AArch64TargetTransformInfo.cpp

Go to the documentation of this file.

//===-- AArch64TargetTransformInfo.cpp - AArch64 specific TTI -------------===//
//
// 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
//
//===----------------------------------------------------------------------===//
 
#include "AArch64TargetTransformInfo.h"
#include "AArch64ExpandImm.h"
#include "AArch64PerfectShuffle.h"
#include "AArch64SMEAttributes.h"
#include "MCTargetDesc/AArch64AddressingModes.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/bit.h"
#include "llvm/Analysis/LoopInfo.h"
#include "llvm/Analysis/TargetTransformInfo.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/CodeGen/BasicTTIImpl.h"
#include "llvm/CodeGen/CostTable.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsAArch64.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/Support/Debug.h"
#include "llvm/TargetParser/AArch64TargetParser.h"
#include "llvm/Transforms/InstCombine/InstCombiner.h"
#include "llvm/Transforms/Utils/UnrollLoop.h"
#include "llvm/Transforms/Vectorize/LoopVectorizationLegality.h"
#include <algorithm>
#include <optional>
using namespace llvm;
using namespace llvm::PatternMatch;
 
#define DEBUG_TYPE "aarch64tti"
 
namespace {
class TailFoldingOption {
  // These bitfields will only ever be set to something non-zero in operator=,
  // when setting the -sve-tail-folding option. This option should always be of
  // the form (default|simple|all|disable)[+(Flag1|Flag2|etc)], where here
  // InitialBits is one of (disabled|all|simple). EnableBits represents
  // additional flags we're enabling, and DisableBits for those flags we're
  // disabling. The default flag is tracked in the variable NeedsDefault, since
  // at the time of setting the option we may not know what the default value
  // for the CPU is.
  TailFoldingOpts InitialBits = TailFoldingOpts::Disabled;
  TailFoldingOpts EnableBits = TailFoldingOpts::Disabled;
  TailFoldingOpts DisableBits = TailFoldingOpts::Disabled;
 
  // This value needs to be initialised to true in case the user does not
  // explicitly set the -sve-tail-folding option.
  bool NeedsDefault = true;
 
  void setInitialBits(TailFoldingOpts Bits) { InitialBits = Bits; }
 
  void setNeedsDefault(bool V) { NeedsDefault = V; }
 
  void setEnableBit(TailFoldingOpts Bit) {
    EnableBits |= Bit;
    DisableBits &= ~Bit;
  }
 
  void setDisableBit(TailFoldingOpts Bit) {
    EnableBits &= ~Bit;
    DisableBits |= Bit;
  }
 
  TailFoldingOpts getBits(TailFoldingOpts DefaultBits) const {
    TailFoldingOpts Bits = TailFoldingOpts::Disabled;
 
    assert((InitialBits == TailFoldingOpts::Disabled || !NeedsDefault) &&
           "Initial bits should only include one of "
           "(disabled|all|simple|default)");
    Bits = NeedsDefault ? DefaultBits : InitialBits;
    Bits |= EnableBits;
    Bits &= ~DisableBits;
 
    return Bits;
  }
 
  void reportError(std::string Opt) {
    errs() << "invalid argument '" << Opt
           << "' to -sve-tail-folding=; the option should be of the form\n"
              "  (disabled|all|default|simple)[+(reductions|recurrences"
              "|reverse|noreductions|norecurrences|noreverse)]\n";
    report_fatal_error("Unrecognised tail-folding option");
  }
 
public:
 
  void operator=(const std::string &Val) {
    // If the user explicitly sets -sve-tail-folding= then treat as an error.
    if (Val.empty()) {
      reportError("");
      return;
    }
 
    // Since the user is explicitly setting the option we don't automatically
    // need the default unless they require it.
    setNeedsDefault(false);
 
    SmallVector<StringRef, 4> TailFoldTypes;
    StringRef(Val).split(TailFoldTypes, '+', -1, false);
 
    unsigned StartIdx = 1;
    if (TailFoldTypes[0] == "disabled")
      setInitialBits(TailFoldingOpts::Disabled);
    else if (TailFoldTypes[0] == "all")
      setInitialBits(TailFoldingOpts::All);
    else if (TailFoldTypes[0] == "default")
      setNeedsDefault(true);
    else if (TailFoldTypes[0] == "simple")
      setInitialBits(TailFoldingOpts::Simple);
    else {
      StartIdx = 0;
      setInitialBits(TailFoldingOpts::Disabled);
    }
 
    for (unsigned I = StartIdx; I < TailFoldTypes.size(); I++) {
      if (TailFoldTypes[I] == "reductions")
        setEnableBit(TailFoldingOpts::Reductions);
      else if (TailFoldTypes[I] == "recurrences")
        setEnableBit(TailFoldingOpts::Recurrences);
      else if (TailFoldTypes[I] == "reverse")
        setEnableBit(TailFoldingOpts::Reverse);
      else if (TailFoldTypes[I] == "noreductions")
        setDisableBit(TailFoldingOpts::Reductions);
      else if (TailFoldTypes[I] == "norecurrences")
        setDisableBit(TailFoldingOpts::Recurrences);
      else if (TailFoldTypes[I] == "noreverse")
        setDisableBit(TailFoldingOpts::Reverse);
      else
        reportError(Val);
    }
  }
 
  bool isDisabled(TailFoldingOpts DefaultBits) const {
    return getBits(DefaultBits) == TailFoldingOpts::Disabled;
  }
 
  bool satisfies(TailFoldingOpts DefaultBits, TailFoldingOpts Required) const {
    return (getBits(DefaultBits) & Required) == Required;
  }
};
} // namespace
 
TailFoldingOption TailFoldingOptionLoc;
 
static cl::opt<TailFoldingOption, true, cl::parser<std::string>> SVETailFolding(
    "sve-tail-folding",
    cl::desc(
        "Control the use of vectorisation using tail-folding for SVE where the"
        " option is specified in the form (Initial)[+(Flag1|Flag2|...)]:"
        "\ndisabled      (Initial) No loop types will vectorize using "
        "tail-folding"
        "\ndefault       (Initial) Uses the default tail-folding settings for "
        "the target CPU"
        "\nall           (Initial) All legal loop types will vectorize using "
        "tail-folding"
        "\nsimple        (Initial) Use tail-folding for simple loops (not "
        "reductions or recurrences)"
        "\nreductions    Use tail-folding for loops containing reductions"
        "\nnoreductions  Inverse of above"
        "\nrecurrences   Use tail-folding for loops containing fixed order "
        "recurrences"
        "\nnorecurrences Inverse of above"
        "\nreverse       Use tail-folding for loops requiring reversed "
        "predicates"
        "\nnoreverse     Inverse of above"),
    cl::location(TailFoldingOptionLoc));
 
/// Returns true if the function has explicit operations that can only be
/// lowered using incompatible instructions for the selected mode. This also
/// returns true if the function F may use or modify ZA state.
bool AArch64TTIImpl::hasPossibleIncompatibleOps(const Function *F,
                                                bool ConsiderZA,
                                                bool ConsiderZT,
                                                bool ConsiderSM) const {
  if (!ConsiderZA && !ConsiderZT && !ConsiderSM)
    return false;
 
  bool IsAlwaysInline = F->hasFnAttribute(Attribute::AlwaysInline);
  // If the interface has vscale-dependent arguments/return value, then the ACLE
  // describes that the function can only have defined behaviour if vscale is
  // the same for both streaming/non-streaming mode, meaning that we can assume
  // vscale is equivalent.
  // FIXME: Handle the case where vscale_range is known and equivalent for
  // caller and callee.
  bool AssumeVScaleIsEquivalent =
      F->getReturnType()->isScalableTy() ||
      any_of(F->getFunctionType()->params(),
             [](const Type *T) { return T->isScalableTy(); });
 
  for (const BasicBlock &BB : *F) {
    for (const Instruction &I : BB) {
      if (auto *CB = dyn_cast<CallBase>(&I)) {
        SMECallAttrs CallAttrs(*CB, &TLI->getRuntimeLibcallsInfo());
 
        // Inline asm must be rejected as it could use SME state.
        if (CB->isInlineAsm())
          return true;
 
        if (CallAttrs.callee().isSMEABIRoutine())
          return true;
 
        // If the callee has calls to streaming compatible functions, then those
        // may have statements which depend on SME state (e.g. vscale or
        // explicit reads of PSTATE.SM). If we were to inline those calls,
        // behaviour may change after inlining because the streaming-compatible
        // function would be executed in a different streaming mode.
        // Conservatively avoid inlining for now.
        if (ConsiderSM && CallAttrs.callee().hasStreamingCompatibleInterface())
          return true;
      }
 
      if (!ConsiderSM)
        continue;
 
      auto *II = dyn_cast<IntrinsicInst>(&I);
 
      // FIXME: Be less conservative about target intrinsics.
      if (II && II->getCalledFunction()->isTargetIntrinsic())
        return true;
 
      // Determine if the operation is vscale dependent by looking at it's
      // operand/result types.
      if (!AssumeVScaleIsEquivalent &&
          ((II && II->getIntrinsicID() == Intrinsic::vscale) ||
           I.getType()->isScalableTy() ||
           any_of(
               I.operand_values(),
               [](const Value *V) { return V->getType()->isScalableTy(); }) ||
           (isa<GetElementPtrInst>(I) && cast<GetElementPtrInst>(I)
                                             .getSourceElementType()
                                             ->isScalableTy()) ||
           (isa<AllocaInst>(I) && cast<AllocaInst>(I).isScalable())))
        return true;
 
      // Inlining operations on fixed-length vectors when the streaming
      // mode does not match is rejected because performance may be impacted.
      // This decision should eventually be moved to the cost-model.
      if (!IsAlwaysInline && (isa<FixedVectorType>(I.getType()) ||
                              any_of(I.operand_values(), [](const Value *V) {
                                return isa<FixedVectorType>(V->getType());
                              })))
        return true;
 
      // If the operation is invalid in the current mode, it should not be
      // inlined.
      SmallVector<Value *> Opnds(I.op_begin(), I.op_end());
      if (!getInstructionCost(&I, Opnds, TTI::TCK_RecipThroughput).isValid())
        return true;
    }
  }
 
  return false;
}
 
static void extractAttrFeatures(const Function &F, const AArch64TTIImpl *TTI,
                                SmallVectorImpl<StringRef> &Features) {
  StringRef AttributeStr =
      TTI->isMultiversionedFunction(F) ? "fmv-features" : "target-features";
  StringRef FeatureStr = F.getFnAttribute(AttributeStr).getValueAsString();
  FeatureStr.split(Features, ",");
}
 
APInt AArch64TTIImpl::getFeatureMask(const Function &F) const {
  SmallVector<StringRef, 8> Features;
  extractAttrFeatures(F, this, Features);
  return AArch64::getCpuSupportsMask(Features);
}
 
APInt AArch64TTIImpl::getPriorityMask(const Function &F) const {
  SmallVector<StringRef, 8> Features;
  extractAttrFeatures(F, this, Features);
  return AArch64::getFMVPriority(Features);
}
 
bool AArch64TTIImpl::isMultiversionedFunction(const Function &F) const {
  return F.hasFnAttribute("fmv-features");
}
 
/// The compiler must not inline when that may alter the behavior of the
/// program. This is especially relevant around SME which implements different
/// runtime modes and maintains external state through attributes.
///
/// The compiler must not inline when:
//
/// * The module is compiled with 'strict-fp' and the called function
///   implements a different FP environment than the caller.
///
/// * The called function has operations that are incompatible in the mode
///   of the caller, e.g.:
///   * inlining non-streaming-only instructions into a streaming function.
///   * inlining streaming-only instructions into a non-streaming function.
///
///   Inline asm blocks must be entered and exited in the [streaming] mode of
///   the parent function. There is no language-level mechanism to inform the
///   compiler that a particular inline asm block is streaming compatible, so
///   the compiler must reject this as a candidate for inlining.
///
/// * The called function contains vscale-dependent operations but otherwise
///   does not take/return VL-dependent arguments (see definition in the
///   ACLE (Arm C/C++ Language Extensions)).
///
/// * The called function sets up new ZA/ZT state into a function that already
///   has ZA or ZT state, as that is not valid as per the ACLE.
///
/// The compiler should not inline when:
//
/// * The called function has fixed-length vectors and the caller is in
///   streaming mode, as this may cause performance regressions. This should
///   never result in an error to be reported.
///
/// * The called function sets up new ZA/ZT state into a function that has no
///   ZA and no ZT state, as the compiler currently cannot transfer the
///   attribute to the caller. It may also impact performance, but that should
///   be covered by AArch64TTIImpl::getInlineCallPenalty().
///
/// If a callee is incompatible with its caller, then the inliner should prevent
/// inlining. LLVM's `alwaysinline` attribute has no impact on whether a callee
/// is inline compatible. This is especially relevant with streaming-mode
/// attributes, where  `alwaysinline`ing a function into a function with a
/// different streaming mode may otherwise impact runtime behaviour.
bool AArch64TTIImpl::areInlineCompatible(const Function *Caller,
                                         const Function *Callee) const {
  SMECallAttrs CallAttrs(*Caller, *Callee);
 
  // Never inline a function explicitly marked as being streaming,
  // into a non-streaming function. Assume it was marked as streaming
  // for a reason.
  if (CallAttrs.caller().hasNonStreamingInterfaceAndBody() &&
      CallAttrs.callee().hasStreamingInterfaceOrBody())
    return false;
 
  if (CallAttrs.callee().isNewZA() || CallAttrs.callee().isNewZT0())
    return false;
 
  // When inlining, we should consider the body of the function, not the
  // interface.
  if (CallAttrs.callee().hasStreamingBody()) {
    CallAttrs.callee().set(SMEAttrs::SM_Compatible, false);
    CallAttrs.callee().set(SMEAttrs::SM_Enabled, true);
  }
 
  bool ConsiderZA = CallAttrs.requiresZASave();
  bool ConsiderZT = CallAttrs.requiresPreservingZT0() ||
                    CallAttrs.requiresPreservingAllZAState();
  bool ConsiderSM = CallAttrs.requiresSMChange();
 
  // FP environment is interpreted differently between streaming mode and
  // non-streaming mode, so are not inline-compatible.
  // FIXME: Analyze whether the callee actually has any FP operations.
  if (ConsiderSM && (Caller->isStrictFP() || Callee->isStrictFP()))
    return false;
 
  if (hasPossibleIncompatibleOps(Callee, ConsiderZA, ConsiderZT, ConsiderSM))
    return false;
 
  return BaseT::areInlineCompatible(Caller, Callee);
}
 
bool AArch64TTIImpl::areTypesABICompatible(const Function *Caller,
                                           const Function *Callee,
                                           ArrayRef<Type *> Types) const {
  if (!BaseT::areTypesABICompatible(Caller, Callee, Types))
    return false;
 
  // We need to ensure that argument promotion does not attempt to promote
  // pointers to fixed-length vector types larger than 128 bits like
  // <8 x float> (and pointers to aggregate types which have such fixed-length
  // vector type members) into the values of the pointees. Such vector types
  // are used for SVE VLS but there is no ABI for SVE VLS arguments and the
  // backend cannot lower such value arguments. The 128-bit fixed-length SVE
  // types can be safely treated as 128-bit NEON types and they cannot be
  // distinguished in IR.
  if (ST->useSVEForFixedLengthVectors() && llvm::any_of(Types, [](Type *Ty) {
        auto FVTy = dyn_cast<FixedVectorType>(Ty);
        return FVTy &&
               FVTy->getScalarSizeInBits() * FVTy->getNumElements() > 128;
      }))
    return false;
 
  return true;
}
 
unsigned
AArch64TTIImpl::getInlineCallPenalty(const Function *F, const CallBase &Call,
                                     unsigned DefaultCallPenalty) const {
  // This function calculates a penalty for executing Call in F.
  //
  // There are two ways this function can be called:
  // (1)  F:
  //       call from F -> G (the call here is Call)
  //
  // For (1), Call.getCaller() == F, so it will always return a high cost if
  // a streaming-mode change is required (thus promoting the need to inline the
  // function)
  //
  // (2)  F:
  //       call from F -> G (the call here is not Call)
  //      G:
  //       call from G -> H (the call here is Call)
  //
  // For (2), if after inlining the body of G into F the call to H requires a
  // streaming-mode change, and the call to G from F would also require a
  // streaming-mode change, then there is benefit to do the streaming-mode
  // change only once and avoid inlining of G into F.
 
  SMEAttrs FAttrs(*F);
  SMECallAttrs CallAttrs(Call, &getTLI()->getRuntimeLibcallsInfo());
 
  if (SMECallAttrs(FAttrs, CallAttrs.callee()).requiresSMChange()) {
    if (F == Call.getCaller()) // (1)
      return ST->getCLOpts().call_penalty_sm_change * DefaultCallPenalty;
    if (SMECallAttrs(FAttrs, CallAttrs.caller()).requiresSMChange()) // (2)
      return ST->getCLOpts().inline_call_penalty_sm_change * DefaultCallPenalty;
  }
 
  return DefaultCallPenalty;
}
 
bool AArch64TTIImpl::shouldMaximizeVectorBandwidth(
    TargetTransformInfo::VectorKind K, Type *ElementTy) const {
  // Don't maximize bandwidth for i1s. The legal width of predicates is derived
  // from the data vectors they are used with. Maximizing the predicate width
  // will just result in excessive VFs with a lot of unpacking.
  if (ElementTy && ElementTy->isIntegerTy(1))
    return false;
 
  // If we don't know the specific class of vector just return true. Later VF
  // selection will ask us about fixed or scalable vectors.
  if (K == TargetTransformInfo::VectorKind::FixedOrScalable)
    return true;
 
  if (K == TargetTransformInfo::VectorKind::FixedWidth && ST->isNeonAvailable())
    return true;
 
  return K == TargetTransformInfo::VectorKind::Scalable &&
         ST->isSVEorStreamingSVEAvailable() &&
         !ST->disableMaximizeScalableBandwidth();
}
 
/// Calculate the cost of materializing a 64-bit value. This helper
/// method might only calculate a fraction of a larger immediate. Therefore it
/// is valid to return a cost of ZERO.
InstructionCost AArch64TTIImpl::getIntImmCost(int64_t Val) const {
  // Check if the immediate can be encoded within an instruction.
  if (Val == 0 || AArch64_AM::isLogicalImmediate(Val, 64))
    return 0;
 
  if (Val < 0)
    Val = ~Val;
 
  // Calculate how many moves we will need to materialize this constant.
  SmallVector<AArch64_IMM::ImmInsnModel, 4> Insn;
  AArch64_IMM::expandMOVImm(Val, 64, Insn);
  return Insn.size();
}
 
/// Calculate the cost of materializing the given constant.
InstructionCost
AArch64TTIImpl::getIntImmCost(const APInt &Imm, Type *Ty,
                              TTI::TargetCostKind CostKind) const {
  assert(Ty->isIntegerTy());
 
  unsigned BitSize = Ty->getPrimitiveSizeInBits();
  if (BitSize == 0)
    return ~0U;
 
  // Sign-extend all constants to a multiple of 64-bit.
  APInt ImmVal = Imm;
  if (BitSize & 0x3f)
    ImmVal = Imm.sext((BitSize + 63) & ~0x3fU);
 
  // Split the constant into 64-bit chunks and calculate the cost for each
  // chunk.
  InstructionCost Cost = 0;
  for (unsigned ShiftVal = 0; ShiftVal < BitSize; ShiftVal += 64) {
    APInt Tmp = ImmVal.ashr(ShiftVal).sextOrTrunc(64);
    int64_t Val = Tmp.getSExtValue();
    Cost += getIntImmCost(Val);
  }
  // We need at least one instruction to materialze the constant.
  return std::max<InstructionCost>(1, Cost);
}
 
InstructionCost AArch64TTIImpl::getIntImmCostInst(unsigned Opcode, unsigned Idx,
                                                  const APInt &Imm, Type *Ty,
                                                  TTI::TargetCostKind CostKind,
                                                  Instruction *Inst) const {
  assert(Ty->isIntegerTy());
 
  unsigned BitSize = Ty->getPrimitiveSizeInBits();
  // There is no cost model for constants with a bit size of 0. Return TCC_Free
  // here, so that constant hoisting will ignore this constant.
  if (BitSize == 0)
    return TTI::TCC_Free;
 
  unsigned ImmIdx = ~0U;
  switch (Opcode) {
  default:
    return TTI::TCC_Free;
  case Instruction::GetElementPtr:
    // Always hoist the base address of a GetElementPtr.
    if (Idx == 0)
      return 2 * TTI::TCC_Basic;
    return TTI::TCC_Free;
  case Instruction::Store:
    ImmIdx = 0;
    break;
  case Instruction::Add:
  case Instruction::Sub:
  case Instruction::Mul:
  case Instruction::UDiv:
  case Instruction::SDiv:
  case Instruction::URem:
  case Instruction::SRem:
  case Instruction::And:
  case Instruction::Or:
  case Instruction::Xor:
  case Instruction::ICmp:
    ImmIdx = 1;
    break;
  // Always return TCC_Free for the shift value of a shift instruction.
  case Instruction::Shl:
  case Instruction::LShr:
  case Instruction::AShr:
    if (Idx == 1)
      return TTI::TCC_Free;
    break;
  case Instruction::Trunc:
  case Instruction::ZExt:
  case Instruction::SExt:
  case Instruction::IntToPtr:
  case Instruction::PtrToInt:
  case Instruction::BitCast:
  case Instruction::PHI:
  case Instruction::Call:
  case Instruction::Select:
  case Instruction::Ret:
  case Instruction::Load:
    break;
  }
 
  if (Idx == ImmIdx) {
    int NumConstants = (BitSize + 63) / 64;
    InstructionCost Cost = AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
    return (Cost <= NumConstants * TTI::TCC_Basic)
               ? static_cast<int>(TTI::TCC_Free)
               : Cost;
  }
  return AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
}
 
InstructionCost
AArch64TTIImpl::getIntImmCostIntrin(Intrinsic::ID IID, unsigned Idx,
                                    const APInt &Imm, Type *Ty,
                                    TTI::TargetCostKind CostKind) const {
  assert(Ty->isIntegerTy());
 
  unsigned BitSize = Ty->getPrimitiveSizeInBits();
  // There is no cost model for constants with a bit size of 0. Return TCC_Free
  // here, so that constant hoisting will ignore this constant.
  if (BitSize == 0)
    return TTI::TCC_Free;
 
  // Most (all?) AArch64 intrinsics do not support folding immediates into the
  // selected instruction, so we compute the materialization cost for the
  // immediate directly.
  if (IID >= Intrinsic::aarch64_addg && IID <= Intrinsic::aarch64_udiv)
    return AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
 
  switch (IID) {
  default:
    return TTI::TCC_Free;
  case Intrinsic::sadd_with_overflow:
  case Intrinsic::uadd_with_overflow:
  case Intrinsic::ssub_with_overflow:
  case Intrinsic::usub_with_overflow:
  case Intrinsic::smul_with_overflow:
  case Intrinsic::umul_with_overflow:
    if (Idx == 1) {
      int NumConstants = (BitSize + 63) / 64;
      InstructionCost Cost = AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
      return (Cost <= NumConstants * TTI::TCC_Basic)
                 ? static_cast<int>(TTI::TCC_Free)
                 : Cost;
    }
    break;
  case Intrinsic::experimental_stackmap:
    if ((Idx < 2) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
      return TTI::TCC_Free;
    break;
  case Intrinsic::experimental_patchpoint_void:
  case Intrinsic::experimental_patchpoint:
    if ((Idx < 4) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
      return TTI::TCC_Free;
    break;
  case Intrinsic::experimental_gc_statepoint:
    if ((Idx < 5) || (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue())))
      return TTI::TCC_Free;
    break;
  }
  return AArch64TTIImpl::getIntImmCost(Imm, Ty, CostKind);
}
 
TargetTransformInfo::PopcntSupportKind
AArch64TTIImpl::getPopcntSupport(unsigned TyWidth) const {
  assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
  if (TyWidth == 32 || TyWidth == 64)
    return TTI::PSK_FastHardware;
  // TODO: AArch64TargetLowering::LowerCTPOP() supports 128bit popcount.
  return TTI::PSK_Software;
}
 
InstructionCost AArch64TTIImpl::getBranchMispredictPenalty() const {
  // MispredictPenalty is defined per-CPU in AArch64Sched*.td (e.g.,
  // AArch64SchedNeoverseV2.td).
  return ST->getMispredictionPenalty();
}
 
static bool isUnpackedVectorVT(EVT VecVT) {
  return VecVT.isScalableVector() &&
         VecVT.getSizeInBits().getKnownMinValue() < AArch64::SVEBitsPerBlock;
}
 
static InstructionCost getHistogramCost(const AArch64Subtarget *ST,
                                        const IntrinsicCostAttributes &ICA) {
  // We need to know at least the number of elements in the vector of buckets
  // and the size of each element to update.
  if (ICA.getArgTypes().size() < 2)
    return InstructionCost::getInvalid();
 
  // Only interested in costing for the hardware instruction from SVE2.
  if (!ST->isSVEAvailable() || !ST->hasSVE2())
    return InstructionCost::getInvalid();
 
  Type *BucketPtrsTy = ICA.getArgTypes()[0]; // Type of vector of pointers
  Type *EltTy = ICA.getArgTypes()[1];        // Type of bucket elements
  unsigned TotalHistCnts = 1;
 
  unsigned EltSize = EltTy->getScalarSizeInBits();
  // Only allow (up to 64b) integers or pointers
  if ((!EltTy->isIntegerTy() && !EltTy->isPointerTy()) || EltSize > 64)
    return InstructionCost::getInvalid();
 
  // FIXME: We should be able to generate histcnt for fixed-length vectors
  //        using ptrue with a specific VL.
  if (VectorType *VTy = dyn_cast<VectorType>(BucketPtrsTy)) {
    unsigned EC = VTy->getElementCount().getKnownMinValue();
    if (!isPowerOf2_64(EC) || !VTy->isScalableTy() || EC == 1)
      return InstructionCost::getInvalid();
 
    // HistCnt only supports 32b and 64b element types
    unsigned LegalEltSize = EltSize <= 32 ? 32 : 64;
 
    if (EC == 2 || (LegalEltSize == 32 && EC == 4))
      return InstructionCost(ST->getCLOpts().base_histcnt_cost);
 
    unsigned NaturalVectorWidth = AArch64::SVEBitsPerBlock / LegalEltSize;
    TotalHistCnts = EC / NaturalVectorWidth;
 
    return InstructionCost(ST->getCLOpts().base_histcnt_cost * TotalHistCnts);
  }
 
  return InstructionCost::getInvalid();
}
 
InstructionCost
AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
                                      TTI::TargetCostKind CostKind) const {
  // The code-generator is currently not able to handle scalable vectors
  // of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
  // it. This change will be removed when code-generation for these types is
  // sufficiently reliable.
  // Only allow masked ld/st to pass through to getMemIntrinsicInstrCost().
  auto *RetTy = ICA.getReturnType();
  if (auto *VTy = dyn_cast<ScalableVectorType>(RetTy))
    if (VTy->getElementCount() == ElementCount::getScalable(1) &&
        !is_contained({Intrinsic::masked_load, Intrinsic::masked_store},
                      ICA.getID()))
      return InstructionCost::getInvalid();
 
  switch (ICA.getID()) {
  case Intrinsic::experimental_vector_histogram_add: {
    InstructionCost HistCost = getHistogramCost(ST, ICA);
    // If the cost isn't valid, we may still be able to scalarize
    if (HistCost.isValid())
      return HistCost;
    break;
  }
  case Intrinsic::experimental_vector_histogram_uadd_sat:
  case Intrinsic::experimental_vector_histogram_umax:
  case Intrinsic::experimental_vector_histogram_umin:
    // No hardware support for these histogram variants on AArch64.
    return InstructionCost::getInvalid();
  case Intrinsic::clmul: {
    auto LT = getTypeLegalizationCost(RetTy);
 
    // PMUL v8i8/v16i8 is always available on AArch64
    if (ST->hasNEON()) {
      if (LT.second == MVT::v8i8 || LT.second == MVT::v16i8)
        return LT.first;
 
      // Scalar i8 lowers through scalar/vector moves around PMUL.
      if (TLI->getValueType(DL, RetTy, true) == MVT::i8) {
        auto *VecTy =
            FixedVectorType::get(Type::getInt8Ty(RetTy->getContext()), 8);
        return 1 +
               getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind,
                                  -1, nullptr, nullptr) *
                   2 +
               getVectorInstrCost(Instruction::InsertElement, VecTy, CostKind,
                                  -1, nullptr, nullptr);
      }
    }
 
    if (LT.second.SimpleTy == MVT::nxv2i64)
      if (ST->hasSVEAES() && (ST->isSVEAvailable() || ST->hasSSVE_AES()))
        return LT.first * 3;
 
    if (ST->hasSVE2() || ST->hasSME()) {
      switch (LT.second.SimpleTy) {
      case MVT::nxv16i8:
        return LT.first;
      case MVT::nxv8i16:
        return LT.first * 6;
      case MVT::nxv4i32:
        return LT.first * 3;
      case MVT::nxv2i64:
        return LT.first * 8;
      default:
        break;
      }
    }
 
    // Avoid +sve giving this cost 2 due to custom lowering: It's very slow
    if (LT.second.SimpleTy == MVT::nxv2i64)
      return 192;
 
    if (ST->hasAES()) {
      switch (LT.second.SimpleTy) {
      case MVT::i16:
      case MVT::i32:
      case MVT::i64:
      case MVT::i128: {
        auto *VecTy =
            FixedVectorType::get(Type::getInt64Ty(RetTy->getContext()), 1);
        return LT.first *
               (1 +
                getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind,
                                   -1, nullptr, nullptr) *
                    2 +
                getVectorInstrCost(Instruction::InsertElement, VecTy, CostKind,
                                   -1, nullptr, nullptr));
      }
      case MVT::v1i64:
        return LT.first;
      case MVT::v2i64:
        return LT.first * 3;
      case MVT::v2i32:
        return LT.first * 6;
      case MVT::v4i32:
        return LT.first * 11;
      case MVT::v4i16:
        return LT.first * 14;
      default:
        break;
      }
    }
    break;
  }
  case Intrinsic::smulh:
  case Intrinsic::umulh: {
    InstructionCost MulCost =
        getArithmeticInstrCost(Instruction::Mul, RetTy, CostKind);
 
    // Wide types like i128 need every partial product of a wide multiply
    // and the carry chain between them. This is a crude cost approximation.
    if (RetTy->getScalarSizeInBits() > 64)
      return 2 * MulCost;
 
    auto LT = getTypeLegalizationCost(RetTy);
    MVT MTy = LT.second;
 
    // Per-register cost to extend the operands.
    InstructionCost ExtraCost =
        LT.first *
        (MTy.getScalarSizeInBits() != RetTy->getScalarSizeInBits() ? 2 : 0);
 
    // MULH and MUL lower similarly for 64-bit products. For vectors, they
    // also share the reduced vector multiply bandwidth seen on some cores.
    if (MTy.getScalarSizeInBits() == 64)
      return MulCost + ExtraCost;
 
    // Scalable vectors (and fixed-length vectors, when possible) use SVE.
    if (MTy.isScalableVector() ||
        (MTy.isFixedLengthVector() && ST->isSVEorStreamingSVEAvailable()))
      return LT.first + ExtraCost;
 
    static const CostTblEntry MulHighCostTbl[] = {
        {ISD::MULHU, MVT::i32, 2}, // [su]mull + lsr
 
        {ISD::MULHU, MVT::v8i8, 2},  // [su]mull + shrn
        {ISD::MULHU, MVT::v4i16, 2}, // "
        {ISD::MULHU, MVT::v2i32, 2}, // "
 
        {ISD::MULHU, MVT::v16i8, 3}, // [su]mull + [su]mull2 + uzp2
        {ISD::MULHU, MVT::v8i16, 3}, // "
        {ISD::MULHU, MVT::v4i32, 3}, // "
    };
 
    // Currently we assume SMULH has the same cost as UMULH.
    if (const auto *Entry = CostTableLookup(MulHighCostTbl, ISD::MULHU, MTy))
      return LT.first * Entry->Cost + ExtraCost;
 
    break;
  }
  case Intrinsic::umin:
  case Intrinsic::umax:
  case Intrinsic::smin:
  case Intrinsic::smax: {
    static const auto ValidMinMaxTys = {MVT::v8i8,  MVT::v16i8, MVT::v4i16,
                                        MVT::v8i16, MVT::v2i32, MVT::v4i32,
                                        MVT::nxv16i8, MVT::nxv8i16, MVT::nxv4i32,
                                        MVT::nxv2i64};
    auto LT = getTypeLegalizationCost(RetTy);
    // Type promotion for v2i8 and v2i16 types have a heavy cost when
    // vectorising. Account for this cost to avoid vectorising unprofitable
    // examples when vectorising loops with low trip counts.
    bool IsSigned =
        ICA.getID() == Intrinsic::smin || ICA.getID() == Intrinsic::smax;
    EVT VT = TLI->getValueType(DL, RetTy, /*AllowUnknown=*/true);
    if (VT == MVT::v2i8 || VT == MVT::v2i16 || VT == MVT::v4i8)
      return LT.first * (IsSigned ? 5 : 3);
    // v2i64 types get converted to cmp+bif hence the cost of 2
    if (LT.second == MVT::v2i64)
      return LT.first * 2;
    if (any_of(ValidMinMaxTys, equal_to(LT.second)))
      return LT.first;
    break;
  }
  case Intrinsic::scmp:
  case Intrinsic::ucmp: {
    static const CostTblEntry BitreverseTbl[] = {
        {Intrinsic::scmp, MVT::i32, 3},   // cmp+cset+csinv
        {Intrinsic::scmp, MVT::i64, 3},   // cmp+cset+csinv
        {Intrinsic::scmp, MVT::v8i8, 3},  // cmgt+cmgt+sub
        {Intrinsic::scmp, MVT::v16i8, 3}, // cmgt+cmgt+sub
        {Intrinsic::scmp, MVT::v4i16, 3}, // cmgt+cmgt+sub
        {Intrinsic::scmp, MVT::v8i16, 3}, // cmgt+cmgt+sub
        {Intrinsic::scmp, MVT::v2i32, 3}, // cmgt+cmgt+sub
        {Intrinsic::scmp, MVT::v4i32, 3}, // cmgt+cmgt+sub
        {Intrinsic::scmp, MVT::v1i64, 3}, // cmgt+cmgt+sub
        {Intrinsic::scmp, MVT::v2i64, 3}, // cmgt+cmgt+sub
    };
    const auto LT = getTypeLegalizationCost(RetTy);
    const auto *Entry =
        CostTableLookup(BitreverseTbl, Intrinsic::scmp, LT.second);
    if (Entry)
      return Entry->Cost * LT.first;
    break;
  }
  case Intrinsic::sadd_sat:
  case Intrinsic::ssub_sat:
  case Intrinsic::uadd_sat:
  case Intrinsic::usub_sat: {
    static const auto ValidSatTys = {MVT::v8i8,  MVT::v16i8, MVT::v4i16,
                                     MVT::v8i16, MVT::v2i32, MVT::v4i32,
                                     MVT::v2i64};
    auto LT = getTypeLegalizationCost(RetTy);
    // This is a base cost of 1 for the vadd, plus 3 extract shifts if we
    // need to extend the type, as it uses shr(qadd(shl, shl)).
    unsigned Instrs =
        LT.second.getScalarSizeInBits() == RetTy->getScalarSizeInBits() ? 1 : 4;
    if (any_of(ValidSatTys, equal_to(LT.second)))
      return LT.first * Instrs;
 
    TypeSize TS = getDataLayout().getTypeSizeInBits(RetTy);
    uint64_t VectorSize = TS.getKnownMinValue();
 
    if (ST->isSVEAvailable() && VectorSize >= 128 && isPowerOf2_64(VectorSize))
      return LT.first * Instrs;
 
    break;
  }
  case Intrinsic::abs: {
    static const auto ValidAbsTys = {MVT::v8i8,    MVT::v16i8,   MVT::v4i16,
                                     MVT::v8i16,   MVT::v2i32,   MVT::v4i32,
                                     MVT::v2i64,   MVT::nxv16i8, MVT::nxv8i16,
                                     MVT::nxv4i32, MVT::nxv2i64};
    auto LT = getTypeLegalizationCost(RetTy);
    if (any_of(ValidAbsTys, equal_to(LT.second)))
      return LT.first;
    break;
  }
  case Intrinsic::bswap: {
    static const auto ValidAbsTys = {MVT::v4i16, MVT::v8i16, MVT::v2i32,
                                     MVT::v4i32, MVT::v2i64};
    auto LT = getTypeLegalizationCost(RetTy);
    if (any_of(ValidAbsTys, equal_to(LT.second)) &&
        LT.second.getScalarSizeInBits() == RetTy->getScalarSizeInBits())
      return LT.first;
    break;
  }
  case Intrinsic::fma:
  case Intrinsic::fmuladd: {
    // Given a fma or fmuladd, cost it the same as a fmul instruction which are
    // usually the same for costs. TODO: Add fp16 and bf16 expansion costs.
    Type *EltTy = RetTy->getScalarType();
    if (EltTy->isFloatTy() || EltTy->isDoubleTy() ||
        (EltTy->isHalfTy() && ST->hasFullFP16()))
      return getArithmeticInstrCost(Instruction::FMul, RetTy, CostKind);
    break;
  }
  case Intrinsic::stepvector: {
    InstructionCost Cost = 1; // Cost of the `index' instruction
    auto LT = getTypeLegalizationCost(RetTy);
    // Legalisation of illegal vectors involves an `index' instruction plus
    // (LT.first - 1) vector adds.
    if (LT.first > 1) {
      Type *LegalVTy = EVT(LT.second).getTypeForEVT(RetTy->getContext());
      InstructionCost AddCost =
          getArithmeticInstrCost(Instruction::Add, LegalVTy, CostKind);
      Cost += AddCost * (LT.first - 1);
    }
    return Cost;
  }
  case Intrinsic::vector_extract:
  case Intrinsic::vector_insert: {
    // If both the vector and subvector types are legal types and the index
    // is 0, then this should be a no-op or simple operation; return a
    // relatively low cost.
 
    // If arguments aren't actually supplied, then we cannot determine the
    // value of the index. We also want to skip predicate types.
    if (ICA.getArgs().size() != ICA.getArgTypes().size() ||
        ICA.getReturnType()->getScalarType()->isIntegerTy(1))
      break;
 
    LLVMContext &C = RetTy->getContext();
    EVT VecVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
    bool IsExtract = ICA.getID() == Intrinsic::vector_extract;
    EVT SubVecVT = IsExtract ? getTLI()->getValueType(DL, RetTy)
                             : getTLI()->getValueType(DL, ICA.getArgTypes()[1]);
    // Skip this if either the vector or subvector types are unpacked
    // SVE types; they may get lowered to stack stores and loads.
    if (isUnpackedVectorVT(VecVT) || isUnpackedVectorVT(SubVecVT))
      break;
 
    TargetLoweringBase::LegalizeKind SubVecLK =
        getTLI()->getTypeConversion(C, SubVecVT);
    TargetLoweringBase::LegalizeKind VecLK =
        getTLI()->getTypeConversion(C, VecVT);
    const Value *Idx = IsExtract ? ICA.getArgs()[1] : ICA.getArgs()[2];
    const ConstantInt *CIdx = cast<ConstantInt>(Idx);
    if (SubVecLK.first == TargetLoweringBase::TypeLegal &&
        VecLK.first == TargetLoweringBase::TypeLegal && CIdx->isZero())
      return TTI::TCC_Free;
    break;
  }
  case Intrinsic::bitreverse: {
    static const CostTblEntry BitreverseTbl[] = {
        {Intrinsic::bitreverse, MVT::i32, 1},
        {Intrinsic::bitreverse, MVT::i64, 1},
        {Intrinsic::bitreverse, MVT::v8i8, 1},
        {Intrinsic::bitreverse, MVT::v16i8, 1},
        {Intrinsic::bitreverse, MVT::v4i16, 2},
        {Intrinsic::bitreverse, MVT::v8i16, 2},
        {Intrinsic::bitreverse, MVT::v2i32, 2},
        {Intrinsic::bitreverse, MVT::v4i32, 2},
        {Intrinsic::bitreverse, MVT::v1i64, 2},
        {Intrinsic::bitreverse, MVT::v2i64, 2},
    };
    const auto LegalisationCost = getTypeLegalizationCost(RetTy);
    const auto *Entry =
        CostTableLookup(BitreverseTbl, ICA.getID(), LegalisationCost.second);
    if (Entry) {
      // Cost Model is using the legal type(i32) that i8 and i16 will be
      // converted to +1 so that we match the actual lowering cost
      if (TLI->getValueType(DL, RetTy, true) == MVT::i8 ||
          TLI->getValueType(DL, RetTy, true) == MVT::i16)
        return LegalisationCost.first * Entry->Cost + 1;
 
      return LegalisationCost.first * Entry->Cost;
    }
    break;
  }
  case Intrinsic::pdep:
  case Intrinsic::pext: {
    unsigned BW = RetTy->getScalarSizeInBits();
    if (!RetTy->isVectorTy() && ST->isSVEBitPermAvailable() && BW <= 64) {
      Type *VecTy = VectorType::get(RetTy, 64 / BW, false);
      InstructionCost Cost =
          getVectorInstrCost(Instruction::InsertElement, VecTy, CostKind, 0,
                             nullptr, nullptr) +
          getVectorInstrCost(Instruction::InsertElement, VecTy, CostKind, 0,
                             nullptr, nullptr) +
          1 +
          getVectorInstrCost(Instruction::ExtractElement, VecTy, CostKind, 0,
                             nullptr, nullptr);
      return Cost;
    }
    break;
  }
  case Intrinsic::ctpop: {
    auto LT = getTypeLegalizationCost(RetTy);
    MVT MTy = LT.second;
 
    if (ST->hasCSSC() && !RetTy->isVectorTy()) {
      int ExtraCost =
          MTy.getScalarSizeInBits() != RetTy->getScalarSizeInBits() ? 1 : 0;
      return LT.first + ExtraCost;
    }
    if (!ST->hasNEON()) {
      // 32-bit or 64-bit ctpop without NEON is 12 instructions.
      return getTypeLegalizationCost(RetTy).first * 12;
    }
    static const CostTblEntry CtpopCostTbl[] = {
        {ISD::CTPOP, MVT::v2i64, 4},
        {ISD::CTPOP, MVT::v4i32, 3},
        {ISD::CTPOP, MVT::v8i16, 2},
        {ISD::CTPOP, MVT::v16i8, 1},
        {ISD::CTPOP, MVT::i64, 4},
        {ISD::CTPOP, MVT::v2i32, 3},
        {ISD::CTPOP, MVT::v4i16, 2},
        {ISD::CTPOP, MVT::v8i8, 1},
        {ISD::CTPOP, MVT::i32, 5},
        // SVE types (For targets that override NEON for fixed length vectors)
        {ISD::CTPOP, MVT::nxv2i64, 1},
        {ISD::CTPOP, MVT::nxv4i32, 1},
        {ISD::CTPOP, MVT::nxv8i16, 1},
        {ISD::CTPOP, MVT::nxv16i8, 1},
    };
 
    // When SVE is available CNT will be used for fixed and scalable vectors.
    if (ST->isSVEorStreamingSVEAvailable() && MTy.isFixedLengthVector())
      MTy = MVT::getScalableVectorVT(MTy.getVectorElementType(),
                                     128 / MTy.getScalarSizeInBits());
 
    if (const auto *Entry = CostTableLookup(CtpopCostTbl, ISD::CTPOP, MTy)) {
      // Extra cost of +1 when illegal vector types are legalized by promoting
      // the integer type.
      int ExtraCost = MTy.isVector() && MTy.getScalarSizeInBits() !=
                                            RetTy->getScalarSizeInBits()
                          ? 1
                          : 0;
      return LT.first * Entry->Cost + ExtraCost;
    }
    break;
  }
  case Intrinsic::sadd_with_overflow:
  case Intrinsic::uadd_with_overflow:
  case Intrinsic::ssub_with_overflow:
  case Intrinsic::usub_with_overflow:
  case Intrinsic::smul_with_overflow:
  case Intrinsic::umul_with_overflow: {
    static const CostTblEntry WithOverflowCostTbl[] = {
        {Intrinsic::sadd_with_overflow, MVT::i8, 3},
        {Intrinsic::uadd_with_overflow, MVT::i8, 3},
        {Intrinsic::sadd_with_overflow, MVT::i16, 3},
        {Intrinsic::uadd_with_overflow, MVT::i16, 3},
        {Intrinsic::sadd_with_overflow, MVT::i32, 1},
        {Intrinsic::uadd_with_overflow, MVT::i32, 1},
        {Intrinsic::sadd_with_overflow, MVT::i64, 1},
        {Intrinsic::uadd_with_overflow, MVT::i64, 1},
        {Intrinsic::ssub_with_overflow, MVT::i8, 3},
        {Intrinsic::usub_with_overflow, MVT::i8, 3},
        {Intrinsic::ssub_with_overflow, MVT::i16, 3},
        {Intrinsic::usub_with_overflow, MVT::i16, 3},
        {Intrinsic::ssub_with_overflow, MVT::i32, 1},
        {Intrinsic::usub_with_overflow, MVT::i32, 1},
        {Intrinsic::ssub_with_overflow, MVT::i64, 1},
        {Intrinsic::usub_with_overflow, MVT::i64, 1},
        {Intrinsic::smul_with_overflow, MVT::i8, 5},
        {Intrinsic::umul_with_overflow, MVT::i8, 4},
        {Intrinsic::smul_with_overflow, MVT::i16, 5},
        {Intrinsic::umul_with_overflow, MVT::i16, 4},
        {Intrinsic::smul_with_overflow, MVT::i32, 2}, // eg umull;tst
        {Intrinsic::umul_with_overflow, MVT::i32, 2}, // eg umull;cmp sxtw
        {Intrinsic::smul_with_overflow, MVT::i64, 3}, // eg mul;smulh;cmp
        {Intrinsic::umul_with_overflow, MVT::i64, 3}, // eg mul;umulh;cmp asr
    };
    EVT MTy = TLI->getValueType(DL, RetTy->getContainedType(0), true);
    if (MTy.isSimple())
      if (const auto *Entry = CostTableLookup(WithOverflowCostTbl, ICA.getID(),
                                              MTy.getSimpleVT()))
        return Entry->Cost;
    break;
  }
  case Intrinsic::fptosi_sat:
  case Intrinsic::fptoui_sat: {
    if (ICA.getArgTypes().empty())
      break;
    bool IsSigned = ICA.getID() == Intrinsic::fptosi_sat;
    auto LT = getTypeLegalizationCost(ICA.getArgTypes()[0]);
    EVT MTy = TLI->getValueType(DL, RetTy);
    // Check for the legal types, which are where the size of the input and the
    // output are the same, or we are using cvt f64->i32 or f32->i64.
    if ((LT.second == MVT::f32 || LT.second == MVT::f64 ||
         LT.second == MVT::v2f32 || LT.second == MVT::v4f32 ||
         LT.second == MVT::v2f64)) {
      if ((LT.second.getScalarSizeInBits() == MTy.getScalarSizeInBits() ||
           (LT.second == MVT::f64 && MTy == MVT::i32) ||
           (LT.second == MVT::f32 && MTy == MVT::i64)))
        return LT.first;
      // Extending vector types v2f32->v2i64, fcvtl*2 + fcvt*2
      if (LT.second.getScalarType() == MVT::f32 && MTy.isFixedLengthVector() &&
          MTy.getScalarSizeInBits() == 64)
        return LT.first * (MTy.getVectorNumElements() > 2 ? 4 : 2);
    }
    // Similarly for fp16 sizes. Without FullFP16 we generally need to fcvt to
    // f32.
    if (LT.second.getScalarType() == MVT::f16 && !ST->hasFullFP16())
      return LT.first + getIntrinsicInstrCost(
                            {ICA.getID(),
                             RetTy,
                             {ICA.getArgTypes()[0]->getWithNewType(
                                 Type::getFloatTy(RetTy->getContext()))}},
                            CostKind);
    if ((LT.second == MVT::f16 && MTy == MVT::i32) ||
        (LT.second == MVT::f16 && MTy == MVT::i64) ||
        ((LT.second == MVT::v4f16 || LT.second == MVT::v8f16) &&
         (LT.second.getScalarSizeInBits() == MTy.getScalarSizeInBits())))
      return LT.first;
    // Extending vector types v8f16->v8i32, fcvtl*2 + fcvt*2
    if (LT.second.getScalarType() == MVT::f16 && MTy.isFixedLengthVector() &&
        MTy.getScalarSizeInBits() == 32)
      return LT.first * (MTy.getVectorNumElements() > 4 ? 4 : 2);
    // Extending vector types v8f16->v8i32. These current scalarize but the
    // codegen could be better.
    if (LT.second.getScalarType() == MVT::f16 && MTy.isFixedLengthVector() &&
        MTy.getScalarSizeInBits() == 64)
      return MTy.getVectorNumElements() * 3;
 
    // If we can we use a legal convert followed by a min+max
    if ((LT.second.getScalarType() == MVT::f32 ||
         LT.second.getScalarType() == MVT::f64 ||
         LT.second.getScalarType() == MVT::f16) &&
        LT.second.getScalarSizeInBits() >= MTy.getScalarSizeInBits()) {
      Type *LegalTy =
          Type::getIntNTy(RetTy->getContext(), LT.second.getScalarSizeInBits());
      if (LT.second.isVector())
        LegalTy = VectorType::get(LegalTy, LT.second.getVectorElementCount());
      InstructionCost Cost = 1;
      IntrinsicCostAttributes Attrs1(IsSigned ? Intrinsic::smin
                                              : Intrinsic::umin,
                                     LegalTy, {LegalTy, LegalTy});
      Cost += getIntrinsicInstrCost(Attrs1, CostKind);
      IntrinsicCostAttributes Attrs2(IsSigned ? Intrinsic::smax
                                              : Intrinsic::umax,
                                     LegalTy, {LegalTy, LegalTy});
      Cost += getIntrinsicInstrCost(Attrs2, CostKind);
      return LT.first * Cost +
             ((LT.second.getScalarType() != MVT::f16 || ST->hasFullFP16()) ? 0
                                                                           : 1);
    }
    // Otherwise we need to follow the default expansion that clamps the value
    // using a float min/max with a fcmp+sel for nan handling when signed.
    Type *FPTy = ICA.getArgTypes()[0]->getScalarType();
    RetTy = RetTy->getScalarType();
    if (LT.second.isVector()) {
      FPTy = VectorType::get(FPTy, LT.second.getVectorElementCount());
      RetTy = VectorType::get(RetTy, LT.second.getVectorElementCount());
    }
    IntrinsicCostAttributes Attrs1(Intrinsic::minnum, FPTy, {FPTy, FPTy});
    InstructionCost Cost = getIntrinsicInstrCost(Attrs1, CostKind);
    IntrinsicCostAttributes Attrs2(Intrinsic::maxnum, FPTy, {FPTy, FPTy});
    Cost += getIntrinsicInstrCost(Attrs2, CostKind);
    Cost +=
        getCastInstrCost(IsSigned ? Instruction::FPToSI : Instruction::FPToUI,
                         RetTy, FPTy, TTI::CastContextHint::None, CostKind);
    if (IsSigned) {
      Type *CondTy = RetTy->getWithNewBitWidth(1);
      Cost += getCmpSelInstrCost(BinaryOperator::FCmp, FPTy, CondTy,
                                 CmpInst::FCMP_UNO, CostKind);
      Cost += getCmpSelInstrCost(BinaryOperator::Select, RetTy, CondTy,
                                 CmpInst::FCMP_UNO, CostKind);
    }
    return LT.first * Cost;
  }
  case Intrinsic::fshl:
  case Intrinsic::fshr: {
    if (ICA.getArgs().empty())
      break;
 
    const TTI::OperandValueInfo OpInfoZ = TTI::getOperandInfo(ICA.getArgs()[2]);
 
    // ROTR / ROTL is a funnel shift with equal first and second operand. For
    // ROTR on integer registers (i32/i64) this can be done in a single ror
    // instruction. A fshl with a non-constant shift uses a neg + ror.
    if (RetTy->isIntegerTy() && ICA.getArgs()[0] == ICA.getArgs()[1] &&
        (RetTy->getPrimitiveSizeInBits() == 32 ||
         RetTy->getPrimitiveSizeInBits() == 64)) {
      InstructionCost NegCost =
          (ICA.getID() == Intrinsic::fshl && !OpInfoZ.isConstant()) ? 1 : 0;
      return 1 + NegCost;
    }
 
    // TODO: Add handling for fshl where third argument is not a constant.
    if (!OpInfoZ.isConstant())
      break;
 
    const auto LegalisationCost = getTypeLegalizationCost(RetTy);
    if (OpInfoZ.isUniform()) {
      static const CostTblEntry FshlTbl[] = {
          {Intrinsic::fshl, MVT::v4i32, 2}, // shl + usra
          {Intrinsic::fshl, MVT::v2i64, 2}, {Intrinsic::fshl, MVT::v16i8, 2},
          {Intrinsic::fshl, MVT::v8i16, 2}, {Intrinsic::fshl, MVT::v2i32, 2},
          {Intrinsic::fshl, MVT::v8i8, 2},  {Intrinsic::fshl, MVT::v4i16, 2}};
      // Costs for both fshl & fshr are the same, so just pass Intrinsic::fshl
      // to avoid having to duplicate the costs.
      const auto *Entry =
          CostTableLookup(FshlTbl, Intrinsic::fshl, LegalisationCost.second);
      if (Entry)
        return LegalisationCost.first * Entry->Cost;
    }
 
    auto TyL = getTypeLegalizationCost(RetTy);
    if (!RetTy->isIntegerTy())
      break;
 
    // Estimate cost manually, as types like i8 and i16 will get promoted to
    // i32 and CostTableLookup will ignore the extra conversion cost.
    bool HigherCost = (RetTy->getScalarSizeInBits() != 32 &&
                       RetTy->getScalarSizeInBits() < 64) ||
                      (RetTy->getScalarSizeInBits() % 64 != 0);
    unsigned ExtraCost = HigherCost ? 1 : 0;
    if (RetTy->getScalarSizeInBits() == 32 ||
        RetTy->getScalarSizeInBits() == 64)
      ExtraCost = 0; // fhsl/fshr for i32 and i64 can be lowered to a single
                     // extr instruction.
    else if (HigherCost)
      ExtraCost = 1;
    else
      break;
    return TyL.first + ExtraCost;
  }
  case Intrinsic::get_active_lane_mask: {
    auto RetTy = cast<VectorType>(ICA.getReturnType());
    EVT RetVT = getTLI()->getValueType(DL, RetTy);
    EVT OpVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
    if (getTLI()->shouldExpandGetActiveLaneMask(RetVT, OpVT))
      break;
 
    if (RetTy->isScalableTy()) {
      if (TLI->getTypeAction(RetTy->getContext(), RetVT) !=
          TargetLowering::TypeSplitVector)
        break;
 
      auto LT = getTypeLegalizationCost(RetTy);
      InstructionCost Cost = LT.first;
      // When SVE2p1 or SME2 is available, we can halve getTypeLegalizationCost
      // as get_active_lane_mask may lower to the sve_whilelo_x2 intrinsic, e.g.
      //   nxv32i1 = get_active_lane_mask(base, idx) ->
      //    {nxv16i1, nxv16i1} = sve_whilelo_x2(base, idx)
      if (ST->hasSVE2p1() || ST->hasSME2()) {
        Cost /= 2;
        if (Cost == 1)
          return Cost;
      }
 
      // If more than one whilelo intrinsic is required, include the extra cost
      // required by the saturating add & select required to increment the
      // start value after the first intrinsic call.
      Type *OpTy = ICA.getArgTypes()[0];
      IntrinsicCostAttributes AddAttrs(Intrinsic::uadd_sat, OpTy, {OpTy, OpTy});
      InstructionCost SplitCost = getIntrinsicInstrCost(AddAttrs, CostKind);
      Type *CondTy = OpTy->getWithNewBitWidth(1);
      SplitCost += getCmpSelInstrCost(Instruction::Select, OpTy, CondTy,
                                      CmpInst::ICMP_UGT, CostKind);
      return Cost + (SplitCost * (Cost - 1));
    } else if (!getTLI()->isTypeLegal(RetVT)) {
      // We don't have enough context at this point to determine if the mask
      // is going to be kept live after the block, which will force the vXi1
      // type to be expanded to legal vectors of integers, e.g. v4i1->v4i32.
      // For now, we just assume the vectorizer created this intrinsic and
      // the result will be the input for a PHI. In this case the cost will
      // be extremely high for fixed-width vectors.
      // NOTE: getScalarizationOverhead returns a cost that's far too
      // pessimistic for the actual generated codegen. In reality there are
      // two instructions generated per lane.
      return cast<FixedVectorType>(RetTy)->getNumElements() * 2;
    }
    break;
  }
  case Intrinsic::experimental_vector_match: {
    if (!ST->hasSVE2() || !ST->isSVEAvailable())
      break;
 
    auto *NeedleTy = cast<FixedVectorType>(ICA.getArgTypes()[1]);
 
    // We expand vector.matches with <= 2 elements to a chain of compares.
    unsigned SearchSize = NeedleTy->getNumElements();
    if (SearchSize <= 2)
      break;
 
    auto [LegalParts, SearchVT] = getTypeLegalizationCost(ICA.getArgTypes()[0]);
    if (!is_contained(
            {MVT::nxv8i16, MVT::nxv16i8, MVT::v8i16, MVT::v16i8, MVT::v8i8},
            SearchVT.SimpleTy))
      break;
 
    unsigned ElementSizeInBits = SearchVT.getScalarSizeInBits();
 
    // Number of needle elements we can compare per `match` instruction.
    unsigned NeedleEltsPerMatch = AArch64::SVEBitsPerBlock / ElementSizeInBits;
 
    // How many `match` instructions we need to match `SearchSize` elements.
    unsigned MatchesRequiredForNeedle =
        llvm::divideCeil(SearchSize, NeedleEltsPerMatch);
 
    // Base cost for MATCH instructions. At least on the Neoverse V2 and
    // Neoverse V3, these are cheap operations with the same latency as a
    // vector ADD. In most cases, however, we also need to do an extra DUP.
    // For fixed-length vectors we currently need an extra five--six
    // instructions besides the MATCH.
    InstructionCost Cost = 4;
    if (isa<FixedVectorType>(RetTy))
      Cost += 10;
 
    return Cost * LegalParts * MatchesRequiredForNeedle;
  }
  case Intrinsic::cttz: {
    auto LT = getTypeLegalizationCost(ICA.getArgTypes()[0]);
    if (LT.second == MVT::v8i8 || LT.second == MVT::v16i8)
      return LT.first * 2;
    if (LT.second == MVT::v4i16 || LT.second == MVT::v8i16 ||
        LT.second == MVT::v2i32 || LT.second == MVT::v4i32)
      return LT.first * 3;
    break;
  }
  case Intrinsic::experimental_cttz_elts: {
    EVT ArgVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
    if (getTLI()->isOperationCustom(ISD::CTTZ_ELTS, ArgVT)) {
      // This will consist of a SVE brkb and a cntp instruction. These
      // typically have the same latency and half the throughput as a vector
      // add instruction.
      return 4;
    }
    break;
  }
  case Intrinsic::mask_beforefirst: {
    if (!ST->isSVEorStreamingSVEAvailable() ||
        !isa<ScalableVectorType>(RetTy) ||
        !TLI->isTypeLegal(getTLI()->getValueType(DL, RetTy)))
      break;
 
    // mask_beforefirst(mask) -> brkb_z(ptrue, mask) but we ignore ptrue
    // because it's typically used by other instructions in the function.
    return 1;
  }
  case Intrinsic::loop_dependence_raw_mask:
  case Intrinsic::loop_dependence_war_mask: {
    // The whilewr/rw instructions require SVE2 or SME.
    if (ST->hasSVE2() || ST->hasSME()) {
      EVT VecVT = getTLI()->getValueType(DL, RetTy);
      unsigned EltSizeInBytes =
          cast<ConstantInt>(ICA.getArgs()[2])->getZExtValue();
      if (!is_contained({1u, 2u, 4u, 8u}, EltSizeInBytes) ||
          VecVT.getVectorMinNumElements() != (16 / EltSizeInBytes))
        break;
      // For fixed-vector types we need to AND the mask with a ptrue vl<N>.
      return isa<FixedVectorType>(RetTy) ? 2 : 1;
    }
    break;
  }
  case Intrinsic::experimental_vector_extract_last_active:
    if (ST->isSVEorStreamingSVEAvailable()) {
      auto [LegalCost, _] = getTypeLegalizationCost(ICA.getArgTypes()[0]);
      // This should turn into chained clastb instructions.
      return LegalCost;
    }
    break;
  case Intrinsic::pow: {
    // For scalar calls we know the target has the libcall, and for fixed-width
    // vectors we know for the worst case it can be scalarised.
    EVT VT = getTLI()->getValueType(DL, RetTy);
    RTLIB::Libcall LC = RTLIB::getPOW(VT);
    bool HasLibcall =
        getTLI()->getLibcallLoweringInfo().getVectorLibcallImplOrMasked(LC) !=
        RTLIB::Unsupported;
    bool CanLowerWithLibcalls = !isa<ScalableVectorType>(RetTy) || HasLibcall;
 
    // If we know that the call can be lowered with libcalls then it's safe to
    // reduce the costs in some cases. This is important for scalable vectors,
    // since we cannot scalarize the call in the absence of a vector math
    // library.
    if (CanLowerWithLibcalls && ICA.getInst() && !ICA.getArgs().empty()) {
      // If we know the fast math flags and the exponent is a constant then the
      // cost may be less for some exponents like 0.25 and 0.75.
      const Constant *ExpC = dyn_cast<Constant>(ICA.getArgs()[1]);
      if (ExpC && isa<VectorType>(ExpC->getType()))
        ExpC = ExpC->getSplatValue();
      if (auto *ExpF = dyn_cast_or_null<ConstantFP>(ExpC)) {
        // The argument must be a FP constant.
        bool Is025 = ExpF->getValueAPF().isExactlyValue(0.25);
        bool Is075 = ExpF->getValueAPF().isExactlyValue(0.75);
        FastMathFlags FMF = ICA.getInst()->getFastMathFlags();
        if ((Is025 || Is075) && FMF.noInfs() && FMF.approxFunc() &&
            (!Is025 || FMF.noSignedZeros())) {
          IntrinsicCostAttributes Attrs(Intrinsic::sqrt, RetTy, {RetTy}, FMF);
          InstructionCost Sqrt = getIntrinsicInstrCost(Attrs, CostKind);
          if (Is025)
            return 2 * Sqrt;
          InstructionCost FMul =
              getArithmeticInstrCost(Instruction::FMul, RetTy, CostKind);
          return (Sqrt * 2) + FMul;
        }
        // TODO: For 1/3 exponents we expect the cbrt call to be slightly
        // cheaper than pow.
      }
    }
 
    if (HasLibcall)
      return getCallInstrCost(nullptr, RetTy, ICA.getArgTypes(), CostKind);
    break;
  }
  case Intrinsic::sqrt:
  case Intrinsic::fabs:
  case Intrinsic::ceil:
  case Intrinsic::floor:
  case Intrinsic::nearbyint:
  case Intrinsic::round:
  case Intrinsic::rint:
  case Intrinsic::roundeven:
  case Intrinsic::trunc:
  case Intrinsic::minnum:
  case Intrinsic::maxnum:
  case Intrinsic::minimum:
  case Intrinsic::maximum: {
    if (isa<ScalableVectorType>(RetTy) && ST->isSVEorStreamingSVEAvailable()) {
      auto LT = getTypeLegalizationCost(RetTy);
      return LT.first;
    }
    break;
  }
  default:
    break;
  }
  return BaseT::getIntrinsicInstrCost(ICA, CostKind);
}
 
/// The function will remove redundant reinterprets casting in the presence
/// of the control flow
static std::optional<Instruction *> processPhiNode(InstCombiner &IC,
                                                   IntrinsicInst &II) {
  SmallVector<Instruction *, 32> Worklist;
  auto RequiredType = II.getType();
 
  auto *PN = dyn_cast<PHINode>(II.getArgOperand(0));
  assert(PN && "Expected Phi Node!");
 
  // Don't create a new Phi unless we can remove the old one.
  if (!PN->hasOneUse())
    return std::nullopt;
 
  for (Value *IncValPhi : PN->incoming_values()) {
    auto *Reinterpret = dyn_cast<IntrinsicInst>(IncValPhi);
    if (!Reinterpret ||
        Reinterpret->getIntrinsicID() !=
            Intrinsic::aarch64_sve_convert_to_svbool ||
        RequiredType != Reinterpret->getArgOperand(0)->getType())
      return std::nullopt;
  }
 
  // Create the new Phi
  IC.Builder.SetInsertPoint(PN);
  PHINode *NPN = IC.Builder.CreatePHI(RequiredType, PN->getNumIncomingValues());
  Worklist.push_back(PN);
 
  for (unsigned I = 0; I < PN->getNumIncomingValues(); I++) {
    auto *Reinterpret = cast<Instruction>(PN->getIncomingValue(I));
    NPN->addIncoming(Reinterpret->getOperand(0), PN->getIncomingBlock(I));
    Worklist.push_back(Reinterpret);
  }
 
  // Cleanup Phi Node and reinterprets
  return IC.replaceInstUsesWith(II, NPN);
}
 
// A collection of properties common to SVE intrinsics that allow for combines
// to be written without needing to know the specific intrinsic.
struct SVEIntrinsicInfo {
  //
  // Helper routines for common intrinsic definitions.
  //
 
  // e.g. llvm.aarch64.sve.add pg, op1, op2
  //        with IID ==> llvm.aarch64.sve.add_u
  static SVEIntrinsicInfo
  defaultMergingOp(Intrinsic::ID IID = Intrinsic::not_intrinsic) {
    return SVEIntrinsicInfo()
        .setGoverningPredicateOperandIdx(0)
        .setOperandIdxInactiveLanesTakenFrom(1)
        .setMatchingUndefIntrinsic(IID);
  }
 
  // e.g. llvm.aarch64.sve.neg inactive, pg, op
  static SVEIntrinsicInfo defaultMergingUnaryOp() {
    return SVEIntrinsicInfo()
        .setGoverningPredicateOperandIdx(1)
        .setOperandIdxInactiveLanesTakenFrom(0)
        .setOperandIdxWithNoActiveLanes(0);
  }
 
  // e.g. llvm.aarch64.sve.fcvtnt inactive, pg, op
  static SVEIntrinsicInfo defaultMergingUnaryNarrowingTopOp() {
    return SVEIntrinsicInfo()
        .setGoverningPredicateOperandIdx(1)
        .setOperandIdxInactiveLanesTakenFrom(0);
  }
 
  // e.g. llvm.aarch64.sve.add_u pg, op1, op2
  static SVEIntrinsicInfo defaultUndefOp() {
    return SVEIntrinsicInfo()
        .setGoverningPredicateOperandIdx(0)
        .setInactiveLanesAreNotDefined();
  }
 
  // e.g. llvm.aarch64.sve.prf pg, ptr        (GPIndex = 0)
  //      llvm.aarch64.sve.st1 data, pg, ptr  (GPIndex = 1)
  static SVEIntrinsicInfo defaultVoidOp(unsigned GPIndex) {
    return SVEIntrinsicInfo()
        .setGoverningPredicateOperandIdx(GPIndex)
        .setInactiveLanesAreUnused();
  }
 
  // e.g. llvm.aarch64.sve.cmpeq pg, op1, op2
  //      llvm.aarch64.sve.ld1 pg, ptr
  static SVEIntrinsicInfo defaultZeroingOp() {
    return SVEIntrinsicInfo()
        .setGoverningPredicateOperandIdx(0)
        .setInactiveLanesAreUnused()
        .setResultIsZeroInitialized();
  }
 
  // All properties relate to predication and thus having a general predicate
  // is the minimum requirement to say there is intrinsic info to act on.
  explicit operator bool() const { return hasGoverningPredicate(); }
 
  //
  // Properties relating to the governing predicate.
  //
 
  bool hasGoverningPredicate() const {
    return GoverningPredicateIdx != std::numeric_limits<unsigned>::max();
  }
 
  unsigned getGoverningPredicateOperandIdx() const {
    assert(hasGoverningPredicate() && "Property not set!");
    return GoverningPredicateIdx;
  }
 
  SVEIntrinsicInfo &setGoverningPredicateOperandIdx(unsigned Index) {
    assert(!hasGoverningPredicate() && "Cannot set property twice!");
    GoverningPredicateIdx = Index;
    return *this;
  }
 
  //
  // Properties relating to operations the intrinsic could be transformed into.
  // NOTE: This does not mean such a transformation is always possible, but the
  // knowledge makes it possible to reuse existing optimisations without needing
  // to embed specific handling for each intrinsic. For example, instruction
  // simplification can be used to optimise an intrinsic's active lanes.
  //
 
  //
  // Intrinsic that produces the same result for active lanes.
  //
 
  bool hasMatchingUndefIntrinsic() const {
    return UndefIntrinsic != Intrinsic::not_intrinsic;
  }
 
  Intrinsic::ID getMatchingUndefIntrinsic() const {
    assert(hasMatchingUndefIntrinsic() && "Property not set!");
    return UndefIntrinsic;
  }
 
  SVEIntrinsicInfo &setMatchingUndefIntrinsic(Intrinsic::ID IID) {
    assert(!hasMatchingUndefIntrinsic() && "Cannot set property twice!");
    UndefIntrinsic = IID;
    return *this;
  }
 
  //
  // Instruction where active lanes produce the same result.
  //
 
  bool hasMatchingIROpode() const { return IROpcode != 0; }
 
  unsigned getMatchingIROpode() const {
    assert(hasMatchingIROpode() && "Property not set!");
    return IROpcode;
  }
 
  SVEIntrinsicInfo &setMatchingIROpcode(unsigned Opcode) {
    assert(!hasMatchingIROpode() && "Cannot set property twice!");
    IROpcode = Opcode;
    return *this;
  }
 
  bool hasCmpPredicate() const {
    return CmpPredicate != CmpInst::BAD_ICMP_PREDICATE;
  }
 
  CmpInst::Predicate getCmpPredicate() const {
    assert(hasCmpPredicate() && "Property not set!");
    return CmpPredicate;
  }
 
  SVEIntrinsicInfo &setCmpPredicate(CmpInst::Predicate Pred) {
    assert(!hasCmpPredicate() && "Cannot set property twice!");
    CmpPredicate = Pred;
 
    if (CmpInst::isFPPredicate(Pred))
      return setMatchingIROpcode(Instruction::FCmp);
 
    if (CmpInst::isIntPredicate(Pred))
      return setMatchingIROpcode(Instruction::ICmp);
 
    llvm_unreachable("Unsupported compare predicate!");
  }
 
  //
  // Properties relating to the result of inactive lanes.
  //
 
  bool inactiveLanesTakenFromOperand() const {
    return ResultLanes == InactiveLanesTakenFromOperand;
  }
 
  unsigned getOperandIdxInactiveLanesTakenFrom() const {
    assert(inactiveLanesTakenFromOperand() && "Property not set!");
    return OperandIdxForInactiveLanes;
  }
 
  SVEIntrinsicInfo &setOperandIdxInactiveLanesTakenFrom(unsigned Index) {
    assert(ResultLanes == Uninitialized && "Cannot set property twice!");
    ResultLanes = InactiveLanesTakenFromOperand;
    OperandIdxForInactiveLanes = Index;
    return *this;
  }
 
  bool inactiveLanesAreNotDefined() const {
    return ResultLanes == InactiveLanesAreNotDefined;
  }
 
  SVEIntrinsicInfo &setInactiveLanesAreNotDefined() {
    assert(ResultLanes == Uninitialized && "Cannot set property twice!");
    ResultLanes = InactiveLanesAreNotDefined;
    return *this;
  }
 
  bool inactiveLanesAreUnused() const {
    return ResultLanes == InactiveLanesAreUnused;
  }
 
  SVEIntrinsicInfo &setInactiveLanesAreUnused() {
    assert(ResultLanes == Uninitialized && "Cannot set property twice!");
    ResultLanes = InactiveLanesAreUnused;
    return *this;
  }
 
  // NOTE: Whilst not limited to only inactive lanes, the common use case is:
  // inactiveLanesAreZeroed =
  //     resultIsZeroInitialized() && inactiveLanesAreUnused()
  bool resultIsZeroInitialized() const { return ResultIsZeroInitialized; }
 
  SVEIntrinsicInfo &setResultIsZeroInitialized() {
    ResultIsZeroInitialized = true;
    return *this;
  }
 
  //
  // The first operand of unary merging operations is typically only used to
  // set the result for inactive lanes. Knowing this allows us to deadcode the
  // operand when we can prove there are no inactive lanes.
  //
 
  bool hasOperandWithNoActiveLanes() const {
    return OperandIdxWithNoActiveLanes != std::numeric_limits<unsigned>::max();
  }
 
  unsigned getOperandIdxWithNoActiveLanes() const {
    assert(hasOperandWithNoActiveLanes() && "Property not set!");
    return OperandIdxWithNoActiveLanes;
  }
 
  SVEIntrinsicInfo &setOperandIdxWithNoActiveLanes(unsigned Index) {
    assert(!hasOperandWithNoActiveLanes() && "Cannot set property twice!");
    OperandIdxWithNoActiveLanes = Index;
    return *this;
  }
 
private:
  unsigned GoverningPredicateIdx = std::numeric_limits<unsigned>::max();
 
  Intrinsic::ID UndefIntrinsic = Intrinsic::not_intrinsic;
  unsigned IROpcode = 0;
  CmpInst::Predicate CmpPredicate = CmpInst::BAD_ICMP_PREDICATE;
 
  enum PredicationStyle {
    Uninitialized,
    InactiveLanesTakenFromOperand,
    InactiveLanesAreNotDefined,
    InactiveLanesAreUnused
  } ResultLanes = Uninitialized;
 
  bool ResultIsZeroInitialized = false;
  unsigned OperandIdxForInactiveLanes = std::numeric_limits<unsigned>::max();
  unsigned OperandIdxWithNoActiveLanes = std::numeric_limits<unsigned>::max();
};
 
static SVEIntrinsicInfo constructSVEIntrinsicInfo(IntrinsicInst &II) {
  // Some SVE intrinsics do not use scalable vector types, but since they are
  // not relevant from an SVEIntrinsicInfo perspective, they are also ignored.
  if (!isa<ScalableVectorType>(II.getType()) &&
      all_of(II.args(), [&](const Value *V) {
        return !isa<ScalableVectorType>(V->getType());
      }))
    return SVEIntrinsicInfo();
 
  Intrinsic::ID IID = II.getIntrinsicID();
  switch (IID) {
  default:
    break;
  case Intrinsic::aarch64_sve_fcvt_bf16f32_v2:
  case Intrinsic::aarch64_sve_fcvt_f16f32:
  case Intrinsic::aarch64_sve_fcvt_f16f64:
  case Intrinsic::aarch64_sve_fcvt_f32f16:
  case Intrinsic::aarch64_sve_fcvt_f32f64:
  case Intrinsic::aarch64_sve_fcvt_f64f16:
  case Intrinsic::aarch64_sve_fcvt_f64f32:
  case Intrinsic::aarch64_sve_fcvtlt_f32f16:
  case Intrinsic::aarch64_sve_fcvtlt_f64f32:
  case Intrinsic::aarch64_sve_fcvtx_f32f64:
  case Intrinsic::aarch64_sve_fcvtzs:
  case Intrinsic::aarch64_sve_fcvtzs_i32f16:
  case Intrinsic::aarch64_sve_fcvtzs_i32f64:
  case Intrinsic::aarch64_sve_fcvtzs_i64f16:
  case Intrinsic::aarch64_sve_fcvtzs_i64f32:
  case Intrinsic::aarch64_sve_fcvtzu:
  case Intrinsic::aarch64_sve_fcvtzu_i32f16:
  case Intrinsic::aarch64_sve_fcvtzu_i32f64:
  case Intrinsic::aarch64_sve_fcvtzu_i64f16:
  case Intrinsic::aarch64_sve_fcvtzu_i64f32:
  case Intrinsic::aarch64_sve_revb:
  case Intrinsic::aarch64_sve_revh:
  case Intrinsic::aarch64_sve_revw:
  case Intrinsic::aarch64_sve_revd:
  case Intrinsic::aarch64_sve_scvtf:
  case Intrinsic::aarch64_sve_scvtf_f16i32:
  case Intrinsic::aarch64_sve_scvtf_f16i64:
  case Intrinsic::aarch64_sve_scvtf_f32i64:
  case Intrinsic::aarch64_sve_scvtf_f64i32:
  case Intrinsic::aarch64_sve_ucvtf:
  case Intrinsic::aarch64_sve_ucvtf_f16i32:
  case Intrinsic::aarch64_sve_ucvtf_f16i64:
  case Intrinsic::aarch64_sve_ucvtf_f32i64:
  case Intrinsic::aarch64_sve_ucvtf_f64i32:
    return SVEIntrinsicInfo::defaultMergingUnaryOp();
 
  case Intrinsic::aarch64_sve_fcvtnt_bf16f32_v2:
  case Intrinsic::aarch64_sve_fcvtnt_f16f32:
  case Intrinsic::aarch64_sve_fcvtnt_f32f64:
  case Intrinsic::aarch64_sve_fcvtxnt_f32f64:
    return SVEIntrinsicInfo::defaultMergingUnaryNarrowingTopOp();
 
  case Intrinsic::aarch64_sve_fabd:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fabd_u);
  case Intrinsic::aarch64_sve_fadd:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fadd_u)
        .setMatchingIROpcode(Instruction::FAdd);
  case Intrinsic::aarch64_sve_fdiv:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fdiv_u)
        .setMatchingIROpcode(Instruction::FDiv);
  case Intrinsic::aarch64_sve_fmax:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmax_u);
  case Intrinsic::aarch64_sve_fmaxnm:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmaxnm_u);
  case Intrinsic::aarch64_sve_fmin:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmin_u);
  case Intrinsic::aarch64_sve_fminnm:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fminnm_u);
  case Intrinsic::aarch64_sve_fmla:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmla_u);
  case Intrinsic::aarch64_sve_fmls:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmls_u);
  case Intrinsic::aarch64_sve_fmul:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmul_u)
        .setMatchingIROpcode(Instruction::FMul);
  case Intrinsic::aarch64_sve_fmulx:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fmulx_u);
  case Intrinsic::aarch64_sve_fnmla:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fnmla_u);
  case Intrinsic::aarch64_sve_fnmls:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fnmls_u);
  case Intrinsic::aarch64_sve_fsub:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_fsub_u)
        .setMatchingIROpcode(Instruction::FSub);
  case Intrinsic::aarch64_sve_add:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_add_u)
        .setMatchingIROpcode(Instruction::Add);
  case Intrinsic::aarch64_sve_mla:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_mla_u);
  case Intrinsic::aarch64_sve_mls:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_mls_u);
  case Intrinsic::aarch64_sve_mul:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_mul_u)
        .setMatchingIROpcode(Instruction::Mul);
  case Intrinsic::aarch64_sve_sabd:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sabd_u);
  case Intrinsic::aarch64_sve_sdiv:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sdiv_u)
        .setMatchingIROpcode(Instruction::SDiv);
  case Intrinsic::aarch64_sve_smax:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_smax_u);
  case Intrinsic::aarch64_sve_smin:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_smin_u);
  case Intrinsic::aarch64_sve_smulh:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_smulh_u);
  case Intrinsic::aarch64_sve_sub:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sub_u)
        .setMatchingIROpcode(Instruction::Sub);
  case Intrinsic::aarch64_sve_uabd:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_uabd_u);
  case Intrinsic::aarch64_sve_udiv:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_udiv_u)
        .setMatchingIROpcode(Instruction::UDiv);
  case Intrinsic::aarch64_sve_umax:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_umax_u);
  case Intrinsic::aarch64_sve_umin:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_umin_u);
  case Intrinsic::aarch64_sve_umulh:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_umulh_u);
  case Intrinsic::aarch64_sve_asr:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_asr_u)
        .setMatchingIROpcode(Instruction::AShr);
  case Intrinsic::aarch64_sve_lsl:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_lsl_u)
        .setMatchingIROpcode(Instruction::Shl);
  case Intrinsic::aarch64_sve_lsr:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_lsr_u)
        .setMatchingIROpcode(Instruction::LShr);
  case Intrinsic::aarch64_sve_and:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_and_u)
        .setMatchingIROpcode(Instruction::And);
  case Intrinsic::aarch64_sve_bic:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_bic_u);
  case Intrinsic::aarch64_sve_eor:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_eor_u)
        .setMatchingIROpcode(Instruction::Xor);
  case Intrinsic::aarch64_sve_orr:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_orr_u)
        .setMatchingIROpcode(Instruction::Or);
  case Intrinsic::aarch64_sve_shsub:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_shsub_u);
  case Intrinsic::aarch64_sve_shsubr:
    return SVEIntrinsicInfo::defaultMergingOp();
  case Intrinsic::aarch64_sve_sqrshl:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sqrshl_u);
  case Intrinsic::aarch64_sve_sqshl:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sqshl_u);
  case Intrinsic::aarch64_sve_sqsub:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_sqsub_u);
  case Intrinsic::aarch64_sve_srshl:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_srshl_u);
  case Intrinsic::aarch64_sve_uhsub:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_uhsub_u);
  case Intrinsic::aarch64_sve_uhsubr:
    return SVEIntrinsicInfo::defaultMergingOp();
  case Intrinsic::aarch64_sve_uqrshl:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_uqrshl_u);
  case Intrinsic::aarch64_sve_uqshl:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_uqshl_u);
  case Intrinsic::aarch64_sve_uqsub:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_uqsub_u);
  case Intrinsic::aarch64_sve_urshl:
    return SVEIntrinsicInfo::defaultMergingOp(Intrinsic::aarch64_sve_urshl_u);
 
  case Intrinsic::aarch64_sve_add_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::Add);
  case Intrinsic::aarch64_sve_and_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::And);
  case Intrinsic::aarch64_sve_asr_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::AShr);
  case Intrinsic::aarch64_sve_eor_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::Xor);
  case Intrinsic::aarch64_sve_fadd_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::FAdd);
  case Intrinsic::aarch64_sve_fdiv_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::FDiv);
  case Intrinsic::aarch64_sve_fmul_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::FMul);
  case Intrinsic::aarch64_sve_fsub_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::FSub);
  case Intrinsic::aarch64_sve_lsl_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::Shl);
  case Intrinsic::aarch64_sve_lsr_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::LShr);
  case Intrinsic::aarch64_sve_mul_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::Mul);
  case Intrinsic::aarch64_sve_orr_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::Or);
  case Intrinsic::aarch64_sve_sdiv_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::SDiv);
  case Intrinsic::aarch64_sve_sub_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::Sub);
  case Intrinsic::aarch64_sve_udiv_u:
    return SVEIntrinsicInfo::defaultUndefOp().setMatchingIROpcode(
        Instruction::UDiv);
 
  case Intrinsic::aarch64_sve_addqv:
  case Intrinsic::aarch64_sve_bic_z:
  case Intrinsic::aarch64_sve_brka_z:
  case Intrinsic::aarch64_sve_brkb_z:
  case Intrinsic::aarch64_sve_brkn_z:
  case Intrinsic::aarch64_sve_brkpa_z:
  case Intrinsic::aarch64_sve_brkpb_z:
  case Intrinsic::aarch64_sve_cntp:
  case Intrinsic::aarch64_sve_compact:
  case Intrinsic::aarch64_sve_eorv:
  case Intrinsic::aarch64_sve_eorqv:
  case Intrinsic::aarch64_sve_nand_z:
  case Intrinsic::aarch64_sve_nor_z:
  case Intrinsic::aarch64_sve_orn_z:
  case Intrinsic::aarch64_sve_orv:
  case Intrinsic::aarch64_sve_orqv:
  case Intrinsic::aarch64_sve_pnext:
  case Intrinsic::aarch64_sve_rdffr_z:
  case Intrinsic::aarch64_sve_saddv:
  case Intrinsic::aarch64_sve_uaddv:
  case Intrinsic::aarch64_sve_umaxv:
  case Intrinsic::aarch64_sve_umaxqv:
  case Intrinsic::aarch64_sve_facge:
  case Intrinsic::aarch64_sve_facgt:
  case Intrinsic::aarch64_sve_ld1:
  case Intrinsic::aarch64_sve_ld1_gather:
  case Intrinsic::aarch64_sve_ld1_gather_index:
  case Intrinsic::aarch64_sve_ld1_gather_scalar_offset:
  case Intrinsic::aarch64_sve_ld1_gather_sxtw:
  case Intrinsic::aarch64_sve_ld1_gather_sxtw_index:
  case Intrinsic::aarch64_sve_ld1_gather_uxtw:
  case Intrinsic::aarch64_sve_ld1_gather_uxtw_index:
  case Intrinsic::aarch64_sve_ld1q_gather_index:
  case Intrinsic::aarch64_sve_ld1q_gather_scalar_offset:
  case Intrinsic::aarch64_sve_ld1q_gather_vector_offset:
  case Intrinsic::aarch64_sve_ld1ro:
  case Intrinsic::aarch64_sve_ld1rq:
  case Intrinsic::aarch64_sve_ld1udq:
  case Intrinsic::aarch64_sve_ld1uwq:
  case Intrinsic::aarch64_sve_ld2_sret:
  case Intrinsic::aarch64_sve_ld2q_sret:
  case Intrinsic::aarch64_sve_ld3_sret:
  case Intrinsic::aarch64_sve_ld3q_sret:
  case Intrinsic::aarch64_sve_ld4_sret:
  case Intrinsic::aarch64_sve_ld4q_sret:
  case Intrinsic::aarch64_sve_ldff1:
  case Intrinsic::aarch64_sve_ldff1_gather:
  case Intrinsic::aarch64_sve_ldff1_gather_index:
  case Intrinsic::aarch64_sve_ldff1_gather_scalar_offset:
  case Intrinsic::aarch64_sve_ldff1_gather_sxtw:
  case Intrinsic::aarch64_sve_ldff1_gather_sxtw_index:
  case Intrinsic::aarch64_sve_ldff1_gather_uxtw:
  case Intrinsic::aarch64_sve_ldff1_gather_uxtw_index:
  case Intrinsic::aarch64_sve_ldnf1:
  case Intrinsic::aarch64_sve_ldnt1:
  case Intrinsic::aarch64_sve_ldnt1_gather:
  case Intrinsic::aarch64_sve_ldnt1_gather_index:
  case Intrinsic::aarch64_sve_ldnt1_gather_scalar_offset:
  case Intrinsic::aarch64_sve_ldnt1_gather_uxtw:
    return SVEIntrinsicInfo::defaultZeroingOp();
 
  case Intrinsic::aarch64_sve_and_z:
    return SVEIntrinsicInfo::defaultZeroingOp().setMatchingIROpcode(
        Instruction::And);
  case Intrinsic::aarch64_sve_orr_z:
    return SVEIntrinsicInfo::defaultZeroingOp().setMatchingIROpcode(
        Instruction::Or);
  case Intrinsic::aarch64_sve_eor_z:
    return SVEIntrinsicInfo::defaultZeroingOp().setMatchingIROpcode(
        Instruction::Xor);
 
  case Intrinsic::aarch64_sve_cmpeq:
  case Intrinsic::aarch64_sve_cmpeq_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_EQ);
  case Intrinsic::aarch64_sve_cmpge:
  case Intrinsic::aarch64_sve_cmpge_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_SGE);
  case Intrinsic::aarch64_sve_cmpgt:
  case Intrinsic::aarch64_sve_cmpgt_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_SGT);
  case Intrinsic::aarch64_sve_cmphi:
  case Intrinsic::aarch64_sve_cmphi_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_UGT);
  case Intrinsic::aarch64_sve_cmphs:
  case Intrinsic::aarch64_sve_cmphs_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_UGE);
  case Intrinsic::aarch64_sve_cmple_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_SLE);
  case Intrinsic::aarch64_sve_cmplo_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_ULT);
  case Intrinsic::aarch64_sve_cmpls_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_ULE);
  case Intrinsic::aarch64_sve_cmplt_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_SLT);
  case Intrinsic::aarch64_sve_cmpne:
  case Intrinsic::aarch64_sve_cmpne_wide:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::ICMP_NE);
  case Intrinsic::aarch64_sve_fcmpeq:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::FCMP_OEQ);
  case Intrinsic::aarch64_sve_fcmpge:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::FCMP_OGE);
  case Intrinsic::aarch64_sve_fcmpgt:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::FCMP_OGT);
  case Intrinsic::aarch64_sve_fcmpne:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::FCMP_UNE);
  case Intrinsic::aarch64_sve_fcmpuo:
    return SVEIntrinsicInfo::defaultZeroingOp().setCmpPredicate(
        CmpInst::FCMP_UNO);
 
  case Intrinsic::aarch64_sve_prf:
  case Intrinsic::aarch64_sve_prfb_gather_index:
  case Intrinsic::aarch64_sve_prfb_gather_scalar_offset:
  case Intrinsic::aarch64_sve_prfb_gather_sxtw_index:
  case Intrinsic::aarch64_sve_prfb_gather_uxtw_index:
  case Intrinsic::aarch64_sve_prfd_gather_index:
  case Intrinsic::aarch64_sve_prfd_gather_scalar_offset:
  case Intrinsic::aarch64_sve_prfd_gather_sxtw_index:
  case Intrinsic::aarch64_sve_prfd_gather_uxtw_index:
  case Intrinsic::aarch64_sve_prfh_gather_index:
  case Intrinsic::aarch64_sve_prfh_gather_scalar_offset:
  case Intrinsic::aarch64_sve_prfh_gather_sxtw_index:
  case Intrinsic::aarch64_sve_prfh_gather_uxtw_index:
  case Intrinsic::aarch64_sve_prfw_gather_index:
  case Intrinsic::aarch64_sve_prfw_gather_scalar_offset:
  case Intrinsic::aarch64_sve_prfw_gather_sxtw_index:
  case Intrinsic::aarch64_sve_prfw_gather_uxtw_index:
    return SVEIntrinsicInfo::defaultVoidOp(0);
 
  case Intrinsic::aarch64_sve_st1_scatter:
  case Intrinsic::aarch64_sve_st1_scatter_scalar_offset:
  case Intrinsic::aarch64_sve_st1_scatter_sxtw:
  case Intrinsic::aarch64_sve_st1_scatter_sxtw_index:
  case Intrinsic::aarch64_sve_st1_scatter_uxtw:
  case Intrinsic::aarch64_sve_st1_scatter_uxtw_index:
  case Intrinsic::aarch64_sve_st1dq:
  case Intrinsic::aarch64_sve_st1q_scatter_index:
  case Intrinsic::aarch64_sve_st1q_scatter_scalar_offset:
  case Intrinsic::aarch64_sve_st1q_scatter_vector_offset:
  case Intrinsic::aarch64_sve_st1wq:
  case Intrinsic::aarch64_sve_stnt1:
  case Intrinsic::aarch64_sve_stnt1_scatter:
  case Intrinsic::aarch64_sve_stnt1_scatter_index:
  case Intrinsic::aarch64_sve_stnt1_scatter_scalar_offset:
  case Intrinsic::aarch64_sve_stnt1_scatter_uxtw:
    return SVEIntrinsicInfo::defaultVoidOp(1);
  case Intrinsic::aarch64_sve_st2:
  case Intrinsic::aarch64_sve_st2q:
    return SVEIntrinsicInfo::defaultVoidOp(2);
  case Intrinsic::aarch64_sve_st3:
  case Intrinsic::aarch64_sve_st3q:
    return SVEIntrinsicInfo::defaultVoidOp(3);
  case Intrinsic::aarch64_sve_st4:
  case Intrinsic::aarch64_sve_st4q:
    return SVEIntrinsicInfo::defaultVoidOp(4);
  }
 
  return SVEIntrinsicInfo();
}
 
static bool isAllActivePredicate(Value *Pred) {
  Value *UncastedPred;
 
  // Look through predicate casts that only remove lanes.
  if (match(Pred, m_Intrinsic<Intrinsic::aarch64_sve_convert_from_svbool>(
                      m_Value(UncastedPred)))) {
    auto *OrigPredTy = cast<ScalableVectorType>(Pred->getType());
    Pred = UncastedPred;
 
    if (match(Pred, m_Intrinsic<Intrinsic::aarch64_sve_convert_to_svbool>(
                        m_Value(UncastedPred))))
      // If the predicate has the same or less lanes than the uncasted predicate
      // then we know the casting has no effect.
      if (OrigPredTy->getMinNumElements() <=
          cast<ScalableVectorType>(UncastedPred->getType())
              ->getMinNumElements())
        Pred = UncastedPred;
  }
 
  auto *C = dyn_cast<Constant>(Pred);
  return C && C->isAllOnesValue();
}
 
// Simplify `V` by only considering the operations that affect active lanes.
// This function should only return existing Values or newly created Constants.
static Value *stripInactiveLanes(Value *V, const Value *Pg) {
  auto *Dup = dyn_cast<IntrinsicInst>(V);
  if (Dup && Dup->getIntrinsicID() == Intrinsic::aarch64_sve_dup &&
      Dup->getOperand(1) == Pg && isa<Constant>(Dup->getOperand(2)))
    return ConstantVector::getSplat(
        cast<VectorType>(V->getType())->getElementCount(),
        cast<Constant>(Dup->getOperand(2)));
 
  return V;
}
 
static std::optional<Instruction *>
simplifySVEIntrinsicBinOp(InstCombiner &IC, IntrinsicInst &II,
                          const SVEIntrinsicInfo &IInfo) {
  const unsigned Opc = IInfo.getMatchingIROpode();
  assert(Instruction::isBinaryOp(Opc) && "Expected a binary operation!");
 
  Value *Pg = II.getOperand(0);
  Value *Op1 = II.getOperand(1);
  Value *Op2 = II.getOperand(2);
  const DataLayout &DL = II.getDataLayout();
 
  // Canonicalise constants to the RHS.
  if (Instruction::isCommutative(Opc) && IInfo.inactiveLanesAreNotDefined() &&
      isa<Constant>(Op1) && !isa<Constant>(Op2)) {
    IC.replaceOperand(II, 1, Op2);
    IC.replaceOperand(II, 2, Op1);
    return &II;
  }
 
  // Only active lanes matter when simplifying the operation.
  Op1 = stripInactiveLanes(Op1, Pg);
  Op2 = stripInactiveLanes(Op2, Pg);
 
  Value *SimpleII;
  if (auto FII = dyn_cast<FPMathOperator>(&II))
    SimpleII = simplifyBinOp(Opc, Op1, Op2, FII->getFastMathFlags(), DL);
  else
    SimpleII = simplifyBinOp(Opc, Op1, Op2, DL);
 
  // If both operands are convert.to.svbool from the same narrower predicate
  // type, try to simplify the operation at that narrower type. This is valid
  // because the conversions zero the lanes not represented by the narrower
  // type, so those lanes of the result are zero either way.
  Value *NarrowOp1, *NarrowOp2;
  if (!SimpleII &&
      match(Op1, m_Intrinsic<Intrinsic::aarch64_sve_convert_to_svbool>(
                     m_Value(NarrowOp1))) &&
      match(Op2, m_Intrinsic<Intrinsic::aarch64_sve_convert_to_svbool>(
                     m_Value(NarrowOp2))) &&
      NarrowOp1->getType() == NarrowOp2->getType()) {
    Value *SimpleNarrow = simplifyBinOp(Opc, NarrowOp1, NarrowOp2, DL);
    if (SimpleNarrow && !isa<UndefValue>(SimpleNarrow)) {
      if (match(SimpleNarrow, m_ZeroInt()))
        SimpleII = Constant::getNullValue(II.getType());
      else if (SimpleNarrow == NarrowOp1)
        SimpleII = Op1;
      else if (SimpleNarrow == NarrowOp2)
        SimpleII = Op2;
      else
        SimpleII =
            IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_convert_to_svbool,
                                       SimpleNarrow->getType(), SimpleNarrow);
    }
  }
 
  // An SVE intrinsic's result is always defined. However, this is not the case
  // for its equivalent IR instruction (e.g. when shifting by an amount more
  // than the data's bitwidth). Simplifications to an undefined result must be
  // ignored to preserve the intrinsic's expected behaviour.
  if (!SimpleII || isa<UndefValue>(SimpleII))
    return std::nullopt;
 
  if (IInfo.inactiveLanesAreNotDefined())
    return IC.replaceInstUsesWith(II, SimpleII);
 
  Value *Inactive =
      IInfo.resultIsZeroInitialized()
          ? Constant::getNullValue(II.getType())
          : II.getOperand(IInfo.getOperandIdxInactiveLanesTakenFrom());
 
  // The intrinsic does nothing (e.g. sve.mul(pg, A, 1.0)).
  if (SimpleII == Inactive)
    return IC.replaceInstUsesWith(II, SimpleII);
 
  // Inactive lanes must be preserved.
  SimpleII = IC.Builder.CreateSelect(Pg, SimpleII, Inactive);
  return IC.replaceInstUsesWith(II, SimpleII);
}
 
static std::optional<Instruction *>
simplifySVEIntrinsicCompare(InstCombiner &IC, IntrinsicInst &II,
                            const SVEIntrinsicInfo &IInfo) {
  const unsigned Opc = IInfo.getMatchingIROpode();
  assert((Opc == Instruction::ICmp || Opc == Instruction::FCmp) &&
         "Expected a compare operation!");
 
  Value *Pg = II.getOperand(0);
  Value *LHS = II.getOperand(1);
  Value *RHS = II.getOperand(2);
  CmpInst::Predicate CmpPred = IInfo.getCmpPredicate();
  bool IsWideICmp =
      Opc == Instruction::ICmp && LHS->getType() != RHS->getType();
  assert((IsWideICmp || LHS->getType() == RHS->getType()) &&
         "Unexpected wide compare!");
 
  // Canonicalise constants to the RHS.
  if ((ICmpInst::isCommutative(CmpPred) || FCmpInst::isCommutative(CmpPred)) &&
      isa<Constant>(LHS) && !isa<Constant>(RHS) && !IsWideICmp) {
    IC.replaceOperand(II, 1, RHS);
    IC.replaceOperand(II, 2, LHS);
    return &II;
  }
 
  // Only active lanes matter when simplifying the operation.
  LHS = stripInactiveLanes(LHS, Pg);
  RHS = stripInactiveLanes(RHS, Pg);
 
  if (IsWideICmp) {
    // We can do more for wide compares, but not using simplifyCmpInst.
    const APInt *LHSVal, *RHSVal;
    if (!match(LHS, m_APInt(LHSVal)) || !match(RHS, m_APInt(RHSVal)))
      return std::nullopt;
 
    // Consider cmpge.wide(..., <vscale x 4 x i32> LHS, <vscale x 2 x i64> RHS),
    // we must reconstruct the constants because LHS has the wrong element type,
    // and RHS the wrong element count.
    Type *WideVT = VectorType::get(RHS->getType()->getScalarType(),
                                   cast<VectorType>(LHS->getType()));
    // NOTE: Wide equality comparisons are signed.
    if (ICmpInst::isUnsigned(CmpPred)) {
      LHS = ConstantInt::get(WideVT, LHSVal->getZExtValue());
      RHS = ConstantInt::get(WideVT, RHSVal->getZExtValue());
    } else {
      LHS = ConstantInt::get(WideVT, LHSVal->getSExtValue());
      RHS = ConstantInt::get(WideVT, RHSVal->getSExtValue());
    }
  }
 
  // TODO: Allow fast-math flags for calls to compare intrinsics.
  const DataLayout &DL = II.getDataLayout();
  Value *SimpleII = simplifyCmpInst(CmpPred, LHS, RHS, DL);
 
  // No simplification happened.
  if (!SimpleII)
    return std::nullopt;
 
  assert(IInfo.resultIsZeroInitialized() && "Expected a zeroing operation!");
 
  if (match(SimpleII, m_ZeroInt()))
    return IC.replaceInstUsesWith(II, SimpleII);
 
  // Inactive lanes must be zeroed.
  SimpleII = IC.Builder.CreateLogicalAnd(Pg, SimpleII);
  return IC.replaceInstUsesWith(II, SimpleII);
}
 
// Use SVE intrinsic info to eliminate redundant operands and/or canonicalise
// to operations with less strict inactive lane requirements.
static std::optional<Instruction *>
simplifySVEIntrinsic(InstCombiner &IC, IntrinsicInst &II,
                     const SVEIntrinsicInfo &IInfo) {
  if (!IInfo.hasGoverningPredicate())
    return std::nullopt;
 
  auto *OpPredicate = II.getOperand(IInfo.getGoverningPredicateOperandIdx());
 
  // If there are no active lanes.
  if (match(OpPredicate, m_ZeroInt())) {
    if (IInfo.inactiveLanesTakenFromOperand())
      return IC.replaceInstUsesWith(
          II, II.getOperand(IInfo.getOperandIdxInactiveLanesTakenFrom()));
 
    if (IInfo.inactiveLanesAreUnused()) {
      if (IInfo.resultIsZeroInitialized())
        IC.replaceInstUsesWith(II, Constant::getNullValue(II.getType()));
 
      return IC.eraseInstFromFunction(II);
    }
  }
 
  // If there are no inactive lanes.
  if (isAllActivePredicate(OpPredicate)) {
    if (IInfo.hasOperandWithNoActiveLanes()) {
      unsigned OpIdx = IInfo.getOperandIdxWithNoActiveLanes();
      if (!isa<UndefValue>(II.getOperand(OpIdx)))
        return IC.replaceOperand(II, OpIdx, UndefValue::get(II.getType()));
    }
 
    if (IInfo.hasMatchingUndefIntrinsic()) {
      auto *NewDecl = Intrinsic::getOrInsertDeclaration(
          II.getModule(), IInfo.getMatchingUndefIntrinsic(), {II.getType()});
      II.setCalledFunction(NewDecl);
      return &II;
    }
  }
 
  if (!IInfo.hasMatchingIROpode())
    return std::nullopt;
 
  //
  // Operation specific simplifications.
  //
 
  unsigned Opc = IInfo.getMatchingIROpode();
 
  if (Instruction::isBinaryOp(Opc))
    return simplifySVEIntrinsicBinOp(IC, II, IInfo);
 
  if (Opc == Instruction::FCmp || Opc == Instruction::ICmp)
    return simplifySVEIntrinsicCompare(IC, II, IInfo);
 
  return std::nullopt;
}
 
// (from_svbool (binop (to_svbool pred) (svbool_t _) (svbool_t _))))
// => (binop (pred) (from_svbool _) (from_svbool _))
//
// The above transformation eliminates a `to_svbool` in the predicate
// operand of bitwise operation `binop` by narrowing the vector width of
// the operation. For example, it would convert a `<vscale x 16 x i1>
// and` into a `<vscale x 4 x i1> and`. This is profitable because
// to_svbool must zero the new lanes during widening, whereas
// from_svbool is free.
static std::optional<Instruction *>
tryCombineFromSVBoolBinOp(InstCombiner &IC, IntrinsicInst &II) {
  auto m_ConvertToSVBool = [](auto P) {
    return m_Intrinsic<Intrinsic::aarch64_sve_convert_to_svbool>(P);
  };
  constexpr Intrinsic::ID ConvertFromSVBool =
      Intrinsic::aarch64_sve_convert_from_svbool;
 
  Type *Ty = II.getType();
  Value *LHS, *RHS, *NarrowLHS, *NarrowRHS;
 
  if (match(II.getOperand(0),
            m_LogicalAnd(m_Value(LHS),
                         m_ConvertToSVBool(m_SpecificType(Ty, NarrowRHS))))) {
    NarrowLHS = IC.Builder.CreateIntrinsic(ConvertFromSVBool, Ty, LHS);
    Value *NarrowAnd = IC.Builder.CreateLogicalAnd(NarrowLHS, NarrowRHS);
    return IC.replaceInstUsesWith(II, NarrowAnd);
  }
 
  if (match(II.getOperand(0),
            m_LogicalAnd(m_ConvertToSVBool(m_SpecificType(Ty, NarrowLHS)),
                         m_Value(RHS)))) {
    NarrowRHS = IC.Builder.CreateIntrinsic(ConvertFromSVBool, Ty, RHS);
    Value *NarrowAnd = IC.Builder.CreateLogicalAnd(NarrowLHS, NarrowRHS);
    return IC.replaceInstUsesWith(II, NarrowAnd);
  }
 
  auto BinOp = dyn_cast<IntrinsicInst>(II.getOperand(0));
  if (!BinOp)
    return std::nullopt;
 
  Intrinsic::ID BinOpIID = BinOp->getIntrinsicID();
  switch (BinOpIID) {
  case Intrinsic::aarch64_sve_and_z:
  case Intrinsic::aarch64_sve_bic_z:
  case Intrinsic::aarch64_sve_eor_z:
  case Intrinsic::aarch64_sve_nand_z:
  case Intrinsic::aarch64_sve_nor_z:
  case Intrinsic::aarch64_sve_orn_z:
  case Intrinsic::aarch64_sve_orr_z:
    break;
  default:
    return std::nullopt;
  }
 
  Value *BinOpPred = BinOp->getOperand(0);
  Value *BinOpOp1 = BinOp->getOperand(1);
  Value *BinOpOp2 = BinOp->getOperand(2);
 
  Value *NarrowBinOpPred;
  if (!match(BinOpPred, m_ConvertToSVBool(m_SpecificType(Ty, NarrowBinOpPred))))
    return std::nullopt;
 
  Value *NarrowBinOpOp1 =
      IC.Builder.CreateIntrinsic(ConvertFromSVBool, Ty, BinOpOp1);
  Value *NarrowBinOpOp2 = NarrowBinOpOp1;
  if (BinOpOp1 != BinOpOp2)
    NarrowBinOpOp2 =
        IC.Builder.CreateIntrinsic(ConvertFromSVBool, Ty, BinOpOp2);
  Value *NarrowedBinOp = IC.Builder.CreateIntrinsic(
      BinOpIID, Ty, {NarrowBinOpPred, NarrowBinOpOp1, NarrowBinOpOp2});
  return IC.replaceInstUsesWith(II, NarrowedBinOp);
}
 
static std::optional<Instruction *>
instCombineConvertFromSVBool(InstCombiner &IC, IntrinsicInst &II) {
  // If the reinterpret instruction operand is a PHI Node
  if (isa<PHINode>(II.getArgOperand(0)))
    return processPhiNode(IC, II);
 
  if (auto BinOpCombine = tryCombineFromSVBoolBinOp(IC, II))
    return BinOpCombine;
 
  // Ignore converts to/from svcount_t.
  if (isa<TargetExtType>(II.getArgOperand(0)->getType()) ||
      isa<TargetExtType>(II.getType()))
    return std::nullopt;
 
  SmallVector<Instruction *, 32> CandidatesForRemoval;
  Value *Cursor = II.getOperand(0), *EarliestReplacement = nullptr;
 
  const auto *IVTy = cast<VectorType>(II.getType());
 
  // Walk the chain of conversions.
  while (Cursor) {
    // If the type of the cursor has fewer lanes than the final result, zeroing
    // must take place, which breaks the equivalence chain.
    const auto *CursorVTy = cast<VectorType>(Cursor->getType());
    if (CursorVTy->getElementCount().getKnownMinValue() <
        IVTy->getElementCount().getKnownMinValue())
      break;
 
    // If the cursor has the same type as I, it is a viable replacement.
    if (Cursor->getType() == IVTy)
      EarliestReplacement = Cursor;
 
    auto *IntrinsicCursor = dyn_cast<IntrinsicInst>(Cursor);
 
    // If this is not an SVE conversion intrinsic, this is the end of the chain.
    if (!IntrinsicCursor || !(IntrinsicCursor->getIntrinsicID() ==
                                  Intrinsic::aarch64_sve_convert_to_svbool ||
                              IntrinsicCursor->getIntrinsicID() ==
                                  Intrinsic::aarch64_sve_convert_from_svbool))
      break;
 
    CandidatesForRemoval.insert(CandidatesForRemoval.begin(), IntrinsicCursor);
    Cursor = IntrinsicCursor->getOperand(0);
  }
 
  // If no viable replacement in the conversion chain was found, there is
  // nothing to do.
  if (!EarliestReplacement)
    return std::nullopt;
 
  return IC.replaceInstUsesWith(II, EarliestReplacement);
}
 
static std::optional<Instruction *> instCombineSVESel(InstCombiner &IC,
                                                      IntrinsicInst &II) {
  // svsel(ptrue, x, y) => x
  auto *OpPredicate = II.getOperand(0);
  if (isAllActivePredicate(OpPredicate))
    return IC.replaceInstUsesWith(II, II.getOperand(1));
 
  auto Select =
      IC.Builder.CreateSelect(OpPredicate, II.getOperand(1), II.getOperand(2));
  return IC.replaceInstUsesWith(II, Select);
}
 
static std::optional<Instruction *> instCombineSVEDup(InstCombiner &IC,
                                                      IntrinsicInst &II) {
  Value *Pg = II.getOperand(1);
 
  // sve.dup(V, all_active, X) ==> splat(X)
  if (isAllActivePredicate(Pg)) {
    auto *RetTy = cast<ScalableVectorType>(II.getType());
    Value *Splat = IC.Builder.CreateVectorSplat(RetTy->getElementCount(),
                                                II.getArgOperand(2));
    return IC.replaceInstUsesWith(II, Splat);
  }
 
  if (!match(Pg, m_Intrinsic<Intrinsic::aarch64_sve_ptrue>(
                     m_SpecificInt(AArch64SVEPredPattern::vl1))))
    return std::nullopt;
 
  // sve.dup(V, sve.ptrue(vl1), X) ==> insertelement V, X, 0
  Value *Insert = IC.Builder.CreateInsertElement(
      II.getArgOperand(0), II.getArgOperand(2), uint64_t(0));
  return IC.replaceInstUsesWith(II, Insert);
}
 
static std::optional<Instruction *> instCombineSVEDupX(InstCombiner &IC,
                                                       IntrinsicInst &II) {
  // Replace DupX with a regular IR splat.
  auto *RetTy = cast<ScalableVectorType>(II.getType());
  Value *Splat = IC.Builder.CreateVectorSplat(RetTy->getElementCount(),
                                              II.getArgOperand(0));
  Splat->takeName(&II);
  return IC.replaceInstUsesWith(II, Splat);
}
 
// xor(cmpne(%pg, %lhs, %rhs), %pg)
// -> cmpeq(%pg, %lhs, %rhs)
static std::optional<Instruction *> instCombineXorSVECmpCC(InstCombiner &IC,
                                                           IntrinsicInst &II) {
  if (!II.hasOneUse())
    return std::nullopt;
  auto *User = cast<Instruction>(*II.user_begin());
  if (!match(User, m_c_Xor(m_Specific(&II), m_Specific(II.getOperand(0)))))
    return std::nullopt;
 
  Intrinsic::ID IID;
  switch (II.getIntrinsicID()) {
  case Intrinsic::aarch64_sve_cmpne:
    IID = Intrinsic::aarch64_sve_cmpeq;
    break;
  case Intrinsic::aarch64_sve_cmpne_wide:
    IID = Intrinsic::aarch64_sve_cmpeq_wide;
    break;
  case Intrinsic::aarch64_sve_cmpeq:
    IID = Intrinsic::aarch64_sve_cmpne;
    break;
  case Intrinsic::aarch64_sve_cmpeq_wide:
    IID = Intrinsic::aarch64_sve_cmpne_wide;
    break;
  default:
    return std::nullopt;
  }
 
  IC.Builder.SetInsertPoint(User);
  Value *CMPCC = IC.Builder.CreateIntrinsic(
      IID, II.getOperand(1)->getType(),
      {II.getOperand(0), II.getOperand(1), II.getOperand(2)});
  IC.replaceInstUsesWith(*User, CMPCC);
  IC.eraseInstFromFunction(*User);
  return &II;
}
 
// zext(cmpne(ptrue, %v, 0))
// -> umin(%pg, %v, 1)
static std::optional<Instruction *> instCombineZExtSVECmpNE(InstCombiner &IC,
                                                            IntrinsicInst &II) {
  if (!isAllActivePredicate(II.getOperand(0)) ||
      !match(II.getOperand(2), m_Zero()))
    return std::nullopt;
 
  for (auto *U : II.users()) {
    if (match(U, m_ZExt(m_Specific(&II)))) {
      auto *User = cast<Instruction>(U);
      Type *Ty = II.getOperand(1)->getType();
      if (User->getType() != Ty)
        continue;
      IC.Builder.SetInsertPoint(User);
      Value *UMin = IC.Builder.CreateIntrinsic(
          Intrinsic::aarch64_sve_umin, Ty,
          {II.getOperand(0), II.getOperand(1), ConstantInt::get(Ty, 1)});
      IC.replaceInstUsesWith(*User, UMin);
      IC.eraseInstFromFunction(*User);
      return &II;
    }
  }
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVECmpNE(InstCombiner &IC,
                                                        IntrinsicInst &II) {
  LLVMContext &Ctx = II.getContext();
 
  if (auto Res = instCombineXorSVECmpCC(IC, II))
    return Res;
 
  if (auto Res = instCombineZExtSVECmpNE(IC, II))
    return Res;
 
  if (!isAllActivePredicate(II.getArgOperand(0)))
    return std::nullopt;
 
  // Check that we have a compare of zero..
  auto *SplatValue =
      dyn_cast_or_null<ConstantInt>(getSplatValue(II.getArgOperand(2)));
  if (!SplatValue || !SplatValue->isZero())
    return std::nullopt;
 
  // ..against a dupq
  auto *DupQLane = dyn_cast<IntrinsicInst>(II.getArgOperand(1));
  if (!DupQLane ||
      DupQLane->getIntrinsicID() != Intrinsic::aarch64_sve_dupq_lane)
    return std::nullopt;
 
  // Where the dupq is a lane 0 replicate of a vector insert
  auto *DupQLaneIdx = dyn_cast<ConstantInt>(DupQLane->getArgOperand(1));
  if (!DupQLaneIdx || !DupQLaneIdx->isZero())
    return std::nullopt;
 
  auto *VecIns = dyn_cast<IntrinsicInst>(DupQLane->getArgOperand(0));
  if (!VecIns || VecIns->getIntrinsicID() != Intrinsic::vector_insert)
    return std::nullopt;
 
  // Where the vector insert is a fixed constant vector insert into undef at
  // index zero
  if (!isa<UndefValue>(VecIns->getArgOperand(0)))
    return std::nullopt;
 
  if (!cast<ConstantInt>(VecIns->getArgOperand(2))->isZero())
    return std::nullopt;
 
  auto *ConstVec = dyn_cast<Constant>(VecIns->getArgOperand(1));
  if (!ConstVec)
    return std::nullopt;
 
  auto *VecTy = dyn_cast<FixedVectorType>(ConstVec->getType());
  auto *OutTy = dyn_cast<ScalableVectorType>(II.getType());
  if (!VecTy || !OutTy || VecTy->getNumElements() != OutTy->getMinNumElements())
    return std::nullopt;
 
  unsigned NumElts = VecTy->getNumElements();
  unsigned PredicateBits = 0;
 
  // Expand intrinsic operands to a 16-bit byte level predicate
  for (unsigned I = 0; I < NumElts; ++I) {
    auto *Arg = dyn_cast<ConstantInt>(ConstVec->getAggregateElement(I));
    if (!Arg)
      return std::nullopt;
    if (!Arg->isZero())
      PredicateBits |= 1 << (I * (16 / NumElts));
  }
 
  // If all bits are zero bail early with an empty predicate
  if (PredicateBits == 0) {
    auto *PFalse = Constant::getNullValue(II.getType());
    PFalse->takeName(&II);
    return IC.replaceInstUsesWith(II, PFalse);
  }
 
  // Calculate largest predicate type used (where byte predicate is largest)
  unsigned Mask = 8;
  for (unsigned I = 0; I < 16; ++I)
    if ((PredicateBits & (1 << I)) != 0)
      Mask |= (I % 8);
 
  unsigned PredSize = Mask & -Mask;
  auto *PredType = ScalableVectorType::get(
      Type::getInt1Ty(Ctx), AArch64::SVEBitsPerBlock / (PredSize * 8));
 
  // Ensure all relevant bits are set
  for (unsigned I = 0; I < 16; I += PredSize)
    if ((PredicateBits & (1 << I)) == 0)
      return std::nullopt;
 
  auto *ConvertToSVBool =
      IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_convert_to_svbool,
                                 PredType, ConstantInt::getTrue(PredType));
  auto *ConvertFromSVBool =
      IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_convert_from_svbool,
                                 II.getType(), ConvertToSVBool);
 
  ConvertFromSVBool->takeName(&II);
  return IC.replaceInstUsesWith(II, ConvertFromSVBool);
}
 
static std::optional<Instruction *> instCombineSVELast(InstCombiner &IC,
                                                       IntrinsicInst &II) {
  Value *Pg = II.getArgOperand(0);
  Value *Vec = II.getArgOperand(1);
  auto IntrinsicID = II.getIntrinsicID();
  bool IsAfter = IntrinsicID == Intrinsic::aarch64_sve_lasta;
 
  // lastX(splat(X)) --> X
  if (auto *SplatVal = getSplatValue(Vec))
    return IC.replaceInstUsesWith(II, SplatVal);
 
  // If x and/or y is a splat value then:
  // lastX (binop (x, y)) --> binop(lastX(x), lastX(y))
  Value *LHS, *RHS;
  if (match(Vec, m_OneUse(m_BinOp(m_Value(LHS), m_Value(RHS))))) {
    if (isSplatValue(LHS) || isSplatValue(RHS)) {
      auto *OldBinOp = cast<BinaryOperator>(Vec);
      auto OpC = OldBinOp->getOpcode();
      auto *NewLHS =
          IC.Builder.CreateIntrinsic(IntrinsicID, {Vec->getType()}, {Pg, LHS});
      auto *NewRHS =
          IC.Builder.CreateIntrinsic(IntrinsicID, {Vec->getType()}, {Pg, RHS});
      auto *NewBinOp = BinaryOperator::CreateWithCopiedFlags(
          OpC, NewLHS, NewRHS, OldBinOp, OldBinOp->getName(), II.getIterator());
      return IC.replaceInstUsesWith(II, NewBinOp);
    }
  }
 
  auto *C = dyn_cast<Constant>(Pg);
  if (IsAfter && C && C->isNullValue()) {
    // The intrinsic is extracting lane 0 so use an extract instead.
    auto *IdxTy = Type::getInt64Ty(II.getContext());
    auto *Extract = ExtractElementInst::Create(Vec, ConstantInt::get(IdxTy, 0));
    Extract->insertBefore(II.getIterator());
    Extract->takeName(&II);
    return IC.replaceInstUsesWith(II, Extract);
  }
 
  auto *IntrPG = dyn_cast<IntrinsicInst>(Pg);
  if (!IntrPG)
    return std::nullopt;
 
  if (IntrPG->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)
    return std::nullopt;
 
  const auto PTruePattern =
      cast<ConstantInt>(IntrPG->getOperand(0))->getZExtValue();
 
  // Can the intrinsic's predicate be converted to a known constant index?
  unsigned MinNumElts = getNumElementsFromSVEPredPattern(PTruePattern);
  if (!MinNumElts)
    return std::nullopt;
 
  unsigned Idx = MinNumElts - 1;
  // Increment the index if extracting the element after the last active
  // predicate element.
  if (IsAfter)
    ++Idx;
 
  // Ignore extracts whose index is larger than the known minimum vector
  // length. NOTE: This is an artificial constraint where we prefer to
  // maintain what the user asked for until an alternative is proven faster.
  auto *PgVTy = cast<ScalableVectorType>(Pg->getType());
  if (Idx >= PgVTy->getMinNumElements())
    return std::nullopt;
 
  // The intrinsic is extracting a fixed lane so use an extract instead.
  auto *IdxTy = Type::getInt64Ty(II.getContext());
  auto *Extract = ExtractElementInst::Create(Vec, ConstantInt::get(IdxTy, Idx));
  Extract->insertBefore(II.getIterator());
  Extract->takeName(&II);
  return IC.replaceInstUsesWith(II, Extract);
}
 
static std::optional<Instruction *> instCombineSVECondLast(InstCombiner &IC,
                                                           IntrinsicInst &II) {
  // The SIMD&FP variant of CLAST[AB] is significantly faster than the scalar
  // integer variant across a variety of micro-architectures. Replace scalar
  // integer CLAST[AB] intrinsic with optimal SIMD&FP variant. A simple
  // bitcast-to-fp + clast[ab] + bitcast-to-int will cost a cycle or two more
  // depending on the micro-architecture, but has been observed as generally
  // being faster, particularly when the CLAST[AB] op is a loop-carried
  // dependency.
  Value *Pg = II.getArgOperand(0);
  Value *Fallback = II.getArgOperand(1);
  Value *Vec = II.getArgOperand(2);
  Type *Ty = II.getType();
 
  if (!Ty->isIntegerTy())
    return std::nullopt;
 
  Type *FPTy;
  switch (cast<IntegerType>(Ty)->getBitWidth()) {
  default:
    return std::nullopt;
  case 16:
    FPTy = IC.Builder.getHalfTy();
    break;
  case 32:
    FPTy = IC.Builder.getFloatTy();
    break;
  case 64:
    FPTy = IC.Builder.getDoubleTy();
    break;
  }
 
  Value *FPFallBack = IC.Builder.CreateBitCast(Fallback, FPTy);
  auto *FPVTy = VectorType::get(
      FPTy, cast<VectorType>(Vec->getType())->getElementCount());
  Value *FPVec = IC.Builder.CreateBitCast(Vec, FPVTy);
  auto *FPII = IC.Builder.CreateIntrinsic(
      II.getIntrinsicID(), {FPVec->getType()}, {Pg, FPFallBack, FPVec});
  Value *FPIItoInt = IC.Builder.CreateBitCast(FPII, II.getType());
  return IC.replaceInstUsesWith(II, FPIItoInt);
}
 
static std::optional<Instruction *> instCombineRDFFR(InstCombiner &IC,
                                                     IntrinsicInst &II) {
  // Replace rdffr with predicated rdffr.z intrinsic, so that optimizePTestInstr
  // can work with RDFFR_PP for ptest elimination.
  auto *RDFFR = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_rdffr_z,
                                           ConstantInt::getTrue(II.getType()));
  RDFFR->takeName(&II);
  return IC.replaceInstUsesWith(II, RDFFR);
}
 
static std::optional<Instruction *>
instCombineSVECntElts(InstCombiner &IC, IntrinsicInst &II, unsigned NumElts) {
  const auto Pattern = cast<ConstantInt>(II.getArgOperand(0))->getZExtValue();
 
  if (Pattern == AArch64SVEPredPattern::all) {
    Value *Cnt = IC.Builder.CreateElementCount(
        II.getType(), ElementCount::getScalable(NumElts));
    Cnt->takeName(&II);
    return IC.replaceInstUsesWith(II, Cnt);
  }
 
  unsigned MinNumElts = getNumElementsFromSVEPredPattern(Pattern);
 
  return MinNumElts && NumElts >= MinNumElts
             ? std::optional<Instruction *>(IC.replaceInstUsesWith(
                   II, ConstantInt::get(II.getType(), MinNumElts)))
             : std::nullopt;
}
 
static std::optional<Instruction *>
instCombineSMECntsd(InstCombiner &IC, IntrinsicInst &II,
                    const AArch64Subtarget *ST) {
  if (!ST->isStreaming())
    return std::nullopt;
 
  // In streaming-mode, aarch64_sme_cntds is equivalent to aarch64_sve_cntd
  // with SVEPredPattern::all
  Value *Cnt =
      IC.Builder.CreateElementCount(II.getType(), ElementCount::getScalable(2));
  Cnt->takeName(&II);
  return IC.replaceInstUsesWith(II, Cnt);
}
 
static std::optional<Instruction *> instCombineSVEPTest(InstCombiner &IC,
                                                        IntrinsicInst &II) {
  Value *PgVal = II.getArgOperand(0);
  Value *OpVal = II.getArgOperand(1);
 
  // PTEST_<FIRST|LAST>(X, X) is equivalent to PTEST_ANY(X, X).
  // Later optimizations prefer this form.
  if (PgVal == OpVal &&
      (II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_first ||
       II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_last)) {
    Value *Ops[] = {PgVal, OpVal};
    Type *Tys[] = {PgVal->getType()};
 
    auto *PTest =
        IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_ptest_any, Tys, Ops);
    PTest->takeName(&II);
 
    return IC.replaceInstUsesWith(II, PTest);
  }
 
  IntrinsicInst *Pg = dyn_cast<IntrinsicInst>(PgVal);
  IntrinsicInst *Op = dyn_cast<IntrinsicInst>(OpVal);
 
  if (!Pg || !Op)
    return std::nullopt;
 
  Intrinsic::ID OpIID = Op->getIntrinsicID();
 
  if (Pg->getIntrinsicID() == Intrinsic::aarch64_sve_convert_to_svbool &&
      OpIID == Intrinsic::aarch64_sve_convert_to_svbool &&
      Pg->getArgOperand(0)->getType() == Op->getArgOperand(0)->getType()) {
    Value *Ops[] = {Pg->getArgOperand(0), Op->getArgOperand(0)};
    Type *Tys[] = {Pg->getArgOperand(0)->getType()};
 
    auto *PTest = IC.Builder.CreateIntrinsic(II.getIntrinsicID(), Tys, Ops);
 
    PTest->takeName(&II);
    return IC.replaceInstUsesWith(II, PTest);
  }
 
  // Transform PTEST_ANY(X=OP(PG,...), X) -> PTEST_ANY(PG, X)).
  // Later optimizations may rewrite sequence to use the flag-setting variant
  // of instruction X to remove PTEST.
  if ((Pg == Op) && (II.getIntrinsicID() == Intrinsic::aarch64_sve_ptest_any) &&
      ((OpIID == Intrinsic::aarch64_sve_brka_z) ||
       (OpIID == Intrinsic::aarch64_sve_brkb_z) ||
       (OpIID == Intrinsic::aarch64_sve_brkpa_z) ||
       (OpIID == Intrinsic::aarch64_sve_brkpb_z) ||
       (OpIID == Intrinsic::aarch64_sve_rdffr_z) ||
       (OpIID == Intrinsic::aarch64_sve_and_z) ||
       (OpIID == Intrinsic::aarch64_sve_bic_z) ||
       (OpIID == Intrinsic::aarch64_sve_eor_z) ||
       (OpIID == Intrinsic::aarch64_sve_nand_z) ||
       (OpIID == Intrinsic::aarch64_sve_nor_z) ||
       (OpIID == Intrinsic::aarch64_sve_orn_z) ||
       (OpIID == Intrinsic::aarch64_sve_orr_z))) {
    Value *Ops[] = {Pg->getArgOperand(0), Pg};
    Type *Tys[] = {Pg->getType()};
 
    auto *PTest = IC.Builder.CreateIntrinsic(II.getIntrinsicID(), Tys, Ops);
    PTest->takeName(&II);
 
    return IC.replaceInstUsesWith(II, PTest);
  }
 
  return std::nullopt;
}
 
template <Intrinsic::ID MulOpc, Intrinsic::ID FuseOpc>
static std::optional<Instruction *>
instCombineSVEVectorFuseMulAddSub(InstCombiner &IC, IntrinsicInst &II,
                                  bool MergeIntoAddendOp) {
  Value *P = II.getOperand(0);
  Value *MulOp0, *MulOp1, *AddendOp, *Mul;
  if (MergeIntoAddendOp) {
    AddendOp = II.getOperand(1);
    Mul = II.getOperand(2);
  } else {
    AddendOp = II.getOperand(2);
    Mul = II.getOperand(1);
  }
 
  if (!match(Mul, m_Intrinsic<MulOpc>(m_Specific(P), m_Value(MulOp0),
                                      m_Value(MulOp1))))
    return std::nullopt;
 
  if (!Mul->hasOneUse())
    return std::nullopt;
 
  Instruction *FMFSource = nullptr;
  if (II.getType()->isFPOrFPVectorTy()) {
    llvm::FastMathFlags FAddFlags = II.getFastMathFlags();
    // Stop the combine when the flags on the inputs differ in case dropping
    // flags would lead to us missing out on more beneficial optimizations.
    if (FAddFlags != cast<CallInst>(Mul)->getFastMathFlags())
      return std::nullopt;
    if (!FAddFlags.allowContract())
      return std::nullopt;
    FMFSource = &II;
  }
 
  Value *Res;
  if (MergeIntoAddendOp)
    Res = IC.Builder.CreateIntrinsic(FuseOpc, {II.getType()},
                                     {P, AddendOp, MulOp0, MulOp1}, FMFSource);
  else
    Res = IC.Builder.CreateIntrinsic(FuseOpc, {II.getType()},
                                     {P, MulOp0, MulOp1, AddendOp}, FMFSource);
 
  return IC.replaceInstUsesWith(II, Res);
}
 
static std::optional<Instruction *>
instCombineSVELD1(InstCombiner &IC, IntrinsicInst &II, const DataLayout &DL) {
  Value *Pred = II.getOperand(0);
  Value *PtrOp = II.getOperand(1);
  Type *VecTy = II.getType();
 
  if (isAllActivePredicate(Pred)) {
    LoadInst *Load = IC.Builder.CreateLoad(VecTy, PtrOp);
    Load->copyMetadata(II);
    return IC.replaceInstUsesWith(II, Load);
  }
 
  CallInst *MaskedLoad =
      IC.Builder.CreateMaskedLoad(VecTy, PtrOp, PtrOp->getPointerAlignment(DL),
                                  Pred, ConstantAggregateZero::get(VecTy));
  MaskedLoad->copyMetadata(II);
  return IC.replaceInstUsesWith(II, MaskedLoad);
}
 
static std::optional<Instruction *>
instCombineSVEST1(InstCombiner &IC, IntrinsicInst &II, const DataLayout &DL) {
  Value *VecOp = II.getOperand(0);
  Value *Pred = II.getOperand(1);
  Value *PtrOp = II.getOperand(2);
 
  if (isAllActivePredicate(Pred)) {
    StoreInst *Store = IC.Builder.CreateStore(VecOp, PtrOp);
    Store->copyMetadata(II);
    return IC.eraseInstFromFunction(II);
  }
 
  CallInst *MaskedStore = IC.Builder.CreateMaskedStore(
      VecOp, PtrOp, PtrOp->getPointerAlignment(DL), Pred);
  MaskedStore->copyMetadata(II);
  return IC.eraseInstFromFunction(II);
}
 
static Instruction::BinaryOps intrinsicIDToBinOpCode(unsigned Intrinsic) {
  switch (Intrinsic) {
  case Intrinsic::aarch64_sve_fmul_u:
    return Instruction::BinaryOps::FMul;
  case Intrinsic::aarch64_sve_fadd_u:
    return Instruction::BinaryOps::FAdd;
  case Intrinsic::aarch64_sve_fsub_u:
    return Instruction::BinaryOps::FSub;
  default:
    return Instruction::BinaryOpsEnd;
  }
}
 
static std::optional<Instruction *>
instCombineSVEVectorBinOp(InstCombiner &IC, IntrinsicInst &II) {
  // Bail due to missing support for ISD::STRICT_ scalable vector operations.
  if (II.isStrictFP())
    return std::nullopt;
 
  auto *OpPredicate = II.getOperand(0);
  auto BinOpCode = intrinsicIDToBinOpCode(II.getIntrinsicID());
  if (BinOpCode == Instruction::BinaryOpsEnd ||
      !isAllActivePredicate(OpPredicate))
    return std::nullopt;
  auto BinOp = IC.Builder.CreateBinOpFMF(
      BinOpCode, II.getOperand(1), II.getOperand(2), II.getFastMathFlags());
  return IC.replaceInstUsesWith(II, BinOp);
}
 
static std::optional<Instruction *>
instCombineSVEVectorMlaU(InstCombiner &IC, IntrinsicInst &II) {
  assert(II.getIntrinsicID() == Intrinsic::aarch64_sve_mla_u &&
         "Expected MLA_U intrinsic");
  Value *Acc = II.getArgOperand(1);
  Value *MulOp0 = II.getArgOperand(2);
  Value *MulOp1 = II.getArgOperand(3);
 
  // For mla_u, inactive lanes are undefined, so it is valid to drop the
  // predicate when replacing mla_u(acc, x, 1) with add(acc, x) or
  // mla_u(acc, x, -1) with sub(acc, x).
  if (match(MulOp0, m_One()))
    return IC.replaceInstUsesWith(II, IC.Builder.CreateAdd(Acc, MulOp1));
  if (match(MulOp1, m_One()))
    return IC.replaceInstUsesWith(II, IC.Builder.CreateAdd(Acc, MulOp0));
  if (match(MulOp0, m_AllOnes()))
    return IC.replaceInstUsesWith(II, IC.Builder.CreateSub(Acc, MulOp1));
  if (match(MulOp1, m_AllOnes()))
    return IC.replaceInstUsesWith(II, IC.Builder.CreateSub(Acc, MulOp0));
 
  if (isa<Constant>(MulOp0) && !isa<Constant>(MulOp1)) {
    II.setArgOperand(2, MulOp1);
    II.setArgOperand(3, MulOp0);
    return &II;
  }
 
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineSVEPairwiseAddLong(InstCombiner &IC, IntrinsicInst &II) {
  assert((II.getIntrinsicID() == Intrinsic::aarch64_sve_sadalp ||
          II.getIntrinsicID() == Intrinsic::aarch64_sve_uadalp) &&
         "Expected SADALP or UADALP intrinsic");
 
  // Simplify add(adalp(pg, zeroinitializer, in), wide_acc)
  //       -> adalp(pg, wide_acc, in)
  auto *User = dyn_cast_or_null<Instruction>(II.getUniqueUndroppableUser());
  if (!User || !match(II.getArgOperand(1), m_Zero()))
    return std::nullopt;
 
  Value *Acc;
  if (!match(User, m_c_Add(m_Specific(&II), m_Value(Acc))))
    return std::nullopt;
 
  IC.Builder.SetInsertPoint(User);
  Value *PairwiseAddLong = IC.Builder.CreateIntrinsic(
      II.getIntrinsicID(), {II.getType()},
      {II.getArgOperand(0), Acc, II.getArgOperand(2)});
 
  IC.replaceInstUsesWith(*User, PairwiseAddLong);
  IC.eraseInstFromFunction(*User);
  return &II; // II is now trivially dead and will get erased.
}
 
static std::optional<Instruction *> instCombineSVEVectorAdd(InstCombiner &IC,
                                                            IntrinsicInst &II) {
  if (auto MLA = instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul,
                                                   Intrinsic::aarch64_sve_mla>(
          IC, II, true))
    return MLA;
  if (auto MAD = instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul,
                                                   Intrinsic::aarch64_sve_mad>(
          IC, II, false))
    return MAD;
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineSVEVectorFAdd(InstCombiner &IC, IntrinsicInst &II) {
  if (auto FMLA =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fmla>(IC, II,
                                                                         true))
    return FMLA;
  if (auto FMAD =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fmad>(IC, II,
                                                                         false))
    return FMAD;
  if (auto FMLA =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
                                            Intrinsic::aarch64_sve_fmla>(IC, II,
                                                                         true))
    return FMLA;
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineSVEVectorFAddU(InstCombiner &IC, IntrinsicInst &II) {
  if (auto FMLA =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fmla>(IC, II,
                                                                         true))
    return FMLA;
  if (auto FMAD =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fmad>(IC, II,
                                                                         false))
    return FMAD;
  if (auto FMLA_U =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
                                            Intrinsic::aarch64_sve_fmla_u>(
              IC, II, true))
    return FMLA_U;
  return instCombineSVEVectorBinOp(IC, II);
}
 
static std::optional<Instruction *>
instCombineSVEVectorFSub(InstCombiner &IC, IntrinsicInst &II) {
  if (auto FMLS =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fmls>(IC, II,
                                                                         true))
    return FMLS;
  if (auto FMSB =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fnmsb>(
              IC, II, false))
    return FMSB;
  if (auto FMLS =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
                                            Intrinsic::aarch64_sve_fmls>(IC, II,
                                                                         true))
    return FMLS;
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineSVEVectorFSubU(InstCombiner &IC, IntrinsicInst &II) {
  if (auto FMLS =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fmls>(IC, II,
                                                                         true))
    return FMLS;
  if (auto FMSB =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul,
                                            Intrinsic::aarch64_sve_fnmsb>(
              IC, II, false))
    return FMSB;
  if (auto FMLS_U =
          instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_fmul_u,
                                            Intrinsic::aarch64_sve_fmls_u>(
              IC, II, true))
    return FMLS_U;
  return instCombineSVEVectorBinOp(IC, II);
}
 
static std::optional<Instruction *> instCombineSVEVectorSub(InstCombiner &IC,
                                                            IntrinsicInst &II) {
  if (auto MLS = instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul,
                                                   Intrinsic::aarch64_sve_mls>(
          IC, II, true))
    return MLS;
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVEUnpack(InstCombiner &IC,
                                                         IntrinsicInst &II) {
  Value *UnpackArg = II.getArgOperand(0);
  auto *RetTy = cast<ScalableVectorType>(II.getType());
  bool IsSigned = II.getIntrinsicID() == Intrinsic::aarch64_sve_sunpkhi ||
                  II.getIntrinsicID() == Intrinsic::aarch64_sve_sunpklo;
 
  // Hi = uunpkhi(splat(X)) --> Hi = splat(extend(X))
  // Lo = uunpklo(splat(X)) --> Lo = splat(extend(X))
  if (auto *ScalarArg = getSplatValue(UnpackArg)) {
    ScalarArg =
        IC.Builder.CreateIntCast(ScalarArg, RetTy->getScalarType(), IsSigned);
    Value *NewVal =
        IC.Builder.CreateVectorSplat(RetTy->getElementCount(), ScalarArg);
    NewVal->takeName(&II);
    return IC.replaceInstUsesWith(II, NewVal);
  }
 
  return std::nullopt;
}
static std::optional<Instruction *> instCombineSVETBL(InstCombiner &IC,
                                                      IntrinsicInst &II) {
  auto *OpVal = II.getOperand(0);
  auto *OpIndices = II.getOperand(1);
  VectorType *VTy = cast<VectorType>(II.getType());
 
  // Check whether OpIndices is a constant splat value < minimal element count
  // of result.
  auto *SplatValue = dyn_cast_or_null<ConstantInt>(getSplatValue(OpIndices));
  if (!SplatValue ||
      SplatValue->getValue().uge(VTy->getElementCount().getKnownMinValue()))
    return std::nullopt;
 
  // Convert sve_tbl(OpVal sve_dup_x(SplatValue)) to
  // splat_vector(extractelement(OpVal, SplatValue)) for further optimization.
  auto *Extract = IC.Builder.CreateExtractElement(OpVal, SplatValue);
  auto *VectorSplat =
      IC.Builder.CreateVectorSplat(VTy->getElementCount(), Extract);
 
  VectorSplat->takeName(&II);
  return IC.replaceInstUsesWith(II, VectorSplat);
}
 
static std::optional<Instruction *> instCombineSVEUzp1(InstCombiner &IC,
                                                       IntrinsicInst &II) {
  Value *A, *B;
  Type *RetTy = II.getType();
  constexpr Intrinsic::ID FromSVB = Intrinsic::aarch64_sve_convert_from_svbool;
  constexpr Intrinsic::ID ToSVB = Intrinsic::aarch64_sve_convert_to_svbool;
 
  // uzp1(to_svbool(A), to_svbool(B)) --> <A, B>
  // uzp1(from_svbool(to_svbool(A)), from_svbool(to_svbool(B))) --> <A, B>
  if ((match(II.getArgOperand(0),
             m_Intrinsic<FromSVB>(m_Intrinsic<ToSVB>(m_Value(A)))) &&
       match(II.getArgOperand(1),
             m_Intrinsic<FromSVB>(m_Intrinsic<ToSVB>(m_Value(B))))) ||
      (match(II.getArgOperand(0), m_Intrinsic<ToSVB>(m_Value(A))) &&
       match(II.getArgOperand(1), m_Intrinsic<ToSVB>(m_Value(B))))) {
    auto *TyA = cast<ScalableVectorType>(A->getType());
    if (TyA == B->getType() &&
        RetTy == ScalableVectorType::getDoubleElementsVectorType(TyA)) {
      auto *SubVec = IC.Builder.CreateInsertVector(
          RetTy, PoisonValue::get(RetTy), A, uint64_t(0));
      auto *ConcatVec = IC.Builder.CreateInsertVector(RetTy, SubVec, B,
                                                      TyA->getMinNumElements());
      ConcatVec->takeName(&II);
      return IC.replaceInstUsesWith(II, ConcatVec);
    }
  }
 
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVEZip(InstCombiner &IC,
                                                      IntrinsicInst &II) {
  // zip1(uzp1(A, B), uzp2(A, B)) --> A
  // zip2(uzp1(A, B), uzp2(A, B)) --> B
  Value *A, *B;
  if (match(II.getArgOperand(0),
            m_Intrinsic<Intrinsic::aarch64_sve_uzp1>(m_Value(A), m_Value(B))) &&
      match(II.getArgOperand(1), m_Intrinsic<Intrinsic::aarch64_sve_uzp2>(
                                     m_Specific(A), m_Specific(B))))
    return IC.replaceInstUsesWith(
        II, (II.getIntrinsicID() == Intrinsic::aarch64_sve_zip1 ? A : B));
 
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineLD1GatherIndex(InstCombiner &IC, IntrinsicInst &II) {
  Value *Mask = II.getOperand(0);
  Value *BasePtr = II.getOperand(1);
  Value *Index = II.getOperand(2);
  Type *Ty = II.getType();
  Value *PassThru = ConstantAggregateZero::get(Ty);
 
  // Contiguous gather => masked load.
  // (sve.ld1.gather.index Mask BasePtr (sve.index IndexBase 1))
  // => (masked.load (gep BasePtr IndexBase) Align Mask zeroinitializer)
  Value *IndexBase;
  if (match(Index, m_Intrinsic<Intrinsic::aarch64_sve_index>(m_Value(IndexBase),
                                                             m_One()))) {
    Align Alignment =
        BasePtr->getPointerAlignment(II.getDataLayout());
 
    Value *Ptr = IC.Builder.CreateGEP(cast<VectorType>(Ty)->getElementType(),
                                      BasePtr, IndexBase);
    CallInst *MaskedLoad =
        IC.Builder.CreateMaskedLoad(Ty, Ptr, Alignment, Mask, PassThru);
    MaskedLoad->takeName(&II);
    return IC.replaceInstUsesWith(II, MaskedLoad);
  }
 
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineST1ScatterIndex(InstCombiner &IC, IntrinsicInst &II) {
  Value *Val = II.getOperand(0);
  Value *Mask = II.getOperand(1);
  Value *BasePtr = II.getOperand(2);
  Value *Index = II.getOperand(3);
  Type *Ty = Val->getType();
 
  // Contiguous scatter => masked store.
  // (sve.st1.scatter.index Value Mask BasePtr (sve.index IndexBase 1))
  // => (masked.store Value (gep BasePtr IndexBase) Align Mask)
  Value *IndexBase;
  if (match(Index, m_Intrinsic<Intrinsic::aarch64_sve_index>(m_Value(IndexBase),
                                                             m_One()))) {
    Align Alignment =
        BasePtr->getPointerAlignment(II.getDataLayout());
 
    Value *Ptr = IC.Builder.CreateGEP(cast<VectorType>(Ty)->getElementType(),
                                      BasePtr, IndexBase);
    (void)IC.Builder.CreateMaskedStore(Val, Ptr, Alignment, Mask);
 
    return IC.eraseInstFromFunction(II);
  }
 
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVESDIV(InstCombiner &IC,
                                                       IntrinsicInst &II) {
  Type *Int32Ty = IC.Builder.getInt32Ty();
  Value *Pred = II.getOperand(0);
  Value *Vec = II.getOperand(1);
  Value *DivVec = II.getOperand(2);
 
  Value *SplatValue = getSplatValue(DivVec);
  ConstantInt *SplatConstantInt = dyn_cast_or_null<ConstantInt>(SplatValue);
  if (!SplatConstantInt)
    return std::nullopt;
 
  APInt Divisor = SplatConstantInt->getValue();
  const int64_t DivisorValue = Divisor.getSExtValue();
  if (DivisorValue == -1)
    return std::nullopt;
  if (DivisorValue == 1)
    IC.replaceInstUsesWith(II, Vec);
 
  if (Divisor.isPowerOf2()) {
    Constant *DivisorLog2 = ConstantInt::get(Int32Ty, Divisor.logBase2());
    auto ASRD = IC.Builder.CreateIntrinsic(
        Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, DivisorLog2});
    return IC.replaceInstUsesWith(II, ASRD);
  }
  if (Divisor.isNegatedPowerOf2()) {
    Divisor.negate();
    Constant *DivisorLog2 = ConstantInt::get(Int32Ty, Divisor.logBase2());
    auto ASRD = IC.Builder.CreateIntrinsic(
        Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, DivisorLog2});
    auto NEG = IC.Builder.CreateIntrinsic(
        Intrinsic::aarch64_sve_neg, {ASRD->getType()}, {ASRD, Pred, ASRD});
    return IC.replaceInstUsesWith(II, NEG);
  }
 
  return std::nullopt;
}
 
/// Match Or(And(MASK, A), And(Not(MASK), B)), which ISel selects as
/// BSL/BIT/BIF. Both Ands must have one use and be in the Or's block. MainAnd,
///
/// OtherAnd, A and B are out parameters with the corresponding instructions.
/// MainAnd refers to And that has Not as one of its operands.
static bool matchBSL(const Instruction *I, Instruction *&MainAnd,
                     Instruction *&OtherAnd, Value *&A, Value *&B) {
  Value *MaskValue;
  if (!match(I,
             m_c_Or(m_OneUse(m_Instruction(OtherAnd)),
                    m_OneUse(m_c_And(m_Not(m_Value(MaskValue)), m_Value(A))))))
    return false;
 
  if (!match(OtherAnd, m_c_And(m_Specific(MaskValue), m_Value(B))))
    return false;
 
  MainAnd = I->getOperand(0) == OtherAnd ? cast<Instruction>(I->getOperand(1))
                                         : cast<Instruction>(I->getOperand(0));
 
  // Both Ands should be in same basic block as Or
  if (I->getParent() != MainAnd->getParent() ||
      I->getParent() != OtherAnd->getParent())
    return false;
 
  return true;
}
 
bool SimplifyValuePattern(SmallVector<Value *> &Vec, bool AllowPoison) {
  size_t VecSize = Vec.size();
  if (VecSize == 1)
    return true;
  if (!isPowerOf2_64(VecSize))
    return false;
  size_t HalfVecSize = VecSize / 2;
 
  for (auto LHS = Vec.begin(), RHS = Vec.begin() + HalfVecSize;
       RHS != Vec.end(); LHS++, RHS++) {
    if (*LHS != nullptr && *RHS != nullptr) {
      if (*LHS == *RHS)
        continue;
      else
        return false;
    }
    if (!AllowPoison)
      return false;
    if (*LHS == nullptr && *RHS != nullptr)
      *LHS = *RHS;
  }
 
  Vec.resize(HalfVecSize);
  SimplifyValuePattern(Vec, AllowPoison);
  return true;
}
 
// Try to simplify dupqlane patterns like dupqlane(f32 A, f32 B, f32 A, f32 B)
// to dupqlane(f64(C)) where C is A concatenated with B
static std::optional<Instruction *> instCombineSVEDupqLane(InstCombiner &IC,
                                                           IntrinsicInst &II) {
  // An out-of-range index generates an all zero result. This means only an
  // index of zero has the same defined result across all vector lengths.
  if (!match(II.getOperand(1), m_Zero()))
    return std::nullopt;
 
  Value *CurrentInsertElt = nullptr, *Default = nullptr;
  if (!match(II.getOperand(0),
             m_Intrinsic<Intrinsic::vector_insert>(
                 m_Value(Default), m_Value(CurrentInsertElt), m_Zero())) ||
      !isa<FixedVectorType>(CurrentInsertElt->getType()))
    return std::nullopt;
  auto IIScalableTy = cast<ScalableVectorType>(II.getType());
 
  // Insert the scalars into a container ordered by InsertElement index
  SmallVector<Value *> Elts(IIScalableTy->getMinNumElements(), nullptr);
  while (auto InsertElt = dyn_cast<InsertElementInst>(CurrentInsertElt)) {
    auto Idx = cast<ConstantInt>(InsertElt->getOperand(2));
    Elts[Idx->getValue().getZExtValue()] = InsertElt->getOperand(1);
    CurrentInsertElt = InsertElt->getOperand(0);
  }
 
  bool AllowPoison =
      isa<PoisonValue>(CurrentInsertElt) && isa<PoisonValue>(Default);
  if (!SimplifyValuePattern(Elts, AllowPoison))
    return std::nullopt;
 
  // Rebuild the simplified chain of InsertElements. e.g. (a, b, a, b) as (a, b)
  Value *InsertEltChain = PoisonValue::get(CurrentInsertElt->getType());
  for (size_t I = 0; I < Elts.size(); I++) {
    if (Elts[I] == nullptr)
      continue;
    InsertEltChain = IC.Builder.CreateInsertElement(InsertEltChain, Elts[I],
                                                    IC.Builder.getInt64(I));
  }
  if (InsertEltChain == nullptr)
    return std::nullopt;
 
  // Splat the simplified sequence, e.g. (f16 a, f16 b, f16 c, f16 d) as one i64
  // value or (f16 a, f16 b) as one i32 value. This requires an InsertSubvector
  // be bitcast to a type wide enough to fit the sequence, be splatted, and then
  // be narrowed back to the original type.
  unsigned PatternWidth = IIScalableTy->getScalarSizeInBits() * Elts.size();
  unsigned PatternElementCount = IIScalableTy->getScalarSizeInBits() *
                                 IIScalableTy->getMinNumElements() /
                                 PatternWidth;
 
  IntegerType *WideTy = IC.Builder.getIntNTy(PatternWidth);
  auto *WideScalableTy = ScalableVectorType::get(WideTy, PatternElementCount);
  auto *WideShuffleMaskTy =
      ScalableVectorType::get(IC.Builder.getInt32Ty(), PatternElementCount);
 
  auto InsertSubvector = IC.Builder.CreateInsertVector(
      II.getType(), PoisonValue::get(II.getType()), InsertEltChain,
      uint64_t(0));
  auto WideBitcast =
      IC.Builder.CreateBitOrPointerCast(InsertSubvector, WideScalableTy);
  auto WideShuffleMask = ConstantAggregateZero::get(WideShuffleMaskTy);
  auto WideShuffle = IC.Builder.CreateShuffleVector(
      WideBitcast, PoisonValue::get(WideScalableTy), WideShuffleMask);
  auto NarrowBitcast =
      IC.Builder.CreateBitOrPointerCast(WideShuffle, II.getType());
 
  return IC.replaceInstUsesWith(II, NarrowBitcast);
}
 
static std::optional<Instruction *> instCombineMaxMinNM(InstCombiner &IC,
                                                        IntrinsicInst &II) {
  Value *A = II.getArgOperand(0);
  Value *B = II.getArgOperand(1);
  if (A == B)
    return IC.replaceInstUsesWith(II, A);
 
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVESrshl(InstCombiner &IC,
                                                        IntrinsicInst &II) {
  Value *Pred = II.getOperand(0);
  Value *Vec = II.getOperand(1);
  Value *Shift = II.getOperand(2);
 
  // Convert SRSHL into the simpler LSL intrinsic when fed by an ABS intrinsic.
  Value *AbsPred, *MergedValue;
  if (!match(Vec, m_Intrinsic<Intrinsic::aarch64_sve_sqabs>(
                      m_Value(MergedValue), m_Value(AbsPred), m_Value())) &&
      !match(Vec, m_Intrinsic<Intrinsic::aarch64_sve_abs>(
                      m_Value(MergedValue), m_Value(AbsPred), m_Value())))
 
    return std::nullopt;
 
  // Transform is valid if any of the following are true:
  // * The ABS merge value is an undef or non-negative
  // * The ABS predicate is all active
  // * The ABS predicate and the SRSHL predicates are the same
  if (!isa<UndefValue>(MergedValue) && !match(MergedValue, m_NonNegative()) &&
      AbsPred != Pred && !isAllActivePredicate(AbsPred))
    return std::nullopt;
 
  // Only valid when the shift amount is non-negative, otherwise the rounding
  // behaviour of SRSHL cannot be ignored.
  if (!match(Shift, m_NonNegative()))
    return std::nullopt;
 
  auto LSL = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_lsl,
                                        {II.getType()}, {Pred, Vec, Shift});
 
  return IC.replaceInstUsesWith(II, LSL);
}
 
static std::optional<Instruction *> instCombineSVEInsr(InstCombiner &IC,
                                                       IntrinsicInst &II) {
  Value *Vec = II.getOperand(0);
 
  if (getSplatValue(Vec) == II.getOperand(1))
    return IC.replaceInstUsesWith(II, Vec);
 
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineDMB(InstCombiner &IC, IntrinsicInst &II,
               unsigned LookaheadThreshold) {
  // If this barrier is post-dominated by identical one we can remove it
  auto *NI = II.getNextNode();
  auto CanSkipOver = [](Instruction *I) {
    return !I->mayReadOrWriteMemory() && !I->mayHaveSideEffects();
  };
  while (LookaheadThreshold-- && CanSkipOver(NI)) {
    auto *NIBB = NI->getParent();
    NI = NI->getNextNode();
    if (!NI) {
      if (auto *SuccBB = NIBB->getUniqueSuccessor())
        NI = &*SuccBB->getFirstNonPHIOrLifetime();
      else
        break;
    }
  }
  auto *NextII = dyn_cast_or_null<IntrinsicInst>(NI);
  if (NextII && II.isIdenticalTo(NextII))
    return IC.eraseInstFromFunction(II);
 
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineWhilelo(InstCombiner &IC,
                                                       IntrinsicInst &II) {
  return IC.replaceInstUsesWith(
      II,
      IC.Builder.CreateIntrinsic(Intrinsic::get_active_lane_mask,
                                 {II.getType(), II.getOperand(0)->getType()},
                                 {II.getOperand(0), II.getOperand(1)}));
}
 
static std::optional<Instruction *> instCombinePTrue(InstCombiner &IC,
                                                     IntrinsicInst &II) {
  unsigned PredPattern = cast<ConstantInt>(II.getOperand(0))->getZExtValue();
  // SVE vector length is a power-of-two, thus pow2 is synonymous with all.
  if (PredPattern == AArch64SVEPredPattern::all ||
      PredPattern == AArch64SVEPredPattern::pow2)
    return IC.replaceInstUsesWith(II, ConstantInt::getTrue(II.getType()));
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVEUxt(InstCombiner &IC,
                                                      IntrinsicInst &II,
                                                      unsigned NumBits) {
  Value *Passthru = II.getOperand(0);
  Value *Pg = II.getOperand(1);
  Value *Op = II.getOperand(2);
 
  // Convert UXT[BHW] to AND.
  if (isa<UndefValue>(Passthru) || isAllActivePredicate(Pg)) {
    auto *Ty = cast<VectorType>(II.getType());
    auto MaskValue = APInt::getLowBitsSet(Ty->getScalarSizeInBits(), NumBits);
    auto *Mask = ConstantInt::get(Ty, MaskValue);
    auto *And = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_and_u, {Ty},
                                           {Pg, Op, Mask});
    return IC.replaceInstUsesWith(II, And);
  }
 
  return std::nullopt;
}
 
static std::optional<Instruction *>
instCombineInStreamingMode(InstCombiner &IC, IntrinsicInst &II) {
  SMEAttrs FnSMEAttrs(*II.getFunction());
  bool IsStreaming = FnSMEAttrs.hasStreamingInterfaceOrBody();
  if (IsStreaming || !FnSMEAttrs.hasStreamingCompatibleInterface())
    return IC.replaceInstUsesWith(
        II, ConstantInt::getBool(II.getType(), IsStreaming));
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVEUMin(InstCombiner &IC,
                                                       IntrinsicInst &II) {
  // umin(umin(A, 1), umin(B, 1)) -> umin(umin(A,B), 1)
  constexpr Intrinsic::ID UMinID = Intrinsic::aarch64_sve_umin_u;
  Value *A, *B;
  Value *Pg = II.getOperand(0);
  if (match(II.getOperand(1), m_OneUse(m_Intrinsic<UMinID>(
                                  m_Specific(Pg), m_Value(A), m_One()))) &&
      match(II.getOperand(2), m_OneUse(m_Intrinsic<UMinID>(
                                  m_Specific(Pg), m_Value(B), m_One())))) {
    Value *NewUMin =
        IC.Builder.CreateIntrinsic(UMinID, II.getType(), {Pg, A, B});
    Value *NewLogicalUMin = IC.Builder.CreateIntrinsic(
        UMinID, II.getType(), {Pg, NewUMin, ConstantInt::get(II.getType(), 1)});
    return IC.replaceInstUsesWith(II, NewLogicalUMin);
  }
 
  // umin(umin(A, 1), 1) -> umin(A, 1)
  if (match(II.getOperand(1),
            m_Intrinsic<UMinID>(m_Specific(Pg), m_Value(), m_One())) &&
      match(II.getOperand(2), m_One()))
    return IC.replaceInstUsesWith(II, II.getOperand(1));
 
  return std::nullopt;
}
 
static std::optional<Instruction *> instCombineSVEOrr(InstCombiner &IC,
                                                      IntrinsicInst &II) {
  // orr(umin(A, 1), umin(B, 1)) -> umin(orr(A, B), 1)
  constexpr Intrinsic::ID UMinID = Intrinsic::aarch64_sve_umin_u;
  Value *Pg = II.getOperand(0);
 
  Value *A, *B;
  if (!match(II.getOperand(1), m_OneUse(m_Intrinsic<UMinID>(
                                   m_Specific(Pg), m_Value(A), m_One()))) ||
      !match(II.getOperand(2), m_OneUse(m_Intrinsic<UMinID>(
                                   m_Specific(Pg), m_Value(B), m_One()))))
    return std::nullopt;
 
  Value *NewOrr = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_orr_u,
                                             II.getType(), {Pg, A, B});
  Value *NewUMin = IC.Builder.CreateIntrinsic(
      UMinID, II.getType(), {Pg, NewOrr, ConstantInt::get(II.getType(), 1)});
  return IC.replaceInstUsesWith(II, NewUMin);
}
 
static std::optional<Instruction *> instCombineSVEAnd(InstCombiner &IC,
                                                      IntrinsicInst &II) {
  // and(cmphs(pg, ConstA, A), cmphs(pg, A, ConstB))
  // ->
  // cmphs(pg, ConstA - ConstB, sub(pg, A, ConstB))
  constexpr Intrinsic::ID CmphsID = Intrinsic::aarch64_sve_cmphs;
  Value *Pg = II.getOperand(0);
  Value *LHS = II.getOperand(1);
  Value *RHS = II.getOperand(2);
 
  Value *A, *PgLHS, *PgRHS;
  uint64_t ConstA, ConstB;
  if (!match(LHS, m_Intrinsic<CmphsID>(m_Value(PgLHS), m_ConstantInt(ConstA),
                                       m_Value(A))) ||
      !match(RHS, m_Intrinsic<CmphsID>(m_Value(PgRHS), m_Specific(A),
                                       m_ConstantInt(ConstB))) ||
      !LHS->hasOneUser() || !RHS->hasOneUser())
    return std::nullopt;
 
  // Always false regardless of predication
  if (ConstB > ConstA)
    return IC.replaceInstUsesWith(II, Constant::getNullValue(II.getType()));
 
  // The predicate for both CMPHSs must match.
  // The predicate for the AND can either be equal to the CMPHS predicates, or
  // either of the CMPHS values.
  if (PgLHS != PgRHS || (Pg != LHS && Pg != RHS && Pg != PgLHS))
    return std::nullopt;
 
  Type *VecTy = A->getType();
  Constant *Base = ConstantInt::get(VecTy, ConstB);
  Value *Sub = IC.Builder.CreateIntrinsic(Intrinsic::aarch64_sve_sub_u, VecTy,
                                          {PgLHS, A, Base});
  Constant *Limit = ConstantInt::get(VecTy, ConstA - ConstB);
  Value *NewCmphs =
      IC.Builder.CreateIntrinsic(CmphsID, VecTy, {PgLHS, Limit, Sub});
 
  return IC.replaceInstUsesWith(II, NewCmphs);
}
 
std::optional<Instruction *>
AArch64TTIImpl::instCombineIntrinsic(InstCombiner &IC,
                                     IntrinsicInst &II) const {
  const SVEIntrinsicInfo &IInfo = constructSVEIntrinsicInfo(II);
  if (std::optional<Instruction *> I = simplifySVEIntrinsic(IC, II, IInfo))
    return I;
 
  Intrinsic::ID IID = II.getIntrinsicID();
  switch (IID) {
  default:
    break;
  case Intrinsic::aarch64_dmb:
    return instCombineDMB(IC, II, ST->getCLOpts().dmb_lookahead_threshold);
  case Intrinsic::aarch64_neon_fmaxnm:
  case Intrinsic::aarch64_neon_fminnm:
    return instCombineMaxMinNM(IC, II);
  case Intrinsic::aarch64_sve_convert_from_svbool:
    return instCombineConvertFromSVBool(IC, II);
  case Intrinsic::aarch64_sve_dup:
    return instCombineSVEDup(IC, II);
  case Intrinsic::aarch64_sve_dup_x:
    return instCombineSVEDupX(IC, II);
  case Intrinsic::aarch64_sve_cmpeq:
  case Intrinsic::aarch64_sve_cmpeq_wide:
    return instCombineXorSVECmpCC(IC, II);
  case Intrinsic::aarch64_sve_cmpne:
  case Intrinsic::aarch64_sve_cmpne_wide:
    return instCombineSVECmpNE(IC, II);
  case Intrinsic::aarch64_sve_rdffr:
    return instCombineRDFFR(IC, II);
  case Intrinsic::aarch64_sve_lasta:
  case Intrinsic::aarch64_sve_lastb:
    return instCombineSVELast(IC, II);
  case Intrinsic::aarch64_sve_clasta_n:
  case Intrinsic::aarch64_sve_clastb_n:
    return instCombineSVECondLast(IC, II);
  case Intrinsic::aarch64_sve_cntd:
    return instCombineSVECntElts(IC, II, 2);
  case Intrinsic::aarch64_sve_cntw:
    return instCombineSVECntElts(IC, II, 4);
  case Intrinsic::aarch64_sve_cnth:
    return instCombineSVECntElts(IC, II, 8);
  case Intrinsic::aarch64_sve_cntb:
    return instCombineSVECntElts(IC, II, 16);
  case Intrinsic::aarch64_sme_cntsd:
    return instCombineSMECntsd(IC, II, ST);
  case Intrinsic::aarch64_sve_ptest_any:
  case Intrinsic::aarch64_sve_ptest_first:
  case Intrinsic::aarch64_sve_ptest_last:
    return instCombineSVEPTest(IC, II);
  case Intrinsic::aarch64_sve_fadd:
    return instCombineSVEVectorFAdd(IC, II);
  case Intrinsic::aarch64_sve_fadd_u:
    return instCombineSVEVectorFAddU(IC, II);
  case Intrinsic::aarch64_sve_fmul_u:
    return instCombineSVEVectorBinOp(IC, II);
  case Intrinsic::aarch64_sve_fsub:
    return instCombineSVEVectorFSub(IC, II);
  case Intrinsic::aarch64_sve_fsub_u:
    return instCombineSVEVectorFSubU(IC, II);
  case Intrinsic::aarch64_sve_add:
    return instCombineSVEVectorAdd(IC, II);
  case Intrinsic::aarch64_sve_add_u:
    return instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul_u,
                                             Intrinsic::aarch64_sve_mla_u>(
        IC, II, true);
  case Intrinsic::aarch64_sve_mla_u:
    return instCombineSVEVectorMlaU(IC, II);
  case Intrinsic::aarch64_sve_sadalp:
  case Intrinsic::aarch64_sve_uadalp:
    return instCombineSVEPairwiseAddLong(IC, II);
  case Intrinsic::aarch64_sve_sub:
    return instCombineSVEVectorSub(IC, II);
  case Intrinsic::aarch64_sve_sub_u:
    return instCombineSVEVectorFuseMulAddSub<Intrinsic::aarch64_sve_mul_u,
                                             Intrinsic::aarch64_sve_mls_u>(
        IC, II, true);
  case Intrinsic::aarch64_sve_tbl:
    return instCombineSVETBL(IC, II);
  case Intrinsic::aarch64_sve_uunpkhi:
  case Intrinsic::aarch64_sve_uunpklo:
  case Intrinsic::aarch64_sve_sunpkhi:
  case Intrinsic::aarch64_sve_sunpklo:
    return instCombineSVEUnpack(IC, II);
  case Intrinsic::aarch64_sve_uzp1:
    return instCombineSVEUzp1(IC, II);
  case Intrinsic::aarch64_sve_zip1:
  case Intrinsic::aarch64_sve_zip2:
    return instCombineSVEZip(IC, II);
  case Intrinsic::aarch64_sve_ld1_gather_index:
    return instCombineLD1GatherIndex(IC, II);
  case Intrinsic::aarch64_sve_st1_scatter_index:
    return instCombineST1ScatterIndex(IC, II);
  case Intrinsic::aarch64_sve_ld1:
    return instCombineSVELD1(IC, II, DL);
  case Intrinsic::aarch64_sve_st1:
    return instCombineSVEST1(IC, II, DL);
  case Intrinsic::aarch64_sve_sdiv:
    return instCombineSVESDIV(IC, II);
  case Intrinsic::aarch64_sve_sel:
    return instCombineSVESel(IC, II);
  case Intrinsic::aarch64_sve_srshl:
    return instCombineSVESrshl(IC, II);
  case Intrinsic::aarch64_sve_dupq_lane:
    return instCombineSVEDupqLane(IC, II);
  case Intrinsic::aarch64_sve_insr:
    return instCombineSVEInsr(IC, II);
  case Intrinsic::aarch64_sve_whilelo:
    return instCombineWhilelo(IC, II);
  case Intrinsic::aarch64_sve_ptrue:
    return instCombinePTrue(IC, II);
  case Intrinsic::aarch64_sve_uxtb:
    return instCombineSVEUxt(IC, II, 8);
  case Intrinsic::aarch64_sve_uxth:
    return instCombineSVEUxt(IC, II, 16);
  case Intrinsic::aarch64_sve_uxtw:
    return instCombineSVEUxt(IC, II, 32);
  case Intrinsic::aarch64_sme_in_streaming_mode:
    return instCombineInStreamingMode(IC, II);
  case Intrinsic::aarch64_sve_umin_u:
    return instCombineSVEUMin(IC, II);
  case Intrinsic::aarch64_sve_orr_u:
    return instCombineSVEOrr(IC, II);
  case Intrinsic::aarch64_sve_and_z:
    return instCombineSVEAnd(IC, II);
  }
 
  return std::nullopt;
}
 
std::optional<Value *> AArch64TTIImpl::simplifyDemandedVectorEltsIntrinsic(
    InstCombiner &IC, IntrinsicInst &II, APInt OrigDemandedElts,
    APInt &UndefElts, APInt &UndefElts2, APInt &UndefElts3,
    std::function<void(Instruction *, unsigned, APInt, APInt &)>
        SimplifyAndSetOp) const {
  switch (II.getIntrinsicID()) {
  default:
    break;
  case Intrinsic::aarch64_neon_fcvtxn:
  case Intrinsic::aarch64_neon_rshrn:
  case Intrinsic::aarch64_neon_sqrshrn:
  case Intrinsic::aarch64_neon_sqrshrun:
  case Intrinsic::aarch64_neon_sqshrn:
  case Intrinsic::aarch64_neon_sqshrun:
  case Intrinsic::aarch64_neon_sqxtn:
  case Intrinsic::aarch64_neon_sqxtun:
  case Intrinsic::aarch64_neon_uqrshrn:
  case Intrinsic::aarch64_neon_uqshrn:
  case Intrinsic::aarch64_neon_uqxtn:
    SimplifyAndSetOp(&II, 0, OrigDemandedElts, UndefElts);
    break;
  }
 
  return std::nullopt;
}
 
bool AArch64TTIImpl::enableScalableVectorization() const {
  return ST->isSVEAvailable() ||
         (ST->isSVEorStreamingSVEAvailable() &&
          ST->getCLOpts().enable_scalable_autovec_in_streaming_mode);
}
 
TypeSize
AArch64TTIImpl::getRegisterBitWidth(TargetTransformInfo::RegisterKind K) const {
  switch (K) {
  case TargetTransformInfo::RGK_Scalar:
    return TypeSize::getFixed(64);
  case TargetTransformInfo::RGK_FixedWidthVector:
    if (ST->useSVEForFixedLengthVectors() &&
        (ST->isSVEAvailable() ||
         ST->getCLOpts().enable_fixedwidth_autovec_in_streaming_mode))
      return TypeSize::getFixed(
          std::max(ST->getMinSVEVectorSizeInBits(), 128u));
    else if (ST->isNeonAvailable())
      return TypeSize::getFixed(128);
    else
      return TypeSize::getFixed(0);
  case TargetTransformInfo::RGK_ScalableVector:
    if (ST->isSVEAvailable() ||
        (ST->isSVEorStreamingSVEAvailable() &&
         ST->getCLOpts().enable_scalable_autovec_in_streaming_mode))
      return TypeSize::getScalable(128);
    else
      return TypeSize::getScalable(0);
  }
  llvm_unreachable("Unsupported register kind");
}
 
bool AArch64TTIImpl::isSingleExtWideningInstruction(
    unsigned Opcode, Type *DstTy, ArrayRef<const Value *> Args,
    Type *SrcOverrideTy) const {
  // A helper that returns a vector type from the given type. The number of
  // elements in type Ty determines the vector width.
  auto toVectorTy = [&](Type *ArgTy) {
    return VectorType::get(ArgTy->getScalarType(),
                           cast<VectorType>(DstTy)->getElementCount());
  };
 
  // Exit early if DstTy is not a vector type whose elements are one of [i16,
  // i32, i64]. SVE doesn't generally have the same set of instructions to
  // perform an extend with the add/sub/mul. There are SMULLB style
  // instructions, but they operate on top/bottom, requiring some sort of lane
  // interleaving to be used with zext/sext.
  unsigned DstEltSize = DstTy->getScalarSizeInBits();
  if (!useNeonVector(DstTy) || Args.size() != 2 ||
      (DstEltSize != 16 && DstEltSize != 32 && DstEltSize != 64))
    return false;
 
  Type *SrcTy = SrcOverrideTy;
  switch (Opcode) {
  case Instruction::Add:   // UADDW(2), SADDW(2).
  case Instruction::Sub: { // USUBW(2), SSUBW(2).
    // The second operand needs to be an extend
    if (isa<SExtInst>(Args[1]) || isa<ZExtInst>(Args[1])) {
      if (!SrcTy)
        SrcTy =
            toVectorTy(cast<Instruction>(Args[1])->getOperand(0)->getType());
      break;
    }
 
    if (Opcode == Instruction::Sub)
      return false;
 
    // UADDW(2), SADDW(2) can be commutted.
    if (isa<SExtInst>(Args[0]) || isa<ZExtInst>(Args[0])) {
      if (!SrcTy)
        SrcTy =
            toVectorTy(cast<Instruction>(Args[0])->getOperand(0)->getType());
      break;
    }
    return false;
  }
  default:
    return false;
  }
 
  // Legalize the destination type and ensure it can be used in a widening
  // operation.
  auto DstTyL = getTypeLegalizationCost(DstTy);
  if (!DstTyL.second.isVector() || DstEltSize != DstTy->getScalarSizeInBits())
    return false;
 
  // Legalize the source type and ensure it can be used in a widening
  // operation.
  assert(SrcTy && "Expected some SrcTy");
  auto SrcTyL = getTypeLegalizationCost(SrcTy);
  unsigned SrcElTySize = SrcTyL.second.getScalarSizeInBits();
  if (!SrcTyL.second.isVector() || SrcElTySize != SrcTy->getScalarSizeInBits())
    return false;
 
  // Get the total number of vector elements in the legalized types.
  InstructionCost NumDstEls =
      DstTyL.first * DstTyL.second.getVectorMinNumElements();
  InstructionCost NumSrcEls =
      SrcTyL.first * SrcTyL.second.getVectorMinNumElements();
 
  // Return true if the legalized types have the same number of vector elements
  // and the destination element type size is twice that of the source type.
  return NumDstEls == NumSrcEls && 2 * SrcElTySize == DstEltSize;
}
 
Type *AArch64TTIImpl::isBinExtWideningInstruction(unsigned Opcode, Type *DstTy,
                                                  ArrayRef<const Value *> Args,
                                                  Type *SrcOverrideTy) const {
  if (Opcode != Instruction::Add && Opcode != Instruction::Sub &&
      Opcode != Instruction::Mul)
    return nullptr;
 
  // Exit early if DstTy is not a vector type whose elements are one of [i16,
  // i32, i64]. SVE doesn't generally have the same set of instructions to
  // perform an extend with the add/sub/mul. There are SMULLB style
  // instructions, but they operate on top/bottom, requiring some sort of lane
  // interleaving to be used with zext/sext.
  unsigned DstEltSize = DstTy->getScalarSizeInBits();
  if (!useNeonVector(DstTy) || Args.size() != 2 ||
      (DstEltSize != 16 && DstEltSize != 32 && DstEltSize != 64))
    return nullptr;
 
  auto getScalarSizeWithOverride = [&](const Value *V) {
    if (SrcOverrideTy)
      return SrcOverrideTy->getScalarSizeInBits();
    return cast<Instruction>(V)
        ->getOperand(0)
        ->getType()
        ->getScalarSizeInBits();
  };
 
  unsigned MaxEltSize = 0;
  if ((isa<SExtInst>(Args[0]) && isa<SExtInst>(Args[1])) ||
      (isa<ZExtInst>(Args[0]) && isa<ZExtInst>(Args[1]))) {
    unsigned EltSize0 = getScalarSizeWithOverride(Args[0]);
    unsigned EltSize1 = getScalarSizeWithOverride(Args[1]);
    MaxEltSize = std::max(EltSize0, EltSize1);
  } else if (isa<SExtInst, ZExtInst>(Args[0]) &&
             isa<SExtInst, ZExtInst>(Args[1])) {
    unsigned EltSize0 = getScalarSizeWithOverride(Args[0]);
    unsigned EltSize1 = getScalarSizeWithOverride(Args[1]);
    // mul(sext, zext) will become smull(sext, zext) if the extends are large
    // enough.
    if (EltSize0 >= DstEltSize / 2 || EltSize1 >= DstEltSize / 2)
      return nullptr;
    MaxEltSize = DstEltSize / 2;
  } else if (Opcode == Instruction::Mul &&
             (isa<ZExtInst>(Args[0]) || isa<ZExtInst>(Args[1]))) {
    // If one of the operands is a Zext and the other has enough zero bits
    // to be treated as unsigned, we can still generate a umull, meaning the
    // zext is free.
    KnownBits Known =
        computeKnownBits(isa<ZExtInst>(Args[0]) ? Args[1] : Args[0], DL);
    if (Args[0]->getType()->getScalarSizeInBits() -
            Known.Zero.countLeadingOnes() >
        DstTy->getScalarSizeInBits() / 2)
      return nullptr;
 
    MaxEltSize =
        getScalarSizeWithOverride(isa<ZExtInst>(Args[0]) ? Args[0] : Args[1]);
  } else
    return nullptr;
 
  if (MaxEltSize * 2 > DstEltSize)
    return nullptr;
 
  Type *ExtTy = DstTy->getWithNewBitWidth(MaxEltSize * 2);
  if (ExtTy->getPrimitiveSizeInBits() <= 64)
    return nullptr;
  return ExtTy;
}
 
// s/urhadd instructions implement the following pattern, making the
// extends free:
//   %x = add ((zext i8 -> i16), 1)
//   %y = (zext i8 -> i16)
//   trunc i16 (lshr (add %x, %y), 1) -> i8
//
bool AArch64TTIImpl::isExtPartOfAvgExpr(const Instruction *ExtUser, Type *Dst,
                                        Type *Src) const {
  // The source should be a legal vector type.
  if (!Src->isVectorTy() || !TLI->isTypeLegal(TLI->getValueType(DL, Src)) ||
      (Src->isScalableTy() && !ST->hasSVE2()))
    return false;
 
  if (ExtUser->getOpcode() != Instruction::Add || !ExtUser->hasOneUse())
    return false;
 
  // Look for trunc/shl/add before trying to match the pattern.
  const Instruction *Add = ExtUser;
  auto *AddUser =
      dyn_cast_or_null<Instruction>(Add->getUniqueUndroppableUser());
  if (AddUser && AddUser->getOpcode() == Instruction::Add)
    Add = AddUser;
 
  auto *Shr = dyn_cast_or_null<Instruction>(Add->getUniqueUndroppableUser());
  if (!Shr || Shr->getOpcode() != Instruction::LShr)
    return false;
 
  auto *Trunc = dyn_cast_or_null<Instruction>(Shr->getUniqueUndroppableUser());
  if (!Trunc || Trunc->getOpcode() != Instruction::Trunc ||
      Src->getScalarSizeInBits() !=
          cast<CastInst>(Trunc)->getDestTy()->getScalarSizeInBits())
    return false;
 
  // Try to match the whole pattern. Ext could be either the first or second
  // m_ZExtOrSExt matched.
  Instruction *Ex1, *Ex2;
  if (!(match(Add, m_c_Add(m_Instruction(Ex1),
                           m_c_Add(m_Instruction(Ex2), m_One())))))
    return false;
 
  // Ensure both extends are of the same type
  if (match(Ex1, m_ZExtOrSExt(m_Value())) &&
      Ex1->getOpcode() == Ex2->getOpcode())
    return true;
 
  return false;
}
 
InstructionCost AArch64TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
                                                 Type *Src,
                                                 TTI::CastContextHint CCH,
                                                 TTI::TargetCostKind CostKind,
                                                 const Instruction *I) const {
  int ISD = TLI->InstructionOpcodeToISD(Opcode);
  assert(ISD && "Invalid opcode");
 
  // Codegen is not able to select all <vscale x 1 x Ty> casts yet.
  if (auto *VTy = dyn_cast<ScalableVectorType>(Dst);
      VTy && VTy->getElementCount() == ElementCount::getScalable(1)) {
    if (!is_contained({ISD::TRUNCATE, ISD::SIGN_EXTEND, ISD::ZERO_EXTEND,
                       ISD::BITCAST, ISD::ADDRSPACECAST},
                      ISD))
      return InstructionCost::getInvalid();
  }
 
  // A non-power-of-2 i1 vector that fits in a single register has no bitmask
  // lowering, so its lanes are extracted and combined one by one.
  if (auto *SrcVTy = dyn_cast<FixedVectorType>(Src);
      ISD == ISD::BITCAST && SrcVTy && Dst->isIntegerTy() &&
      Src->isIntOrIntVectorTy(1) && !isPowerOf2_32(SrcVTy->getNumElements()) &&
      getTypeLegalizationCost(SrcVTy).first == 1) {
    unsigned NumElts = SrcVTy->getNumElements();
    return getScalarizationOverhead(SrcVTy, APInt::getAllOnes(NumElts),
                                    /*Insert=*/false, /*Extract=*/true,
                                    CostKind) +
           NumElts - 1;
  }
 
  // If the cast is observable, and it is used by a widening instruction (e.g.,
  // uaddl, saddw, etc.), it may be free.
  if (I && !I->users().empty()) {
    // Determine whether Usr can absorb the cast from Src to Dst into a
    // widening instruction (e.g. uaddl, saddw, urhadd), making the cast free
    // with respect to that user.
    auto GetUserAbsorbedCastCost =
        [&](const Instruction *Usr) -> std::optional<InstructionCost> {
      SmallVector<const Value *, 4> Operands(Usr->operand_values());
 
      if (Type *ExtTy = isBinExtWideningInstruction(
              Usr->getOpcode(), Dst, Operands,
              Src != I->getOperand(0)->getType() ? Src : nullptr)) {
        // The cost from Src->Src*2 needs to be added if required, the cost
        // from Src*2->ExtTy is free.
        if (ExtTy->getScalarSizeInBits() > Src->getScalarSizeInBits() * 2) {
          Type *DoubleSrcTy =
              Src->getWithNewBitWidth(Src->getScalarSizeInBits() * 2);
          return getCastInstrCost(Opcode, DoubleSrcTy, Src,
                                  TTI::CastContextHint::None, CostKind);
        }
 
        return 0;
      }
 
      if (isSingleExtWideningInstruction(
              Usr->getOpcode(), Dst, Operands,
              Src != I->getOperand(0)->getType() ? Src : nullptr)) {
        // For adds only count the second operand as free if both operands
        // are extends but not the same operation. (i.e both operands are
        // not free in add(sext, zext)).
        if (Usr->getOpcode() == Instruction::Add) {
          if (I == Usr->getOperand(1) ||
              (isa<CastInst>(Usr->getOperand(1)) &&
               cast<CastInst>(Usr->getOperand(1))->getOpcode() == Opcode))
            return 0;
        } else {
          // Others are free so long as isSingleExtWideningInstruction
          // returned true.
          return 0;
        }
      }
 
      if (!isa<ZExtInst, SExtInst>(I))
        return std::nullopt;
 
      // The cast will be free for the s/urhadd instructions
      if (isExtPartOfAvgExpr(Usr, Dst, Src))
        return 0;
 
      // [US]SHLL/SHLL fold a doubling extend into a shift left by a splat
      // immediate of at most the source element width.
      const APInt *ShAmt;
      unsigned SrcBits = Src->getScalarSizeInBits();
      if (match(Usr, m_Shl(m_Specific(I), m_APInt(ShAmt))) &&
          ShAmt->ule(SrcBits) && useNeonVector(Dst) &&
          Dst->getScalarSizeInBits() == 2 * SrcBits &&
          getTypeLegalizationCost(Src).second.getScalarSizeInBits() == SrcBits)
        return 0;
 
      return std::nullopt;
    };
 
    InstructionCost MaxAbsorbedCost = 0;
    bool AllUsersAbsorbCast = true;
    for (const User *U : I->users()) {
      auto *Usr = cast<Instruction>(U);
      std::optional<InstructionCost> UserCost = GetUserAbsorbedCastCost(Usr);
      if (!UserCost) {
        AllUsersAbsorbCast = false;
        break;
      }
      MaxAbsorbedCost = std::max(MaxAbsorbedCost, *UserCost);
    }
 
    if (AllUsersAbsorbCast)
      return MaxAbsorbedCost;
  }
 
  // A vector compare already produces a lane mask (CMxx/FCMxx), so sign
  // extending it to the compare width is free. That only holds while the
  // result stays a vector; scalarized lanes still need a CSETM each.
  const Instruction *Cmp;
  const Value *CmpOp;
  if (I && useNeonVector(Dst) &&
      match(I, m_SExt(m_Instruction(Cmp, m_Cmp(m_Value(CmpOp), m_Value())))) &&
      Cmp->getParent() == I->getParent() &&
      CmpOp->getType()->getScalarSizeInBits() == Dst->getScalarSizeInBits() &&
      getTypeLegalizationCost(Dst).second.isVector())
    return 0;
 
  EVT SrcTy = TLI->getValueType(DL, Src);
  EVT DstTy = TLI->getValueType(DL, Dst);
 
  // SVE has no elements wider than 64 bits. The generic cost for a bitcast
  // between unsupported scalable types can otherwise appear to be free.
  if ((SrcTy.isScalableVector() && SrcTy.getScalarSizeInBits() > 64) ||
      (DstTy.isScalableVector() && DstTy.getScalarSizeInBits() > 64))
    return InstructionCost::getInvalid();
 
  // A bitcast that crosses register banks is costed as a vector insert/extract.
  // Between the same register bank it is free. Note: Some costs are not handled
  // well by this, like i1 vectors that need to be packed.
  if (ISD == ISD::BITCAST && ST->isLittleEndian()) {
    std::pair<InstructionCost, MVT> SrcLT = getTypeLegalizationCost(Src);
    std::pair<InstructionCost, MVT> DstLT = getTypeLegalizationCost(Dst);
    bool SrcIsFloatOrVec =
        SrcLT.second.isVector() || SrcLT.second.isFloatingPoint();
    bool DstIsFloatOrVec =
        DstLT.second.isVector() || DstLT.second.isFloatingPoint();
 
    // Consider the bitcast free if they are both the same type and have the
    // same size after legalization. Scalable vectors are handled specially as
    // the type legalization cost of v2i16 is v2i64.
    if (SrcIsFloatOrVec == DstIsFloatOrVec &&
        (SrcLT.second.getSizeInBits() == DstLT.second.getSizeInBits() ||
         (SrcLT.second.isScalableVector() &&
          SrcLT.second.getScalarSizeInBits() != 1 &&
          DstLT.second.getScalarSizeInBits() != 1)))
      return 0;
 
    return std::max(SrcLT.first, DstLT.first) *
           (CostKind == TTI::TCK_CodeSize
                ? 1
                : ST->getVectorInsertExtractBaseCost());
  }
 
  // From a vector to a scalarized vector will be an series of extract-element
  // and extends.
  if ((ISD == ISD::ZERO_EXTEND || ISD == ISD::SIGN_EXTEND) &&
      DstTy.getScalarSizeInBits() > 64 && SrcTy.getScalarSizeInBits() <= 64 &&
      DstTy.isFixedLengthVector()) {
    InstructionCost LaneCost = getVectorInstrCost(
        Instruction::ExtractElement, Src, CostKind, -1, nullptr, nullptr);
    InstructionCost ExtCost = getCastInstrCost(
        Opcode, Dst->getScalarType(), Src->getScalarType(), CCH, CostKind);
    return DstTy.getVectorNumElements() * (LaneCost + ExtCost);
  }
 
  if (!SrcTy.isSimple() || !DstTy.isSimple())
    return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
 
  // For the moment we do not have lowering for SVE1-only fptrunc f64->bf16 as
  // we use fcvtx under SVE2. Give them invalid costs.
  if (!ST->hasSVE2() && !ST->isStreamingSVEAvailable() &&
      ISD == ISD::FP_ROUND && SrcTy.isScalableVector() &&
      DstTy.getScalarType() == MVT::bf16 && SrcTy.getScalarType() == MVT::f64)
    return InstructionCost::getInvalid();
 
  static const TypeConversionCostTblEntry BF16Tbl[] = {
      {ISD::FP_ROUND, MVT::bf16, MVT::f32, 1},     // bfcvt
      {ISD::FP_ROUND, MVT::bf16, MVT::f64, 1},     // bfcvt
      {ISD::FP_ROUND, MVT::v4bf16, MVT::v4f32, 1}, // bfcvtn
      {ISD::FP_ROUND, MVT::v8bf16, MVT::v8f32, 2}, // bfcvtn+bfcvtn2
      {ISD::FP_ROUND, MVT::v2bf16, MVT::v2f64, 2}, // bfcvtn+fcvtn
      {ISD::FP_ROUND, MVT::v4bf16, MVT::v4f64, 3}, // fcvtn+fcvtl2+bfcvtn
      {ISD::FP_ROUND, MVT::v8bf16, MVT::v8f64, 6}, // 2 * fcvtn+fcvtn2+bfcvtn
      {ISD::FP_ROUND, MVT::nxv2bf16, MVT::nxv2f32, 1},  // bfcvt
      {ISD::FP_ROUND, MVT::nxv4bf16, MVT::nxv4f32, 1},  // bfcvt
      {ISD::FP_ROUND, MVT::nxv8bf16, MVT::nxv8f32, 3},  // bfcvt+bfcvt+uzp1
      {ISD::FP_ROUND, MVT::nxv2bf16, MVT::nxv2f64, 2},  // fcvtx+bfcvt
      {ISD::FP_ROUND, MVT::nxv4bf16, MVT::nxv4f64, 5},  // 2*fcvtx+2*bfcvt+uzp1
      {ISD::FP_ROUND, MVT::nxv8bf16, MVT::nxv8f64, 11}, // 4*fcvt+4*bfcvt+3*uzp
  };
 
  if (ST->hasBF16())
    if (const auto *Entry = ConvertCostTableLookup(
            BF16Tbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
      return Entry->Cost;
 
  // We have to estimate a cost of fixed length operation upon
  // SVE registers(operations) with the number of registers required
  // for a fixed type to be represented upon SVE registers.
  EVT WiderTy = SrcTy.bitsGT(DstTy) ? SrcTy : DstTy;
  if (SrcTy.isFixedLengthVector() && DstTy.isFixedLengthVector() &&
      SrcTy.getVectorNumElements() == DstTy.getVectorNumElements() &&
      ST->useSVEForFixedLengthVectors(WiderTy)) {
    std::pair<InstructionCost, MVT> LT =
        getTypeLegalizationCost(WiderTy.getTypeForEVT(Dst->getContext()));
    unsigned NumElements =
        AArch64::SVEBitsPerBlock / LT.second.getScalarSizeInBits();
    return LT.first *
           getCastInstrCost(
               Opcode,
               ScalableVectorType::get(Dst->getScalarType(), NumElements),
               ScalableVectorType::get(Src->getScalarType(), NumElements), CCH,
               CostKind, I);
  }
 
  // Symbolic constants for the SVE sitofp/uitofp entries in the table below
  // The cost of unpacking twice is artificially increased for now in order
  // to avoid regressions against NEON, which will use tbl instructions directly
  // instead of multiple layers of [s|u]unpk[lo|hi].
  // We use the unpacks in cases where the destination type is illegal and
  // requires splitting of the input, even if the input type itself is legal.
  const unsigned int SVE_EXT_COST = 1;
  const unsigned int SVE_FCVT_COST = 1;
  const unsigned int SVE_UNPACK_ONCE = 4;
  const unsigned int SVE_UNPACK_TWICE = 16;
 
  static const TypeConversionCostTblEntry ConversionTbl[] = {
      {ISD::TRUNCATE, MVT::v2i8, MVT::v2i64, 1},    // xtn
      {ISD::TRUNCATE, MVT::v2i16, MVT::v2i64, 1},   // xtn
      {ISD::TRUNCATE, MVT::v2i32, MVT::v2i64, 1},   // xtn
      {ISD::TRUNCATE, MVT::v4i8, MVT::v4i32, 1},    // xtn
      {ISD::TRUNCATE, MVT::v4i8, MVT::v4i64, 3},    // 2 xtn + 1 uzp1
      {ISD::TRUNCATE, MVT::v4i16, MVT::v4i32, 1},   // xtn
      {ISD::TRUNCATE, MVT::v4i16, MVT::v4i64, 2},   // 1 uzp1 + 1 xtn
      {ISD::TRUNCATE, MVT::v4i32, MVT::v4i64, 1},   // 1 uzp1
      {ISD::TRUNCATE, MVT::v8i8, MVT::v8i16, 1},    // 1 xtn
      {ISD::TRUNCATE, MVT::v8i8, MVT::v8i32, 2},    // 1 uzp1 + 1 xtn
      {ISD::TRUNCATE, MVT::v8i8, MVT::v8i64, 4},    // 3 x uzp1 + xtn
      {ISD::TRUNCATE, MVT::v8i16, MVT::v8i32, 1},   // 1 uzp1
      {ISD::TRUNCATE, MVT::v8i16, MVT::v8i64, 3},   // 3 x uzp1
      {ISD::TRUNCATE, MVT::v8i32, MVT::v8i64, 2},   // 2 x uzp1
      {ISD::TRUNCATE, MVT::v16i8, MVT::v16i16, 1},  // uzp1
      {ISD::TRUNCATE, MVT::v16i8, MVT::v16i32, 3},  // (2 + 1) x uzp1
      {ISD::TRUNCATE, MVT::v16i8, MVT::v16i64, 7},  // (4 + 2 + 1) x uzp1
      {ISD::TRUNCATE, MVT::v16i16, MVT::v16i32, 2}, // 2 x uzp1
      {ISD::TRUNCATE, MVT::v16i16, MVT::v16i64, 6}, // (4 + 2) x uzp1
      {ISD::TRUNCATE, MVT::v16i32, MVT::v16i64, 4}, // 4 x uzp1
 
      // Truncations on nxvmiN
      {ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i8, 2},
      {ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i16, 2},
      {ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i32, 2},
      {ISD::TRUNCATE, MVT::nxv2i1, MVT::nxv2i64, 2},
      {ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i8, 2},
      {ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i16, 2},
      {ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i32, 2},
      {ISD::TRUNCATE, MVT::nxv4i1, MVT::nxv4i64, 5},
      {ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i8, 2},
      {ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i16, 2},
      {ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i32, 5},
      {ISD::TRUNCATE, MVT::nxv8i1, MVT::nxv8i64, 11},
      {ISD::TRUNCATE, MVT::nxv16i1, MVT::nxv16i8, 2},
      {ISD::TRUNCATE, MVT::nxv2i8, MVT::nxv2i16, 0},
      {ISD::TRUNCATE, MVT::nxv2i8, MVT::nxv2i32, 0},
      {ISD::TRUNCATE, MVT::nxv2i8, MVT::nxv2i64, 0},
      {ISD::TRUNCATE, MVT::nxv2i16, MVT::nxv2i32, 0},
      {ISD::TRUNCATE, MVT::nxv2i16, MVT::nxv2i64, 0},
      {ISD::TRUNCATE, MVT::nxv2i32, MVT::nxv2i64, 0},
      {ISD::TRUNCATE, MVT::nxv4i8, MVT::nxv4i16, 0},
      {ISD::TRUNCATE, MVT::nxv4i8, MVT::nxv4i32, 0},
      {ISD::TRUNCATE, MVT::nxv4i8, MVT::nxv4i64, 1},
      {ISD::TRUNCATE, MVT::nxv4i16, MVT::nxv4i32, 0},
      {ISD::TRUNCATE, MVT::nxv4i16, MVT::nxv4i64, 1},
      {ISD::TRUNCATE, MVT::nxv4i32, MVT::nxv4i64, 1},
      {ISD::TRUNCATE, MVT::nxv8i8, MVT::nxv8i16, 0},
      {ISD::TRUNCATE, MVT::nxv8i8, MVT::nxv8i32, 1},
      {ISD::TRUNCATE, MVT::nxv8i8, MVT::nxv8i64, 3},
      {ISD::TRUNCATE, MVT::nxv8i16, MVT::nxv8i32, 1},
      {ISD::TRUNCATE, MVT::nxv8i16, MVT::nxv8i64, 3},
      {ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i16, 1},
      {ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i32, 3},
      {ISD::TRUNCATE, MVT::nxv16i8, MVT::nxv16i64, 7},
 
      // The number of shll instructions for the extension.
      {ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i16, 3},
      {ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i16, 3},
      {ISD::SIGN_EXTEND, MVT::v4i64, MVT::v4i32, 2},
      {ISD::ZERO_EXTEND, MVT::v4i64, MVT::v4i32, 2},
      {ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 3},
      {ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 3},
      {ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i16, 2},
      {ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i16, 2},
      {ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i8, 7},
      {ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i8, 7},
      {ISD::SIGN_EXTEND, MVT::v8i64, MVT::v8i16, 6},
      {ISD::ZERO_EXTEND, MVT::v8i64, MVT::v8i16, 6},
      {ISD::SIGN_EXTEND, MVT::v16i16, MVT::v16i8, 2},
      {ISD::ZERO_EXTEND, MVT::v16i16, MVT::v16i8, 2},
      {ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 6},
      {ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 6},
 
      // FP Ext and trunc
      {ISD::FP_EXTEND, MVT::f64, MVT::f32, 1},     // fcvt
      {ISD::FP_EXTEND, MVT::v2f64, MVT::v2f32, 1}, // fcvtl
      {ISD::FP_EXTEND, MVT::v4f64, MVT::v4f32, 2}, // fcvtl+fcvtl2
      //   FP16
      {ISD::FP_EXTEND, MVT::f32, MVT::f16, 1},     // fcvt
      {ISD::FP_EXTEND, MVT::f64, MVT::f16, 1},     // fcvt
      {ISD::FP_EXTEND, MVT::v4f32, MVT::v4f16, 1}, // fcvtl
      {ISD::FP_EXTEND, MVT::v8f32, MVT::v8f16, 2}, // fcvtl+fcvtl2
      {ISD::FP_EXTEND, MVT::v2f64, MVT::v2f16, 2}, // fcvtl+fcvtl
      {ISD::FP_EXTEND, MVT::v4f64, MVT::v4f16, 3}, // fcvtl+fcvtl2+fcvtl
      {ISD::FP_EXTEND, MVT::v8f64, MVT::v8f16, 6}, // 2 * fcvtl+fcvtl2+fcvtl
      //   BF16 (uses shift)
      {ISD::FP_EXTEND, MVT::f32, MVT::bf16, 1},     // shl
      {ISD::FP_EXTEND, MVT::f64, MVT::bf16, 2},     // shl+fcvt
      {ISD::FP_EXTEND, MVT::v4f32, MVT::v4bf16, 1}, // shll
      {ISD::FP_EXTEND, MVT::v8f32, MVT::v8bf16, 2}, // shll+shll2
      {ISD::FP_EXTEND, MVT::v2f64, MVT::v2bf16, 2}, // shll+fcvtl
      {ISD::FP_EXTEND, MVT::v4f64, MVT::v4bf16, 3}, // shll+fcvtl+fcvtl2
      {ISD::FP_EXTEND, MVT::v8f64, MVT::v8bf16, 6}, // 2 * shll+fcvtl+fcvtl2
      // FP Ext and trunc
      {ISD::FP_ROUND, MVT::f32, MVT::f64, 1},     // fcvt
      {ISD::FP_ROUND, MVT::v2f32, MVT::v2f64, 1}, // fcvtn
      {ISD::FP_ROUND, MVT::v4f32, MVT::v4f64, 2}, // fcvtn+fcvtn2
      //   FP16
      {ISD::FP_ROUND, MVT::f16, MVT::f32, 1},     // fcvt
      {ISD::FP_ROUND, MVT::f16, MVT::f64, 1},     // fcvt
      {ISD::FP_ROUND, MVT::v4f16, MVT::v4f32, 1}, // fcvtn
      {ISD::FP_ROUND, MVT::v8f16, MVT::v8f32, 2}, // fcvtn+fcvtn2
      {ISD::FP_ROUND, MVT::v2f16, MVT::v2f64, 2}, // fcvtn+fcvtn
      {ISD::FP_ROUND, MVT::v4f16, MVT::v4f64, 3}, // fcvtn+fcvtn2+fcvtn
      {ISD::FP_ROUND, MVT::v8f16, MVT::v8f64, 6}, // 2 * fcvtn+fcvtn2+fcvtn
      //   BF16 (more complex, with +bf16 is handled above)
      {ISD::FP_ROUND, MVT::bf16, MVT::f32, 8}, // Expansion is ~8 insns
      {ISD::FP_ROUND, MVT::bf16, MVT::f64, 9}, // fcvtn + above
      {ISD::FP_ROUND, MVT::v2bf16, MVT::v2f32, 8},
      {ISD::FP_ROUND, MVT::v4bf16, MVT::v4f32, 8},
      {ISD::FP_ROUND, MVT::v8bf16, MVT::v8f32, 15},
      {ISD::FP_ROUND, MVT::v2bf16, MVT::v2f64, 9},
      {ISD::FP_ROUND, MVT::v4bf16, MVT::v4f64, 10},
      {ISD::FP_ROUND, MVT::v8bf16, MVT::v8f64, 19},
 
      // LowerVectorINT_TO_FP:
      {ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i32, 1},
      {ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i32, 1},
      {ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i64, 1},
      {ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i32, 1},
      {ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i32, 1},
      {ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i64, 1},
 
      // SVE: to nxv2f16
      {ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i16, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i32, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f16, MVT::nxv2i64, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i16, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i32, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f16, MVT::nxv2i64, SVE_FCVT_COST},
 
      // SVE: to nxv4f16
      {ISD::SINT_TO_FP, MVT::nxv4f16, MVT::nxv4i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv4f16, MVT::nxv4i16, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv4f16, MVT::nxv4i32, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f16, MVT::nxv4i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f16, MVT::nxv4i16, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f16, MVT::nxv4i32, SVE_FCVT_COST},
 
      // SVE: to nxv8f16
      {ISD::SINT_TO_FP, MVT::nxv8f16, MVT::nxv8i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv8f16, MVT::nxv8i16, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv8f16, MVT::nxv8i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv8f16, MVT::nxv8i16, SVE_FCVT_COST},
 
      // SVE: to nxv16f16
      {ISD::SINT_TO_FP, MVT::nxv16f16, MVT::nxv16i8,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv16f16, MVT::nxv16i8,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
 
      // Complex: to v2f32
      {ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i8, 3},
      {ISD::SINT_TO_FP, MVT::v2f32, MVT::v2i16, 3},
      {ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i8, 3},
      {ISD::UINT_TO_FP, MVT::v2f32, MVT::v2i16, 3},
 
      // SVE: to nxv2f32
      {ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i16, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i32, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f32, MVT::nxv2i64, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i16, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i32, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f32, MVT::nxv2i64, SVE_FCVT_COST},
 
      // Complex: to v4f32
      {ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i8, 4},
      {ISD::SINT_TO_FP, MVT::v4f32, MVT::v4i16, 2},
      {ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i8, 3},
      {ISD::UINT_TO_FP, MVT::v4f32, MVT::v4i16, 2},
 
      // SVE: to nxv4f32
      {ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i16, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i32, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f32, MVT::nxv4i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f32, MVT::nxv4i16, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv4f32, MVT::nxv4i32, SVE_FCVT_COST},
 
      // Complex: to v8f32
      {ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i8, 10},
      {ISD::SINT_TO_FP, MVT::v8f32, MVT::v8i16, 4},
      {ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i8, 10},
      {ISD::UINT_TO_FP, MVT::v8f32, MVT::v8i16, 4},
 
      // SVE: to nxv8f32
      {ISD::SINT_TO_FP, MVT::nxv8f32, MVT::nxv8i8,
       SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv8f32, MVT::nxv8i16,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv8f32, MVT::nxv8i8,
       SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv8f32, MVT::nxv8i16,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
 
      // SVE: to nxv16f32
      {ISD::SINT_TO_FP, MVT::nxv16f32, MVT::nxv16i8,
       SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv16f32, MVT::nxv16i8,
       SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
 
      // Complex: to v16f32
      {ISD::SINT_TO_FP, MVT::v16f32, MVT::v16i8, 21},
      {ISD::UINT_TO_FP, MVT::v16f32, MVT::v16i8, 21},
 
      // Complex: to v2f64
      {ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i8, 4},
      {ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i16, 4},
      {ISD::SINT_TO_FP, MVT::v2f64, MVT::v2i32, 2},
      {ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i8, 4},
      {ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i16, 4},
      {ISD::UINT_TO_FP, MVT::v2f64, MVT::v2i32, 2},
 
      // SVE: to nxv2f64
      {ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i16, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i32, SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv2f64, MVT::nxv2i64, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i8,
       SVE_EXT_COST + SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i16, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i32, SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv2f64, MVT::nxv2i64, SVE_FCVT_COST},
 
      // Complex: to v4f64
      {ISD::SINT_TO_FP, MVT::v4f64, MVT::v4i32, 4},
      {ISD::UINT_TO_FP, MVT::v4f64, MVT::v4i32, 4},
 
      // SVE: to nxv4f64
      {ISD::SINT_TO_FP, MVT::nxv4f64, MVT::nxv4i8,
       SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv4f64, MVT::nxv4i16,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv4f64, MVT::nxv4i32,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f64, MVT::nxv4i8,
       SVE_EXT_COST + SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f64, MVT::nxv4i16,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv4f64, MVT::nxv4i32,
       SVE_UNPACK_ONCE + 2 * SVE_FCVT_COST},
 
      // SVE: to nxv8f64
      {ISD::SINT_TO_FP, MVT::nxv8f64, MVT::nxv8i8,
       SVE_EXT_COST + SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
      {ISD::SINT_TO_FP, MVT::nxv8f64, MVT::nxv8i16,
       SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv8f64, MVT::nxv8i8,
       SVE_EXT_COST + SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
      {ISD::UINT_TO_FP, MVT::nxv8f64, MVT::nxv8i16,
       SVE_UNPACK_TWICE + 4 * SVE_FCVT_COST},
 
      // LowerVectorFP_TO_INT
      {ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f32, 1},
      {ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f32, 1},
      {ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f64, 1},
      {ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f32, 1},
      {ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f32, 1},
      {ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f64, 1},
 
      // Complex, from v2f32: legal type is v2i32 (no cost) or v2i64 (1 ext).
      {ISD::FP_TO_SINT, MVT::v2i64, MVT::v2f32, 2},
      {ISD::FP_TO_SINT, MVT::v2i16, MVT::v2f32, 1},
      {ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f32, 1},
      {ISD::FP_TO_UINT, MVT::v2i64, MVT::v2f32, 2},
      {ISD::FP_TO_UINT, MVT::v2i16, MVT::v2f32, 1},
      {ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f32, 1},
 
      // Complex, from v4f32: legal type is v4i16, 1 narrowing => ~2
      {ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f32, 2},
      {ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f32, 2},
      {ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f32, 2},
      {ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f32, 2},
 
      // Complex, from v2f64: legal type is v2i32, 1 narrowing => ~2.
      {ISD::FP_TO_SINT, MVT::v2i32, MVT::v2f64, 2},
      {ISD::FP_TO_SINT, MVT::v2i16, MVT::v2f64, 2},
      {ISD::FP_TO_SINT, MVT::v2i8, MVT::v2f64, 2},
      {ISD::FP_TO_UINT, MVT::v2i32, MVT::v2f64, 2},
      {ISD::FP_TO_UINT, MVT::v2i16, MVT::v2f64, 2},
      {ISD::FP_TO_UINT, MVT::v2i8, MVT::v2f64, 2},
 
      // Complex, from nxv2f32.
      {ISD::FP_TO_SINT, MVT::nxv2i64, MVT::nxv2f32, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i32, MVT::nxv2f32, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i16, MVT::nxv2f32, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i8, MVT::nxv2f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i64, MVT::nxv2f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i32, MVT::nxv2f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i16, MVT::nxv2f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i8, MVT::nxv2f32, 1},
 
      // Complex, from nxv2f64.
      {ISD::FP_TO_SINT, MVT::nxv2i64, MVT::nxv2f64, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i32, MVT::nxv2f64, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i16, MVT::nxv2f64, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i8, MVT::nxv2f64, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i1, MVT::nxv2f64, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i64, MVT::nxv2f64, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i32, MVT::nxv2f64, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i16, MVT::nxv2f64, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i8, MVT::nxv2f64, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i1, MVT::nxv2f64, 1},
 
      // Complex, from nxv4f32.
      {ISD::FP_TO_SINT, MVT::nxv4i64, MVT::nxv4f32, 4},
      {ISD::FP_TO_SINT, MVT::nxv4i32, MVT::nxv4f32, 1},
      {ISD::FP_TO_SINT, MVT::nxv4i16, MVT::nxv4f32, 1},
      {ISD::FP_TO_SINT, MVT::nxv4i8, MVT::nxv4f32, 1},
      {ISD::FP_TO_SINT, MVT::nxv4i1, MVT::nxv4f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv4i64, MVT::nxv4f32, 4},
      {ISD::FP_TO_UINT, MVT::nxv4i32, MVT::nxv4f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv4i16, MVT::nxv4f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv4i8, MVT::nxv4f32, 1},
      {ISD::FP_TO_UINT, MVT::nxv4i1, MVT::nxv4f32, 1},
 
      // Complex, from nxv8f64. Illegal -> illegal conversions not required.
      {ISD::FP_TO_SINT, MVT::nxv8i16, MVT::nxv8f64, 7},
      {ISD::FP_TO_SINT, MVT::nxv8i8, MVT::nxv8f64, 7},
      {ISD::FP_TO_UINT, MVT::nxv8i16, MVT::nxv8f64, 7},
      {ISD::FP_TO_UINT, MVT::nxv8i8, MVT::nxv8f64, 7},
 
      // Complex, from nxv4f64. Illegal -> illegal conversions not required.
      {ISD::FP_TO_SINT, MVT::nxv4i32, MVT::nxv4f64, 3},
      {ISD::FP_TO_SINT, MVT::nxv4i16, MVT::nxv4f64, 3},
      {ISD::FP_TO_SINT, MVT::nxv4i8, MVT::nxv4f64, 3},
      {ISD::FP_TO_UINT, MVT::nxv4i32, MVT::nxv4f64, 3},
      {ISD::FP_TO_UINT, MVT::nxv4i16, MVT::nxv4f64, 3},
      {ISD::FP_TO_UINT, MVT::nxv4i8, MVT::nxv4f64, 3},
 
      // Complex, from nxv8f32. Illegal -> illegal conversions not required.
      {ISD::FP_TO_SINT, MVT::nxv8i16, MVT::nxv8f32, 3},
      {ISD::FP_TO_SINT, MVT::nxv8i8, MVT::nxv8f32, 3},
      {ISD::FP_TO_UINT, MVT::nxv8i16, MVT::nxv8f32, 3},
      {ISD::FP_TO_UINT, MVT::nxv8i8, MVT::nxv8f32, 3},
 
      // Complex, from nxv8f16.
      {ISD::FP_TO_SINT, MVT::nxv8i64, MVT::nxv8f16, 10},
      {ISD::FP_TO_SINT, MVT::nxv8i32, MVT::nxv8f16, 4},
      {ISD::FP_TO_SINT, MVT::nxv8i16, MVT::nxv8f16, 1},
      {ISD::FP_TO_SINT, MVT::nxv8i8, MVT::nxv8f16, 1},
      {ISD::FP_TO_SINT, MVT::nxv8i1, MVT::nxv8f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv8i64, MVT::nxv8f16, 10},
      {ISD::FP_TO_UINT, MVT::nxv8i32, MVT::nxv8f16, 4},
      {ISD::FP_TO_UINT, MVT::nxv8i16, MVT::nxv8f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv8i8, MVT::nxv8f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv8i1, MVT::nxv8f16, 1},
 
      // Complex, from nxv4f16.
      {ISD::FP_TO_SINT, MVT::nxv4i64, MVT::nxv4f16, 4},
      {ISD::FP_TO_SINT, MVT::nxv4i32, MVT::nxv4f16, 1},
      {ISD::FP_TO_SINT, MVT::nxv4i16, MVT::nxv4f16, 1},
      {ISD::FP_TO_SINT, MVT::nxv4i8, MVT::nxv4f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv4i64, MVT::nxv4f16, 4},
      {ISD::FP_TO_UINT, MVT::nxv4i32, MVT::nxv4f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv4i16, MVT::nxv4f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv4i8, MVT::nxv4f16, 1},
 
      // Complex, from nxv2f16.
      {ISD::FP_TO_SINT, MVT::nxv2i64, MVT::nxv2f16, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i32, MVT::nxv2f16, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i16, MVT::nxv2f16, 1},
      {ISD::FP_TO_SINT, MVT::nxv2i8, MVT::nxv2f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i64, MVT::nxv2f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i32, MVT::nxv2f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i16, MVT::nxv2f16, 1},
      {ISD::FP_TO_UINT, MVT::nxv2i8, MVT::nxv2f16, 1},
 
      // Truncate from nxvmf32 to nxvmf16.
      {ISD::FP_ROUND, MVT::nxv2f16, MVT::nxv2f32, 1},
      {ISD::FP_ROUND, MVT::nxv4f16, MVT::nxv4f32, 1},
      {ISD::FP_ROUND, MVT::nxv8f16, MVT::nxv8f32, 3},
 
      // Truncate from nxvmf32 to nxvmbf16.
      {ISD::FP_ROUND, MVT::nxv2bf16, MVT::nxv2f32, 8},
      {ISD::FP_ROUND, MVT::nxv4bf16, MVT::nxv4f32, 8},
      {ISD::FP_ROUND, MVT::nxv8bf16, MVT::nxv8f32, 17},
 
      // Truncate from nxvmf64 to nxvmf16.
      {ISD::FP_ROUND, MVT::nxv2f16, MVT::nxv2f64, 1},
      {ISD::FP_ROUND, MVT::nxv4f16, MVT::nxv4f64, 3},
      {ISD::FP_ROUND, MVT::nxv8f16, MVT::nxv8f64, 7},
 
      // Truncate from nxvmf64 to nxvmbf16.
      {ISD::FP_ROUND, MVT::nxv2bf16, MVT::nxv2f64, 9},
      {ISD::FP_ROUND, MVT::nxv4bf16, MVT::nxv4f64, 19},
      {ISD::FP_ROUND, MVT::nxv8bf16, MVT::nxv8f64, 39},
 
      // Truncate from nxvmf64 to nxvmf32.
      {ISD::FP_ROUND, MVT::nxv2f32, MVT::nxv2f64, 1},
      {ISD::FP_ROUND, MVT::nxv4f32, MVT::nxv4f64, 3},
      {ISD::FP_ROUND, MVT::nxv8f32, MVT::nxv8f64, 6},
 
      // Extend from nxvmf16 to nxvmf32.
      {ISD::FP_EXTEND, MVT::nxv2f32, MVT::nxv2f16, 1},
      {ISD::FP_EXTEND, MVT::nxv4f32, MVT::nxv4f16, 1},
      {ISD::FP_EXTEND, MVT::nxv8f32, MVT::nxv8f16, 2},
 
      // Extend from nxvmbf16 to nxvmf32.
      {ISD::FP_EXTEND, MVT::nxv2f32, MVT::nxv2bf16, 1}, // lsl
      {ISD::FP_EXTEND, MVT::nxv4f32, MVT::nxv4bf16, 1}, // lsl
      {ISD::FP_EXTEND, MVT::nxv8f32, MVT::nxv8bf16, 4}, // unpck+unpck+lsl+lsl
 
      // Extend from nxvmf16 to nxvmf64.
      {ISD::FP_EXTEND, MVT::nxv2f64, MVT::nxv2f16, 1},
      {ISD::FP_EXTEND, MVT::nxv4f64, MVT::nxv4f16, 2},
      {ISD::FP_EXTEND, MVT::nxv8f64, MVT::nxv8f16, 4},
 
      // Extend from nxvmbf16 to nxvmf64.
      {ISD::FP_EXTEND, MVT::nxv2f64, MVT::nxv2bf16, 2},  // lsl+fcvt
      {ISD::FP_EXTEND, MVT::nxv4f64, MVT::nxv4bf16, 6},  // 2*unpck+2*lsl+2*fcvt
      {ISD::FP_EXTEND, MVT::nxv8f64, MVT::nxv8bf16, 14}, // 6*unpck+4*lsl+4*fcvt
 
      // Extend from nxvmf32 to nxvmf64.
      {ISD::FP_EXTEND, MVT::nxv2f64, MVT::nxv2f32, 1},
      {ISD::FP_EXTEND, MVT::nxv4f64, MVT::nxv4f32, 2},
      {ISD::FP_EXTEND, MVT::nxv8f64, MVT::nxv8f32, 6},
 
      // Add cost for extending to illegal -too wide- scalable vectors.
      // zero/sign extend are implemented by multiple unpack operations,
      // where each operation has a cost of 1.
      {ISD::ZERO_EXTEND, MVT::nxv16i16, MVT::nxv16i8, 2},
      {ISD::ZERO_EXTEND, MVT::nxv16i32, MVT::nxv16i8, 6},
      {ISD::ZERO_EXTEND, MVT::nxv16i64, MVT::nxv16i8, 14},
      {ISD::ZERO_EXTEND, MVT::nxv8i32, MVT::nxv8i16, 2},
      {ISD::ZERO_EXTEND, MVT::nxv8i64, MVT::nxv8i16, 6},
      {ISD::ZERO_EXTEND, MVT::nxv4i64, MVT::nxv4i32, 2},
 
      {ISD::SIGN_EXTEND, MVT::nxv16i16, MVT::nxv16i8, 2},
      {ISD::SIGN_EXTEND, MVT::nxv16i32, MVT::nxv16i8, 6},
      {ISD::SIGN_EXTEND, MVT::nxv16i64, MVT::nxv16i8, 14},
      {ISD::SIGN_EXTEND, MVT::nxv8i32, MVT::nxv8i16, 2},
      {ISD::SIGN_EXTEND, MVT::nxv8i64, MVT::nxv8i16, 6},
      {ISD::SIGN_EXTEND, MVT::nxv4i64, MVT::nxv4i32, 2},
  };
 
  if (const auto *Entry = ConvertCostTableLookup(
          ConversionTbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
    return Entry->Cost;
 
  static const TypeConversionCostTblEntry FP16Tbl[] = {
      {ISD::FP_TO_SINT, MVT::v4i8, MVT::v4f16, 1}, // fcvtzs
      {ISD::FP_TO_UINT, MVT::v4i8, MVT::v4f16, 1},
      {ISD::FP_TO_SINT, MVT::v4i16, MVT::v4f16, 1}, // fcvtzs
      {ISD::FP_TO_UINT, MVT::v4i16, MVT::v4f16, 1},
      {ISD::FP_TO_SINT, MVT::v4i32, MVT::v4f16, 2}, // fcvtl+fcvtzs
      {ISD::FP_TO_UINT, MVT::v4i32, MVT::v4f16, 2},
      {ISD::FP_TO_SINT, MVT::v8i8, MVT::v8f16, 2}, // fcvtzs+xtn
      {ISD::FP_TO_UINT, MVT::v8i8, MVT::v8f16, 2},
      {ISD::FP_TO_SINT, MVT::v8i16, MVT::v8f16, 1}, // fcvtzs
      {ISD::FP_TO_UINT, MVT::v8i16, MVT::v8f16, 1},
      {ISD::FP_TO_SINT, MVT::v8i32, MVT::v8f16, 4}, // 2*fcvtl+2*fcvtzs
      {ISD::FP_TO_UINT, MVT::v8i32, MVT::v8f16, 4},
      {ISD::FP_TO_SINT, MVT::v16i8, MVT::v16f16, 3}, // 2*fcvtzs+xtn
      {ISD::FP_TO_UINT, MVT::v16i8, MVT::v16f16, 3},
      {ISD::FP_TO_SINT, MVT::v16i16, MVT::v16f16, 2}, // 2*fcvtzs
      {ISD::FP_TO_UINT, MVT::v16i16, MVT::v16f16, 2},
      {ISD::FP_TO_SINT, MVT::v16i32, MVT::v16f16, 8}, // 4*fcvtl+4*fcvtzs
      {ISD::FP_TO_UINT, MVT::v16i32, MVT::v16f16, 8},
      {ISD::UINT_TO_FP, MVT::v8f16, MVT::v8i8, 2},   // ushll + ucvtf
      {ISD::SINT_TO_FP, MVT::v8f16, MVT::v8i8, 2},   // sshll + scvtf
      {ISD::UINT_TO_FP, MVT::v16f16, MVT::v16i8, 4}, // 2 * ushl(2) + 2 * ucvtf
      {ISD::SINT_TO_FP, MVT::v16f16, MVT::v16i8, 4}, // 2 * sshl(2) + 2 * scvtf
  };
 
  if (ST->hasFullFP16())
    if (const auto *Entry = ConvertCostTableLookup(
            FP16Tbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
      return Entry->Cost;
 
  // INT_TO_FP of i64->f32 will scalarize, which is required to avoid
  // double-rounding issues.
  if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) &&
      DstTy.getScalarType() == MVT::f32 && SrcTy.getScalarSizeInBits() > 32 &&
      isa<FixedVectorType>(Dst) && isa<FixedVectorType>(Src))
    return cast<FixedVectorType>(Dst)->getNumElements() *
               getCastInstrCost(Opcode, Dst->getScalarType(),
                                Src->getScalarType(), CCH, CostKind) +
           BaseT::getScalarizationOverhead(cast<FixedVectorType>(Src), false,
                                           true, CostKind) +
           BaseT::getScalarizationOverhead(cast<FixedVectorType>(Dst), true,
                                           false, CostKind);
 
  if ((ISD == ISD::ZERO_EXTEND || ISD == ISD::SIGN_EXTEND) &&
      CCH == TTI::CastContextHint::Masked &&
      ST->isSVEorStreamingSVEAvailable() &&
      TLI->getTypeAction(Src->getContext(), SrcTy) ==
          TargetLowering::TypePromoteInteger &&
      TLI->getTypeAction(Dst->getContext(), DstTy) ==
          TargetLowering::TypeSplitVector) {
    // The standard behaviour in the backend for these cases is to split the
    // extend up into two parts:
    //  1. Perform an extending load or masked load up to the legal type.
    //  2. Extend the loaded data to the final type.
    std::pair<InstructionCost, MVT> SrcLT = getTypeLegalizationCost(Src);
    Type *LegalTy = EVT(SrcLT.second).getTypeForEVT(Src->getContext());
    InstructionCost Part1 = AArch64TTIImpl::getCastInstrCost(
        Opcode, LegalTy, Src, CCH, CostKind, I);
    InstructionCost Part2 = AArch64TTIImpl::getCastInstrCost(
        Opcode, Dst, LegalTy, TTI::CastContextHint::None, CostKind, I);
    return Part1 + Part2;
  }
 
  // The BasicTTIImpl version only deals with CCH==TTI::CastContextHint::Normal,
  // but we also want to include the TTI::CastContextHint::Masked case too.
  if ((ISD == ISD::ZERO_EXTEND || ISD == ISD::SIGN_EXTEND) &&
      CCH == TTI::CastContextHint::Masked &&
      ST->isSVEorStreamingSVEAvailable() && TLI->isTypeLegal(DstTy))
    CCH = TTI::CastContextHint::Normal;
 
  return BaseT::getCastInstrCost(Opcode, Dst, Src, CCH, CostKind, I);
}
 
InstructionCost
AArch64TTIImpl::getExtractWithExtendCost(unsigned Opcode, Type *Dst,
                                         VectorType *VecTy, unsigned Index,
                                         TTI::TargetCostKind CostKind) const {
 
  // Make sure we were given a valid extend opcode.
  assert((Opcode == Instruction::SExt || Opcode == Instruction::ZExt) &&
         "Invalid opcode");
 
  // We are extending an element we extract from a vector, so the source type
  // of the extend is the element type of the vector.
  auto *Src = VecTy->getElementType();
 
  // Sign- and zero-extends are for integer types only.
  assert(isa<IntegerType>(Dst) && isa<IntegerType>(Src) && "Invalid type");
 
  // Get the cost for the extract. We compute the cost (if any) for the extend
  // below.
  InstructionCost Cost = getVectorInstrCost(Instruction::ExtractElement, VecTy,
                                            CostKind, Index, nullptr, nullptr);
 
  // Legalize the types.
  auto VecLT = getTypeLegalizationCost(VecTy);
  auto DstVT = TLI->getValueType(DL, Dst);
  auto SrcVT = TLI->getValueType(DL, Src);
 
  // If the resulting type is still a vector and the destination type is legal,
  // we may get the extension for free. If not, get the default cost for the
  // extend.
  if (!VecLT.second.isVector() || !TLI->isTypeLegal(DstVT))
    return Cost + getCastInstrCost(Opcode, Dst, Src, TTI::CastContextHint::None,
                                   CostKind);
 
  // The destination type should be larger than the element type. If not, get
  // the default cost for the extend.
  if (DstVT.getFixedSizeInBits() < SrcVT.getFixedSizeInBits())
    return Cost + getCastInstrCost(Opcode, Dst, Src, TTI::CastContextHint::None,
                                   CostKind);
 
  switch (Opcode) {
  default:
    llvm_unreachable("Opcode should be either SExt or ZExt");
 
  // For sign-extends, we only need a smov, which performs the extension
  // automatically.
  case Instruction::SExt:
    return Cost;
 
  // For zero-extends, the extend is performed automatically by a umov unless
  // the destination type is i64 and the element type is i8 or i16.
  case Instruction::ZExt:
    if (DstVT.getSizeInBits() != 64u || SrcVT.getSizeInBits() == 32u)
      return Cost;
  }
 
  // If we are unable to perform the extend for free, get the default cost.
  return Cost + getCastInstrCost(Opcode, Dst, Src, TTI::CastContextHint::None,
                                 CostKind);
}
 
InstructionCost AArch64TTIImpl::getCFInstrCost(unsigned Opcode,
                                               TTI::TargetCostKind CostKind,
                                               const Instruction *I) const {
  if (CostKind != TTI::TCK_RecipThroughput)
    return Opcode == Instruction::PHI ? 0 : 1;
  assert(CostKind == TTI::TCK_RecipThroughput && "unexpected CostKind");
  // Branches are assumed to be predicted.
  return 0;
}
 
InstructionCost AArch64TTIImpl::getVectorInstrCostHelper(
    unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, unsigned Index,
    const Instruction *I, Value *Scalar,
    ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx,
    TTI::VectorInstrContext VIC) const {
  assert(Ty->isVectorTy() && "This must be a vector type");
 
  if (Index != -1U) {
    // Legalize the type.
    std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
 
    // This type is legalized to a scalar type.
    if (!LT.second.isVector())
      return 0;
 
    // The type may be split. For fixed-width vectors we can normalize the
    // index to the new type.
    if (LT.second.isFixedLengthVector()) {
      unsigned Width = LT.second.getVectorNumElements();
      Index = Index % Width;
    }
 
    // The element at index zero is already inside the vector.
    // - For a insert-element or extract-element
    // instruction that extracts integers, an explicit FPR -> GPR move is
    // needed. So it has non-zero cost.
    if (Index == 0 && !Ty->getScalarType()->isIntegerTy())
      return 0;
 
    // SVE has no scalar move to an arbitrary lane above the low 128-bit portion
    // of a Z register, e.g. there is no equivalent of "mov z0.d[9], d0".
    // Fixed-length vectors wider than 128 bits therefore need
    // [splice]/index/pred/splat/cmp/pred-mov when scalarizing inserts for those
    // lanes, so model them as more expensive than ordinary NEON lane accesses.
    if (ST->useSVEForFixedLengthVectors()) {
      InstructionCost Cost = CostKind == TTI::TCK_CodeSize
                                 ? 1
                                 : ST->getVectorInsertExtractBaseCost();
      if (Index * Ty->getScalarSizeInBits() < 128)
        return Cost;
      if (Index * Ty->getScalarSizeInBits() < 512 &&
          Opcode == Instruction::ExtractElement)
        // Integer extracts (>128b, <512b) require extra mov from FPR -> GPR.
        return Ty->getScalarType()->isIntegerTy() ? Cost + 1 : Cost;
      if (Opcode == Instruction::ExtractElement)
        return Cost + 2; // cost of mov imm + whilels + lastb
      if (Opcode == Instruction::InsertElement)
        return Cost + 3; // cost of insert with cmp/splice
      llvm_unreachable("unexpected opcode");
    }
 
    // This is recognising a LD1 single-element structure to one lane of one
    // register instruction. I.e., if this is an `insertelement` instruction,
    // and its second operand is a load, then we will generate a LD1, which
    // are expensive instructions on some uArchs.
    if (VIC == TTI::VectorInstrContext::Load) {
      if (ST->hasFastLD1Single())
        return 0;
      return CostKind == TTI::TCK_CodeSize
                 ? 0
                 : ST->getVectorInsertExtractBaseCost() + 1;
    }
 
    // i1 inserts and extract will include an extra cset or cmp of the vector
    // value. Increase the cost by 1 to account.
    if (Ty->getScalarSizeInBits() == 1)
      return CostKind == TTI::TCK_CodeSize
                 ? 2
                 : ST->getVectorInsertExtractBaseCost() + 1;
 
    // FIXME:
    // If the extract-element and insert-element instructions could be
    // simplified away (e.g., could be combined into users by looking at use-def
    // context), they have no cost. This is not done in the first place for
    // compile-time considerations.
  }
 
  // In case of Neon, if there exists extractelement from lane != 0 such that
  // 1. extractelement does not necessitate a move from vector_reg -> GPR.
  // 2. extractelement result feeds into fmul.
  // 3. Other operand of fmul is an extractelement from lane 0 or lane
  // equivalent to 0.
  // then the extractelement can be merged with fmul in the backend and it
  // incurs no cost.
  // e.g.
  // define double @foo(<2 x double> %a) {
  //   %1 = extractelement <2 x double> %a, i32 0
  //   %2 = extractelement <2 x double> %a, i32 1
  //   %res = fmul double %1, %2
  //   ret double %res
  // }
  // %2 and %res can be merged in the backend to generate fmul d0, d0, v1.d[1]
  auto ExtractCanFuseWithFmul = [&]() {
    // We bail out if the extract is from lane 0.
    if (Index == 0)
      return false;
 
    // Check if the scalar element type of the vector operand of ExtractElement
    // instruction is one of the allowed types.
    auto IsAllowedScalarTy = [&](const Type *T) {
      return T->isFloatTy() || T->isDoubleTy() ||
             (T->isHalfTy() && ST->hasFullFP16());
    };
 
    // Check if the extractelement user is scalar fmul.
    auto IsUserFMulScalarTy = [](const Value *EEUser) {
      // Check if the user is scalar fmul.
      const auto *BO = dyn_cast<BinaryOperator>(EEUser);
      return BO && BO->getOpcode() == BinaryOperator::FMul &&
             !BO->getType()->isVectorTy();
    };
 
    // Check if the extract index is from lane 0 or lane equivalent to 0 for a
    // certain scalar type and a certain vector register width.
    auto IsExtractLaneEquivalentToZero = [&](unsigned Idx, unsigned EltSz) {
      auto RegWidth =
          getRegisterBitWidth(TargetTransformInfo::RGK_FixedWidthVector)
              .getFixedValue();
      return Idx == 0 || (RegWidth != 0 && (Idx * EltSz) % RegWidth == 0);
    };
 
    // Check if the type constraints on input vector type and result scalar type
    // of extractelement instruction are satisfied.
    if (!isa<FixedVectorType>(Ty) || !IsAllowedScalarTy(Ty->getScalarType()))
      return false;
 
    if (Scalar) {
      DenseMap<User *, unsigned> UserToExtractIdx;
      for (auto *U : Scalar->users()) {
        if (!IsUserFMulScalarTy(U))
          return false;
        // Recording entry for the user is important. Index value is not
        // important.
        UserToExtractIdx[U];
      }
      if (UserToExtractIdx.empty())
        return false;
      for (auto &[S, U, L] : ScalarUserAndIdx) {
        for (auto *U : S->users()) {
          if (UserToExtractIdx.contains(U)) {
            auto *FMul = cast<BinaryOperator>(U);
            auto *Op0 = FMul->getOperand(0);
            auto *Op1 = FMul->getOperand(1);
            if ((Op0 == S && Op1 == S) || Op0 != S || Op1 != S) {
              UserToExtractIdx[U] = L;
              break;
            }
          }
        }
      }
      for (auto &[U, L] : UserToExtractIdx) {
        if (!IsExtractLaneEquivalentToZero(Index, Ty->getScalarSizeInBits()) &&
            !IsExtractLaneEquivalentToZero(L, Ty->getScalarSizeInBits()))
          return false;
      }
    } else {
      const auto *EE = cast<ExtractElementInst>(I);
 
      const auto *IdxOp = dyn_cast<ConstantInt>(EE->getIndexOperand());
      if (!IdxOp)
        return false;
 
      return !EE->users().empty() && all_of(EE->users(), [&](const User *U) {
        if (!IsUserFMulScalarTy(U))
          return false;
 
        // Check if the other operand of extractelement is also extractelement
        // from lane equivalent to 0.
        const auto *BO = cast<BinaryOperator>(U);
        const auto *OtherEE = dyn_cast<ExtractElementInst>(
            BO->getOperand(0) == EE ? BO->getOperand(1) : BO->getOperand(0));
        if (OtherEE) {
          const auto *IdxOp = dyn_cast<ConstantInt>(OtherEE->getIndexOperand());
          if (!IdxOp)
            return false;
          return IsExtractLaneEquivalentToZero(
              cast<ConstantInt>(OtherEE->getIndexOperand())
                  ->getValue()
                  .getZExtValue(),
              OtherEE->getType()->getScalarSizeInBits());
        }
        return true;
      });
    }
    return true;
  };
 
  if (Opcode == Instruction::ExtractElement && (I || Scalar) &&
      ExtractCanFuseWithFmul())
    return 0;
 
  // All other insert/extracts cost this much.
  return CostKind == TTI::TCK_CodeSize ? 1
                                       : ST->getVectorInsertExtractBaseCost();
}
 
InstructionCost AArch64TTIImpl::getVectorInstrCost(
    unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, unsigned Index,
    const Value *Op0, const Value *Op1, TTI::VectorInstrContext VIC) const {
  // Treat insert at lane 0 into a poison vector as having zero cost. This
  // ensures vector broadcasts via an insert + shuffle (and will be lowered to a
  // single dup) are treated as cheap.
  if (Opcode == Instruction::InsertElement && Index == 0 && Op0 &&
      isa<PoisonValue>(Op0))
    return 0;
  return getVectorInstrCostHelper(Opcode, Ty, CostKind, Index, nullptr, nullptr,
                                  {}, VIC);
}
 
InstructionCost AArch64TTIImpl::getVectorInstrCost(
    unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind, unsigned Index,
    Value *Scalar, ArrayRef<std::tuple<Value *, User *, int>> ScalarUserAndIdx,
    TTI::VectorInstrContext VIC) const {
  return getVectorInstrCostHelper(Opcode, Ty, CostKind, Index, nullptr, Scalar,
                                  ScalarUserAndIdx, VIC);
}
 
InstructionCost
AArch64TTIImpl::getVectorInstrCost(const Instruction &I, Type *Ty,
                                   TTI::TargetCostKind CostKind, unsigned Index,
                                   TTI::VectorInstrContext VIC) const {
  return getVectorInstrCostHelper(I.getOpcode(), Ty, CostKind, Index, &I,
                                  nullptr, {}, VIC);
}
 
InstructionCost
AArch64TTIImpl::getIndexedVectorInstrCostFromEnd(unsigned Opcode, Type *Ty,
                                                 TTI::TargetCostKind CostKind,
                                                 unsigned Index) const {
  if (isa<FixedVectorType>(Ty))
    return BaseT::getIndexedVectorInstrCostFromEnd(Opcode, Ty, CostKind, Index);
 
  // This typically requires both while and lastb instructions in order
  // to extract the last element. If this is in a loop the while
  // instruction can at least be hoisted out, although it will consume a
  // predicate register. The cost should be more expensive than the base
  // extract cost, which is 2 for most CPUs.
  return CostKind == TTI::TCK_CodeSize
             ? 2
             : ST->getVectorInsertExtractBaseCost() + 1;
}
 
InstructionCost AArch64TTIImpl::getScalarizationOverhead(
    VectorType *Ty, const APInt &DemandedElts, bool Insert, bool Extract,
    TTI::TargetCostKind CostKind, bool ForPoisonSrc, ArrayRef<Value *> VL,
    TTI::VectorInstrContext VIC) const {
  if (isa<ScalableVectorType>(Ty))
    return InstructionCost::getInvalid();
 
  // There's no scalarization overhead if ld1/st1 is cheap and the
  // insert/extracts can be folded into the load/stores.
  if (ST->hasFastLD1Single()) {
    if ((VIC == TTI::VectorInstrContext::Store && Extract) ||
        (VIC == TTI::VectorInstrContext::Load && Insert))
      return 0;
  }
 
  if (Ty->getElementType()->isFloatingPointTy())
    return BaseT::getScalarizationOverhead(Ty, DemandedElts, Insert, Extract,
                                           CostKind);
  unsigned VecInstCost =
      CostKind == TTI::TCK_CodeSize ? 1 : ST->getVectorInsertExtractBaseCost();
  return DemandedElts.popcount() * (Insert + Extract) * VecInstCost;
}
 
std::optional<InstructionCost> AArch64TTIImpl::getFP16BF16PromoteCost(
    Type *Ty, TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info,
    TTI::OperandValueInfo Op2Info, bool IncludeTrunc, bool CanUseSVE,
    std::function<InstructionCost(Type *)> InstCost) const {
  if (!Ty->getScalarType()->isHalfTy() && !Ty->getScalarType()->isBFloatTy())
    return std::nullopt;
  if (Ty->getScalarType()->isHalfTy() && ST->hasFullFP16())
    return std::nullopt;
  // If we have +sve-b16b16 the operation can be promoted to SVE.
  if (CanUseSVE && ST->hasSVEB16B16() && ST->isNonStreamingSVEorSME2Available())
    return std::nullopt;
 
  Type *PromotedTy = Ty->getWithNewType(Type::getFloatTy(Ty->getContext()));
  InstructionCost Cost = getCastInstrCost(Instruction::FPExt, PromotedTy, Ty,
                                          TTI::CastContextHint::None, CostKind);
  if (!Op1Info.isConstant() && !Op2Info.isConstant())
    Cost *= 2;
  Cost += InstCost(PromotedTy);
  if (IncludeTrunc)
    Cost += getCastInstrCost(Instruction::FPTrunc, Ty, PromotedTy,
                             TTI::CastContextHint::None, CostKind);
  return Cost;
}
 
InstructionCost AArch64TTIImpl::getArithmeticInstrCost(
    unsigned Opcode, Type *Ty, TTI::TargetCostKind CostKind,
    TTI::OperandValueInfo Op1Info, TTI::OperandValueInfo Op2Info,
    ArrayRef<const Value *> Args, const Instruction *CtxI) const {
 
  // The code-generator is currently not able to handle scalable vectors
  // of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
  // it until all instructions are vetted.
  int ISD = TLI->InstructionOpcodeToISD(Opcode);
  if (auto *VTy = dyn_cast<ScalableVectorType>(Ty))
    if (VTy->getElementCount() == ElementCount::getScalable(1))
      if (!is_contained({ISD::ADD, ISD::SUB, ISD::MUL}, ISD))
        return InstructionCost::getInvalid();
 
  if (CtxI && CtxI->getOpcode() == Opcode && !CtxI->use_empty() &&
      useNeonVector(Ty)) {
    // The Ands of a bit select are folded into the BSL that the Or becomes.
    Instruction *MainAnd, *OtherAnd;
    Value *A, *B;
    if (Opcode == Instruction::And && CtxI->hasOneUse() &&
        matchBSL(CtxI->user_back(), MainAnd, OtherAnd, A, B))
      return 0;
 
    // A Not is folded into the BIC, ORN or BSL of each And/Or using it, unless
    // that And/Or negates both of its operands.
    if (match(CtxI, m_Not(m_Value())) &&
        all_of(CtxI->users(), [&](const User *U) {
          const auto *UI = cast<Instruction>(U);
          return is_contained({Instruction::And, Instruction::Or},
                              UI->getOpcode()) &&
                 UI->getParent() == CtxI->getParent() &&
                 !match(UI, m_BinOp(m_Not(m_Value()), m_Not(m_Value())));
        }))
      return 0;
  }
 
  // Legalize the type.
  std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
 
  // TODO: Handle more cost kinds for floating point operations.
  if (ISD == ISD::FADD || ISD == ISD::FSUB || ISD == ISD::FMUL ||
      ISD == ISD::FDIV || ISD == ISD::FREM || ISD == ISD::FNEG)
    if (CostKind != TTI::TCK_RecipThroughput)
      return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
                                           Op2Info, Args, CtxI);
 
  if (ISD == ISD::FADD || ISD == ISD::FSUB || ISD == ISD::FMUL ||
      ISD == ISD::FDIV || ISD == ISD::FREM) {
    // Increase the cost for half and bfloat types if not architecturally
    // supported.
    if (auto PromotedCost = getFP16BF16PromoteCost(
            Ty, CostKind, Op1Info, Op2Info, /*IncludeTrunc=*/true,
            // There is not native support for fdiv/frem even with +sve-b16b16.
            /*CanUseSVE=*/ISD != ISD::FDIV && ISD != ISD::FREM,
            [&](Type *PromotedTy) {
              return getArithmeticInstrCost(Opcode, PromotedTy, CostKind,
                                            Op1Info, Op2Info);
            }))
      return *PromotedCost;
 
    // fp128 all go via libcalls
    if (Ty->getScalarType()->isFP128Ty())
      return (CostKind == TTI::TCK_CodeSize ? 1 : 10) * LT.first;
  }
 
  // If the operation is a widening instruction (smull or umull) and both
  // operands are extends the cost can be cheaper by considering that the
  // operation will operate on the narrowest type size possible (double the
  // largest input size) and a further extend.
  if (Type *ExtTy = isBinExtWideningInstruction(Opcode, Ty, Args)) {
    if (ExtTy != Ty)
      return getArithmeticInstrCost(Opcode, ExtTy, CostKind) +
             getCastInstrCost(Instruction::ZExt, Ty, ExtTy,
                              TTI::CastContextHint::None, CostKind);
    return LT.first;
  }
 
  switch (ISD) {
  default:
    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
                                         Op2Info);
  case ISD::ADD:
  case ISD::SUB:
    return LT.first; // Also works for i128
  case ISD::MUL: {
    // i128 multiply is umulh + 2*madd + mul and grows ~O(Bitwidth^2). For
    // scalable vectors the cost of LT.first will be invalid, leading to an
    // invalid cost overall.
    unsigned Mul64CostFactor = (CostKind == TTI::TCK_RecipThroughput &&
                                ST->hasLimited64bitVectorMulBandwidth())
                                   ? 4
                                   : 1;
    if (Ty->getScalarSizeInBits() > 64) {
      unsigned NumLanes = isa<FixedVectorType>(Ty)
                              ? cast<FixedVectorType>(Ty)->getNumElements()
                              : 1;
      InstructionCost CostPerLane = LT.first / NumLanes;
      return CostPerLane * CostPerLane * NumLanes * Mul64CostFactor;
    }
 
    if (LT.second == MVT::v2i64) {
      // When SVE is available, then we can lower the v2i64 operation using
      // the SVE mul instruction, which has a lower cost.
      if (ST->hasSVE())
        return LT.first * Mul64CostFactor;
 
      // When SVE is not available, there is no MUL.2d instruction,
      // which means mul <2 x i64> is expensive as elements are extracted
      // from the vectors and the muls scalarized.
      // As getScalarizationOverhead is a bit too pessimistic, we
      // estimate the cost for a i64 vector directly here, which is:
      // - four 2-cost i64 extracts,
      // - two 2-cost i64 inserts, and
      // - two 1-cost muls.
      // So, for a v2i64 with LT.First = 1 the cost is 14, and for a v4i64 with
      // LT.first = 2 the cost is 28.
      return cast<VectorType>(Ty)->getElementCount().getKnownMinValue() *
             (getArithmeticInstrCost(Opcode, Ty->getScalarType(), CostKind) +
              getVectorInstrCost(Instruction::ExtractElement, Ty, CostKind, -1,
                                 nullptr, nullptr) *
                  2 +
              getVectorInstrCost(Instruction::InsertElement, Ty, CostKind, -1,
                                 nullptr, nullptr));
    }
 
    if (LT.second == MVT::nxv2i64)
      return LT.first * Mul64CostFactor;
 
    return LT.first;
  }
  case ISD::SREM:
  case ISD::SDIV:
    /*
    Notes for sdiv/srem specific costs:
    1. This only considers the cases where the divisor is constant, uniform and
    (pow-of-2/non-pow-of-2). Other cases are not important since they either
    result in some form of (ldr + adrp), corresponding to constant vectors, or
    scalarization of the division operation.
    2. Constant divisors, either negative in whole or partially, don't result in
    significantly different codegen as compared to positive constant divisors.
    So, we don't consider negative divisors separately.
    3. If the codegen is significantly different with SVE, it has been indicated
    using comments at appropriate places.
 
    sdiv specific cases:
    -----------------------------------------------------------------------
    codegen                       | pow-of-2               | Type
    -----------------------------------------------------------------------
    add + cmp + csel + asr        | Y                      | i64
    add + cmp + csel + asr        | Y                      | i32
    -----------------------------------------------------------------------
 
    srem specific cases:
    -----------------------------------------------------------------------
    codegen                       | pow-of-2               | Type
    -----------------------------------------------------------------------
    negs + and + and + csneg      | Y                      | i64
    negs + and + and + csneg      | Y                      | i32
    -----------------------------------------------------------------------
 
    other sdiv/srem cases:
    -------------------------------------------------------------------------
    common codegen            | + srem     | + sdiv     | pow-of-2  | Type
    -------------------------------------------------------------------------
    smulh + asr + add + add   | -          | -          | N         | i64
    smull + lsr + add + add   | -          | -          | N         | i32
    usra                      | and + sub  | sshr       | Y         | <2 x i64>
    2 * (scalar code)         | -          | -          | N         | <2 x i64>
    usra                      | bic + sub  | sshr + neg | Y         | <4 x i32>
    smull2 + smull + uzp2     | mls        | -          | N         | <4 x i32>
           + sshr  + usra     |            |            |           |
    -------------------------------------------------------------------------
    */
    if (Op2Info.isConstant() && Op2Info.isUniform()) {
      InstructionCost AddCost =
          getArithmeticInstrCost(Instruction::Add, Ty, CostKind,
                                 Op1Info.getNoProps(), Op2Info.getNoProps());
      InstructionCost AsrCost =
          getArithmeticInstrCost(Instruction::AShr, Ty, CostKind,
                                 Op1Info.getNoProps(), Op2Info.getNoProps());
      InstructionCost MulCost =
          getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
                                 Op1Info.getNoProps(), Op2Info.getNoProps());
      // add/cmp/csel/csneg should have similar cost while asr/negs/and should
      // have similar cost.
      auto VT = TLI->getValueType(DL, Ty);
      if (VT.isScalarInteger() && VT.getSizeInBits() <= 64) {
        if (Op2Info.isPowerOf2() || Op2Info.isNegatedPowerOf2()) {
          // Neg can be folded into the asr instruction.
          return ISD == ISD::SDIV ? (3 * AddCost + AsrCost)
                                  : (3 * AsrCost + AddCost);
        } else {
          return MulCost + AsrCost + 2 * AddCost;
        }
      } else if (VT.isVector()) {
        InstructionCost UsraCost = 2 * AsrCost;
        if (Op2Info.isPowerOf2() || Op2Info.isNegatedPowerOf2()) {
          // Division with scalable types corresponds to native 'asrd'
          // instruction when SVE is available.
          // e.g. %1 = sdiv <vscale x 4 x i32> %a, splat (i32 8)
 
          // One more for the negation in SDIV
          InstructionCost Cost =
              (Op2Info.isNegatedPowerOf2() && ISD == ISD::SDIV) ? AsrCost : 0;
          if (Ty->isScalableTy() && ST->hasSVE())
            Cost += 2 * AsrCost;
          else {
            Cost +=
                UsraCost +
                (ISD == ISD::SDIV
                     ? (LT.second.getScalarType() == MVT::i64 ? 1 : 2) * AsrCost
                     : 2 * AddCost);
          }
          return Cost;
        } else if (LT.second == MVT::v2i64) {
          return VT.getVectorNumElements() *
                 getArithmeticInstrCost(Opcode, Ty->getScalarType(), CostKind,
                                        Op1Info.getNoProps(),
                                        Op2Info.getNoProps());
        } else {
          // When SVE is available, we get:
          // smulh + lsr + add/sub + asr + add/sub.
          if (Ty->isScalableTy() && ST->hasSVE())
            return MulCost /*smulh cost*/ + 2 * AddCost + 2 * AsrCost;
          return 2 * MulCost + AddCost /*uzp2 cost*/ + AsrCost + UsraCost;
        }
      }
    }
    if (Op2Info.isConstant() && !Op2Info.isUniform() &&
        LT.second.isFixedLengthVector()) {
      // FIXME: When the constant vector is non-uniform, this may result in
      // loading the vector from constant pool or in some cases, may also result
      // in scalarization. For now, we are approximating this with the
      // scalarization cost.
      auto ExtractCost = 2 * getVectorInstrCost(Instruction::ExtractElement, Ty,
                                                CostKind, -1, nullptr, nullptr);
      auto InsertCost = getVectorInstrCost(Instruction::InsertElement, Ty,
                                           CostKind, -1, nullptr, nullptr);
      unsigned NElts = cast<FixedVectorType>(Ty)->getNumElements();
      return ExtractCost + InsertCost +
             NElts * getArithmeticInstrCost(Opcode, Ty->getScalarType(),
                                            CostKind, Op1Info.getNoProps(),
                                            Op2Info.getNoProps());
    }
    [[fallthrough]];
  case ISD::UDIV:
  case ISD::UREM: {
    auto VT = TLI->getValueType(DL, Ty);
    if (Op2Info.isConstant()) {
      // If the operand is a power of 2 we can use the shift or and cost.
      if (ISD == ISD::UDIV && Op2Info.isPowerOf2())
        return getArithmeticInstrCost(Instruction::LShr, Ty, CostKind,
                                      Op1Info.getNoProps(),
                                      Op2Info.getNoProps());
      if (ISD == ISD::UREM && Op2Info.isPowerOf2())
        return getArithmeticInstrCost(Instruction::And, Ty, CostKind,
                                      Op1Info.getNoProps(),
                                      Op2Info.getNoProps());
 
      if (ISD == ISD::UDIV || ISD == ISD::UREM) {
        // Divides by a constant are expanded to MULHU + SUB + SRL + ADD + SRL.
        // The MULHU will be expanded to UMULL for the types not listed below,
        // and will become a pair of UMULL+MULL2 for 128bit vectors.
        bool HasMULH = VT == MVT::i64 || LT.second == MVT::nxv2i64 ||
                       LT.second == MVT::nxv4i32 || LT.second == MVT::nxv8i16 ||
                       LT.second == MVT::nxv16i8;
        bool Is128bit = LT.second.is128BitVector();
 
        InstructionCost MulCost =
            getArithmeticInstrCost(Instruction::Mul, Ty, CostKind,
                                   Op1Info.getNoProps(), Op2Info.getNoProps());
        InstructionCost AddCost =
            getArithmeticInstrCost(Instruction::Add, Ty, CostKind,
                                   Op1Info.getNoProps(), Op2Info.getNoProps());
        InstructionCost ShrCost =
            getArithmeticInstrCost(Instruction::AShr, Ty, CostKind,
                                   Op1Info.getNoProps(), Op2Info.getNoProps());
        InstructionCost DivCost = MulCost * (Is128bit ? 2 : 1) + // UMULL/UMULH
                                  (HasMULH ? 0 : ShrCost) +      // UMULL shift
                                  AddCost * 2 + ShrCost;
        return DivCost + (ISD == ISD::UREM ? MulCost + AddCost : 0);
      }
    }
 
    // div i128's are lowered as libcalls.  Pass nullptr as (u)divti3 calls are
    // emitted by the backend even when those functions are not declared in the
    // module.
    if (!VT.isVector() && VT.getSizeInBits() > 64)
      return getCallInstrCost(/*Function*/ nullptr, Ty, {Ty, Ty}, CostKind);
 
    InstructionCost Cost = BaseT::getArithmeticInstrCost(
        Opcode, Ty, CostKind, Op1Info, Op2Info);
    if (Ty->isVectorTy() && (ISD == ISD::SDIV || ISD == ISD::UDIV)) {
      if (TLI->isOperationLegalOrCustom(ISD, LT.second) && ST->hasSVE()) {
        // SDIV/UDIV operations are lowered using SVE, then we can have less
        // costs.
        if (VT.isSimple() && isa<FixedVectorType>(Ty) &&
            Ty->getPrimitiveSizeInBits().getFixedValue() < 128) {
          static const CostTblEntry DivTbl[]{
              {ISD::SDIV, MVT::v2i8, 5},  {ISD::SDIV, MVT::v4i8, 8},
              {ISD::SDIV, MVT::v8i8, 8},  {ISD::SDIV, MVT::v2i16, 5},
              {ISD::SDIV, MVT::v4i16, 5}, {ISD::SDIV, MVT::v2i32, 1},
              {ISD::UDIV, MVT::v2i8, 5},  {ISD::UDIV, MVT::v4i8, 8},
              {ISD::UDIV, MVT::v8i8, 8},  {ISD::UDIV, MVT::v2i16, 5},
              {ISD::UDIV, MVT::v4i16, 5}, {ISD::UDIV, MVT::v2i32, 1}};
 
          const auto *Entry = CostTableLookup(DivTbl, ISD, VT.getSimpleVT());
          if (nullptr != Entry)
            return Entry->Cost;
        }
        // A non-power-of-2 count can't divide as a single whole-register op
        // (an inactive lane's leftover value could be a zero divisor and
        // trap), so the legalizer emits one div per whole register plus one
        // per set bit of the remainder (e.g. <7 x i32> emits 3 divs, not 2).
        if (auto *FVTy = dyn_cast<FixedVectorType>(Ty);
            FVTy && LT.second.isFixedLengthVector()) {
          unsigned NumElts = FVTy->getNumElements();
          unsigned RegElts = LT.second.getVectorNumElements();
          if (RegElts > 0)
            Cost = (NumElts / RegElts + popcount(NumElts % RegElts)) * 2;
        }
        // For 8/16-bit elements, the cost is higher because the type
        // requires promotion and possibly splitting:
        if (LT.second.getScalarType() == MVT::i8)
          Cost *= 8;
        else if (LT.second.getScalarType() == MVT::i16)
          Cost *= 4;
        return Cost;
      } else {
        // If one of the operands is a uniform constant then the cost for each
        // element is Cost for insertion, extraction and division.
        // Insertion cost = 2, Extraction Cost = 2, Division = cost for the
        // operation with scalar type
        if ((Op1Info.isConstant() && Op1Info.isUniform()) ||
            (Op2Info.isConstant() && Op2Info.isUniform())) {
          if (auto *VTy = dyn_cast<FixedVectorType>(Ty)) {
            InstructionCost DivCost = BaseT::getArithmeticInstrCost(
                Opcode, Ty->getScalarType(), CostKind, Op1Info, Op2Info);
            return (4 + DivCost) * VTy->getNumElements();
          }
        }
        // On AArch64, without SVE, vector divisions are expanded
        // into scalar divisions of each pair of elements.
        Cost += getVectorInstrCost(Instruction::ExtractElement, Ty, CostKind,
                                   -1, nullptr, nullptr);
        Cost += getVectorInstrCost(Instruction::InsertElement, Ty, CostKind, -1,
                                   nullptr, nullptr);
      }
 
      // TODO: if one of the arguments is scalar, then it's not necessary to
      // double the cost of handling the vector elements.
      Cost += Cost;
    }
    return Cost;
  }
  case ISD::XOR:
  case ISD::OR:
  case ISD::AND:
    // TODO: revisit these costs as it's not accurate enough for non-uniform
    // constant.
    return LT.first;
  case ISD::SRL:
  case ISD::SRA:
  case ISD::SHL: {
    // Immediate vector shifts require uniform shift amounts. Non-uniform
    // constants therefore use variable shifts and require materializing the
    // shift vector. Account for a shift and materialization per legalized
    // vector, together with shared setup.
    // This cost is for (ldr, shl) + adrp
    // TODO: These costs are based on CodeSize only, consider other CostKinds.
    if (Op2Info.isConstant() && !Op2Info.isUniform() &&
        LT.second.isFixedLengthVector())
      return 2 * LT.first + 1;
 
    // Marked 'custom' for combining purposes; a uniform shift amount still
    // lowers to a single legal instruction.
    return LT.first;
  }
 
  case ISD::FNEG:
    // Scalar fmul(fneg) or fneg(fmul) can be converted to fnmul
    if ((Ty->isFloatTy() || Ty->isDoubleTy() ||
         (Ty->isHalfTy() && ST->hasFullFP16())) &&
        CtxI &&
        ((CtxI->hasOneUse() &&
          match(*CtxI->user_begin(), m_FMul(m_Value(), m_Value()))) ||
         match(CtxI->getOperand(0), m_FMul(m_Value(), m_Value()))))
      return 0;
    [[fallthrough]];
  case ISD::FADD:
  case ISD::FSUB:
    if (!Ty->getScalarType()->isFP128Ty())
      return LT.first;
    [[fallthrough]];
  case ISD::FMUL:
  case ISD::FDIV:
    // These nodes are marked as 'custom' just to lower them to SVE.
    // We know said lowering will incur no additional cost.
    if (!Ty->getScalarType()->isFP128Ty())
      return 2 * LT.first;
 
    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
                                         Op2Info);
  case ISD::FREM:
    // Pass nullptr as fmod/fmodf calls are emitted by the backend even when
    // those functions are not declared in the module.
    if (!Ty->isVectorTy())
      return getCallInstrCost(/*Function*/ nullptr, Ty, {Ty, Ty}, CostKind);
    return BaseT::getArithmeticInstrCost(Opcode, Ty, CostKind, Op1Info,
                                         Op2Info);
  }
}
 
InstructionCost
AArch64TTIImpl::getAddressComputationCost(Type *PtrTy, ScalarEvolution *SE,
                                          const SCEV *Ptr,
                                          TTI::TargetCostKind CostKind) const {
  // Address computations in vectorized code with non-consecutive addresses will
  // likely result in more instructions compared to scalar code where the
  // computation can more often be merged into the index mode. The resulting
  // extra micro-ops can significantly decrease throughput.
  unsigned NumVectorInstToHideOverhead =
      ST->getCLOpts().neon_nonconst_stride_overhead;
  int MaxMergeDistance = 64;
 
  if (PtrTy->isVectorTy() && SE &&
      !BaseT::isConstantStridedAccessLessThan(SE, Ptr, MaxMergeDistance + 1))
    return NumVectorInstToHideOverhead;
 
  // In many cases the address computation is not merged into the instruction
  // addressing mode.
  return 1;
}
 
/// Check whether Opcode1 has less throughput according to the scheduling
/// model than Opcode2. For variant scheduling classes, the lowest throughput
/// of Opcode1 and the highest throughput of Opcode2 are compared among all the
/// scheduling classes they may be resolved to.
bool AArch64TTIImpl::hasKnownLowerThroughputFromSchedulingModel(
    unsigned Opcode1, unsigned Opcode2) const {
  const MCSchedModel &Sched = ST->getSchedModel();
  const TargetInstrInfo *TII = ST->getInstrInfo();
  if (!Sched.hasInstrSchedModel())
    return false;
 
  auto GetRThroughput = [&](unsigned Opcode, bool Highest) {
    std::optional<double> Res;
    SmallVector<unsigned> Worklist = {TII->get(Opcode).getSchedClass()};
    while (!Worklist.empty()) {
      unsigned SchedClass = Worklist.pop_back_val();
      const MCSchedClassDesc *SCD = Sched.getSchedClassDesc(SchedClass);
      if (SCD->isVariant()) {
        append_range(Worklist, AArch64_MC::getVariantSchedClassTargets(
                                   SchedClass, Sched.getProcessorID()));
        continue;
      }
      if (!SCD->isValid())
        continue;
      double RThroughput = MCSchedModel::getReciprocalThroughput(*ST, *SCD);
      if (!Res || (Highest ? RThroughput > *Res : RThroughput < *Res))
        Res = RThroughput;
    }
    return Res;
  };
  std::optional<double> RThroughput1 =
      GetRThroughput(Opcode1, /*Highest=*/true);
  std::optional<double> RThroughput2 =
      GetRThroughput(Opcode2, /*Highest=*/false);
  return RThroughput1 && RThroughput2 && *RThroughput1 > *RThroughput2;
}
 
InstructionCost AArch64TTIImpl::getCmpSelInstrCost(
    unsigned Opcode, Type *ValTy, Type *CondTy, CmpInst::Predicate VecPred,
    TTI::TargetCostKind CostKind, TTI::OperandValueInfo Op1Info,
    TTI::OperandValueInfo Op2Info, const Instruction *I) const {
  // We don't lower some vector selects well that are wider than the register
  // width. TODO: Improve this with different cost kinds.
  if (isa<FixedVectorType>(ValTy) && Opcode == Instruction::Select) {
    // We would need this many instructions to hide the scalarization happening.
    const int AmortizationCost = 20;
 
    // If VecPred is not set, check if we can get a predicate from the context
    // instruction, if its type matches the requested ValTy.
    if (VecPred == CmpInst::BAD_ICMP_PREDICATE && I && I->getType() == ValTy) {
      CmpPredicate CurrentPred;
      if (match(I, m_Select(m_Cmp(CurrentPred, m_Value(), m_Value()), m_Value(),
                            m_Value())))
        VecPred = CurrentPred;
    }
    // Check if we have a compare/select chain that can be lowered using
    // a (F)CMxx & BFI pair.
    if (CmpInst::isIntPredicate(VecPred) || VecPred == CmpInst::FCMP_OLE ||
        VecPred == CmpInst::FCMP_OLT || VecPred == CmpInst::FCMP_OGT ||
        VecPred == CmpInst::FCMP_OGE || VecPred == CmpInst::FCMP_OEQ ||
        VecPred == CmpInst::FCMP_UNE) {
      static const auto ValidMinMaxTys = {
          MVT::v8i8,  MVT::v16i8, MVT::v4i16, MVT::v8i16, MVT::v2i32,
          MVT::v4i32, MVT::v2i64, MVT::v2f32, MVT::v4f32, MVT::v2f64};
      static const auto ValidFP16MinMaxTys = {MVT::v4f16, MVT::v8f16};
 
      auto LT = getTypeLegalizationCost(ValTy);
      if (any_of(ValidMinMaxTys, equal_to(LT.second)) ||
          (ST->hasFullFP16() &&
           any_of(ValidFP16MinMaxTys, equal_to(LT.second))))
        return LT.first;
    }
 
    static const TypeConversionCostTblEntry VectorSelectTbl[] = {
        {Instruction::Select, MVT::v2i1, MVT::v2f32, 2},
        {Instruction::Select, MVT::v2i1, MVT::v2f64, 2},
        {Instruction::Select, MVT::v4i1, MVT::v4f32, 2},
        {Instruction::Select, MVT::v4i1, MVT::v4f16, 2},
        {Instruction::Select, MVT::v8i1, MVT::v8f16, 2},
        {Instruction::Select, MVT::v16i1, MVT::v16i16, 16},
        {Instruction::Select, MVT::v8i1, MVT::v8i32, 8},
        {Instruction::Select, MVT::v16i1, MVT::v16i32, 16},
        {Instruction::Select, MVT::v4i1, MVT::v4i64, 4 * AmortizationCost},
        {Instruction::Select, MVT::v8i1, MVT::v8i64, 8 * AmortizationCost},
        {Instruction::Select, MVT::v16i1, MVT::v16i64, 16 * AmortizationCost}};
 
    EVT SelCondTy = TLI->getValueType(DL, CondTy);
    EVT SelValTy = TLI->getValueType(DL, ValTy);
    if (SelCondTy.isSimple() && SelValTy.isSimple()) {
      if (const auto *Entry = ConvertCostTableLookup(VectorSelectTbl, Opcode,
                                                     SelCondTy.getSimpleVT(),
                                                     SelValTy.getSimpleVT()))
        return Entry->Cost;
    }
  }
 
  if (Opcode == Instruction::FCmp) {
    if (auto PromotedCost = getFP16BF16PromoteCost(
            ValTy, CostKind, Op1Info, Op2Info, /*IncludeTrunc=*/false,
            // TODO: Consider costing SVE FCMPs.
            /*CanUseSVE=*/false, [&](Type *PromotedTy) {
              InstructionCost Cost =
                  getCmpSelInstrCost(Opcode, PromotedTy, CondTy, VecPred,
                                     CostKind, Op1Info, Op2Info);
              if (isa<VectorType>(PromotedTy))
                Cost += getCastInstrCost(
                    Instruction::Trunc,
                    VectorType::getInteger(cast<VectorType>(ValTy)),
                    VectorType::getInteger(cast<VectorType>(PromotedTy)),
                    TTI::CastContextHint::None, CostKind);
              return Cost;
            }))
      return *PromotedCost;
 
    auto LT = getTypeLegalizationCost(ValTy);
    // Model unknown fp compares as a libcall.
    if (LT.second.getScalarType() != MVT::f64 &&
        LT.second.getScalarType() != MVT::f32 &&
        LT.second.getScalarType() != MVT::f16)
      return LT.first * getCallInstrCost(/*Function*/ nullptr, ValTy,
                                         {ValTy, ValTy}, CostKind);
 
    // Some comparison operators require expanding to multiple compares + or.
    unsigned Factor = 1;
    if (!CondTy->isVectorTy() &&
        (VecPred == FCmpInst::FCMP_ONE || VecPred == FCmpInst::FCMP_UEQ))
      Factor = 2; // fcmp with 2 selects
    else if (isa<FixedVectorType>(ValTy) &&
             (VecPred == FCmpInst::FCMP_ONE || VecPred == FCmpInst::FCMP_UEQ ||
              VecPred == FCmpInst::FCMP_ORD || VecPred == FCmpInst::FCMP_UNO))
      Factor = 3; // fcmxx+fcmyy+or
    else if (isa<ScalableVectorType>(ValTy) &&
             (VecPred == FCmpInst::FCMP_ONE || VecPred == FCmpInst::FCMP_UEQ))
      Factor = 3; // fcmxx+fcmyy+or
 
    if (isa<ScalableVectorType>(ValTy) &&
        CostKind == TTI::TCK_RecipThroughput &&
        hasKnownLowerThroughputFromSchedulingModel(AArch64::FCMEQ_PPzZZ_S,
                                                   AArch64::FCMEQv4f32))
      Factor *= 2;
 
    return Factor * (CostKind == TTI::TCK_Latency ? 2 : LT.first);
  }
 
  // Treat the icmp in icmp(and, 0) or icmp(and, -1/1) when it can be folded to
  // icmp(and, 0) as free, as we can make use of ands, but only if the
  // comparison is not unsigned. FIXME: Enable for non-throughput cost kinds
  // providing it will not cause performance regressions.
  if (CostKind == TTI::TCK_RecipThroughput && ValTy->isIntegerTy() &&
      Opcode == Instruction::ICmp && I && !CmpInst::isUnsigned(VecPred) &&
      TLI->isTypeLegal(TLI->getValueType(DL, ValTy)) &&
      match(I->getOperand(0), m_And(m_Value(), m_Value()))) {
    if (match(I->getOperand(1), m_Zero()))
      return 0;
 
    // x >= 1 / x < 1 -> x > 0 / x <= 0
    if (match(I->getOperand(1), m_One()) &&
        (VecPred == CmpInst::ICMP_SLT || VecPred == CmpInst::ICMP_SGE))
      return 0;
 
    // x <= -1 / x > -1 -> x > 0 / x <= 0
    if (match(I->getOperand(1), m_AllOnes()) &&
        (VecPred == CmpInst::ICMP_SLE || VecPred == CmpInst::ICMP_SGT))
      return 0;
  }
 
  // The base case handles scalable vectors fine for now, since it treats the
  // cost as 1 * legalization cost.
  InstructionCost Cost = BaseT::getCmpSelInstrCost(
      Opcode, ValTy, CondTy, VecPred, CostKind, Op1Info, Op2Info, I);
  if (Opcode == Instruction::ICmp && isa<ScalableVectorType>(ValTy) &&
      CostKind == TTI::TCK_RecipThroughput &&
      hasKnownLowerThroughputFromSchedulingModel(AArch64::CMPEQ_PPzZZ_S,
                                                 AArch64::CMEQv4i32))
    Cost *= 2;
  return Cost;
}
 
AArch64TTIImpl::TTI::MemCmpExpansionOptions
AArch64TTIImpl::enableMemCmpExpansion(bool OptSize, bool IsZeroCmp) const {
  TTI::MemCmpExpansionOptions Options;
  if (ST->requiresStrictAlign()) {
    // TODO: Add cost modeling for strict align. Misaligned loads expand to
    // a bunch of instructions when strict align is enabled.
    return Options;
  }
  Options.AllowOverlappingLoads = true;
  Options.MaxNumLoads = TLI->getMaxExpandSizeMemcmp(OptSize);
  Options.NumLoadsPerBlock = IsZeroCmp ? Options.MaxNumLoads : 1;
  // TODO: Though vector loads usually perform well on AArch64, in some targets
  // they may wake up the FP unit, which raises the power consumption.  Perhaps
  // they could be used with no holds barred (-O3).
  Options.LoadSizes = {8, 4, 2, 1};
  Options.AllowedTailExpansions = {3, 5, 6};
  return Options;
}
 
bool AArch64TTIImpl::prefersVectorizedAddressing() const {
  return ST->hasSVE();
}
 
InstructionCost
AArch64TTIImpl::getMemIntrinsicInstrCost(const MemIntrinsicCostAttributes &MICA,
                                         TTI::TargetCostKind CostKind) const {
  switch (MICA.getID()) {
  case Intrinsic::masked_scatter:
  case Intrinsic::masked_gather:
    return getGatherScatterOpCost(MICA, CostKind);
  case Intrinsic::masked_load:
  case Intrinsic::masked_store:
  case Intrinsic::masked_expandload:
  case Intrinsic::masked_compressstore:
    return getMaskedMemoryOpCost(MICA, CostKind);
  }
  return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
}
 
InstructionCost
AArch64TTIImpl::getMaskedMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
                                      TTI::TargetCostKind CostKind) const {
  Type *Src = MICA.getDataType();
 
  if (useNeonVector(Src))
    return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
  auto LT = getTypeLegalizationCost(Src);
  if (!LT.first.isValid())
    return InstructionCost::getInvalid();
 
  // Return an invalid cost for element types that we are unable to lower.
  auto *VT = cast<VectorType>(Src);
  if (VT->getElementType()->isIntegerTy(1))
    return InstructionCost::getInvalid();
 
  // <vscale x 1 x eltty> operations require mask adaptation.
  // Allow it for normal masked ld/st
  if (VT->getElementCount() == ElementCount::getScalable(1))
    return is_contained({Intrinsic::masked_load, Intrinsic::masked_store},
                        MICA.getID())
               ? LT.first + 1
               : InstructionCost::getInvalid();
 
  InstructionCost MemOpCost = LT.first;
  if (MICA.getID() == Intrinsic::masked_expandload) {
    // Fall back to the base cost (we can expand fixed-width expandloads).
    if (!isLegalMaskedExpandLoad(Src, MICA.getAlignment()))
      return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
 
    // Operation will be split into expand of masked.load
    // Something like:
    // cntp     x8, p0, p0.s
    // whilelo  p1.s, xzr, x8
    // ld1w     { z1.s }, p1/z, [x0]
    // expand   z1.s, p0, z1.s
    MemOpCost *= 4;
  }
 
  if (MICA.getID() == Intrinsic::masked_compressstore) {
    // Fall back to the base cost (we can expand fixed-width compressstores).
    if (!isLegalMaskedCompressStore(Src, MICA.getAlignment()))
      return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
 
    // A compress store lowers to something like:
    //  compact  z0.s, p0, z0.s
    //  cntp     x8, p1, p0.s
    //  whilelo  p0.s, xzr, x8
    //  st1w     { z0.s }, p0, [x0]
    MemOpCost *= 4;
  }
 
  // If we need to split the memory operation, we will also need to split the
  // mask. This will likely lead to overestimating the cost in some cases if
  // multiple memory operations use the same mask, but we often don't have
  // enough context to figure that out here.
  //
  // If the elements being loaded are bytes then the mask will already be split,
  // since the number of bits in a P register matches the number of bytes in a
  // Z register.
  if (LT.first > 1 && LT.second.getScalarSizeInBits() > 8)
    return MemOpCost * 2;
 
  return MemOpCost;
}
 
// This function returns gather/scatter overhead either from
// user-provided value or specialized values per-target from \p ST.
static unsigned getSVEGatherScatterOverhead(unsigned Opcode,
                                            const AArch64Subtarget *ST) {
  assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
         "Should be called on only load or stores.");
  switch (Opcode) {
  case Instruction::Load:
    return ST->getCLOpts().sve_gather_overhead.value_or(
        ST->getGatherOverhead());
    break;
  case Instruction::Store:
    return ST->getCLOpts().sve_scatter_overhead.value_or(
        ST->getScatterOverhead());
    break;
  default:
    llvm_unreachable("Shouldn't have reached here");
  }
}
 
InstructionCost
AArch64TTIImpl::getGatherScatterOpCost(const MemIntrinsicCostAttributes &MICA,
                                       TTI::TargetCostKind CostKind) const {
 
  unsigned Opcode = (MICA.getID() == Intrinsic::masked_gather ||
                     MICA.getID() == Intrinsic::vp_gather)
                        ? Instruction::Load
                        : Instruction::Store;
 
  Type *DataTy = MICA.getDataType();
  Align Alignment = MICA.getAlignment();
  const Instruction *I = MICA.getInst();
 
  // No need to call useNeonVector here since isLegalMaskedGatherScatter
  // determines if we're using SVE for fixed-length vectors.
  if (!isLegalMaskedGatherScatter(DataTy))
    return BaseT::getMemIntrinsicInstrCost(MICA, CostKind);
  auto *VT = cast<VectorType>(DataTy);
  auto LT = getTypeLegalizationCost(DataTy);
  if (!LT.first.isValid())
    return InstructionCost::getInvalid();
 
  // Return an invalid cost for element types that we are unable to lower.
  if (!LT.second.isVector() ||
      !isElementTypeLegalForScalableVector(VT->getElementType()) ||
      VT->getElementType()->isIntegerTy(1))
    return InstructionCost::getInvalid();
 
  // The code-generator is currently not able to handle scalable vectors
  // of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
  // it. This change will be removed when code-generation for these types is
  // sufficiently reliable.
  if (VT->getElementCount() == ElementCount::getScalable(1))
    return InstructionCost::getInvalid();
 
  ElementCount LegalVF = LT.second.getVectorElementCount();
  InstructionCost MemOpCost =
      getMemoryOpCost(Opcode, VT->getElementType(), Alignment, 0, CostKind,
                      {TTI::OK_AnyValue, TTI::OP_None}, I);
  // Add on an overhead cost for using gathers/scatters.
  MemOpCost *= getSVEGatherScatterOverhead(Opcode, ST);
  return LT.first * MemOpCost * getMaxNumElements(LegalVF);
}
 
bool AArch64TTIImpl::useNeonVector(const Type *Ty) const {
  return isa<FixedVectorType>(Ty) && !ST->useSVEForFixedLengthVectors();
}
 
InstructionCost AArch64TTIImpl::getMemoryOpCost(unsigned Opcode, Type *Ty,
                                                Align Alignment,
                                                unsigned AddressSpace,
                                                TTI::TargetCostKind CostKind,
                                                TTI::OperandValueInfo OpInfo,
                                                const Instruction *I) const {
  EVT VT = TLI->getValueType(DL, Ty, true);
  // Type legalization can't handle structs
  if (VT == MVT::Other)
    return BaseT::getMemoryOpCost(Opcode, Ty, Alignment, AddressSpace,
                                  CostKind);
 
  auto LT = getTypeLegalizationCost(Ty);
  if (!LT.first.isValid())
    return InstructionCost::getInvalid();
 
  if (auto *VTy = dyn_cast<ScalableVectorType>(Ty)) {
 
    // We only support full register predicate loads and stores.
    if (VTy->getElementType()->isIntegerTy(1) &&
        !VTy->getElementCount().isKnownMultipleOf(
            ElementCount::getScalable(16)))
      return InstructionCost::getInvalid();
 
    // <vscale x 1 x eltty> operations require crafting a new mask.
    if (VTy->getElementCount() == ElementCount::getScalable(1)) {
      Intrinsic::ID IID = Opcode == Instruction::Load ? Intrinsic::masked_load
                                                      : Intrinsic::masked_store;
      return getMaskedMemoryOpCost(
                 MemIntrinsicCostAttributes(IID, Ty, Alignment, AddressSpace),
                 CostKind) +
             1;
    }
  }
 
  // TODO: consider latency as well for TCK_SizeAndLatency.
  if (CostKind == TTI::TCK_CodeSize || CostKind == TTI::TCK_SizeAndLatency)
    return LT.first;
 
  if (CostKind == TTI::TCK_Latency) {
    // Latency doesn't make much sense for stores, so just return 1
    if (Opcode == Instruction::Store)
      return 1;
    // If the subtarget has overridden the load latency then use that instead of
    // querying the SchedModel.
    if (ST->getFixedLoadLatency())
      return (LT.first - 1) + ST->getFixedLoadLatency();
    // We expect the load to become LT.first loads of type LT.second. The
    // latency will be the latency of the last load plus the time it gets to get
    // there, which will be the amount of other loads before that (i.e. total
    // loads - 1) multiplied by how long it takes to get through them (the
    // reciprocal of the throughput). We get the latency and reciprocal
    // throughput from the SchedModel, and assume that the loads become the
    // variant with unsigned integer offset.
    unsigned Inst = 0;
    if (LT.second.isScalableVector() ||
        ST->useSVEForFixedLengthVectors(LT.second)) {
      Inst = AArch64::LDR_ZXI;
    } else if (LT.second.isVector() || LT.second.isFloatingPoint()) {
      switch (LT.second.getSizeInBits()) {
      case 8:
        Inst = AArch64::LDRBui;
        break;
      case 16:
        Inst = AArch64::LDRHui;
        break;
      case 32:
        Inst = AArch64::LDRSui;
        break;
      case 64:
        Inst = AArch64::LDRDui;
        break;
      case 128:
        Inst = AArch64::LDRQui;
        break;
      default:
        llvm_unreachable("Unexpected float or vector type");
      }
    } else {
      switch (LT.second.getSizeInBits()) {
      case 8:
        Inst = AArch64::LDRBBui;
        break;
      case 16:
        Inst = AArch64::LDRHHui;
        break;
      case 32:
        Inst = AArch64::LDRWui;
        break;
      case 64:
        Inst = AArch64::LDRXui;
        break;
      default:
        llvm_unreachable("Unexpected integer type");
      }
    }
    const MCSchedModel &Sched = ST->getSchedModel();
    const TargetInstrInfo *TII = ST->getInstrInfo();
    unsigned SchedClass = TII->get(Inst).getSchedClass();
    const MCSchedClassDesc *SCD = Sched.hasInstrSchedModel()
                                      ? Sched.getSchedClassDesc(SchedClass)
                                      : nullptr;
    // If the cpu has no scheduling model, or it doesn't describe the load, then
    // fall back to the default load latency. Variant scheduling classes can't
    // be resolved without a MachineInstr, so treat them the same way.
    if (!SCD || !SCD->isValid() || SCD->isVariant())
      return (LT.first - 1) + ST->getLoadLatency();
    // We need to convert the number of loads before the last to a float here,
    // as the reciprocal throughput may be fractional.
    float NumLoads = (LT.first - 1).getValue();
    return NumLoads * Sched.getReciprocalThroughput(*ST, *SCD) +
           Sched.computeInstrLatency(*ST, *SCD);
  }
 
  if (ST->isMisaligned128StoreSlow() && Opcode == Instruction::Store &&
      LT.second.is128BitVector() && Alignment < Align(16)) {
    // Unaligned stores are extremely inefficient. We don't split all
    // unaligned 128-bit stores because the negative impact that has shown in
    // practice on inlined block copy code.
    // We make such stores expensive so that we will only vectorize if there
    // are 6 other instructions getting vectorized.
    const int AmortizationCost = 6;
 
    return LT.first * 2 * AmortizationCost;
  }
 
  // Opaque ptr or ptr vector types are i64s and can be lowered to STP/LDPs.
  if (Ty->isPtrOrPtrVectorTy())
    return LT.first;
 
  if (useNeonVector(Ty)) {
    // Check truncating stores and extending loads.
    if (Ty->getScalarSizeInBits() != LT.second.getScalarSizeInBits()) {
      // v4i8 types are lowered to scalar a load/store and sshll/xtn.
      if (VT == MVT::v4i8)
        return 2;
      // Otherwise we need to scalarize.
      return cast<FixedVectorType>(Ty)->getNumElements() * 2;
    }
    EVT EltVT = VT.getVectorElementType();
    unsigned EltSize = EltVT.getScalarSizeInBits();
    if (!isPowerOf2_32(EltSize) || EltSize < 8 || EltSize > 64 ||
        Alignment != Align(1))
      return LT.first;
    // FIXME: v3i8 lowering currently is very inefficient, due to automatic
    // widening to v4i8, which produces suboptimal results.
    if (VT.getVectorNumElements() == 3 && EltVT == MVT::i8)
      return LT.first;
 
    // Check non-power-of-2 loads/stores for legal vector element types with
    // NEON. Non-power-of-2 memory ops will get broken down to a set of
    // operations on smaller power-of-2 ops, including ld1/st1.
    InstructionCost Cost = VT.getVectorNumElements() / (128 / EltSize);
    unsigned Remainder = VT.getVectorNumElements() % (128 / EltSize);
    if (Remainder != 0) {
      SmallVector<std::pair<unsigned, unsigned>> TypeWorklist;
      TypeWorklist.push_back({Remainder, 0});
      while (!TypeWorklist.empty()) {
        auto [CurrNumElements, Offset] = TypeWorklist.pop_back_val();
        if (isPowerOf2_32(CurrNumElements)) {
          // 1 per load/store + possible lane insert.
          Cost += 1 + (Offset == 0 ? 0 : 1);
          continue;
        }
 
        unsigned PrevPow2 = NextPowerOf2(CurrNumElements) / 2;
        TypeWorklist.push_back({PrevPow2, Offset});
        TypeWorklist.push_back({CurrNumElements - PrevPow2, Offset + PrevPow2});
      }
    }
    return Cost;
  }
 
  return LT.first;
}
 
InstructionCost AArch64TTIImpl::getInterleavedMemoryOpCost(
    unsigned Opcode, Type *VecTy, unsigned Factor, ArrayRef<unsigned> Indices,
    Align Alignment, unsigned AddressSpace, TTI::TargetCostKind CostKind,
    bool UseMaskForCond, bool UseMaskForGaps) const {
  assert(Factor >= 2 && "Invalid interleave factor");
  auto *VecVTy = cast<VectorType>(VecTy);
 
  if (VecTy->isScalableTy() && !ST->hasSVE())
    return InstructionCost::getInvalid();
 
  // Scalable VFs emit vector.[de]interleave intrinsics, for which the target
  // supports factors up to the maximum supported interleave factor.
  if (VecTy->isScalableTy()) {
    if (Factor > TLI->getMaxSupportedInterleaveFactor())
      return InstructionCost::getInvalid();
 
    if (Factor == 3 &&
        DL.getTypeSizeInBits(VecTy).getKnownMinValue() != (3 * 128))
      return InstructionCost::getInvalid();
  }
 
  auto LT = getTypeLegalizationCost(VecTy);
  unsigned MaxNativeInterleaveFactor = TLI->getMaxSupportedInterleaveFactor();
  // Vectorization for masked interleaved accesses is only enabled for scalable
  // VF. For fixed-length SVE, avoid non-native interleave factor because
  // the generic fallback costs wide fixed-vector shuffles too optimistically.
  if (!VecTy->isScalableTy() &&
      (UseMaskForCond || UseMaskForGaps ||
       (Factor > MaxNativeInterleaveFactor &&
        TLI->useSVEForFixedLengthVectorVT(LT.second))))
    return InstructionCost::getInvalid();
 
  if (!UseMaskForGaps && Factor <= MaxNativeInterleaveFactor) {
    ElementCount EC = VecVTy->getElementCount();
    auto *SubVecTy = VectorType::get(VecVTy->getElementType(),
                                     EC.divideCoefficientBy(Factor));
 
    // ldN/stN only support legal vector types of size 64 or 128 in bits.
    // Accesses having vector types that are a multiple of 128 bits can be
    // matched to more than one ldN/stN instruction.
    bool UseScalable;
    if (EC.isKnownMultipleOf(Factor) &&
        TLI->isLegalInterleavedAccessType(SubVecTy, DL, UseScalable))
      return Factor * TLI->getNumInterleavedAccesses(SubVecTy, DL, UseScalable);
 
    // Cost the alternative approach for scalable vectors where the interleave
    // factor is larger than the VF: use a contiguous load/store of the full
    // wide vector followed by deinterleave/interleave shuffles.
    if (VecTy->isScalableTy() && EC.isKnownMultipleOf(Factor)) {
      if (SubVecTy->getElementCount() == ElementCount::getScalable(1))
        return InstructionCost::getInvalid();
 
      // Cost of the contiguous memory operation on the wide vector.
      InstructionCost MemCost;
      if (UseMaskForCond) {
        unsigned IID = Opcode == Instruction::Load ? Intrinsic::masked_load
                                                   : Intrinsic::masked_store;
        MemCost = getMemIntrinsicInstrCost(
            MemIntrinsicCostAttributes(IID, VecTy, Alignment, AddressSpace),
            CostKind);
      } else {
        MemCost =
            getMemoryOpCost(Opcode, VecTy, Alignment, AddressSpace, CostKind);
      }
 
      // llvm.vector.deinterleaveN is lowered as a binary tree of deinterleave2
      // operations. The tree has Log2(Factor) levels, with Factor UZP/ZIP
      // operations at each level, giving a total shuffle cost of
      // Factor * Log2(Factor).
      auto SubVecCost = getTypeLegalizationCost(SubVecTy);
      auto ResultCost = getTypeLegalizationCost(VecTy);
      llvm::InstructionCost LegalizationCost = SubVecCost.first;
 
      // FIXME: A temporary increase to the cost in cases where the input
      // element type is 4x the output type. Otherwise it produces an SVE tail
      // loop which is significantly larger than the NEON equivalent.
      if (Opcode == Instruction::Store && Factor == 4 &&
          SubVecCost.second.getScalarSizeInBits() ==
              (4 * ResultCost.second.getScalarSizeInBits()))
        LegalizationCost *= 4;
 
      return MemCost + (Factor * LegalizationCost) + (Factor * Log2_64(Factor));
    }
  }
 
  return BaseT::getInterleavedMemoryOpCost(Opcode, VecTy, Factor, Indices,
                                           Alignment, AddressSpace, CostKind,
                                           UseMaskForCond, UseMaskForGaps);
}
 
InstructionCost
AArch64TTIImpl::getCostOfKeepingLiveOverCall(ArrayRef<Type *> Tys) const {
  InstructionCost Cost = 0;
  TTI::TargetCostKind CostKind = TTI::TCK_RecipThroughput;
  for (auto *I : Tys) {
    if (!I->isVectorTy())
      continue;
    if (I->getScalarSizeInBits() * cast<FixedVectorType>(I)->getNumElements() ==
        128)
      Cost += getMemoryOpCost(Instruction::Store, I, Align(128), 0, CostKind) +
              getMemoryOpCost(Instruction::Load, I, Align(128), 0, CostKind);
  }
  return Cost;
}
 
bool AArch64TTIImpl::isLegalMaskedExpandLoad(Type *DataTy,
                                             Align Alignment) const {
  // Neon types should be scalarised when we are not choosing to use SVE.
  if (useNeonVector(DataTy))
    return false;
 
  if (!isElementTypeLegalForScalableVector(DataTy->getScalarType()))
    return false;
 
  // Return true only if we are able to lower using the SVE2p2/SME2p2
  // expand instruction.
  return (ST->isSVEAvailable() && ST->hasSVE2p2()) ||
         (ST->isSVEorStreamingSVEAvailable() && ST->hasSME2p2());
}
 
bool AArch64TTIImpl::isLegalSpeculativeLoad(Type *DataType,
                                            unsigned AddressSpace) const {
  // Matches AArch64TargetLowering::emitCanLoadSpeculatively: only address
  // space 0 and power-of-2 sizes up to the 16-byte MTE tag granule.
  // TODO: Support scalable vectors.
  if (AddressSpace != 0)
    return false;
  TypeSize Size = DL.getTypeStoreSize(DataType);
  return !Size.isScalable() && isPowerOf2_64(Size.getFixedValue()) &&
         Size.getFixedValue() <= 16;
}
 
bool AArch64TTIImpl::hasMultiVectorLoadStore(
    unsigned NumVectors, TTI::MaskSource Mask, VectorType *VectorTy,
    bool IsStore, std::optional<Instruction::CastOps> CastHint) const {
  if (NumVectors <= 1 || !ST->enableSubRegLiveness() || !ST->hasSVE2p1())
    return false;
 
  switch (Mask) {
  case TTI::MaskSource::ArbitraryPredicate:
    return false;
  case TTI::MaskSource::None:
  case TTI::MaskSource::ActiveLaneMask:
    break;
  }
 
  // A null vector type queries whether the target supports multi-vector memory
  // operations in general.
  if (!VectorTy)
    return true;
 
  if (!isa<ScalableVectorType>(VectorTy))
    return false;
 
  // Conservatively, avoid using multi-vector loads when it's possible we could
  // use extending loads instead. Note: We can ignore stores as we only use
  // truncating stores when the store vector-width is < a full SVE vector.
  if (!IsStore &&
      (CastHint == Instruction::ZExt || CastHint == Instruction::SExt))
    return false;
 
  TypeSize VectorWidth = DL.getTypeSizeInBits(VectorTy);
  if (!VectorWidth.isKnownMultipleOf(AArch64::SVEBitsPerBlock))
    return false;
 
  // For masked loads/stores only allow operations that map to two or four
  // vectors (as these can use a single predicate-as-counter).
  if (Mask == TTI::MaskSource::ActiveLaneMask) {
    unsigned NumSVEVectors = (VectorWidth * NumVectors).getKnownMinValue() /
                             AArch64::SVEBitsPerBlock;
    return NumSVEVectors == 2 || NumSVEVectors == 4;
  }
 
  // For unpredicated loads/stores allow any pow-of-two multiple of a vector >=
  // to a single z-register. We can split operations wider than a single
  // multi-vector load/store during ISEL.
  return isPowerOf2_32(NumVectors);
}
 
unsigned
AArch64TTIImpl::getMaxInterleaveFactor(ElementCount VF,
                                       bool HasUnorderedReductions) const {
  if (VF.isScalar() || (HasUnorderedReductions && VF.getKnownMinValue() <= 4))
    return 4;
  return ST->getMaxInterleaveFactor();
}
 
// For Falkor, we want to avoid having too many strided loads in a loop since
// that can exhaust the HW prefetcher resources.  We adjust the unroller
// MaxCount preference below to attempt to ensure unrolling doesn't create too
// many strided loads.
static void
getFalkorUnrollingPreferences(Loop *L, ScalarEvolution &SE,
                              TargetTransformInfo::UnrollingPreferences &UP) {
  enum { MaxStridedLoads = 7 };
  auto countStridedLoads = [](Loop *L, ScalarEvolution &SE) {
    int StridedLoads = 0;
    // FIXME? We could make this more precise by looking at the CFG and
    // e.g. not counting loads in each side of an if-then-else diamond.
    for (const auto BB : L->blocks()) {
      for (auto &I : *BB) {
        LoadInst *LMemI = dyn_cast<LoadInst>(&I);
        if (!LMemI)
          continue;
 
        Value *PtrValue = LMemI->getPointerOperand();
        if (L->isLoopInvariant(PtrValue))
          continue;
 
        const SCEV *LSCEV = SE.getSCEV(PtrValue);
        const SCEVAddRecExpr *LSCEVAddRec = dyn_cast<SCEVAddRecExpr>(LSCEV);
        if (!LSCEVAddRec || !LSCEVAddRec->isAffine())
          continue;
 
        // FIXME? We could take pairing of unrolled load copies into account
        // by looking at the AddRec, but we would probably have to limit this
        // to loops with no stores or other memory optimization barriers.
        ++StridedLoads;
        // We've seen enough strided loads that seeing more won't make a
        // difference.
        if (StridedLoads > MaxStridedLoads / 2)
          return StridedLoads;
      }
    }
    return StridedLoads;
  };
 
  int StridedLoads = countStridedLoads(L, SE);
  LLVM_DEBUG(dbgs() << "falkor-hwpf: detected " << StridedLoads
                    << " strided loads\n");
  // Pick the largest power of 2 unroll count that won't result in too many
  // strided loads.
  if (StridedLoads) {
    UP.MaxCount = 1 << Log2_32(MaxStridedLoads / StridedLoads);
    LLVM_DEBUG(dbgs() << "falkor-hwpf: setting unroll MaxCount to "
                      << UP.MaxCount << '\n');
  }
}
 
// This function returns true if the loop:
//  1. Has a valid cost, and
//  2. Has a cost within the supplied budget.
// Otherwise it returns false.
static bool isLoopSizeWithinBudget(Loop *L, const AArch64TTIImpl &TTI,
                                   InstructionCost Budget,
                                   unsigned *FinalSize) {
  // Estimate the size of the loop.
  InstructionCost LoopCost = 0;
 
  for (auto *BB : L->getBlocks()) {
    for (auto &I : *BB) {
      SmallVector<const Value *, 4> Operands(I.operand_values());
      InstructionCost Cost =
          TTI.getInstructionCost(&I, Operands, TTI::TCK_CodeSize);
      // This can happen with intrinsics that don't currently have a cost model
      // or for some operations that require SVE.
      if (!Cost.isValid())
        return false;
 
      LoopCost += Cost;
      if (LoopCost > Budget)
        return false;
    }
  }
 
  if (FinalSize)
    *FinalSize = LoopCost.getValue();
  return true;
}
 
static bool shouldUnrollMultiExitLoop(Loop *L, ScalarEvolution &SE,
                                      const AArch64TTIImpl &TTI) {
  // Only consider loops with unknown trip counts for which we can determine
  // a symbolic expression. Multi-exit loops with small known trip counts will
  // likely be unrolled anyway.
  const SCEV *BTC = SE.getSymbolicMaxBackedgeTakenCount(L);
  if (isa<SCEVConstant>(BTC) || isa<SCEVCouldNotCompute>(BTC))
    return false;
 
  // It might not be worth unrolling loops with low max trip counts. Restrict
  // this to max trip counts > 32 for now.
  unsigned MaxTC = SE.getSmallConstantMaxTripCount(L);
  if (MaxTC > 0 && MaxTC <= 32)
    return false;
 
  // Make sure the loop size is <= 5.
  if (!isLoopSizeWithinBudget(L, TTI, 5, nullptr))
    return false;
 
  // Small search loops with multiple exits can be highly beneficial to unroll.
  // We only care about loops with exactly two exiting blocks, although each
  // block could jump to the same exit block.
  ArrayRef<BasicBlock *> Blocks = L->getBlocks();
  if (Blocks.size() != 2)
    return false;
 
  if (any_of(Blocks, [](BasicBlock *BB) {
        return !isa<UncondBrInst, CondBrInst>(BB->getTerminator());
      }))
    return false;
 
  return true;
}
 
/// For Apple CPUs, we want to runtime-unroll loops to make better use if the
/// OOO engine's wide instruction window and various predictors.
static void
getAppleRuntimeUnrollPreferences(Loop *L, ScalarEvolution &SE,
                                 TargetTransformInfo::UnrollingPreferences &UP,
                                 const AArch64TTIImpl &TTI) {
  // Limit loops with structure that is highly likely to benefit from runtime
  // unrolling; that is we exclude outer loops and loops with many blocks (i.e.
  // likely with complex control flow). Note that the heuristics here may be
  // overly conservative and we err on the side of avoiding runtime unrolling
  // rather than unroll excessively. They are all subject to further refinement.
  if (!L->isInnermost() || L->getNumBlocks() > 8)
    return;
 
  // Loops with multiple exits are handled by common code.
  if (!L->getExitBlock())
    return;
 
  // Check if the loop contains any reductions that could be parallelized when
  // unrolling. If so, enable partial unrolling, if the trip count is know to be
  // a multiple of 2.
  bool HasParellelizableReductions =
      L->getNumBlocks() == 1 &&
      any_of(L->getHeader()->phis(),
             [&SE, L](PHINode &Phi) {
               return canParallelizeReductionWhenUnrolling(Phi, L, &SE);
             }) &&
      isLoopSizeWithinBudget(L, TTI, 12, nullptr);
  if (HasParellelizableReductions &&
      SE.getSmallConstantTripMultiple(L, L->getExitingBlock()) % 2 == 0) {
    UP.Partial = true;
    UP.MaxCount = 4;
    UP.AddAdditionalAccumulators = true;
  }
 
  const SCEV *BTC = SE.getSymbolicMaxBackedgeTakenCount(L);
  if (isa<SCEVConstant>(BTC) || isa<SCEVCouldNotCompute>(BTC) ||
      (SE.getSmallConstantMaxTripCount(L) > 0 &&
       SE.getSmallConstantMaxTripCount(L) <= 32))
    return;
 
  if (getBooleanLoopAttribute(L, "llvm.loop.unroll.runtime.disable"))
    return;
 
  if (SE.getSymbolicMaxBackedgeTakenCount(L) != SE.getBackedgeTakenCount(L))
    return;
 
  // Limit to loops with trip counts that are cheap to expand.
  UP.SCEVExpansionBudget = 1;
 
  if (HasParellelizableReductions) {
    UP.Runtime = true;
    UP.DefaultUnrollRuntimeCount = 4;
    UP.AddAdditionalAccumulators = true;
  }
 
  // Try to unroll small, single-block loops with low budget, if they have
  // load/store dependencies, to expose more parallel memory access streams,
  // or if they do little work inside a block (i.e. load -> X -> store pattern).
  BasicBlock *Header = L->getHeader();
  BasicBlock *Latch = L->getLoopLatch();
  if (Header == Latch) {
    // Estimate the size of the loop.
    unsigned Size;
    unsigned Width = 10;
    if (!isLoopSizeWithinBudget(L, TTI, Width, &Size))
      return;
 
    // Try to find an unroll count that maximizes the use of the instruction
    // window, i.e. trying to fetch as many instructions per cycle as possible.
    unsigned MaxInstsPerLine = 16;
    unsigned UC = 1;
    unsigned BestUC = 1;
    unsigned SizeWithBestUC = BestUC * Size;
    while (UC <= 8) {
      unsigned SizeWithUC = UC * Size;
      if (SizeWithUC > 48)
        break;
      if ((SizeWithUC % MaxInstsPerLine) == 0 ||
          (SizeWithBestUC % MaxInstsPerLine) < (SizeWithUC % MaxInstsPerLine)) {
        BestUC = UC;
        SizeWithBestUC = BestUC * Size;
      }
      UC++;
    }
 
    if (BestUC == 1)
      return;
 
    SmallPtrSet<Value *, 8> LoadedValuesPlus;
    SmallVector<StoreInst *> Stores;
    for (auto *BB : L->blocks()) {
      for (auto &I : *BB) {
        Value *Ptr = getLoadStorePointerOperand(&I);
        if (!Ptr)
          continue;
        const SCEV *PtrSCEV = SE.getSCEV(Ptr);
        if (SE.isLoopInvariant(PtrSCEV, L))
          continue;
        if (isa<LoadInst>(&I)) {
          LoadedValuesPlus.insert(&I);
          // Include in-loop 1st users of loaded values.
          for (auto *U : I.users())
            if (L->contains(cast<Instruction>(U)))
              LoadedValuesPlus.insert(U);
        } else
          Stores.push_back(cast<StoreInst>(&I));
      }
    }
 
    if (none_of(Stores, [&LoadedValuesPlus](StoreInst *SI) {
          return LoadedValuesPlus.contains(SI->getOperand(0));
        }))
      return;
 
    UP.Runtime = true;
    UP.DefaultUnrollRuntimeCount = BestUC;
    return;
  }
 
  // Try to runtime-unroll loops with early-continues depending on loop-varying
  // loads; this helps with branch-prediction for the early-continues.
  auto *Term = dyn_cast<CondBrInst>(Header->getTerminator());
  SmallVector<BasicBlock *> Preds(predecessors(Latch));
  if (!Term || Preds.size() == 1 || !llvm::is_contained(Preds, Header) ||
      none_of(Preds, [L](BasicBlock *Pred) { return L->contains(Pred); }))
    return;
 
  std::function<bool(Instruction *, unsigned)> DependsOnLoopLoad =
      [&](Instruction *I, unsigned Depth) -> bool {
    if (isa<PHINode>(I) || L->isLoopInvariant(I) || Depth > 8)
      return false;
 
    if (isa<LoadInst>(I))
      return true;
 
    return any_of(I->operands(), [&](Value *V) {
      auto *I = dyn_cast<Instruction>(V);
      return I && DependsOnLoopLoad(I, Depth + 1);
    });
  };
  CmpPredicate Pred;
  Instruction *I;
  if (match(Term, m_Br(m_ICmp(Pred, m_Instruction(I), m_Value()), m_Value(),
                       m_Value())) &&
      DependsOnLoopLoad(I, 0)) {
    UP.Runtime = true;
  }
}
 
void AArch64TTIImpl::getUnrollingPreferences(
    Loop *L, ScalarEvolution &SE, TTI::UnrollingPreferences &UP,
    OptimizationRemarkEmitter *ORE) const {
  // Enable partial unrolling and runtime unrolling.
  BaseT::getUnrollingPreferences(L, SE, UP, ORE);
 
  UP.UpperBound = true;
 
  // A loop can have a small maximum trip count while SCEV still cannot
  // form an exact backedge count - typically a data-dependent exit, e.g.
  // shifting a value until it reaches zero. Unlike for counted loops, the
  // unrolled body keeps an exit test per iteration, and whether that pays
  // off depends on how many iterations the loop usually runs, which is
  // unknown at compile time; the code growth and extra branches are certain.
  // Be conservative and hold such loops to a lower upper bound; 5 still lets
  // smaller early-exit loops unroll. Also disable runtime unrolling, which
  // would clamp the unroll count to the known maximum trip count and produce
  // the same complete unroll.
  if (L->getExitingBlock() && !SE.isBackedgeTakenCountMaxOrZero(L) &&
      isa<SCEVCouldNotCompute>(SE.getBackedgeTakenCount(L))) {
    UP.MaxUpperBound = 5;
    UP.Runtime = false;
  }
 
  // For inner loop, it is more likely to be a hot one, and the runtime check
  // can be promoted out from LICM pass, so the overhead is less, let's try
  // a larger threshold to unroll more loops.
  if (L->getLoopDepth() > 1)
    UP.PartialThreshold *= 2;
 
  // Disable partial & runtime unrolling on -Os.
  UP.PartialOptSizeThreshold = 0;
 
  // Scan the loop: don't unroll loops with calls as this could prevent
  // inlining. Don't unroll auto-vectorized loops either, though do allow
  // unrolling of the scalar remainder.
  bool IsVectorized = getBooleanLoopAttribute(L, "llvm.loop.isvectorized");
  InstructionCost Cost = 0;
  for (auto *BB : L->getBlocks()) {
    for (auto &I : *BB) {
      // Both auto-vectorized loops and the scalar remainder have the
      // isvectorized attribute, so differentiate between them by the presence
      // of vector instructions.
      if (IsVectorized && I.getType()->isVectorTy())
        return;
      if (isa<CallBase>(I)) {
        if (isa<CallInst>(I) || isa<InvokeInst>(I))
          if (const Function *F = cast<CallBase>(I).getCalledFunction())
            if (!isLoweredToCall(F))
              continue;
        return;
      }
 
      // The cost is only compared against Aarch64ForceUnrollThreshold below.
      if (Cost >= ST->getCLOpts().force_unroll_threshold)
        continue;
      SmallVector<const Value *, 4> Operands(I.operand_values());
      Cost += getInstructionCost(&I, Operands,
                                 TargetTransformInfo::TCK_SizeAndLatency);
    }
  }
 
  // Apply subtarget-specific unrolling preferences.
  if (ST->isAppleMLike())
    getAppleRuntimeUnrollPreferences(L, SE, UP, *this);
  else if (ST->getProcFamily() == AArch64Subtarget::Falkor &&
           ST->getCLOpts().enable_falkor_hwpf_unroll_fix)
    getFalkorUnrollingPreferences(L, SE, UP);
 
  // If this is a small, multi-exit loop similar to something like std::find,
  // then there is typically a performance improvement achieved by unrolling.
  if (!L->getExitBlock() && shouldUnrollMultiExitLoop(L, SE, *this)) {
    UP.RuntimeUnrollMultiExit = true;
    UP.Runtime = true;
    // Limit unroll count.
    UP.DefaultUnrollRuntimeCount = 4;
    // Allow slightly more costly trip-count expansion to catch search loops
    // with pointer inductions.
    UP.SCEVExpansionBudget = 5;
    return;
  }
 
  // Enable runtime unrolling for in-order models
  // If mcpu is omitted, getProcFamily() returns AArch64Subtarget::Others, so by
  // checking for that case, we can ensure that the default behaviour is
  // unchanged
  if (ST->getProcFamily() != AArch64Subtarget::Generic &&
      !ST->getSchedModel().isOutOfOrder()) {
    UP.Runtime = true;
    UP.Partial = true;
    UP.UnrollRemainder = true;
    UP.DefaultUnrollRuntimeCount = 4;
 
    UP.UnrollAndJam = true;
    UP.UnrollAndJamInnerLoopThreshold = 60;
  }
 
  // Force unrolling small loops can be very useful because of the branch
  // taken cost of the backedge.
  if (Cost < ST->getCLOpts().force_unroll_threshold)
    UP.Force = true;
}
 
void AArch64TTIImpl::getPeelingPreferences(Loop *L, ScalarEvolution &SE,
                                           TTI::PeelingPreferences &PP) const {
  BaseT::getPeelingPreferences(L, SE, PP);
}
 
Value *AArch64TTIImpl::getOrCreateResultFromMemIntrinsic(IntrinsicInst *Inst,
                                                         Type *ExpectedType,
                                                         bool CanCreate) const {
  switch (Inst->getIntrinsicID()) {
  default:
    return nullptr;
  case Intrinsic::aarch64_neon_st1x2:
  case Intrinsic::aarch64_neon_st1x3:
  case Intrinsic::aarch64_neon_st1x4:
  case Intrinsic::aarch64_neon_st2:
  case Intrinsic::aarch64_neon_st3:
  case Intrinsic::aarch64_neon_st4: {
    // Create a struct type
    StructType *ST = dyn_cast<StructType>(ExpectedType);
    if (!CanCreate || !ST)
      return nullptr;
    unsigned NumElts = Inst->arg_size() - 1;
    if (ST->getNumElements() != NumElts)
      return nullptr;
    for (unsigned i = 0, e = NumElts; i != e; ++i) {
      if (Inst->getArgOperand(i)->getType() != ST->getElementType(i))
        return nullptr;
    }
    Value *Res = PoisonValue::get(ExpectedType);
    IRBuilder<> Builder(Inst);
    for (unsigned i = 0, e = NumElts; i != e; ++i) {
      Value *L = Inst->getArgOperand(i);
      Res = Builder.CreateInsertValue(Res, L, i);
    }
    return Res;
  }
  case Intrinsic::aarch64_neon_ld1x2:
  case Intrinsic::aarch64_neon_ld1x3:
  case Intrinsic::aarch64_neon_ld1x4:
  case Intrinsic::aarch64_neon_ld2:
  case Intrinsic::aarch64_neon_ld3:
  case Intrinsic::aarch64_neon_ld4:
    if (Inst->getType() == ExpectedType)
      return Inst;
    return nullptr;
  }
}
 
bool AArch64TTIImpl::getTgtMemIntrinsic(IntrinsicInst *Inst,
                                        MemIntrinsicInfo &Info) const {
  switch (Inst->getIntrinsicID()) {
  default:
    break;
  case Intrinsic::aarch64_neon_ld1x2:
  case Intrinsic::aarch64_neon_ld1x3:
  case Intrinsic::aarch64_neon_ld1x4:
  case Intrinsic::aarch64_neon_ld2:
  case Intrinsic::aarch64_neon_ld3:
  case Intrinsic::aarch64_neon_ld4:
    Info.ReadMem = true;
    Info.WriteMem = false;
    Info.PtrVal = Inst->getArgOperand(0);
    break;
  case Intrinsic::aarch64_neon_st1x2:
  case Intrinsic::aarch64_neon_st1x3:
  case Intrinsic::aarch64_neon_st1x4:
  case Intrinsic::aarch64_neon_st2:
  case Intrinsic::aarch64_neon_st3:
  case Intrinsic::aarch64_neon_st4:
    Info.ReadMem = false;
    Info.WriteMem = true;
    Info.PtrVal = Inst->getArgOperand(Inst->arg_size() - 1);
    break;
  }
 
  // Use the ID of neon load as the "matching id".
  switch (Inst->getIntrinsicID()) {
  default:
    return false;
  case Intrinsic::aarch64_neon_ld1x2:
  case Intrinsic::aarch64_neon_st1x2:
    Info.MatchingId = Intrinsic::aarch64_neon_ld1x2;
    break;
  case Intrinsic::aarch64_neon_ld1x3:
  case Intrinsic::aarch64_neon_st1x3:
    Info.MatchingId = Intrinsic::aarch64_neon_ld1x3;
    break;
  case Intrinsic::aarch64_neon_ld1x4:
  case Intrinsic::aarch64_neon_st1x4:
    Info.MatchingId = Intrinsic::aarch64_neon_ld1x4;
    break;
  case Intrinsic::aarch64_neon_ld2:
  case Intrinsic::aarch64_neon_st2:
    Info.MatchingId = Intrinsic::aarch64_neon_ld2;
    break;
  case Intrinsic::aarch64_neon_ld3:
  case Intrinsic::aarch64_neon_st3:
    Info.MatchingId = Intrinsic::aarch64_neon_ld3;
    break;
  case Intrinsic::aarch64_neon_ld4:
  case Intrinsic::aarch64_neon_st4:
    Info.MatchingId = Intrinsic::aarch64_neon_ld4;
    break;
  }
  return true;
}
 
/// See if \p I should be considered for address type promotion. We check if \p
/// I is a sext with right type and used in memory accesses. If it used in a
/// "complex" getelementptr, we allow it to be promoted without finding other
/// sext instructions that sign extended the same initial value. A getelementptr
/// is considered as "complex" if it has more than 2 operands.
bool AArch64TTIImpl::shouldConsiderAddressTypePromotion(
    const Instruction &I, bool &AllowPromotionWithoutCommonHeader) const {
  bool Considerable = false;
  AllowPromotionWithoutCommonHeader = false;
  if (!isa<SExtInst>(&I))
    return false;
  Type *ConsideredSExtType =
      Type::getInt64Ty(I.getParent()->getParent()->getContext());
  if (I.getType() != ConsideredSExtType)
    return false;
  // See if the sext is the one with the right type and used in at least one
  // GetElementPtrInst.
  for (const User *U : I.users()) {
    if (const GetElementPtrInst *GEPInst = dyn_cast<GetElementPtrInst>(U)) {
      Considerable = true;
      // A getelementptr is considered as "complex" if it has more than 2
      // operands. We will promote a SExt used in such complex GEP as we
      // expect some computation to be merged if they are done on 64 bits.
      if (GEPInst->getNumOperands() > 2) {
        AllowPromotionWithoutCommonHeader = true;
        break;
      }
    }
  }
  return Considerable;
}
 
bool AArch64TTIImpl::isLegalToVectorizeReduction(
    const RecurrenceDescriptor &RdxDesc, ElementCount VF) const {
  if (!VF.isScalable())
    return true;
 
  Type *Ty = RdxDesc.getRecurrenceType();
  if (Ty->isBFloatTy() || !isElementTypeLegalForScalableVector(Ty))
    return false;
 
  switch (RdxDesc.getRecurrenceKind()) {
  case RecurKind::Sub:
  case RecurKind::FSub:
  case RecurKind::AddChainWithSubs:
  case RecurKind::FAddChainWithSubs:
  case RecurKind::Add:
  case RecurKind::FAdd:
  case RecurKind::And:
  case RecurKind::Or:
  case RecurKind::Xor:
  case RecurKind::SMin:
  case RecurKind::SMax:
  case RecurKind::UMin:
  case RecurKind::UMax:
  case RecurKind::FMin:
  case RecurKind::FMax:
  case RecurKind::FMulAdd:
  case RecurKind::AnyOf:
  case RecurKind::FindLast:
    return true;
  default:
    return false;
  }
}
 
InstructionCost
AArch64TTIImpl::getMinMaxReductionCost(Intrinsic::ID IID, VectorType *Ty,
                                       FastMathFlags FMF,
                                       TTI::TargetCostKind CostKind) const {
  // The code-generator is currently not able to handle scalable vectors
  // of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
  // it. This change will be removed when code-generation for these types is
  // sufficiently reliable.
  if (auto *VTy = dyn_cast<ScalableVectorType>(Ty))
    if (VTy->getElementCount() == ElementCount::getScalable(1))
      return InstructionCost::getInvalid();
 
  std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Ty);
 
  if (LT.second.getScalarType() == MVT::f16 && !ST->hasFullFP16())
    return BaseT::getMinMaxReductionCost(IID, Ty, FMF, CostKind);
 
  InstructionCost LegalizationCost = 0;
  if (LT.first > 1) {
    Type *LegalVTy = EVT(LT.second).getTypeForEVT(Ty->getContext());
    IntrinsicCostAttributes Attrs(IID, LegalVTy, {LegalVTy, LegalVTy}, FMF);
    LegalizationCost = getIntrinsicInstrCost(Attrs, CostKind) * (LT.first - 1);
  }
 
  return LegalizationCost + /*Cost of horizontal reduction*/ 2;
}
 
InstructionCost AArch64TTIImpl::getArithmeticReductionCostSVE(
    unsigned Opcode, VectorType *ValTy, TTI::TargetCostKind CostKind) const {
  std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
  InstructionCost LegalizationCost = 0;
  if (LT.first > 1) {
    Type *LegalVTy = EVT(LT.second).getTypeForEVT(ValTy->getContext());
    LegalizationCost = getArithmeticInstrCost(Opcode, LegalVTy, CostKind);
    LegalizationCost *= LT.first - 1;
  }
 
  int ISD = TLI->InstructionOpcodeToISD(Opcode);
  assert(ISD && "Invalid opcode");
  // Add the final reduction cost for the legal horizontal reduction
  switch (ISD) {
  case ISD::ADD:
  case ISD::AND:
  case ISD::OR:
  case ISD::XOR:
  case ISD::FADD:
    return LegalizationCost + 2;
  default:
    return InstructionCost::getInvalid();
  }
}
 
InstructionCost
AArch64TTIImpl::getArithmeticReductionCost(unsigned Opcode, VectorType *ValTy,
                                           std::optional<FastMathFlags> FMF,
                                           TTI::TargetCostKind CostKind) const {
  // The code-generator is currently not able to handle scalable vectors
  // of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
  // it. This change will be removed when code-generation for these types is
  // sufficiently reliable.
  if (auto *VTy = dyn_cast<ScalableVectorType>(ValTy))
    if (VTy->getElementCount() == ElementCount::getScalable(1))
      return InstructionCost::getInvalid();
 
  if (TTI::requiresOrderedReduction(FMF)) {
    if (auto *FixedVTy = dyn_cast<FixedVectorType>(ValTy)) {
      InstructionCost BaseCost =
          BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
      // Add on extra cost to reflect the extra overhead on some CPUs. We still
      // end up vectorizing for more computationally intensive loops.
      return BaseCost + FixedVTy->getNumElements();
    }
 
    if (Opcode != Instruction::FAdd || ValTy->getElementType()->isBFloatTy())
      return InstructionCost::getInvalid();
 
    auto *VTy = cast<ScalableVectorType>(ValTy);
    InstructionCost Cost =
        getArithmeticInstrCost(Opcode, VTy->getScalarType(), CostKind);
    Cost *= getMaxNumElements(VTy->getElementCount());
    return Cost;
  }
 
  std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(ValTy);
  MVT MTy = LT.second;
 
  if (isa<ScalableVectorType>(ValTy) || TLI->useSVEForFixedLengthVectorVT(MTy))
    return getArithmeticReductionCostSVE(Opcode, ValTy, CostKind);
 
  int ISD = TLI->InstructionOpcodeToISD(Opcode);
  assert(ISD && "Invalid opcode");
 
  // Horizontal adds can use the 'addv' instruction. We model the cost of these
  // instructions as twice a normal vector add, plus 1 for each legalization
  // step (LT.first). This is the only arithmetic vector reduction operation for
  // which we have an instruction.
  // OR, XOR and AND costs should match the codegen from:
  // OR: llvm/test/CodeGen/AArch64/reduce-or.ll
  // XOR: llvm/test/CodeGen/AArch64/reduce-xor.ll
  // AND: llvm/test/CodeGen/AArch64/reduce-and.ll
  static const CostTblEntry CostTblNoPairwise[]{
      {ISD::ADD, MVT::v8i8, 2},
      {ISD::ADD, MVT::v16i8, 2},
      {ISD::ADD, MVT::v4i16, 2},
      {ISD::ADD, MVT::v8i16, 2},
      {ISD::ADD, MVT::v2i32, 2},
      {ISD::ADD, MVT::v4i32, 2},
      {ISD::ADD, MVT::v2i64, 2},
      {ISD::OR, MVT::v8i8, 5},  // fmov + orr_lsr + orr_lsr + lsr + orr
      {ISD::OR, MVT::v16i8, 7}, // ext + orr + same as v8i8
      {ISD::OR, MVT::v4i16, 4}, // fmov + orr_lsr + lsr + orr
      {ISD::OR, MVT::v8i16, 6}, // ext + orr + same as v4i16
      {ISD::OR, MVT::v2i32, 3}, // fmov + lsr + orr
      {ISD::OR, MVT::v4i32, 5}, // ext + orr + same as v2i32
      {ISD::OR, MVT::v2i64, 3}, // ext + orr + fmov
      {ISD::XOR, MVT::v8i8, 5}, // Same as above for or...
      {ISD::XOR, MVT::v16i8, 7},
      {ISD::XOR, MVT::v4i16, 4},
      {ISD::XOR, MVT::v8i16, 6},
      {ISD::XOR, MVT::v2i32, 3},
      {ISD::XOR, MVT::v4i32, 5},
      {ISD::XOR, MVT::v2i64, 3},
      {ISD::AND, MVT::v8i8, 5}, // Same as above for or...
      {ISD::AND, MVT::v16i8, 7},
      {ISD::AND, MVT::v4i16, 4},
      {ISD::AND, MVT::v8i16, 6},
      {ISD::AND, MVT::v2i32, 3},
      {ISD::AND, MVT::v4i32, 5},
      {ISD::AND, MVT::v2i64, 3},
  };
  switch (ISD) {
  default:
    break;
  case ISD::FADD:
    if (Type *EltTy = ValTy->getScalarType();
        // FIXME: For half types without fullfp16 support, this could extend and
        // use a fp32 faddp reduction but current codegen unrolls.
        MTy.isVector() && (EltTy->isFloatTy() || EltTy->isDoubleTy() ||
                           (EltTy->isHalfTy() && ST->hasFullFP16()))) {
      const unsigned NElts = MTy.getVectorNumElements();
      if (ValTy->getElementCount().getFixedValue() >= 2 && NElts >= 2 &&
          isPowerOf2_32(NElts))
        // Reduction corresponding to series of fadd instructions is lowered to
        // series of faddp instructions. faddp has latency/throughput that
        // matches fadd instruction and hence, every faddp instruction can be
        // considered to have a relative cost = 1 with
        // CostKind = TCK_RecipThroughput.
        // An faddp will pairwise add vector elements, so the size of input
        // vector reduces by half every time, requiring
        // #(faddp instructions) = log2_32(NElts).
        return (LT.first - 1) + /*No of faddp instructions*/ Log2_32(NElts);
    }
    break;
  case ISD::ADD:
    if (const auto *Entry = CostTableLookup(CostTblNoPairwise, ISD, MTy))
      return (LT.first - 1) + Entry->Cost;
    break;
  case ISD::XOR:
  case ISD::AND:
  case ISD::OR:
    const auto *Entry = CostTableLookup(CostTblNoPairwise, ISD, MTy);
    if (!Entry)
      break;
    auto *ValVTy = cast<FixedVectorType>(ValTy);
    if (MTy.getVectorNumElements() <= ValVTy->getNumElements() &&
        isPowerOf2_32(ValVTy->getNumElements())) {
      InstructionCost ExtraCost = 0;
      if (LT.first != 1) {
        // Type needs to be split, so there is an extra cost of LT.first - 1
        // arithmetic ops.
        auto *Ty = FixedVectorType::get(ValTy->getElementType(),
                                        MTy.getVectorNumElements());
        ExtraCost = getArithmeticInstrCost(Opcode, Ty, CostKind);
        ExtraCost *= LT.first - 1;
      }
      // All and/or/xor of i1 will be lowered with maxv/minv/addv + fmov
      auto Cost = ValVTy->getElementType()->isIntegerTy(1) ? 2 : Entry->Cost;
      return Cost + ExtraCost;
    }
    break;
  }
  return BaseT::getArithmeticReductionCost(Opcode, ValTy, FMF, CostKind);
}
 
InstructionCost AArch64TTIImpl::getExtendedReductionCost(
    unsigned Opcode, bool IsUnsigned, Type *ResTy, VectorType *VecTy,
    std::optional<FastMathFlags> FMF, TTI::TargetCostKind CostKind) const {
  EVT VecVT = TLI->getValueType(DL, VecTy);
  EVT ResVT = TLI->getValueType(DL, ResTy);
 
  if (IsUnsigned && Opcode == Instruction::Add &&
      VecVT.isScalableVectorOf(MVT::i1) && ResVT.isSimple() &&
      getTLI()->isTypeLegal(VecVT)) {
    // This is vecreduce.add(ZExt(mask)) which can lower as cntp(mask).
    return 1;
  }
 
  if (Opcode == Instruction::Add && VecVT.isSimple() && ResVT.isSimple() &&
      !VecVT.isScalableVT() && VecVT.getFixedSizeInBits() >= 64) {
    std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(VecTy);
 
    // The legal cases are:
    //   UADDLV 8/16/32->32
    //   UADDLP 32->64
    unsigned RevVTSize = ResVT.getSizeInBits();
    if (((LT.second == MVT::v8i8 || LT.second == MVT::v16i8) &&
         RevVTSize <= 32) ||
        ((LT.second == MVT::v4i16 || LT.second == MVT::v8i16) &&
         RevVTSize <= 32) ||
        ((LT.second == MVT::v2i32 || LT.second == MVT::v4i32) &&
         RevVTSize <= 64))
      return (LT.first - 1) * 2 + 2;
  }
 
  return BaseT::getExtendedReductionCost(Opcode, IsUnsigned, ResTy, VecTy, FMF,
                                         CostKind);
}
 
InstructionCost
AArch64TTIImpl::getMulAccReductionCost(bool IsUnsigned, unsigned RedOpcode,
                                       Type *ResTy, VectorType *VecTy,
                                       TTI::TargetCostKind CostKind) const {
  EVT VecVT = TLI->getValueType(DL, VecTy);
  EVT ResVT = TLI->getValueType(DL, ResTy);
 
  if (ST->hasDotProd() && VecVT.isSimple() && ResVT.isSimple() &&
      RedOpcode == Instruction::Add) {
    std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(VecTy);
 
    // The legal cases with dotprod are
    //   UDOT 8->32
    // Which requires an additional uaddv to sum the i32 values.
    if ((LT.second == MVT::v8i8 || LT.second == MVT::v16i8) &&
         ResVT == MVT::i32)
      return LT.first + 2;
  }
 
  return BaseT::getMulAccReductionCost(IsUnsigned, RedOpcode, ResTy, VecTy,
                                       CostKind);
}
 
InstructionCost
AArch64TTIImpl::getSpliceCost(VectorType *Tp, int Index,
                              TTI::TargetCostKind CostKind) const {
  static const CostTblEntry ShuffleTbl[] = {
      { TTI::SK_Splice, MVT::nxv16i8,  1 },
      { TTI::SK_Splice, MVT::nxv8i16,  1 },
      { TTI::SK_Splice, MVT::nxv4i32,  1 },
      { TTI::SK_Splice, MVT::nxv2i64,  1 },
      { TTI::SK_Splice, MVT::nxv2f16,  1 },
      { TTI::SK_Splice, MVT::nxv4f16,  1 },
      { TTI::SK_Splice, MVT::nxv8f16,  1 },
      { TTI::SK_Splice, MVT::nxv2bf16, 1 },
      { TTI::SK_Splice, MVT::nxv4bf16, 1 },
      { TTI::SK_Splice, MVT::nxv8bf16, 1 },
      { TTI::SK_Splice, MVT::nxv2f32,  1 },
      { TTI::SK_Splice, MVT::nxv4f32,  1 },
      { TTI::SK_Splice, MVT::nxv2f64,  1 },
  };
 
  // The code-generator is currently not able to handle scalable vectors
  // of <vscale x 1 x eltty> yet, so return an invalid cost to avoid selecting
  // it. This change will be removed when code-generation for these types is
  // sufficiently reliable.
  if (Tp->getElementCount() == ElementCount::getScalable(1))
    return InstructionCost::getInvalid();
 
  std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(Tp);
  Type *LegalVTy = EVT(LT.second).getTypeForEVT(Tp->getContext());
  EVT PromotedVT = LT.second.getScalarType() == MVT::i1
                       ? TLI->getPromotedVTForPredicate(EVT(LT.second))
                       : LT.second;
  Type *PromotedVTy = EVT(PromotedVT).getTypeForEVT(Tp->getContext());
  InstructionCost LegalizationCost = 0;
  if (Index < 0) {
    LegalizationCost =
        getCmpSelInstrCost(Instruction::ICmp, PromotedVTy, PromotedVTy,
                           CmpInst::BAD_ICMP_PREDICATE, CostKind) +
        getCmpSelInstrCost(Instruction::Select, PromotedVTy, LegalVTy,
                           CmpInst::BAD_ICMP_PREDICATE, CostKind);
  }
 
  // Predicated splice are promoted when lowering. See AArch64ISelLowering.cpp
  // Cost performed on a promoted type.
  if (LT.second.getScalarType() == MVT::i1) {
    LegalizationCost +=
        getCastInstrCost(Instruction::ZExt, PromotedVTy, LegalVTy,
                         TTI::CastContextHint::None, CostKind) +
        getCastInstrCost(Instruction::Trunc, LegalVTy, PromotedVTy,
                         TTI::CastContextHint::None, CostKind);
  }
  const auto *Entry =
      CostTableLookup(ShuffleTbl, TTI::SK_Splice, PromotedVT.getSimpleVT());
  assert(Entry && "Illegal Type for Splice");
  LegalizationCost += Entry->Cost;
  return LegalizationCost * LT.first;
}
 
InstructionCost AArch64TTIImpl::getPartialReductionCost(
    unsigned Opcode, Type *InputTypeA, Type *InputTypeB, Type *AccumType,
    ElementCount VF, TTI::PartialReductionExtendKind OpAExtend,
    TTI::PartialReductionExtendKind OpBExtend, std::optional<unsigned> BinOp,
    TTI::TargetCostKind CostKind, std::optional<FastMathFlags> FMF) const {
  InstructionCost Invalid = InstructionCost::getInvalid();
 
  if (CostKind != TTI::TCK_RecipThroughput)
    return Invalid;
 
  if ((Opcode != Instruction::Add && Opcode != Instruction::Sub &&
       Opcode != Instruction::FAdd && Opcode != Instruction::FSub))
    return Invalid;
 
  // Floating-point partial reductions are invalid if `reassoc` and `contract`
  // are not allowed.
  if (AccumType->isFloatingPointTy()) {
    assert(FMF && "Missing FastMathFlags for floating-point partial reduction");
    if (!FMF->allowReassoc() || !FMF->allowContract())
      return Invalid;
  } else {
    assert(!FMF &&
           "FastMathFlags only apply to floating-point partial reductions");
  }
 
  assert((BinOp || (OpBExtend == TTI::PR_None && !InputTypeB)) &&
         (!BinOp || (OpBExtend != TTI::PR_None && InputTypeB)) &&
         "Unexpected values for OpBExtend or InputTypeB");
 
  // We only support multiply binary operations for now, and for muls we
  // require the types being extended to be the same.
  if (BinOp && ((*BinOp != Instruction::Mul && *BinOp != Instruction::FMul) ||
                InputTypeA != InputTypeB))
    return Invalid;
 
  // We only support the following element sizes.
  if (!is_contained({8u, 16u, 32u, 64u}, AccumType->getScalarSizeInBits()))
    return Invalid;
 
  // If none of the operands are extended and there's no extra BinOp, just
  // cost this as the equivalent arithmetic instruction.
  // TODO: Depending on VF and element type, we may be able to improve on this.
  if (!OpAExtend) {
    assert(!OpBExtend && "Extended second operand without extended first.");
    assert(InputTypeA == AccumType && "Type mismatch with no extensions.");
 
    VectorType *VTy = VectorType::get(AccumType, VF);
    return getArithmeticInstrCost(Opcode, VTy, CostKind);
  }
 
  bool IsUSDot = OpBExtend != TTI::PR_None && OpAExtend != OpBExtend;
  // USDot is natively supported with +i8mm. With plain +dotprod, SUMLA is
  // lowered to two udots plus an eor and a sub.
  if (IsUSDot && !ST->hasMatMulInt8() && !ST->hasDotProd())
    // FIXME: Remove this early bailout in favour of expand cost.
    return Invalid;
 
  unsigned Ratio =
      AccumType->getScalarSizeInBits() / InputTypeA->getScalarSizeInBits();
  if (VF.getKnownMinValue() <= Ratio)
    return Invalid;
 
  VectorType *InputVectorType = VectorType::get(InputTypeA, VF);
  VectorType *AccumVectorType =
      VectorType::get(AccumType, VF.divideCoefficientBy(Ratio));
  // We don't yet support all kinds of legalization.
  auto TC = TLI->getTypeConversion(AccumVectorType->getContext(),
                                   EVT::getEVT(AccumVectorType));
  switch (TC.first) {
  default:
    return Invalid;
  case TargetLowering::TypeLegal:
  case TargetLowering::TypePromoteInteger:
  case TargetLowering::TypeSplitVector:
    // The legalised type (e.g. after splitting) must be legal too.
    if (TLI->getTypeAction(AccumVectorType->getContext(), TC.second) !=
        TargetLowering::TypeLegal)
      return Invalid;
    break;
  }
 
  std::pair<InstructionCost, MVT> AccumLT =
      getTypeLegalizationCost(AccumVectorType);
  std::pair<InstructionCost, MVT> InputLT =
      getTypeLegalizationCost(InputVectorType);
 
  // Returns true if the subtarget supports the operation for a given type.
  auto IsSupported = [&](bool SVEPred, bool NEONPred) -> bool {
    return (ST->isSVEorStreamingSVEAvailable() && SVEPred) ||
           (AccumLT.second.isFixedLengthVector() &&
            AccumLT.second.getSizeInBits() <= 128 && ST->isNeonAvailable() &&
            NEONPred);
  };
 
  bool IsSub = Opcode == Instruction::Sub || Opcode == Instruction::FSub;
  InstructionCost Cost = InputLT.first * TTI::TCC_Basic;
  // Integer partial sub-reductions that don't map to a specific instruction,
  // carry an extra cost for implementing a double negation:
  //      partial_reduce_umls  acc, lhs, rhs
  // <=> -partial_reduce_umla -acc, lhs, rhs
  InstructionCost INegCost = IsSub ? 2 * InputLT.first * TTI::TCC_Basic : 0;
 
  if (AccumLT.second.getScalarType() == MVT::i32 &&
      InputLT.second.getScalarType() == MVT::i8) {
    // i8 -> i32 is natively supported with udot/sdot for both NEON and SVE.
    if (!IsUSDot && IsSupported(true, ST->hasDotProd()))
      return Cost + INegCost;
    // i8 -> i32 usdot requires +i8mm
    if (IsUSDot && IsSupported(ST->hasMatMulInt8(), ST->hasMatMulInt8()))
      return Cost + INegCost;
    // Without +i8mm, lower SUMLA via two udots plus an eor and a sub on plain
    // +dotprod targets. Note that this is only implemented for NEON, as all
    // modern CPUs with SVE also have +i8mm. Charge an extra factor for the
    // expansion.
    if (IsUSDot && IsSupported(false, ST->hasDotProd()))
      return Cost * 3 + INegCost;
  }
 
  if (ST->isSVEorStreamingSVEAvailable() && !IsUSDot) {
    // i16 -> i64 is natively supported for udot/sdot
    if (AccumLT.second.getScalarType() == MVT::i64 &&
        InputLT.second.getScalarType() == MVT::i16)
      return Cost + INegCost;
    // i16 -> i32 is natively supported with SVE2p1 udot/sdot.
    // For sub-reductions, we prefer using the *mlslb/t instructions.
    if (AccumLT.second.getScalarType() == MVT::i32 &&
        InputLT.second.getScalarType() == MVT::i16 &&
        (ST->hasSVE2p1() || ST->hasSME2()) && !IsSub)
      return Cost;
    // i8 -> i64 is supported with an extra level of extends
    if (AccumLT.second.getScalarType() == MVT::i64 &&
        InputLT.second.getScalarType() == MVT::i8)
      // FIXME: Also account for folding the i32 dot product results into the
      // i64 accumulator, as the NEON case below does.
      return Cost + INegCost;
    // i8 -> i16 is natively supported with SVE2p3 udot/sdot
    // For sub-reductions, we prefer using the *mlslb/t instructions.
    if (AccumLT.second.getScalarType() == MVT::i16 &&
        InputLT.second.getScalarType() == MVT::i8 &&
        (ST->hasSVE2p3() || ST->hasSME2p3()) && !IsSub)
      return Cost;
  }
 
  // With +dotprod, NEON lowers an extended i8 -> i64 partial reduction to a
  // udot/sdot into a zeroed i32 vector followed by an [SU]ADALP into the i64
  // accumulator. Charge one basic cost for each of the two steps per legal
  // input vector. This must stay after the SVE cases, as IsSupported's NEON
  // predicate also holds for fixed-length vectors on SVE targets.
  // TODO: Handle zero-extended multiplies and sub-reductions. In loops,
  // CodeGenPrepare turns zero extends that do not directly feed the partial
  // reduction (such as the operands of a multiply, or of the negation in a
  // sub-reduction) into TBLs, which hides the udot from instruction selection.
  // Sign-extended sub-reductions have not been measured yet.
  // TODO: Without +dotprod, pure adds with a ratio of 4 or 8 lower to a chain
  // of [SU]ADDLPs and an [SU]ADALP, which is not costed yet either.
  bool IsSExtMul = BinOp && OpAExtend == TTI::PR_SignExtend &&
                   OpBExtend == TTI::PR_SignExtend;
  if ((!BinOp || IsSExtMul) && !IsSub &&
      AccumLT.second.getScalarType() == MVT::i64 &&
      InputLT.second.getScalarType() == MVT::i8 &&
      IsSupported(false, ST->hasDotProd()))
    return Cost * 2;
 
  // f16 -> f32 is natively supported for fdot using either
  // SVE or NEON instruction.
  if (Opcode == Instruction::FAdd && !IsSub &&
      IsSupported(ST->hasSME2() || ST->hasSVE2p1(), ST->hasF16F32DOT()) &&
      AccumLT.second.getScalarType() == MVT::f32 &&
      InputLT.second.getScalarType() == MVT::f16)
    return Cost;
 
  // For a ratio of 2, we can use *mlal and *mlsl top/bottom instructions.
  if (Ratio == 2 && !IsUSDot) {
    MVT InVT = InputLT.second.getScalarType();
 
    // SVE2 [us]ml[as]lb/t and NEON [us]ml[as]l(2). A pure widening add with a
    // ratio of 2 can use [SU]ADALP instead.
    if (IsSupported(ST->hasSVE2() || ST->hasSME(), true) &&
        llvm::is_contained({MVT::i8, MVT::i16, MVT::i32}, InVT.SimpleTy))
      return (BinOp || IsSub) ? Cost * 2 : Cost;
 
    // SVE2 fml[as]lb/t and NEON fml[as]l(2)
    if (IsSupported(ST->hasSVE2(), ST->hasFP16FML()) && InVT == MVT::f16)
      return Cost * 2;
 
    // SME2/SVE2p1 bfmlslb/t
    if (IsSupported(ST->hasSVE2p1() || ST->hasSME2(), false) &&
        InVT == MVT::bf16 && IsSub)
      return Cost * 2;
 
    // FP partial sub-reductions that don't map to a specific instruction,
    // carry an extra cost for implementing an extra negation:
    //      partial_reduce_fmls  acc, lhs, rhs
    // <=>  partial_reduce_fmla  acc, lhs, -rhs
    InstructionCost FNegCost = IsSub ? InputLT.first * TTI::TCC_Basic : 0;
 
    // SVE and NEON bfmlalb/t
    if (IsSupported(ST->hasBF16(), ST->hasBF16()) && InVT == MVT::bf16)
      return Cost * 2 + FNegCost;
  }
 
  return BaseT::getPartialReductionCost(Opcode, InputTypeA, InputTypeB,
                                        AccumType, VF, OpAExtend, OpBExtend,
                                        BinOp, CostKind, FMF);
}
 
InstructionCost AArch64TTIImpl::getShuffleCost(
    TTI::ShuffleKind Kind, VectorType *DstTy, VectorType *SrcTy,
    TTI::TargetCostKind CostKind, ArrayRef<int> Mask, int Index,
    VectorType *SubTp, ArrayRef<const Value *> Args, const Instruction *CtxI,
    TTI::VectorInstrContext VIC) const {
  assert((Mask.empty() || DstTy->isScalableTy() ||
          Mask.size() == DstTy->getElementCount().getKnownMinValue()) &&
         "Expected the Mask to match the return size if given");
  assert(SrcTy->getScalarType() == DstTy->getScalarType() &&
         "Expected the same scalar types");
  std::pair<InstructionCost, MVT> LT = getTypeLegalizationCost(SrcTy);
 
  // If we have a Mask, and the LT is being legalized somehow, split the Mask
  // into smaller vectors and sum the cost of each shuffle.
  if (!Mask.empty() && isa<FixedVectorType>(SrcTy) && LT.second.isVector() &&
      LT.second.getScalarSizeInBits() * Mask.size() > 128 &&
      SrcTy->getScalarSizeInBits() == LT.second.getScalarSizeInBits() &&
      Mask.size() > LT.second.getVectorNumElements() && !Index && !SubTp) {
    // Check for LD3/LD4 instructions, which are represented in llvm IR as
    // deinterleaving-shuffle(load). The shuffle cost could potentially be free,
    // but we model it with a cost of LT.first so that LD3/LD4 have a higher
    // cost than just the load.
    if (Args.size() >= 1 && isa<LoadInst>(Args[0]) &&
        (ShuffleVectorInst::isDeInterleaveMaskOfFactor(Mask, 3) ||
         ShuffleVectorInst::isDeInterleaveMaskOfFactor(Mask, 4)))
      return std::max<InstructionCost>(1, LT.first / 4);
 
    // Check for ST3/ST4 instructions, which are represented in llvm IR as
    // store(interleaving-shuffle). The shuffle cost could potentially be free,
    // but we model it with a cost of LT.first so that ST3/ST4 have a higher
    // cost than just the store.
    if (CtxI && CtxI->hasOneUse() && isa<StoreInst>(*CtxI->user_begin()) &&
        (ShuffleVectorInst::isInterleaveMask(
             Mask, 4, SrcTy->getElementCount().getKnownMinValue() * 2) ||
         ShuffleVectorInst::isInterleaveMask(
             Mask, 3, SrcTy->getElementCount().getKnownMinValue() * 2)))
      return LT.first;
 
    unsigned TpNumElts = Mask.size();
    unsigned LTNumElts = LT.second.getVectorNumElements();
    unsigned NumVecs = (TpNumElts + LTNumElts - 1) / LTNumElts;
    VectorType *NTp = VectorType::get(SrcTy->getScalarType(),
                                      LT.second.getVectorElementCount());
    InstructionCost Cost;
    std::map<std::tuple<unsigned, unsigned, SmallVector<int>>, InstructionCost>
        PreviousCosts;
    for (unsigned N = 0; N < NumVecs; N++) {
      SmallVector<int> NMask;
      // Split the existing mask into chunks of size LTNumElts. Track the source
      // sub-vectors to ensure the result has at most 2 inputs.
      unsigned Source1 = -1U, Source2 = -1U;
      unsigned NumSources = 0;
      for (unsigned E = 0; E < LTNumElts; E++) {
        int MaskElt = (N * LTNumElts + E < TpNumElts) ? Mask[N * LTNumElts + E]
                                                      : PoisonMaskElem;
        if (MaskElt < 0) {
          NMask.push_back(PoisonMaskElem);
          continue;
        }
 
        // Calculate which source from the input this comes from and whether it
        // is new to us.
        unsigned Source = MaskElt / LTNumElts;
        if (NumSources == 0) {
          Source1 = Source;
          NumSources = 1;
        } else if (NumSources == 1 && Source != Source1) {
          Source2 = Source;
          NumSources = 2;
        } else if (NumSources >= 2 && Source != Source1 && Source != Source2) {
          NumSources++;
        }
 
        // Add to the new mask. For the NumSources>2 case these are not correct,
        // but are only used for the modular lane number.
        if (Source == Source1)
          NMask.push_back(MaskElt % LTNumElts);
        else if (Source == Source2)
          NMask.push_back(MaskElt % LTNumElts + LTNumElts);
        else
          NMask.push_back(MaskElt % LTNumElts);
      }
      // Check if we have already generated this sub-shuffle, which means we
      // will have already generated the output. For example a <16 x i32> splat
      // will be the same sub-splat 4 times, which only needs to be generated
      // once and reused.
      auto Result =
          PreviousCosts.insert({std::make_tuple(Source1, Source2, NMask), 0});
      // Check if it was already in the map (already costed).
      if (!Result.second)
        continue;
      // If the sub-mask has at most 2 input sub-vectors then re-cost it using
      // getShuffleCost. If not then cost it using the worst case as the number
      // of element moves into a new vector.
      InstructionCost NCost =
          NumSources <= 2
              ? getShuffleCost(NumSources <= 1 ? TTI::SK_PermuteSingleSrc
                                               : TTI::SK_PermuteTwoSrc,
                               NTp, NTp, CostKind, NMask, 0, nullptr, Args,
                               CtxI)
              : LTNumElts;
      Result.first->second = NCost;
      Cost += NCost;
    }
    return Cost;
  }
 
  Kind = improveShuffleKindFromMask(Kind, Mask, SrcTy, Index, SubTp);
  bool IsExtractSubvector = Kind == TTI::SK_ExtractSubvector;
  // A subvector extract can be implemented with a NEON/SVE ext (or trivial
  // extract, if from lane 0) for 128-bit NEON vectors or legal SVE vectors.
  // This currently only handles low or high extracts to prevent SLP vectorizer
  // regressions.
  // Note that SVE's ext instruction is destructive, but it can be fused with
  // a movprfx to act like a constructive instruction.
  if (IsExtractSubvector && LT.second.isFixedLengthVector()) {
    if (LT.second.getFixedSizeInBits() >= 128 &&
        cast<FixedVectorType>(SubTp)->getNumElements() ==
            LT.second.getVectorNumElements() / 2) {
      if (Index == 0)
        return 0;
      if (Index == (int)LT.second.getVectorNumElements() / 2)
        return 1;
    }
    Kind = TTI::SK_PermuteSingleSrc;
  }
  // FIXME: This was added to keep the costs equal when adding DstTys. Update
  // the code to handle length-changing shuffles.
  if (Kind == TTI::SK_InsertSubvector) {
    LT = getTypeLegalizationCost(DstTy);
    SrcTy = DstTy;
  }
 
  // Check for identity masks, which we can treat as free for both fixed and
  // scalable vector paths.
  if (!Mask.empty() && LT.second.isFixedLengthVector() &&
      (Kind == TTI::SK_PermuteTwoSrc || Kind == TTI::SK_PermuteSingleSrc) &&
      all_of(enumerate(Mask), [](const auto &M) {
        return M.value() < 0 || M.value() == (int)M.index();
      }))
    return 0;
 
  // Segmented shuffle matching.
  if (Kind == TTI::SK_PermuteSingleSrc && isa<FixedVectorType>(SrcTy) &&
      !Mask.empty() && SrcTy->getPrimitiveSizeInBits().isNonZero() &&
      SrcTy->getPrimitiveSizeInBits().isKnownMultipleOf(
          AArch64::SVEBitsPerBlock)) {
 
    FixedVectorType *VTy = cast<FixedVectorType>(SrcTy);
    unsigned Segments =
        VTy->getPrimitiveSizeInBits() / AArch64::SVEBitsPerBlock;
    unsigned SegmentElts = VTy->getNumElements() / Segments;
 
    // dupq zd.t, zn.t[idx]
    if ((ST->hasSVE2p1() || ST->hasSME2p1()) &&
        ST->isSVEorStreamingSVEAvailable() &&
        isDUPQMask(Mask, Segments, SegmentElts))
      return LT.first;
 
    // mov zd.q, vn
    if (ST->isSVEorStreamingSVEAvailable() &&
        isDUPFirstSegmentMask(Mask, Segments, SegmentElts))
      return LT.first;
  }
 
  // Check for broadcast loads, which are supported by the LD1R instruction.
  // In terms of code-size, the shuffle vector is free when a load + dup get
  // folded into a LD1R. That's what we check and return here. For performance
  // and reciprocal throughput, a LD1R is not completely free. In this case, we
  // return the cost for the broadcast below (i.e. 1 for most/all types), so
  // that we model the load + dup sequence slightly higher because LD1R is a
  // high latency instruction.
  if (CostKind == TTI::TCK_CodeSize && Kind == TTI::SK_Broadcast) {
    bool IsLoad = !Args.empty() && isa<LoadInst>(Args[0]);
    if (IsLoad && LT.second.isVector() &&
        isLegalBroadcastLoad(SrcTy->getElementType(),
                             LT.second.getVectorElementCount()))
      return 0;
  }
 
  // If we have 4 elements for the shuffle and a Mask, get the cost straight
  // from the perfect shuffle tables.
  if (Mask.size() == 4 &&
      SrcTy->getElementCount() == ElementCount::getFixed(4) &&
      (SrcTy->getScalarSizeInBits() == 16 ||
       SrcTy->getScalarSizeInBits() == 32) &&
      all_of(Mask, [](int E) { return E < 8; }))
    return getPerfectShuffleCost(Mask);
 
  // Check for other shuffles that are not SK_ kinds but we have native
  // instructions for, for example ZIP and UZP.
  unsigned Unused;
  if (LT.second.isFixedLengthVector() &&
      LT.second.getVectorNumElements() == Mask.size() &&
      (Kind == TTI::SK_PermuteTwoSrc || Kind == TTI::SK_PermuteSingleSrc ||
       // Discrepancies between isTRNMask and ShuffleVectorInst::isTransposeMask
       // mean that we can end up with shuffles that satisfy isTRNMask, but end
       // up labelled as TTI::SK_InsertSubvector. (e.g. {2, 0}).
       Kind == TTI::SK_InsertSubvector) &&
      (isZIPMask(Mask, LT.second.getVectorNumElements(), Unused, Unused) ||
       isTRNMask(Mask, LT.second.getVectorNumElements(), Unused, Unused) ||
       isUZPMask(Mask, LT.second.getVectorNumElements(), Unused) ||
       isREVMask(Mask, LT.second.getScalarSizeInBits(),
                 LT.second.getVectorNumElements(), 16) ||
       isREVMask(Mask, LT.second.getScalarSizeInBits(),
                 LT.second.getVectorNumElements(), 32) ||
       isREVMask(Mask, LT.second.getScalarSizeInBits(),
                 LT.second.getVectorNumElements(), 64) ||
       // Check for non-zero lane splats
       all_of(drop_begin(Mask),
              [&Mask](int M) { return M < 0 || M == Mask[0]; })))
    return 1;
 
  if (Kind == TTI::SK_Broadcast || Kind == TTI::SK_Transpose ||
      Kind == TTI::SK_Select || Kind == TTI::SK_PermuteSingleSrc ||
      Kind == TTI::SK_Reverse || Kind == TTI::SK_Splice) {
    static const CostTblEntry ShuffleTbl[] = {
        // Broadcast shuffle kinds can be performed with 'dup'.
        {TTI::SK_Broadcast, MVT::v8i8, 1},
        {TTI::SK_Broadcast, MVT::v16i8, 1},
        {TTI::SK_Broadcast, MVT::v4i16, 1},
        {TTI::SK_Broadcast, MVT::v8i16, 1},
        {TTI::SK_Broadcast, MVT::v2i32, 1},
        {TTI::SK_Broadcast, MVT::v4i32, 1},
        {TTI::SK_Broadcast, MVT::v2i64, 1},
        {TTI::SK_Broadcast, MVT::v4f16, 1},
        {TTI::SK_Broadcast, MVT::v8f16, 1},
        {TTI::SK_Broadcast, MVT::v4bf16, 1},
        {TTI::SK_Broadcast, MVT::v8bf16, 1},
        {TTI::SK_Broadcast, MVT::v2f32, 1},
        {TTI::SK_Broadcast, MVT::v4f32, 1},
        {TTI::SK_Broadcast, MVT::v2f64, 1},
        // Transpose shuffle kinds can be performed with 'trn1/trn2' and
        // 'zip1/zip2' instructions.
        {TTI::SK_Transpose, MVT::v8i8, 1},
        {TTI::SK_Transpose, MVT::v16i8, 1},
        {TTI::SK_Transpose, MVT::v4i16, 1},
        {TTI::SK_Transpose, MVT::v8i16, 1},
        {TTI::SK_Transpose, MVT::v2i32, 1},
        {TTI::SK_Transpose, MVT::v4i32, 1},
        {TTI::SK_Transpose, MVT::v2i64, 1},
        {TTI::SK_Transpose, MVT::v4f16, 1},
        {TTI::SK_Transpose, MVT::v8f16, 1},
        {TTI::SK_Transpose, MVT::v4bf16, 1},
        {TTI::SK_Transpose, MVT::v8bf16, 1},
        {TTI::SK_Transpose, MVT::v2f32, 1},
        {TTI::SK_Transpose, MVT::v4f32, 1},
        {TTI::SK_Transpose, MVT::v2f64, 1},
        // Select shuffle kinds.
        // TODO: handle vXi8/vXi16.
        {TTI::SK_Select, MVT::v2i32, 1}, // mov.
        {TTI::SK_Select, MVT::v4i32, 2}, // rev+trn (or similar).
        {TTI::SK_Select, MVT::v2i64, 1}, // mov.
        {TTI::SK_Select, MVT::v2f32, 1}, // mov.
        {TTI::SK_Select, MVT::v4f32, 2}, // rev+trn (or similar).
        {TTI::SK_Select, MVT::v2f64, 1}, // mov.
        // PermuteSingleSrc shuffle kinds.
        {TTI::SK_PermuteSingleSrc, MVT::v2i32, 1}, // mov.
        {TTI::SK_PermuteSingleSrc, MVT::v4i32, 3}, // perfectshuffle worst case.
        {TTI::SK_PermuteSingleSrc, MVT::v2i64, 1}, // mov.
        {TTI::SK_PermuteSingleSrc, MVT::v2f32, 1}, // mov.
        {TTI::SK_PermuteSingleSrc, MVT::v4f32, 3}, // perfectshuffle worst case.
        {TTI::SK_PermuteSingleSrc, MVT::v2f64, 1}, // mov.
        {TTI::SK_PermuteSingleSrc, MVT::v4i16, 3}, // perfectshuffle worst case.
        {TTI::SK_PermuteSingleSrc, MVT::v4f16, 3}, // perfectshuffle worst case.
        {TTI::SK_PermuteSingleSrc, MVT::v4bf16, 3}, // same
        {TTI::SK_PermuteSingleSrc, MVT::v8i16, 8},  // constpool + load + tbl
        {TTI::SK_PermuteSingleSrc, MVT::v8f16, 8},  // constpool + load + tbl
        {TTI::SK_PermuteSingleSrc, MVT::v8bf16, 8}, // constpool + load + tbl
        {TTI::SK_PermuteSingleSrc, MVT::v8i8, 8},   // constpool + load + tbl
        {TTI::SK_PermuteSingleSrc, MVT::v16i8, 8},  // constpool + load + tbl
        // Reverse can be lowered with `rev`.
        {TTI::SK_Reverse, MVT::v2i32, 1}, // REV64
        {TTI::SK_Reverse, MVT::v4i32, 2}, // REV64; EXT
        {TTI::SK_Reverse, MVT::v2i64, 1}, // EXT
        {TTI::SK_Reverse, MVT::v2f32, 1}, // REV64
        {TTI::SK_Reverse, MVT::v4f32, 2}, // REV64; EXT
        {TTI::SK_Reverse, MVT::v2f64, 1}, // EXT
        {TTI::SK_Reverse, MVT::v8f16, 2}, // REV64; EXT
        {TTI::SK_Reverse, MVT::v8bf16, 2}, // REV64; EXT
        {TTI::SK_Reverse, MVT::v8i16, 2}, // REV64; EXT
        {TTI::SK_Reverse, MVT::v16i8, 2}, // REV64; EXT
        {TTI::SK_Reverse, MVT::v4f16, 1}, // REV64
        {TTI::SK_Reverse, MVT::v4bf16, 1}, // REV64
        {TTI::SK_Reverse, MVT::v4i16, 1}, // REV64
        {TTI::SK_Reverse, MVT::v8i8, 1},  // REV64
        // Splice can all be lowered as `ext`.
        {TTI::SK_Splice, MVT::v2i32, 1},
        {TTI::SK_Splice, MVT::v4i32, 1},
        {TTI::SK_Splice, MVT::v2i64, 1},
        {TTI::SK_Splice, MVT::v2f32, 1},
        {TTI::SK_Splice, MVT::v4f32, 1},
        {TTI::SK_Splice, MVT::v2f64, 1},
        {TTI::SK_Splice, MVT::v8f16, 1},
        {TTI::SK_Splice, MVT::v8bf16, 1},
        {TTI::SK_Splice, MVT::v8i16, 1},
        {TTI::SK_Splice, MVT::v16i8, 1},
        {TTI::SK_Splice, MVT::v4f16, 1},
        {TTI::SK_Splice, MVT::v4bf16, 1},
        {TTI::SK_Splice, MVT::v4i16, 1},
        {TTI::SK_Splice, MVT::v8i8, 1},
        // Broadcast shuffle kinds for scalable vectors
        {TTI::SK_Broadcast, MVT::nxv16i8, 1},
        {TTI::SK_Broadcast, MVT::nxv8i16, 1},
        {TTI::SK_Broadcast, MVT::nxv4i32, 1},
        {TTI::SK_Broadcast, MVT::nxv2i64, 1},
        {TTI::SK_Broadcast, MVT::nxv2f16, 1},
        {TTI::SK_Broadcast, MVT::nxv4f16, 1},
        {TTI::SK_Broadcast, MVT::nxv8f16, 1},
        {TTI::SK_Broadcast, MVT::nxv2bf16, 1},
        {TTI::SK_Broadcast, MVT::nxv4bf16, 1},
        {TTI::SK_Broadcast, MVT::nxv8bf16, 1},
        {TTI::SK_Broadcast, MVT::nxv2f32, 1},
        {TTI::SK_Broadcast, MVT::nxv4f32, 1},
        {TTI::SK_Broadcast, MVT::nxv2f64, 1},
        {TTI::SK_Broadcast, MVT::nxv16i1, 1},
        {TTI::SK_Broadcast, MVT::nxv8i1, 1},
        {TTI::SK_Broadcast, MVT::nxv4i1, 1},
        {TTI::SK_Broadcast, MVT::nxv2i1, 1},
        // Handle the cases for vector.reverse with scalable vectors
        {TTI::SK_Reverse, MVT::nxv16i8, 1},
        {TTI::SK_Reverse, MVT::nxv8i16, 1},
        {TTI::SK_Reverse, MVT::nxv4i32, 1},
        {TTI::SK_Reverse, MVT::nxv2i64, 1},
        {TTI::SK_Reverse, MVT::nxv2f16, 1},
        {TTI::SK_Reverse, MVT::nxv4f16, 1},
        {TTI::SK_Reverse, MVT::nxv8f16, 1},
        {TTI::SK_Reverse, MVT::nxv2bf16, 1},
        {TTI::SK_Reverse, MVT::nxv4bf16, 1},
        {TTI::SK_Reverse, MVT::nxv8bf16, 1},
        {TTI::SK_Reverse, MVT::nxv2f32, 1},
        {TTI::SK_Reverse, MVT::nxv4f32, 1},
        {TTI::SK_Reverse, MVT::nxv2f64, 1},
        {TTI::SK_Reverse, MVT::nxv16i1, 1},
        {TTI::SK_Reverse, MVT::nxv8i1, 1},
        {TTI::SK_Reverse, MVT::nxv4i1, 1},
        {TTI::SK_Reverse, MVT::nxv2i1, 1},
    };
    if (const auto *Entry = CostTableLookup(ShuffleTbl, Kind, LT.second))
      return LT.first * Entry->Cost;
  }
 
  if (Kind == TTI::SK_Splice && isa<ScalableVectorType>(SrcTy))
    return getSpliceCost(SrcTy, Index, CostKind);
 
  // Inserting a subvector can often be done with either a D, S or H register
  // move, so long as the inserted vector is "aligned".
  if (Kind == TTI::SK_InsertSubvector && LT.second.isFixedLengthVector() &&
      LT.second.getSizeInBits() <= 128 && SubTp) {
    std::pair<InstructionCost, MVT> SubLT = getTypeLegalizationCost(SubTp);
    if (SubLT.second.isVector()) {
      int NumElts = LT.second.getVectorNumElements();
      int NumSubElts = SubLT.second.getVectorNumElements();
      if ((Index % NumSubElts) == 0 && (NumElts % NumSubElts) == 0)
        return SubLT.first;
    }
  }
 
  // Restore optimal kind.
  if (IsExtractSubvector)
    Kind = TTI::SK_ExtractSubvector;
  return BaseT::getShuffleCost(Kind, DstTy, SrcTy, CostKind, Mask, Index, SubTp,
                               Args, CtxI);
}
 
static bool containsDecreasingPointers(Loop *TheLoop,
                                       PredicatedScalarEvolution *PSE,
                                       const DominatorTree &DT) {
  const auto &Strides = SymbolicStrideMap();
  for (BasicBlock *BB : TheLoop->blocks()) {
    // Scan the instructions in the block and look for addresses that are
    // consecutive and decreasing.
    for (Instruction &I : *BB) {
      if (isa<LoadInst>(&I) || isa<StoreInst>(&I)) {
        Value *Ptr = getLoadStorePointerOperand(&I);
        Type *AccessTy = getLoadStoreType(&I);
        // Analyze assuming predicates will be added, but discard them; this
        // query only guides tail-folding and must not add runtime checks to the
        // loop.
        SmallVector<const SCEVPredicate *> Predicates;
        if (getPtrStride(*PSE, AccessTy, Ptr, TheLoop, DT, Strides,
                         /*ShouldCheckWrap=*/false, &Predicates)
                .value_or(0) < 0)
          return true;
      }
    }
  }
  return false;
}
 
bool AArch64TTIImpl::preferFixedOverScalableIfEqualCost() const {
  return valueOr(ST->getCLOpts().sve_prefer_fixed_over_scalable_if_equal,
                 ST->useFixedOverScalableIfEqualCost());
}
 
unsigned AArch64TTIImpl::getEpilogueVectorizationMinVF() const {
  return ST->getEpilogueVectorizationMinVF();
}
 
bool AArch64TTIImpl::preferTailFoldingOverEpilogue(TailFoldingInfo *TFI) const {
  TailFoldingOpts DefaultOpts = ST->getSVETailFoldingDefaultOpts();
  if (!ST->hasSVE() || TailFoldingOptionLoc.isDisabled(DefaultOpts))
    return false;
 
  // We don't currently support vectorisation with interleaving for SVE - with
  // such loops we're better off not using tail-folding. This gives us a chance
  // to fall back on fixed-width vectorisation using NEON's ld2/st2/etc.
  if (TFI->IAI->hasGroups())
    return false;
 
  TailFoldingOpts Required = TailFoldingOpts::Disabled;
  if (TFI->LVL->getReductionVars().size())
    Required |= TailFoldingOpts::Reductions;
  if (TFI->LVL->getFixedOrderRecurrences().size())
    Required |= TailFoldingOpts::Recurrences;
 
  // We call this to discover whether any load/store pointers in the loop have
  // negative strides. This will require extra work to reverse the loop
  // predicate, which may be expensive.
  if (containsDecreasingPointers(TFI->LVL->getLoop(),
                                 TFI->LVL->getPredicatedScalarEvolution(),
                                 *TFI->LVL->getDominatorTree()))
    Required |= TailFoldingOpts::Reverse;
  if (Required == TailFoldingOpts::Disabled)
    Required |= TailFoldingOpts::Simple;
 
  if (!TailFoldingOptionLoc.satisfies(DefaultOpts, Required))
    return false;
 
  // Don't tail-fold for tight loops where we would be better off interleaving
  // with an unpredicated loop.
  unsigned NumInsns = 0;
  for (BasicBlock *BB : TFI->LVL->getLoop()->blocks()) {
    NumInsns += BB->size();
  }
 
  // We expect 4 of these to be a IV PHI, IV add, IV compare and branch.
  return NumInsns >= ST->getCLOpts().sve_tail_folding_insn_threshold;
}
 
InstructionCost
AArch64TTIImpl::getScalingFactorCost(Type *Ty, GlobalValue *BaseGV,
                                     StackOffset BaseOffset, bool HasBaseReg,
                                     int64_t Scale, unsigned AddrSpace) const {
  // Scaling factors are not free at all.
  // Operands                     | Rt Latency
  // -------------------------------------------
  // Rt, [Xn, Xm]                 | 4
  // -------------------------------------------
  // Rt, [Xn, Xm, lsl #imm]       | Rn: 4 Rm: 5
  // Rt, [Xn, Wm, <extend> #imm]  |
  TargetLoweringBase::AddrMode AM;
  AM.BaseGV = BaseGV;
  AM.BaseOffs = BaseOffset.getFixed();
  AM.HasBaseReg = HasBaseReg;
  AM.Scale = Scale;
  AM.ScalableOffset = BaseOffset.getScalable();
  if (getTLI()->isLegalAddressingMode(DL, AM, Ty, AddrSpace))
    // Scale represents reg2 * scale, thus account for 1 if
    // it is not equal to 0 or 1.
    return AM.Scale != 0 && AM.Scale != 1;
  return InstructionCost::getInvalid();
}
 
bool AArch64TTIImpl::shouldTreatInstructionLikeSelect(
    const Instruction *I) const {
  if (ST->getCLOpts().enable_aarch64_or_like_select) {
    // For the binary operators (e.g. or) we need to be more careful than
    // selects, here we only transform them if they are already at a natural
    // break point in the code - the end of a block with an unconditional
    // terminator.
    if (I->getOpcode() == Instruction::Or &&
        isa<UncondBrInst>(I->getNextNode()))
      return true;
 
    if (I->getOpcode() == Instruction::Add ||
        I->getOpcode() == Instruction::Sub)
      return true;
  }
  return BaseT::shouldTreatInstructionLikeSelect(I);
}
 
bool AArch64TTIImpl::isLSRCostLess(
    const TargetTransformInfo::LSRCost &C1,
    const TargetTransformInfo::LSRCost &C2) const {
  // AArch64 specific here is adding the number of instructions to the
  // comparison (though not as the first consideration, as some targets do),
  // preferring lower number of pre/post-index if numregs and insns are equal,
  // as typically pre/post-index instructions use extra processor resources,
  // along with changing the priority of the base additions.
  // TODO: Maybe a more nuanced tradeoff between instruction count
  // and number of registers? To be investigated at a later date.
  if (ST->getCLOpts().enable_aarch64_lsr_cost_opt)
    return std::tie(C1.NumRegs, C1.Insns, C1.NumPrePostIndex, C1.NumBaseAdds,
                    C1.AddRecCost, C1.NumIVMuls, C1.ScaleCost, C1.ImmCost,
                    C1.SetupCost) <
           std::tie(C2.NumRegs, C2.Insns, C2.NumPrePostIndex, C2.NumBaseAdds,
                    C2.AddRecCost, C2.NumIVMuls, C2.ScaleCost, C2.ImmCost,
                    C2.SetupCost);
 
  return TargetTransformInfoImplBase::isLSRCostLess(C1, C2);
}
 
static bool isSplatShuffle(Value *V) {
  if (auto *Shuf = dyn_cast<ShuffleVectorInst>(V))
    return all_equal(Shuf->getShuffleMask());
  return false;
}
 
/// Check if both Op1 and Op2 are shufflevector extracts of either the lower
/// or upper half of the vector elements.
static bool areExtractShuffleVectors(Value *Op1, Value *Op2,
                                     bool AllowSplat = false) {
  // Scalable types can't be extract shuffle vectors.
  if (Op1->getType()->isScalableTy() || Op2->getType()->isScalableTy())
    return false;
 
  auto areTypesHalfed = [](Value *FullV, Value *HalfV) {
    auto *FullTy = FullV->getType();
    auto *HalfTy = HalfV->getType();
    return FullTy->getPrimitiveSizeInBits().getFixedValue() ==
           2 * HalfTy->getPrimitiveSizeInBits().getFixedValue();
  };
 
  auto extractHalf = [](Value *FullV, Value *HalfV) {
    auto *FullVT = cast<FixedVectorType>(FullV->getType());
    auto *HalfVT = cast<FixedVectorType>(HalfV->getType());
    return FullVT->getNumElements() == 2 * HalfVT->getNumElements();
  };
 
  ArrayRef<int> M1, M2;
  Value *S1Op1 = nullptr, *S2Op1 = nullptr;
  if (!match(Op1, m_Shuffle(m_Value(S1Op1), m_Undef(), m_Mask(M1))) ||
      !match(Op2, m_Shuffle(m_Value(S2Op1), m_Undef(), m_Mask(M2))))
    return false;
 
  // If we allow splats, set S1Op1/S2Op1 to nullptr for the relevant arg so that
  // it is not checked as an extract below.
  if (AllowSplat && isSplatShuffle(Op1))
    S1Op1 = nullptr;
  if (AllowSplat && isSplatShuffle(Op2))
    S2Op1 = nullptr;
 
  // Check that the operands are half as wide as the result and we extract
  // half of the elements of the input vectors.
  if ((S1Op1 && (!areTypesHalfed(S1Op1, Op1) || !extractHalf(S1Op1, Op1))) ||
      (S2Op1 && (!areTypesHalfed(S2Op1, Op2) || !extractHalf(S2Op1, Op2))))
    return false;
 
  // Check the mask extracts either the lower or upper half of vector
  // elements.
  int M1Start = 0;
  int M2Start = 0;
  int NumElements = cast<FixedVectorType>(Op1->getType())->getNumElements() * 2;
  if ((S1Op1 &&
       !ShuffleVectorInst::isExtractSubvectorMask(M1, NumElements, M1Start)) ||
      (S2Op1 &&
       !ShuffleVectorInst::isExtractSubvectorMask(M2, NumElements, M2Start)))
    return false;
 
  if ((M1Start != 0 && M1Start != (NumElements / 2)) ||
      (M2Start != 0 && M2Start != (NumElements / 2)))
    return false;
  if (S1Op1 && S2Op1 && M1Start != M2Start)
    return false;
 
  return true;
}
 
/// Check if Ext1 and Ext2 are extends of the same type, doubling the bitwidth
/// of the vector elements.
static bool areExtractExts(Value *Ext1, Value *Ext2) {
  auto areExtDoubled = [](Instruction *Ext) {
    return Ext->getType()->getScalarSizeInBits() ==
           2 * Ext->getOperand(0)->getType()->getScalarSizeInBits();
  };
 
  if (!match(Ext1, m_ZExtOrSExt(m_Value())) ||
      !match(Ext2, m_ZExtOrSExt(m_Value())) ||
      !areExtDoubled(cast<Instruction>(Ext1)) ||
      !areExtDoubled(cast<Instruction>(Ext2)))
    return false;
 
  return true;
}
 
/// Check if Op could be used with vmull_high_p64 intrinsic.
static bool isOperandOfVmullHighP64(Value *Op) {
  Value *VectorOperand = nullptr;
  ConstantInt *ElementIndex = nullptr;
  return match(Op, m_ExtractElt(m_Value(VectorOperand),
                                m_ConstantInt(ElementIndex))) &&
         ElementIndex->getValue() == 1 &&
         isa<FixedVectorType>(VectorOperand->getType()) &&
         cast<FixedVectorType>(VectorOperand->getType())->getNumElements() == 2;
}
 
/// Check if Op1 and Op2 could be used with vmull_high_p64 intrinsic.
static bool areOperandsOfVmullHighP64(Value *Op1, Value *Op2) {
  return isOperandOfVmullHighP64(Op1) && isOperandOfVmullHighP64(Op2);
}
 
static bool shouldSinkVectorOfPtrs(Value *Ptrs, SmallVectorImpl<Use *> &Ops) {
  // Restrict ourselves to the form CodeGenPrepare typically constructs.
  auto *GEP = dyn_cast<GetElementPtrInst>(Ptrs);
  if (!GEP || GEP->getNumOperands() != 2)
    return false;
 
  Value *Base = GEP->getOperand(0);
  Value *Offsets = GEP->getOperand(1);
 
  // We only care about scalar_base+vector_offsets.
  if (Base->getType()->isVectorTy() || !Offsets->getType()->isVectorTy())
    return false;
 
  // Sink extends that would allow us to use 32-bit offset vectors.
  if (isa<SExtInst>(Offsets) || isa<ZExtInst>(Offsets)) {
    auto *OffsetsInst = cast<Instruction>(Offsets);
    if (OffsetsInst->getType()->getScalarSizeInBits() > 32 &&
        OffsetsInst->getOperand(0)->getType()->getScalarSizeInBits() <= 32)
      Ops.push_back(&GEP->getOperandUse(1));
  }
 
  // Sink the GEP.
  return true;
}
 
/// We want to sink following cases:
/// (add|sub|gep) A, ((mul|shl) vscale, imm); (add|sub|gep) A, vscale;
/// (add|sub|gep) A, ((mul|shl) zext(vscale), imm);
static bool shouldSinkVScale(Value *Op, SmallVectorImpl<Use *> &Ops) {
  if (match(Op, m_VScale()))
    return true;
  if (match(Op, m_Shl(m_VScale(), m_ConstantInt())) ||
      match(Op, m_Mul(m_VScale(), m_ConstantInt()))) {
    Ops.push_back(&cast<Instruction>(Op)->getOperandUse(0));
    return true;
  }
  if (match(Op, m_Shl(m_ZExt(m_VScale()), m_ConstantInt())) ||
      match(Op, m_Mul(m_ZExt(m_VScale()), m_ConstantInt()))) {
    Value *ZExtOp = cast<Instruction>(Op)->getOperand(0);
    Ops.push_back(&cast<Instruction>(ZExtOp)->getOperandUse(0));
    Ops.push_back(&cast<Instruction>(Op)->getOperandUse(0));
    return true;
  }
  return false;
}
 
static bool isFNeg(Value *Op) {
  // TODO: Confirm whether m_FNegOrFSubNegZero is correct.
  return match(Op, m_FNegOrFSubNegZero(m_Value()));
}
 
/// Check if sinking \p I's operands to I's basic block is profitable, because
/// the operands can be folded into a target instruction, e.g.
/// shufflevectors extracts and/or sext/zext can be folded into (u,s)subl(2).
bool AArch64TTIImpl::isProfitableToSinkOperands(Instruction *I,
                                                SmallVectorImpl<Use *> &Ops,
                                                const LoopInfo *LI) const {
  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
    switch (II->getIntrinsicID()) {
    case Intrinsic::aarch64_neon_smull:
    case Intrinsic::aarch64_neon_umull:
      if (areExtractShuffleVectors(II->getOperand(0), II->getOperand(1),
                                   /*AllowSplat=*/true)) {
        Ops.push_back(&II->getOperandUse(0));
        Ops.push_back(&II->getOperandUse(1));
        return true;
      }
      [[fallthrough]];
 
    case Intrinsic::fma:
    case Intrinsic::fmuladd:
      if (isa<VectorType>(I->getType()) &&
          cast<VectorType>(I->getType())->getElementType()->isHalfTy() &&
          !ST->hasFullFP16())
        return false;
 
      if (isFNeg(II->getOperand(0)))
        Ops.push_back(&II->getOperandUse(0));
      if (isFNeg(II->getOperand(1)))
        Ops.push_back(&II->getOperandUse(1));
 
      [[fallthrough]];
    case Intrinsic::aarch64_neon_sqdmull:
    case Intrinsic::aarch64_neon_sqdmulh:
    case Intrinsic::aarch64_neon_sqrdmulh:
      // Sink splats for index lane variants
      if (isSplatShuffle(II->getOperand(0)))
        Ops.push_back(&II->getOperandUse(0));
      if (isSplatShuffle(II->getOperand(1)))
        Ops.push_back(&II->getOperandUse(1));
      return !Ops.empty();
    case Intrinsic::aarch64_neon_fmlal:
    case Intrinsic::aarch64_neon_fmlal2:
    case Intrinsic::aarch64_neon_fmlsl:
    case Intrinsic::aarch64_neon_fmlsl2:
      // Sink splats for index lane variants
      if (isSplatShuffle(II->getOperand(1)))
        Ops.push_back(&II->getOperandUse(1));
      if (isSplatShuffle(II->getOperand(2)))
        Ops.push_back(&II->getOperandUse(2));
      return !Ops.empty();
    case Intrinsic::aarch64_sve_ptest_first:
    case Intrinsic::aarch64_sve_ptest_last:
      if (auto *IIOp = dyn_cast<IntrinsicInst>(II->getOperand(0)))
        if (IIOp->getIntrinsicID() == Intrinsic::aarch64_sve_ptrue)
          Ops.push_back(&II->getOperandUse(0));
      return !Ops.empty();
    case Intrinsic::aarch64_sme_write_horiz:
    case Intrinsic::aarch64_sme_write_vert:
    case Intrinsic::aarch64_sme_writeq_horiz:
    case Intrinsic::aarch64_sme_writeq_vert: {
      auto *Idx = dyn_cast<Instruction>(II->getOperand(1));
      if (!Idx || Idx->getOpcode() != Instruction::Add)
        return false;
      Ops.push_back(&II->getOperandUse(1));
      return true;
    }
    case Intrinsic::aarch64_sme_read_horiz:
    case Intrinsic::aarch64_sme_read_vert:
    case Intrinsic::aarch64_sme_readq_horiz:
    case Intrinsic::aarch64_sme_readq_vert:
    case Intrinsic::aarch64_sme_ld1b_vert:
    case Intrinsic::aarch64_sme_ld1h_vert:
    case Intrinsic::aarch64_sme_ld1w_vert:
    case Intrinsic::aarch64_sme_ld1d_vert:
    case Intrinsic::aarch64_sme_ld1q_vert:
    case Intrinsic::aarch64_sme_st1b_vert:
    case Intrinsic::aarch64_sme_st1h_vert:
    case Intrinsic::aarch64_sme_st1w_vert:
    case Intrinsic::aarch64_sme_st1d_vert:
    case Intrinsic::aarch64_sme_st1q_vert:
    case Intrinsic::aarch64_sme_ld1b_horiz:
    case Intrinsic::aarch64_sme_ld1h_horiz:
    case Intrinsic::aarch64_sme_ld1w_horiz:
    case Intrinsic::aarch64_sme_ld1d_horiz:
    case Intrinsic::aarch64_sme_ld1q_horiz:
    case Intrinsic::aarch64_sme_st1b_horiz:
    case Intrinsic::aarch64_sme_st1h_horiz:
    case Intrinsic::aarch64_sme_st1w_horiz:
    case Intrinsic::aarch64_sme_st1d_horiz:
    case Intrinsic::aarch64_sme_st1q_horiz: {
      auto *Idx = dyn_cast<Instruction>(II->getOperand(3));
      if (!Idx || Idx->getOpcode() != Instruction::Add)
        return false;
      Ops.push_back(&II->getOperandUse(3));
      return true;
    }
    case Intrinsic::aarch64_neon_pmull:
      if (!areExtractShuffleVectors(II->getOperand(0), II->getOperand(1)))
        return false;
      Ops.push_back(&II->getOperandUse(0));
      Ops.push_back(&II->getOperandUse(1));
      return true;
    case Intrinsic::aarch64_neon_pmull64:
      if (!areOperandsOfVmullHighP64(II->getArgOperand(0),
                                     II->getArgOperand(1)))
        return false;
      Ops.push_back(&II->getArgOperandUse(0));
      Ops.push_back(&II->getArgOperandUse(1));
      return true;
    case Intrinsic::masked_gather:
      if (!shouldSinkVectorOfPtrs(II->getArgOperand(0), Ops))
        return false;
      Ops.push_back(&II->getArgOperandUse(0));
      return true;
    case Intrinsic::masked_scatter:
      if (!shouldSinkVectorOfPtrs(II->getArgOperand(1), Ops))
        return false;
      Ops.push_back(&II->getArgOperandUse(1));
      return true;
    default:
      return false;
    }
  }
 
  auto ShouldSinkCondition = [](Value *Cond,
                                SmallVectorImpl<Use *> &Ops) -> bool {
    if (!isa<IntrinsicInst>(Cond))
      return false;
    auto *II = dyn_cast<IntrinsicInst>(Cond);
    if (II->getIntrinsicID() != Intrinsic::vector_reduce_or ||
        !isa<ScalableVectorType>(II->getOperand(0)->getType()))
      return false;
    if (isa<CmpInst>(II->getOperand(0)))
      Ops.push_back(&II->getOperandUse(0));
    return true;
  };
 
  switch (I->getOpcode()) {
  case Instruction::Store: {
    auto *SI = cast<StoreInst>(I);
    auto *II = dyn_cast<IntrinsicInst>(SI->getValueOperand());
    if (!SI->isSimple() || !II)
      break;
 
    if (!LI ||
        LI->getLoopFor(SI->getParent()) != LI->getLoopFor(II->getParent()))
      break;
 
    unsigned Factor = getInterleaveIntrinsicFactor(II->getIntrinsicID());
    if (Factor) {
      Ops.push_back(&SI->getOperandUse(0));
      return true;
    }
    break;
  }
  case Instruction::GetElementPtr:
  case Instruction::Add:
  case Instruction::Sub:
    // Sink vscales closer to uses for better isel
    for (unsigned Op = 0; Op < I->getNumOperands(); ++Op) {
      if (shouldSinkVScale(I->getOperand(Op), Ops)) {
        Ops.push_back(&I->getOperandUse(Op));
        return true;
      }
    }
    break;
  case Instruction::Select: {
    if (!ShouldSinkCondition(I->getOperand(0), Ops))
      return false;
 
    Ops.push_back(&I->getOperandUse(0));
    return true;
  }
  case Instruction::UncondBr:
    return false;
  case Instruction::CondBr: {
    if (!ShouldSinkCondition(cast<CondBrInst>(I)->getCondition(), Ops))
      return false;
 
    Ops.push_back(&I->getOperandUse(0));
    return true;
  }
  case Instruction::FMul:
    // fmul with contract flag can be combined with fadd into fma.
    // Sinking fneg into this block enables fmls pattern.
    if (cast<FPMathOperator>(I)->hasAllowContract()) {
      if (isFNeg(I->getOperand(0)))
        Ops.push_back(&I->getOperandUse(0));
      if (isFNeg(I->getOperand(1)))
        Ops.push_back(&I->getOperandUse(1));
    }
    break;
 
  // Type            |    BIC    |    ORN    |    EON
  // ----------------+-----------+-----------+-----------
  // scalar          |    Base   |    Base   |    Base
  // scalar w/shift  |     -     |     -     |     -
  // fixed vector    | NEON/Base | NEON/Base | BSL2N/Base
  // scalable vector |    SVE    |     -     |   BSL2N
  case Instruction::Xor:
    // EON only for scalars (possibly expanded fixed vectors)
    // and vectors using the SVE2/SME BSL2N instruction.
    if (I->getType()->isVectorTy() && ST->isNeonAvailable()) {
      bool HasBSL2N =
          ST->isSVEorStreamingSVEAvailable() && (ST->hasSVE2() || ST->hasSME());
      if (!HasBSL2N)
        break;
    }
    [[fallthrough]];
  case Instruction::And:
  case Instruction::Or:
    // Even though we could use the SVE2/SME BSL2N instruction,
    // it might pessimize with an extra MOV depending on register allocation.
    if (I->getOpcode() == Instruction::Or &&
        isa<ScalableVectorType>(I->getType()))
      break;
    // Shift can be fold into scalar AND/ORR/EOR,
    // but not the non-negated operand of BIC/ORN/EON.
    if (!(I->getType()->isVectorTy() && ST->hasNEON()) &&
        match(I, m_c_BinOp(m_Shift(m_Value(), m_ConstantInt()), m_Value())))
      break;
    for (auto &Op : I->operands()) {
      // (and/or/xor X, (not Y)) -> (bic/orn/eon X, Y)
      if (match(Op.get(), m_Not(m_Value()))) {
        Ops.push_back(&Op);
        return true;
      }
      // (and/or/xor X, (splat (not Y))) -> (bic/orn/eon X, (splat Y))
      if (match(Op.get(),
                m_Shuffle(m_InsertElt(m_Value(), m_Not(m_Value()), m_ZeroInt()),
                          m_Value(), m_ZeroMask()))) {
        Use &InsertElt = cast<Instruction>(Op)->getOperandUse(0);
        Use &Not = cast<Instruction>(InsertElt)->getOperandUse(1);
        Ops.push_back(&Not);
        Ops.push_back(&InsertElt);
        Ops.push_back(&Op);
        return true;
      }
    }
    break;
  default:
    break;
  }
 
  if (!I->getType()->isVectorTy())
    return !Ops.empty();
 
  switch (I->getOpcode()) {
  case Instruction::Sub:
  case Instruction::Add: {
    if (!areExtractExts(I->getOperand(0), I->getOperand(1)))
      return false;
 
    // If the exts' operands extract either the lower or upper elements, we
    // can sink them too.
    auto Ext1 = cast<Instruction>(I->getOperand(0));
    auto Ext2 = cast<Instruction>(I->getOperand(1));
    if (areExtractShuffleVectors(Ext1->getOperand(0), Ext2->getOperand(0))) {
      Ops.push_back(&Ext1->getOperandUse(0));
      Ops.push_back(&Ext2->getOperandUse(0));
    }
 
    Ops.push_back(&I->getOperandUse(0));
    Ops.push_back(&I->getOperandUse(1));
 
    return true;
  }
  case Instruction::Or: {
    // Pattern: Or(And(MaskValue, A), And(Not(MaskValue), B)) ->
    // bitselect(MaskValue, A, B) where Not(MaskValue) = Xor(MaskValue, -1)
    if (ST->hasNEON()) {
      // MainAnd refers to And instruction that has 'Not' as one of its operands
      Instruction *MainAnd, *OtherAnd;
      Value *A, *B;
      if (matchBSL(I, MainAnd, OtherAnd, A, B)) {
        Ops.push_back(
            &MainAnd->getOperandUse(MainAnd->getOperand(0) == A ? 1 : 0));
        Ops.push_back(&I->getOperandUse(0));
        Ops.push_back(&I->getOperandUse(1));
 
        return true;
      }
    }
 
    return false;
  }
  case Instruction::Mul: {
    auto ShouldSinkSplatForIndexedVariant = [](Value *V) {
      auto *Ty = cast<VectorType>(V->getType());
      // For SVE the lane-indexing is within 128-bits, so we can't fold splats.
      if (Ty->isScalableTy())
        return false;
 
      // Indexed variants of Mul exist for i16 and i32 element types only.
      return Ty->getScalarSizeInBits() == 16 || Ty->getScalarSizeInBits() == 32;
    };
 
    int NumZExts = 0, NumSExts = 0;
    for (auto &Op : I->operands()) {
      // Make sure we are not already sinking this operand
      if (any_of(Ops, [&](Use *U) { return U->get() == Op; }))
        continue;
 
      if (match(&Op, m_ZExtOrSExt(m_Value()))) {
        auto *Ext = cast<Instruction>(Op);
        auto *ExtOp = Ext->getOperand(0);
        if (isSplatShuffle(ExtOp) && ShouldSinkSplatForIndexedVariant(ExtOp))
          Ops.push_back(&Ext->getOperandUse(0));
        Ops.push_back(&Op);
 
        if (isa<SExtInst>(Ext)) {
          NumSExts++;
        } else {
          NumZExts++;
          // A zext(a) is also a sext(zext(a)), if we take more than 2 steps.
          if (Ext->getOperand(0)->getType()->getScalarSizeInBits() * 2 <
              I->getType()->getScalarSizeInBits())
            NumSExts++;
        }
 
        continue;
      }
 
      ShuffleVectorInst *Shuffle = dyn_cast<ShuffleVectorInst>(Op);
      if (!Shuffle)
        continue;
 
      // If the Shuffle is a splat and the operand is a zext/sext, sinking the
      // operand and the s/zext can help create indexed s/umull. This is
      // especially useful to prevent i64 mul being scalarized.
      if (isSplatShuffle(Shuffle) &&
          match(Shuffle->getOperand(0), m_ZExtOrSExt(m_Value()))) {
        Ops.push_back(&Shuffle->getOperandUse(0));
        Ops.push_back(&Op);
        if (match(Shuffle->getOperand(0), m_SExt(m_Value())))
          NumSExts++;
        else
          NumZExts++;
        continue;
      }
 
      Value *ShuffleOperand = Shuffle->getOperand(0);
      InsertElementInst *Insert = dyn_cast<InsertElementInst>(ShuffleOperand);
      if (!Insert)
        continue;
 
      Instruction *OperandInstr = dyn_cast<Instruction>(Insert->getOperand(1));
      if (!OperandInstr)
        continue;
 
      ConstantInt *ElementConstant =
          dyn_cast<ConstantInt>(Insert->getOperand(2));
      // Check that the insertelement is inserting into element 0
      if (!ElementConstant || !ElementConstant->isZero())
        continue;
 
      unsigned Opcode = OperandInstr->getOpcode();
      if (Opcode == Instruction::SExt)
        NumSExts++;
      else if (Opcode == Instruction::ZExt)
        NumZExts++;
      else {
        // If we find that the top bits are known 0, then we can sink and allow
        // the backend to generate a umull.
        unsigned Bitwidth = I->getType()->getScalarSizeInBits();
        APInt UpperMask = APInt::getHighBitsSet(Bitwidth, Bitwidth / 2);
        if (!MaskedValueIsZero(OperandInstr, UpperMask, DL))
          continue;
        NumZExts++;
      }
 
      // And(Load) is excluded to prevent CGP getting stuck in a loop of sinking
      // the And, just to hoist it again back to the load.
      if (!match(OperandInstr, m_And(m_Load(m_Value()), m_Value())))
        Ops.push_back(&Insert->getOperandUse(1));
      Ops.push_back(&Shuffle->getOperandUse(0));
      Ops.push_back(&Op);
    }
 
    // It is profitable to sink if we found two of the same type of extends.
    if (!Ops.empty() && (NumSExts == 2 || NumZExts == 2))
      return true;
 
    // Otherwise, see if we should sink splats for indexed variants.
    if (!ShouldSinkSplatForIndexedVariant(I))
      return false;
 
    Ops.clear();
    if (isSplatShuffle(I->getOperand(0)))
      Ops.push_back(&I->getOperandUse(0));
    if (isSplatShuffle(I->getOperand(1)))
      Ops.push_back(&I->getOperandUse(1));
 
    return !Ops.empty();
  }
  case Instruction::FMul: {
    // For SVE the lane-indexing is within 128-bits, so we can't fold splats.
    if (I->getType()->isScalableTy())
      return !Ops.empty();
 
    if (cast<VectorType>(I->getType())->getElementType()->isHalfTy() &&
        !ST->hasFullFP16())
      return !Ops.empty();
 
    // Sink splats for index lane variants
    if (isSplatShuffle(I->getOperand(0)))
      Ops.push_back(&I->getOperandUse(0));
    if (isSplatShuffle(I->getOperand(1)))
      Ops.push_back(&I->getOperandUse(1));
    return !Ops.empty();
  }
  default:
    return false;
  }
  return false;
}
 
bool AArch64TTIImpl::isLegalMaskedCompressStore(Type *DataType,
                                                Align Alignment) const {
  if (!(ST->isSVEAvailable() ||
        (ST->isSVEorStreamingSVEAvailable() && ST->hasSME2p2())))
    return false;
 
  if (isa<FixedVectorType>(DataType) &&
      DataType->getPrimitiveSizeInBits().getFixedValue() < 128)
    return false;
 
  if (!isa<VectorType>(DataType))
    return isElementTypeLegalForScalableVector(DataType);
 
  auto LT = getTypeLegalizationCost(DataType);
  if (!LT.first.isValid() || !LT.second.isVector())
    return false;
 
  // Use the i32 or i64 compact instructions for f16/bf16 unpacked types.
  LLVMContext &Ctx = DataType->getContext();
  switch (LT.second.SimpleTy) {
  case MVT::nxv2f16:
  case MVT::nxv2bf16:
    return isElementTypeLegalForCompressStore(Type::getInt64Ty(Ctx));
  case MVT::nxv4f16:
  case MVT::nxv4bf16:
    return isElementTypeLegalForCompressStore(Type::getInt32Ty(Ctx));
  default:
    break;
  }
 
  return isElementTypeLegalForCompressStore(
      EVT(LT.second.getScalarType()).getTypeForEVT(Ctx));
}