) -> AND cttz bitwidth-1
  // CSEL cttz(X), 0, ne(X, 0) -> AND cttz bitwidth-1
  if (SDValue Folded = foldCSELofCTTZ(N, DAG))
		return Folded;

  // CSEL a, b, cc, SUBS(x, y) -> CSEL a, b, swapped(cc), SUBS(y, x)
  // if SUB(y, x) already exists and we can produce a swapped predicate for cc.
  SDValue Cond = N->getOperand(3);
  if (DCI.isAfterLegalizeDAG() && Cond.getOpcode() == AArch64ISD::SUBS &&
      Cond.hasOneUse() && Cond->hasNUsesOfValue(0, 0) &&
      DAG.doesNodeExist(ISD::SUB, N->getVTList(),
                        {Cond.getOperand(1), Cond.getOperand(0)}) &&
      !DAG.doesNodeExist(ISD::SUB, N->getVTList(),
                         {Cond.getOperand(0), Cond.getOperand(1)}) &&
      !isNullConstant(Cond.getOperand(1))) {
    AArch64CC::CondCode OldCond =
        static_cast<AArch64CC::CondCode>(N->getConstantOperandVal(2));
    AArch64CC::CondCode NewCond = getSwappedCondition(OldCond);
    if (NewCond != AArch64CC::AL) {
      SDLoc DL(N);
      SDValue Sub = DAG.getNode(AArch64ISD::SUBS, DL, Cond->getVTList(),
                                Cond.getOperand(1), Cond.getOperand(0));
      return DAG.getNode(AArch64ISD::CSEL, DL, N->getVTList(), N->getOperand(0),
                         N->getOperand(1),
                         DAG.getConstant(NewCond, DL, MVT::i32),
                         Sub.getValue(1));
    }
  }

  // CSEL (LASTB P, Z), X, NE(ANY P) -> CLASTB P, X, Z
  if (SDValue CondLast = foldCSELofLASTB(N, DAG))
    return CondLast;

  return performCONDCombine(N, DCI, DAG, 2, 3);
}

// Try to re-use an already extended operand of a vector SetCC feeding a
// extended select. Doing so avoids requiring another full extension of the
// SET_CC result when lowering the select.
static SDValue tryToWidenSetCCOperands(SDNode *Op, SelectionDAG &DAG) {
  EVT Op0MVT = Op->getOperand(0).getValueType();
  if (!Op0MVT.isVector() || Op->use_empty())
    return SDValue();

  // Make sure that all uses of Op are VSELECTs with result matching types where
  // the result type has a larger element type than the SetCC operand.
  SDNode *FirstUse = *Op->user_begin();
  if (FirstUse->getOpcode() != ISD::VSELECT)
    return SDValue();
  EVT UseMVT = FirstUse->getValueType(0);
  if (UseMVT.getScalarSizeInBits() <= Op0MVT.getScalarSizeInBits())
    return SDValue();
  if (any_of(Op->users(), [&UseMVT](const SDNode *N) {
        return N->getOpcode() != ISD::VSELECT || N->getValueType(0) != UseMVT;
      }))
    return SDValue();

  APInt V;
  if (!ISD::isConstantSplatVector(Op->getOperand(1).getNode(), V))
    return SDValue();

  SDLoc DL(Op);
  SDValue Op0ExtV;
  SDValue Op1ExtV;
  ISD::CondCode CC = cast<CondCodeSDNode>(Op->getOperand(2))->get();
  // Check if the first operand of the SET_CC is already extended. If it is,
  // split the SET_CC and re-use the extended version of the operand.
  SDNode *Op0SExt = DAG.getNodeIfExists(ISD::SIGN_EXTEND, DAG.getVTList(UseMVT),
                                        Op->getOperand(0));
  SDNode *Op0ZExt = DAG.getNodeIfExists(ISD::ZERO_EXTEND, DAG.getVTList(UseMVT),
                                        Op->getOperand(0));
  if (Op0SExt && (isSignedIntSetCC(CC) || isIntEqualitySetCC(CC))) {
    Op0ExtV = SDValue(Op0SExt, 0);
    Op1ExtV = DAG.getNode(ISD::SIGN_EXTEND, DL, UseMVT, Op->getOperand(1));
  } else if (Op0ZExt && (isUnsignedIntSetCC(CC) || isIntEqualitySetCC(CC))) {
    Op0ExtV = SDValue(Op0ZExt, 0);
    Op1ExtV = DAG.getNode(ISD::ZERO_EXTEND, DL, UseMVT, Op->getOperand(1));
  } else
    return SDValue();

  return DAG.getNode(ISD::SETCC, DL, UseMVT.changeVectorElementType(MVT::i1),
                     Op0ExtV, Op1ExtV, Op->getOperand(2));
}

static SDValue
performVecReduceBitwiseCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                               SelectionDAG &DAG) {
  SDValue Vec = N->getOperand(0);
  if (DCI.isBeforeLegalize() &&
      Vec.getValueType().getVectorElementType() == MVT::i1 &&
      Vec.getValueType().isFixedLengthVector() &&
      Vec.getValueType().isPow2VectorType()) {
    SDLoc DL(N);
    return getVectorBitwiseReduce(N->getOpcode(), Vec, N->getValueType(0), DL,
                                  DAG);
  }

  return SDValue();
}

static SDValue performSETCCCombine(SDNode *N,
                                   TargetLowering::DAGCombinerInfo &DCI,
                                   SelectionDAG &DAG) {
  assert(N->getOpcode() == ISD::SETCC && "Unexpected opcode!");
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(2))->get();
  SDLoc DL(N);
  EVT VT = N->getValueType(0);

  if (SDValue V = tryToWidenSetCCOperands(N, DAG))
    return V;

  // setcc (csel 0, 1, cond, X), 1, ne ==> csel 0, 1, !cond, X
  if (Cond == ISD::SETNE && isOneConstant(RHS) &&
      LHS->getOpcode() == AArch64ISD::CSEL &&
      isNullConstant(LHS->getOperand(0)) && isOneConstant(LHS->getOperand(1)) &&
      LHS->hasOneUse()) {
    // Invert CSEL's condition.
    auto OldCond =
        static_cast<AArch64CC::CondCode>(LHS.getConstantOperandVal(2));
    auto NewCond = getInvertedCondCode(OldCond);

    // csel 0, 1, !cond, X
    SDValue CSEL =
        DAG.getNode(AArch64ISD::CSEL, DL, LHS.getValueType(), LHS.getOperand(0),
                    LHS.getOperand(1), DAG.getConstant(NewCond, DL, MVT::i32),
                    LHS.getOperand(3));
    return DAG.getZExtOrTrunc(CSEL, DL, VT);
  }

  // setcc (srl x, imm), 0, ne ==> setcc (and x, (-1 << imm)), 0, ne
  if (Cond == ISD::SETNE && isNullConstant(RHS) &&
      LHS->getOpcode() == ISD::SRL && isa<ConstantSDNode>(LHS->getOperand(1)) &&
      LHS->hasOneUse()) {
    EVT TstVT = LHS->getValueType(0);
    if (TstVT.isScalarInteger() && TstVT.getFixedSizeInBits() <= 64 &&
        LHS->getConstantOperandVal(1) < TstVT.getFixedSizeInBits()) {
      // this pattern will get better opt in emitComparison
      uint64_t TstImm = -1ULL << LHS->getConstantOperandVal(1);
      SDValue TST = DAG.getNode(ISD::AND, DL, TstVT, LHS->getOperand(0),
                                DAG.getSignedConstant(TstImm, DL, TstVT));
      return DAG.getNode(ISD::SETCC, DL, VT, TST, RHS, N->getOperand(2));
    }
  }

  // setcc (iN (bitcast (vNi1 X))), 0, (eq|ne)
  //   ==> setcc (iN (zext (i1 (vecreduce_or (vNi1 X))))), 0, (eq|ne)
  // setcc (iN (bitcast (vNi1 X))), -1, (eq|ne)
  //   ==> setcc (iN (sext (i1 (vecreduce_and (vNi1 X))))), -1, (eq|ne)
  if (DCI.isBeforeLegalize() && VT.isScalarInteger() &&
      (Cond == ISD::SETEQ || Cond == ISD::SETNE) &&
      (isNullConstant(RHS) || isAllOnesConstant(RHS)) &&
      LHS->getOpcode() == ISD::BITCAST) {
    EVT ToVT = LHS->getValueType(0);
    EVT FromVT = LHS->getOperand(0).getValueType();
    if (FromVT.isFixedLengthVector() &&
        FromVT.getVectorElementType() == MVT::i1) {
      bool IsNull = isNullConstant(RHS);
      LHS = DAG.getNode(IsNull ? ISD::VECREDUCE_OR : ISD::VECREDUCE_AND,
                        DL, MVT::i1, LHS->getOperand(0));
      LHS = DAG.getNode(IsNull ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND, DL, ToVT,
                        LHS);
      return DAG.getSetCC(DL, VT, LHS, RHS, Cond);
    }
  }

  // Try to perform the memcmp when the result is tested for [in]equality with 0
  if (SDValue V = performOrXorChainCombine(N, DAG))
    return V;

  EVT CmpVT = LHS.getValueType();

  // NOTE: This exists as a combine only because it proved too awkward to match
  // splat(1) across all the NEON types during isel.
  APInt SplatLHSVal;
  if (CmpVT.isInteger() && Cond == ISD::SETGT &&
      ISD::isConstantSplatVector(LHS.getNode(), SplatLHSVal) &&
      SplatLHSVal.isOne())
    return DAG.getSetCC(DL, VT, DAG.getConstant(0, DL, CmpVT), RHS, ISD::SETGE);

  return SDValue();
}

// Replace a flag-setting operator (eg ANDS) with the generic version
// (eg AND) if the flag is unused.
static SDValue performFlagSettingCombine(SDNode *N,
                                         TargetLowering::DAGCombinerInfo &DCI,
                                         unsigned GenericOpcode) {
  SDLoc DL(N);
  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);
  EVT VT = N->getValueType(0);

  // If the flag result isn't used, convert back to a generic opcode.
  if (!N->hasAnyUseOfValue(1)) {
    SDValue Res = DCI.DAG.getNode(GenericOpcode, DL, VT, N->ops());
    return DCI.DAG.getMergeValues({Res, DCI.DAG.getConstant(0, DL, MVT::i32)},
                                  DL);
  }

  // Combine identical generic nodes into this node, re-using the result.
  if (SDNode *Generic = DCI.DAG.getNodeIfExists(
          GenericOpcode, DCI.DAG.getVTList(VT), {LHS, RHS}))
    DCI.CombineTo(Generic, SDValue(N, 0));

  return SDValue();
}

static SDValue performSetCCPunpkCombine(SDNode *N, SelectionDAG &DAG) {
  // setcc_merge_zero pred
  //   (sign_extend (extract_subvector (setcc_merge_zero ... pred ...))), 0, ne
  //   => extract_subvector (inner setcc_merge_zero)
  SDValue Pred = N->getOperand(0);
  SDValue LHS = N->getOperand(1);
  SDValue RHS = N->getOperand(2);
  ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(3))->get();

  if (Cond != ISD::SETNE || !isZerosVector(RHS.getNode()) ||
      LHS->getOpcode() != ISD::SIGN_EXTEND)
    return SDValue();

  SDValue Extract = LHS->getOperand(0);
  if (Extract->getOpcode() != ISD::EXTRACT_SUBVECTOR ||
      Extract->getValueType(0) != N->getValueType(0) ||
      Extract->getConstantOperandVal(1) != 0)
    return SDValue();

  SDValue InnerSetCC = Extract->getOperand(0);
  if (InnerSetCC->getOpcode() != AArch64ISD::SETCC_MERGE_ZERO)
    return SDValue();

  // By this point we've effectively got
  // zero_inactive_lanes_and_trunc_i1(sext_i1(A)). If we can prove A's inactive
  // lanes are already zero then the trunc(sext()) sequence is redundant and we
  // can operate on A directly.
  SDValue InnerPred = InnerSetCC.getOperand(0);
  if (Pred.getOpcode() == AArch64ISD::PTRUE &&
      InnerPred.getOpcode() == AArch64ISD::PTRUE &&
      Pred.getConstantOperandVal(0) == InnerPred.getConstantOperandVal(0) &&
      Pred->getConstantOperandVal(0) >= AArch64SVEPredPattern::vl1 &&
      Pred->getConstantOperandVal(0) <= AArch64SVEPredPattern::vl256)
    return Extract;

  return SDValue();
}

static SDValue
performSetccMergeZeroCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
  assert(N->getOpcode() == AArch64ISD::SETCC_MERGE_ZERO &&
         "Unexpected opcode!");

  SelectionDAG &DAG = DCI.DAG;
  SDValue Pred = N->getOperand(0);
  SDValue LHS = N->getOperand(1);
  SDValue RHS = N->getOperand(2);
  ISD::CondCode Cond = cast<CondCodeSDNode>(N->getOperand(3))->get();

  if (SDValue V = performSetCCPunpkCombine(N, DAG))
    return V;

  if (Cond == ISD::SETNE && isZerosVector(RHS.getNode()) &&
      LHS->getOpcode() == ISD::SIGN_EXTEND &&
      LHS->getOperand(0)->getValueType(0) == N->getValueType(0)) {
    //    setcc_merge_zero(
    //       pred, extend(setcc_merge_zero(pred, ...)), != splat(0))
    // => setcc_merge_zero(pred, ...)
    if (LHS->getOperand(0)->getOpcode() == AArch64ISD::SETCC_MERGE_ZERO &&
        LHS->getOperand(0)->getOperand(0) == Pred)
      return LHS->getOperand(0);

    //    setcc_merge_zero(
    //        all_active, extend(nxvNi1 ...), != splat(0))
    // -> nxvNi1 ...
    if (isAllActivePredicate(DAG, Pred))
      return LHS->getOperand(0);

    //    setcc_merge_zero(
    //        pred, extend(nxvNi1 ...), != splat(0))
    // -> nxvNi1 and(pred, ...)
    if (DCI.isAfterLegalizeDAG())
      // Do this after legalization to allow more folds on setcc_merge_zero
      // to be recognized.
      return DAG.getNode(ISD::AND, SDLoc(N), N->getValueType(0),
                         LHS->getOperand(0), Pred);
  }

  return SDValue();
}

// Optimize some simple tbz/tbnz cases.  Returns the new operand and bit to test
// as well as whether the test should be inverted.  This code is required to
// catch these cases (as opposed to standard dag combines) because
// AArch64ISD::TBZ is matched during legalization.
static SDValue getTestBitOperand(SDValue Op, unsigned &Bit, bool &Invert,
                                 SelectionDAG &DAG) {

  if (!Op->hasOneUse())
    return Op;

  // We don't handle undef/constant-fold cases below, as they should have
  // already been taken care of (e.g. and of 0, test of undefined shifted bits,
  // etc.)

  // (tbz (trunc x), b) -> (tbz x, b)
  // This case is just here to enable more of the below cases to be caught.
  if (Op->getOpcode() == ISD::TRUNCATE &&
      Bit < Op->getValueType(0).getSizeInBits()) {
    return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
  }

  // (tbz (any_ext x), b) -> (tbz x, b) if we don't use the extended bits.
  if (Op->getOpcode() == ISD::ANY_EXTEND &&
      Bit < Op->getOperand(0).getValueSizeInBits()) {
    return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
  }

  if (Op->getNumOperands() != 2)
    return Op;

  auto *C = dyn_cast<ConstantSDNode>(Op->getOperand(1));
  if (!C)
    return Op;

  switch (Op->getOpcode()) {
  default:
    return Op;

  // (tbz (and x, m), b) -> (tbz x, b)
  case ISD::AND:
    if ((C->getZExtValue() >> Bit) & 1)
      return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
    return Op;

  // (tbz (shl x, c), b) -> (tbz x, b-c)
  case ISD::SHL:
    if (C->getZExtValue() <= Bit &&
        (Bit - C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) {
      Bit = Bit - C->getZExtValue();
      return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
    }
    return Op;

  // (tbz (sra x, c), b) -> (tbz x, b+c) or (tbz x, msb) if b+c is > # bits in x
  case ISD::SRA:
    Bit = Bit + C->getZExtValue();
    if (Bit >= Op->getValueType(0).getSizeInBits())
      Bit = Op->getValueType(0).getSizeInBits() - 1;
    return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);

  // (tbz (srl x, c), b) -> (tbz x, b+c)
  case ISD::SRL:
    if ((Bit + C->getZExtValue()) < Op->getValueType(0).getSizeInBits()) {
      Bit = Bit + C->getZExtValue();
      return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
    }
    return Op;

  // (tbz (xor x, -1), b) -> (tbnz x, b)
  case ISD::XOR:
    if ((C->getZExtValue() >> Bit) & 1)
      Invert = !Invert;
    return getTestBitOperand(Op->getOperand(0), Bit, Invert, DAG);
  }
}

// Optimize test single bit zero/non-zero and branch.
static SDValue performTBZCombine(SDNode *N,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 SelectionDAG &DAG) {
  unsigned Bit = N->getConstantOperandVal(2);
  bool Invert = false;
  SDValue TestSrc = N->getOperand(1);
  SDValue NewTestSrc = getTestBitOperand(TestSrc, Bit, Invert, DAG);

  if (TestSrc == NewTestSrc)
    return SDValue();

  unsigned NewOpc = N->getOpcode();
  if (Invert) {
    if (NewOpc == AArch64ISD::TBZ)
      NewOpc = AArch64ISD::TBNZ;
    else {
      assert(NewOpc == AArch64ISD::TBNZ);
      NewOpc = AArch64ISD::TBZ;
    }
  }

  SDLoc DL(N);
  return DAG.getNode(NewOpc, DL, MVT::Other, N->getOperand(0), NewTestSrc,
                     DAG.getConstant(Bit, DL, MVT::i64), N->getOperand(3));
}

// Swap vselect operands where it may allow a predicated operation to achieve
// the `sel`.
//
//     (vselect (setcc ( condcode) (_) (_)) (a)          (op (a) (b)))
//  => (vselect (setcc (!condcode) (_) (_)) (op (a) (b)) (a))
static SDValue trySwapVSelectOperands(SDNode *N, SelectionDAG &DAG) {
  auto SelectA = N->getOperand(1);
  auto SelectB = N->getOperand(2);
  auto NTy = N->getValueType(0);

  if (!NTy.isScalableVector())
    return SDValue();
  SDValue SetCC = N->getOperand(0);
  if (SetCC.getOpcode() != ISD::SETCC || !SetCC.hasOneUse())
    return SDValue();

  switch (SelectB.getOpcode()) {
  default:
    return SDValue();
  case ISD::FMUL:
  case ISD::FSUB:
  case ISD::FADD:
    break;
  }
  if (SelectA != SelectB.getOperand(0))
    return SDValue();

  ISD::CondCode CC = cast<CondCodeSDNode>(SetCC.getOperand(2))->get();
  ISD::CondCode InverseCC =
      ISD::getSetCCInverse(CC, SetCC.getOperand(0).getValueType());
  auto InverseSetCC =
      DAG.getSetCC(SDLoc(SetCC), SetCC.getValueType(), SetCC.getOperand(0),
                   SetCC.getOperand(1), InverseCC);

  return DAG.getNode(ISD::VSELECT, SDLoc(N), NTy,
                     {InverseSetCC, SelectB, SelectA});
}

// vselect (v1i1 setcc) ->
//     vselect (v1iXX setcc)  (XX is the size of the compared operand type)
// FIXME: Currently the type legalizer can't handle VSELECT having v1i1 as
// condition. If it can legalize "VSELECT v1i1" correctly, no need to combine
// such VSELECT.
static SDValue performVSelectCombine(SDNode *N, SelectionDAG &DAG) {
  if (auto SwapResult = trySwapVSelectOperands(N, DAG))
    return SwapResult;

  SDValue N0 = N->getOperand(0);
  SDValue IfTrue = N->getOperand(1);
  SDValue IfFalse = N->getOperand(2);
  EVT ResVT = N->getValueType(0);
  EVT CCVT = N0.getValueType();

  if (isAllActivePredicate(DAG, N0))
    return N->getOperand(1);

  if (isAllInactivePredicate(N0))
    return N->getOperand(2);

  if (isMergePassthruOpcode(IfTrue.getOpcode()) && IfTrue.hasOneUse()) {
    // vselect A, (merge_pasthru_op all_active, B,{Bn,} -), C
    // vselect A, (merge_pasthru_op -, B,{Bn,} undef), C
    // vselect A, (merge_pasthru_op A, B,{Bn,} -), C
    //   -> merge_pasthru_op A, B,{Bn,} C
    if (isAllActivePredicate(DAG, IfTrue->getOperand(0)) ||
        IfTrue->getOperand(IfTrue.getNumOperands() - 1).isUndef() ||
        IfTrue->getOperand(0) == N0) {
      SmallVector<SDValue, 4> Ops(IfTrue->op_values());
      Ops[0] = N0;
      Ops[IfTrue.getNumOperands() - 1] = IfFalse;

      return DAG.getNode(IfTrue.getOpcode(), SDLoc(N), ResVT, Ops);
    }
  }

  // Check for sign pattern (VSELECT setgt, iN lhs, -1, 1, -1) and transform
  // into (OR (ASR lhs, N-1), 1), which requires less instructions for the
  // supported types.
  SDValue SetCC = N->getOperand(0);
  if (SetCC.getOpcode() == ISD::SETCC &&
      SetCC.getOperand(2) == DAG.getCondCode(ISD::SETGT)) {
    SDValue CmpLHS = SetCC.getOperand(0);
    EVT VT = CmpLHS.getValueType();
    SDNode *CmpRHS = SetCC.getOperand(1).getNode();
    SDNode *SplatLHS = N->getOperand(1).getNode();
    SDNode *SplatRHS = N->getOperand(2).getNode();
    APInt SplatLHSVal;
    if (CmpLHS.getValueType() == N->getOperand(1).getValueType() &&
        VT.isSimple() &&
        is_contained(ArrayRef({MVT::v8i8, MVT::v16i8, MVT::v4i16, MVT::v8i16,
                               MVT::v2i32, MVT::v4i32, MVT::v2i64}),
                     VT.getSimpleVT().SimpleTy) &&
        ISD::isConstantSplatVector(SplatLHS, SplatLHSVal) &&
        SplatLHSVal.isOne() && ISD::isConstantSplatVectorAllOnes(CmpRHS) &&
        ISD::isConstantSplatVectorAllOnes(SplatRHS)) {
      unsigned NumElts = VT.getVectorNumElements();
      SmallVector<SDValue, 8> Ops(
          NumElts, DAG.getConstant(VT.getScalarSizeInBits() - 1, SDLoc(N),
                                   VT.getScalarType()));
      SDValue Val = DAG.getBuildVector(VT, SDLoc(N), Ops);

      auto Shift = DAG.getNode(ISD::SRA, SDLoc(N), VT, CmpLHS, Val);
      auto Or = DAG.getNode(ISD::OR, SDLoc(N), VT, Shift, N->getOperand(1));
      return Or;
    }
  }

  EVT CmpVT = N0.getOperand(0).getValueType();
  if (N0.getOpcode() != ISD::SETCC ||
      CCVT.getVectorElementCount() != ElementCount::getFixed(1) ||
      CCVT.getVectorElementType() != MVT::i1 ||
      CmpVT.getVectorElementType().isFloatingPoint())
    return SDValue();

  // Only combine when the result type is of the same size as the compared
  // operands.
  if (ResVT.getSizeInBits() != CmpVT.getSizeInBits())
    return SDValue();

  SetCC = DAG.getSetCC(SDLoc(N), CmpVT.changeVectorElementTypeToInteger(),
                       N0.getOperand(0), N0.getOperand(1),
                       cast<CondCodeSDNode>(N0.getOperand(2))->get());
  return DAG.getNode(ISD::VSELECT, SDLoc(N), ResVT, SetCC,
                     IfTrue, IfFalse);
}

/// A vector select: "(select vL, vR, (setcc LHS, RHS))" is best performed with
/// the compare-mask instructions rather than going via NZCV, even if LHS and
/// RHS are really scalar. This replaces any scalar setcc in the above pattern
/// with a vector one followed by a DUP shuffle on the result.
static SDValue performSelectCombine(SDNode *N,
                                    TargetLowering::DAGCombinerInfo &DCI) {
  SelectionDAG &DAG = DCI.DAG;
  SDValue N0 = N->getOperand(0);
  EVT ResVT = N->getValueType(0);

  if (N0.getOpcode() != ISD::SETCC)
    return SDValue();

  if (ResVT.isScalableVT())
    return SDValue();

  // Make sure the SETCC result is either i1 (initial DAG), or i32, the lowered
  // scalar SetCCResultType. We also don't expect vectors, because we assume
  // that selects fed by vector SETCCs are canonicalized to VSELECT.
  assert((N0.getValueType() == MVT::i1 || N0.getValueType() == MVT::i32) &&
         "Scalar-SETCC feeding SELECT has unexpected result type!");

  // If NumMaskElts == 0, the comparison is larger than select result. The
  // largest real NEON comparison is 64-bits per lane, which means the result is
  // at most 32-bits and an illegal vector. Just bail out for now.
  EVT SrcVT = N0.getOperand(0).getValueType();

  // Don't try to do this optimization when the setcc itself has i1 operands.
  // There are no legal vectors of i1, so this would be pointless. v1f16 is
  // ruled out to prevent the creation of setcc that need to be scalarized.
  if (SrcVT == MVT::i1 ||
      (SrcVT.isFloatingPoint() && SrcVT.getSizeInBits() <= 16))
    return SDValue();

  int NumMaskElts = ResVT.getSizeInBits() / SrcVT.getSizeInBits();
  if (!ResVT.isVector() || NumMaskElts == 0)
    return SDValue();

  SrcVT = EVT::getVectorVT(*DAG.getContext(), SrcVT, NumMaskElts);
  EVT CCVT = SrcVT.changeVectorElementTypeToInteger();

  // Also bail out if the vector CCVT isn't the same size as ResVT.
  // This can happen if the SETCC operand size doesn't divide the ResVT size
  // (e.g., f64 vs v3f32).
  if (CCVT.getSizeInBits() != ResVT.getSizeInBits())
    return SDValue();

  // Make sure we didn't create illegal types, if we're not supposed to.
  assert(DCI.isBeforeLegalize() ||
         DAG.getTargetLoweringInfo().isTypeLegal(SrcVT));

  // First perform a vector comparison, where lane 0 is the one we're interested
  // in.
  SDLoc DL(N0);
  SDValue LHS =
      DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(0));
  SDValue RHS =
      DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, SrcVT, N0.getOperand(1));
  SDValue SetCC = DAG.getNode(ISD::SETCC, DL, CCVT, LHS, RHS, N0.getOperand(2));

  // Now duplicate the comparison mask we want across all other lanes.
  SmallVector<int, 8> DUPMask(CCVT.getVectorNumElements(), 0);
  SDValue Mask = DAG.getVectorShuffle(CCVT, DL, SetCC, SetCC, DUPMask);
  Mask = DAG.getNode(ISD::BITCAST, DL,
                     ResVT.changeVectorElementTypeToInteger(), Mask);

  return DAG.getSelect(DL, ResVT, Mask, N->getOperand(1), N->getOperand(2));
}

static SDValue performDUPCombine(SDNode *N,
                                 TargetLowering::DAGCombinerInfo &DCI) {
  EVT VT = N->getValueType(0);
  SDLoc DL(N);
  // If "v2i32 DUP(x)" and "v4i32 DUP(x)" both exist, use an extract from the
  // 128bit vector version.
  if (VT.is64BitVector() && DCI.isAfterLegalizeDAG()) {
    EVT LVT = VT.getDoubleNumVectorElementsVT(*DCI.DAG.getContext());
    SmallVector<SDValue> Ops(N->ops());
    if (SDNode *LN = DCI.DAG.getNodeIfExists(N->getOpcode(),
                                             DCI.DAG.getVTList(LVT), Ops)) {
      return DCI.DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SDValue(LN, 0),
                             DCI.DAG.getConstant(0, DL, MVT::i64));
    }
  }

  if (N->getOpcode() == AArch64ISD::DUP) {
    // If the instruction is known to produce a scalar in SIMD registers, we can
    // duplicate it across the vector lanes using DUPLANE instead of moving it
    // to a GPR first. For example, this allows us to handle:
    //   v4i32 = DUP (i32 (FCMGT (f32, f32)))
    SDValue Op = N->getOperand(0);
    // FIXME: Ideally, we should be able to handle all instructions that
    // produce a scalar value in FPRs.
    if (Op.getOpcode() == AArch64ISD::FCMEQ ||
        Op.getOpcode() == AArch64ISD::FCMGE ||
        Op.getOpcode() == AArch64ISD::FCMGT) {
      EVT ElemVT = VT.getVectorElementType();
      EVT ExpandedVT = VT;
      // Insert into a 128-bit vector to match DUPLANE's pattern.
      if (VT.getSizeInBits() != 128)
        ExpandedVT = EVT::getVectorVT(*DCI.DAG.getContext(), ElemVT,
                                      128 / ElemVT.getSizeInBits());
      SDValue Zero = DCI.DAG.getConstant(0, DL, MVT::i64);
      SDValue Vec = DCI.DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ExpandedVT,
                                    DCI.DAG.getUNDEF(ExpandedVT), Op, Zero);
      return DCI.DAG.getNode(getDUPLANEOp(ElemVT), DL, VT, Vec, Zero);
    }

    if (DCI.isAfterLegalizeDAG()) {
      // If scalar dup's operand is extract_vector_elt, try to combine them into
      // duplane. For example,
      //
      //    t21: i32 = extract_vector_elt t19, Constant:i64<0>
      //  t18: v4i32 = AArch64ISD::DUP t21
      //  ==>
      //  t22: v4i32 = AArch64ISD::DUPLANE32 t19, Constant:i64<0>
      SDValue EXTRACT_VEC_ELT = N->getOperand(0);
      if (EXTRACT_VEC_ELT.getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
        if (VT == EXTRACT_VEC_ELT.getOperand(0).getValueType()) {
          unsigned Opcode = getDUPLANEOp(VT.getVectorElementType());
          return DCI.DAG.getNode(Opcode, DL, VT, EXTRACT_VEC_ELT.getOperand(0),
                                 EXTRACT_VEC_ELT.getOperand(1));
        }
      }
    }

    return performPostLD1Combine(N, DCI, false);
  }

  return SDValue();
}

/// Get rid of unnecessary NVCASTs (that don't change the type).
static SDValue performNVCASTCombine(SDNode *N, SelectionDAG &DAG) {
  if (N->getValueType(0) == N->getOperand(0).getValueType())
    return N->getOperand(0);
  if (N->getOperand(0).getOpcode() == AArch64ISD::NVCAST)
    return DAG.getNode(AArch64ISD::NVCAST, SDLoc(N), N->getValueType(0),
                       N->getOperand(0).getOperand(0));

  return SDValue();
}

// If all users of the globaladdr are of the form (globaladdr + constant), find
// the smallest constant, fold it into the globaladdr's offset and rewrite the
// globaladdr as (globaladdr + constant) - constant.
static SDValue performGlobalAddressCombine(SDNode *N, SelectionDAG &DAG,
                                           const AArch64Subtarget *Subtarget,
                                           const TargetMachine &TM) {
  auto *GN = cast<GlobalAddressSDNode>(N);
  if (Subtarget->ClassifyGlobalReference(GN->getGlobal(), TM) !=
      AArch64II::MO_NO_FLAG)
    return SDValue();

  uint64_t MinOffset = -1ull;
  for (SDNode *N : GN->users()) {
    if (N->getOpcode() != ISD::ADD)
      return SDValue();
    auto *C = dyn_cast<ConstantSDNode>(N->getOperand(0));
    if (!C)
      C = dyn_cast<ConstantSDNode>(N->getOperand(1));
    if (!C)
      return SDValue();
    MinOffset = std::min(MinOffset, C->getZExtValue());
  }
  uint64_t Offset = MinOffset + GN->getOffset();

  // Require that the new offset is larger than the existing one. Otherwise, we
  // can end up oscillating between two possible DAGs, for example,
  // (add (add globaladdr + 10, -1), 1) and (add globaladdr + 9, 1).
  if (Offset <= uint64_t(GN->getOffset()))
    return SDValue();

  // Check whether folding this offset is legal. It must not go out of bounds of
  // the referenced object to avoid violating the code model, and must be
  // smaller than 2^20 because this is the largest offset expressible in all
  // object formats. (The IMAGE_REL_ARM64_PAGEBASE_REL21 relocation in COFF
  // stores an immediate signed 21 bit offset.)
  //
  // This check also prevents us from folding negative offsets, which will end
  // up being treated in the same way as large positive ones. They could also
  // cause code model violations, and aren't really common enough to matter.
  if (Offset >= (1 << 20))
    return SDValue();

  const GlobalValue *GV = GN->getGlobal();
  Type *T = GV->getValueType();
  if (!T->isSized() ||
      Offset > GV->getDataLayout().getTypeAllocSize(T))
    return SDValue();

  SDLoc DL(GN);
  SDValue Result = DAG.getGlobalAddress(GV, DL, MVT::i64, Offset);
  return DAG.getNode(ISD::SUB, DL, MVT::i64, Result,
                     DAG.getConstant(MinOffset, DL, MVT::i64));
}

static SDValue performCTLZCombine(SDNode *N, SelectionDAG &DAG,
                                  const AArch64Subtarget *Subtarget) {
  SDValue BR = N->getOperand(0);
  if (!Subtarget->hasCSSC() || BR.getOpcode() != ISD::BITREVERSE ||
      !BR.getValueType().isScalarInteger())
    return SDValue();

  SDLoc DL(N);
  return DAG.getNode(ISD::CTTZ, DL, BR.getValueType(), BR.getOperand(0));
}

// Turns the vector of indices into a vector of byte offstes by scaling Offset
// by (BitWidth / 8).
static SDValue getScaledOffsetForBitWidth(SelectionDAG &DAG, SDValue Offset,
                                          SDLoc DL, unsigned BitWidth) {
  assert(Offset.getValueType().isScalableVector() &&
         "This method is only for scalable vectors of offsets");

  SDValue Shift = DAG.getConstant(Log2_32(BitWidth / 8), DL, MVT::i64);
  SDValue SplatShift = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, Shift);

  return DAG.getNode(ISD::SHL, DL, MVT::nxv2i64, Offset, SplatShift);
}

/// Check if the value of \p OffsetInBytes can be used as an immediate for
/// the gather load/prefetch and scatter store instructions with vector base and
/// immediate offset addressing mode:
///
///      [<Zn>.[S|D]{, #<imm>}]
///
/// where <imm> = sizeof(<T>) * k, for k = 0, 1, ..., 31.
inline static bool isValidImmForSVEVecImmAddrMode(unsigned OffsetInBytes,
                                                  unsigned ScalarSizeInBytes) {
  // The immediate is not a multiple of the scalar size.
  if (OffsetInBytes % ScalarSizeInBytes)
    return false;

  // The immediate is out of range.
  if (OffsetInBytes / ScalarSizeInBytes > 31)
    return false;

  return true;
}

/// Check if the value of \p Offset represents a valid immediate for the SVE
/// gather load/prefetch and scatter store instructiona with vector base and
/// immediate offset addressing mode:
///
///      [<Zn>.[S|D]{, #<imm>}]
///
/// where <imm> = sizeof(<T>) * k, for k = 0, 1, ..., 31.
static bool isValidImmForSVEVecImmAddrMode(SDValue Offset,
                                           unsigned ScalarSizeInBytes) {
  ConstantSDNode *OffsetConst = dyn_cast<ConstantSDNode>(Offset.getNode());
  return OffsetConst && isValidImmForSVEVecImmAddrMode(
                            OffsetConst->getZExtValue(), ScalarSizeInBytes);
}

static SDValue performScatterStoreCombine(SDNode *N, SelectionDAG &DAG,
                                          unsigned Opcode,
                                          bool OnlyPackedOffsets = true) {
  const SDValue Src = N->getOperand(2);
  const EVT SrcVT = Src->getValueType(0);
  assert(SrcVT.isScalableVector() &&
         "Scatter stores are only possible for SVE vectors");

  SDLoc DL(N);
  MVT SrcElVT = SrcVT.getVectorElementType().getSimpleVT();

  // Make sure that source data will fit into an SVE register
  if (SrcVT.getSizeInBits().getKnownMinValue() > AArch64::SVEBitsPerBlock)
    return SDValue();

  // For FPs, ACLE only supports _packed_ single and double precision types.
  // SST1Q_[INDEX_]PRED is the ST1Q for sve2p1 and should allow all sizes.
  if (SrcElVT.isFloatingPoint())
    if ((SrcVT != MVT::nxv4f32) && (SrcVT != MVT::nxv2f64) &&
        ((Opcode != AArch64ISD::SST1Q_PRED &&
          Opcode != AArch64ISD::SST1Q_INDEX_PRED) ||
         ((SrcVT != MVT::nxv8f16) && (SrcVT != MVT::nxv8bf16))))
      return SDValue();

  // Depending on the addressing mode, this is either a pointer or a vector of
  // pointers (that fits into one register)
  SDValue Base = N->getOperand(4);
  // Depending on the addressing mode, this is either a single offset or a
  // vector of offsets  (that fits into one register)
  SDValue Offset = N->getOperand(5);

  // For "scalar + vector of indices", just scale the indices. This only
  // applies to non-temporal scatters because there's no instruction that takes
  // indices.
  if (Opcode == AArch64ISD::SSTNT1_INDEX_PRED) {
    Offset =
        getScaledOffsetForBitWidth(DAG, Offset, DL, SrcElVT.getSizeInBits());
    Opcode = AArch64ISD::SSTNT1_PRED;
  } else if (Opcode == AArch64ISD::SST1Q_INDEX_PRED) {
    Offset =
        getScaledOffsetForBitWidth(DAG, Offset, DL, SrcElVT.getSizeInBits());
    Opcode = AArch64ISD::SST1Q_PRED;
  }

  // In the case of non-temporal gather loads there's only one SVE instruction
  // per data-size: "scalar + vector", i.e.
  //    * stnt1{b|h|w|d} { z0.s }, p0/z, [z0.s, x0]
  // Since we do have intrinsics that allow the arguments to be in a different
  // order, we may need to swap them to match the spec.
  if ((Opcode == AArch64ISD::SSTNT1_PRED || Opcode == AArch64ISD::SST1Q_PRED) &&
      Offset.getValueType().isVector())
    std::swap(Base, Offset);

  // SST1_IMM requires that the offset is an immediate that is:
  //    * a multiple of #SizeInBytes,
  //    * in the range [0, 31 x #SizeInBytes],
  // where #SizeInBytes is the size in bytes of the stored items. For
  // immediates outside that range and non-immediate scalar offsets use SST1 or
  // SST1_UXTW instead.
  if (Opcode == AArch64ISD::SST1_IMM_PRED) {
    if (!isValidImmForSVEVecImmAddrMode(Offset,
                                        SrcVT.getScalarSizeInBits() / 8)) {
      if (MVT::nxv4i32 == Base.getValueType().getSimpleVT().SimpleTy)
        Opcode = AArch64ISD::SST1_UXTW_PRED;
      else
        Opcode = AArch64ISD::SST1_PRED;

      std::swap(Base, Offset);
    }
  }

  auto &TLI = DAG.getTargetLoweringInfo();
  if (!TLI.isTypeLegal(Base.getValueType()))
    return SDValue();

  // Some scatter store variants allow unpacked offsets, but only as nxv2i32
  // vectors. These are implicitly sign (sxtw) or zero (zxtw) extend to
  // nxv2i64. Legalize accordingly.
  if (!OnlyPackedOffsets &&
      Offset.getValueType().getSimpleVT().SimpleTy == MVT::nxv2i32)
    Offset = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::nxv2i64, Offset).getValue(0);

  if (!TLI.isTypeLegal(Offset.getValueType()))
    return SDValue();

  // Source value type that is representable in hardware
  EVT HwSrcVt = getSVEContainerType(SrcVT);

  // Keep the original type of the input data to store - this is needed to be
  // able to select the correct instruction, e.g. ST1B, ST1H, ST1W and ST1D. For
  // FP values we want the integer equivalent, so just use HwSrcVt.
  SDValue InputVT = DAG.getValueType(SrcVT);
  if (SrcVT.isFloatingPoint())
    InputVT = DAG.getValueType(HwSrcVt);

  SDVTList VTs = DAG.getVTList(MVT::Other);
  SDValue SrcNew;

  if (Src.getValueType().isFloatingPoint())
    SrcNew = DAG.getNode(ISD::BITCAST, DL, HwSrcVt, Src);
  else
    SrcNew = DAG.getNode(ISD::ANY_EXTEND, DL, HwSrcVt, Src);

  SDValue Ops[] = {N->getOperand(0), // Chain
                   SrcNew,
                   N->getOperand(3), // Pg
                   Base,
                   Offset,
                   InputVT};

  return DAG.getNode(Opcode, DL, VTs, Ops);
}

static SDValue performGatherLoadCombine(SDNode *N, SelectionDAG &DAG,
                                        unsigned Opcode,
                                        bool OnlyPackedOffsets = true) {
  const EVT RetVT = N->getValueType(0);
  assert(RetVT.isScalableVector() &&
         "Gather loads are only possible for SVE vectors");

  SDLoc DL(N);

  // Make sure that the loaded data will fit into an SVE register
  if (RetVT.getSizeInBits().getKnownMinValue() > AArch64::SVEBitsPerBlock)
    return SDValue();

  // Depending on the addressing mode, this is either a pointer or a vector of
  // pointers (that fits into one register)
  SDValue Base = N->getOperand(3);
  // Depending on the addressing mode, this is either a single offset or a
  // vector of offsets  (that fits into one register)
  SDValue Offset = N->getOperand(4);

  // For "scalar + vector of indices", scale the indices to obtain unscaled
  // offsets. This applies to non-temporal and quadword gathers, which do not
  // have an addressing mode with scaled offset.
  if (Opcode == AArch64ISD::GLDNT1_INDEX_MERGE_ZERO) {
    Offset = getScaledOffsetForBitWidth(DAG, Offset, DL,
                                        RetVT.getScalarSizeInBits());
    Opcode = AArch64ISD::GLDNT1_MERGE_ZERO;
  } else if (Opcode == AArch64ISD::GLD1Q_INDEX_MERGE_ZERO) {
    Offset = getScaledOffsetForBitWidth(DAG, Offset, DL,
                                        RetVT.getScalarSizeInBits());
    Opcode = AArch64ISD::GLD1Q_MERGE_ZERO;
  }

  // In the case of non-temporal gather loads and quadword gather loads there's
  // only one addressing mode : "vector + scalar", e.g.
  //   ldnt1{b|h|w|d} { z0.s }, p0/z, [z0.s, x0]
  // Since we do have intrinsics that allow the arguments to be in a different
  // order, we may need to swap them to match the spec.
  if ((Opcode == AArch64ISD::GLDNT1_MERGE_ZERO ||
       Opcode == AArch64ISD::GLD1Q_MERGE_ZERO) &&
      Offset.getValueType().isVector())
    std::swap(Base, Offset);

  // GLD{FF}1_IMM requires that the offset is an immediate that is:
  //    * a multiple of #SizeInBytes,
  //    * in the range [0, 31 x #SizeInBytes],
  // where #SizeInBytes is the size in bytes of the loaded items. For
  // immediates outside that range and non-immediate scalar offsets use
  // GLD1_MERGE_ZERO or GLD1_UXTW_MERGE_ZERO instead.
  if (Opcode == AArch64ISD::GLD1_IMM_MERGE_ZERO ||
      Opcode == AArch64ISD::GLDFF1_IMM_MERGE_ZERO) {
    if (!isValidImmForSVEVecImmAddrMode(Offset,
                                        RetVT.getScalarSizeInBits() / 8)) {
      if (MVT::nxv4i32 == Base.getValueType().getSimpleVT().SimpleTy)
        Opcode = (Opcode == AArch64ISD::GLD1_IMM_MERGE_ZERO)
                     ? AArch64ISD::GLD1_UXTW_MERGE_ZERO
                     : AArch64ISD::GLDFF1_UXTW_MERGE_ZERO;
      else
        Opcode = (Opcode == AArch64ISD::GLD1_IMM_MERGE_ZERO)
                     ? AArch64ISD::GLD1_MERGE_ZERO
                     : AArch64ISD::GLDFF1_MERGE_ZERO;

      std::swap(Base, Offset);
    }
  }

  auto &TLI = DAG.getTargetLoweringInfo();
  if (!TLI.isTypeLegal(Base.getValueType()))
    return SDValue();

  // Some gather load variants allow unpacked offsets, but only as nxv2i32
  // vectors. These are implicitly sign (sxtw) or zero (zxtw) extend to
  // nxv2i64. Legalize accordingly.
  if (!OnlyPackedOffsets &&
      Offset.getValueType().getSimpleVT().SimpleTy == MVT::nxv2i32)
    Offset = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::nxv2i64, Offset).getValue(0);

  // Return value type that is representable in hardware
  EVT HwRetVt = getSVEContainerType(RetVT);

  // Keep the original output value type around - this is needed to be able to
  // select the correct instruction, e.g. LD1B, LD1H, LD1W and LD1D. For FP
  // values we want the integer equivalent, so just use HwRetVT.
  SDValue OutVT = DAG.getValueType(RetVT);
  if (RetVT.isFloatingPoint())
    OutVT = DAG.getValueType(HwRetVt);

  SDVTList VTs = DAG.getVTList(HwRetVt, MVT::Other);
  SDValue Ops[] = {N->getOperand(0), // Chain
                   N->getOperand(2), // Pg
                   Base, Offset, OutVT};

  SDValue Load = DAG.getNode(Opcode, DL, VTs, Ops);
  SDValue LoadChain = SDValue(Load.getNode(), 1);

  if (RetVT.isInteger() && (RetVT != HwRetVt))
    Load = DAG.getNode(ISD::TRUNCATE, DL, RetVT, Load.getValue(0));

  // If the original return value was FP, bitcast accordingly. Doing it here
  // means that we can avoid adding TableGen patterns for FPs.
  if (RetVT.isFloatingPoint())
    Load = DAG.getNode(ISD::BITCAST, DL, RetVT, Load.getValue(0));

  return DAG.getMergeValues({Load, LoadChain}, DL);
}

static SDValue
performSignExtendInRegCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                              SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue Src = N->getOperand(0);
  unsigned Opc = Src->getOpcode();

  // Sign extend of an unsigned unpack -> signed unpack
  if (Opc == AArch64ISD::UUNPKHI || Opc == AArch64ISD::UUNPKLO) {

    unsigned SOpc = Opc == AArch64ISD::UUNPKHI ? AArch64ISD::SUNPKHI
                                               : AArch64ISD::SUNPKLO;

    // Push the sign extend to the operand of the unpack
    // This is necessary where, for example, the operand of the unpack
    // is another unpack:
    // 4i32 sign_extend_inreg (4i32 uunpklo(8i16 uunpklo (16i8 opnd)), from 4i8)
    // ->
    // 4i32 sunpklo (8i16 sign_extend_inreg(8i16 uunpklo (16i8 opnd), from 8i8)
    // ->
    // 4i32 sunpklo(8i16 sunpklo(16i8 opnd))
    SDValue ExtOp = Src->getOperand(0);
    auto VT = cast<VTSDNode>(N->getOperand(1))->getVT();
    EVT EltTy = VT.getVectorElementType();
    (void)EltTy;

    assert((EltTy == MVT::i8 || EltTy == MVT::i16 || EltTy == MVT::i32) &&
           "Sign extending from an invalid type");

    EVT ExtVT = VT.getDoubleNumVectorElementsVT(*DAG.getContext());

    SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, ExtOp.getValueType(),
                              ExtOp, DAG.getValueType(ExtVT));

    return DAG.getNode(SOpc, DL, N->getValueType(0), Ext);
  }

  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  if (!EnableCombineMGatherIntrinsics)
    return SDValue();

  // SVE load nodes (e.g. AArch64ISD::GLD1) are straightforward candidates
  // for DAG Combine with SIGN_EXTEND_INREG. Bail out for all other nodes.
  unsigned NewOpc;
  unsigned MemVTOpNum = 4;
  switch (Opc) {
  case AArch64ISD::LD1_MERGE_ZERO:
    NewOpc = AArch64ISD::LD1S_MERGE_ZERO;
    MemVTOpNum = 3;
    break;
  case AArch64ISD::LDNF1_MERGE_ZERO:
    NewOpc = AArch64ISD::LDNF1S_MERGE_ZERO;
    MemVTOpNum = 3;
    break;
  case AArch64ISD::LDFF1_MERGE_ZERO:
    NewOpc = AArch64ISD::LDFF1S_MERGE_ZERO;
    MemVTOpNum = 3;
    break;
  case AArch64ISD::GLD1_MERGE_ZERO:
    NewOpc = AArch64ISD::GLD1S_MERGE_ZERO;
    break;
  case AArch64ISD::GLD1_SCALED_MERGE_ZERO:
    NewOpc = AArch64ISD::GLD1S_SCALED_MERGE_ZERO;
    break;
  case AArch64ISD::GLD1_SXTW_MERGE_ZERO:
    NewOpc = AArch64ISD::GLD1S_SXTW_MERGE_ZERO;
    break;
  case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO:
    NewOpc = AArch64ISD::GLD1S_SXTW_SCALED_MERGE_ZERO;
    break;
  case AArch64ISD::GLD1_UXTW_MERGE_ZERO:
    NewOpc = AArch64ISD::GLD1S_UXTW_MERGE_ZERO;
    break;
  case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO:
    NewOpc = AArch64ISD::GLD1S_UXTW_SCALED_MERGE_ZERO;
    break;
  case AArch64ISD::GLD1_IMM_MERGE_ZERO:
    NewOpc = AArch64ISD::GLD1S_IMM_MERGE_ZERO;
    break;
  case AArch64ISD::GLDFF1_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDFF1S_MERGE_ZERO;
    break;
  case AArch64ISD::GLDFF1_SCALED_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDFF1S_SCALED_MERGE_ZERO;
    break;
  case AArch64ISD::GLDFF1_SXTW_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDFF1S_SXTW_MERGE_ZERO;
    break;
  case AArch64ISD::GLDFF1_SXTW_SCALED_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDFF1S_SXTW_SCALED_MERGE_ZERO;
    break;
  case AArch64ISD::GLDFF1_UXTW_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDFF1S_UXTW_MERGE_ZERO;
    break;
  case AArch64ISD::GLDFF1_UXTW_SCALED_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDFF1S_UXTW_SCALED_MERGE_ZERO;
    break;
  case AArch64ISD::GLDFF1_IMM_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDFF1S_IMM_MERGE_ZERO;
    break;
  case AArch64ISD::GLDNT1_MERGE_ZERO:
    NewOpc = AArch64ISD::GLDNT1S_MERGE_ZERO;
    break;
  default:
    return SDValue();
  }

  EVT SignExtSrcVT = cast<VTSDNode>(N->getOperand(1))->getVT();
  EVT SrcMemVT = cast<VTSDNode>(Src->getOperand(MemVTOpNum))->getVT();

  if ((SignExtSrcVT != SrcMemVT) || !Src.hasOneUse())
    return SDValue();

  EVT DstVT = N->getValueType(0);
  SDVTList VTs = DAG.getVTList(DstVT, MVT::Other);

  SmallVector<SDValue, 5> Ops;
  for (unsigned I = 0; I < Src->getNumOperands(); ++I)
    Ops.push_back(Src->getOperand(I));

  SDValue ExtLoad = DAG.getNode(NewOpc, SDLoc(N), VTs, Ops);
  DCI.CombineTo(N, ExtLoad);
  DCI.CombineTo(Src.getNode(), ExtLoad, ExtLoad.getValue(1));

  // Return N so it doesn't get rechecked
  return SDValue(N, 0);
}

/// Legalize the gather prefetch (scalar + vector addressing mode) when the
/// offset vector is an unpacked 32-bit scalable vector. The other cases (Offset
/// != nxv2i32) do not need legalization.
static SDValue legalizeSVEGatherPrefetchOffsVec(SDNode *N, SelectionDAG &DAG) {
  const unsigned OffsetPos = 4;
  SDValue Offset = N->getOperand(OffsetPos);

  // Not an unpacked vector, bail out.
  if (Offset.getValueType().getSimpleVT().SimpleTy != MVT::nxv2i32)
    return SDValue();

  // Extend the unpacked offset vector to 64-bit lanes.
  SDLoc DL(N);
  Offset = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::nxv2i64, Offset);
  SmallVector<SDValue, 5> Ops(N->ops());
  // Replace the offset operand with the 64-bit one.
  Ops[OffsetPos] = Offset;

  return DAG.getNode(N->getOpcode(), DL, DAG.getVTList(MVT::Other), Ops);
}

/// Combines a node carrying the intrinsic
/// `aarch64_sve_prf<T>_gather_scalar_offset` into a node that uses
/// `aarch64_sve_prfb_gather_uxtw_index` when the scalar offset passed to
/// `aarch64_sve_prf<T>_gather_scalar_offset` is not a valid immediate for the
/// sve gather prefetch instruction with vector plus immediate addressing mode.
static SDValue combineSVEPrefetchVecBaseImmOff(SDNode *N, SelectionDAG &DAG,
                                               unsigned ScalarSizeInBytes) {
  const unsigned ImmPos = 4, OffsetPos = 3;
  // No need to combine the node if the immediate is valid...
  if (isValidImmForSVEVecImmAddrMode(N->getOperand(ImmPos), ScalarSizeInBytes))
    return SDValue();

  // ...otherwise swap the offset base with the offset...
  SmallVector<SDValue, 5> Ops(N->ops());
  std::swap(Ops[ImmPos], Ops[OffsetPos]);
  // ...and remap the intrinsic `aarch64_sve_prf<T>_gather_scalar_offset` to
  // `aarch64_sve_prfb_gather_uxtw_index`.
  SDLoc DL(N);
  Ops[1] = DAG.getConstant(Intrinsic::aarch64_sve_prfb_gather_uxtw_index, DL,
                           MVT::i64);

  return DAG.getNode(N->getOpcode(), DL, DAG.getVTList(MVT::Other), Ops);
}

// Return true if the vector operation can guarantee only the first lane of its
// result contains data, with all bits in other lanes set to zero.
static bool isLanes1toNKnownZero(SDValue Op) {
  switch (Op.getOpcode()) {
  default:
    return false;
  case AArch64ISD::ANDV_PRED:
  case AArch64ISD::EORV_PRED:
  case AArch64ISD::FADDA_PRED:
  case AArch64ISD::FADDV_PRED:
  case AArch64ISD::FMAXNMV_PRED:
  case AArch64ISD::FMAXV_PRED:
  case AArch64ISD::FMINNMV_PRED:
  case AArch64ISD::FMINV_PRED:
  case AArch64ISD::ORV_PRED:
  case AArch64ISD::SADDV_PRED:
  case AArch64ISD::SMAXV_PRED:
  case AArch64ISD::SMINV_PRED:
  case AArch64ISD::UADDV_PRED:
  case AArch64ISD::UMAXV_PRED:
  case AArch64ISD::UMINV_PRED:
    return true;
  }
}

static SDValue removeRedundantInsertVectorElt(SDNode *N) {
  assert(N->getOpcode() == ISD::INSERT_VECTOR_ELT && "Unexpected node!");
  SDValue InsertVec = N->getOperand(0);
  SDValue InsertElt = N->getOperand(1);
  SDValue InsertIdx = N->getOperand(2);

  // We only care about inserts into the first element...
  if (!isNullConstant(InsertIdx))
    return SDValue();
  // ...of a zero'd vector...
  if (!ISD::isConstantSplatVectorAllZeros(InsertVec.getNode()))
    return SDValue();
  // ...where the inserted data was previously extracted...
  if (InsertElt.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
    return SDValue();

  SDValue ExtractVec = InsertElt.getOperand(0);
  SDValue ExtractIdx = InsertElt.getOperand(1);

  // ...from the first element of a vector.
  if (!isNullConstant(ExtractIdx))
    return SDValue();

  // If we get here we are effectively trying to zero lanes 1-N of a vector.

  // Ensure there's no type conversion going on.
  if (N->getValueType(0) != ExtractVec.getValueType())
    return SDValue();

  if (!isLanes1toNKnownZero(ExtractVec))
    return SDValue();

  // The explicit zeroing is redundant.
  return ExtractVec;
}

static SDValue
performInsertVectorEltCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
  if (SDValue Res = removeRedundantInsertVectorElt(N))
    return Res;

  return performPostLD1Combine(N, DCI, true);
}

static SDValue performFPExtendCombine(SDNode *N, SelectionDAG &DAG,
                                      TargetLowering::DAGCombinerInfo &DCI,
                                      const AArch64Subtarget *Subtarget) {
  SDValue N0 = N->getOperand(0);
  EVT VT = N->getValueType(0);

  // If this is fp_round(fpextend), don't fold it, allow ourselves to be folded.
  if (N->hasOneUse() && N->user_begin()->getOpcode() == ISD::FP_ROUND)
    return SDValue();

  auto hasValidElementTypeForFPExtLoad = [](EVT VT) {
    EVT EltVT = VT.getVectorElementType();
    return EltVT == MVT::f32 || EltVT == MVT::f64;
  };

  // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
  // We purposefully don't care about legality of the nodes here as we know
  // they can be split down into something legal.
  if (DCI.isBeforeLegalizeOps() && ISD::isNormalLoad(N0.getNode()) &&
      N0.hasOneUse() && Subtarget->useSVEForFixedLengthVectors() &&
      VT.isFixedLengthVector() && hasValidElementTypeForFPExtLoad(VT) &&
      VT.getFixedSizeInBits() >= Subtarget->getMinSVEVectorSizeInBits()) {
    LoadSDNode *LN0 = cast<LoadSDNode>(N0);
    SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
                                     LN0->getChain(), LN0->getBasePtr(),
                                     N0.getValueType(), LN0->getMemOperand());
    DCI.CombineTo(N, ExtLoad);
    DCI.CombineTo(
        N0.getNode(),
        DAG.getNode(ISD::FP_ROUND, SDLoc(N0), N0.getValueType(), ExtLoad,
                    DAG.getIntPtrConstant(1, SDLoc(N0), /*isTarget=*/true)),
        ExtLoad.getValue(1));
    return SDValue(N, 0); // Return N so it doesn't get rechecked!
  }

  return SDValue();
}

static SDValue performBSPExpandForSVE(SDNode *N, SelectionDAG &DAG,
                                      const AArch64Subtarget *Subtarget) {
  EVT VT = N->getValueType(0);

  // Don't expand for NEON, SVE2 or SME
  if (!VT.isScalableVector() || Subtarget->hasSVE2() || Subtarget->hasSME())
    return SDValue();

  SDLoc DL(N);

  SDValue Mask = N->getOperand(0);
  SDValue In1 = N->getOperand(1);
  SDValue In2 = N->getOperand(2);

  SDValue InvMask = DAG.getNOT(DL, Mask, VT);
  SDValue Sel = DAG.getNode(ISD::AND, DL, VT, Mask, In1);
  SDValue SelInv = DAG.getNode(ISD::AND, DL, VT, InvMask, In2);
  return DAG.getNode(ISD::OR, DL, VT, Sel, SelInv);
}

static SDValue performDupLane128Combine(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);

  SDValue Insert = N->getOperand(0);
  if (Insert.getOpcode() != ISD::INSERT_SUBVECTOR)
    return SDValue();

  if (!Insert.getOperand(0).isUndef())
    return SDValue();

  uint64_t IdxInsert = Insert.getConstantOperandVal(2);
  uint64_t IdxDupLane = N->getConstantOperandVal(1);
  if (IdxInsert != 0 || IdxDupLane != 0)
    return SDValue();

  SDValue Bitcast = Insert.getOperand(1);
  if (Bitcast.getOpcode() != ISD::BITCAST)
    return SDValue();

  SDValue Subvec = Bitcast.getOperand(0);
  EVT SubvecVT = Subvec.getValueType();
  if (!SubvecVT.is128BitVector())
    return SDValue();
  EVT NewSubvecVT =
      getPackedSVEVectorVT(Subvec.getValueType().getVectorElementType());

  SDLoc DL(N);
  SDValue NewInsert =
      DAG.getNode(ISD::INSERT_SUBVECTOR, DL, NewSubvecVT,
                  DAG.getUNDEF(NewSubvecVT), Subvec, Insert->getOperand(2));
  SDValue NewDuplane128 = DAG.getNode(AArch64ISD::DUPLANE128, DL, NewSubvecVT,
                                      NewInsert, N->getOperand(1));
  return DAG.getNode(ISD::BITCAST, DL, VT, NewDuplane128);
}

// Try to combine mull with uzp1.
static SDValue tryCombineMULLWithUZP1(SDNode *N,
                                      TargetLowering::DAGCombinerInfo &DCI,
                                      SelectionDAG &DAG) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  SDValue LHS = N->getOperand(0);
  SDValue RHS = N->getOperand(1);

  SDValue ExtractHigh;
  SDValue ExtractLow;
  SDValue TruncHigh;
  SDValue TruncLow;
  SDLoc DL(N);

  // Check the operands are trunc and extract_high.
  if (isEssentiallyExtractHighSubvector(LHS) &&
      RHS.getOpcode() == ISD::TRUNCATE) {
    TruncHigh = RHS;
    if (LHS.getOpcode() == ISD::BITCAST)
      ExtractHigh = LHS.getOperand(0);
    else
      ExtractHigh = LHS;
  } else if (isEssentiallyExtractHighSubvector(RHS) &&
             LHS.getOpcode() == ISD::TRUNCATE) {
    TruncHigh = LHS;
    if (RHS.getOpcode() == ISD::BITCAST)
      ExtractHigh = RHS.getOperand(0);
    else
      ExtractHigh = RHS;
  } else
    return SDValue();

  // If the truncate's operand is BUILD_VECTOR with DUP, do not combine the op
  // with uzp1.
  // You can see the regressions on test/CodeGen/AArch64/aarch64-smull.ll
  SDValue TruncHighOp = TruncHigh.getOperand(0);
  EVT TruncHighOpVT = TruncHighOp.getValueType();
  if (TruncHighOp.getOpcode() == AArch64ISD::DUP ||
      DAG.isSplatValue(TruncHighOp, false))
    return SDValue();

  // Check there is other extract_high with same source vector.
  // For example,
  //
  //    t18: v4i16 = extract_subvector t2, Constant:i64<0>
  //    t12: v4i16 = truncate t11
  //  t31: v4i32 = AArch64ISD::SMULL t18, t12
  //    t23: v4i16 = extract_subvector t2, Constant:i64<4>
  //    t16: v4i16 = truncate t15
  //  t30: v4i32 = AArch64ISD::SMULL t23, t1
  //
  // This dagcombine assumes the two extract_high uses same source vector in
  // order to detect the pair of the mull. If they have different source vector,
  // this code will not work.
  // TODO: Should also try to look through a bitcast.
  bool HasFoundMULLow = true;
  SDValue ExtractHighSrcVec = ExtractHigh.getOperand(0);
  if (ExtractHighSrcVec->use_size() != 2)
    HasFoundMULLow = false;

  // Find ExtractLow.
  for (SDNode *User : ExtractHighSrcVec.getNode()->users()) {
    if (User == ExtractHigh.getNode())
      continue;

    if (User->getOpcode() != ISD::EXTRACT_SUBVECTOR ||
        !isNullConstant(User->getOperand(1))) {
      HasFoundMULLow = false;
      break;
    }

    ExtractLow.setNode(User);
  }

  if (!ExtractLow || !ExtractLow->hasOneUse())
    HasFoundMULLow = false;

  // Check ExtractLow's user.
  if (HasFoundMULLow) {
    SDNode *ExtractLowUser = *ExtractLow.getNode()->user_begin();
    if (ExtractLowUser->getOpcode() != N->getOpcode()) {
      HasFoundMULLow = false;
    } else {
      if (ExtractLowUser->getOperand(0) == ExtractLow) {
        if (ExtractLowUser->getOperand(1).getOpcode() == ISD::TRUNCATE)
          TruncLow = ExtractLowUser->getOperand(1);
        else
          HasFoundMULLow = false;
      } else {
        if (ExtractLowUser->getOperand(0).getOpcode() == ISD::TRUNCATE)
          TruncLow = ExtractLowUser->getOperand(0);
        else
          HasFoundMULLow = false;
      }
    }
  }

  // If the truncate's operand is BUILD_VECTOR with DUP, do not combine the op
  // with uzp1.
  // You can see the regressions on test/CodeGen/AArch64/aarch64-smull.ll
  EVT TruncHighVT = TruncHigh.getValueType();
  EVT UZP1VT = TruncHighVT.getDoubleNumVectorElementsVT(*DAG.getContext());
  SDValue TruncLowOp =
      HasFoundMULLow ? TruncLow.getOperand(0) : DAG.getUNDEF(UZP1VT);
  EVT TruncLowOpVT = TruncLowOp.getValueType();
  if (HasFoundMULLow && (TruncLowOp.getOpcode() == AArch64ISD::DUP ||
                         DAG.isSplatValue(TruncLowOp, false)))
    return SDValue();

  // Create uzp1, extract_high and extract_low.
  if (TruncHighOpVT != UZP1VT)
    TruncHighOp = DAG.getNode(ISD::BITCAST, DL, UZP1VT, TruncHighOp);
  if (TruncLowOpVT != UZP1VT)
    TruncLowOp = DAG.getNode(ISD::BITCAST, DL, UZP1VT, TruncLowOp);

  SDValue UZP1 =
      DAG.getNode(AArch64ISD::UZP1, DL, UZP1VT, TruncLowOp, TruncHighOp);
  SDValue HighIdxCst =
      DAG.getConstant(TruncHighVT.getVectorNumElements(), DL, MVT::i64);
  SDValue NewTruncHigh =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, TruncHighVT, UZP1, HighIdxCst);
  DAG.ReplaceAllUsesWith(TruncHigh, NewTruncHigh);

  if (HasFoundMULLow) {
    EVT TruncLowVT = TruncLow.getValueType();
    SDValue NewTruncLow = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, TruncLowVT,
                                      UZP1, ExtractLow.getOperand(1));
    DAG.ReplaceAllUsesWith(TruncLow, NewTruncLow);
  }

  return SDValue(N, 0);
}

static SDValue performMULLCombine(SDNode *N,
                                  TargetLowering::DAGCombinerInfo &DCI,
                                  SelectionDAG &DAG) {
  if (SDValue Val =
          tryCombineLongOpWithDup(Intrinsic::not_intrinsic, N, DCI, DAG))
    return Val;

  if (SDValue Val = tryCombineMULLWithUZP1(N, DCI, DAG))
    return Val;

  return SDValue();
}

static SDValue
performScalarToVectorCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI,
                             SelectionDAG &DAG) {
  // Let's do below transform.
  //
  //         t34: v4i32 = AArch64ISD::UADDLV t2
  //       t35: i32 = extract_vector_elt t34, Constant:i64<0>
  //     t7: i64 = zero_extend t35
  //   t20: v1i64 = scalar_to_vector t7
  // ==>
  //      t34: v4i32 = AArch64ISD::UADDLV t2
  //    t39: v2i32 = extract_subvector t34, Constant:i64<0>
  //  t40: v1i64 = AArch64ISD::NVCAST t39
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  EVT VT = N->getValueType(0);
  if (VT != MVT::v1i64)
    return SDValue();

  SDValue ZEXT = N->getOperand(0);
  if (ZEXT.getOpcode() != ISD::ZERO_EXTEND || ZEXT.getValueType() != MVT::i64)
    return SDValue();

  SDValue EXTRACT_VEC_ELT = ZEXT.getOperand(0);
  if (EXTRACT_VEC_ELT.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
      EXTRACT_VEC_ELT.getValueType() != MVT::i32)
    return SDValue();

  if (!isNullConstant(EXTRACT_VEC_ELT.getOperand(1)))
    return SDValue();

  SDValue UADDLV = EXTRACT_VEC_ELT.getOperand(0);
  if (UADDLV.getOpcode() != AArch64ISD::UADDLV ||
      UADDLV.getValueType() != MVT::v4i32 ||
      UADDLV.getOperand(0).getValueType() != MVT::v8i8)
    return SDValue();

  // Let's generate new sequence with AArch64ISD::NVCAST.
  SDLoc DL(N);
  SDValue EXTRACT_SUBVEC =
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, UADDLV,
                  DAG.getConstant(0, DL, MVT::i64));
  SDValue NVCAST =
      DAG.getNode(AArch64ISD::NVCAST, DL, MVT::v1i64, EXTRACT_SUBVEC);

  return NVCAST;
}

/// If the operand is a bitwise AND with a constant RHS, and the shift has a
/// constant RHS and is the only use, we can pull it out of the shift, i.e.
///
///   (shl (and X, C1), C2) -> (and (shl X, C2), (shl C1, C2))
///
/// We prefer this canonical form to match existing isel patterns.
static SDValue performSHLCombine(SDNode *N,
                                 TargetLowering::DAGCombinerInfo &DCI,
                                 SelectionDAG &DAG) {
  if (DCI.isBeforeLegalizeOps())
    return SDValue();

  SDValue Op0 = N->getOperand(0);
  if (Op0.getOpcode() != ISD::AND || !Op0.hasOneUse())
    return SDValue();

  SDValue C1 = Op0->getOperand(1);
  SDValue C2 = N->getOperand(1);
  if (!isa<ConstantSDNode>(C1) || !isa<ConstantSDNode>(C2))
    return SDValue();

  // Might be folded into shifted op, do not lower.
  if (N->hasOneUse()) {
    unsigned UseOpc = N->user_begin()->getOpcode();
    if (UseOpc == ISD::ADD || UseOpc == ISD::SUB || UseOpc == ISD::SETCC ||
        UseOpc == AArch64ISD::ADDS || UseOpc == AArch64ISD::SUBS)
      return SDValue();
  }

  SDLoc DL(N);
  EVT VT = N->getValueType(0);

  // Don't combine unless (shl C1, C2) can be constant folded. Otherwise,
  // DAGCombiner will simplify (and (op x...), (op y...)) -> (op (and x, y))
  // causing infinite loop. Result may also be worse.
  SDValue NewRHS = DAG.getNode(ISD::SHL, DL, VT, C1, C2);
  if (!isa<ConstantSDNode>(NewRHS))
    return SDValue();

  SDValue X = Op0->getOperand(0);
  SDValue NewShift = DAG.getNode(ISD::SHL, DL, VT, X, C2);
  return DAG.getNode(ISD::AND, DL, VT, NewShift, NewRHS);
}

SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
                                                 DAGCombinerInfo &DCI) const {
  SelectionDAG &DAG = DCI.DAG;
  switch (N->getOpcode()) {
  default:
    LLVM_DEBUG(dbgs() << "Custom combining: skipping\n");
    break;
  case ISD::VECREDUCE_AND:
  case ISD::VECREDUCE_OR:
  case ISD::VECREDUCE_XOR:
    return performVecReduceBitwiseCombine(N, DCI, DAG);
  case ISD::ADD:
  case ISD::SUB:
    return performAddSubCombine(N, DCI);
  case ISD::BUILD_VECTOR:
    return performBuildVectorCombine(N, DCI, DAG);
  case ISD::SMIN:
    return performSMINCombine(N, DAG);
  case ISD::TRUNCATE:
    return performTruncateCombine(N, DAG, DCI);
  case AArch64ISD::ANDS:
    return performFlagSettingCombine(N, DCI, ISD::AND);
  case AArch64ISD::ADC:
    if (auto R = foldOverflowCheck(N, DAG, /* IsAdd */ true))
      return R;
    return foldADCToCINC(N, DAG);
  case AArch64ISD::SBC:
    return foldOverflowCheck(N, DAG, /* IsAdd */ false);
  case AArch64ISD::ADCS:
    if (auto R = foldOverflowCheck(N, DAG, /* IsAdd */ true))
      return R;
    return performFlagSettingCombine(N, DCI, AArch64ISD::ADC);
  case AArch64ISD::SBCS:
    if (auto R = foldOverflowCheck(N, DAG, /* IsAdd */ false))
      return R;
    return performFlagSettingCombine(N, DCI, AArch64ISD::SBC);
  case AArch64ISD::BICi: {
    APInt DemandedBits =
        APInt::getAllOnes(N->getValueType(0).getScalarSizeInBits());
    APInt DemandedElts =
        APInt::getAllOnes(N->getValueType(0).getVectorNumElements());

    if (DAG.getTargetLoweringInfo().SimplifyDemandedBits(
            SDValue(N, 0), DemandedBits, DemandedElts, DCI))
      return SDValue();

    break;
  }
  case ISD::XOR:
    return performXorCombine(N, DAG, DCI, Subtarget);
  case ISD::MUL:
    return performMulCombine(N, DAG, DCI, Subtarget);
  case ISD::SINT_TO_FP:
  case ISD::UINT_TO_FP:
    return performIntToFpCombine(N, DAG, DCI, Subtarget);
  case ISD::FP_TO_SINT:
  case ISD::FP_TO_UINT:
  case ISD::FP_TO_SINT_SAT:
  case ISD::FP_TO_UINT_SAT:
    return performFpToIntCombine(N, DAG, DCI, Subtarget);
  case ISD::OR:
    return performORCombine(N, DCI, Subtarget, *this);
  case ISD::AND:
    return performANDCombine(N, DCI);
  case ISD::FADD:
    return performFADDCombine(N, DCI);
  case ISD::INTRINSIC_WO_CHAIN:
    return performIntrinsicCombine(N, DCI, Subtarget);
  case ISD::ANY_EXTEND:
  case ISD::ZERO_EXTEND:
  case ISD::SIGN_EXTEND:
    return performExtendCombine(N, DCI, DAG);
  case ISD::SIGN_EXTEND_INREG:
    return performSignExtendInRegCombine(N, DCI, DAG);
  case ISD::CONCAT_VECTORS:
    return performConcatVectorsCombine(N, DCI, DAG);
  case ISD::EXTRACT_SUBVECTOR:
    return performExtractSubvectorCombine(N, DCI, DAG);
  case ISD::INSERT_SUBVECTOR:
    return performInsertSubvectorCombine(N, DCI, DAG);
  case ISD::SELECT:
    return performSelectCombine(N, DCI);
  case ISD::VSELECT:
    return performVSelectCombine(N, DCI.DAG);
  case ISD::SETCC:
    return performSETCCCombine(N, DCI, DAG);
  case ISD::LOAD:
    return performLOADCombine(N, DCI, DAG, Subtarget);
  case ISD::STORE:
    return performSTORECombine(N, DCI, DAG, Subtarget);
  case ISD::MSTORE:
    return performMSTORECombine(N, DCI, DAG, Subtarget);
  case ISD::MGATHER:
  case ISD::MSCATTER:
  case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM:
    return performMaskedGatherScatterCombine(N, DCI, DAG);
  case ISD::FP_EXTEND:
    return performFPExtendCombine(N, DAG, DCI, Subtarget);
  case AArch64ISD::BRCOND:
    return performBRCONDCombine(N, DCI, DAG);
  case AArch64ISD::TBNZ:
  case AArch64ISD::TBZ:
    return performTBZCombine(N, DCI, DAG);
  case AArch64ISD::CSEL:
    return performCSELCombine(N, DCI, DAG);
  case AArch64ISD::DUP:
  case AArch64ISD::DUPLANE8:
  case AArch64ISD::DUPLANE16:
  case AArch64ISD::DUPLANE32:
  case AArch64ISD::DUPLANE64:
    return performDUPCombine(N, DCI);
  case AArch64ISD::DUPLANE128:
    return performDupLane128Combine(N, DAG);
  case AArch64ISD::NVCAST:
    return performNVCASTCombine(N, DAG);
  case AArch64ISD::SPLICE:
    return performSpliceCombine(N, DAG);
  case AArch64ISD::UUNPKLO:
  case AArch64ISD::UUNPKHI:
    return performUnpackCombine(N, DAG, Subtarget);
  case AArch64ISD::UZP1:
  case AArch64ISD::UZP2:
    return performUzpCombine(N, DAG, Subtarget);
  case AArch64ISD::SETCC_MERGE_ZERO:
    return performSetccMergeZeroCombine(N, DCI);
  case AArch64ISD::REINTERPRET_CAST:
    return performReinterpretCastCombine(N);
  case AArch64ISD::GLD1_MERGE_ZERO:
  case AArch64ISD::GLD1_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1_UXTW_MERGE_ZERO:
  case AArch64ISD::GLD1_SXTW_MERGE_ZERO:
  case AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1_IMM_MERGE_ZERO:
  case AArch64ISD::GLD1S_MERGE_ZERO:
  case AArch64ISD::GLD1S_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1S_UXTW_MERGE_ZERO:
  case AArch64ISD::GLD1S_SXTW_MERGE_ZERO:
  case AArch64ISD::GLD1S_UXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1S_SXTW_SCALED_MERGE_ZERO:
  case AArch64ISD::GLD1S_IMM_MERGE_ZERO:
    return performGLD1Combine(N, DAG);
  case AArch64ISD::VASHR:
  case AArch64ISD::VLSHR:
    return performVectorShiftCombine(N, *this, DCI);
  case AArch64ISD::SUNPKLO:
    return performSunpkloCombine(N, DAG);
  case AArch64ISD::BSP:
    return performBSPExpandForSVE(N, DAG, Subtarget);
  case ISD::INSERT_VECTOR_ELT:
    return performInsertVectorEltCombine(N, DCI);
  case ISD::EXTRACT_VECTOR_ELT:
    return performExtractVectorEltCombine(N, DCI, Subtarget);
  case ISD::VECREDUCE_ADD:
    return performVecReduceAddCombine(N, DCI.DAG, Subtarget);
  case ISD::GET_ACTIVE_LANE_MASK:
    return performActiveLaneMaskCombine(N, DCI, Subtarget);
  case AArch64ISD::UADDV:
    return performUADDVCombine(N, DAG);
  case AArch64ISD::SMULL:
  case AArch64ISD::UMULL:
  case AArch64ISD::PMULL:
    return performMULLCombine(N, DCI, DAG);
  case ISD::INTRINSIC_VOID:
  case ISD::INTRINSIC_W_CHAIN:
    switch (N->getConstantOperandVal(1)) {
    case Intrinsic::aarch64_sve_prfb_gather_scalar_offset:
      return combineSVEPrefetchVecBaseImmOff(N, DAG, 1 /*=ScalarSizeInBytes*/);
    case Intrinsic::aarch64_sve_prfh_gather_scalar_offset:
      return combineSVEPrefetchVecBaseImmOff(N, DAG, 2 /*=ScalarSizeInBytes*/);
    case Intrinsic::aarch64_sve_prfw_gather_scalar_offset:
      return combineSVEPrefetchVecBaseImmOff(N, DAG, 4 /*=ScalarSizeInBytes*/);
    case Intrinsic::aarch64_sve_prfd_gather_scalar_offset:
      return combineSVEPrefetchVecBaseImmOff(N, DAG, 8 /*=ScalarSizeInBytes*/);
    case Intrinsic::aarch64_sve_prfb_gather_uxtw_index:
    case Intrinsic::aarch64_sve_prfb_gather_sxtw_index:
    case Intrinsic::aarch64_sve_prfh_gather_uxtw_index:
    case Intrinsic::aarch64_sve_prfh_gather_sxtw_index:
    case Intrinsic::aarch64_sve_prfw_gather_uxtw_index:
    case Intrinsic::aarch64_sve_prfw_gather_sxtw_index:
    case Intrinsic::aarch64_sve_prfd_gather_uxtw_index:
    case Intrinsic::aarch64_sve_prfd_gather_sxtw_index:
      return legalizeSVEGatherPrefetchOffsVec(N, DAG);
    case Intrinsic::aarch64_neon_ld2:
    case Intrinsic::aarch64_neon_ld3:
    case Intrinsic::aarch64_neon_ld4:
    case Intrinsic::aarch64_neon_ld1x2:
    case Intrinsic::aarch64_neon_ld1x3:
    case Intrinsic::aarch64_neon_ld1x4:
    case Intrinsic::aarch64_neon_ld2lane:
    case Intrinsic::aarch64_neon_ld3lane:
    case Intrinsic::aarch64_neon_ld4lane:
    case Intrinsic::aarch64_neon_ld2r:
    case Intrinsic::aarch64_neon_ld3r:
    case Intrinsic::aarch64_neon_ld4r:
    case Intrinsic::aarch64_neon_st2:
    case Intrinsic::aarch64_neon_st3:
    case Intrinsic::aarch64_neon_st4:
    case Intrinsic::aarch64_neon_st1x2:
    case Intrinsic::aarch64_neon_st1x3:
    case Intrinsic::aarch64_neon_st1x4:
    case Intrinsic::aarch64_neon_st2lane:
    case Intrinsic::aarch64_neon_st3lane:
    case Intrinsic::aarch64_neon_st4lane:
      return performNEONPostLDSTCombine(N, DCI, DAG);
    case Intrinsic::aarch64_sve_ldnt1:
      return performLDNT1Combine(N, DAG);
    case Intrinsic::aarch64_sve_ld1rq:
      return performLD1ReplicateCombine<AArch64ISD::LD1RQ_MERGE_ZERO>(N, DAG);
    case Intrinsic::aarch64_sve_ld1ro:
      return performLD1ReplicateCombine<AArch64ISD::LD1RO_MERGE_ZERO>(N, DAG);
    case Intrinsic::aarch64_sve_ldnt1_gather_scalar_offset:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLDNT1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldnt1_gather:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLDNT1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldnt1_gather_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLDNT1_INDEX_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldnt1_gather_uxtw:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLDNT1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ld1:
      return performLD1Combine(N, DAG, AArch64ISD::LD1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldnf1:
      return performLD1Combine(N, DAG, AArch64ISD::LDNF1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldff1:
      return performLD1Combine(N, DAG, AArch64ISD::LDFF1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_st1:
      return performST1Combine(N, DAG);
    case Intrinsic::aarch64_sve_stnt1:
      return performSTNT1Combine(N, DAG);
    case Intrinsic::aarch64_sve_stnt1_scatter_scalar_offset:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_PRED);
    case Intrinsic::aarch64_sve_stnt1_scatter_uxtw:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_PRED);
    case Intrinsic::aarch64_sve_stnt1_scatter:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_PRED);
    case Intrinsic::aarch64_sve_stnt1_scatter_index:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SSTNT1_INDEX_PRED);
    case Intrinsic::aarch64_sve_ld1_gather:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ld1q_gather_scalar_offset:
    case Intrinsic::aarch64_sve_ld1q_gather_vector_offset:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1Q_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ld1q_gather_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLD1Q_INDEX_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ld1_gather_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLD1_SCALED_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ld1_gather_sxtw:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_SXTW_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ld1_gather_uxtw:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_UXTW_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ld1_gather_sxtw_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLD1_SXTW_SCALED_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ld1_gather_uxtw_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLD1_UXTW_SCALED_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ld1_gather_scalar_offset:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLD1_IMM_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldff1_gather:
      return performGatherLoadCombine(N, DAG, AArch64ISD::GLDFF1_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldff1_gather_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLDFF1_SCALED_MERGE_ZERO);
    case Intrinsic::aarch64_sve_ldff1_gather_sxtw:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLDFF1_SXTW_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ldff1_gather_uxtw:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLDFF1_UXTW_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ldff1_gather_sxtw_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLDFF1_SXTW_SCALED_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ldff1_gather_uxtw_index:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLDFF1_UXTW_SCALED_MERGE_ZERO,
                                      /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_ldff1_gather_scalar_offset:
      return performGatherLoadCombine(N, DAG,
                                      AArch64ISD::GLDFF1_IMM_MERGE_ZERO);
    case Intrinsic::aarch64_sve_st1q_scatter_scalar_offset:
    case Intrinsic::aarch64_sve_st1q_scatter_vector_offset:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SST1Q_PRED);
    case Intrinsic::aarch64_sve_st1q_scatter_index:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SST1Q_INDEX_PRED);
    case Intrinsic::aarch64_sve_st1_scatter:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_PRED);
    case Intrinsic::aarch64_sve_st1_scatter_index:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_SCALED_PRED);
    case Intrinsic::aarch64_sve_st1_scatter_sxtw:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_SXTW_PRED,
                                        /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_st1_scatter_uxtw:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_UXTW_PRED,
                                        /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_st1_scatter_sxtw_index:
      return performScatterStoreCombine(N, DAG,
                                        AArch64ISD::SST1_SXTW_SCALED_PRED,
                                        /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_st1_scatter_uxtw_index:
      return performScatterStoreCombine(N, DAG,
                                        AArch64ISD::SST1_UXTW_SCALED_PRED,
                                        /*OnlyPackedOffsets=*/false);
    case Intrinsic::aarch64_sve_st1_scatter_scalar_offset:
      return performScatterStoreCombine(N, DAG, AArch64ISD::SST1_IMM_PRED);
    case Intrinsic::aarch64_rndr:
    case Intrinsic::aarch64_rndrrs: {
      unsigned IntrinsicID = N->getConstantOperandVal(1);
      auto Register =
          (IntrinsicID == Intrinsic::aarch64_rndr ? AArch64SysReg::RNDR
                                                  : AArch64SysReg::RNDRRS);
      SDLoc DL(N);
      SDValue A = DAG.getNode(
          AArch64ISD::MRS, DL, DAG.getVTList(MVT::i64, MVT::i32, MVT::Other),
          N->getOperand(0), DAG.getConstant(Register, DL, MVT::i32));
      SDValue B = DAG.getNode(
          AArch64ISD::CSINC, DL, MVT::i32, DAG.getConstant(0, DL, MVT::i32),
          DAG.getConstant(0, DL, MVT::i32),
          DAG.getConstant(AArch64CC::NE, DL, MVT::i32), A.getValue(1));
      return DAG.getMergeValues(
          {A, DAG.getZExtOrTrunc(B, DL, MVT::i1), A.getValue(2)}, DL);
    }
    case Intrinsic::aarch64_sme_ldr_zt:
      return DAG.getNode(AArch64ISD::RESTORE_ZT, SDLoc(N),
                         DAG.getVTList(MVT::Other), N->getOperand(0),
                         N->getOperand(2), N->getOperand(3));
    case Intrinsic::aarch64_sme_str_zt:
      return DAG.getNode(AArch64ISD::SAVE_ZT, SDLoc(N),
                         DAG.getVTList(MVT::Other), N->getOperand(0),
                         N->getOperand(2), N->getOperand(3));
    default:
      break;
    }
    break;
  case ISD::GlobalAddress:
    return performGlobalAddressCombine(N, DAG, Subtarget, getTargetMachine());
  case ISD::CTLZ:
    return performCTLZCombine(N, DAG, Subtarget);
  case ISD::SCALAR_TO_VECTOR:
    return performScalarToVectorCombine(N, DCI, DAG);
  case ISD::SHL:
    return performSHLCombine(N, DCI, DAG);
  }
  return SDValue();
}

// Check if the return value is used as only a return value, as otherwise
// we can't perform a tail-call. In particular, we need to check for
// target ISD nodes that are returns and any other "odd" constructs
// that the generic analysis code won't necessarily catch.
bool AArch64TargetLowering::isUsedByReturnOnly(SDNode *N,
                                               SDValue &Chain) const {
  if (N->getNumValues() != 1)
    return false;
  if (!N->hasNUsesOfValue(1, 0))
    return false;

  SDValue TCChain = Chain;
  SDNode *Copy = *N->user_begin();
  if (Copy->getOpcode() == ISD::CopyToReg) {
    // If the copy has a glue operand, we conservatively assume it isn't safe to
    // perform a tail call.
    if (Copy->getOperand(Copy->getNumOperands() - 1).getValueType() ==
        MVT::Glue)
      return false;
    TCChain = Copy->getOperand(0);
  } else if (Copy->getOpcode() != ISD::FP_EXTEND)
    return false;

  bool HasRet = false;
  for (SDNode *Node : Copy->users()) {
    if (Node->getOpcode() != AArch64ISD::RET_GLUE)
      return false;
    HasRet = true;
  }

  if (!HasRet)
    return false;

  Chain = TCChain;
  return true;
}

// Return whether the an instruction can potentially be optimized to a tail
// call. This will cause the optimizers to attempt to move, or duplicate,
// return instructions to help enable tail call optimizations for this
// instruction.
bool AArch64TargetLowering::mayBeEmittedAsTailCall(const CallInst *CI) const {
  return CI->isTailCall();
}

bool AArch64TargetLowering::isIndexingLegal(MachineInstr &MI, Register Base,
                                            Register Offset, bool IsPre,
                                            MachineRegisterInfo &MRI) const {
  auto CstOffset = getIConstantVRegVal(Offset, MRI);
  if (!CstOffset || CstOffset->isZero())
    return false;

  // All of the indexed addressing mode instructions take a signed 9 bit
  // immediate offset. Our CstOffset is a G_PTR_ADD offset so it already
  // encodes the sign/indexing direction.
  return isInt<9>(CstOffset->getSExtValue());
}

bool AArch64TargetLowering::getIndexedAddressParts(SDNode *N, SDNode *Op,
                                                   SDValue &Base,
                                                   SDValue &Offset,
                                                   SelectionDAG &DAG) const {
  if (Op->getOpcode() != ISD::ADD && Op->getOpcode() != ISD::SUB)
    return false;

  // Non-null if there is exactly one user of the loaded value (ignoring chain).
  SDNode *ValOnlyUser = nullptr;
  for (SDUse &U : N->uses()) {
    if (U.getResNo() == 1)
      continue; // Ignore chain.
    if (ValOnlyUser == nullptr)
      ValOnlyUser = U.getUser();
    else {
      ValOnlyUser = nullptr; // Multiple non-chain uses, bail out.
      break;
    }
  }

  auto IsUndefOrZero = [](SDValue V) {
    return V.isUndef() || isNullOrNullSplat(V, /*AllowUndefs*/ true);
  };

  // If the only user of the value is a scalable vector splat, it is
  // preferable to do a replicating load (ld1r*).
  if (ValOnlyUser && ValOnlyUser->getValueType(0).isScalableVector() &&
      (ValOnlyUser->getOpcode() == ISD::SPLAT_VECTOR ||
       (ValOnlyUser->getOpcode() == AArch64ISD::DUP_MERGE_PASSTHRU &&
        IsUndefOrZero(ValOnlyUser->getOperand(2)))))
    return false;

  Base = Op->getOperand(0);
  // All of the indexed addressing mode instructions take a signed
  // 9 bit immediate offset.
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(Op->getOperand(1))) {
    int64_t RHSC = RHS->getSExtValue();
    if (Op->getOpcode() == ISD::SUB)
      RHSC = -(uint64_t)RHSC;
    if (!isInt<9>(RHSC))
      return false;
    // When big-endian VLD1/VST1 are used for vector load and store, and these
    // only allow an offset that's equal to the store size.
    EVT MemType = cast<MemSDNode>(N)->getMemoryVT();
    if (!Subtarget->isLittleEndian() && MemType.isVector() &&
        (uint64_t)RHSC != MemType.getStoreSize())
      return false;
    // Always emit pre-inc/post-inc addressing mode. Use negated constant offset
    // when dealing with subtraction.
    Offset = DAG.getConstant(RHSC, SDLoc(N), RHS->getValueType(0));
    return true;
  }
  return false;
}

bool AArch64TargetLowering::getPreIndexedAddressParts(SDNode *N, SDValue &Base,
                                                      SDValue &Offset,
                                                      ISD::MemIndexedMode &AM,
                                                      SelectionDAG &DAG) const {
  EVT VT;
  SDValue Ptr;
  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
    VT = LD->getMemoryVT();
    Ptr = LD->getBasePtr();
  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
    VT = ST->getMemoryVT();
    Ptr = ST->getBasePtr();
  } else
    return false;

  if (!getIndexedAddressParts(N, Ptr.getNode(), Base, Offset, DAG))
    return false;
  AM = ISD::PRE_INC;
  return true;
}

bool AArch64TargetLowering::getPostIndexedAddressParts(
    SDNode *N, SDNode *Op, SDValue &Base, SDValue &Offset,
    ISD::MemIndexedMode &AM, SelectionDAG &DAG) const {
  EVT VT;
  SDValue Ptr;
  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
    VT = LD->getMemoryVT();
    Ptr = LD->getBasePtr();
  } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
    VT = ST->getMemoryVT();
    Ptr = ST->getBasePtr();
  } else
    return false;

  if (!getIndexedAddressParts(N, Op, Base, Offset, DAG))
    return false;
  // Post-indexing updates the base, so it's not a valid transform
  // if that's not the same as the load's pointer.
  if (Ptr != Base)
    return false;
  AM = ISD::POST_INC;
  return true;
}

static void replaceBoolVectorBitcast(SDNode *N,
                                     SmallVectorImpl<SDValue> &Results,
                                     SelectionDAG &DAG) {
  SDLoc DL(N);
  SDValue Op = N->getOperand(0);
  EVT VT = N->getValueType(0);
  [[maybe_unused]] EVT SrcVT = Op.getValueType();
  assert(SrcVT.isVector() && SrcVT.getVectorElementType() == MVT::i1 &&
         "Must be bool vector.");

  // Special handling for Clang's __builtin_convertvector. For vectors with <8
  // elements, it adds a vector concatenation with undef(s). If we encounter
  // this here, we can skip the concat.
  if (Op.getOpcode() == ISD::CONCAT_VECTORS && !Op.getOperand(0).isUndef()) {
    bool AllUndef = true;
    for (unsigned I = 1; I < Op.getNumOperands(); ++I)
      AllUndef &= Op.getOperand(I).isUndef();

    if (AllUndef)
      Op = Op.getOperand(0);
  }

  SDValue VectorBits = vectorToScalarBitmask(Op.getNode(), DAG);
  if (VectorBits)
    Results.push_back(DAG.getZExtOrTrunc(VectorBits, DL, VT));
}

static void CustomNonLegalBITCASTResults(SDNode *N,
                                         SmallVectorImpl<SDValue> &Results,
                                         SelectionDAG &DAG, EVT ExtendVT,
                                         EVT CastVT) {
  SDLoc DL(N);
  SDValue Op = N->getOperand(0);
  EVT VT = N->getValueType(0);

  // Use SCALAR_TO_VECTOR for lane zero
  SDValue Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtendVT, Op);
  SDValue CastVal = DAG.getNode(ISD::BITCAST, DL, CastVT, Vec);
  SDValue IdxZero = DAG.getVectorIdxConstant(0, DL);
  Results.push_back(
      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, CastVal, IdxZero));
}

void AArch64TargetLowering::ReplaceBITCASTResults(
    SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
  SDLoc DL(N);
  SDValue Op = N->getOperand(0);
  EVT VT = N->getValueType(0);
  EVT SrcVT = Op.getValueType();

  if (VT == MVT::v2i16 && SrcVT == MVT::i32) {
    CustomNonLegalBITCASTResults(N, Results, DAG, MVT::v2i32, MVT::v4i16);
    return;
  }

  if (VT == MVT::v4i8 && SrcVT == MVT::i32) {
    CustomNonLegalBITCASTResults(N, Results, DAG, MVT::v2i32, MVT::v8i8);
    return;
  }

  if (VT == MVT::v2i8 && SrcVT == MVT::i16) {
    CustomNonLegalBITCASTResults(N, Results, DAG, MVT::v4i16, MVT::v8i8);
    return;
  }

  if (VT.isScalableVector() && !isTypeLegal(VT) && isTypeLegal(SrcVT)) {
    assert(!VT.isFloatingPoint() && SrcVT.isFloatingPoint() &&
           "Expected fp->int bitcast!");

    // Bitcasting between unpacked vector types of different element counts is
    // not a NOP because the live elements are laid out differently.
    //                01234567
    // e.g. nxv2i32 = XX??XX??
    //      nxv4f16 = X?X?X?X?
    if (VT.getVectorElementCount() != SrcVT.getVectorElementCount())
      return;

    SDValue CastResult = getSVESafeBitCast(getSVEContainerType(VT), Op, DAG);
    Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, CastResult));
    return;
  }

  if (SrcVT.isVector() && SrcVT.getVectorElementType() == MVT::i1 &&
      !VT.isVector())
    return replaceBoolVectorBitcast(N, Results, DAG);

  if (VT != MVT::i16 || (SrcVT != MVT::f16 && SrcVT != MVT::bf16))
    return;

  Op = DAG.getTargetInsertSubreg(AArch64::hsub, DL, MVT::f32,
                                 DAG.getUNDEF(MVT::i32), Op);
  Op = DAG.getNode(ISD::BITCAST, DL, MVT::i32, Op);
  Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Op));
}

static void ReplaceAddWithADDP(SDNode *N, SmallVectorImpl<SDValue> &Results,
                               SelectionDAG &DAG,
                               const AArch64Subtarget *Subtarget) {
  EVT VT = N->getValueType(0);
  if (!VT.is256BitVector() ||
      (VT.getScalarType().isFloatingPoint() &&
       !N->getFlags().hasAllowReassociation()) ||
      (VT.getScalarType() == MVT::f16 && !Subtarget->hasFullFP16()) ||
      VT.getScalarType() == MVT::bf16)
    return;

  SDValue X = N->getOperand(0);
  auto *Shuf = dyn_cast<ShuffleVectorSDNode>(N->getOperand(1));
  if (!Shuf) {
    Shuf = dyn_cast<ShuffleVectorSDNode>(N->getOperand(0));
    X = N->getOperand(1);
    if (!Shuf)
      return;
  }

  if (Shuf->getOperand(0) != X || !Shuf->getOperand(1)->isUndef())
    return;

  // Check the mask is 1,0,3,2,5,4,...
  ArrayRef<int> Mask = Shuf->getMask();
  for (int I = 0, E = Mask.size(); I < E; I++)
    if (Mask[I] != (I % 2 == 0 ? I + 1 : I - 1))
      return;

  SDLoc DL(N);
  auto LoHi = DAG.SplitVector(X, DL);
  assert(LoHi.first.getValueType() == LoHi.second.getValueType());
  SDValue Addp = DAG.getNode(AArch64ISD::ADDP, N, LoHi.first.getValueType(),
                             LoHi.first, LoHi.second);

  // Shuffle the elements back into order.
  SmallVector<int> NMask;
  for (unsigned I = 0, E = VT.getVectorNumElements() / 2; I < E; I++) {
    NMask.push_back(I);
    NMask.push_back(I);
  }
  Results.push_back(
      DAG.getVectorShuffle(VT, DL,
                           DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Addp,
                                       DAG.getUNDEF(LoHi.first.getValueType())),
                           DAG.getUNDEF(VT), NMask));
}

static void ReplaceReductionResults(SDNode *N,
                                    SmallVectorImpl<SDValue> &Results,
                                    SelectionDAG &DAG, unsigned InterOp,
                                    unsigned AcrossOp) {
  EVT LoVT, HiVT;
  SDValue Lo, Hi;
  SDLoc DL(N);
  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(N->getValueType(0));
  std::tie(Lo, Hi) = DAG.SplitVectorOperand(N, 0);
  SDValue InterVal = DAG.getNode(InterOp, DL, LoVT, Lo, Hi);
  SDValue SplitVal = DAG.getNode(AcrossOp, DL, LoVT, InterVal);
  Results.push_back(SplitVal);
}

void AArch64TargetLowering::ReplaceExtractSubVectorResults(
    SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
  SDValue In = N->getOperand(0);
  EVT InVT = In.getValueType();

  // Common code will handle these just fine.
  if (!InVT.isScalableVector() || !InVT.isInteger())
    return;

  SDLoc DL(N);
  EVT VT = N->getValueType(0);

  // The following checks bail if this is not a halving operation.

  ElementCount ResEC = VT.getVectorElementCount();

  if (InVT.getVectorElementCount() != (ResEC * 2))
    return;

  auto *CIndex = dyn_cast<ConstantSDNode>(N->getOperand(1));
  if (!CIndex)
    return;

  unsigned Index = CIndex->getZExtValue();
  if ((Index != 0) && (Index != ResEC.getKnownMinValue()))
    return;

  unsigned Opcode = (Index == 0) ? AArch64ISD::UUNPKLO : AArch64ISD::UUNPKHI;
  EVT ExtendedHalfVT = VT.widenIntegerVectorElementType(*DAG.getContext());

  SDValue Half = DAG.getNode(Opcode, DL, ExtendedHalfVT, N->getOperand(0));
  Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, Half));
}

void AArch64TargetLowering::ReplaceGetActiveLaneMaskResults(
    SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
  assert((Subtarget->hasSVE2p1() ||
          (Subtarget->hasSME2() && Subtarget->isStreaming())) &&
         "Custom lower of get.active.lane.mask missing required feature.");

  assert(N->getValueType(0) == MVT::nxv32i1 &&
         "Unexpected result type for get.active.lane.mask");

  SDLoc DL(N);
  SDValue Idx = N->getOperand(0);
  SDValue TC = N->getOperand(1);

  assert(Idx.getValueType().getFixedSizeInBits() <= 64 &&
         "Unexpected operand type for get.active.lane.mask");

  if (Idx.getValueType() != MVT::i64) {
    Idx = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Idx);
    TC = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, TC);
  }

  SDValue ID =
      DAG.getTargetConstant(Intrinsic::aarch64_sve_whilelo_x2, DL, MVT::i64);
  EVT HalfVT = N->getValueType(0).getHalfNumVectorElementsVT(*DAG.getContext());
  auto WideMask =
      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, {HalfVT, HalfVT}, {ID, Idx, TC});

  Results.push_back(DAG.getNode(ISD::CONCAT_VECTORS, DL, N->getValueType(0),
                                {WideMask.getValue(0), WideMask.getValue(1)}));
}

// Create an even/odd pair of X registers holding integer value V.
static SDValue createGPRPairNode(SelectionDAG &DAG, SDValue V) {
  SDLoc DL(V.getNode());
  auto [VLo, VHi] = DAG.SplitScalar(V, DL, MVT::i64, MVT::i64);
  if (DAG.getDataLayout().isBigEndian())
    std::swap (VLo, VHi);
  SDValue RegClass =
      DAG.getTargetConstant(AArch64::XSeqPairsClassRegClassID, DL, MVT::i32);
  SDValue SubReg0 = DAG.getTargetConstant(AArch64::sube64, DL, MVT::i32);
  SDValue SubReg1 = DAG.getTargetConstant(AArch64::subo64, DL, MVT::i32);
  const SDValue Ops[] = { RegClass, VLo, SubReg0, VHi, SubReg1 };
  return SDValue(
      DAG.getMachineNode(TargetOpcode::REG_SEQUENCE, DL, MVT::Untyped, Ops), 0);
}

static void ReplaceCMP_SWAP_128Results(SDNode *N,
                                       SmallVectorImpl<SDValue> &Results,
                                       SelectionDAG &DAG,
                                       const AArch64Subtarget *Subtarget) {
  assert(N->getValueType(0) == MVT::i128 &&
         "AtomicCmpSwap on types less than 128 should be legal");

  MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
  if (Subtarget->hasLSE() || Subtarget->outlineAtomics()) {
    // LSE has a 128-bit compare and swap (CASP), but i128 is not a legal type,
    // so lower it here, wrapped in REG_SEQUENCE and EXTRACT_SUBREG.
    SDValue Ops[] = {
        createGPRPairNode(DAG, N->getOperand(2)), // Compare value
        createGPRPairNode(DAG, N->getOperand(3)), // Store value
        N->getOperand(1), // Ptr
        N->getOperand(0), // Chain in
    };

    unsigned Opcode;
    switch (MemOp->getMergedOrdering()) {
    case AtomicOrdering::Monotonic:
      Opcode = AArch64::CASPX;
      break;
    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:
      llvm_unreachable("Unexpected ordering!");
    }

    MachineSDNode *CmpSwap = DAG.getMachineNode(
        Opcode, SDLoc(N), DAG.getVTList(MVT::Untyped, MVT::Other), Ops);
    DAG.setNodeMemRefs(CmpSwap, {MemOp});

    unsigned SubReg1 = AArch64::sube64, SubReg2 = AArch64::subo64;
    if (DAG.getDataLayout().isBigEndian())
      std::swap(SubReg1, SubReg2);
    SDValue Lo = DAG.getTargetExtractSubreg(SubReg1, SDLoc(N), MVT::i64,
                                            SDValue(CmpSwap, 0));
    SDValue Hi = DAG.getTargetExtractSubreg(SubReg2, SDLoc(N), MVT::i64,
                                            SDValue(CmpSwap, 0));
    Results.push_back(
        DAG.getNode(ISD::BUILD_PAIR, SDLoc(N), MVT::i128, Lo, Hi));
    Results.push_back(SDValue(CmpSwap, 1)); // Chain out
    return;
  }

  unsigned Opcode;
  switch (MemOp->getMergedOrdering()) {
  case AtomicOrdering::Monotonic:
    Opcode = AArch64::CMP_SWAP_128_MONOTONIC;
    break;
  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:
    llvm_unreachable("Unexpected ordering!");
  }

  SDLoc DL(N);
  auto Desired = DAG.SplitScalar(N->getOperand(2), DL, MVT::i64, MVT::i64);
  auto New = DAG.SplitScalar(N->getOperand(3), DL, MVT::i64, MVT::i64);
  SDValue Ops[] = {N->getOperand(1), Desired.first, Desired.second,
                   New.first,        New.second,    N->getOperand(0)};
  SDNode *CmpSwap = DAG.getMachineNode(
      Opcode, SDLoc(N), DAG.getVTList(MVT::i64, MVT::i64, MVT::i32, MVT::Other),
      Ops);
  DAG.setNodeMemRefs(cast<MachineSDNode>(CmpSwap), {MemOp});

  Results.push_back(DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128,
                                SDValue(CmpSwap, 0), SDValue(CmpSwap, 1)));
  Results.push_back(SDValue(CmpSwap, 3));
}

static unsigned getAtomicLoad128Opcode(unsigned ISDOpcode,
                                       AtomicOrdering Ordering) {
  // ATOMIC_LOAD_CLR only appears when lowering ATOMIC_LOAD_AND (see
  // LowerATOMIC_LOAD_AND). We can't take that approach with 128-bit, because
  // the type is not legal. Therefore we shouldn't expect to see a 128-bit
  // ATOMIC_LOAD_CLR at any point.
  assert(ISDOpcode != ISD::ATOMIC_LOAD_CLR &&
         "ATOMIC_LOAD_AND should be lowered to LDCLRP directly");
  assert(ISDOpcode != ISD::ATOMIC_LOAD_ADD && "There is no 128 bit LDADD");
  assert(ISDOpcode != ISD::ATOMIC_LOAD_SUB && "There is no 128 bit LDSUB");

  if (ISDOpcode == ISD::ATOMIC_LOAD_AND) {
    // The operand will need to be XORed in a separate step.
    switch (Ordering) {
    case AtomicOrdering::Monotonic:
      return AArch64::LDCLRP;
      break;
    case AtomicOrdering::Acquire:
      return AArch64::LDCLRPA;
      break;
    case AtomicOrdering::Release:
      return AArch64::LDCLRPL;
      break;
    case AtomicOrdering::AcquireRelease:
    case AtomicOrdering::SequentiallyConsistent:
      return AArch64::LDCLRPAL;
      break;
    default:
      llvm_unreachable("Unexpected ordering!");
    }
  }

  if (ISDOpcode == ISD::ATOMIC_LOAD_OR) {
    switch (Ordering) {
    case AtomicOrdering::Monotonic:
      return AArch64::LDSETP;
      break;
    case AtomicOrdering::Acquire:
      return AArch64::LDSETPA;
      break;
    case AtomicOrdering::Release:
      return AArch64::LDSETPL;
      break;
    case AtomicOrdering::AcquireRelease:
    case AtomicOrdering::SequentiallyConsistent:
      return AArch64::LDSETPAL;
      break;
    default:
      llvm_unreachable("Unexpected ordering!");
    }
  }

  if (ISDOpcode == ISD::ATOMIC_SWAP) {
    switch (Ordering) {
    case AtomicOrdering::Monotonic:
      return AArch64::SWPP;
      break;
    case AtomicOrdering::Acquire:
      return AArch64::SWPPA;
      break;
    case AtomicOrdering::Release:
      return AArch64::SWPPL;
      break;
    case AtomicOrdering::AcquireRelease:
    case AtomicOrdering::SequentiallyConsistent:
      return AArch64::SWPPAL;
      break;
    default:
      llvm_unreachable("Unexpected ordering!");
    }
  }

  llvm_unreachable("Unexpected ISDOpcode!");
}

static void ReplaceATOMIC_LOAD_128Results(SDNode *N,
                                          SmallVectorImpl<SDValue> &Results,
                                          SelectionDAG &DAG,
                                          const AArch64Subtarget *Subtarget) {
  // LSE128 has a 128-bit RMW ops, but i128 is not a legal type, so lower it
  // here. This follows the approach of the CMP_SWAP_XXX pseudo instructions
  // rather than the CASP instructions, because CASP has register classes for
  // the pairs of registers and therefore uses REG_SEQUENCE and EXTRACT_SUBREG
  // to present them as single operands. LSE128 instructions use the GPR64
  // register class (because the pair does not have to be sequential), like
  // CMP_SWAP_XXX, and therefore we use TRUNCATE and BUILD_PAIR.

  assert(N->getValueType(0) == MVT::i128 &&
         "AtomicLoadXXX on types less than 128 should be legal");

  if (!Subtarget->hasLSE128())
    return;

  MachineMemOperand *MemOp = cast<MemSDNode>(N)->getMemOperand();
  const SDValue &Chain = N->getOperand(0);
  const SDValue &Ptr = N->getOperand(1);
  const SDValue &Val128 = N->getOperand(2);
  std::pair<SDValue, SDValue> Val2x64 =
      DAG.SplitScalar(Val128, SDLoc(Val128), MVT::i64, MVT::i64);

  const unsigned ISDOpcode = N->getOpcode();
  const unsigned MachineOpcode =
      getAtomicLoad128Opcode(ISDOpcode, MemOp->getMergedOrdering());

  if (ISDOpcode == ISD::ATOMIC_LOAD_AND) {
    SDLoc DL(Val128);
    Val2x64.first =
        DAG.getNode(ISD::XOR, DL, MVT::i64,
                    DAG.getAllOnesConstant(DL, MVT::i64), Val2x64.first);
    Val2x64.second =
        DAG.getNode(ISD::XOR, DL, MVT::i64,
                    DAG.getAllOnesConstant(DL, MVT::i64), Val2x64.second);
  }

  SDValue Ops[] = {Val2x64.first, Val2x64.second, Ptr, Chain};
  if (DAG.getDataLayout().isBigEndian())
    std::swap(Ops[0], Ops[1]);

  MachineSDNode *AtomicInst =
      DAG.getMachineNode(MachineOpcode, SDLoc(N),
                         DAG.getVTList(MVT::i64, MVT::i64, MVT::Other), Ops);

  DAG.setNodeMemRefs(AtomicInst, {MemOp});

  SDValue Lo = SDValue(AtomicInst, 0), Hi = SDValue(AtomicInst, 1);
  if (DAG.getDataLayout().isBigEndian())
    std::swap(Lo, Hi);

  Results.push_back(DAG.getNode(ISD::BUILD_PAIR, SDLoc(N), MVT::i128, Lo, Hi));
  Results.push_back(SDValue(AtomicInst, 2)); // Chain out
}

void AArch64TargetLowering::ReplaceNodeResults(
    SDNode *N, SmallVectorImpl<SDValue> &Results, SelectionDAG &DAG) const {
  switch (N->getOpcode()) {
  default:
    llvm_unreachable("Don't know how to custom expand this");
  case ISD::BITCAST:
    ReplaceBITCASTResults(N, Results, DAG);
    return;
  case ISD::VECREDUCE_ADD:
  case ISD::VECREDUCE_SMAX:
  case ISD::VECREDUCE_SMIN:
  case ISD::VECREDUCE_UMAX:
  case ISD::VECREDUCE_UMIN:
    Results.push_back(LowerVECREDUCE(SDValue(N, 0), DAG));
    return;
  case ISD::VECTOR_COMPRESS:
    if (SDValue Res = LowerVECTOR_COMPRESS(SDValue(N, 0), DAG))
      Results.push_back(Res);
    return;
  case ISD::ADD:
  case ISD::FADD:
    ReplaceAddWithADDP(N, Results, DAG, Subtarget);
    return;

  case ISD::CTPOP:
  case ISD::PARITY:
    if (SDValue Result = LowerCTPOP_PARITY(SDValue(N, 0), DAG))
      Results.push_back(Result);
    return;
  case AArch64ISD::SADDV:
    ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::SADDV);
    return;
  case AArch64ISD::UADDV:
    ReplaceReductionResults(N, Results, DAG, ISD::ADD, AArch64ISD::UADDV);
    return;
  case AArch64ISD::SMINV:
    ReplaceReductionResults(N, Results, DAG, ISD::SMIN, AArch64ISD::SMINV);
    return;
  case AArch64ISD::UMINV:
    ReplaceReductionResults(N, Results, DAG, ISD::UMIN, AArch64ISD::UMINV);
    return;
  case AArch64ISD::SMAXV:
    ReplaceReductionResults(N, Results, DAG, ISD::SMAX, AArch64ISD::SMAXV);
    return;
  case AArch64ISD::UMAXV:
    ReplaceReductionResults(N, Results, DAG, ISD::UMAX, AArch64ISD::UMAXV);
    return;
  case ISD::MULHS:
    if (useSVEForFixedLengthVectorVT(SDValue(N, 0).getValueType()))
      Results.push_back(
          LowerToPredicatedOp(SDValue(N, 0), DAG, AArch64ISD::MULHS_PRED));
    return;
  case ISD::MULHU:
    if (useSVEForFixedLengthVectorVT(SDValue(N, 0).getValueType()))
      Results.push_back(
          LowerToPredicatedOp(SDValue(N, 0), DAG, AArch64ISD::MULHU_PRED));
    return;
  case ISD::FP_TO_UINT:
  case ISD::FP_TO_SINT:
  case ISD::STRICT_FP_TO_SINT:
  case ISD::STRICT_FP_TO_UINT:
    assert(N->getValueType(0) == MVT::i128 && "unexpected illegal conversion");
    // Let normal code take care of it by not adding anything to Results.
    return;
  case ISD::ATOMIC_CMP_SWAP:
    ReplaceCMP_SWAP_128Results(N, Results, DAG, Subtarget);
    return;
  case ISD::ATOMIC_LOAD_CLR:
    assert(N->getValueType(0) != MVT::i128 &&
           "128-bit ATOMIC_LOAD_AND should be lowered directly to LDCLRP");
    break;
  case ISD::ATOMIC_LOAD_AND:
  case ISD::ATOMIC_LOAD_OR:
  case ISD::ATOMIC_SWAP: {
    assert(cast<AtomicSDNode>(N)->getVal().getValueType() == MVT::i128 &&
           "Expected 128-bit atomicrmw.");
    // These need custom type legalisation so we go directly to instruction.
    ReplaceATOMIC_LOAD_128Results(N, Results, DAG, Subtarget);
    return;
  }
  case ISD::ADDRSPACECAST: {
    SDValue V = LowerADDRSPACECAST(SDValue(N, 0), DAG);
    Results.push_back(V);
    return;
  }
  case ISD::ATOMIC_LOAD:
  case ISD::LOAD: {
    MemSDNode *LoadNode = cast<MemSDNode>(N);
    EVT MemVT = LoadNode->getMemoryVT();
    // Handle lowering 256 bit non temporal loads into LDNP for little-endian
    // targets.
    if (LoadNode->isNonTemporal() && Subtarget->isLittleEndian() &&
        MemVT.getSizeInBits() == 256u &&
        (MemVT.getScalarSizeInBits() == 8u ||
         MemVT.getScalarSizeInBits() == 16u ||
         MemVT.getScalarSizeInBits() == 32u ||
         MemVT.getScalarSizeInBits() == 64u)) {

      SDValue Result = DAG.getMemIntrinsicNode(
          AArch64ISD::LDNP, SDLoc(N),
          DAG.getVTList({MemVT.getHalfNumVectorElementsVT(*DAG.getContext()),
                         MemVT.getHalfNumVectorElementsVT(*DAG.getContext()),
                         MVT::Other}),
          {LoadNode->getChain(), LoadNode->getBasePtr()},
          LoadNode->getMemoryVT(), LoadNode->getMemOperand());

      SDValue Pair = DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(N), MemVT,
                                 Result.getValue(0), Result.getValue(1));
      Results.append({Pair, Result.getValue(2) /* Chain */});
      return;
    }

    if ((!LoadNode->isVolatile() && !LoadNode->isAtomic()) ||
        LoadNode->getMemoryVT() != MVT::i128) {
      // Non-volatile or atomic loads are optimized later in AArch64's load/store
      // optimizer.
      return;
    }

    if (SDValue(N, 0).getValueType() == MVT::i128) {
      auto *AN = dyn_cast<AtomicSDNode>(LoadNode);
      bool isLoadAcquire =
          AN && AN->getSuccessOrdering() == AtomicOrdering::Acquire;
      unsigned Opcode = isLoadAcquire ? AArch64ISD::LDIAPP : AArch64ISD::LDP;

      if (isLoadAcquire)
        assert(Subtarget->hasFeature(AArch64::FeatureRCPC3));

      SDValue Result = DAG.getMemIntrinsicNode(
          Opcode, SDLoc(N), DAG.getVTList({MVT::i64, MVT::i64, MVT::Other}),
          {LoadNode->getChain(), LoadNode->getBasePtr()},
          LoadNode->getMemoryVT(), LoadNode->getMemOperand());

      unsigned FirstRes = DAG.getDataLayout().isBigEndian() ? 1 : 0;

      SDValue Pair =
          DAG.getNode(ISD::BUILD_PAIR, SDLoc(N), MVT::i128,
                      Result.getValue(FirstRes), Result.getValue(1 - FirstRes));
      Results.append({Pair, Result.getValue(2) /* Chain */});
    }
    return;
  }
  case ISD::EXTRACT_SUBVECTOR:
    ReplaceExtractSubVectorResults(N, Results, DAG);
    return;
  case ISD::INSERT_SUBVECTOR:
  case ISD::CONCAT_VECTORS:
    // Custom lowering has been requested for INSERT_SUBVECTOR and
    // CONCAT_VECTORS -- but delegate to common code for result type
    // legalisation
    return;
  case ISD::GET_ACTIVE_LANE_MASK:
    ReplaceGetActiveLaneMaskResults(N, Results, DAG);
    return;
  case ISD::INTRINSIC_WO_CHAIN: {
    EVT VT = N->getValueType(0);

    Intrinsic::ID IntID =
        static_cast<Intrinsic::ID>(N->getConstantOperandVal(0));
    switch (IntID) {
    default:
      return;
    case Intrinsic::aarch64_sve_clasta_n: {
      assert((VT == MVT::i8 || VT == MVT::i16) &&
             "custom lowering for unexpected type");
      SDLoc DL(N);
      auto Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, N->getOperand(2));
      auto V = DAG.getNode(AArch64ISD::CLASTA_N, DL, MVT::i32,
                           N->getOperand(1), Op2, N->getOperand(3));
      Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V));
      return;
    }
    case Intrinsic::aarch64_sve_clastb_n: {
      assert((VT == MVT::i8 || VT == MVT::i16) &&
             "custom lowering for unexpected type");
      SDLoc DL(N);
      auto Op2 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, N->getOperand(2));
      auto V = DAG.getNode(AArch64ISD::CLASTB_N, DL, MVT::i32,
                           N->getOperand(1), Op2, N->getOperand(3));
      Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V));
      return;
    }
    case Intrinsic::aarch64_sve_lasta: {
      assert((VT == MVT::i8 || VT == MVT::i16) &&
             "custom lowering for unexpected type");
      SDLoc DL(N);
      auto V = DAG.getNode(AArch64ISD::LASTA, DL, MVT::i32,
                           N->getOperand(1), N->getOperand(2));
      Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V));
      return;
    }
    case Intrinsic::aarch64_sve_lastb: {
      assert((VT == MVT::i8 || VT == MVT::i16) &&
             "custom lowering for unexpected type");
      SDLoc DL(N);
      auto V = DAG.getNode(AArch64ISD::LASTB, DL, MVT::i32,
                           N->getOperand(1), N->getOperand(2));
      Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V));
      return;
    }
    case Intrinsic::aarch64_sme_in_streaming_mode: {
      SDLoc DL(N);
      SDValue Chain = DAG.getEntryNode();
      SDValue RuntimePStateSM =
          getRuntimePStateSM(DAG, Chain, DL, N->getValueType(0));
      Results.push_back(
          DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, RuntimePStateSM));
      return;
    }
    case Intrinsic::experimental_vector_match: {
      if (!VT.isFixedLengthVector() || VT.getVectorElementType() != MVT::i1)
        return;

      // NOTE: Only trivial type promotion is supported.
      EVT NewVT = getTypeToTransformTo(*DAG.getContext(), VT);
      if (NewVT.getVectorNumElements() != VT.getVectorNumElements())
        return;

      SDLoc DL(N);
      auto V = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, NewVT, N->ops());
      Results.push_back(DAG.getNode(ISD::TRUNCATE, DL, VT, V));
      return;
    }
    }
  }
  case ISD::READ_REGISTER: {
    SDLoc DL(N);
    assert(N->getValueType(0) == MVT::i128 &&
           "READ_REGISTER custom lowering is only for 128-bit sysregs");
    SDValue Chain = N->getOperand(0);
    SDValue SysRegName = N->getOperand(1);

    SDValue Result = DAG.getNode(
        AArch64ISD::MRRS, DL, DAG.getVTList({MVT::i64, MVT::i64, MVT::Other}),
        Chain, SysRegName);

    // Sysregs are not endian. Result.getValue(0) always contains the lower half
    // of the 128-bit System Register value.
    SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i128,
                               Result.getValue(0), Result.getValue(1));
    Results.push_back(Pair);
    Results.push_back(Result.getValue(2)); // Chain
    return;
  }
  }
}

bool AArch64TargetLowering::useLoadStackGuardNode(const Module &M) const {
  if (Subtarget->isTargetAndroid() || Subtarget->isTargetFuchsia())
    return TargetLowering::useLoadStackGuardNode(M);
  return true;
}

unsigned AArch64TargetLowering::combineRepeatedFPDivisors() const {
  // Combine multiple FDIVs with the same divisor into multiple FMULs by the
  // reciprocal if there are three or more FDIVs.
  return 3;
}

TargetLoweringBase::LegalizeTypeAction
AArch64TargetLowering::getPreferredVectorAction(MVT VT) const {
  // During type legalization, we prefer to widen v1i8, v1i16, v1i32  to v8i8,
  // v4i16, v2i32 instead of to promote.
  if (VT == MVT::v1i8 || VT == MVT::v1i16 || VT == MVT::v1i32 ||
      VT == MVT::v1f32)
    return TypeWidenVector;

  return TargetLoweringBase::getPreferredVectorAction(VT);
}

// In v8.4a, ldp and stp instructions are guaranteed to be single-copy atomic
// provided the address is 16-byte aligned.
bool AArch64TargetLowering::isOpSuitableForLDPSTP(const Instruction *I) const {
  if (!Subtarget->hasLSE2())
    return false;

  if (auto LI = dyn_cast<LoadInst>(I))
    return LI->getType()->getPrimitiveSizeInBits() == 128 &&
           LI->getAlign() >= Align(16);

  if (auto SI = dyn_cast<StoreInst>(I))
    return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128 &&
           SI->getAlign() >= Align(16);

  return false;
}

bool AArch64TargetLowering::isOpSuitableForLSE128(const Instruction *I) const {
  if (!Subtarget->hasLSE128())
    return false;

  // Only use SWPP for stores where LSE2 would require a fence. Unlike STP, SWPP
  // will clobber the two registers.
  if (const auto *SI = dyn_cast<StoreInst>(I))
    return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128 &&
           SI->getAlign() >= Align(16) &&
           (SI->getOrdering() == AtomicOrdering::Release ||
            SI->getOrdering() == AtomicOrdering::SequentiallyConsistent);

  if (const auto *RMW = dyn_cast<AtomicRMWInst>(I))
    return RMW->getValOperand()->getType()->getPrimitiveSizeInBits() == 128 &&
           RMW->getAlign() >= Align(16) &&
           (RMW->getOperation() == AtomicRMWInst::Xchg ||
            RMW->getOperation() == AtomicRMWInst::And ||
            RMW->getOperation() == AtomicRMWInst::Or);

  return false;
}

bool AArch64TargetLowering::isOpSuitableForRCPC3(const Instruction *I) const {
  if (!Subtarget->hasLSE2() || !Subtarget->hasRCPC3())
    return false;

  if (auto LI = dyn_cast<LoadInst>(I))
    return LI->getType()->getPrimitiveSizeInBits() == 128 &&
           LI->getAlign() >= Align(16) &&
           LI->getOrdering() == AtomicOrdering::Acquire;

  if (auto SI = dyn_cast<StoreInst>(I))
    return SI->getValueOperand()->getType()->getPrimitiveSizeInBits() == 128 &&
           SI->getAlign() >= Align(16) &&
           SI->getOrdering() == AtomicOrdering::Release;

  return false;
}

bool AArch64TargetLowering::shouldInsertFencesForAtomic(
    const Instruction *I) const {
  if (isOpSuitableForRCPC3(I))
    return false;
  if (isOpSuitableForLSE128(I))
    return false;
  if (isOpSuitableForLDPSTP(I))
    return true;
  return false;
}

bool AArch64TargetLowering::shouldInsertTrailingFenceForAtomicStore(
    const Instruction *I) const {
  // Store-Release instructions only provide seq_cst guarantees when paired with
  // Load-Acquire instructions. MSVC CRT does not use these instructions to
  // implement seq_cst loads and stores, so we need additional explicit fences
  // after memory writes.
  if (!Subtarget->getTargetTriple().isWindowsMSVCEnvironment())
    return false;

  switch (I->getOpcode()) {
  default:
    return false;
  case Instruction::AtomicCmpXchg:
    return cast<AtomicCmpXchgInst>(I)->getSuccessOrdering() ==
           AtomicOrdering::SequentiallyConsistent;
  case Instruction::AtomicRMW:
    return cast<AtomicRMWInst>(I)->getOrdering() ==
           AtomicOrdering::SequentiallyConsistent;
  case Instruction::Store:
    return cast<StoreInst>(I)->getOrdering() ==
           AtomicOrdering::SequentiallyConsistent;
  }
}

// Loads and stores less than 128-bits are already atomic; ones above that
// are doomed anyway, so defer to the default libcall and blame the OS when
// things go wrong.
TargetLoweringBase::AtomicExpansionKind
AArch64TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
  unsigned Size = SI->getValueOperand()->getType()->getPrimitiveSizeInBits();
  if (Size != 128)
    return AtomicExpansionKind::None;
  if (isOpSuitableForRCPC3(SI))
    return AtomicExpansionKind::None;
  if (isOpSuitableForLSE128(SI))
    return AtomicExpansionKind::Expand;
  if (isOpSuitableForLDPSTP(SI))
    return AtomicExpansionKind::None;
  return AtomicExpansionKind::Expand;
}

// Loads and stores less than 128-bits are already atomic; ones above that
// are doomed anyway, so defer to the default libcall and blame the OS when
// things go wrong.
TargetLowering::AtomicExpansionKind
AArch64TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
  unsigned Size = LI->getType()->getPrimitiveSizeInBits();

  if (Size != 128)
    return AtomicExpansionKind::None;
  if (isOpSuitableForRCPC3(LI))
    return AtomicExpansionKind::None;
  // No LSE128 loads
  if (isOpSuitableForLDPSTP(LI))
    return AtomicExpansionKind::None;

  // At -O0, fast-regalloc cannot cope with the live vregs necessary to
  // implement atomicrmw without spilling. If the target address is also on the
  // stack and close enough to the spill slot, this can lead to a situation
  // where the monitor always gets cleared and the atomic operation can never
  // succeed. So at -O0 lower this operation to a CAS loop.
  if (getTargetMachine().getOptLevel() == CodeGenOptLevel::None)
    return AtomicExpansionKind::CmpXChg;

  // Using CAS for an atomic load has a better chance of succeeding under high
  // contention situations. So use it if available.
  return Subtarget->hasLSE() ? AtomicExpansionKind::CmpXChg
                             : AtomicExpansionKind::LLSC;
}

// Return true if the atomic operation expansion will lower to use a library
// call, and is thus ineligible to use an LLSC expansion.
static bool rmwOpMayLowerToLibcall(const AArch64Subtarget &Subtarget,
                                   const AtomicRMWInst *RMW) {
  if (!RMW->isFloatingPointOperation())
    return false;
  switch (RMW->getType()->getScalarType()->getTypeID()) {
  case Type::FloatTyID:
  case Type::DoubleTyID:
  case Type::HalfTyID:
  case Type::BFloatTyID:
    // Will use soft float
    return !Subtarget.hasFPARMv8();
  default:
    // fp128 will emit library calls.
    return true;
  }

  llvm_unreachable("covered type switch");
}

// The "default" for integer RMW operations is to expand to an LL/SC loop.
// However, with the LSE instructions (or outline-atomics mode, which provides
// library routines in place of the LSE-instructions), we can directly emit many
// operations instead.
TargetLowering::AtomicExpansionKind
AArch64TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
  Type *Ty = AI->getType();
  unsigned Size = Ty->getPrimitiveSizeInBits();
  assert(Size <= 128 && "AtomicExpandPass should've handled larger sizes.");

  bool CanUseLSE128 = Subtarget->hasLSE128() && Size == 128 &&
                      (AI->getOperation() == AtomicRMWInst::Xchg ||
                       AI->getOperation() == AtomicRMWInst::Or ||
                       AI->getOperation() == AtomicRMWInst::And);
  if (CanUseLSE128)
    return AtomicExpansionKind::None;

  // If LSFE available, use atomic FP instructions in preference to expansion
  if (Subtarget->hasLSFE() && (AI->getOperation() == AtomicRMWInst::FAdd ||
                               AI->getOperation() == AtomicRMWInst::FMax ||
                               AI->getOperation() == AtomicRMWInst::FMin ||
                               AI->getOperation() == AtomicRMWInst::FMaximum ||
                               AI->getOperation() == AtomicRMWInst::FMinimum))
    return AtomicExpansionKind::None;

  // Nand is not supported in LSE.
  // Leave 128 bits to LLSC or CmpXChg.
  if (AI->getOperation() != AtomicRMWInst::Nand && Size < 128 &&
      !AI->isFloatingPointOperation()) {
    if (Subtarget->hasLSE())
      return AtomicExpansionKind::None;
    if (Subtarget->outlineAtomics()) {
      // [U]Min/[U]Max RWM atomics are used in __sync_fetch_ libcalls so far.
      // Don't outline them unless
      // (1) high level <atomic> support approved:
      //   http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p0493r1.pdf
      // (2) low level libgcc and compiler-rt support implemented by:
      //   min/max outline atomics helpers
      if (AI->getOperation() != AtomicRMWInst::Min &&
          AI->getOperation() != AtomicRMWInst::Max &&
          AI->getOperation() != AtomicRMWInst::UMin &&
          AI->getOperation() != AtomicRMWInst::UMax) {
        return AtomicExpansionKind::None;
      }
    }
  }

  // At -O0, fast-regalloc cannot cope with the live vregs necessary to
  // implement atomicrmw without spilling. If the target address is also on the
  // stack and close enough to the spill slot, this can lead to a situation
  // where the monitor always gets cleared and the atomic operation can never
  // succeed. So at -O0 lower this operation to a CAS loop. Also worthwhile if
  // we have a single CAS instruction that can replace the loop.
  if (getTargetMachine().getOptLevel() == CodeGenOptLevel::None ||
      Subtarget->hasLSE() || rmwOpMayLowerToLibcall(*Subtarget, AI))
    return AtomicExpansionKind::CmpXChg;

  return AtomicExpansionKind::LLSC;
}

TargetLowering::AtomicExpansionKind
AArch64TargetLowering::shouldExpandAtomicCmpXchgInIR(
    AtomicCmpXchgInst *AI) const {
  // If subtarget has LSE, leave cmpxchg intact for codegen.
  if (Subtarget->hasLSE() || Subtarget->outlineAtomics())
    return AtomicExpansionKind::None;
  // At -O0, fast-regalloc cannot cope with the live vregs necessary to
  // implement cmpxchg without spilling. If the address being exchanged is also
  // on the stack and close enough to the spill slot, this can lead to a
  // situation where the monitor always gets cleared and the atomic operation
  // can never succeed. So at -O0 we need a late-expanded pseudo-inst instead.
  if (getTargetMachine().getOptLevel() == CodeGenOptLevel::None)
    return AtomicExpansionKind::None;

  // 128-bit atomic cmpxchg is weird; AtomicExpand doesn't know how to expand
  // it.
  unsigned Size = AI->getCompareOperand()->getType()->getPrimitiveSizeInBits();
  if (Size > 64)
    return AtomicExpansionKind::None;

  return AtomicExpansionKind::LLSC;
}

Value *AArch64TargetLowering::emitLoadLinked(IRBuilderBase &Builder,
                                             Type *ValueTy, Value *Addr,
                                             AtomicOrdering Ord) const {
  Module *M = Builder.GetInsertBlock()->getParent()->getParent();
  bool IsAcquire = isAcquireOrStronger(Ord);

  // Since i128 isn't legal and intrinsics don't get type-lowered, the ldrexd
  // intrinsic must return {i64, i64} and we have to recombine them into a
  // single i128 here.
  if (ValueTy->getPrimitiveSizeInBits() == 128) {
    Intrinsic::ID Int =
        IsAcquire ? Intrinsic::aarch64_ldaxp : Intrinsic::aarch64_ldxp;

    Value *LoHi =
        Builder.CreateIntrinsic(Int, Addr, /*FMFSource=*/nullptr, "lohi");

    Value *Lo = Builder.CreateExtractValue(LoHi, 0, "lo");
    Value *Hi = Builder.CreateExtractValue(LoHi, 1, "hi");

    auto *Int128Ty = Type::getInt128Ty(Builder.getContext());
    Lo = Builder.CreateZExt(Lo, Int128Ty, "lo64");
    Hi = Builder.CreateZExt(Hi, Int128Ty, "hi64");

    Value *Or = Builder.CreateOr(
        Lo, Builder.CreateShl(Hi, ConstantInt::get(Int128Ty, 64)), "val64");
    return Builder.CreateBitCast(Or, ValueTy);
  }

  Type *Tys[] = { Addr->getType() };
  Intrinsic::ID Int =
      IsAcquire ? Intrinsic::aarch64_ldaxr : Intrinsic::aarch64_ldxr;

  const DataLayout &DL = M->getDataLayout();
  IntegerType *IntEltTy = Builder.getIntNTy(DL.getTypeSizeInBits(ValueTy));
  CallInst *CI = Builder.CreateIntrinsic(Int, Tys, Addr);
  CI->addParamAttr(0, Attribute::get(Builder.getContext(),
                                     Attribute::ElementType, IntEltTy));
  Value *Trunc = Builder.CreateTrunc(CI, IntEltTy);

  return Builder.CreateBitCast(Trunc, ValueTy);
}

void AArch64TargetLowering::emitAtomicCmpXchgNoStoreLLBalance(
    IRBuilderBase &Builder) const {
  Builder.CreateIntrinsic(Intrinsic::aarch64_clrex, {});
}

Value *AArch64TargetLowering::emitStoreConditional(IRBuilderBase &Builder,
                                                   Value *Val, Value *Addr,
                                                   AtomicOrdering Ord) const {
  Module *M = Builder.GetInsertBlock()->getParent()->getParent();
  bool IsRelease = isReleaseOrStronger(Ord);

  // Since the intrinsics must have legal type, the i128 intrinsics take two
  // parameters: "i64, i64". We must marshal Val into the appropriate form
  // before the call.
  if (Val->getType()->getPrimitiveSizeInBits() == 128) {
    Intrinsic::ID Int =
        IsRelease ? Intrinsic::aarch64_stlxp : Intrinsic::aarch64_stxp;
    Function *Stxr = Intrinsic::getOrInsertDeclaration(M, Int);
    Type *Int64Ty = Type::getInt64Ty(M->getContext());
    Type *Int128Ty = Type::getInt128Ty(M->getContext());

    Value *CastVal = Builder.CreateBitCast(Val, Int128Ty);

    Value *Lo = Builder.CreateTrunc(CastVal, Int64Ty, "lo");
    Value *Hi =
        Builder.CreateTrunc(Builder.CreateLShr(CastVal, 64), Int64Ty, "hi");
    return Builder.CreateCall(Stxr, {Lo, Hi, Addr});
  }

  Intrinsic::ID Int =
      IsRelease ? Intrinsic::aarch64_stlxr : Intrinsic::aarch64_stxr;
  Type *Tys[] = { Addr->getType() };
  Function *Stxr = Intrinsic::getOrInsertDeclaration(M, Int, Tys);

  const DataLayout &DL = M->getDataLayout();
  IntegerType *IntValTy = Builder.getIntNTy(DL.getTypeSizeInBits(Val->getType()));
  Val = Builder.CreateBitCast(Val, IntValTy);

  CallInst *CI = Builder.CreateCall(
      Stxr, {Builder.CreateZExtOrBitCast(
                 Val, Stxr->getFunctionType()->getParamType(0)),
             Addr});
  CI->addParamAttr(1, Attribute::get(Builder.getContext(),
                                     Attribute::ElementType, Val->getType()));
  return CI;
}

bool AArch64TargetLowering::functionArgumentNeedsConsecutiveRegisters(
    Type *Ty, CallingConv::ID CallConv, bool isVarArg,
    const DataLayout &DL) const {
  if (!Ty->isArrayTy()) {
    const TypeSize &TySize = Ty->getPrimitiveSizeInBits();
    return TySize.isScalable() && TySize.getKnownMinValue() > 128;
  }

  // All non aggregate members of the type must have the same type
  SmallVector<EVT> ValueVTs;
  ComputeValueVTs(*this, DL, Ty, ValueVTs);
  return all_equal(ValueVTs);
}

bool AArch64TargetLowering::shouldNormalizeToSelectSequence(LLVMContext &,
                                                            EVT) const {
  return false;
}

static Value *UseTlsOffset(IRBuilderBase &IRB, unsigned Offset) {
  Module *M = IRB.GetInsertBlock()->getParent()->getParent();
  Function *ThreadPointerFunc = Intrinsic::getOrInsertDeclaration(
      M, Intrinsic::thread_pointer, IRB.getPtrTy());
  return IRB.CreatePointerCast(
      IRB.CreateConstGEP1_32(IRB.getInt8Ty(), IRB.CreateCall(ThreadPointerFunc),
                             Offset),
      IRB.getPtrTy(0));
}

Value *AArch64TargetLowering::getIRStackGuard(IRBuilderBase &IRB) const {
  // Android provides a fixed TLS slot for the stack cookie. See the definition
  // of TLS_SLOT_STACK_GUARD in
  // https://android.googlesource.com/platform/bionic/+/main/libc/platform/bionic/tls_defines.h
  if (Subtarget->isTargetAndroid())
    return UseTlsOffset(IRB, 0x28);

  // Fuchsia is similar.
  // <zircon/tls.h> defines ZX_TLS_STACK_GUARD_OFFSET with this value.
  if (Subtarget->isTargetFuchsia())
    return UseTlsOffset(IRB, -0x10);

  return TargetLowering::getIRStackGuard(IRB);
}

void AArch64TargetLowering::insertSSPDeclarations(Module &M) const {
  // MSVC CRT provides functionalities for stack protection.
  if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment()) {
    // MSVC CRT has a global variable holding security cookie.
    M.getOrInsertGlobal("__security_cookie",
                        PointerType::getUnqual(M.getContext()));

    // MSVC CRT has a function to validate security cookie.
    FunctionCallee SecurityCheckCookie =
        M.getOrInsertFunction(Subtarget->getSecurityCheckCookieName(),
                              Type::getVoidTy(M.getContext()),
                              PointerType::getUnqual(M.getContext()));
    if (Function *F = dyn_cast<Function>(SecurityCheckCookie.getCallee())) {
      F->setCallingConv(CallingConv::Win64);
      F->addParamAttr(0, Attribute::AttrKind::InReg);
    }
    return;
  }
  TargetLowering::insertSSPDeclarations(M);
}

Value *AArch64TargetLowering::getSDagStackGuard(const Module &M) const {
  // MSVC CRT has a global variable holding security cookie.
  if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment())
    return M.getGlobalVariable("__security_cookie");
  return TargetLowering::getSDagStackGuard(M);
}

Function *AArch64TargetLowering::getSSPStackGuardCheck(const Module &M) const {
  // MSVC CRT has a function to validate security cookie.
  if (Subtarget->getTargetTriple().isWindowsMSVCEnvironment())
    return M.getFunction(Subtarget->getSecurityCheckCookieName());
  return TargetLowering::getSSPStackGuardCheck(M);
}

Value *
AArch64TargetLowering::getSafeStackPointerLocation(IRBuilderBase &IRB) const {
  // Android provides a fixed TLS slot for the SafeStack pointer. See the
  // definition of TLS_SLOT_SAFESTACK in
  // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
  if (Subtarget->isTargetAndroid())
    return UseTlsOffset(IRB, 0x48);

  // Fuchsia is similar.
  // <zircon/tls.h> defines ZX_TLS_UNSAFE_SP_OFFSET with this value.
  if (Subtarget->isTargetFuchsia())
    return UseTlsOffset(IRB, -0x8);

  return TargetLowering::getSafeStackPointerLocation(IRB);
}

/// If a physical register, this returns the register that receives the
/// exception address on entry to an EH pad.
Register AArch64TargetLowering::getExceptionPointerRegister(
    const Constant *PersonalityFn) const {
  // FIXME: This is a guess. Has this been defined yet?
  return AArch64::X0;
}

/// If a physical register, this returns the register that receives the
/// exception typeid on entry to a landing pad.
Register AArch64TargetLowering::getExceptionSelectorRegister(
    const Constant *PersonalityFn) const {
  // FIXME: This is a guess. Has this been defined yet?
  return AArch64::X1;
}

bool AArch64TargetLowering::isMaskAndCmp0FoldingBeneficial(
    const Instruction &AndI) const {
  // Only sink 'and' mask to cmp use block if it is masking a single bit, since
  // this is likely to be fold the and/cmp/br into a single tbz instruction.  It
  // may be beneficial to sink in other cases, but we would have to check that
  // the cmp would not get folded into the br to form a cbz for these to be
  // beneficial.
  ConstantInt* Mask = dyn_cast<ConstantInt>(AndI.getOperand(1));
  if (!Mask)
    return false;
  return Mask->getValue().isPowerOf2();
}

bool AArch64TargetLowering::
    shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
        SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
        unsigned OldShiftOpcode, unsigned NewShiftOpcode,
        SelectionDAG &DAG) const {
  // Does baseline recommend not to perform the fold by default?
  if (!TargetLowering::shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
          X, XC, CC, Y, OldShiftOpcode, NewShiftOpcode, DAG))
    return false;
  // Else, if this is a vector shift, prefer 'shl'.
  return X.getValueType().isScalarInteger() || NewShiftOpcode == ISD::SHL;
}

TargetLowering::ShiftLegalizationStrategy
AArch64TargetLowering::preferredShiftLegalizationStrategy(
    SelectionDAG &DAG, SDNode *N, unsigned int ExpansionFactor) const {
  if (DAG.getMachineFunction().getFunction().hasMinSize() &&
      !Subtarget->isTargetWindows() && !Subtarget->isTargetDarwin())
    return ShiftLegalizationStrategy::LowerToLibcall;
  return TargetLowering::preferredShiftLegalizationStrategy(DAG, N,
                                                            ExpansionFactor);
}

void AArch64TargetLowering::initializeSplitCSR(MachineBasicBlock *Entry) const {
  // Update IsSplitCSR in AArch64unctionInfo.
  AArch64FunctionInfo *AFI = Entry->getParent()->getInfo<AArch64FunctionInfo>();
  AFI->setIsSplitCSR(true);
}

void AArch64TargetLowering::insertCopiesSplitCSR(
    MachineBasicBlock *Entry,
    const SmallVectorImpl<MachineBasicBlock *> &Exits) const {
  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  const MCPhysReg *IStart = TRI->getCalleeSavedRegsViaCopy(Entry->getParent());
  if (!IStart)
    return;

  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  MachineRegisterInfo *MRI = &Entry->getParent()->getRegInfo();
  MachineBasicBlock::iterator MBBI = Entry->begin();
  for (const MCPhysReg *I = IStart; *I; ++I) {
    const TargetRegisterClass *RC = nullptr;
    if (AArch64::GPR64RegClass.contains(*I))
      RC = &AArch64::GPR64RegClass;
    else if (AArch64::FPR64RegClass.contains(*I))
      RC = &AArch64::FPR64RegClass;
    else
      llvm_unreachable("Unexpected register class in CSRsViaCopy!");

    Register NewVR = MRI->createVirtualRegister(RC);
    // Create copy from CSR to a virtual register.
    // FIXME: this currently does not emit CFI pseudo-instructions, it works
    // fine for CXX_FAST_TLS since the C++-style TLS access functions should be
    // nounwind. If we want to generalize this later, we may need to emit
    // CFI pseudo-instructions.
    assert(Entry->getParent()->getFunction().hasFnAttribute(
               Attribute::NoUnwind) &&
           "Function should be nounwind in insertCopiesSplitCSR!");
    Entry->addLiveIn(*I);
    BuildMI(*Entry, MBBI, DebugLoc(), TII->get(TargetOpcode::COPY), NewVR)
        .addReg(*I);

    // Insert the copy-back instructions right before the terminator.
    for (auto *Exit : Exits)
      BuildMI(*Exit, Exit->getFirstTerminator(), DebugLoc(),
              TII->get(TargetOpcode::COPY), *I)
          .addReg(NewVR);
  }
}

bool AArch64TargetLowering::isIntDivCheap(EVT VT, AttributeList Attr) const {
  // Integer division on AArch64 is expensive. However, when aggressively
  // optimizing for code size, we prefer to use a div instruction, as it is
  // usually smaller than the alternative sequence.
  // The exception to this is vector division. Since AArch64 doesn't have vector
  // integer division, leaving the division as-is is a loss even in terms of
  // size, because it will have to be scalarized, while the alternative code
  // sequence can be performed in vector form.
  bool OptSize = Attr.hasFnAttr(Attribute::MinSize);
  return OptSize && !VT.isVector();
}

bool AArch64TargetLowering::canMergeStoresTo(unsigned AddressSpace, EVT MemVT,
                                             const MachineFunction &MF) const {
  // Avoid merging stores into fixed-length vectors when Neon is unavailable.
  // In future, we could allow this when SVE is available, but currently,
  // the SVE lowerings for BUILD_VECTOR are limited to a few specific cases (and
  // the general lowering may introduce stack spills/reloads).
  if (MemVT.isFixedLengthVector() && !Subtarget->isNeonAvailable())
    return false;

  // Do not merge to float value size (128 bytes) if no implicit float attribute
  // is set.
  bool NoFloat = MF.getFunction().hasFnAttribute(Attribute::NoImplicitFloat);
  return !NoFloat || MemVT.getSizeInBits() <= 64;
}

bool AArch64TargetLowering::preferIncOfAddToSubOfNot(EVT VT) const {
  // We want inc-of-add for scalars and sub-of-not for vectors.
  return VT.isScalarInteger();
}

bool AArch64TargetLowering::shouldConvertFpToSat(unsigned Op, EVT FPVT,
                                                 EVT VT) const {
  // v8f16 without fp16 need to be extended to v8f32, which is more difficult to
  // legalize.
  if (FPVT == MVT::v8f16 && !Subtarget->hasFullFP16())
    return false;
  if (FPVT == MVT::v8bf16)
    return false;
  return TargetLowering::shouldConvertFpToSat(Op, FPVT, VT);
}

bool AArch64TargetLowering::shouldExpandCmpUsingSelects(EVT VT) const {
  // Expand scalar and SVE operations using selects. Neon vectors prefer sub to
  // avoid vselect becoming bsl / unrolling.
  return !VT.isFixedLengthVector();
}

MachineInstr *
AArch64TargetLowering::EmitKCFICheck(MachineBasicBlock &MBB,
                                     MachineBasicBlock::instr_iterator &MBBI,
                                     const TargetInstrInfo *TII) const {
  assert(MBBI->isCall() && MBBI->getCFIType() &&
         "Invalid call instruction for a KCFI check");

  switch (MBBI->getOpcode()) {
  case AArch64::BLR:
  case AArch64::BLRNoIP:
  case AArch64::TCRETURNri:
  case AArch64::TCRETURNrix16x17:
  case AArch64::TCRETURNrix17:
  case AArch64::TCRETURNrinotx16:
    break;
  default:
    llvm_unreachable("Unexpected CFI call opcode");
  }

  MachineOperand &Target = MBBI->getOperand(0);
  assert(Target.isReg() && "Invalid target operand for an indirect call");
  Target.setIsRenamable(false);

  return BuildMI(MBB, MBBI, MBBI->getDebugLoc(), TII->get(AArch64::KCFI_CHECK))
      .addReg(Target.getReg())
      .addImm(MBBI->getCFIType())
      .getInstr();
}

bool AArch64TargetLowering::enableAggressiveFMAFusion(EVT VT) const {
  return Subtarget->hasAggressiveFMA() && VT.isFloatingPoint();
}

unsigned
AArch64TargetLowering::getVaListSizeInBits(const DataLayout &DL) const {
  if (Subtarget->isTargetDarwin() || Subtarget->isTargetWindows())
    return getPointerTy(DL).getSizeInBits();

  return 3 * getPointerTy(DL).getSizeInBits() + 2 * 32;
}

void AArch64TargetLowering::finalizeLowering(MachineFunction &MF) const {
  MachineFrameInfo &MFI = MF.getFrameInfo();
  // If we have any vulnerable SVE stack objects then the stack protector
  // needs to be placed at the top of the SVE stack area, as the SVE locals
  // are placed above the other locals, so we allocate it as if it were a
  // scalable vector.
  // FIXME: It may be worthwhile having a specific interface for this rather
  // than doing it here in finalizeLowering.
  if (MFI.hasStackProtectorIndex()) {
    for (unsigned int i = 0, e = MFI.getObjectIndexEnd(); i != e; ++i) {
      if (MFI.getStackID(i) == TargetStackID::ScalableVector &&
          MFI.getObjectSSPLayout(i) != MachineFrameInfo::SSPLK_None) {
        MFI.setStackID(MFI.getStackProtectorIndex(),
                       TargetStackID::ScalableVector);
        MFI.setObjectAlignment(MFI.getStackProtectorIndex(), Align(16));
        break;
      }
    }
  }
  MFI.computeMaxCallFrameSize(MF);
  TargetLoweringBase::finalizeLowering(MF);
}

// Unlike X86, we let frame lowering assign offsets to all catch objects.
bool AArch64TargetLowering::needsFixedCatchObjects() const { return true; }

bool AArch64TargetLowering::shouldLocalize(
    const MachineInstr &MI, const TargetTransformInfo *TTI) const {
  auto &MF = *MI.getMF();
  auto &MRI = MF.getRegInfo();
  auto maxUses = [](unsigned RematCost) {
    // A cost of 1 means remats are basically free.
    if (RematCost == 1)
      return std::numeric_limits<unsigned>::max();
    if (RematCost == 2)
      return 2U;

    // Remat is too expensive, only sink if there's one user.
    if (RematCost > 2)
      return 1U;
    llvm_unreachable("Unexpected remat cost");
  };

  unsigned Opc = MI.getOpcode();
  switch (Opc) {
  case TargetOpcode::G_GLOBAL_VALUE: {
    // On Darwin, TLS global vars get selected into function calls, which
    // we don't want localized, as they can get moved into the middle of a
    // another call sequence.
    const GlobalValue &GV = *MI.getOperand(1).getGlobal();
    if (GV.isThreadLocal() && Subtarget->isTargetMachO())
      return false;
    return true; // Always localize G_GLOBAL_VALUE to avoid high reg pressure.
  }
  case TargetOpcode::G_FCONSTANT:
  case TargetOpcode::G_CONSTANT: {
    const ConstantInt *CI;
    unsigned AdditionalCost = 0;

    if (Opc == TargetOpcode::G_CONSTANT)
      CI = MI.getOperand(1).getCImm();
    else {
      LLT Ty = MRI.getType(MI.getOperand(0).getReg());
      // We try to estimate cost of 32/64b fpimms, as they'll likely be
      // materialized as integers.
      if (Ty.getScalarSizeInBits() != 32 && Ty.getScalarSizeInBits() != 64)
        break;
      auto APF = MI.getOperand(1).getFPImm()->getValueAPF();
      bool OptForSize = MF.getFunction().hasOptSize();
      if (isFPImmLegal(APF, EVT::getFloatingPointVT(Ty.getScalarSizeInBits()),
                       OptForSize))
        return true; // Constant should be cheap.
      CI =
          ConstantInt::get(MF.getFunction().getContext(), APF.bitcastToAPInt());
      // FP materialization also costs an extra move, from gpr to fpr.
      AdditionalCost = 1;
    }
    APInt Imm = CI->getValue();
    InstructionCost Cost = TTI->getIntImmCost(
        Imm, CI->getType(), TargetTransformInfo::TCK_CodeSize);
    assert(Cost.isValid() && "Expected a valid imm cost");

    unsigned RematCost = Cost.getValue();
    RematCost += AdditionalCost;
    Register Reg = MI.getOperand(0).getReg();
    unsigned MaxUses = maxUses(RematCost);
    // Don't pass UINT_MAX sentinel value to hasAtMostUserInstrs().
    if (MaxUses == std::numeric_limits<unsigned>::max())
      --MaxUses;
    return MRI.hasAtMostUserInstrs(Reg, MaxUses);
  }
  // If we legalized G_GLOBAL_VALUE into ADRP + G_ADD_LOW, mark both as being
  // localizable.
  case AArch64::ADRP:
  case AArch64::G_ADD_LOW:
  // Need to localize G_PTR_ADD so that G_GLOBAL_VALUE can be localized too.
  case TargetOpcode::G_PTR_ADD:
    return true;
  default:
    break;
  }
  return TargetLoweringBase::shouldLocalize(MI, TTI);
}

bool AArch64TargetLowering::fallBackToDAGISel(const Instruction &Inst) const {
  // Fallback for scalable vectors.
  // Note that if EnableSVEGISel is true, we allow scalable vector types for
  // all instructions, regardless of whether they are actually supported.
  if (!EnableSVEGISel) {
    if (Inst.getType()->isScalableTy()) {
      return true;
    }

    for (unsigned i = 0; i < Inst.getNumOperands(); ++i)
      if (Inst.getOperand(i)->getType()->isScalableTy())
        return true;

    if (const AllocaInst *AI = dyn_cast<AllocaInst>(&Inst)) {
      if (AI->getAllocatedType()->isScalableTy())
        return true;
    }
  }

  // Checks to allow the use of SME instructions
  if (auto *Base = dyn_cast<CallBase>(&Inst)) {
    auto CallAttrs = SMECallAttrs(*Base);
    if (CallAttrs.requiresSMChange() || CallAttrs.requiresLazySave() ||
        CallAttrs.requiresPreservingZT0() ||
        CallAttrs.requiresPreservingAllZAState())
      return true;
  }
  return false;
}

// Return the largest legal scalable vector type that matches VT's element type.
static EVT getContainerForFixedLengthVector(SelectionDAG &DAG, EVT VT) {
  assert(VT.isFixedLengthVector() &&
         DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
         "Expected legal fixed length vector!");
  switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("unexpected element type for SVE container");
  case MVT::i8:
    return EVT(MVT::nxv16i8);
  case MVT::i16:
    return EVT(MVT::nxv8i16);
  case MVT::i32:
    return EVT(MVT::nxv4i32);
  case MVT::i64:
    return EVT(MVT::nxv2i64);
  case MVT::bf16:
    return EVT(MVT::nxv8bf16);
  case MVT::f16:
    return EVT(MVT::nxv8f16);
  case MVT::f32:
    return EVT(MVT::nxv4f32);
  case MVT::f64:
    return EVT(MVT::nxv2f64);
  }
}

// Return a predicate with active lanes corresponding to the extent of VT.
static SDValue getPredicateForFixedLengthVector(SelectionDAG &DAG, SDLoc &DL,
                                                EVT VT) {
  assert(VT.isFixedLengthVector() &&
         DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
         "Expected legal fixed length vector!");

  std::optional<unsigned> PgPattern =
      getSVEPredPatternFromNumElements(VT.getVectorNumElements());
  assert(PgPattern && "Unexpected element count for SVE predicate");

  // For vectors that are exactly getMaxSVEVectorSizeInBits big, we can use
  // AArch64SVEPredPattern::all, which can enable the use of unpredicated
  // variants of instructions when available.
  const auto &Subtarget = DAG.getSubtarget<AArch64Subtarget>();
  unsigned MinSVESize = Subtarget.getMinSVEVectorSizeInBits();
  unsigned MaxSVESize = Subtarget.getMaxSVEVectorSizeInBits();
  if (MaxSVESize && MinSVESize == MaxSVESize &&
      MaxSVESize == VT.getSizeInBits())
    PgPattern = AArch64SVEPredPattern::all;

  MVT MaskVT;
  switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("unexpected element type for SVE predicate");
  case MVT::i8:
    MaskVT = MVT::nxv16i1;
    break;
  case MVT::i16:
  case MVT::f16:
  case MVT::bf16:
    MaskVT = MVT::nxv8i1;
    break;
  case MVT::i32:
  case MVT::f32:
    MaskVT = MVT::nxv4i1;
    break;
  case MVT::i64:
  case MVT::f64:
    MaskVT = MVT::nxv2i1;
    break;
  }

  return getPTrue(DAG, DL, MaskVT, *PgPattern);
}

static SDValue getPredicateForScalableVector(SelectionDAG &DAG, SDLoc &DL,
                                             EVT VT) {
  assert(VT.isScalableVector() && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
         "Expected legal scalable vector!");
  auto PredTy = VT.changeVectorElementType(MVT::i1);
  return getPTrue(DAG, DL, PredTy, AArch64SVEPredPattern::all);
}

static SDValue getPredicateForVector(SelectionDAG &DAG, SDLoc &DL, EVT VT) {
  if (VT.isFixedLengthVector())
    return getPredicateForFixedLengthVector(DAG, DL, VT);

  return getPredicateForScalableVector(DAG, DL, VT);
}

// Grow V to consume an entire SVE register.
static SDValue convertToScalableVector(SelectionDAG &DAG, EVT VT, SDValue V) {
  assert(VT.isScalableVector() &&
         "Expected to convert into a scalable vector!");
  assert(V.getValueType().isFixedLengthVector() &&
         "Expected a fixed length vector operand!");
  SDLoc DL(V);
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V, Zero);
}

// Shrink V so it's just big enough to maintain a VT's worth of data.
static SDValue convertFromScalableVector(SelectionDAG &DAG, EVT VT, SDValue V) {
  assert(VT.isFixedLengthVector() &&
         "Expected to convert into a fixed length vector!");
  assert(V.getValueType().isScalableVector() &&
         "Expected a scalable vector operand!");
  SDLoc DL(V);
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V, Zero);
}

// Convert all fixed length vector loads larger than NEON to masked_loads.
SDValue AArch64TargetLowering::LowerFixedLengthVectorLoadToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  auto Load = cast<LoadSDNode>(Op);

  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
  EVT LoadVT = ContainerVT;
  EVT MemVT = Load->getMemoryVT();

  auto Pg = getPredicateForFixedLengthVector(DAG, DL, VT);

  if (VT.isFloatingPoint()) {
    LoadVT = ContainerVT.changeTypeToInteger();
    MemVT = MemVT.changeTypeToInteger();
  }

  SDValue NewLoad = DAG.getMaskedLoad(
      LoadVT, DL, Load->getChain(), Load->getBasePtr(), Load->getOffset(), Pg,
      DAG.getUNDEF(LoadVT), MemVT, Load->getMemOperand(),
      Load->getAddressingMode(), Load->getExtensionType());

  SDValue Result = NewLoad;
  if (VT.isFloatingPoint() && Load->getExtensionType() == ISD::EXTLOAD) {
    EVT ExtendVT = ContainerVT.changeVectorElementType(
        Load->getMemoryVT().getVectorElementType());

    Result = getSVESafeBitCast(ExtendVT, Result, DAG);
    Result = DAG.getNode(AArch64ISD::FP_EXTEND_MERGE_PASSTHRU, DL, ContainerVT,
                         Pg, Result, DAG.getUNDEF(ContainerVT));
  } else if (VT.isFloatingPoint()) {
    Result = DAG.getNode(ISD::BITCAST, DL, ContainerVT, Result);
  }

  Result = convertFromScalableVector(DAG, VT, Result);
  SDValue MergedValues[2] = {Result, NewLoad.getValue(1)};
  return DAG.getMergeValues(MergedValues, DL);
}

static SDValue convertFixedMaskToScalableVector(SDValue Mask,
                                                SelectionDAG &DAG) {
  SDLoc DL(Mask);
  EVT InVT = Mask.getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT);
  SDValue Pg = getPredicateForFixedLengthVector(DAG, DL, InVT);

  if (ISD::isBuildVectorAllOnes(Mask.getNode()))
    return Pg;

  bool InvertCond = false;
  if (isBitwiseNot(Mask)) {
    InvertCond = true;
    Mask = Mask.getOperand(0);
  }

  SDValue Op1, Op2;
  ISD::CondCode CC;

  // When Mask is the result of a SETCC, it's better to regenerate the compare.
  if (Mask.getOpcode() == ISD::SETCC) {
    Op1 = convertToScalableVector(DAG, ContainerVT, Mask.getOperand(0));
    Op2 = convertToScalableVector(DAG, ContainerVT, Mask.getOperand(1));
    CC = cast<CondCodeSDNode>(Mask.getOperand(2))->get();
  } else {
    Op1 = convertToScalableVector(DAG, ContainerVT, Mask);
    Op2 = DAG.getConstant(0, DL, ContainerVT);
    CC = ISD::SETNE;
  }

  if (InvertCond)
    CC = getSetCCInverse(CC, Op1.getValueType());

  return DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, DL, Pg.getValueType(),
                     {Pg, Op1, Op2, DAG.getCondCode(CC)});
}

// Convert all fixed length vector loads larger than NEON to masked_loads.
SDValue AArch64TargetLowering::LowerFixedLengthVectorMLoadToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  auto Load = cast<MaskedLoadSDNode>(Op);

  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);

  SDValue Mask = Load->getMask();
  // If this is an extending load and the mask type is not the same as
  // load's type then we have to extend the mask type.
  if (VT.getScalarSizeInBits() > Mask.getValueType().getScalarSizeInBits()) {
    assert(Load->getExtensionType() != ISD::NON_EXTLOAD &&
           "Incorrect mask type");
    Mask = DAG.getNode(ISD::SIGN_EXTEND, DL, VT, Mask);
  }
  Mask = convertFixedMaskToScalableVector(Mask, DAG);

  SDValue PassThru;
  bool IsPassThruZeroOrUndef = false;

  if (Load->getPassThru()->isUndef()) {
    PassThru = DAG.getUNDEF(ContainerVT);
    IsPassThruZeroOrUndef = true;
  } else {
    if (ContainerVT.isInteger())
      PassThru = DAG.getConstant(0, DL, ContainerVT);
    else
      PassThru = DAG.getConstantFP(0, DL, ContainerVT);
    if (isZerosVector(Load->getPassThru().getNode()))
      IsPassThruZeroOrUndef = true;
  }

  SDValue NewLoad = DAG.getMaskedLoad(
      ContainerVT, DL, Load->getChain(), Load->getBasePtr(), Load->getOffset(),
      Mask, PassThru, Load->getMemoryVT(), Load->getMemOperand(),
      Load->getAddressingMode(), Load->getExtensionType());

  SDValue Result = NewLoad;
  if (!IsPassThruZeroOrUndef) {
    SDValue OldPassThru =
        convertToScalableVector(DAG, ContainerVT, Load->getPassThru());
    Result = DAG.getSelect(DL, ContainerVT, Mask, Result, OldPassThru);
  }

  Result = convertFromScalableVector(DAG, VT, Result);
  SDValue MergedValues[2] = {Result, NewLoad.getValue(1)};
  return DAG.getMergeValues(MergedValues, DL);
}

// Convert all fixed length vector stores larger than NEON to masked_stores.
SDValue AArch64TargetLowering::LowerFixedLengthVectorStoreToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  auto Store = cast<StoreSDNode>(Op);

  SDLoc DL(Op);
  EVT VT = Store->getValue().getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
  EVT MemVT = Store->getMemoryVT();

  auto Pg = getPredicateForFixedLengthVector(DAG, DL, VT);
  auto NewValue = convertToScalableVector(DAG, ContainerVT, Store->getValue());

  if (VT.isFloatingPoint() && Store->isTruncatingStore()) {
    EVT TruncVT = ContainerVT.changeVectorElementType(
        Store->getMemoryVT().getVectorElementType());
    MemVT = MemVT.changeTypeToInteger();
    NewValue = DAG.getNode(AArch64ISD::FP_ROUND_MERGE_PASSTHRU, DL, TruncVT, Pg,
                           NewValue, DAG.getTargetConstant(0, DL, MVT::i64),
                           DAG.getUNDEF(TruncVT));
    NewValue =
        getSVESafeBitCast(ContainerVT.changeTypeToInteger(), NewValue, DAG);
  } else if (VT.isFloatingPoint()) {
    MemVT = MemVT.changeTypeToInteger();
    NewValue =
        getSVESafeBitCast(ContainerVT.changeTypeToInteger(), NewValue, DAG);
  }

  return DAG.getMaskedStore(Store->getChain(), DL, NewValue,
                            Store->getBasePtr(), Store->getOffset(), Pg, MemVT,
                            Store->getMemOperand(), Store->getAddressingMode(),
                            Store->isTruncatingStore());
}

SDValue AArch64TargetLowering::LowerFixedLengthVectorMStoreToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  auto *Store = cast<MaskedStoreSDNode>(Op);

  SDLoc DL(Op);
  EVT VT = Store->getValue().getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);

  auto NewValue = convertToScalableVector(DAG, ContainerVT, Store->getValue());
  SDValue Mask = convertFixedMaskToScalableVector(Store->getMask(), DAG);

  return DAG.getMaskedStore(
      Store->getChain(), DL, NewValue, Store->getBasePtr(), Store->getOffset(),
      Mask, Store->getMemoryVT(), Store->getMemOperand(),
      Store->getAddressingMode(), Store->isTruncatingStore());
}

SDValue AArch64TargetLowering::LowerFixedLengthVectorIntDivideToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  EVT EltVT = VT.getVectorElementType();

  bool Signed = Op.getOpcode() == ISD::SDIV;
  unsigned PredOpcode = Signed ? AArch64ISD::SDIV_PRED : AArch64ISD::UDIV_PRED;

  bool Negated;
  uint64_t SplatVal;
  if (Signed && isPow2Splat(Op.getOperand(1), SplatVal, Negated)) {
    EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
    SDValue Op1 = convertToScalableVector(DAG, ContainerVT, Op.getOperand(0));
    SDValue Op2 = DAG.getTargetConstant(Log2_64(SplatVal), DL, MVT::i32);

    SDValue Pg = getPredicateForFixedLengthVector(DAG, DL, VT);
    SDValue Res =
        DAG.getNode(AArch64ISD::SRAD_MERGE_OP1, DL, ContainerVT, Pg, Op1, Op2);
    if (Negated)
      Res = DAG.getNode(ISD::SUB, DL, ContainerVT,
                        DAG.getConstant(0, DL, ContainerVT), Res);

    return convertFromScalableVector(DAG, VT, Res);
  }

  // Scalable vector i32/i64 DIV is supported.
  if (EltVT == MVT::i32 || EltVT == MVT::i64)
    return LowerToPredicatedOp(Op, DAG, PredOpcode);

  // Scalable vector i8/i16 DIV is not supported. Promote it to i32.
  EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext());
  EVT PromVT = HalfVT.widenIntegerVectorElementType(*DAG.getContext());
  unsigned ExtendOpcode = Signed ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;

  // If the wider type is legal: extend, op, and truncate.
  EVT WideVT = VT.widenIntegerVectorElementType(*DAG.getContext());
  if (DAG.getTargetLoweringInfo().isTypeLegal(WideVT)) {
    SDValue Op0 = DAG.getNode(ExtendOpcode, DL, WideVT, Op.getOperand(0));
    SDValue Op1 = DAG.getNode(ExtendOpcode, DL, WideVT, Op.getOperand(1));
    SDValue Div = DAG.getNode(Op.getOpcode(), DL, WideVT, Op0, Op1);
    return DAG.getNode(ISD::TRUNCATE, DL, VT, Div);
  }

  auto HalveAndExtendVector = [&DAG, &DL, &HalfVT, &PromVT,
                               &ExtendOpcode](SDValue Op) {
    SDValue IdxZero = DAG.getConstant(0, DL, MVT::i64);
    SDValue IdxHalf =
        DAG.getConstant(HalfVT.getVectorNumElements(), DL, MVT::i64);
    SDValue Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, Op, IdxZero);
    SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, Op, IdxHalf);
    return std::pair<SDValue, SDValue>(
        {DAG.getNode(ExtendOpcode, DL, PromVT, Lo),
         DAG.getNode(ExtendOpcode, DL, PromVT, Hi)});
  };

  // If wider type is not legal: split, extend, op, trunc and concat.
  auto [Op0LoExt, Op0HiExt] = HalveAndExtendVector(Op.getOperand(0));
  auto [Op1LoExt, Op1HiExt] = HalveAndExtendVector(Op.getOperand(1));
  SDValue Lo = DAG.getNode(Op.getOpcode(), DL, PromVT, Op0LoExt, Op1LoExt);
  SDValue Hi = DAG.getNode(Op.getOpcode(), DL, PromVT, Op0HiExt, Op1HiExt);
  SDValue LoTrunc = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, Lo);
  SDValue HiTrunc = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, Hi);
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, {LoTrunc, HiTrunc});
}

SDValue AArch64TargetLowering::LowerFixedLengthVectorIntExtendToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  SDLoc DL(Op);
  SDValue Val = Op.getOperand(0);
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, Val.getValueType());
  Val = convertToScalableVector(DAG, ContainerVT, Val);

  bool Signed = Op.getOpcode() == ISD::SIGN_EXTEND;
  unsigned ExtendOpc = Signed ? AArch64ISD::SUNPKLO : AArch64ISD::UUNPKLO;

  // Repeatedly unpack Val until the result is of the desired element type.
  switch (ContainerVT.getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("unimplemented container type");
  case MVT::nxv16i8:
    Val = DAG.getNode(ExtendOpc, DL, MVT::nxv8i16, Val);
    if (VT.getVectorElementType() == MVT::i16)
      break;
    [[fallthrough]];
  case MVT::nxv8i16:
    Val = DAG.getNode(ExtendOpc, DL, MVT::nxv4i32, Val);
    if (VT.getVectorElementType() == MVT::i32)
      break;
    [[fallthrough]];
  case MVT::nxv4i32:
    Val = DAG.getNode(ExtendOpc, DL, MVT::nxv2i64, Val);
    assert(VT.getVectorElementType() == MVT::i64 && "Unexpected element type!");
    break;
  }

  return convertFromScalableVector(DAG, VT, Val);
}

SDValue AArch64TargetLowering::LowerFixedLengthVectorTruncateToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  SDLoc DL(Op);
  SDValue Val = Op.getOperand(0);
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, Val.getValueType());
  Val = convertToScalableVector(DAG, ContainerVT, Val);

  // Repeatedly truncate Val until the result is of the desired element type.
  switch (ContainerVT.getSimpleVT().SimpleTy) {
  default:
    llvm_unreachable("unimplemented container type");
  case MVT::nxv2i64:
    Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv4i32, Val);
    Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv4i32, Val, Val);
    if (VT.getVectorElementType() == MVT::i32)
      break;
    [[fallthrough]];
  case MVT::nxv4i32:
    Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv8i16, Val);
    Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv8i16, Val, Val);
    if (VT.getVectorElementType() == MVT::i16)
      break;
    [[fallthrough]];
  case MVT::nxv8i16:
    Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv16i8, Val);
    Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv16i8, Val, Val);
    assert(VT.getVectorElementType() == MVT::i8 && "Unexpected element type!");
    break;
  }

  return convertFromScalableVector(DAG, VT, Val);
}

SDValue AArch64TargetLowering::LowerFixedLengthExtractVectorElt(
    SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  EVT InVT = Op.getOperand(0).getValueType();
  assert(InVT.isFixedLengthVector() && "Expected fixed length vector type!");

  SDLoc DL(Op);
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT);
  SDValue Op0 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(0));

  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, Op0, Op.getOperand(1));
}

SDValue AArch64TargetLowering::LowerFixedLengthInsertVectorElt(
    SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  SDLoc DL(Op);
  EVT InVT = Op.getOperand(0).getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT);
  SDValue Op0 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(0));

  auto ScalableRes = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ContainerVT, Op0,
                                 Op.getOperand(1), Op.getOperand(2));

  return convertFromScalableVector(DAG, VT, ScalableRes);
}

// Convert vector operation 'Op' to an equivalent predicated operation whereby
// the original operation's type is used to construct a suitable predicate.
// NOTE: The results for inactive lanes are undefined.
SDValue AArch64TargetLowering::LowerToPredicatedOp(SDValue Op,
                                                   SelectionDAG &DAG,
                                                   unsigned NewOp) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  auto Pg = getPredicateForVector(DAG, DL, VT);

  if (VT.isFixedLengthVector()) {
    assert(isTypeLegal(VT) && "Expected only legal fixed-width types");
    EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);

    // Create list of operands by converting existing ones to scalable types.
    SmallVector<SDValue, 4> Operands = {Pg};
    for (const SDValue &V : Op->op_values()) {
      if (isa<CondCodeSDNode>(V)) {
        Operands.push_back(V);
        continue;
      }

      if (const VTSDNode *VTNode = dyn_cast<VTSDNode>(V)) {
        EVT VTArg = VTNode->getVT().getVectorElementType();
        EVT NewVTArg = ContainerVT.changeVectorElementType(VTArg);
        Operands.push_back(DAG.getValueType(NewVTArg));
        continue;
      }

      assert(isTypeLegal(V.getValueType()) &&
             "Expected only legal fixed-width types");
      Operands.push_back(convertToScalableVector(DAG, ContainerVT, V));
    }

    if (isMergePassthruOpcode(NewOp))
      Operands.push_back(DAG.getUNDEF(ContainerVT));

    auto ScalableRes = DAG.getNode(NewOp, DL, ContainerVT, Operands);
    return convertFromScalableVector(DAG, VT, ScalableRes);
  }

  assert(VT.isScalableVector() && "Only expect to lower scalable vector op!");

  SmallVector<SDValue, 4> Operands = {Pg};
  for (const SDValue &V : Op->op_values()) {
    assert((!V.getValueType().isVector() ||
            V.getValueType().isScalableVector()) &&
           "Only scalable vectors are supported!");
    Operands.push_back(V);
  }

  if (isMergePassthruOpcode(NewOp))
    Operands.push_back(DAG.getUNDEF(VT));

  return DAG.getNode(NewOp, DL, VT, Operands, Op->getFlags());
}

// If a fixed length vector operation has no side effects when applied to
// undefined elements, we can safely use scalable vectors to perform the same
// operation without needing to worry about predication.
SDValue AArch64TargetLowering::LowerToScalableOp(SDValue Op,
                                                 SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && isTypeLegal(VT) &&
         "Only expected to lower fixed length vector operation!");
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);

  // Create list of operands by converting existing ones to scalable types.
  SmallVector<SDValue, 4> Ops;
  for (const SDValue &V : Op->op_values()) {
    assert(!isa<VTSDNode>(V) && "Unexpected VTSDNode node!");

    // Pass through non-vector operands.
    if (!V.getValueType().isVector()) {
      Ops.push_back(V);
      continue;
    }

    // "cast" fixed length vector to a scalable vector.
    assert(V.getValueType().isFixedLengthVector() &&
           isTypeLegal(V.getValueType()) &&
           "Only fixed length vectors are supported!");
    Ops.push_back(convertToScalableVector(DAG, ContainerVT, V));
  }

  auto ScalableRes = DAG.getNode(Op.getOpcode(), SDLoc(Op), ContainerVT, Ops);
  return convertFromScalableVector(DAG, VT, ScalableRes);
}

SDValue AArch64TargetLowering::LowerVECREDUCE_SEQ_FADD(SDValue ScalarOp,
    SelectionDAG &DAG) const {
  SDLoc DL(ScalarOp);
  SDValue AccOp = ScalarOp.getOperand(0);
  SDValue VecOp = ScalarOp.getOperand(1);
  EVT SrcVT = VecOp.getValueType();
  EVT ResVT = SrcVT.getVectorElementType();

  EVT ContainerVT = SrcVT;
  if (SrcVT.isFixedLengthVector()) {
    ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT);
    VecOp = convertToScalableVector(DAG, ContainerVT, VecOp);
  }

  SDValue Pg = getPredicateForVector(DAG, DL, SrcVT);
  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);

  // Convert operands to Scalable.
  AccOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ContainerVT,
                      DAG.getUNDEF(ContainerVT), AccOp, Zero);

  // Perform reduction.
  SDValue Rdx = DAG.getNode(AArch64ISD::FADDA_PRED, DL, ContainerVT,
                            Pg, AccOp, VecOp);

  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, Rdx, Zero);
}

SDValue AArch64TargetLowering::LowerPredReductionToSVE(SDValue ReduceOp,
                                                       SelectionDAG &DAG) const {
  SDLoc DL(ReduceOp);
  SDValue Op = ReduceOp.getOperand(0);
  EVT OpVT = Op.getValueType();
  EVT VT = ReduceOp.getValueType();

  if (!OpVT.isScalableVector() || OpVT.getVectorElementType() != MVT::i1)
    return SDValue();

  SDValue Pg = getPredicateForVector(DAG, DL, OpVT);

  switch (ReduceOp.getOpcode()) {
  default:
    return SDValue();
  case ISD::VECREDUCE_OR:
    if (isAllActivePredicate(DAG, Pg) && OpVT == MVT::nxv16i1)
      // The predicate can be 'Op' because
      // vecreduce_or(Op & <all true>) <=> vecreduce_or(Op).
      return getPTest(DAG, VT, Op, Op, AArch64CC::ANY_ACTIVE);
    else
      return getPTest(DAG, VT, Pg, Op, AArch64CC::ANY_ACTIVE);
  case ISD::VECREDUCE_AND: {
    Op = DAG.getNode(ISD::XOR, DL, OpVT, Op, Pg);
    return getPTest(DAG, VT, Pg, Op, AArch64CC::NONE_ACTIVE);
  }
  case ISD::VECREDUCE_XOR: {
    SDValue ID =
        DAG.getTargetConstant(Intrinsic::aarch64_sve_cntp, DL, MVT::i64);
    if (OpVT == MVT::nxv1i1) {
      // Emulate a CNTP on .Q using .D and a different governing predicate.
      Pg = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, MVT::nxv2i1, Pg);
      Op = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, MVT::nxv2i1, Op);
    }
    SDValue Cntp =
        DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::i64, ID, Pg, Op);
    return DAG.getAnyExtOrTrunc(Cntp, DL, VT);
  }
  }

  return SDValue();
}

SDValue AArch64TargetLowering::LowerReductionToSVE(unsigned Opcode,
                                                   SDValue ScalarOp,
                                                   SelectionDAG &DAG) const {
  SDLoc DL(ScalarOp);
  SDValue VecOp = ScalarOp.getOperand(0);
  EVT SrcVT = VecOp.getValueType();

  if (useSVEForFixedLengthVectorVT(
          SrcVT,
          /*OverrideNEON=*/Subtarget->useSVEForFixedLengthVectors())) {
    EVT ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT);
    VecOp = convertToScalableVector(DAG, ContainerVT, VecOp);
  }

  // Lower VECREDUCE_ADD of nxv2i1-nxv16i1 to CNTP rather than UADDV.
  if (ScalarOp.getOpcode() == ISD::VECREDUCE_ADD &&
      VecOp.getOpcode() == ISD::ZERO_EXTEND) {
    SDValue BoolVec = VecOp.getOperand(0);
    if (BoolVec.getValueType().getVectorElementType() == MVT::i1) {
      // CNTP(BoolVec & BoolVec) <=> CNTP(BoolVec & PTRUE)
      SDValue CntpOp = DAG.getNode(
          ISD::INTRINSIC_WO_CHAIN, DL, MVT::i64,
          DAG.getTargetConstant(Intrinsic::aarch64_sve_cntp, DL, MVT::i64),
          BoolVec, BoolVec);
      return DAG.getAnyExtOrTrunc(CntpOp, DL, ScalarOp.getValueType());
    }
  }

  // UADDV always returns an i64 result.
  EVT ResVT = (Opcode == AArch64ISD::UADDV_PRED) ? MVT::i64 :
                                                   SrcVT.getVectorElementType();
  EVT RdxVT = SrcVT;
  if (SrcVT.isFixedLengthVector() || Opcode == AArch64ISD::UADDV_PRED)
    RdxVT = getPackedSVEVectorVT(ResVT);

  SDValue Pg = getPredicateForVector(DAG, DL, SrcVT);
  SDValue Rdx = DAG.getNode(Opcode, DL, RdxVT, Pg, VecOp);
  SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT,
                            Rdx, DAG.getConstant(0, DL, MVT::i64));

  // The VEC_REDUCE nodes expect an element size result.
  if (ResVT != ScalarOp.getValueType())
    Res = DAG.getAnyExtOrTrunc(Res, DL, ScalarOp.getValueType());

  return Res;
}

SDValue
AArch64TargetLowering::LowerFixedLengthVectorSelectToSVE(SDValue Op,
    SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);

  EVT InVT = Op.getOperand(1).getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT);
  SDValue Op1 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(1));
  SDValue Op2 = convertToScalableVector(DAG, ContainerVT, Op->getOperand(2));

  // Convert the mask to a predicated (NOTE: We don't need to worry about
  // inactive lanes since VSELECT is safe when given undefined elements).
  EVT MaskVT = Op.getOperand(0).getValueType();
  EVT MaskContainerVT = getContainerForFixedLengthVector(DAG, MaskVT);
  auto Mask = convertToScalableVector(DAG, MaskContainerVT, Op.getOperand(0));
  Mask = DAG.getNode(ISD::TRUNCATE, DL,
                     MaskContainerVT.changeVectorElementType(MVT::i1), Mask);

  auto ScalableRes = DAG.getNode(ISD::VSELECT, DL, ContainerVT,
                                Mask, Op1, Op2);

  return convertFromScalableVector(DAG, VT, ScalableRes);
}

SDValue AArch64TargetLowering::LowerFixedLengthVectorSetccToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT InVT = Op.getOperand(0).getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, InVT);

  assert(InVT.isFixedLengthVector() && isTypeLegal(InVT) &&
         "Only expected to lower fixed length vector operation!");
  assert(Op.getValueType() == InVT.changeTypeToInteger() &&
         "Expected integer result of the same bit length as the inputs!");

  auto Op1 = convertToScalableVector(DAG, ContainerVT, Op.getOperand(0));
  auto Op2 = convertToScalableVector(DAG, ContainerVT, Op.getOperand(1));
  auto Pg = getPredicateForFixedLengthVector(DAG, DL, InVT);

  EVT CmpVT = Pg.getValueType();
  auto Cmp = DAG.getNode(AArch64ISD::SETCC_MERGE_ZERO, DL, CmpVT,
                         {Pg, Op1, Op2, Op.getOperand(2)});

  EVT PromoteVT = ContainerVT.changeTypeToInteger();
  auto Promote = DAG.getBoolExtOrTrunc(Cmp, DL, PromoteVT, InVT);
  return convertFromScalableVector(DAG, Op.getValueType(), Promote);
}

SDValue
AArch64TargetLowering::LowerFixedLengthBitcastToSVE(SDValue Op,
                                                    SelectionDAG &DAG) const {
  SDLoc DL(Op);
  auto SrcOp = Op.getOperand(0);
  EVT VT = Op.getValueType();
  EVT ContainerDstVT = getContainerForFixedLengthVector(DAG, VT);
  EVT ContainerSrcVT =
      getContainerForFixedLengthVector(DAG, SrcOp.getValueType());

  SrcOp = convertToScalableVector(DAG, ContainerSrcVT, SrcOp);
  Op = DAG.getNode(ISD::BITCAST, DL, ContainerDstVT, SrcOp);
  return convertFromScalableVector(DAG, VT, Op);
}

SDValue AArch64TargetLowering::LowerFixedLengthConcatVectorsToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  SDLoc DL(Op);
  unsigned NumOperands = Op->getNumOperands();

  assert(NumOperands > 1 && isPowerOf2_32(NumOperands) &&
         "Unexpected number of operands in CONCAT_VECTORS");

  auto SrcOp1 = Op.getOperand(0);
  auto SrcOp2 = Op.getOperand(1);
  EVT VT = Op.getValueType();
  EVT SrcVT = SrcOp1.getValueType();

  // Match a splat of 128b segments that fit in a single register.
  if (SrcVT.is128BitVector() && all_equal(Op.getNode()->op_values())) {
    EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
    SDValue Splat =
        DAG.getNode(AArch64ISD::DUPLANE128, DL, ContainerVT,
                    convertToScalableVector(DAG, ContainerVT, SrcOp1),
                    DAG.getConstant(0, DL, MVT::i64, /*isTarget=*/true));
    return convertFromScalableVector(DAG, VT, Splat);
  }

  if (NumOperands > 2) {
    SmallVector<SDValue, 4> Ops;
    EVT PairVT = SrcVT.getDoubleNumVectorElementsVT(*DAG.getContext());
    for (unsigned I = 0; I < NumOperands; I += 2)
      Ops.push_back(DAG.getNode(ISD::CONCAT_VECTORS, DL, PairVT,
                                Op->getOperand(I), Op->getOperand(I + 1)));

    return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Ops);
  }

  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);

  SDValue Pg = getPredicateForFixedLengthVector(DAG, DL, SrcVT);
  SrcOp1 = convertToScalableVector(DAG, ContainerVT, SrcOp1);
  SrcOp2 = convertToScalableVector(DAG, ContainerVT, SrcOp2);

  Op = DAG.getNode(AArch64ISD::SPLICE, DL, ContainerVT, Pg, SrcOp1, SrcOp2);

  return convertFromScalableVector(DAG, VT, Op);
}

SDValue
AArch64TargetLowering::LowerFixedLengthFPExtendToSVE(SDValue Op,
                                                     SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  SDLoc DL(Op);
  SDValue Val = Op.getOperand(0);
  SDValue Pg = getPredicateForVector(DAG, DL, VT);
  EVT SrcVT = Val.getValueType();
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
  EVT ExtendVT = ContainerVT.changeVectorElementType(
      SrcVT.getVectorElementType());

  Val = DAG.getNode(ISD::BITCAST, DL, SrcVT.changeTypeToInteger(), Val);
  Val = DAG.getNode(ISD::ANY_EXTEND, DL, VT.changeTypeToInteger(), Val);

  Val = convertToScalableVector(DAG, ContainerVT.changeTypeToInteger(), Val);
  Val = getSVESafeBitCast(ExtendVT, Val, DAG);
  Val = DAG.getNode(AArch64ISD::FP_EXTEND_MERGE_PASSTHRU, DL, ContainerVT,
                    Pg, Val, DAG.getUNDEF(ContainerVT));

  return convertFromScalableVector(DAG, VT, Val);
}

SDValue
AArch64TargetLowering::LowerFixedLengthFPRoundToSVE(SDValue Op,
                                                    SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  SDLoc DL(Op);
  SDValue Val = Op.getOperand(0);
  EVT SrcVT = Val.getValueType();
  EVT ContainerSrcVT = getContainerForFixedLengthVector(DAG, SrcVT);
  EVT RoundVT = ContainerSrcVT.changeVectorElementType(
      VT.getVectorElementType());
  SDValue Pg = getPredicateForVector(DAG, DL, RoundVT);

  Val = convertToScalableVector(DAG, ContainerSrcVT, Val);
  Val = DAG.getNode(AArch64ISD::FP_ROUND_MERGE_PASSTHRU, DL, RoundVT, Pg, Val,
                    Op.getOperand(1), DAG.getUNDEF(RoundVT));
  Val = getSVESafeBitCast(ContainerSrcVT.changeTypeToInteger(), Val, DAG);
  Val = convertFromScalableVector(DAG, SrcVT.changeTypeToInteger(), Val);

  Val = DAG.getNode(ISD::TRUNCATE, DL, VT.changeTypeToInteger(), Val);
  return DAG.getNode(ISD::BITCAST, DL, VT, Val);
}

SDValue
AArch64TargetLowering::LowerFixedLengthIntToFPToSVE(SDValue Op,
                                                    SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  bool IsSigned = Op.getOpcode() == ISD::SINT_TO_FP;
  unsigned Opcode = IsSigned ? AArch64ISD::SINT_TO_FP_MERGE_PASSTHRU
                             : AArch64ISD::UINT_TO_FP_MERGE_PASSTHRU;

  SDLoc DL(Op);
  SDValue Val = Op.getOperand(0);
  EVT SrcVT = Val.getValueType();
  EVT ContainerDstVT = getContainerForFixedLengthVector(DAG, VT);
  EVT ContainerSrcVT = getContainerForFixedLengthVector(DAG, SrcVT);

  if (VT.bitsGE(SrcVT)) {
    SDValue Pg = getPredicateForFixedLengthVector(DAG, DL, VT);

    Val = DAG.getNode(IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND, DL,
                      VT.changeTypeToInteger(), Val);

    // Safe to use a larger than specified operand because by promoting the
    // value nothing has changed from an arithmetic point of view.
    Val =
        convertToScalableVector(DAG, ContainerDstVT.changeTypeToInteger(), Val);
    Val = DAG.getNode(Opcode, DL, ContainerDstVT, Pg, Val,
                      DAG.getUNDEF(ContainerDstVT));
    return convertFromScalableVector(DAG, VT, Val);
  } else {
    EVT CvtVT = ContainerSrcVT.changeVectorElementType(
        ContainerDstVT.getVectorElementType());
    SDValue Pg = getPredicateForFixedLengthVector(DAG, DL, SrcVT);

    Val = convertToScalableVector(DAG, ContainerSrcVT, Val);
    Val = DAG.getNode(Opcode, DL, CvtVT, Pg, Val, DAG.getUNDEF(CvtVT));
    Val = getSVESafeBitCast(ContainerSrcVT, Val, DAG);
    Val = convertFromScalableVector(DAG, SrcVT, Val);

    Val = DAG.getNode(ISD::TRUNCATE, DL, VT.changeTypeToInteger(), Val);
    return DAG.getNode(ISD::BITCAST, DL, VT, Val);
  }
}

SDValue
AArch64TargetLowering::LowerVECTOR_DEINTERLEAVE(SDValue Op,
                                                SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT OpVT = Op.getValueType();
  assert(OpVT.isScalableVector() &&
         "Expected scalable vector in LowerVECTOR_DEINTERLEAVE.");

  // Are multi-register uzp instructions available?
  if (Subtarget->hasSME2() && Subtarget->isStreaming() &&
      OpVT.getVectorElementType() != MVT::i1) {
    Intrinsic::ID IntID;
    switch (Op->getNumOperands()) {
    default:
      return SDValue();
    case 2:
      IntID = Intrinsic::aarch64_sve_uzp_x2;
      break;
    case 4:
      if (Subtarget->getMinSVEVectorSizeInBits() < 256 &&
          OpVT.getScalarSizeInBits() == 64)
        return SDValue();
      IntID = Intrinsic::aarch64_sve_uzp_x4;
      break;
    }

    SmallVector<SDValue, 5> Ops;
    Ops.push_back(DAG.getTargetConstant(IntID, DL, MVT::i64));
    Ops.append(Op->op_values().begin(), Op->op_values().end());
    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op->getVTList(), Ops);
  }

  if (Op->getNumOperands() != 2)
    return SDValue();

  SDValue Even = DAG.getNode(AArch64ISD::UZP1, DL, OpVT, Op.getOperand(0),
                             Op.getOperand(1));
  SDValue Odd = DAG.getNode(AArch64ISD::UZP2, DL, OpVT, Op.getOperand(0),
                            Op.getOperand(1));
  return DAG.getMergeValues({Even, Odd}, DL);
}

SDValue AArch64TargetLowering::LowerVECTOR_INTERLEAVE(SDValue Op,
                                                      SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT OpVT = Op.getValueType();
  assert(OpVT.isScalableVector() &&
         "Expected scalable vector in LowerVECTOR_INTERLEAVE.");

  // Are multi-register zip instructions available?
  if (Subtarget->hasSME2() && Subtarget->isStreaming() &&
      OpVT.getVectorElementType() != MVT::i1) {
    Intrinsic::ID IntID;
    switch (Op->getNumOperands()) {
    default:
      return SDValue();
    case 2:
      IntID = Intrinsic::aarch64_sve_zip_x2;
      break;
    case 4:
      if (Subtarget->getMinSVEVectorSizeInBits() < 256 &&
          OpVT.getScalarSizeInBits() == 64)
        return SDValue();
      IntID = Intrinsic::aarch64_sve_zip_x4;
      break;
    }

    SmallVector<SDValue, 5> Ops;
    Ops.push_back(DAG.getTargetConstant(IntID, DL, MVT::i64));
    Ops.append(Op->op_values().begin(), Op->op_values().end());
    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, Op->getVTList(), Ops);
  }

  if (Op->getNumOperands() != 2)
    return SDValue();

  SDValue Lo = DAG.getNode(AArch64ISD::ZIP1, DL, OpVT, Op.getOperand(0),
                           Op.getOperand(1));
  SDValue Hi = DAG.getNode(AArch64ISD::ZIP2, DL, OpVT, Op.getOperand(0),
                           Op.getOperand(1));
  return DAG.getMergeValues({Lo, Hi}, DL);
}

SDValue AArch64TargetLowering::LowerVECTOR_HISTOGRAM(SDValue Op,
                                                     SelectionDAG &DAG) const {
  // FIXME: Maybe share some code with LowerMGather/Scatter?
  MaskedHistogramSDNode *HG = cast<MaskedHistogramSDNode>(Op);
  SDLoc DL(HG);
  SDValue Chain = HG->getChain();
  SDValue Inc = HG->getInc();
  SDValue Mask = HG->getMask();
  SDValue Ptr = HG->getBasePtr();
  SDValue Index = HG->getIndex();
  SDValue Scale = HG->getScale();
  SDValue IntID = HG->getIntID();

  // The Intrinsic ID determines the type of update operation.
  [[maybe_unused]] ConstantSDNode *CID = cast<ConstantSDNode>(IntID.getNode());
  // Right now, we only support 'add' as an update.
  assert(CID->getZExtValue() == Intrinsic::experimental_vector_histogram_add &&
         "Unexpected histogram update operation");

  EVT IndexVT = Index.getValueType();
  LLVMContext &Ctx = *DAG.getContext();
  ElementCount EC = IndexVT.getVectorElementCount();
  EVT MemVT = EVT::getVectorVT(Ctx, HG->getMemoryVT(), EC);
  EVT IncExtVT =
      EVT::getIntegerVT(Ctx, AArch64::SVEBitsPerBlock / EC.getKnownMinValue());
  EVT IncSplatVT = EVT::getVectorVT(Ctx, IncExtVT, EC);
  bool ExtTrunc = IncSplatVT != MemVT;

  SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
  SDValue PassThru = DAG.getSplatVector(IncSplatVT, DL, Zero);
  SDValue IncSplat = DAG.getSplatVector(
      IncSplatVT, DL, DAG.getAnyExtOrTrunc(Inc, DL, IncExtVT));
  SDValue Ops[] = {Chain, PassThru, Mask, Ptr, Index, Scale};

  MachineMemOperand *MMO = HG->getMemOperand();
  // Create an MMO for the gather, without load|store flags.
  MachineMemOperand *GMMO = DAG.getMachineFunction().getMachineMemOperand(
      MMO->getPointerInfo(), MachineMemOperand::MOLoad, MMO->getSize(),
      MMO->getAlign(), MMO->getAAInfo());
  ISD::MemIndexType IndexType = HG->getIndexType();
  SDValue Gather = DAG.getMaskedGather(
      DAG.getVTList(IncSplatVT, MVT::Other), MemVT, DL, Ops, GMMO, IndexType,
      ExtTrunc ? ISD::EXTLOAD : ISD::NON_EXTLOAD);

  SDValue GChain = Gather.getValue(1);

  // Perform the histcnt, multiply by inc, add to bucket data.
  SDValue ID =
      DAG.getTargetConstant(Intrinsic::aarch64_sve_histcnt, DL, IncExtVT);
  SDValue HistCnt =
      DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, IndexVT, ID, Mask, Index, Index);
  SDValue Mul = DAG.getNode(ISD::MUL, DL, IncSplatVT, HistCnt, IncSplat);
  SDValue Add = DAG.getNode(ISD::ADD, DL, IncSplatVT, Gather, Mul);

  // Create an MMO for the scatter, without load|store flags.
  MachineMemOperand *SMMO = DAG.getMachineFunction().getMachineMemOperand(
      MMO->getPointerInfo(), MachineMemOperand::MOStore, MMO->getSize(),
      MMO->getAlign(), MMO->getAAInfo());

  SDValue ScatterOps[] = {GChain, Add, Mask, Ptr, Index, Scale};
  SDValue Scatter = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), MemVT, DL,
                                         ScatterOps, SMMO, IndexType, ExtTrunc);
  return Scatter;
}

/// If a PARTIAL_REDUCE_MLA node comes in with an accumulator-input type pairing
/// of (nx)v2i64/(nx)v16i8, we cannot directly lower it to a (u|s)dot. We can
/// however still make use of the dot product instruction by instead
/// accumulating over two steps: (nx)v16i8 -> (nx)v4i32 -> (nx)v2i64.
/// If available, make use of the (U|S)ADDW(B|T) instructions, otherwise
/// the following pattern is emitted:
/// add(add(Acc, ext(EXTRACT_SUBVECTOR(N, 0)), ext(EXTRACT_SUBVECTOR(N,
/// NTy/2))))
SDValue
AArch64TargetLowering::LowerPARTIAL_REDUCE_MLA(SDValue Op,
                                               SelectionDAG &DAG) const {
  SDLoc DL(Op);

  SDValue Acc = Op.getOperand(0);
  SDValue LHS = Op.getOperand(1);
  SDValue RHS = Op.getOperand(2);
  EVT ResultVT = Op.getValueType();
  EVT OrigResultVT = ResultVT;
  EVT OpVT = LHS.getValueType();

  bool ConvertToScalable =
      ResultVT.isFixedLengthVector() &&
      useSVEForFixedLengthVectorVT(ResultVT, /*OverrideNEON=*/true);

  if (ConvertToScalable) {
    ResultVT = getContainerForFixedLengthVector(DAG, ResultVT);
    OpVT = getContainerForFixedLengthVector(DAG, LHS.getValueType());
    Acc = convertToScalableVector(DAG, ResultVT, Acc);
    LHS = convertToScalableVector(DAG, OpVT, LHS);
    RHS = convertToScalableVector(DAG, OpVT, RHS);
    Op = DAG.getNode(Op.getOpcode(), DL, ResultVT, {Acc, LHS, RHS});
  }

  // Two-way and four-way partial reductions are supported by patterns.
  // We only need to handle the 8-way partial reduction.
  if (ResultVT.getScalarType() != MVT::i64 || OpVT.getScalarType() != MVT::i8)
    return ConvertToScalable ? convertFromScalableVector(DAG, OrigResultVT, Op)
                             : Op;

  EVT DotVT = ResultVT.isScalableVector() ? MVT::nxv4i32 : MVT::v4i32;
  SDValue DotNode = DAG.getNode(Op.getOpcode(), DL, DotVT,
                                DAG.getConstant(0, DL, DotVT), LHS, RHS);

  SDValue Res;
  bool IsUnsigned = Op.getOpcode() == ISD::PARTIAL_REDUCE_UMLA;
  if (Subtarget->hasSVE2() || Subtarget->isStreamingSVEAvailable()) {
    unsigned LoOpcode = IsUnsigned ? AArch64ISD::UADDWB : AArch64ISD::SADDWB;
    unsigned HiOpcode = IsUnsigned ? AArch64ISD::UADDWT : AArch64ISD::SADDWT;
    SDValue Lo = DAG.getNode(LoOpcode, DL, ResultVT, Acc, DotNode);
    Res = DAG.getNode(HiOpcode, DL, ResultVT, Lo, DotNode);
  } else {
    // Fold (nx)v4i32 into (nx)v2i64
    auto [DotNodeLo, DotNodeHi] = DAG.SplitVector(DotNode, DL);
    if (IsUnsigned) {
      DotNodeLo = DAG.getZExtOrTrunc(DotNodeLo, DL, ResultVT);
      DotNodeHi = DAG.getZExtOrTrunc(DotNodeHi, DL, ResultVT);
    } else {
      DotNodeLo = DAG.getSExtOrTrunc(DotNodeLo, DL, ResultVT);
      DotNodeHi = DAG.getSExtOrTrunc(DotNodeHi, DL, ResultVT);
    }
    auto Lo = DAG.getNode(ISD::ADD, DL, ResultVT, Acc, DotNodeLo);
    Res = DAG.getNode(ISD::ADD, DL, ResultVT, Lo, DotNodeHi);
  }

  return ConvertToScalable ? convertFromScalableVector(DAG, OrigResultVT, Res)
                           : Res;
}

SDValue
AArch64TargetLowering::LowerGET_ACTIVE_LANE_MASK(SDValue Op,
                                                 SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  assert(Subtarget->isSVEorStreamingSVEAvailable() &&
         "Lowering fixed length get_active_lane_mask requires SVE!");

  // There are no dedicated fixed-length instructions for GET_ACTIVE_LANE_MASK,
  // but we can use SVE when available.

  SDLoc DL(Op);
  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
  EVT WhileVT = ContainerVT.changeElementType(MVT::i1);

  SDValue Mask = DAG.getNode(ISD::GET_ACTIVE_LANE_MASK, DL, WhileVT,
                             Op.getOperand(0), Op.getOperand(1));
  SDValue MaskAsInt = DAG.getNode(ISD::SIGN_EXTEND, DL, ContainerVT, Mask);
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, MaskAsInt,
                     DAG.getVectorIdxConstant(0, DL));
}

SDValue
AArch64TargetLowering::LowerFixedLengthFPToIntToSVE(SDValue Op,
                                                    SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  bool IsSigned = Op.getOpcode() == ISD::FP_TO_SINT;
  unsigned Opcode = IsSigned ? AArch64ISD::FCVTZS_MERGE_PASSTHRU
                             : AArch64ISD::FCVTZU_MERGE_PASSTHRU;

  SDLoc DL(Op);
  SDValue Val = Op.getOperand(0);
  EVT SrcVT = Val.getValueType();
  EVT ContainerDstVT = getContainerForFixedLengthVector(DAG, VT);
  EVT ContainerSrcVT = getContainerForFixedLengthVector(DAG, SrcVT);

  if (VT.bitsGT(SrcVT)) {
    EVT CvtVT = ContainerDstVT.changeVectorElementType(
      ContainerSrcVT.getVectorElementType());
    SDValue Pg = getPredicateForFixedLengthVector(DAG, DL, VT);

    Val = DAG.getNode(ISD::BITCAST, DL, SrcVT.changeTypeToInteger(), Val);
    Val = DAG.getNode(ISD::ANY_EXTEND, DL, VT, Val);

    Val = convertToScalableVector(DAG, ContainerDstVT, Val);
    Val = getSVESafeBitCast(CvtVT, Val, DAG);
    Val = DAG.getNode(Opcode, DL, ContainerDstVT, Pg, Val,
                      DAG.getUNDEF(ContainerDstVT));
    return convertFromScalableVector(DAG, VT, Val);
  } else {
    EVT CvtVT = ContainerSrcVT.changeTypeToInteger();
    SDValue Pg = getPredicateForFixedLengthVector(DAG, DL, SrcVT);

    // Safe to use a larger than specified result since an fp_to_int where the
    // result doesn't fit into the destination is undefined.
    Val = convertToScalableVector(DAG, ContainerSrcVT, Val);
    Val = DAG.getNode(Opcode, DL, CvtVT, Pg, Val, DAG.getUNDEF(CvtVT));
    Val = convertFromScalableVector(DAG, SrcVT.changeTypeToInteger(), Val);

    return DAG.getNode(ISD::TRUNCATE, DL, VT, Val);
  }
}

static SDValue GenerateFixedLengthSVETBL(SDValue Op, SDValue Op1, SDValue Op2,
                                         ArrayRef<int> ShuffleMask, EVT VT,
                                         EVT ContainerVT, SelectionDAG &DAG) {
  auto &Subtarget = DAG.getSubtarget<AArch64Subtarget>();
  SDLoc DL(Op);
  unsigned MinSVESize = Subtarget.getMinSVEVectorSizeInBits();
  unsigned MaxSVESize = Subtarget.getMaxSVEVectorSizeInBits();
  bool IsSingleOp =
      ShuffleVectorInst::isSingleSourceMask(ShuffleMask, ShuffleMask.size());

  if (!Subtarget.isNeonAvailable() && !MinSVESize)
    MinSVESize = 128;

  // Ignore two operands if no SVE2 or all index numbers couldn't
  // be represented.
  if (!IsSingleOp && !Subtarget.hasSVE2())
    return SDValue();

  EVT VTOp1 = Op.getOperand(0).getValueType();
  unsigned BitsPerElt = VTOp1.getVectorElementType().getSizeInBits();
  unsigned IndexLen = MinSVESize / BitsPerElt;
  unsigned ElementsPerVectorReg = VTOp1.getVectorNumElements();
  uint64_t MaxOffset = maxUIntN(BitsPerElt);
  EVT MaskEltType = VTOp1.getVectorElementType().changeTypeToInteger();
  EVT MaskType = EVT::getVectorVT(*DAG.getContext(), MaskEltType, IndexLen);
  bool MinMaxEqual = (MinSVESize == MaxSVESize);
  assert(ElementsPerVectorReg <= IndexLen && ShuffleMask.size() <= IndexLen &&
         "Incorrectly legalised shuffle operation");

  SmallVector<SDValue, 8> TBLMask;
  // If MinSVESize is not equal to MaxSVESize then we need to know which
  // TBL mask element needs adjustment.
  SmallVector<SDValue, 8> AddRuntimeVLMask;

  // Bail out for 8-bits element types, because with 2048-bit SVE register
  // size 8 bits is only sufficient to index into the first source vector.
  if (!IsSingleOp && !MinMaxEqual && BitsPerElt == 8)
    return SDValue();

  for (int Index : ShuffleMask) {
    // Handling poison index value.
    if (Index < 0)
      Index = 0;
    // If the mask refers to elements in the second operand, then we have to
    // offset the index by the number of elements in a vector. If this is number
    // is not known at compile-time, we need to maintain a mask with 'VL' values
    // to add at runtime.
    if ((unsigned)Index >= ElementsPerVectorReg) {
      if (MinMaxEqual) {
        Index += IndexLen - ElementsPerVectorReg;
      } else {
        Index = Index - ElementsPerVectorReg;
        AddRuntimeVLMask.push_back(DAG.getConstant(1, DL, MVT::i64));
      }
    } else if (!MinMaxEqual)
      AddRuntimeVLMask.push_back(DAG.getConstant(0, DL, MVT::i64));
    // For 8-bit elements and 1024-bit SVE registers and MaxOffset equals
    // to 255, this might point to the last element of in the second operand
    // of the shufflevector, thus we are rejecting this transform.
    if ((unsigned)Index >= MaxOffset)
      return SDValue();
    TBLMask.push_back(DAG.getConstant(Index, DL, MVT::i64));
  }

  // Choosing an out-of-range index leads to the lane being zeroed vs zero
  // value where it would perform first lane duplication for out of
  // index elements. For i8 elements an out-of-range index could be a valid
  // for 2048-bit vector register size.
  for (unsigned i = 0; i < IndexLen - ElementsPerVectorReg; ++i) {
    TBLMask.push_back(DAG.getConstant((int)MaxOffset, DL, MVT::i64));
    if (!MinMaxEqual)
      AddRuntimeVLMask.push_back(DAG.getConstant(0, DL, MVT::i64));
  }

  EVT MaskContainerVT = getContainerForFixedLengthVector(DAG, MaskType);
  SDValue VecMask =
      DAG.getBuildVector(MaskType, DL, ArrayRef(TBLMask.data(), IndexLen));
  SDValue SVEMask = convertToScalableVector(DAG, MaskContainerVT, VecMask);

  SDValue Shuffle;
  if (IsSingleOp)
    Shuffle =
        DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ContainerVT,
                    DAG.getConstant(Intrinsic::aarch64_sve_tbl, DL, MVT::i32),
                    Op1, SVEMask);
  else if (Subtarget.hasSVE2()) {
    if (!MinMaxEqual) {
      unsigned MinNumElts = AArch64::SVEBitsPerBlock / BitsPerElt;
      SDValue VScale = (BitsPerElt == 64)
                           ? DAG.getVScale(DL, MVT::i64, APInt(64, MinNumElts))
                           : DAG.getVScale(DL, MVT::i32, APInt(32, MinNumElts));
      SDValue VecMask =
          DAG.getBuildVector(MaskType, DL, ArrayRef(TBLMask.data(), IndexLen));
      SDValue MulByMask = DAG.getNode(
          ISD::MUL, DL, MaskType,
          DAG.getNode(ISD::SPLAT_VECTOR, DL, MaskType, VScale),
          DAG.getBuildVector(MaskType, DL,
                             ArrayRef(AddRuntimeVLMask.data(), IndexLen)));
      SDValue UpdatedVecMask =
          DAG.getNode(ISD::ADD, DL, MaskType, VecMask, MulByMask);
      SVEMask = convertToScalableVector(
          DAG, getContainerForFixedLengthVector(DAG, MaskType), UpdatedVecMask);
    }
    Shuffle =
        DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ContainerVT,
                    DAG.getConstant(Intrinsic::aarch64_sve_tbl2, DL, MVT::i32),
                    Op1, Op2, SVEMask);
  }
  Shuffle = convertFromScalableVector(DAG, VT, Shuffle);
  return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
}

SDValue AArch64TargetLowering::LowerFixedLengthVECTOR_SHUFFLEToSVE(
    SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");

  auto *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
  auto ShuffleMask = SVN->getMask();

  SDLoc DL(Op);
  SDValue Op1 = Op.getOperand(0);
  SDValue Op2 = Op.getOperand(1);

  EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
  Op1 = convertToScalableVector(DAG, ContainerVT, Op1);
  Op2 = convertToScalableVector(DAG, ContainerVT, Op2);

  auto MinLegalExtractEltScalarTy = [](EVT ScalarTy) -> EVT {
    if (ScalarTy == MVT::i8 || ScalarTy == MVT::i16)
      return MVT::i32;
    return ScalarTy;
  };

  if (SVN->isSplat()) {
    unsigned Lane = std::max(0, SVN->getSplatIndex());
    EVT ScalarTy = MinLegalExtractEltScalarTy(VT.getVectorElementType());
    SDValue SplatEl = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ScalarTy, Op1,
                                  DAG.getConstant(Lane, DL, MVT::i64));
    Op = DAG.getNode(ISD::SPLAT_VECTOR, DL, ContainerVT, SplatEl);
    return convertFromScalableVector(DAG, VT, Op);
  }

  bool ReverseEXT = false;
  unsigned Imm;
  if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm) &&
      Imm == VT.getVectorNumElements() - 1) {
    if (ReverseEXT)
      std::swap(Op1, Op2);
    EVT ScalarTy = MinLegalExtractEltScalarTy(VT.getVectorElementType());
    SDValue Scalar = DAG.getNode(
        ISD::EXTRACT_VECTOR_ELT, DL, ScalarTy, Op1,
        DAG.getConstant(VT.getVectorNumElements() - 1, DL, MVT::i64));
    Op = DAG.getNode(AArch64ISD::INSR, DL, ContainerVT, Op2, Scalar);
    return convertFromScalableVector(DAG, VT, Op);
  }

  unsigned EltSize = VT.getScalarSizeInBits();
  for (unsigned BlockSize : {64U, 32U, 16U}) {
    if (isREVMask(ShuffleMask, EltSize, VT.getVectorNumElements(), BlockSize)) {
      unsigned RevOp;
      if (EltSize == 8)
        RevOp = AArch64ISD::BSWAP_MERGE_PASSTHRU;
      else if (EltSize == 16)
        RevOp = AArch64ISD::REVH_MERGE_PASSTHRU;
      else
        RevOp = AArch64ISD::REVW_MERGE_PASSTHRU;
      EVT BlockedVT =
          getPackedSVEVectorVT(EVT::getIntegerVT(*DAG.getContext(), BlockSize));
      SDValue Pg = getPredicateForVector(DAG, DL, BlockedVT);
      SDValue BlockedOp1 = DAG.getNode(ISD::BITCAST, DL, BlockedVT, Op1);
      SDValue BlockedRev = DAG.getNode(RevOp, DL, BlockedVT, Pg, BlockedOp1,
                                       DAG.getUNDEF(BlockedVT));
      SDValue Container =
          DAG.getNode(ISD::BITCAST, DL, ContainerVT, BlockedRev);
      return convertFromScalableVector(DAG, VT, Container);
    }
  }

  if (Subtarget->hasSVE2p1() && EltSize == 64 &&
      isREVMask(ShuffleMask, EltSize, VT.getVectorNumElements(), 128)) {
    SDValue Pg = getPredicateForVector(DAG, DL, VT);
    SDValue Revd = DAG.getNode(AArch64ISD::REVD_MERGE_PASSTHRU, DL, ContainerVT,
                               Pg, Op1, DAG.getUNDEF(ContainerVT));
    return convertFromScalableVector(DAG, VT, Revd);
  }

  unsigned WhichResult;
  if (isZIPMask(ShuffleMask, VT.getVectorNumElements(), WhichResult) &&
      WhichResult == 0)
    return convertFromScalableVector(
        DAG, VT, DAG.getNode(AArch64ISD::ZIP1, DL, ContainerVT, Op1, Op2));

  if (isTRNMask(ShuffleMask, VT.getVectorNumElements(), WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
    return convertFromScalableVector(
        DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op2));
  }

  if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult) && WhichResult == 0)
    return convertFromScalableVector(
        DAG, VT, DAG.getNode(AArch64ISD::ZIP1, DL, ContainerVT, Op1, Op1));

  if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
    unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
    return convertFromScalableVector(
        DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op1));
  }

  // Functions like isZIPMask return true when a ISD::VECTOR_SHUFFLE's mask
  // represents the same logical operation as performed by a ZIP instruction. In
  // isolation these functions do not mean the ISD::VECTOR_SHUFFLE is exactly
  // equivalent to an AArch64 instruction. There's the extra component of
  // ISD::VECTOR_SHUFFLE's value type to consider. Prior to SVE these functions
  // only operated on 64/128bit vector types that have a direct mapping to a
  // target register and so an exact mapping is implied.
  // However, when using SVE for fixed length vectors, most legal vector types
  // are actually sub-vectors of a larger SVE register. When mapping
  // ISD::VECTOR_SHUFFLE to an SVE instruction care must be taken to consider
  // how the mask's indices translate. Specifically, when the mapping requires
  // an exact meaning for a specific vector index (e.g. Index X is the last
  // vector element in the register) then such mappings are often only safe when
  // the exact SVE register size is know. The main exception to this is when
  // indices are logically relative to the first element of either
  // ISD::VECTOR_SHUFFLE operand because these relative indices don't change
  // when converting from fixed-length to scalable vector types (i.e. the start
  // of a fixed length vector is always the start of a scalable vector).
  unsigned MinSVESize = Subtarget->getMinSVEVectorSizeInBits();
  unsigned MaxSVESize = Subtarget->getMaxSVEVectorSizeInBits();
  if (MinSVESize == MaxSVESize && MaxSVESize == VT.getSizeInBits()) {
    if (ShuffleVectorInst::isReverseMask(ShuffleMask, ShuffleMask.size()) &&
        Op2.isUndef()) {
      Op = DAG.getNode(ISD::VECTOR_REVERSE, DL, ContainerVT, Op1);
      return convertFromScalableVector(DAG, VT, Op);
    }

    if (isZIPMask(ShuffleMask, VT.getVectorNumElements(), WhichResult) &&
        WhichResult != 0)
      return convertFromScalableVector(
          DAG, VT, DAG.getNode(AArch64ISD::ZIP2, DL, ContainerVT, Op1, Op2));

    if (isUZPMask(ShuffleMask, VT.getVectorNumElements(), WhichResult)) {
      unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
      return convertFromScalableVector(
          DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op2));
    }

    if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult) && WhichResult != 0)
      return convertFromScalableVector(
          DAG, VT, DAG.getNode(AArch64ISD::ZIP2, DL, ContainerVT, Op1, Op1));

    if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
      unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
      return convertFromScalableVector(
          DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op1));
    }

    if ((Subtarget->hasSVE2p1() || Subtarget->hasSME2p1()) &&
        Subtarget->isSVEorStreamingSVEAvailable()) {
      assert(VT.getFixedSizeInBits() % AArch64::SVEBitsPerBlock == 0 &&
             "Unsupported SVE vector size");

      unsigned Segments = VT.getFixedSizeInBits() / AArch64::SVEBitsPerBlock;
      unsigned SegmentElts = VT.getVectorNumElements() / Segments;
      if (std::optional<unsigned> Lane =
              isDUPQMask(ShuffleMask, Segments, SegmentElts)) {
        SDValue IID =
            DAG.getConstant(Intrinsic::aarch64_sve_dup_laneq, DL, MVT::i64);
        return convertFromScalableVector(
            DAG, VT,
            DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, ContainerVT,
                        {IID, Op1,
                         DAG.getConstant(*Lane, DL, MVT::i64,
                                         /*isTarget=*/true)}));
      }
    }
  }

  // Try to widen the shuffle before generating a possibly expensive SVE TBL.
  // This may allow the shuffle to be matched as something cheaper like ZIP1.
  if (SDValue WideOp = tryWidenMaskForShuffle(Op, DAG))
    return WideOp;

  // Avoid producing TBL instruction if we don't know SVE register minimal size,
  // unless NEON is not available and we can assume minimal SVE register size is
  // 128-bits.
  if (MinSVESize || !Subtarget->isNeonAvailable())
    return GenerateFixedLengthSVETBL(Op, Op1, Op2, ShuffleMask, VT, ContainerVT,
                                     DAG);

  return SDValue();
}

SDValue AArch64TargetLowering::getSVESafeBitCast(EVT VT, SDValue Op,
                                                 SelectionDAG &DAG) const {
  SDLoc DL(Op);
  EVT InVT = Op.getValueType();

  assert(VT.isScalableVector() && isTypeLegal(VT) &&
         InVT.isScalableVector() && isTypeLegal(InVT) &&
         "Only expect to cast between legal scalable vector types!");
  assert(VT.getVectorElementType() != MVT::i1 &&
         InVT.getVectorElementType() != MVT::i1 &&
         "For predicate bitcasts, use getSVEPredicateBitCast");

  if (InVT == VT)
    return Op;

  EVT PackedVT = getPackedSVEVectorVT(VT.getVectorElementType());
  EVT PackedInVT = getPackedSVEVectorVT(InVT.getVectorElementType());

  // Safe bitcasting between unpacked vector types of different element counts
  // is currently unsupported because the following is missing the necessary
  // work to ensure the result's elements live where they're supposed to within
  // an SVE register.
  //                01234567
  // e.g. nxv2i32 = XX??XX??
  //      nxv4f16 = X?X?X?X?
  assert((VT.getVectorElementCount() == InVT.getVectorElementCount() ||
          VT == PackedVT || InVT == PackedInVT) &&
         "Unexpected bitcast!");

  // Pack input if required.
  if (InVT != PackedInVT)
    Op = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, PackedInVT, Op);

  if (Subtarget->isLittleEndian() ||
      PackedVT.getScalarSizeInBits() == PackedInVT.getScalarSizeInBits())
    Op = DAG.getNode(ISD::BITCAST, DL, PackedVT, Op);
  else {
    EVT PackedVTAsInt = PackedVT.changeTypeToInteger();
    EVT PackedInVTAsInt = PackedInVT.changeTypeToInteger();

    // Simulate the effect of casting through memory.
    Op = DAG.getNode(ISD::BITCAST, DL, PackedInVTAsInt, Op);
    if (PackedInVTAsInt.getScalarSizeInBits() != 8)
      Op = DAG.getNode(ISD::BSWAP, DL, PackedInVTAsInt, Op);
    Op = DAG.getNode(AArch64ISD::NVCAST, DL, PackedVTAsInt, Op);
    if (PackedVTAsInt.getScalarSizeInBits() != 8)
      Op = DAG.getNode(ISD::BSWAP, DL, PackedVTAsInt, Op);
    Op = DAG.getNode(ISD::BITCAST, DL, PackedVT, Op);
  }

  // Unpack result if required.
  if (VT != PackedVT)
    Op = DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, VT, Op);

  return Op;
}

bool AArch64TargetLowering::isAllActivePredicate(SelectionDAG &DAG,
                                                 SDValue N) const {
  return ::isAllActivePredicate(DAG, N);
}

EVT AArch64TargetLowering::getPromotedVTForPredicate(EVT VT) const {
  return ::getPromotedVTForPredicate(VT);
}

bool AArch64TargetLowering::SimplifyDemandedBitsForTargetNode(
    SDValue Op, const APInt &OriginalDemandedBits,
    const APInt &OriginalDemandedElts, KnownBits &Known, TargetLoweringOpt &TLO,
    unsigned Depth) const {

  unsigned Opc = Op.getOpcode();
  switch (Opc) {
  case AArch64ISD::VSHL: {
    // Match (VSHL (VLSHR Val X) X)
    SDValue ShiftL = Op;
    SDValue ShiftR = Op->getOperand(0);
    if (ShiftR->getOpcode() != AArch64ISD::VLSHR)
      return false;

    if (!ShiftL.hasOneUse() || !ShiftR.hasOneUse())
      return false;

    unsigned ShiftLBits = ShiftL->getConstantOperandVal(1);
    unsigned ShiftRBits = ShiftR->getConstantOperandVal(1);

    // Other cases can be handled as well, but this is not
    // implemented.
    if (ShiftRBits != ShiftLBits)
      return false;

    unsigned ScalarSize = Op.getScalarValueSizeInBits();
    assert(ScalarSize > ShiftLBits && "Invalid shift imm");

    APInt ZeroBits = APInt::getLowBitsSet(ScalarSize, ShiftLBits);
    APInt UnusedBits = ~OriginalDemandedBits;

    if ((ZeroBits & UnusedBits) != ZeroBits)
      return false;

    // All bits that are zeroed by (VSHL (VLSHR Val X) X) are not
    // used - simplify to just Val.
    return TLO.CombineTo(Op, ShiftR->getOperand(0));
  }
  case AArch64ISD::BICi: {
    // Fold BICi if all destination bits already known to be zeroed
    SDValue Op0 = Op.getOperand(0);
    KnownBits KnownOp0 =
        TLO.DAG.computeKnownBits(Op0, OriginalDemandedElts, Depth + 1);
    // Op0 &= ~(ConstantOperandVal(1) << ConstantOperandVal(2))
    APInt BitsToClear =
        (Op->getConstantOperandAPInt(1) << Op->getConstantOperandAPInt(2))
            .trunc(KnownOp0.getBitWidth());
    APInt AlreadyZeroedBitsToClear = BitsToClear & KnownOp0.Zero;
    if (BitsToClear.isSubsetOf(AlreadyZeroedBitsToClear))
      return TLO.CombineTo(Op, Op0);

    Known = KnownOp0 & KnownBits::makeConstant(~BitsToClear);
    return false;
  }
  case ISD::INTRINSIC_WO_CHAIN: {
    if (auto ElementSize = IsSVECntIntrinsic(Op)) {
      unsigned MaxSVEVectorSizeInBits = Subtarget->getMaxSVEVectorSizeInBits();
      if (!MaxSVEVectorSizeInBits)
        MaxSVEVectorSizeInBits = AArch64::SVEMaxBitsPerVector;
      unsigned MaxElements = MaxSVEVectorSizeInBits / *ElementSize;
      // The SVE count intrinsics don't support the multiplier immediate so we
      // don't have to account for that here. The value returned may be slightly
      // over the true required bits, as this is based on the "ALL" pattern. The
      // other patterns are also exposed by these intrinsics, but they all
      // return a value that's strictly less than "ALL".
      unsigned RequiredBits = llvm::bit_width(MaxElements);
      unsigned BitWidth = Known.Zero.getBitWidth();
      if (RequiredBits < BitWidth)
        Known.Zero.setHighBits(BitWidth - RequiredBits);
      return false;
    }
  }
  }

  return TargetLowering::SimplifyDemandedBitsForTargetNode(
      Op, OriginalDemandedBits, OriginalDemandedElts, Known, TLO, Depth);
}

bool AArch64TargetLowering::isTargetCanonicalConstantNode(SDValue Op) const {
  return Op.getOpcode() == AArch64ISD::DUP ||
         Op.getOpcode() == AArch64ISD::MOVI ||
         Op.getOpcode() == AArch64ISD::MOVIshift ||
         (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
          Op.getOperand(0).getOpcode() == AArch64ISD::DUP) ||
         TargetLowering::isTargetCanonicalConstantNode(Op);
}

bool AArch64TargetLowering::isComplexDeinterleavingSupported() const {
  return Subtarget->hasSVE() || Subtarget->hasSVE2() ||
         Subtarget->hasComplxNum();
}

bool AArch64TargetLowering::isComplexDeinterleavingOperationSupported(
    ComplexDeinterleavingOperation Operation, Type *Ty) const {
  auto *VTy = dyn_cast<VectorType>(Ty);
  if (!VTy)
    return false;

  // If the vector is scalable, SVE is enabled, implying support for complex
  // numbers. Otherwise, we need to ensure complex number support is available
  if (!VTy->isScalableTy() && !Subtarget->hasComplxNum())
    return false;

  auto *ScalarTy = VTy->getScalarType();
  unsigned NumElements = VTy->getElementCount().getKnownMinValue();

  // We can only process vectors that have a bit size of 128 or higher (with an
  // additional 64 bits for Neon). Additionally, these vectors must have a
  // power-of-2 size, as we later split them into the smallest supported size
  // and merging them back together after applying complex operation.
  unsigned VTyWidth = VTy->getScalarSizeInBits() * NumElements;
  if ((VTyWidth < 128 && (VTy->isScalableTy() || VTyWidth != 64)) ||
      !llvm::isPowerOf2_32(VTyWidth))
    return false;

  if (ScalarTy->isIntegerTy() && Subtarget->hasSVE2() && VTy->isScalableTy()) {
    unsigned ScalarWidth = ScalarTy->getScalarSizeInBits();

    if (Operation == ComplexDeinterleavingOperation::CDot)
      return ScalarWidth == 32 || ScalarWidth == 64;
    return 8 <= ScalarWidth && ScalarWidth <= 64;
  }

  // CDot is not supported outside of scalable/sve scopes
  if (Operation == ComplexDeinterleavingOperation::CDot)
    return false;

  return (ScalarTy->isHalfTy() && Subtarget->hasFullFP16()) ||
         ScalarTy->isFloatTy() || ScalarTy->isDoubleTy();
}

Value *AArch64TargetLowering::createComplexDeinterleavingIR(
    IRBuilderBase &B, ComplexDeinterleavingOperation OperationType,
    ComplexDeinterleavingRotation Rotation, Value *InputA, Value *InputB,
    Value *Accumulator) const {
  VectorType *Ty = cast<VectorType>(InputA->getType());
  if (Accumulator == nullptr)
    Accumulator = Constant::getNullValue(Ty);
  bool IsScalable = Ty->isScalableTy();
  bool IsInt = Ty->getElementType()->isIntegerTy();

  unsigned TyWidth =
      Ty->getScalarSizeInBits() * Ty->getElementCount().getKnownMinValue();

  assert(((TyWidth >= 128 && llvm::isPowerOf2_32(TyWidth)) || TyWidth == 64) &&
         "Vector type must be either 64 or a power of 2 that is at least 128");

  if (TyWidth > 128) {
    int Stride = Ty->getElementCount().getKnownMinValue() / 2;
    int AccStride = cast<VectorType>(Accumulator->getType())
                        ->getElementCount()
                        .getKnownMinValue() /
                    2;
    auto *HalfTy = VectorType::getHalfElementsVectorType(Ty);
    auto *LowerSplitA = B.CreateExtractVector(HalfTy, InputA, uint64_t(0));
    auto *LowerSplitB = B.CreateExtractVector(HalfTy, InputB, uint64_t(0));
    auto *UpperSplitA = B.CreateExtractVector(HalfTy, InputA, Stride);
    auto *UpperSplitB = B.CreateExtractVector(HalfTy, InputB, Stride);
    Value *LowerSplitAcc = nullptr;
    Value *UpperSplitAcc = nullptr;
    Type *FullTy = Ty;
    FullTy = Accumulator->getType();
    auto *HalfAccTy = VectorType::getHalfElementsVectorType(
        cast<VectorType>(Accumulator->getType()));
    LowerSplitAcc = B.CreateExtractVector(HalfAccTy, Accumulator, uint64_t(0));
    UpperSplitAcc = B.CreateExtractVector(HalfAccTy, Accumulator, AccStride);
    auto *LowerSplitInt = createComplexDeinterleavingIR(
        B, OperationType, Rotation, LowerSplitA, LowerSplitB, LowerSplitAcc);
    auto *UpperSplitInt = createComplexDeinterleavingIR(
        B, OperationType, Rotation, UpperSplitA, UpperSplitB, UpperSplitAcc);

    auto *Result = B.CreateInsertVector(FullTy, PoisonValue::get(FullTy),
                                        LowerSplitInt, uint64_t(0));
    return B.CreateInsertVector(FullTy, Result, UpperSplitInt, AccStride);
  }

  if (OperationType == ComplexDeinterleavingOperation::CMulPartial) {
    if (IsScalable) {
      if (IsInt)
        return B.CreateIntrinsic(
            Intrinsic::aarch64_sve_cmla_x, Ty,
            {Accumulator, InputA, InputB, B.getInt32((int)Rotation * 90)});

      auto *Mask = B.getAllOnesMask(Ty->getElementCount());
      return B.CreateIntrinsic(
          Intrinsic::aarch64_sve_fcmla, Ty,
          {Mask, Accumulator, InputA, InputB, B.getInt32((int)Rotation * 90)});
    }

    Intrinsic::ID IdMap[4] = {Intrinsic::aarch64_neon_vcmla_rot0,
                              Intrinsic::aarch64_neon_vcmla_rot90,
                              Intrinsic::aarch64_neon_vcmla_rot180,
                              Intrinsic::aarch64_neon_vcmla_rot270};


    return B.CreateIntrinsic(IdMap[(int)Rotation], Ty,
                             {Accumulator, InputA, InputB});
  }

  if (OperationType == ComplexDeinterleavingOperation::CAdd) {
    if (IsScalable) {
      if (Rotation == ComplexDeinterleavingRotation::Rotation_90 ||
          Rotation == ComplexDeinterleavingRotation::Rotation_270) {
        if (IsInt)
          return B.CreateIntrinsic(
              Intrinsic::aarch64_sve_cadd_x, Ty,
              {InputA, InputB, B.getInt32((int)Rotation * 90)});

        auto *Mask = B.getAllOnesMask(Ty->getElementCount());
        return B.CreateIntrinsic(
            Intrinsic::aarch64_sve_fcadd, Ty,
            {Mask, InputA, InputB, B.getInt32((int)Rotation * 90)});
      }
      return nullptr;
    }

    Intrinsic::ID IntId = Intrinsic::not_intrinsic;
    if (Rotation == ComplexDeinterleavingRotation::Rotation_90)
      IntId = Intrinsic::aarch64_neon_vcadd_rot90;
    else if (Rotation == ComplexDeinterleavingRotation::Rotation_270)
      IntId = Intrinsic::aarch64_neon_vcadd_rot270;

    if (IntId == Intrinsic::not_intrinsic)
      return nullptr;

    return B.CreateIntrinsic(IntId, Ty, {InputA, InputB});
  }

  if (OperationType == ComplexDeinterleavingOperation::CDot && IsInt &&
      IsScalable) {
    return B.CreateIntrinsic(
        Intrinsic::aarch64_sve_cdot, Accumulator->getType(),
        {Accumulator, InputA, InputB, B.getInt32((int)Rotation * 90)});
  }

  return nullptr;
}

bool AArch64TargetLowering::preferScalarizeSplat(SDNode *N) const {
  unsigned Opc = N->getOpcode();
  if (ISD::isExtOpcode(Opc)) {
    if (any_of(N->users(),
               [&](SDNode *Use) { return Use->getOpcode() == ISD::MUL; }))
      return false;
  }
  return true;
}

unsigned AArch64TargetLowering::getMinimumJumpTableEntries() const {
  return Subtarget->getMinimumJumpTableEntries();
}

MVT AArch64TargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
                                                         CallingConv::ID CC,
                                                         EVT VT) const {
  bool NonUnitFixedLengthVector =
      VT.isFixedLengthVector() && !VT.getVectorElementCount().isScalar();
  if (!NonUnitFixedLengthVector || !Subtarget->useSVEForFixedLengthVectors())
    return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);

  EVT VT1;
  MVT RegisterVT;
  unsigned NumIntermediates;
  getVectorTypeBreakdownForCallingConv(Context, CC, VT, VT1, NumIntermediates,
                                       RegisterVT);
  return RegisterVT;
}

unsigned AArch64TargetLowering::getNumRegistersForCallingConv(
    LLVMContext &Context, CallingConv::ID CC, EVT VT) const {
  bool NonUnitFixedLengthVector =
      VT.isFixedLengthVector() && !VT.getVectorElementCount().isScalar();
  if (!NonUnitFixedLengthVector || !Subtarget->useSVEForFixedLengthVectors())
    return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);

  EVT VT1;
  MVT VT2;
  unsigned NumIntermediates;
  return getVectorTypeBreakdownForCallingConv(Context, CC, VT, VT1,
                                              NumIntermediates, VT2);
}

unsigned AArch64TargetLowering::getVectorTypeBreakdownForCallingConv(
    LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
    unsigned &NumIntermediates, MVT &RegisterVT) const {
  int NumRegs = TargetLowering::getVectorTypeBreakdownForCallingConv(
      Context, CC, VT, IntermediateVT, NumIntermediates, RegisterVT);
  if (!RegisterVT.isFixedLengthVector() ||
      RegisterVT.getFixedSizeInBits() <= 128)
    return NumRegs;

  assert(Subtarget->useSVEForFixedLengthVectors() && "Unexpected mode!");
  assert(IntermediateVT == RegisterVT && "Unexpected VT mismatch!");
  assert(RegisterVT.getFixedSizeInBits() % 128 == 0 && "Unexpected size!");

  // A size mismatch here implies either type promotion or widening and would
  // have resulted in scalarisation if larger vectors had not be available.
  if (RegisterVT.getSizeInBits() * NumRegs != VT.getSizeInBits()) {
    EVT EltTy = VT.getVectorElementType();
    EVT NewVT = EVT::getVectorVT(Context, EltTy, ElementCount::getFixed(1));
    if (!isTypeLegal(NewVT))
      NewVT = EltTy;

    IntermediateVT = NewVT;
    NumIntermediates = VT.getVectorNumElements();
    RegisterVT = getRegisterType(Context, NewVT);
    return NumIntermediates;
  }

  // SVE VLS support does not introduce a new ABI so we should use NEON sized
  // types for vector arguments and returns.

  unsigned NumSubRegs = RegisterVT.getFixedSizeInBits() / 128;
  NumIntermediates *= NumSubRegs;
  NumRegs *= NumSubRegs;

  switch (RegisterVT.getVectorElementType().SimpleTy) {
  default:
    llvm_unreachable("unexpected element type for vector");
  case MVT::i8:
    IntermediateVT = RegisterVT = MVT::v16i8;
    break;
  case MVT::i16:
    IntermediateVT = RegisterVT = MVT::v8i16;
    break;
  case MVT::i32:
    IntermediateVT = RegisterVT = MVT::v4i32;
    break;
  case MVT::i64:
    IntermediateVT = RegisterVT = MVT::v2i64;
    break;
  case MVT::f16:
    IntermediateVT = RegisterVT = MVT::v8f16;
    break;
  case MVT::f32:
    IntermediateVT = RegisterVT = MVT::v4f32;
    break;
  case MVT::f64:
    IntermediateVT = RegisterVT = MVT::v2f64;
    break;
  case MVT::bf16:
    IntermediateVT = RegisterVT = MVT::v8bf16;
    break;
  }

  return NumRegs;
}

bool AArch64TargetLowering::hasInlineStackProbe(
    const MachineFunction &MF) const {
  return !Subtarget->isTargetWindows() &&
         MF.getInfo<AArch64FunctionInfo>()->hasStackProbing();
}

bool AArch64TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
  switch (Opc) {
  case ISD::TRUNCATE_SSAT_S:
  case ISD::TRUNCATE_SSAT_U:
  case ISD::TRUNCATE_USAT_U:
    if (VT == MVT::v8i8 || VT == MVT::v4i16 || VT == MVT::v2i32)
      return true;
  }

  return TargetLowering::isTypeDesirableForOp(Opc, VT);
}

bool AArch64TargetLowering::shouldPreservePtrArith(const Function &F,
                                                   EVT VT) const {
  return Subtarget->hasCPA() && UseFEATCPACodegen;
}