X86ISelLowering.cpp

Go to the documentation of this file.

//===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file defines the interfaces that X86 uses to lower LLVM code into a
// selection DAG.
//
//===----------------------------------------------------------------------===//
 
#include "X86ISelLowering.h"
#include "MCTargetDesc/X86ShuffleDecode.h"
#include "X86.h"
#include "X86FrameLowering.h"
#include "X86InstrBuilder.h"
#include "X86IntrinsicsInfo.h"
#include "X86MachineFunctionInfo.h"
#include "X86TargetMachine.h"
#include "llvm/ADT/SmallBitVector.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/ADT/StringSwitch.h"
#include "llvm/Analysis/BlockFrequencyInfo.h"
#include "llvm/Analysis/ProfileSummaryInfo.h"
#include "llvm/Analysis/VectorUtils.h"
#include "llvm/CodeGen/LivePhysRegs.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineLoopInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SDPatternMatch.h"
#include "llvm/CodeGen/SelectionDAGNodes.h"
#include "llvm/CodeGen/TargetLowering.h"
#include "llvm/CodeGen/WinEHFuncInfo.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/EHPersonalities.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalAlias.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/PatternMatch.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCSymbol.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/KnownBits.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
#include <algorithm>
#include <bitset>
#include <cctype>
#include <numeric>
using namespace llvm;
 
#define DEBUG_TYPE "x86-isel"
 
static cl::opt<int> ExperimentalPrefInnermostLoopAlignment(
    "x86-experimental-pref-innermost-loop-alignment", cl::init(4),
    cl::desc(
        "Sets the preferable loop alignment for experiments (as log2 bytes) "
        "for innermost loops only. If specified, this option overrides "
        "alignment set by x86-experimental-pref-loop-alignment."),
    cl::Hidden);
 
static cl::opt<int> BrMergingBaseCostThresh(
    "x86-br-merging-base-cost", cl::init(2),
    cl::desc(
        "Sets the cost threshold for when multiple conditionals will be merged "
        "into one branch versus be split in multiple branches. Merging "
        "conditionals saves branches at the cost of additional instructions. "
        "This value sets the instruction cost limit, below which conditionals "
        "will be merged, and above which conditionals will be split. Set to -1 "
        "to never merge branches."),
    cl::Hidden);
 
static cl::opt<int> BrMergingCcmpBias(
    "x86-br-merging-ccmp-bias", cl::init(6),
    cl::desc("Increases 'x86-br-merging-base-cost' in cases that the target "
             "supports conditional compare instructions."),
    cl::Hidden);
 
static cl::opt<bool>
    WidenShift("x86-widen-shift", cl::init(true),
               cl::desc("Replace narrow shifts with wider shifts."),
               cl::Hidden);
 
static cl::opt<int> BrMergingLikelyBias(
    "x86-br-merging-likely-bias", cl::init(0),
    cl::desc("Increases 'x86-br-merging-base-cost' in cases that it is likely "
             "that all conditionals will be executed. For example for merging "
             "the conditionals (a == b && c > d), if its known that a == b is "
             "likely, then it is likely that if the conditionals are split "
             "both sides will be executed, so it may be desirable to increase "
             "the instruction cost threshold. Set to -1 to never merge likely "
             "branches."),
    cl::Hidden);
 
static cl::opt<int> BrMergingUnlikelyBias(
    "x86-br-merging-unlikely-bias", cl::init(-1),
    cl::desc(
        "Decreases 'x86-br-merging-base-cost' in cases that it is unlikely "
        "that all conditionals will be executed. For example for merging "
        "the conditionals (a == b && c > d), if its known that a == b is "
        "unlikely, then it is unlikely that if the conditionals are split "
        "both sides will be executed, so it may be desirable to decrease "
        "the instruction cost threshold. Set to -1 to never merge unlikely "
        "branches."),
    cl::Hidden);
 
static cl::opt<bool> MulConstantOptimization(
    "mul-constant-optimization", cl::init(true),
    cl::desc("Replace 'mul x, Const' with more effective instructions like "
             "SHIFT, LEA, etc."),
    cl::Hidden);
 
X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
                                     const X86Subtarget &STI)
    : TargetLowering(TM, STI), Subtarget(STI) {
  bool UseX87 = !Subtarget.useSoftFloat() && Subtarget.hasX87();
  MVT PtrVT = MVT::getIntegerVT(TM.getPointerSizeInBits(0));
 
  // Set up the TargetLowering object.
 
  // X86 is weird. It always uses i8 for shift amounts and setcc results.
  setBooleanContents(ZeroOrOneBooleanContent);
  // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
 
  // X86 instruction cache is coherent with its data cache so we can use the
  // default expansion to a no-op.
  setOperationAction(ISD::CLEAR_CACHE, MVT::Other, Expand);
 
  // For 64-bit, since we have so many registers, use the ILP scheduler.
  // For 32-bit, use the register pressure specific scheduling.
  // For Atom, always use ILP scheduling.
  if (Subtarget.isAtom())
    setSchedulingPreference(Sched::ILP);
  else if (Subtarget.is64Bit())
    setSchedulingPreference(Sched::ILP);
  else
    setSchedulingPreference(Sched::RegPressure);
  const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo();
  setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
 
  // Bypass expensive divides and use cheaper ones.
  if (TM.getOptLevel() >= CodeGenOptLevel::Default) {
    if (Subtarget.hasSlowDivide32())
      addBypassSlowDiv(32, 8);
    if (Subtarget.hasSlowDivide64() && Subtarget.is64Bit())
      addBypassSlowDiv(64, 32);
  }
 
  if (Subtarget.canUseCMPXCHG16B())
    setMaxAtomicSizeInBitsSupported(128);
  else if (Subtarget.canUseCMPXCHG8B())
    setMaxAtomicSizeInBitsSupported(64);
  else
    setMaxAtomicSizeInBitsSupported(32);
 
  setMaxDivRemBitWidthSupported(Subtarget.is64Bit() ? 128 : 64);
 
  setMaxLargeFPConvertBitWidthSupported(128);
 
  // Set up the register classes.
  addRegisterClass(MVT::i8, &X86::GR8RegClass);
  addRegisterClass(MVT::i16, &X86::GR16RegClass);
  addRegisterClass(MVT::i32, &X86::GR32RegClass);
  if (Subtarget.is64Bit())
    addRegisterClass(MVT::i64, &X86::GR64RegClass);
 
  for (MVT VT : MVT::integer_valuetypes())
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
 
  // We don't accept any truncstore of integer registers.
  setTruncStoreAction(MVT::i64, MVT::i32, Expand);
  setTruncStoreAction(MVT::i64, MVT::i16, Expand);
  setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
  setTruncStoreAction(MVT::i32, MVT::i16, Expand);
  setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
  setTruncStoreAction(MVT::i16, MVT::i8,  Expand);
 
  setTruncStoreAction(MVT::f64, MVT::f32, Expand);
 
  // SETOEQ and SETUNE require checking two conditions.
  for (auto VT : {MVT::f32, MVT::f64, MVT::f80}) {
    setCondCodeAction(ISD::SETOEQ, VT, Expand);
    setCondCodeAction(ISD::SETUNE, VT, Expand);
  }
 
  // Integer absolute.
  if (Subtarget.canUseCMOV()) {
    setOperationAction(ISD::ABS            , MVT::i16  , Custom);
    setOperationAction(ISD::ABS            , MVT::i32  , Custom);
    if (Subtarget.is64Bit())
      setOperationAction(ISD::ABS          , MVT::i64  , Custom);
  }
 
  // Absolute difference.
  for (auto Op : {ISD::ABDS, ISD::ABDU}) {
    setOperationAction(Op                  , MVT::i8   , Custom);
    setOperationAction(Op                  , MVT::i16  , Custom);
    setOperationAction(Op                  , MVT::i32  , Custom);
    if (Subtarget.is64Bit())
     setOperationAction(Op                 , MVT::i64  , Custom);
  }
 
  // Signed saturation subtraction.
  setOperationAction(ISD::SSUBSAT          , MVT::i8   , Custom);
  setOperationAction(ISD::SSUBSAT          , MVT::i16  , Custom);
  setOperationAction(ISD::SSUBSAT          , MVT::i32  , Custom);
  if (Subtarget.is64Bit())
    setOperationAction(ISD::SSUBSAT        , MVT::i64  , Custom);
 
  // Funnel shifts.
  for (auto ShiftOp : {ISD::FSHL, ISD::FSHR}) {
    // For slow shld targets we only lower for code size.
    LegalizeAction ShiftDoubleAction = Subtarget.isSHLDSlow() ? Custom : Legal;
 
    setOperationAction(ShiftOp             , MVT::i8   , Custom);
    setOperationAction(ShiftOp             , MVT::i16  , Custom);
    setOperationAction(ShiftOp             , MVT::i32  , ShiftDoubleAction);
    if (Subtarget.is64Bit())
      setOperationAction(ShiftOp           , MVT::i64  , ShiftDoubleAction);
  }
 
  if (!Subtarget.useSoftFloat()) {
    // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
    // operation.
    setOperationAction(ISD::UINT_TO_FP,        MVT::i8, Promote);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i8, Promote);
    setOperationAction(ISD::UINT_TO_FP,        MVT::i16, Promote);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i16, Promote);
    // We have an algorithm for SSE2, and we turn this into a 64-bit
    // FILD or VCVTUSI2SS/SD for other targets.
    setOperationAction(ISD::UINT_TO_FP,        MVT::i32, Custom);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i32, Custom);
    // We have an algorithm for SSE2->double, and we turn this into a
    // 64-bit FILD followed by conditional FADD for other targets.
    setOperationAction(ISD::UINT_TO_FP,        MVT::i64, Custom);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i64, Custom);
 
    // Promote i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
    // this operation.
    setOperationAction(ISD::SINT_TO_FP,        MVT::i8, Promote);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i8, Promote);
    // SSE has no i16 to fp conversion, only i32. We promote in the handler
    // to allow f80 to use i16 and f64 to use i16 with sse1 only
    setOperationAction(ISD::SINT_TO_FP,        MVT::i16, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i16, Custom);
    // f32 and f64 cases are Legal with SSE1/SSE2, f80 case is not
    setOperationAction(ISD::SINT_TO_FP,        MVT::i32, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i32, Custom);
    // In 32-bit mode these are custom lowered.  In 64-bit mode F32 and F64
    // are Legal, f80 is custom lowered.
    setOperationAction(ISD::SINT_TO_FP,        MVT::i64, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i64, Custom);
 
    // Promote i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
    // this operation.
    setOperationAction(ISD::FP_TO_SINT,        MVT::i8,  Promote);
    // FIXME: This doesn't generate invalid exception when it should. PR44019.
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i8,  Promote);
    setOperationAction(ISD::FP_TO_SINT,        MVT::i16, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i16, Custom);
    setOperationAction(ISD::FP_TO_SINT,        MVT::i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i32, Custom);
    // In 32-bit mode these are custom lowered.  In 64-bit mode F32 and F64
    // are Legal, f80 is custom lowered.
    setOperationAction(ISD::FP_TO_SINT,        MVT::i64, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i64, Custom);
 
    // Handle FP_TO_UINT by promoting the destination to a larger signed
    // conversion.
    setOperationAction(ISD::FP_TO_UINT,        MVT::i8,  Promote);
    // FIXME: This doesn't generate invalid exception when it should. PR44019.
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i8,  Promote);
    setOperationAction(ISD::FP_TO_UINT,        MVT::i16, Promote);
    // FIXME: This doesn't generate invalid exception when it should. PR44019.
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i16, Promote);
    setOperationAction(ISD::FP_TO_UINT,        MVT::i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i32, Custom);
    setOperationAction(ISD::FP_TO_UINT,        MVT::i64, Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i64, Custom);
 
    setOperationAction(ISD::LRINT,             MVT::f32, Custom);
    setOperationAction(ISD::LRINT,             MVT::f64, Custom);
    setOperationAction(ISD::LLRINT,            MVT::f32, Custom);
    setOperationAction(ISD::LLRINT,            MVT::f64, Custom);
 
    if (!Subtarget.is64Bit() && Subtarget.hasX87()) {
      setOperationAction(ISD::LRINT,  MVT::i64, Custom);
      setOperationAction(ISD::LLRINT, MVT::i64, Custom);
    }
  }
 
  if (Subtarget.hasSSE2()) {
    // Custom lowering for saturating float to int conversions.
    // We handle promotion to larger result types manually.
    for (MVT VT : { MVT::i8, MVT::i16, MVT::i32 }) {
      setOperationAction(ISD::FP_TO_UINT_SAT, VT, Custom);
      setOperationAction(ISD::FP_TO_SINT_SAT, VT, Custom);
    }
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i64, Custom);
      setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i64, Custom);
    }
  }
  if (Subtarget.hasAVX10_2()) {
    for (MVT VT : {MVT::v8i8, MVT::v16i8, MVT::v32i8}) {
      setOperationAction(ISD::FP_TO_UINT_SAT, VT, Custom);
      setOperationAction(ISD::FP_TO_SINT_SAT, VT, Custom);
    }
    setOperationAction(ISD::FP_TO_UINT_SAT, MVT::v2i32, Custom);
    setOperationAction(ISD::FP_TO_SINT_SAT, MVT::v2i32, Custom);
    setOperationAction(ISD::FP_TO_UINT_SAT, MVT::v8i64, Legal);
    setOperationAction(ISD::FP_TO_SINT_SAT, MVT::v8i64, Legal);
    for (MVT VT : {MVT::i32, MVT::v4i32, MVT::v8i32, MVT::v16i32, MVT::v2i64,
                   MVT::v4i64}) {
      setOperationAction(ISD::FP_TO_UINT_SAT, VT, Legal);
      setOperationAction(ISD::FP_TO_SINT_SAT, VT, Legal);
    }
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::FP_TO_UINT_SAT, MVT::i64, Legal);
      setOperationAction(ISD::FP_TO_SINT_SAT, MVT::i64, Legal);
    }
  }
 
  // Handle address space casts between mixed sized pointers.
  setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom);
  setOperationAction(ISD::ADDRSPACECAST, MVT::i64, Custom);
 
  // TODO: when we have SSE, these could be more efficient, by using movd/movq.
  if (!Subtarget.hasSSE2()) {
    setOperationAction(ISD::BITCAST        , MVT::f32  , Expand);
    setOperationAction(ISD::BITCAST        , MVT::i32  , Expand);
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::BITCAST      , MVT::f64  , Expand);
      // Without SSE, i64->f64 goes through memory.
      setOperationAction(ISD::BITCAST      , MVT::i64  , Expand);
    }
  } else if (!Subtarget.is64Bit())
    setOperationAction(ISD::BITCAST      , MVT::i64  , Custom);
 
  // Scalar integer divide and remainder are lowered to use operations that
  // produce two results, to match the available instructions. This exposes
  // the two-result form to trivial CSE, which is able to combine x/y and x%y
  // into a single instruction.
  //
  // Scalar integer multiply-high is also lowered to use two-result
  // operations, to match the available instructions. However, plain multiply
  // (low) operations are left as Legal, as there are single-result
  // instructions for this in x86. Using the two-result multiply instructions
  // when both high and low results are needed must be arranged by dagcombine.
  for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
    setOperationAction(ISD::MULHS, VT, Expand);
    setOperationAction(ISD::MULHU, VT, Expand);
    setOperationAction(ISD::SDIV, VT, Expand);
    setOperationAction(ISD::UDIV, VT, Expand);
    setOperationAction(ISD::SREM, VT, Expand);
    setOperationAction(ISD::UREM, VT, Expand);
  }
 
  setOperationAction(ISD::BR_JT            , MVT::Other, Expand);
  setOperationAction(ISD::BRCOND           , MVT::Other, Custom);
  for (auto VT : { MVT::f32, MVT::f64, MVT::f80, MVT::f128,
                   MVT::i8,  MVT::i16, MVT::i32, MVT::i64 }) {
    setOperationAction(ISD::BR_CC,     VT, Expand);
    setOperationAction(ISD::SELECT_CC, VT, Expand);
  }
  if (Subtarget.is64Bit())
    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16  , Legal);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8   , Legal);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1   , Expand);
 
  setOperationAction(ISD::FREM, MVT::f32, LibCall);
  setOperationAction(ISD::FREM, MVT::f64, LibCall);
  setOperationAction(ISD::FREM, MVT::f80, LibCall);
  setOperationAction(ISD::FREM, MVT::f128, LibCall);
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasX87()) {
    setOperationAction(ISD::GET_ROUNDING   , MVT::i32  , Custom);
    setOperationAction(ISD::SET_ROUNDING   , MVT::Other, Custom);
    setOperationAction(ISD::GET_FPENV_MEM  , MVT::Other, Custom);
    setOperationAction(ISD::SET_FPENV_MEM  , MVT::Other, Custom);
    setOperationAction(ISD::RESET_FPENV    , MVT::Other, Custom);
  }
 
  // Promote the i8 variants and force them on up to i32 which has a shorter
  // encoding.
  setOperationPromotedToType(ISD::CTTZ, MVT::i8, MVT::i32);
  setOperationPromotedToType(ISD::CTTZ_ZERO_POISON, MVT::i8, MVT::i32);
  // Promoted i16. tzcntw has a false dependency on Intel CPUs. For BSF, we emit
  // a REP prefix to encode it as TZCNT for modern CPUs so it makes sense to
  // promote that too.
  setOperationPromotedToType(ISD::CTTZ, MVT::i16, MVT::i32);
  setOperationPromotedToType(ISD::CTTZ_ZERO_POISON, MVT::i16, MVT::i32);
 
  if (!Subtarget.hasBMI()) {
    setOperationAction(ISD::CTTZ, MVT::i32, Custom);
    setOperationAction(ISD::CTTZ_ZERO_POISON, MVT::i32, Legal);
    if (Subtarget.is64Bit()) {
      setOperationAction(ISD::CTTZ, MVT::i64, Custom);
      setOperationAction(ISD::CTTZ_ZERO_POISON, MVT::i64, Legal);
    }
  }
 
  if (Subtarget.hasLZCNT()) {
    // When promoting the i8 variants, force them to i32 for a shorter
    // encoding.
    setOperationPromotedToType(ISD::CTLZ, MVT::i8, MVT::i32);
    setOperationPromotedToType(ISD::CTLZ_ZERO_POISON, MVT::i8, MVT::i32);
  } else {
    for (auto VT : {MVT::i8, MVT::i16, MVT::i32, MVT::i64}) {
      if (VT == MVT::i64 && !Subtarget.is64Bit())
        continue;
      setOperationAction(ISD::CTLZ, VT, Custom);
      setOperationAction(ISD::CTLZ_ZERO_POISON, VT, Custom);
    }
  }
 
  for (auto Op : {ISD::FP16_TO_FP, ISD::STRICT_FP16_TO_FP, ISD::FP_TO_FP16,
                  ISD::STRICT_FP_TO_FP16}) {
    // Special handling for half-precision floating point conversions.
    // If we don't have F16C support, then lower half float conversions
    // into library calls.
    setOperationAction(
        Op, MVT::f32,
        (!Subtarget.useSoftFloat() && Subtarget.hasF16C()) ? Custom : Expand);
    // There's never any support for operations beyond MVT::f32.
    setOperationAction(Op, MVT::f64, Expand);
    setOperationAction(Op, MVT::f80, Expand);
    setOperationAction(Op, MVT::f128, Expand);
  }
 
  for (auto VT : {MVT::f32, MVT::f64, MVT::f80, MVT::f128}) {
    setOperationAction(ISD::STRICT_FP_TO_BF16, VT, Expand);
    setOperationAction(ISD::STRICT_BF16_TO_FP, VT, Expand);
  }
 
  for (MVT VT : {MVT::f32, MVT::f64, MVT::f80, MVT::f128}) {
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::bf16, Expand);
    setTruncStoreAction(VT, MVT::f16, Expand);
    setTruncStoreAction(VT, MVT::bf16, Expand);
 
    setOperationAction(ISD::BF16_TO_FP, VT, Expand);
    setOperationAction(ISD::FP_TO_BF16, VT, Custom);
  }
 
  setOperationAction(ISD::PARITY, MVT::i8, Custom);
  setOperationAction(ISD::PARITY, MVT::i16, Custom);
  setOperationAction(ISD::PARITY, MVT::i32, Custom);
  if (Subtarget.is64Bit())
    setOperationAction(ISD::PARITY, MVT::i64, Custom);
  if (Subtarget.hasPOPCNT()) {
    setOperationPromotedToType(ISD::CTPOP, MVT::i8, MVT::i32);
    // popcntw is longer to encode than popcntl and also has a false dependency
    // on the dest that popcntl hasn't had since Cannon Lake.
    setOperationPromotedToType(ISD::CTPOP, MVT::i16, MVT::i32);
  } else {
    setOperationAction(ISD::CTPOP          , MVT::i8   , Custom);
    setOperationAction(ISD::CTPOP          , MVT::i16  , Custom);
    setOperationAction(ISD::CTPOP          , MVT::i32  , Custom);
    setOperationAction(ISD::CTPOP          , MVT::i64  , Custom);
  }
 
  setOperationAction(ISD::READCYCLECOUNTER , MVT::i64  , Custom);
 
  if (!Subtarget.hasMOVBE())
    setOperationAction(ISD::BSWAP          , MVT::i16  , Expand);
 
  // X86 wants to expand cmov itself.
  for (auto VT : { MVT::f32, MVT::f64, MVT::f80, MVT::f128 }) {
    setOperationAction(ISD::SELECT, VT, Custom);
    setOperationAction(ISD::SETCC, VT, Custom);
    setOperationAction(ISD::STRICT_FSETCC, VT, Custom);
    setOperationAction(ISD::STRICT_FSETCCS, VT, Custom);
  }
  for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
    if (VT == MVT::i64 && !Subtarget.is64Bit())
      continue;
    setOperationAction(ISD::SELECT, VT, Custom);
    setOperationAction(ISD::SETCC,  VT, Custom);
  }
 
  setOperationAction(ISD::COND_LOOP, MVT::Other, Custom);
 
  // Custom action for SELECT MMX and expand action for SELECT_CC MMX
  setOperationAction(ISD::SELECT, MVT::x86mmx, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::x86mmx, Expand);
 
  setOperationAction(ISD::EH_RETURN       , MVT::Other, Custom);
  // NOTE: EH_SJLJ_SETJMP/_LONGJMP are not recommended, since
  // LLVM/Clang supports zero-cost DWARF and SEH exception handling.
  setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
  setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
  setOperationAction(ISD::EH_SJLJ_SETUP_DISPATCH, MVT::Other, Custom);
 
  // Darwin ABI issue.
  for (auto VT : { MVT::i32, MVT::i64 }) {
    if (VT == MVT::i64 && !Subtarget.is64Bit())
      continue;
    setOperationAction(ISD::ConstantPool    , VT, Custom);
    setOperationAction(ISD::JumpTable       , VT, Custom);
    setOperationAction(ISD::GlobalAddress   , VT, Custom);
    setOperationAction(ISD::GlobalTLSAddress, VT, Custom);
    setOperationAction(ISD::ExternalSymbol  , VT, Custom);
    setOperationAction(ISD::BlockAddress    , VT, Custom);
  }
 
  // 64-bit shl, sra, srl (iff 32-bit x86)
  for (auto VT : { MVT::i32, MVT::i64 }) {
    if (VT == MVT::i64 && !Subtarget.is64Bit())
      continue;
    setOperationAction(ISD::SHL_PARTS, VT, Custom);
    setOperationAction(ISD::SRA_PARTS, VT, Custom);
    setOperationAction(ISD::SRL_PARTS, VT, Custom);
  }
 
  if (Subtarget.hasSSEPrefetch())
    setOperationAction(ISD::PREFETCH      , MVT::Other, Custom);
 
  setOperationAction(ISD::ATOMIC_FENCE  , MVT::Other, Custom);
 
  // Expand certain atomics
  for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
    setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_ADD, VT, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_OR, VT, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_XOR, VT, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_AND, VT, Custom);
    setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
  }
 
  if (!Subtarget.is64Bit())
    setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
 
  if (Subtarget.is64Bit() && Subtarget.hasAVX()) {
    // All CPUs supporting AVX will atomically load/store aligned 128-bit
    // values, so we can emit [V]MOVAPS/[V]MOVDQA.
    setOperationAction(ISD::ATOMIC_LOAD, MVT::i128, Custom);
    setOperationAction(ISD::ATOMIC_STORE, MVT::i128, Custom);
  }
 
  if (Subtarget.canUseCMPXCHG16B())
    setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
 
  // FIXME - use subtarget debug flags
  if (!Subtarget.isTargetDarwin() && !Subtarget.isTargetELF() &&
      !Subtarget.isTargetCygMing() && !Subtarget.isTargetWin64() &&
      TM.Options.ExceptionModel != ExceptionHandling::SjLj) {
    setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
  }
 
  setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
  setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
 
  setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
  setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
 
  setOperationAction(ISD::TRAP, MVT::Other, Legal);
  setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
  if (Subtarget.isTargetPS())
    setOperationAction(ISD::UBSANTRAP, MVT::Other, Expand);
  else
    setOperationAction(ISD::UBSANTRAP, MVT::Other, Legal);
 
  // VASTART needs to be custom lowered to use the VarArgsFrameIndex
  setOperationAction(ISD::VASTART           , MVT::Other, Custom);
  setOperationAction(ISD::VAEND             , MVT::Other, Expand);
  bool Is64Bit = Subtarget.is64Bit();
  setOperationAction(ISD::VAARG,  MVT::Other, Is64Bit ? Custom : Expand);
  setOperationAction(ISD::VACOPY, MVT::Other, Is64Bit ? Custom : Expand);
 
  setOperationAction(ISD::STACKSAVE,          MVT::Other, Expand);
  setOperationAction(ISD::STACKRESTORE,       MVT::Other, Expand);
 
  setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
 
  // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
  setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
  setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
 
  setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Legal);
 
  auto setF16Action = [&] (MVT VT, LegalizeAction Action) {
    setOperationAction(ISD::FABS, VT, Action);
    setOperationAction(ISD::FNEG, VT, Action);
    setOperationAction(ISD::FCOPYSIGN, VT, Expand);
    setOperationAction(ISD::FREM, VT, Action);
    setOperationAction(ISD::FMA, VT, Action);
    setOperationAction(ISD::FMINNUM, VT, Action);
    setOperationAction(ISD::FMAXNUM, VT, Action);
    setOperationAction(ISD::FMINIMUM, VT, Action);
    setOperationAction(ISD::FMAXIMUM, VT, Action);
    setOperationAction(ISD::FMINIMUMNUM, VT, Action);
    setOperationAction(ISD::FMAXIMUMNUM, VT, Action);
    setOperationAction(ISD::FSIN, VT, Action);
    setOperationAction(ISD::FCOS, VT, Action);
    setOperationAction(ISD::FSINCOS, VT, Action);
    setOperationAction(ISD::FTAN, VT, Action);
    setOperationAction(ISD::FSQRT, VT, Action);
    setOperationAction(ISD::FPOW, VT, Action);
    setOperationAction(ISD::FPOWI, VT, Action);
    setOperationAction(ISD::FLOG, VT, Action);
    setOperationAction(ISD::FLOG2, VT, Action);
    setOperationAction(ISD::FLOG10, VT, Action);
    setOperationAction(ISD::FEXP, VT, Action);
    setOperationAction(ISD::FEXP2, VT, Action);
    setOperationAction(ISD::FEXP10, VT, Action);
    setOperationAction(ISD::FCEIL, VT, Action);
    setOperationAction(ISD::FFLOOR, VT, Action);
    setOperationAction(ISD::FNEARBYINT, VT, Action);
    setOperationAction(ISD::FRINT, VT, Action);
    setOperationAction(ISD::BR_CC, VT, Action);
    setOperationAction(ISD::SETCC, VT, Action);
    setOperationAction(ISD::SELECT, VT, Custom);
    setOperationAction(ISD::SELECT_CC, VT, Action);
    setOperationAction(ISD::FROUND, VT, Action);
    setOperationAction(ISD::FROUNDEVEN, VT, Action);
    setOperationAction(ISD::FTRUNC, VT, Action);
    setOperationAction(ISD::FLDEXP, VT, Action);
    setOperationAction(ISD::FSINCOSPI, VT, Action);
  };
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasSSE2()) {
    // f16, f32 and f64 use SSE.
    // Set up the FP register classes.
    addRegisterClass(MVT::f16, Subtarget.hasAVX512() ? &X86::FR16XRegClass
                                                     : &X86::FR16RegClass);
    addRegisterClass(MVT::f32, Subtarget.hasAVX512() ? &X86::FR32XRegClass
                                                     : &X86::FR32RegClass);
    addRegisterClass(MVT::f64, Subtarget.hasAVX512() ? &X86::FR64XRegClass
                                                     : &X86::FR64RegClass);
 
    // Disable f32->f64 extload as we can only generate this in one instruction
    // under optsize. So its easier to pattern match (fpext (load)) for that
    // case instead of needing to emit 2 instructions for extload in the
    // non-optsize case.
    setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f32, Expand);
 
    for (auto VT : { MVT::f32, MVT::f64 }) {
      // Use ANDPD to simulate FABS.
      setOperationAction(ISD::FABS, VT, Custom);
 
      // Use XORP to simulate FNEG.
      setOperationAction(ISD::FNEG, VT, Custom);
 
      // Use ANDPD and ORPD to simulate FCOPYSIGN.
      setOperationAction(ISD::FCOPYSIGN, VT, Custom);
 
      // These might be better off as horizontal vector ops.
      setOperationAction(ISD::FADD, VT, Custom);
      setOperationAction(ISD::FSUB, VT, Custom);
 
      // We don't support sin/cos/fmod
      setOperationAction(ISD::FSIN   , VT, Expand);
      setOperationAction(ISD::FCOS   , VT, Expand);
      setOperationAction(ISD::FSINCOS, VT, Expand);
    }
 
    // Half type will be promoted by default.
    setF16Action(MVT::f16, Promote);
    setOperationAction(ISD::FADD, MVT::f16, Promote);
    setOperationAction(ISD::FSUB, MVT::f16, Promote);
    setOperationAction(ISD::FMUL, MVT::f16, Promote);
    setOperationAction(ISD::FDIV, MVT::f16, Promote);
    setOperationAction(ISD::FABS, MVT::f16, Custom);
    setOperationAction(ISD::FNEG, MVT::f16, Custom);
    setOperationAction(ISD::FCOPYSIGN, MVT::f16, Custom);
    setOperationAction(ISD::FP_ROUND, MVT::f16, Custom);
    setOperationAction(ISD::FP_EXTEND, MVT::f32, Custom);
    setOperationAction(ISD::FP_EXTEND, MVT::f64, Custom);
 
    setOperationAction(ISD::STRICT_FADD, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FSUB, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FMUL, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FDIV, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FMA, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FMINNUM, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FMAXNUM, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FMINIMUM, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FMAXIMUM, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FSQRT, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FPOW, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FLDEXP, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FLOG, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FLOG2, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FLOG10, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FEXP, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FEXP2, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FCEIL, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FFLOOR, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FNEARBYINT, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FRINT, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FSETCC, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FROUND, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FROUNDEVEN, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FTRUNC, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f16, Custom);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f32, Custom);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f64, Custom);
 
    setOperationAction(ISD::LLROUND, MVT::f16, Expand);
    setOperationAction(ISD::LROUND, MVT::f16, Expand);
    setOperationAction(ISD::LRINT, MVT::f16, Expand);
    setOperationAction(ISD::LLRINT, MVT::f16, Expand);
 
    setOperationAction(ISD::STRICT_LLROUND, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_LROUND, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_LRINT, MVT::f16, Promote);
    setOperationAction(ISD::STRICT_LLRINT, MVT::f16, Promote);
 
    // Lower this to MOVMSK plus an AND.
    setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
    setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
 
  } else if (!Subtarget.useSoftFloat() && Subtarget.hasSSE1() &&
             (UseX87 || Is64Bit)) {
    // Use SSE for f32, x87 for f64.
    // Set up the FP register classes.
    addRegisterClass(MVT::f32, &X86::FR32RegClass);
    if (UseX87)
      addRegisterClass(MVT::f64, &X86::RFP64RegClass);
 
    // Use ANDPS to simulate FABS.
    setOperationAction(ISD::FABS , MVT::f32, Custom);
 
    // Use XORP to simulate FNEG.
    setOperationAction(ISD::FNEG , MVT::f32, Custom);
 
    if (UseX87)
      setOperationAction(ISD::UNDEF, MVT::f64, Expand);
 
    // Use ANDPS and ORPS to simulate FCOPYSIGN.
    if (UseX87)
      setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
    setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
 
    // We don't support sin/cos/fmod
    setOperationAction(ISD::FSIN   , MVT::f32, Expand);
    setOperationAction(ISD::FCOS   , MVT::f32, Expand);
    setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
 
    if (UseX87) {
      // Always expand sin/cos functions even though x87 has an instruction.
      setOperationAction(ISD::FSIN, MVT::f64, Expand);
      setOperationAction(ISD::FCOS, MVT::f64, Expand);
      setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
    }
  } else if (UseX87) {
    // f32 and f64 in x87.
    // Set up the FP register classes.
    addRegisterClass(MVT::f64, &X86::RFP64RegClass);
    addRegisterClass(MVT::f32, &X86::RFP32RegClass);
 
    for (auto VT : { MVT::f32, MVT::f64 }) {
      setOperationAction(ISD::UNDEF,     VT, Expand);
      setOperationAction(ISD::FCOPYSIGN, VT, Expand);
 
      // Always expand sin/cos functions even though x87 has an instruction.
      setOperationAction(ISD::FSIN   , VT, Expand);
      setOperationAction(ISD::FCOS   , VT, Expand);
      setOperationAction(ISD::FSINCOS, VT, Expand);
    }
  }
 
  // Expand FP32 immediates into loads from the stack, save special cases.
  if (isTypeLegal(MVT::f32)) {
    if (UseX87 && (getRegClassFor(MVT::f32) == &X86::RFP32RegClass)) {
      addLegalFPImmediate(APFloat(+0.0f)); // FLD0
      addLegalFPImmediate(APFloat(+1.0f)); // FLD1
      addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
      addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
    } else // SSE immediates.
      addLegalFPImmediate(APFloat(+0.0f)); // xorps
  }
  // Expand FP64 immediates into loads from the stack, save special cases.
  if (isTypeLegal(MVT::f64)) {
    if (UseX87 && getRegClassFor(MVT::f64) == &X86::RFP64RegClass) {
      addLegalFPImmediate(APFloat(+0.0)); // FLD0
      addLegalFPImmediate(APFloat(+1.0)); // FLD1
      addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
      addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
    } else // SSE immediates.
      addLegalFPImmediate(APFloat(+0.0)); // xorpd
  }
  // Support fp16 0 immediate.
  if (isTypeLegal(MVT::f16))
    addLegalFPImmediate(APFloat::getZero(APFloat::IEEEhalf()));
 
  // Handle constrained floating-point operations of scalar.
  setOperationAction(ISD::STRICT_FADD,      MVT::f32, Legal);
  setOperationAction(ISD::STRICT_FADD,      MVT::f64, Legal);
  setOperationAction(ISD::STRICT_FSUB,      MVT::f32, Legal);
  setOperationAction(ISD::STRICT_FSUB,      MVT::f64, Legal);
  setOperationAction(ISD::STRICT_FMUL,      MVT::f32, Legal);
  setOperationAction(ISD::STRICT_FMUL,      MVT::f64, Legal);
  setOperationAction(ISD::STRICT_FDIV,      MVT::f32, Legal);
  setOperationAction(ISD::STRICT_FDIV,      MVT::f64, Legal);
  setOperationAction(ISD::STRICT_FP_ROUND,  MVT::f32, Legal);
  setOperationAction(ISD::STRICT_FP_ROUND,  MVT::f64, Legal);
  setOperationAction(ISD::STRICT_FSQRT,     MVT::f32, Legal);
  setOperationAction(ISD::STRICT_FSQRT,     MVT::f64, Legal);
 
  // We don't support FMA.
  setOperationAction(ISD::FMA, MVT::f64, Expand);
  setOperationAction(ISD::FMA, MVT::f32, Expand);
 
  // f80 always uses X87.
  if (UseX87) {
    addRegisterClass(MVT::f80, &X86::RFP80RegClass);
    setOperationAction(ISD::UNDEF,     MVT::f80, Expand);
    setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
    {
      APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended());
      addLegalFPImmediate(TmpFlt);  // FLD0
      TmpFlt.changeSign();
      addLegalFPImmediate(TmpFlt);  // FLD0/FCHS
 
      bool ignored;
      APFloat TmpFlt2(+1.0);
      TmpFlt2.convert(APFloat::x87DoubleExtended(), APFloat::rmNearestTiesToEven,
                      &ignored);
      addLegalFPImmediate(TmpFlt2);  // FLD1
      TmpFlt2.changeSign();
      addLegalFPImmediate(TmpFlt2);  // FLD1/FCHS
    }
 
    // Always expand sin/cos functions even though x87 has an instruction.
    // clang-format off
    setOperationAction(ISD::FSIN   , MVT::f80, Expand);
    setOperationAction(ISD::FCOS   , MVT::f80, Expand);
    setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
    setOperationAction(ISD::FTAN   , MVT::f80, Expand);
    setOperationAction(ISD::FASIN  , MVT::f80, Expand);
    setOperationAction(ISD::FACOS  , MVT::f80, Expand);
    setOperationAction(ISD::FATAN  , MVT::f80, Expand);
    setOperationAction(ISD::FATAN2 , MVT::f80, Expand);
    setOperationAction(ISD::FSINH  , MVT::f80, Expand);
    setOperationAction(ISD::FCOSH  , MVT::f80, Expand);
    setOperationAction(ISD::FTANH  , MVT::f80, Expand);
    // clang-format on
 
    setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
    setOperationAction(ISD::FCEIL,  MVT::f80, Expand);
    setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
    setOperationAction(ISD::FRINT,  MVT::f80, Expand);
    setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
    setOperationAction(ISD::FROUNDEVEN, MVT::f80, Expand);
    setOperationAction(ISD::FMA, MVT::f80, Expand);
    setOperationAction(ISD::LROUND, MVT::f80, LibCall);
    setOperationAction(ISD::LLROUND, MVT::f80, LibCall);
    setOperationAction(ISD::LRINT, MVT::f80, Custom);
    setOperationAction(ISD::LLRINT, MVT::f80, Custom);
 
    // Handle constrained floating-point operations of scalar.
    setOperationAction(ISD::STRICT_FADD     , MVT::f80, Legal);
    setOperationAction(ISD::STRICT_FSUB     , MVT::f80, Legal);
    setOperationAction(ISD::STRICT_FMUL     , MVT::f80, Legal);
    setOperationAction(ISD::STRICT_FDIV     , MVT::f80, Legal);
    setOperationAction(ISD::STRICT_FSQRT    , MVT::f80, Legal);
    setOperationAction(ISD::FCANONICALIZE, MVT::f80, Expand);
    if (isTypeLegal(MVT::f16)) {
      setOperationAction(ISD::FP_EXTEND, MVT::f80, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f80, Custom);
    } else {
      setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f80, Legal);
    }
    // FIXME: When the target is 64-bit, STRICT_FP_ROUND will be overwritten
    // as Custom.
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::f80, Legal);
  }
 
  // f128 uses xmm registers, but most operations require libcalls.
  if (!Subtarget.useSoftFloat() && Subtarget.is64Bit() && Subtarget.hasSSE1()) {
    addRegisterClass(MVT::f128, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                   : &X86::VR128RegClass);
 
    addLegalFPImmediate(APFloat::getZero(APFloat::IEEEquad())); // xorps
 
    setOperationAction(ISD::FADD,        MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FADD, MVT::f128, LibCall);
    setOperationAction(ISD::FSUB,        MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FSUB, MVT::f128, LibCall);
    setOperationAction(ISD::FDIV,        MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FDIV, MVT::f128, LibCall);
    setOperationAction(ISD::FMUL,        MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FMUL, MVT::f128, LibCall);
    setOperationAction(ISD::FMA,         MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FMA,  MVT::f128, LibCall);
 
    setOperationAction(ISD::FABS, MVT::f128, Custom);
    setOperationAction(ISD::FNEG, MVT::f128, Custom);
    setOperationAction(ISD::FCOPYSIGN, MVT::f128, Custom);
 
    // clang-format off
    setOperationAction(ISD::FSIN,         MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FSIN,  MVT::f128, LibCall);
    setOperationAction(ISD::FCOS,         MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FCOS,  MVT::f128, LibCall);
    setOperationAction(ISD::FSINCOS,      MVT::f128, LibCall);
    setOperationAction(ISD::FTAN,         MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FTAN,  MVT::f128, LibCall);
    // clang-format on
    // No STRICT_FSINCOS
    setOperationAction(ISD::FSQRT,        MVT::f128, LibCall);
    setOperationAction(ISD::STRICT_FSQRT, MVT::f128, LibCall);
 
    setOperationAction(ISD::FP_EXTEND,        MVT::f128, Custom);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::f128, Custom);
    // We need to custom handle any FP_ROUND with an f128 input, but
    // LegalizeDAG uses the result type to know when to run a custom handler.
    // So we have to list all legal floating point result types here.
    if (isTypeLegal(MVT::f32)) {
      setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
      setOperationAction(ISD::STRICT_FP_ROUND, MVT::f32, Custom);
    }
    if (isTypeLegal(MVT::f64)) {
      setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);
      setOperationAction(ISD::STRICT_FP_ROUND, MVT::f64, Custom);
    }
    if (isTypeLegal(MVT::f80)) {
      setOperationAction(ISD::FP_ROUND, MVT::f80, Custom);
      setOperationAction(ISD::STRICT_FP_ROUND, MVT::f80, Custom);
      setOperationAction(ISD::FCANONICALIZE, MVT::f80, Expand);
    }
 
    setOperationAction(ISD::SETCC, MVT::f128, Custom);
 
    setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f32, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f64, Expand);
    setLoadExtAction(ISD::EXTLOAD, MVT::f128, MVT::f80, Expand);
    setTruncStoreAction(MVT::f128, MVT::f32, Expand);
    setTruncStoreAction(MVT::f128, MVT::f64, Expand);
    setTruncStoreAction(MVT::f128, MVT::f80, Expand);
  }
 
  // Always use a library call for pow.
  setOperationAction(ISD::FPOW             , MVT::f32  , Expand);
  setOperationAction(ISD::FPOW             , MVT::f64  , Expand);
  setOperationAction(ISD::FPOW             , MVT::f80  , Expand);
  setOperationAction(ISD::FPOW             , MVT::f128 , Expand);
 
  setOperationAction(ISD::FLOG, MVT::f80, Expand);
  setOperationAction(ISD::FLOG2, MVT::f80, Expand);
  setOperationAction(ISD::FLOG10, MVT::f80, Expand);
  setOperationAction(ISD::FEXP, MVT::f80, Expand);
  setOperationAction(ISD::FEXP2, MVT::f80, Expand);
  setOperationAction(ISD::FEXP10, MVT::f80, Expand);
  setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
  setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
 
  // Some FP actions are always expanded for vector types.
  for (auto VT : { MVT::v8f16, MVT::v16f16, MVT::v32f16,
                   MVT::v4f32, MVT::v8f32,  MVT::v16f32,
                   MVT::v2f64, MVT::v4f64,  MVT::v8f64 }) {
    // clang-format off
    setOperationAction(ISD::FSIN,      VT, Expand);
    setOperationAction(ISD::FSINCOS,   VT, Expand);
    setOperationAction(ISD::FCOS,      VT, Expand);
    setOperationAction(ISD::FTAN,      VT, Expand);
    setOperationAction(ISD::FREM,      VT, Expand);
    setOperationAction(ISD::FCOPYSIGN, VT, Expand);
    setOperationAction(ISD::FPOW,      VT, Expand);
    setOperationAction(ISD::FLOG,      VT, Expand);
    setOperationAction(ISD::FLOG2,     VT, Expand);
    setOperationAction(ISD::FLOG10,    VT, Expand);
    setOperationAction(ISD::FEXP,      VT, Expand);
    setOperationAction(ISD::FEXP2,     VT, Expand);
    setOperationAction(ISD::FEXP10,    VT, Expand);
    // clang-format on
  }
 
  // First set operation action for all vector types to either promote
  // (for widening) or expand (for scalarization). Then we will selectively
  // turn on ones that can be effectively codegen'd.
  for (MVT VT : MVT::fixedlen_vector_valuetypes()) {
    setOperationAction(ISD::SDIV, VT, Expand);
    setOperationAction(ISD::UDIV, VT, Expand);
    setOperationAction(ISD::SREM, VT, Expand);
    setOperationAction(ISD::UREM, VT, Expand);
    setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
    setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
    setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
    setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
    setOperationAction(ISD::FMA,  VT, Expand);
    setOperationAction(ISD::FFLOOR, VT, Expand);
    setOperationAction(ISD::FCEIL, VT, Expand);
    setOperationAction(ISD::FTRUNC, VT, Expand);
    setOperationAction(ISD::FRINT, VT, Expand);
    setOperationAction(ISD::FNEARBYINT, VT, Expand);
    setOperationAction(ISD::FROUNDEVEN, VT, Expand);
    setOperationAction(ISD::SMUL_LOHI, VT, Expand);
    setOperationAction(ISD::MULHS, VT, Expand);
    setOperationAction(ISD::UMUL_LOHI, VT, Expand);
    setOperationAction(ISD::MULHU, VT, Expand);
    setOperationAction(ISD::SDIVREM, VT, Expand);
    setOperationAction(ISD::UDIVREM, VT, Expand);
    setOperationAction(ISD::CTPOP, VT, Expand);
    setOperationAction(ISD::CTTZ, VT, Expand);
    setOperationAction(ISD::CTLZ, VT, Expand);
    setOperationAction(ISD::ROTL, VT, Expand);
    setOperationAction(ISD::ROTR, VT, Expand);
    setOperationAction(ISD::BSWAP, VT, Expand);
    setOperationAction(ISD::SETCC, VT, Expand);
    setOperationAction(ISD::FP_TO_UINT, VT, Expand);
    setOperationAction(ISD::FP_TO_SINT, VT, Expand);
    setOperationAction(ISD::UINT_TO_FP, VT, Expand);
    setOperationAction(ISD::SINT_TO_FP, VT, Expand);
    setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
    setOperationAction(ISD::TRUNCATE, VT, Expand);
    setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
    setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
    setOperationAction(ISD::ANY_EXTEND, VT, Expand);
    setOperationAction(ISD::SELECT_CC, VT, Expand);
    for (MVT InnerVT : MVT::fixedlen_vector_valuetypes()) {
      setTruncStoreAction(InnerVT, VT, Expand);
 
      setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
      setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
 
      // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
      // types, we have to deal with them whether we ask for Expansion or not.
      // Setting Expand causes its own optimisation problems though, so leave
      // them legal.
      if (VT.getVectorElementType() == MVT::i1)
        setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
 
      // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
      // split/scalarized right now.
      if (VT.getVectorElementType() == MVT::f16 ||
          VT.getVectorElementType() == MVT::bf16)
        setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
    }
  }
 
  // FIXME: In order to prevent SSE instructions being expanded to MMX ones
  // with -msoft-float, disable use of MMX as well.
  if (!Subtarget.useSoftFloat() && Subtarget.hasMMX()) {
    addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
    // No operations on x86mmx supported, everything uses intrinsics.
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasSSE1()) {
    addRegisterClass(MVT::v4f32, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                    : &X86::VR128RegClass);
 
    setOperationAction(ISD::FMAXIMUM,           MVT::f32, Custom);
    setOperationAction(ISD::FMINIMUM,           MVT::f32, Custom);
    setOperationAction(ISD::FMAXIMUMNUM,        MVT::f32, Custom);
    setOperationAction(ISD::FMINIMUMNUM,        MVT::f32, Custom);
 
    setOperationAction(ISD::FNEG,               MVT::v4f32, Custom);
    setOperationAction(ISD::FABS,               MVT::v4f32, Custom);
    setOperationAction(ISD::FCOPYSIGN,          MVT::v4f32, Custom);
    setOperationAction(ISD::BUILD_VECTOR,       MVT::v4f32, Custom);
    setOperationAction(ISD::VECTOR_SHUFFLE,     MVT::v4f32, Custom);
    setOperationAction(ISD::VSELECT,            MVT::v4f32, Custom);
    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
    setOperationAction(ISD::SELECT,             MVT::v4f32, Custom);
    setOperationAction(ISD::FCANONICALIZE, MVT::v4f32, Expand);
 
    setOperationAction(ISD::LOAD,               MVT::v2f32, Custom);
    setOperationAction(ISD::STORE,              MVT::v2f32, Custom);
    setOperationAction(ISD::FCANONICALIZE, MVT::v2f32, Expand);
 
    setOperationAction(ISD::STRICT_FADD,        MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FSUB,        MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FMUL,        MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FDIV,        MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FSQRT,       MVT::v4f32, Legal);
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasSSE2()) {
    addRegisterClass(MVT::v2f64, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                    : &X86::VR128RegClass);
 
    // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
    // registers cannot be used even for integer operations.
    addRegisterClass(MVT::v16i8, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                    : &X86::VR128RegClass);
    addRegisterClass(MVT::v8i16, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                    : &X86::VR128RegClass);
    addRegisterClass(MVT::v8f16, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                    : &X86::VR128RegClass);
    addRegisterClass(MVT::v4i32, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                    : &X86::VR128RegClass);
    addRegisterClass(MVT::v2i64, Subtarget.hasVLX() ? &X86::VR128XRegClass
                                                    : &X86::VR128RegClass);
 
    for (auto VT : { MVT::f64, MVT::v4f32, MVT::v2f64 }) {
      setOperationAction(ISD::FMAXIMUM, VT, Custom);
      setOperationAction(ISD::FMINIMUM, VT, Custom);
      setOperationAction(ISD::FMAXIMUMNUM, VT, Custom);
      setOperationAction(ISD::FMINIMUMNUM, VT, Custom);
    }
 
    setOperationAction(ISD::MUL,                MVT::v2i8,  Custom);
    setOperationAction(ISD::MUL,                MVT::v4i8,  Custom);
    setOperationAction(ISD::MUL,                MVT::v8i8,  Custom);
 
    setOperationAction(ISD::MUL,                MVT::v16i8, Custom);
    setOperationAction(ISD::MUL,                MVT::v4i32, Custom);
    setOperationAction(ISD::MUL,                MVT::v2i64, Custom);
    setOperationAction(ISD::MULHU,              MVT::v4i32, Custom);
    setOperationAction(ISD::MULHS,              MVT::v4i32, Custom);
    setOperationAction(ISD::MULHU,              MVT::v16i8, Custom);
    setOperationAction(ISD::MULHS,              MVT::v16i8, Custom);
    setOperationAction(ISD::MULHU,              MVT::v8i16, Legal);
    setOperationAction(ISD::MULHS,              MVT::v8i16, Legal);
    setOperationAction(ISD::MUL,                MVT::v8i16, Legal);
    setOperationAction(ISD::AVGCEILU,           MVT::v16i8, Legal);
    setOperationAction(ISD::AVGCEILU,           MVT::v8i16, Legal);
 
    setOperationAction(ISD::SMULO,              MVT::v16i8, Custom);
    setOperationAction(ISD::UMULO,              MVT::v16i8, Custom);
    setOperationAction(ISD::UMULO,              MVT::v2i32, Custom);
 
    setOperationAction(ISD::FNEG,               MVT::v2f64, Custom);
    setOperationAction(ISD::FCANONICALIZE, MVT::v2f64, Expand);
    setOperationAction(ISD::FABS,               MVT::v2f64, Custom);
    setOperationAction(ISD::FCOPYSIGN,          MVT::v2f64, Custom);
 
    setOperationAction(ISD::LRINT, MVT::v4f32, Custom);
    setOperationAction(ISD::LRINT, MVT::v2i32, Custom);
 
    setOperationAction(ISD::AND, MVT::i128, Custom);
    setOperationAction(ISD::OR, MVT::i128, Custom);
    setOperationAction(ISD::XOR, MVT::i128, Custom);
    setOperationAction(ISD::SELECT, MVT::i128, Custom);
 
    if (Subtarget.hasPCLMUL()) {
      for (auto VT : {MVT::i64, MVT::v4i32, MVT::v2i64}) {
        setOperationAction(ISD::CLMUL, VT, Custom);
        setOperationAction(ISD::CLMULH, VT, Custom);
      }
      setOperationAction(ISD::CLMUL, MVT::i32, Custom);
      setOperationAction(ISD::CLMUL, MVT::i16, Custom);
      setOperationAction(ISD::CLMUL, MVT::i8, Custom);
    }
 
    for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) {
      setOperationAction(ISD::SMAX, VT, VT == MVT::v8i16 ? Legal : Custom);
      setOperationAction(ISD::SMIN, VT, VT == MVT::v8i16 ? Legal : Custom);
      setOperationAction(ISD::UMAX, VT, VT == MVT::v16i8 ? Legal : Custom);
      setOperationAction(ISD::UMIN, VT, VT == MVT::v16i8 ? Legal : Custom);
      setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
    }
 
    // SSE2 can use basic vector unrolling.
    // SSE41 can use PHMINPOS to perform v16i8/v8i16 minmax reductions.
    // Fallback to ReplaceNodeResults for vXi64 reductions on 32-bit targets.
    for (auto VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::i64}) {
      setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom);
    }
 
    setOperationAction(ISD::UADDSAT,            MVT::v16i8, Legal);
    setOperationAction(ISD::SADDSAT,            MVT::v16i8, Legal);
    setOperationAction(ISD::USUBSAT,            MVT::v16i8, Legal);
    setOperationAction(ISD::SSUBSAT,            MVT::v16i8, Legal);
    setOperationAction(ISD::UADDSAT,            MVT::v8i16, Legal);
    setOperationAction(ISD::SADDSAT,            MVT::v8i16, Legal);
    setOperationAction(ISD::USUBSAT,            MVT::v8i16, Legal);
    setOperationAction(ISD::SSUBSAT,            MVT::v8i16, Legal);
    setOperationAction(ISD::USUBSAT,            MVT::v4i32, Custom);
    setOperationAction(ISD::USUBSAT,            MVT::v2i64, Custom);
 
    setOperationAction(ISD::INSERT_VECTOR_ELT,  MVT::v16i8, Custom);
    setOperationAction(ISD::INSERT_VECTOR_ELT,  MVT::v8i16, Custom);
    setOperationAction(ISD::INSERT_VECTOR_ELT,  MVT::v4i32, Custom);
    setOperationAction(ISD::INSERT_VECTOR_ELT,  MVT::v4f32, Custom);
 
    for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) {
      setOperationAction(ISD::SETCC, VT, Custom);
      setOperationAction(ISD::CTPOP, VT, Custom);
      setOperationAction(ISD::ABS, VT, Custom);
      setOperationAction(ISD::ABDS, VT, Custom);
      setOperationAction(ISD::ABDU, VT, Custom);
 
      // The condition codes aren't legal in SSE/AVX and under AVX512 we use
      // setcc all the way to isel and prefer SETGT in some isel patterns.
      setCondCodeAction(ISD::SETLT, VT, Custom);
      setCondCodeAction(ISD::SETLE, VT, Custom);
    }
 
    setOperationAction(ISD::SETCC,          MVT::v2f64, Custom);
    setOperationAction(ISD::SETCC,          MVT::v4f32, Custom);
    setOperationAction(ISD::STRICT_FSETCC,  MVT::v2f64, Custom);
    setOperationAction(ISD::STRICT_FSETCC,  MVT::v4f32, Custom);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::v2f64, Custom);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::v4f32, Custom);
 
    for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
      setOperationAction(ISD::SCALAR_TO_VECTOR,   VT, Custom);
      setOperationAction(ISD::BUILD_VECTOR,       VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE,     VT, Custom);
      setOperationAction(ISD::VSELECT,            VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
    }
 
    for (auto VT : { MVT::v8f16, MVT::v2f64, MVT::v2i64 }) {
      setOperationAction(ISD::BUILD_VECTOR,       VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE,     VT, Custom);
      setOperationAction(ISD::VSELECT,            VT, Custom);
 
      if (VT == MVT::v2i64 && !Subtarget.is64Bit())
        continue;
 
      setOperationAction(ISD::INSERT_VECTOR_ELT,  VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
    }
    setF16Action(MVT::v8f16, Expand);
    setOperationAction(ISD::FADD, MVT::v8f16, Expand);
    setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
    setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
    setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
    setOperationAction(ISD::FNEG, MVT::v8f16, Custom);
    setOperationAction(ISD::FABS, MVT::v8f16, Custom);
    setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Custom);
 
    // Custom lower v2i64 and v2f64 selects.
    setOperationAction(ISD::SELECT,             MVT::v2f64, Custom);
    setOperationAction(ISD::SELECT,             MVT::v2i64, Custom);
    setOperationAction(ISD::SELECT,             MVT::v4i32, Custom);
    setOperationAction(ISD::SELECT,             MVT::v8i16, Custom);
    setOperationAction(ISD::SELECT,             MVT::v8f16, Custom);
    setOperationAction(ISD::SELECT,             MVT::v16i8, Custom);
 
    setOperationAction(ISD::FP_TO_SINT,         MVT::v4i32, Custom);
    setOperationAction(ISD::FP_TO_UINT,         MVT::v4i32, Custom);
    setOperationAction(ISD::FP_TO_SINT,         MVT::v2i32, Custom);
    setOperationAction(ISD::FP_TO_UINT,         MVT::v2i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT,  MVT::v4i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT,  MVT::v2i32, Custom);
 
    // Custom legalize these to avoid over promotion or custom promotion.
    for (auto VT : {MVT::v2i8, MVT::v4i8, MVT::v8i8, MVT::v2i16, MVT::v4i16}) {
      setOperationAction(ISD::FP_TO_SINT,        VT, Custom);
      setOperationAction(ISD::FP_TO_UINT,        VT, Custom);
      setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Custom);
      setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Custom);
    }
 
    setOperationAction(ISD::SINT_TO_FP,         MVT::v4i32, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP,  MVT::v4i32, Custom);
    setOperationAction(ISD::SINT_TO_FP,         MVT::v2i32, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP,  MVT::v2i32, Custom);
 
    setOperationAction(ISD::UINT_TO_FP,         MVT::v2i32, Custom);
    setOperationAction(ISD::STRICT_UINT_TO_FP,  MVT::v2i32, Custom);
 
    setOperationAction(ISD::UINT_TO_FP,         MVT::v4i32, Custom);
    setOperationAction(ISD::STRICT_UINT_TO_FP,  MVT::v4i32, Custom);
 
    // Fast v2f32 UINT_TO_FP( v2i32 ) custom conversion.
    setOperationAction(ISD::SINT_TO_FP,         MVT::v2f32, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP,  MVT::v2f32, Custom);
    setOperationAction(ISD::UINT_TO_FP,         MVT::v2f32, Custom);
    setOperationAction(ISD::STRICT_UINT_TO_FP,  MVT::v2f32, Custom);
 
    setOperationAction(ISD::FP_EXTEND,          MVT::v2f32, Custom);
    setOperationAction(ISD::STRICT_FP_EXTEND,   MVT::v2f32, Custom);
    setOperationAction(ISD::FP_ROUND,           MVT::v2f32, Custom);
    setOperationAction(ISD::STRICT_FP_ROUND,    MVT::v2f32, Custom);
 
    // We want to legalize this to an f64 load rather than an i64 load on
    // 64-bit targets and two 32-bit loads on a 32-bit target. Similar for
    // store.
    setOperationAction(ISD::LOAD,               MVT::v2i32, Custom);
    setOperationAction(ISD::LOAD,               MVT::v4i16, Custom);
    setOperationAction(ISD::LOAD,               MVT::v8i8,  Custom);
    setOperationAction(ISD::STORE,              MVT::v2i32, Custom);
    setOperationAction(ISD::STORE,              MVT::v4i16, Custom);
    setOperationAction(ISD::STORE,              MVT::v8i8,  Custom);
 
    // Add 32-bit vector stores to help vectorization opportunities.
    setOperationAction(ISD::STORE,              MVT::v2i16, Custom);
    setOperationAction(ISD::STORE,              MVT::v4i8,  Custom);
 
    setOperationAction(ISD::BITCAST,            MVT::v2i32, Custom);
    setOperationAction(ISD::BITCAST,            MVT::v4i16, Custom);
    setOperationAction(ISD::BITCAST,            MVT::v8i8,  Custom);
    if (!Subtarget.hasAVX512())
      setOperationAction(ISD::BITCAST, MVT::v16i1, Custom);
 
    setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
    setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
    setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
 
    setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
 
    setOperationAction(ISD::TRUNCATE,    MVT::v2i8,  Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v2i16, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v2i32, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v2i64, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v4i8,  Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v4i16, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v4i32, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v4i64, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v8i8,  Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v8i16, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v8i32, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v8i64, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v16i8, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v16i16, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v16i32, Custom);
    setOperationAction(ISD::TRUNCATE,    MVT::v16i64, Custom);
 
    // In the customized shift lowering, the legal v4i32/v2i64 cases
    // in AVX2 will be recognized.
    for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) {
      setOperationAction(ISD::SRL,              VT, Custom);
      setOperationAction(ISD::SHL,              VT, Custom);
      setOperationAction(ISD::SRA,              VT, Custom);
      if (VT == MVT::v2i64) continue;
      setOperationAction(ISD::ROTL,             VT, Custom);
      setOperationAction(ISD::ROTR,             VT, Custom);
      setOperationAction(ISD::FSHL,             VT, Custom);
      setOperationAction(ISD::FSHR,             VT, Custom);
    }
 
    setOperationAction(ISD::STRICT_FSQRT,       MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FADD,        MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FSUB,        MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FMUL,        MVT::v2f64, Legal);
    setOperationAction(ISD::STRICT_FDIV,        MVT::v2f64, Legal);
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasGFNI()) {
    setOperationAction(ISD::BITREVERSE, MVT::i8, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::i16, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::i32, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::i64, Custom);
 
    for (auto VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) {
      setOperationAction(ISD::BITREVERSE, VT, Custom);
    }
 
    setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
    setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasSSSE3()) {
    setOperationAction(ISD::ABS,                MVT::v16i8, Legal);
    setOperationAction(ISD::ABS,                MVT::v8i16, Legal);
    setOperationAction(ISD::ABS,                MVT::v4i32, Legal);
 
    for (auto VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64}) {
      setOperationAction(ISD::BITREVERSE,       VT, Custom);
      setOperationAction(ISD::CTLZ,             VT, Custom);
    }
    setOperationAction(ISD::BITREVERSE,        MVT::i128, Custom);
 
    // These might be better off as horizontal vector ops.
    setOperationAction(ISD::ADD,                MVT::i16, Custom);
    setOperationAction(ISD::ADD,                MVT::i32, Custom);
    setOperationAction(ISD::SUB,                MVT::i16, Custom);
    setOperationAction(ISD::SUB,                MVT::i32, Custom);
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasSSE41()) {
    for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
      setOperationAction(ISD::FFLOOR,            RoundedTy,  Legal);
      setOperationAction(ISD::STRICT_FFLOOR,     RoundedTy,  Legal);
      setOperationAction(ISD::FCEIL,             RoundedTy,  Legal);
      setOperationAction(ISD::STRICT_FCEIL,      RoundedTy,  Legal);
      setOperationAction(ISD::FTRUNC,            RoundedTy,  Legal);
      setOperationAction(ISD::STRICT_FTRUNC,     RoundedTy,  Legal);
      setOperationAction(ISD::FRINT,             RoundedTy,  Legal);
      setOperationAction(ISD::STRICT_FRINT,      RoundedTy,  Legal);
      setOperationAction(ISD::FNEARBYINT,        RoundedTy,  Legal);
      setOperationAction(ISD::STRICT_FNEARBYINT, RoundedTy,  Legal);
      setOperationAction(ISD::FROUNDEVEN,        RoundedTy,  Legal);
      setOperationAction(ISD::STRICT_FROUNDEVEN, RoundedTy,  Legal);
 
      setOperationAction(ISD::FROUND,            RoundedTy,  Custom);
    }
 
    setOperationAction(ISD::SMAX,               MVT::v16i8, Legal);
    setOperationAction(ISD::SMAX,               MVT::v4i32, Legal);
    setOperationAction(ISD::UMAX,               MVT::v8i16, Legal);
    setOperationAction(ISD::UMAX,               MVT::v4i32, Legal);
    setOperationAction(ISD::SMIN,               MVT::v16i8, Legal);
    setOperationAction(ISD::SMIN,               MVT::v4i32, Legal);
    setOperationAction(ISD::UMIN,               MVT::v8i16, Legal);
    setOperationAction(ISD::UMIN,               MVT::v4i32, Legal);
 
    setOperationAction(ISD::UADDSAT,            MVT::v4i32, Custom);
    setOperationAction(ISD::SADDSAT,            MVT::v2i64, Custom);
    setOperationAction(ISD::SSUBSAT,            MVT::v2i64, Custom);
 
    // FIXME: Do we need to handle scalar-to-vector here?
    setOperationAction(ISD::MUL,                MVT::v4i32, Legal);
    setOperationAction(ISD::SMULO,              MVT::v2i32, Custom);
 
    // We directly match byte blends in the backend as they match the VSELECT
    // condition form.
    setOperationAction(ISD::VSELECT,            MVT::v16i8, Legal);
 
    // SSE41 brings specific instructions for doing vector sign extend even in
    // cases where we don't have SRA.
    for (auto VT : { MVT::v8i16, MVT::v4i32, MVT::v2i64 }) {
      setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Legal);
      setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Legal);
    }
 
    // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
    for (auto LoadExtOp : { ISD::SEXTLOAD, ISD::ZEXTLOAD }) {
      setLoadExtAction(LoadExtOp, MVT::v8i16, MVT::v8i8,  Legal);
      setLoadExtAction(LoadExtOp, MVT::v4i32, MVT::v4i8,  Legal);
      setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i8,  Legal);
      setLoadExtAction(LoadExtOp, MVT::v4i32, MVT::v4i16, Legal);
      setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i16, Legal);
      setLoadExtAction(LoadExtOp, MVT::v2i64, MVT::v2i32, Legal);
    }
 
    if (Subtarget.is64Bit() && !Subtarget.hasAVX512()) {
      // We need to scalarize v4i64->v432 uint_to_fp using cvtsi2ss, but we can
      // do the pre and post work in the vector domain.
      setOperationAction(ISD::UINT_TO_FP,        MVT::v4i64, Custom);
      setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4i64, Custom);
      // We need to mark SINT_TO_FP as Custom even though we want to expand it
      // so that DAG combine doesn't try to turn it into uint_to_fp.
      setOperationAction(ISD::SINT_TO_FP,        MVT::v4i64, Custom);
      setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4i64, Custom);
    }
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasSSE42()) {
    setOperationAction(ISD::UADDSAT,            MVT::v2i64, Custom);
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasXOP()) {
    for (MVT VT : { MVT::v16i8, MVT::v8i16,  MVT::v4i32, MVT::v2i64,
                    MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) {
      setOperationAction(ISD::ROTL, VT, VT.is128BitVector() ? Legal : Custom);
      setOperationAction(ISD::ROTR, VT, Custom);
    }
 
    // XOP can efficiently perform BITREVERSE with VPPERM.
    for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 })
      setOperationAction(ISD::BITREVERSE, VT, Custom);
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasAVX()) {
    bool HasInt256 = Subtarget.hasInt256();
 
    addRegisterClass(MVT::v32i8,  Subtarget.hasVLX() ? &X86::VR256XRegClass
                                                     : &X86::VR256RegClass);
    addRegisterClass(MVT::v16i16, Subtarget.hasVLX() ? &X86::VR256XRegClass
                                                     : &X86::VR256RegClass);
    addRegisterClass(MVT::v16f16, Subtarget.hasVLX() ? &X86::VR256XRegClass
                                                     : &X86::VR256RegClass);
    addRegisterClass(MVT::v8i32,  Subtarget.hasVLX() ? &X86::VR256XRegClass
                                                     : &X86::VR256RegClass);
    addRegisterClass(MVT::v8f32,  Subtarget.hasVLX() ? &X86::VR256XRegClass
                                                     : &X86::VR256RegClass);
    addRegisterClass(MVT::v4i64,  Subtarget.hasVLX() ? &X86::VR256XRegClass
                                                     : &X86::VR256RegClass);
    addRegisterClass(MVT::v4f64,  Subtarget.hasVLX() ? &X86::VR256XRegClass
                                                     : &X86::VR256RegClass);
 
    for (auto VT : { MVT::v8f32, MVT::v4f64 }) {
      setOperationAction(ISD::FFLOOR,            VT, Legal);
      setOperationAction(ISD::STRICT_FFLOOR,     VT, Legal);
      setOperationAction(ISD::FCEIL,             VT, Legal);
      setOperationAction(ISD::STRICT_FCEIL,      VT, Legal);
      setOperationAction(ISD::FTRUNC,            VT, Legal);
      setOperationAction(ISD::STRICT_FTRUNC,     VT, Legal);
      setOperationAction(ISD::FRINT,             VT, Legal);
      setOperationAction(ISD::STRICT_FRINT,      VT, Legal);
      setOperationAction(ISD::FNEARBYINT,        VT, Legal);
      setOperationAction(ISD::STRICT_FNEARBYINT, VT, Legal);
      setOperationAction(ISD::FROUNDEVEN,        VT, Legal);
      setOperationAction(ISD::STRICT_FROUNDEVEN, VT, Legal);
 
      setOperationAction(ISD::FROUND,            VT, Custom);
 
      setOperationAction(ISD::FNEG,              VT, Custom);
      setOperationAction(ISD::FABS,              VT, Custom);
      setOperationAction(ISD::FCOPYSIGN,         VT, Custom);
 
      setOperationAction(ISD::FMAXIMUM,          VT, Custom);
      setOperationAction(ISD::FMINIMUM,          VT, Custom);
      setOperationAction(ISD::FMAXIMUMNUM,       VT, Custom);
      setOperationAction(ISD::FMINIMUMNUM,       VT, Custom);
      setOperationAction(ISD::FCANONICALIZE, VT, Expand);
    }
 
    setOperationAction(ISD::LRINT, MVT::v8f32, Custom);
    setOperationAction(ISD::LRINT, MVT::v4f64, Custom);
 
    setOperationAction(ISD::AND, MVT::i256, Custom);
    setOperationAction(ISD::OR, MVT::i256, Custom);
    setOperationAction(ISD::XOR, MVT::i256, Custom);
    setOperationAction(ISD::SELECT, MVT::i256, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::i256, Custom);
 
    // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
    // even though v8i16 is a legal type.
    setOperationPromotedToType(ISD::FP_TO_SINT,        MVT::v8i16, MVT::v8i32);
    setOperationPromotedToType(ISD::FP_TO_UINT,        MVT::v8i16, MVT::v8i32);
    setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v8i16, MVT::v8i32);
    setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v8i16, MVT::v8i32);
    setOperationAction(ISD::FP_TO_SINT,                MVT::v8i32, Custom);
    setOperationAction(ISD::FP_TO_UINT,                MVT::v8i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT,         MVT::v8i32, Custom);
 
    setOperationAction(ISD::SINT_TO_FP,         MVT::v8i32, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP,  MVT::v8i32, Custom);
    setOperationAction(ISD::FP_EXTEND,          MVT::v8f32, Expand);
    setOperationAction(ISD::FP_ROUND,           MVT::v8f16, Expand);
    setOperationAction(ISD::FP_EXTEND,          MVT::v4f64, Custom);
    setOperationAction(ISD::STRICT_FP_EXTEND,   MVT::v4f64, Custom);
 
    setOperationAction(ISD::STRICT_FP_ROUND,    MVT::v4f32, Legal);
    setOperationAction(ISD::STRICT_FADD,        MVT::v8f32, Legal);
    setOperationAction(ISD::STRICT_FADD,        MVT::v4f64, Legal);
    setOperationAction(ISD::STRICT_FSUB,        MVT::v8f32, Legal);
    setOperationAction(ISD::STRICT_FSUB,        MVT::v4f64, Legal);
    setOperationAction(ISD::STRICT_FMUL,        MVT::v8f32, Legal);
    setOperationAction(ISD::STRICT_FMUL,        MVT::v4f64, Legal);
    setOperationAction(ISD::STRICT_FDIV,        MVT::v8f32, Legal);
    setOperationAction(ISD::STRICT_FDIV,        MVT::v4f64, Legal);
    setOperationAction(ISD::STRICT_FSQRT,       MVT::v8f32, Legal);
    setOperationAction(ISD::STRICT_FSQRT,       MVT::v4f64, Legal);
 
    if (!Subtarget.hasAVX512())
      setOperationAction(ISD::BITCAST, MVT::v32i1, Custom);
 
    // In the customized shift lowering, the legal v8i32/v4i64 cases
    // in AVX2 will be recognized.
    for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) {
      setOperationAction(ISD::SRL,             VT, Custom);
      setOperationAction(ISD::SHL,             VT, Custom);
      setOperationAction(ISD::SRA,             VT, Custom);
      setOperationAction(ISD::ABDS,            VT, Custom);
      setOperationAction(ISD::ABDU,            VT, Custom);
      setOperationAction(ISD::VECREDUCE_AND,   VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR,    VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR,   VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMAX,  VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN,  VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX,  VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN,  VT, Custom);
      if (VT == MVT::v4i64) continue;
      setOperationAction(ISD::ROTL,            VT, Custom);
      setOperationAction(ISD::ROTR,            VT, Custom);
      setOperationAction(ISD::FSHL,            VT, Custom);
      setOperationAction(ISD::FSHR,            VT, Custom);
    }
 
    // These types need custom splitting if their input is a 128-bit vector.
    setOperationAction(ISD::SIGN_EXTEND,       MVT::v8i64,  Custom);
    setOperationAction(ISD::SIGN_EXTEND,       MVT::v16i32, Custom);
    setOperationAction(ISD::ZERO_EXTEND,       MVT::v8i64,  Custom);
    setOperationAction(ISD::ZERO_EXTEND,       MVT::v16i32, Custom);
 
    setOperationAction(ISD::SELECT,            MVT::v4f64, Custom);
    setOperationAction(ISD::SELECT,            MVT::v4i64, Custom);
    setOperationAction(ISD::SELECT,            MVT::v8i32, Custom);
    setOperationAction(ISD::SELECT,            MVT::v16i16, Custom);
    setOperationAction(ISD::SELECT,            MVT::v16f16, Custom);
    setOperationAction(ISD::SELECT,            MVT::v32i8, Custom);
    setOperationAction(ISD::SELECT,            MVT::v8f32, Custom);
 
    for (auto VT : { MVT::v16i16, MVT::v8i32, MVT::v4i64 }) {
      setOperationAction(ISD::SIGN_EXTEND,     VT, Custom);
      setOperationAction(ISD::ZERO_EXTEND,     VT, Custom);
      setOperationAction(ISD::ANY_EXTEND,      VT, Custom);
    }
 
    setOperationAction(ISD::TRUNCATE,          MVT::v32i8, Custom);
    setOperationAction(ISD::TRUNCATE,          MVT::v32i16, Custom);
    setOperationAction(ISD::TRUNCATE,          MVT::v32i32, Custom);
    setOperationAction(ISD::TRUNCATE,          MVT::v32i64, Custom);
 
    for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) {
      setOperationAction(ISD::SETCC,           VT, Custom);
      setOperationAction(ISD::CTPOP,           VT, Custom);
      setOperationAction(ISD::CTLZ,            VT, Custom);
      setOperationAction(ISD::BITREVERSE,      VT, Custom);
 
      // The condition codes aren't legal in SSE/AVX and under AVX512 we use
      // setcc all the way to isel and prefer SETGT in some isel patterns.
      setCondCodeAction(ISD::SETLT, VT, Custom);
      setCondCodeAction(ISD::SETLE, VT, Custom);
    }
 
    setOperationAction(ISD::SETCC,          MVT::v4f64, Custom);
    setOperationAction(ISD::SETCC,          MVT::v8f32, Custom);
    setOperationAction(ISD::STRICT_FSETCC,  MVT::v4f64, Custom);
    setOperationAction(ISD::STRICT_FSETCC,  MVT::v8f32, Custom);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::v4f64, Custom);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::v8f32, Custom);
 
    if (Subtarget.hasAnyFMA()) {
      for (auto VT : { MVT::f32, MVT::f64, MVT::v4f32, MVT::v8f32,
                       MVT::v2f64, MVT::v4f64 }) {
        setOperationAction(ISD::FMA, VT, Legal);
        setOperationAction(ISD::STRICT_FMA, VT, Legal);
      }
    }
 
    for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 }) {
      setOperationAction(ISD::ADD, VT, HasInt256 ? Legal : Custom);
      setOperationAction(ISD::SUB, VT, HasInt256 ? Legal : Custom);
    }
 
    setOperationAction(ISD::MUL,       MVT::v4i64,  Custom);
    setOperationAction(ISD::MUL,       MVT::v8i32,  HasInt256 ? Legal : Custom);
    setOperationAction(ISD::MUL,       MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::MUL,       MVT::v32i8,  Custom);
 
    setOperationAction(ISD::MULHU,     MVT::v8i32,  Custom);
    setOperationAction(ISD::MULHS,     MVT::v8i32,  Custom);
    setOperationAction(ISD::MULHU,     MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::MULHS,     MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::MULHU,     MVT::v32i8,  Custom);
    setOperationAction(ISD::MULHS,     MVT::v32i8,  Custom);
    setOperationAction(ISD::AVGCEILU,  MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::AVGCEILU,  MVT::v32i8,  HasInt256 ? Legal : Custom);
 
    setOperationAction(ISD::SMULO,     MVT::v32i8, Custom);
    setOperationAction(ISD::UMULO,     MVT::v32i8, Custom);
 
    setOperationAction(ISD::ABS,       MVT::v4i64,  Custom);
    setOperationAction(ISD::SMAX,      MVT::v4i64,  Custom);
    setOperationAction(ISD::UMAX,      MVT::v4i64,  Custom);
    setOperationAction(ISD::SMIN,      MVT::v4i64,  Custom);
    setOperationAction(ISD::UMIN,      MVT::v4i64,  Custom);
 
    setOperationAction(ISD::UADDSAT,   MVT::v32i8,  HasInt256 ? Legal : Custom);
    setOperationAction(ISD::SADDSAT,   MVT::v32i8,  HasInt256 ? Legal : Custom);
    setOperationAction(ISD::USUBSAT,   MVT::v32i8,  HasInt256 ? Legal : Custom);
    setOperationAction(ISD::SSUBSAT,   MVT::v32i8,  HasInt256 ? Legal : Custom);
    setOperationAction(ISD::UADDSAT,   MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::SADDSAT,   MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::USUBSAT,   MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::SSUBSAT,   MVT::v16i16, HasInt256 ? Legal : Custom);
    setOperationAction(ISD::UADDSAT,   MVT::v8i32, Custom);
    setOperationAction(ISD::USUBSAT,   MVT::v8i32, Custom);
    setOperationAction(ISD::UADDSAT,   MVT::v4i64, Custom);
    setOperationAction(ISD::USUBSAT,   MVT::v4i64, Custom);
 
    for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
      setOperationAction(ISD::ABS,  VT, HasInt256 ? Legal : Custom);
      setOperationAction(ISD::SMAX, VT, HasInt256 ? Legal : Custom);
      setOperationAction(ISD::UMAX, VT, HasInt256 ? Legal : Custom);
      setOperationAction(ISD::SMIN, VT, HasInt256 ? Legal : Custom);
      setOperationAction(ISD::UMIN, VT, HasInt256 ? Legal : Custom);
    }
 
    for (auto VT : {MVT::v16i16, MVT::v8i32, MVT::v4i64}) {
      setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom);
      setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom);
    }
 
    if (HasInt256) {
      // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
      // when we have a 256bit-wide blend with immediate.
      setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
      setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v8i32, Custom);
 
      // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
      for (auto LoadExtOp : { ISD::SEXTLOAD, ISD::ZEXTLOAD }) {
        setLoadExtAction(LoadExtOp, MVT::v16i16, MVT::v16i8, Legal);
        setLoadExtAction(LoadExtOp, MVT::v8i32,  MVT::v8i8,  Legal);
        setLoadExtAction(LoadExtOp, MVT::v4i64,  MVT::v4i8,  Legal);
        setLoadExtAction(LoadExtOp, MVT::v8i32,  MVT::v8i16, Legal);
        setLoadExtAction(LoadExtOp, MVT::v4i64,  MVT::v4i16, Legal);
        setLoadExtAction(LoadExtOp, MVT::v4i64,  MVT::v4i32, Legal);
      }
    }
 
    for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64,
                     MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 }) {
      setOperationAction(ISD::MLOAD,  VT, Subtarget.hasVLX() ? Legal : Custom);
      setOperationAction(ISD::MSTORE, VT, Legal);
    }
 
    // Extract subvector is special because the value type
    // (result) is 128-bit but the source is 256-bit wide.
    for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64,
                     MVT::v8f16, MVT::v4f32, MVT::v2f64 }) {
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
    }
 
    // Custom lower several nodes for 256-bit types.
    for (MVT VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64,
                    MVT::v16f16, MVT::v8f32, MVT::v4f64 }) {
      setOperationAction(ISD::BUILD_VECTOR,       VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE,     VT, Custom);
      setOperationAction(ISD::VSELECT,            VT, Custom);
      setOperationAction(ISD::INSERT_VECTOR_ELT,  VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::SCALAR_TO_VECTOR,   VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR,   VT, Legal);
      setOperationAction(ISD::CONCAT_VECTORS,     VT, Custom);
      setOperationAction(ISD::STORE,              VT, Custom);
    }
    setF16Action(MVT::v16f16, Expand);
    setOperationAction(ISD::FNEG, MVT::v16f16, Custom);
    setOperationAction(ISD::FABS, MVT::v16f16, Custom);
    setOperationAction(ISD::FCOPYSIGN, MVT::v16f16, Custom);
    setOperationAction(ISD::FADD, MVT::v16f16, Expand);
    setOperationAction(ISD::FSUB, MVT::v16f16, Expand);
    setOperationAction(ISD::FMUL, MVT::v16f16, Expand);
    setOperationAction(ISD::FDIV, MVT::v16f16, Expand);
 
    // Only PCLMUL required as we always unroll clmul vectors.
    if (Subtarget.hasPCLMUL()) {
      for (auto VT : {MVT::v8i32, MVT::v4i64}) {
        setOperationAction(ISD::CLMUL, VT, Custom);
        setOperationAction(ISD::CLMULH, VT, Custom);
      }
    }
 
    if (HasInt256) {
      setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
 
      // Custom legalize 2x32 to get a little better code.
      setOperationAction(ISD::MGATHER, MVT::v2f32, Custom);
      setOperationAction(ISD::MGATHER, MVT::v2i32, Custom);
 
      for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64,
                       MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 })
        setOperationAction(ISD::MGATHER,  VT, Custom);
    }
 
    if (Subtarget.hasGFNI()) {
      setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
      setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
    }
  }
 
  if (!Subtarget.useSoftFloat() && !Subtarget.hasFP16() &&
      Subtarget.hasF16C()) {
    for (MVT VT : { MVT::f16, MVT::v2f16, MVT::v4f16, MVT::v8f16 }) {
      setOperationAction(ISD::FP_ROUND,           VT, Custom);
      setOperationAction(ISD::STRICT_FP_ROUND,    VT, Custom);
    }
    for (MVT VT : { MVT::f32, MVT::v2f32, MVT::v4f32, MVT::v8f32 }) {
      setOperationAction(ISD::FP_EXTEND,          VT, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND,   VT, Custom);
    }
    for (unsigned Opc : {ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FDIV}) {
      setOperationPromotedToType(Opc, MVT::v8f16, MVT::v8f32);
      setOperationPromotedToType(Opc, MVT::v16f16, MVT::v16f32);
    }
    setOperationAction(ISD::SETCC, MVT::v8f16, Custom);
    setOperationAction(ISD::SETCC, MVT::v16f16, Custom);
  }
 
  // This block controls legalization of the mask vector sizes that are
  // available with AVX512. 512-bit vectors are in a separate block controlled
  // by useAVX512Regs.
  if (!Subtarget.useSoftFloat() && Subtarget.hasAVX512()) {
    addRegisterClass(MVT::v1i1,   &X86::VK1RegClass);
    addRegisterClass(MVT::v2i1,   &X86::VK2RegClass);
    addRegisterClass(MVT::v4i1,   &X86::VK4RegClass);
    addRegisterClass(MVT::v8i1,   &X86::VK8RegClass);
    addRegisterClass(MVT::v16i1,  &X86::VK16RegClass);
 
    setOperationAction(ISD::SELECT,             MVT::v1i1, Custom);
    setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v1i1, Custom);
    setOperationAction(ISD::BUILD_VECTOR,       MVT::v1i1, Custom);
 
    setOperationPromotedToType(ISD::FP_TO_SINT,        MVT::v8i1,  MVT::v8i32);
    setOperationPromotedToType(ISD::FP_TO_UINT,        MVT::v8i1,  MVT::v8i32);
    setOperationPromotedToType(ISD::FP_TO_SINT,        MVT::v4i1,  MVT::v4i32);
    setOperationPromotedToType(ISD::FP_TO_UINT,        MVT::v4i1,  MVT::v4i32);
    setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v8i1,  MVT::v8i32);
    setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v8i1,  MVT::v8i32);
    setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v4i1,  MVT::v4i32);
    setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v4i1,  MVT::v4i32);
    setOperationAction(ISD::FP_TO_SINT,                MVT::v2i1,  Custom);
    setOperationAction(ISD::FP_TO_UINT,                MVT::v2i1,  Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT,         MVT::v2i1,  Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT,         MVT::v2i1,  Custom);
    setOperationAction(ISD::FCANONICALIZE, MVT::v8f16, Expand);
    setOperationAction(ISD::FCANONICALIZE, MVT::v16f16, Expand);
    setOperationAction(ISD::FCANONICALIZE, MVT::v32f16, Expand);
 
    // There is no byte sized k-register load or store without AVX512DQ.
    if (!Subtarget.hasDQI()) {
      setOperationAction(ISD::LOAD, MVT::v1i1, Custom);
      setOperationAction(ISD::LOAD, MVT::v2i1, Custom);
      setOperationAction(ISD::LOAD, MVT::v4i1, Custom);
      setOperationAction(ISD::LOAD, MVT::v8i1, Custom);
 
      setOperationAction(ISD::STORE, MVT::v1i1, Custom);
      setOperationAction(ISD::STORE, MVT::v2i1, Custom);
      setOperationAction(ISD::STORE, MVT::v4i1, Custom);
      setOperationAction(ISD::STORE, MVT::v8i1, Custom);
    }
 
    // Extends of v16i1/v8i1/v4i1/v2i1 to 128-bit vectors.
    for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64 }) {
      setOperationAction(ISD::SIGN_EXTEND, VT, Custom);
      setOperationAction(ISD::ZERO_EXTEND, VT, Custom);
      setOperationAction(ISD::ANY_EXTEND,  VT, Custom);
    }
 
    for (auto VT : { MVT::v1i1, MVT::v2i1, MVT::v4i1, MVT::v8i1, MVT::v16i1 })
      setOperationAction(ISD::VSELECT,          VT, Expand);
 
    for (auto VT : { MVT::v2i1, MVT::v4i1, MVT::v8i1, MVT::v16i1 }) {
      setOperationAction(ISD::SETCC,            VT, Custom);
      setOperationAction(ISD::SELECT,           VT, Custom);
      setOperationAction(ISD::TRUNCATE,         VT, Custom);
 
      setOperationAction(ISD::BUILD_VECTOR,     VT, Custom);
      setOperationAction(ISD::CONCAT_VECTORS,   VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
      setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE,   VT,  Custom);
    }
 
    for (auto VT : { MVT::v1i1, MVT::v2i1, MVT::v4i1, MVT::v8i1 })
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
  }
  if (Subtarget.hasDQI() && Subtarget.hasVLX()) {
    for (MVT VT : {MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64}) {
      setOperationAction(ISD::LRINT, VT, Legal);
      setOperationAction(ISD::LLRINT, VT, Legal);
    }
  }
 
  // This block controls legalization for 512-bit operations with 8/16/32/64 bit
  // elements. 512-bits can be disabled based on prefer-vector-width and
  // required-vector-width function attributes.
  if (!Subtarget.useSoftFloat() && Subtarget.useAVX512Regs()) {
    bool HasBWI = Subtarget.hasBWI();
 
    addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
    addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
    addRegisterClass(MVT::v8i64,  &X86::VR512RegClass);
    addRegisterClass(MVT::v8f64,  &X86::VR512RegClass);
    addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
    addRegisterClass(MVT::v32f16, &X86::VR512RegClass);
    addRegisterClass(MVT::v64i8,  &X86::VR512RegClass);
 
    for (auto ExtType : {ISD::ZEXTLOAD, ISD::SEXTLOAD}) {
      setLoadExtAction(ExtType, MVT::v16i32, MVT::v16i8,  Legal);
      setLoadExtAction(ExtType, MVT::v16i32, MVT::v16i16, Legal);
      setLoadExtAction(ExtType, MVT::v8i64,  MVT::v8i8,   Legal);
      setLoadExtAction(ExtType, MVT::v8i64,  MVT::v8i16,  Legal);
      setLoadExtAction(ExtType, MVT::v8i64,  MVT::v8i32,  Legal);
      if (HasBWI)
        setLoadExtAction(ExtType, MVT::v32i16, MVT::v32i8, Legal);
    }
 
    for (MVT VT : { MVT::v16f32, MVT::v8f64 }) {
      setOperationAction(ISD::FMAXIMUM, VT, Custom);
      setOperationAction(ISD::FMINIMUM, VT, Custom);
      setOperationAction(ISD::FMAXIMUMNUM, VT, Custom);
      setOperationAction(ISD::FMINIMUMNUM, VT, Custom);
      setOperationAction(ISD::FNEG,  VT, Custom);
      setOperationAction(ISD::FABS,  VT, Custom);
      setOperationAction(ISD::FMA,   VT, Legal);
      setOperationAction(ISD::STRICT_FMA, VT, Legal);
      setOperationAction(ISD::FCOPYSIGN, VT, Custom);
      setOperationAction(ISD::FCANONICALIZE, VT, Expand);
    }
    setOperationAction(ISD::LRINT, MVT::v16f32,
                       Subtarget.hasDQI() ? Legal : Custom);
    setOperationAction(ISD::LRINT, MVT::v8f64,
                       Subtarget.hasDQI() ? Legal : Custom);
    if (Subtarget.hasDQI())
      setOperationAction(ISD::LLRINT, MVT::v8f64, Legal);
 
    setOperationAction(ISD::AND, MVT::i512, Custom);
    setOperationAction(ISD::OR, MVT::i512, Custom);
    setOperationAction(ISD::XOR, MVT::i512, Custom);
    setOperationAction(ISD::ADD, MVT::i512, Custom);
    setOperationAction(ISD::SUB, MVT::i512, Custom);
    setOperationAction(ISD::SRL, MVT::i512, Custom);
    setOperationAction(ISD::SHL, MVT::i512, Custom);
    setOperationAction(ISD::SRA, MVT::i512, Custom);
    setOperationAction(ISD::FSHR, MVT::i512, Custom);
    setOperationAction(ISD::FSHL, MVT::i512, Custom);
    setOperationAction(ISD::FSHR, MVT::i256, Custom);
    setOperationAction(ISD::FSHL, MVT::i256, Custom);
    setOperationAction(ISD::SELECT, MVT::i512, Custom);
    setOperationAction(ISD::BITREVERSE, MVT::i512, Custom);
 
    for (MVT VT : { MVT::v16i1, MVT::v16i8 }) {
      setOperationPromotedToType(ISD::FP_TO_SINT       , VT, MVT::v16i32);
      setOperationPromotedToType(ISD::FP_TO_UINT       , VT, MVT::v16i32);
      setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, VT, MVT::v16i32);
      setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, VT, MVT::v16i32);
    }
 
    for (MVT VT : { MVT::v16i16, MVT::v16i32 }) {
      setOperationAction(ISD::FP_TO_SINT,        VT, Custom);
      setOperationAction(ISD::FP_TO_UINT,        VT, Custom);
      setOperationAction(ISD::STRICT_FP_TO_SINT, VT, Custom);
      setOperationAction(ISD::STRICT_FP_TO_UINT, VT, Custom);
    }
 
    setOperationAction(ISD::SINT_TO_FP,        MVT::v16i32, Custom);
    setOperationAction(ISD::UINT_TO_FP,        MVT::v16i32, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v16i32, Custom);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v16i32, Custom);
    setOperationAction(ISD::FP_EXTEND,         MVT::v8f64,  Custom);
    setOperationAction(ISD::STRICT_FP_EXTEND,  MVT::v8f64,  Custom);
 
    setOperationAction(ISD::STRICT_FADD,      MVT::v16f32, Legal);
    setOperationAction(ISD::STRICT_FADD,      MVT::v8f64,  Legal);
    setOperationAction(ISD::STRICT_FSUB,      MVT::v16f32, Legal);
    setOperationAction(ISD::STRICT_FSUB,      MVT::v8f64,  Legal);
    setOperationAction(ISD::STRICT_FMUL,      MVT::v16f32, Legal);
    setOperationAction(ISD::STRICT_FMUL,      MVT::v8f64,  Legal);
    setOperationAction(ISD::STRICT_FDIV,      MVT::v16f32, Legal);
    setOperationAction(ISD::STRICT_FDIV,      MVT::v8f64,  Legal);
    setOperationAction(ISD::STRICT_FSQRT,     MVT::v16f32, Legal);
    setOperationAction(ISD::STRICT_FSQRT,     MVT::v8f64,  Legal);
    setOperationAction(ISD::STRICT_FP_ROUND,  MVT::v8f32,  Legal);
 
    setTruncStoreAction(MVT::v8i64,   MVT::v8i8,   Legal);
    setTruncStoreAction(MVT::v8i64,   MVT::v8i16,  Legal);
    setTruncStoreAction(MVT::v8i64,   MVT::v8i32,  Legal);
    setTruncStoreAction(MVT::v16i32,  MVT::v16i8,  Legal);
    setTruncStoreAction(MVT::v16i32,  MVT::v16i16, Legal);
    if (HasBWI)
      setTruncStoreAction(MVT::v32i16,  MVT::v32i8, Legal);
 
    // With 512-bit vectors and no VLX, we prefer to widen MLOAD/MSTORE
    // to 512-bit rather than use the AVX2 instructions so that we can use
    // k-masks.
    if (!Subtarget.hasVLX()) {
      for (auto VT : {MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64,
           MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64}) {
        setOperationAction(ISD::MLOAD,  VT, Custom);
        setOperationAction(ISD::MSTORE, VT, Custom);
      }
    }
 
    setOperationAction(ISD::TRUNCATE,    MVT::v8i32,  Legal);
    setOperationAction(ISD::TRUNCATE,    MVT::v16i16, Legal);
    setOperationAction(ISD::TRUNCATE,    MVT::v32i8,  HasBWI ? Legal : Custom);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64,  Custom);
    setOperationAction(ISD::ANY_EXTEND,  MVT::v32i16, Custom);
    setOperationAction(ISD::ANY_EXTEND,  MVT::v16i32, Custom);
    setOperationAction(ISD::ANY_EXTEND,  MVT::v8i64,  Custom);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64,  Custom);
 
    if (HasBWI) {
      // Extends from v64i1 masks to 512-bit vectors.
      setOperationAction(ISD::SIGN_EXTEND,        MVT::v64i8, Custom);
      setOperationAction(ISD::ZERO_EXTEND,        MVT::v64i8, Custom);
      setOperationAction(ISD::ANY_EXTEND,         MVT::v64i8, Custom);
    }
 
    for (auto VT : { MVT::v16f32, MVT::v8f64 }) {
      setOperationAction(ISD::FFLOOR,            VT, Legal);
      setOperationAction(ISD::STRICT_FFLOOR,     VT, Legal);
      setOperationAction(ISD::FCEIL,             VT, Legal);
      setOperationAction(ISD::STRICT_FCEIL,      VT, Legal);
      setOperationAction(ISD::FTRUNC,            VT, Legal);
      setOperationAction(ISD::STRICT_FTRUNC,     VT, Legal);
      setOperationAction(ISD::FRINT,             VT, Legal);
      setOperationAction(ISD::STRICT_FRINT,      VT, Legal);
      setOperationAction(ISD::FNEARBYINT,        VT, Legal);
      setOperationAction(ISD::STRICT_FNEARBYINT, VT, Legal);
      setOperationAction(ISD::FROUNDEVEN,        VT, Legal);
      setOperationAction(ISD::STRICT_FROUNDEVEN, VT, Legal);
 
      setOperationAction(ISD::FROUND,            VT, Custom);
    }
 
    for (auto VT : {MVT::v32i16, MVT::v16i32, MVT::v8i64}) {
      setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, VT, Custom);
      setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Custom);
    }
 
    setOperationAction(ISD::ADD, MVT::v32i16, HasBWI ? Legal : Custom);
    setOperationAction(ISD::SUB, MVT::v32i16, HasBWI ? Legal : Custom);
    setOperationAction(ISD::ADD, MVT::v64i8,  HasBWI ? Legal : Custom);
    setOperationAction(ISD::SUB, MVT::v64i8,  HasBWI ? Legal : Custom);
 
    setOperationAction(ISD::MUL, MVT::v8i64,  Custom);
    setOperationAction(ISD::MUL, MVT::v16i32, Legal);
    setOperationAction(ISD::MUL, MVT::v32i16, HasBWI ? Legal : Custom);
    setOperationAction(ISD::MUL, MVT::v64i8,  Custom);
 
    setOperationAction(ISD::MULHU, MVT::v16i32, Custom);
    setOperationAction(ISD::MULHS, MVT::v16i32, Custom);
    setOperationAction(ISD::MULHS, MVT::v32i16, HasBWI ? Legal : Custom);
    setOperationAction(ISD::MULHU, MVT::v32i16, HasBWI ? Legal : Custom);
    setOperationAction(ISD::MULHS, MVT::v64i8,  Custom);
    setOperationAction(ISD::MULHU, MVT::v64i8,  Custom);
    setOperationAction(ISD::AVGCEILU, MVT::v32i16, HasBWI ? Legal : Custom);
    setOperationAction(ISD::AVGCEILU, MVT::v64i8,  HasBWI ? Legal : Custom);
 
    setOperationAction(ISD::SMULO, MVT::v64i8, Custom);
    setOperationAction(ISD::UMULO, MVT::v64i8, Custom);
 
    for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32, MVT::v8i64 }) {
      setOperationAction(ISD::SRL,              VT, Custom);
      setOperationAction(ISD::SHL,              VT, Custom);
      setOperationAction(ISD::SRA,              VT, Custom);
      setOperationAction(ISD::ROTL,             VT, Custom);
      setOperationAction(ISD::ROTR,             VT, Custom);
      setOperationAction(ISD::SETCC,            VT, Custom);
      setOperationAction(ISD::ABDS,             VT, Custom);
      setOperationAction(ISD::ABDU,             VT, Custom);
      setOperationAction(ISD::BITREVERSE,       VT, Custom);
      setOperationAction(ISD::VECREDUCE_AND,    VT, Custom);
      setOperationAction(ISD::VECREDUCE_OR,     VT, Custom);
      setOperationAction(ISD::VECREDUCE_XOR,    VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMAX,   VT, Custom);
      setOperationAction(ISD::VECREDUCE_SMIN,   VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMAX,   VT, Custom);
      setOperationAction(ISD::VECREDUCE_UMIN,   VT, Custom);
 
      // The condition codes aren't legal in SSE/AVX and under AVX512 we use
      // setcc all the way to isel and prefer SETGT in some isel patterns.
      setCondCodeAction(ISD::SETLT, VT, Custom);
      setCondCodeAction(ISD::SETLE, VT, Custom);
    }
 
    setOperationAction(ISD::SETCC,          MVT::v8f64, Custom);
    setOperationAction(ISD::SETCC,          MVT::v16f32, Custom);
    setOperationAction(ISD::STRICT_FSETCC,  MVT::v8f64, Custom);
    setOperationAction(ISD::STRICT_FSETCC,  MVT::v16f32, Custom);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::v8f64, Custom);
    setOperationAction(ISD::STRICT_FSETCCS, MVT::v16f32, Custom);
 
    for (auto VT : { MVT::v16i32, MVT::v8i64 }) {
      setOperationAction(ISD::SMAX,             VT, Legal);
      setOperationAction(ISD::UMAX,             VT, Legal);
      setOperationAction(ISD::SMIN,             VT, Legal);
      setOperationAction(ISD::UMIN,             VT, Legal);
      setOperationAction(ISD::ROTL,             VT, Legal);
      setOperationAction(ISD::ROTR,             VT, Legal);
      setOperationAction(ISD::ABS,              VT, Legal);
      setOperationAction(ISD::CTPOP,            VT, Custom);
    }
 
    for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
      setOperationAction(ISD::ABS,     VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::CTPOP,   VT, Subtarget.hasBITALG() ? Legal : Custom);
      setOperationAction(ISD::CTLZ,    VT, Custom);
      setOperationAction(ISD::SMAX,    VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::UMAX,    VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::SMIN,    VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::UMIN,    VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::UADDSAT, VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::SADDSAT, VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::USUBSAT, VT, HasBWI ? Legal : Custom);
      setOperationAction(ISD::SSUBSAT, VT, HasBWI ? Legal : Custom);
    }
 
    setOperationAction(ISD::FSHL,       MVT::v64i8, Custom);
    setOperationAction(ISD::FSHR,       MVT::v64i8, Custom);
    setOperationAction(ISD::FSHL,      MVT::v32i16, Custom);
    setOperationAction(ISD::FSHR,      MVT::v32i16, Custom);
    setOperationAction(ISD::FSHL,      MVT::v16i32, Custom);
    setOperationAction(ISD::FSHR,      MVT::v16i32, Custom);
 
    if (Subtarget.hasDQI() || Subtarget.hasFP16())
      for (auto Opc : {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::STRICT_SINT_TO_FP,
                       ISD::STRICT_UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT,
                       ISD::STRICT_FP_TO_SINT, ISD::STRICT_FP_TO_UINT})
        setOperationAction(Opc,           MVT::v8i64, Custom);
 
    if (Subtarget.hasDQI())
      setOperationAction(ISD::MUL,        MVT::v8i64, Legal);
 
    if (Subtarget.hasCDI()) {
      // NonVLX sub-targets extend 128/256 vectors to use the 512 version.
      for (auto VT : { MVT::v16i32, MVT::v8i64} ) {
        setOperationAction(ISD::CTLZ,            VT, Legal);
      }
    } // Subtarget.hasCDI()
 
    if (Subtarget.hasVPOPCNTDQ()) {
      for (auto VT : { MVT::v16i32, MVT::v8i64 })
        setOperationAction(ISD::CTPOP, VT, Legal);
      setOperationAction(ISD::CTPOP, MVT::i512, Custom);
    }
 
    // Extract subvector is special because the value type
    // (result) is 256-bit but the source is 512-bit wide.
    // 128-bit was made Legal under AVX1.
    for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64,
                     MVT::v16f16, MVT::v8f32, MVT::v4f64 })
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
 
    for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32, MVT::v8i64,
                     MVT::v32f16, MVT::v16f32, MVT::v8f64 }) {
      setOperationAction(ISD::CONCAT_VECTORS,     VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR,   VT, Legal);
      setOperationAction(ISD::SELECT,             VT, Custom);
      setOperationAction(ISD::VSELECT,            VT, Custom);
      setOperationAction(ISD::BUILD_VECTOR,       VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE,     VT, Custom);
      setOperationAction(ISD::SCALAR_TO_VECTOR,   VT, Custom);
      setOperationAction(ISD::INSERT_VECTOR_ELT,  VT, Custom);
    }
    setF16Action(MVT::v32f16, Expand);
    setOperationAction(ISD::FP_ROUND, MVT::v16f16, Custom);
    setOperationAction(ISD::STRICT_FP_ROUND, MVT::v16f16, Custom);
    setOperationAction(ISD::FP_EXTEND, MVT::v16f32, Custom);
    setOperationAction(ISD::STRICT_FP_EXTEND, MVT::v16f32, Custom);
    for (unsigned Opc : {ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FDIV})
      setOperationPromotedToType(Opc, MVT::v32f16, MVT::v32f32);
    setOperationAction(ISD::SETCC, MVT::v32f16, Custom);
 
    for (auto VT : { MVT::v16i32, MVT::v8i64, MVT::v16f32, MVT::v8f64 }) {
      setOperationAction(ISD::MLOAD,               VT, Legal);
      setOperationAction(ISD::MSTORE,              VT, Legal);
      setOperationAction(ISD::MGATHER,             VT, Custom);
      setOperationAction(ISD::MSCATTER,            VT, Custom);
    }
    if (HasBWI) {
      for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
        setOperationAction(ISD::MLOAD,        VT, Legal);
        setOperationAction(ISD::MSTORE,       VT, Legal);
      }
    } else {
      setOperationAction(ISD::STORE, MVT::v32i16, Custom);
      setOperationAction(ISD::STORE, MVT::v64i8,  Custom);
    }
 
    if (Subtarget.hasVBMI2()) {
      for (auto VT : {MVT::v32i16, MVT::v16i32, MVT::v8i64}) {
        setOperationAction(ISD::FSHL, VT, Legal);
        setOperationAction(ISD::FSHR, VT, Legal);
      }
 
      setOperationAction(ISD::ROTL, MVT::v32i16, Legal);
      setOperationAction(ISD::ROTR, MVT::v32i16, Legal);
    }
 
    // Only PCLMUL required as we always unroll clmul vectors.
    if (Subtarget.hasPCLMUL()) {
      for (auto VT : {MVT::v16i32, MVT::v8i64}) {
        setOperationAction(ISD::CLMUL, VT, Custom);
        setOperationAction(ISD::CLMULH, VT, Custom);
      }
    }
 
    setOperationAction(ISD::FNEG, MVT::v32f16, Custom);
    setOperationAction(ISD::FABS, MVT::v32f16, Custom);
    setOperationAction(ISD::FCOPYSIGN, MVT::v32f16, Custom);
 
    if (Subtarget.hasGFNI()) {
      setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
      setOperationAction(ISD::CTTZ, MVT::v64i8, Custom);
    }
  }// useAVX512Regs
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasVBMI2()) {
    for (auto VT : {MVT::v8i16, MVT::v4i32, MVT::v2i64, MVT::v16i16, MVT::v8i32,
                    MVT::v4i64}) {
      setOperationAction(ISD::FSHL, VT, Legal);
      setOperationAction(ISD::FSHR, VT, Legal);
    }
 
    setOperationAction(ISD::ROTL, MVT::v16i16, Legal);
    setOperationAction(ISD::ROTR, MVT::v16i16, Legal);
    setOperationAction(ISD::ROTL, MVT::v8i16, Legal);
    setOperationAction(ISD::ROTR, MVT::v8i16, Legal);
  }
 
  // This block controls legalization for operations that don't have
  // pre-AVX512 equivalents. Without VLX we use 512-bit operations for
  // narrower widths.
  if (!Subtarget.useSoftFloat() && Subtarget.hasAVX512()) {
    for (MVT VT : {MVT::f16, MVT::f32, MVT::f64, MVT::v8f16, MVT::v4f32,
                   MVT::v2f64, MVT::v16f16, MVT::v8f32, MVT::v4f64, MVT::v32f16,
                   MVT::v16f32, MVT::v8f64})
      setOperationAction(ISD::FLDEXP, VT, Custom);
 
    // These operations are handled on non-VLX by artificially widening in
    // isel patterns.
    setOperationAction(ISD::STRICT_FP_TO_UINT,  MVT::v8i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT,  MVT::v4i32, Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT,  MVT::v2i32, Custom);
 
    if (Subtarget.hasDQI()) {
      // Fast v2f32 SINT_TO_FP( v2i64 ) custom conversion.
      // v2f32 UINT_TO_FP is already custom under SSE2.
      assert(isOperationCustom(ISD::UINT_TO_FP, MVT::v2f32) &&
             isOperationCustom(ISD::STRICT_UINT_TO_FP, MVT::v2f32) &&
             "Unexpected operation action!");
      // v2i64 FP_TO_S/UINT(v2f32) custom conversion.
      setOperationAction(ISD::FP_TO_SINT,        MVT::v2f32, Custom);
      setOperationAction(ISD::FP_TO_UINT,        MVT::v2f32, Custom);
      setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2f32, Custom);
      setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2f32, Custom);
    }
 
    for (auto VT : { MVT::v2i64, MVT::v4i64 }) {
      setOperationAction(ISD::SMAX, VT, Legal);
      setOperationAction(ISD::UMAX, VT, Legal);
      setOperationAction(ISD::SMIN, VT, Legal);
      setOperationAction(ISD::UMIN, VT, Legal);
      setOperationAction(ISD::ABS,  VT, Legal);
    }
 
    for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64 }) {
      setOperationAction(ISD::ROTL,     VT, Legal);
      setOperationAction(ISD::ROTR,     VT, Legal);
    }
 
    // Custom legalize 2x32 to get a little better code.
    setOperationAction(ISD::MSCATTER, MVT::v2f32, Custom);
    setOperationAction(ISD::MSCATTER, MVT::v2i32, Custom);
 
    for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64,
                     MVT::v4f32, MVT::v8f32, MVT::v2f64, MVT::v4f64 })
      setOperationAction(ISD::MSCATTER, VT, Custom);
 
    if (Subtarget.hasDQI()) {
      for (auto Opc : {ISD::SINT_TO_FP, ISD::UINT_TO_FP, ISD::STRICT_SINT_TO_FP,
                       ISD::STRICT_UINT_TO_FP, ISD::FP_TO_SINT, ISD::FP_TO_UINT,
                       ISD::STRICT_FP_TO_SINT, ISD::STRICT_FP_TO_UINT}) {
        setOperationAction(Opc, MVT::v2i64, Custom);
        setOperationAction(Opc, MVT::v4i64, Custom);
      }
      setOperationAction(ISD::MUL, MVT::v2i64, Legal);
      setOperationAction(ISD::MUL, MVT::v4i64, Legal);
    }
 
    if (Subtarget.hasCDI()) {
      for (auto VT : {MVT::i256, MVT::i512}) {
        if (VT == MVT::i512 && !Subtarget.useAVX512Regs())
          continue;
        setOperationAction(ISD::CTLZ, VT, Custom);
        setOperationAction(ISD::CTTZ, VT, Custom);
        setOperationAction(ISD::CTLZ_ZERO_POISON, VT, Custom);
        setOperationAction(ISD::CTTZ_ZERO_POISON, VT, Custom);
      }
      for (auto VT : { MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64 }) {
        setOperationAction(ISD::CTLZ,            VT, Legal);
      }
    } // Subtarget.hasCDI()
 
    if (Subtarget.hasVPOPCNTDQ()) {
      for (auto VT : {MVT::v4i32, MVT::v8i32, MVT::v2i64, MVT::v4i64})
        setOperationAction(ISD::CTPOP, VT, Legal);
      setOperationAction(ISD::CTPOP, MVT::i256, Custom);
    }
 
    // We can try to convert vectors to different sizes to leverage legal
    // `vpcompress` cases. So we mark these supported vector sizes as Custom and
    // then specialize to Legal below.
    for (MVT VT : {MVT::v8i32, MVT::v8f32, MVT::v4i32, MVT::v4f32, MVT::v4i64,
                   MVT::v4f64, MVT::v2i64, MVT::v2f64, MVT::v16i8, MVT::v8i16,
                   MVT::v16i16, MVT::v8i8})
      setOperationAction(ISD::VECTOR_COMPRESS, VT, Custom);
 
    // Legal vpcompress depends on various AVX512 extensions.
    // Legal in AVX512F
    for (MVT VT : {MVT::v16i32, MVT::v16f32, MVT::v8i64, MVT::v8f64})
      setOperationAction(ISD::VECTOR_COMPRESS, VT, Legal);
 
    // Legal in AVX512F + AVX512VL
    if (Subtarget.hasVLX())
      for (MVT VT : {MVT::v8i32, MVT::v8f32, MVT::v4i32, MVT::v4f32, MVT::v4i64,
                     MVT::v4f64, MVT::v2i64, MVT::v2f64})
        setOperationAction(ISD::VECTOR_COMPRESS, VT, Legal);
 
    // Legal in AVX512F + AVX512VBMI2
    if (Subtarget.hasVBMI2())
      for (MVT VT : {MVT::v32i16, MVT::v64i8})
        setOperationAction(ISD::VECTOR_COMPRESS, VT, Legal);
 
    // Legal in AVX512F + AVX512VL + AVX512VBMI2
    if (Subtarget.hasVBMI2() && Subtarget.hasVLX())
      for (MVT VT : {MVT::v16i8, MVT::v8i16, MVT::v32i8, MVT::v16i16})
        setOperationAction(ISD::VECTOR_COMPRESS, VT, Legal);
  }
 
  // This block control legalization of v32i1/v64i1 which are available with
  // AVX512BW..
  if (!Subtarget.useSoftFloat() && Subtarget.hasBWI()) {
    addRegisterClass(MVT::v32i1,  &X86::VK32RegClass);
    addRegisterClass(MVT::v64i1,  &X86::VK64RegClass);
 
    for (auto VT : { MVT::v32i1, MVT::v64i1 }) {
      setOperationAction(ISD::VSELECT,            VT, Expand);
      setOperationAction(ISD::TRUNCATE,           VT, Custom);
      setOperationAction(ISD::SETCC,              VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::INSERT_VECTOR_ELT,  VT, Custom);
      setOperationAction(ISD::SELECT,             VT, Custom);
      setOperationAction(ISD::BUILD_VECTOR,       VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE,     VT, Custom);
      setOperationAction(ISD::CONCAT_VECTORS,     VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR,   VT, Custom);
    }
 
    for (auto VT : { MVT::v16i1, MVT::v32i1 })
      setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
 
    // Extends from v32i1 masks to 256-bit vectors.
    setOperationAction(ISD::SIGN_EXTEND,        MVT::v32i8, Custom);
    setOperationAction(ISD::ZERO_EXTEND,        MVT::v32i8, Custom);
    setOperationAction(ISD::ANY_EXTEND,         MVT::v32i8, Custom);
 
    for (auto VT : {MVT::v32i8, MVT::v16i8, MVT::v16i16, MVT::v8i16,
                    MVT::v16f16, MVT::v8f16}) {
      setOperationAction(ISD::MLOAD,  VT, Subtarget.hasVLX() ? Legal : Custom);
      setOperationAction(ISD::MSTORE, VT, Subtarget.hasVLX() ? Legal : Custom);
    }
 
    // These operations are handled on non-VLX by artificially widening in
    // isel patterns.
    // TODO: Custom widen in lowering on non-VLX and drop the isel patterns?
 
    if (Subtarget.hasBITALG()) {
      for (auto VT : { MVT::v16i8, MVT::v32i8, MVT::v8i16, MVT::v16i16 })
        setOperationAction(ISD::CTPOP, VT, Legal);
    }
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasFP16()) {
    auto setGroup = [&] (MVT VT) {
      setOperationAction(ISD::FADD,               VT, Legal);
      setOperationAction(ISD::STRICT_FADD,        VT, Legal);
      setOperationAction(ISD::FSUB,               VT, Legal);
      setOperationAction(ISD::STRICT_FSUB,        VT, Legal);
      setOperationAction(ISD::FMUL,               VT, Legal);
      setOperationAction(ISD::STRICT_FMUL,        VT, Legal);
      setOperationAction(ISD::FDIV,               VT, Legal);
      setOperationAction(ISD::STRICT_FDIV,        VT, Legal);
      setOperationAction(ISD::FSQRT,              VT, Legal);
      setOperationAction(ISD::STRICT_FSQRT,       VT, Legal);
 
      setOperationAction(ISD::FFLOOR,             VT, Legal);
      setOperationAction(ISD::STRICT_FFLOOR,      VT, Legal);
      setOperationAction(ISD::FCEIL,              VT, Legal);
      setOperationAction(ISD::STRICT_FCEIL,       VT, Legal);
      setOperationAction(ISD::FTRUNC,             VT, Legal);
      setOperationAction(ISD::STRICT_FTRUNC,      VT, Legal);
      setOperationAction(ISD::FRINT,              VT, Legal);
      setOperationAction(ISD::STRICT_FRINT,       VT, Legal);
      setOperationAction(ISD::FNEARBYINT,         VT, Legal);
      setOperationAction(ISD::STRICT_FNEARBYINT,  VT, Legal);
      setOperationAction(ISD::FROUNDEVEN, VT, Legal);
      setOperationAction(ISD::STRICT_FROUNDEVEN, VT, Legal);
 
      setOperationAction(ISD::FROUND,             VT, Custom);
 
      setOperationAction(ISD::LOAD,               VT, Legal);
      setOperationAction(ISD::STORE,              VT, Legal);
 
      setOperationAction(ISD::FMA,                VT, Legal);
      setOperationAction(ISD::STRICT_FMA,         VT, Legal);
      setOperationAction(ISD::VSELECT,            VT, Legal);
      setOperationAction(ISD::BUILD_VECTOR,       VT, Custom);
      setOperationAction(ISD::SELECT,             VT, Custom);
 
      setOperationAction(ISD::FNEG,               VT, Custom);
      setOperationAction(ISD::FABS,               VT, Custom);
      setOperationAction(ISD::FCOPYSIGN,          VT, Custom);
      setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE,     VT, Custom);
 
      setOperationAction(ISD::SETCC,              VT, Custom);
      setOperationAction(ISD::STRICT_FSETCC,      VT, Custom);
      setOperationAction(ISD::STRICT_FSETCCS,     VT, Custom);
    };
 
    // AVX512_FP16 scalar operations
    setGroup(MVT::f16);
    setOperationAction(ISD::FREM,                 MVT::f16, Promote);
    setOperationAction(ISD::STRICT_FREM,          MVT::f16, Promote);
    setOperationAction(ISD::SELECT_CC,            MVT::f16, Expand);
    setOperationAction(ISD::BR_CC,                MVT::f16, Expand);
    setOperationAction(ISD::STRICT_FROUND,        MVT::f16, Promote);
    setOperationAction(ISD::FROUNDEVEN,           MVT::f16, Legal);
    setOperationAction(ISD::STRICT_FROUNDEVEN,    MVT::f16, Legal);
    setOperationAction(ISD::FP_ROUND,             MVT::f16, Custom);
    setOperationAction(ISD::STRICT_FP_ROUND,      MVT::f16, Custom);
    setOperationAction(ISD::FMAXIMUM,             MVT::f16, Custom);
    setOperationAction(ISD::FMINIMUM,             MVT::f16, Custom);
    setOperationAction(ISD::FMAXIMUMNUM,          MVT::f16, Custom);
    setOperationAction(ISD::FMINIMUMNUM,          MVT::f16, Custom);
    setOperationAction(ISD::FP_EXTEND,            MVT::f32, Legal);
    setOperationAction(ISD::STRICT_FP_EXTEND,     MVT::f32, Legal);
    setOperationAction(ISD::LRINT,                MVT::f16, Legal);
    setOperationAction(ISD::LLRINT,               MVT::f16, Legal);
 
    setCondCodeAction(ISD::SETOEQ, MVT::f16, Expand);
    setCondCodeAction(ISD::SETUNE, MVT::f16, Expand);
 
    if (Subtarget.useAVX512Regs()) {
      setGroup(MVT::v32f16);
      setOperationAction(ISD::SCALAR_TO_VECTOR,       MVT::v32f16, Custom);
      setOperationAction(ISD::SINT_TO_FP,             MVT::v32i16, Legal);
      setOperationAction(ISD::STRICT_SINT_TO_FP,      MVT::v32i16, Legal);
      setOperationAction(ISD::UINT_TO_FP,             MVT::v32i16, Legal);
      setOperationAction(ISD::STRICT_UINT_TO_FP,      MVT::v32i16, Legal);
      setOperationAction(ISD::FP_ROUND,               MVT::v16f16, Legal);
      setOperationAction(ISD::STRICT_FP_ROUND,        MVT::v16f16, Legal);
      setOperationAction(ISD::FP_EXTEND,              MVT::v16f32, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND,       MVT::v16f32, Legal);
      setOperationAction(ISD::FP_EXTEND,              MVT::v8f64,  Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND,       MVT::v8f64,  Legal);
      setOperationAction(ISD::INSERT_VECTOR_ELT,      MVT::v32f16, Custom);
 
      setOperationAction(ISD::FP_TO_SINT,             MVT::v32i16, Custom);
      setOperationAction(ISD::STRICT_FP_TO_SINT,      MVT::v32i16, Custom);
      setOperationAction(ISD::FP_TO_UINT,             MVT::v32i16, Custom);
      setOperationAction(ISD::STRICT_FP_TO_UINT,      MVT::v32i16, Custom);
      setOperationPromotedToType(ISD::FP_TO_SINT,     MVT::v32i8,  MVT::v32i16);
      setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v32i8,
                                 MVT::v32i16);
      setOperationPromotedToType(ISD::FP_TO_UINT,     MVT::v32i8,  MVT::v32i16);
      setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v32i8,
                                 MVT::v32i16);
      setOperationPromotedToType(ISD::FP_TO_SINT,     MVT::v32i1,  MVT::v32i16);
      setOperationPromotedToType(ISD::STRICT_FP_TO_SINT, MVT::v32i1,
                                 MVT::v32i16);
      setOperationPromotedToType(ISD::FP_TO_UINT,     MVT::v32i1,  MVT::v32i16);
      setOperationPromotedToType(ISD::STRICT_FP_TO_UINT, MVT::v32i1,
                                 MVT::v32i16);
 
      setOperationAction(ISD::EXTRACT_SUBVECTOR,      MVT::v16f16, Legal);
      setOperationAction(ISD::INSERT_SUBVECTOR,       MVT::v32f16, Legal);
      setOperationAction(ISD::CONCAT_VECTORS,         MVT::v32f16, Custom);
 
      setLoadExtAction(ISD::EXTLOAD, MVT::v8f64,  MVT::v8f16,  Legal);
      setLoadExtAction(ISD::EXTLOAD, MVT::v16f32, MVT::v16f16, Legal);
 
      setOperationAction(ISD::FMINIMUM, MVT::v32f16, Custom);
      setOperationAction(ISD::FMAXIMUM, MVT::v32f16, Custom);
      setOperationAction(ISD::FMINIMUMNUM, MVT::v32f16, Custom);
      setOperationAction(ISD::FMAXIMUMNUM, MVT::v32f16, Custom);
      setOperationAction(ISD::LRINT, MVT::v32f16, Legal);
      setOperationAction(ISD::LLRINT, MVT::v8f16, Legal);
    }
 
    setOperationAction(ISD::FP_TO_SINT,         MVT::v8i16, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT,  MVT::v8i16, Custom);
    setOperationAction(ISD::FP_TO_UINT,         MVT::v8i16, Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT,  MVT::v8i16, Custom);
 
    if (Subtarget.hasVLX()) {
      setGroup(MVT::v8f16);
      setGroup(MVT::v16f16);
 
      setOperationAction(ISD::SCALAR_TO_VECTOR,   MVT::v8f16,  Legal);
      setOperationAction(ISD::SCALAR_TO_VECTOR,   MVT::v16f16, Custom);
      setOperationAction(ISD::SINT_TO_FP,         MVT::v16i16, Legal);
      setOperationAction(ISD::STRICT_SINT_TO_FP,  MVT::v16i16, Legal);
      setOperationAction(ISD::SINT_TO_FP,         MVT::v8i16,  Legal);
      setOperationAction(ISD::STRICT_SINT_TO_FP,  MVT::v8i16,  Legal);
      setOperationAction(ISD::UINT_TO_FP,         MVT::v16i16, Legal);
      setOperationAction(ISD::STRICT_UINT_TO_FP,  MVT::v16i16, Legal);
      setOperationAction(ISD::UINT_TO_FP,         MVT::v8i16,  Legal);
      setOperationAction(ISD::STRICT_UINT_TO_FP,  MVT::v8i16,  Legal);
 
      setOperationAction(ISD::FP_ROUND,           MVT::v8f16, Legal);
      setOperationAction(ISD::STRICT_FP_ROUND,    MVT::v8f16, Legal);
      setOperationAction(ISD::FP_EXTEND,          MVT::v8f32, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND,   MVT::v8f32, Legal);
      setOperationAction(ISD::FP_EXTEND,          MVT::v4f64, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND,   MVT::v4f64, Legal);
 
      // INSERT_VECTOR_ELT v8f16 extended to VECTOR_SHUFFLE
      setOperationAction(ISD::INSERT_VECTOR_ELT,    MVT::v8f16,  Custom);
      setOperationAction(ISD::INSERT_VECTOR_ELT,    MVT::v16f16, Custom);
 
      setOperationAction(ISD::EXTRACT_SUBVECTOR,    MVT::v8f16, Legal);
      setOperationAction(ISD::INSERT_SUBVECTOR,     MVT::v16f16, Legal);
      setOperationAction(ISD::CONCAT_VECTORS,       MVT::v16f16, Custom);
 
      setLoadExtAction(ISD::EXTLOAD, MVT::v4f64, MVT::v4f16, Legal);
      setLoadExtAction(ISD::EXTLOAD, MVT::v2f64, MVT::v2f16, Legal);
      setLoadExtAction(ISD::EXTLOAD, MVT::v8f32, MVT::v8f16, Legal);
      setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, MVT::v4f16, Legal);
 
      // Need to custom widen these to prevent scalarization.
      setOperationAction(ISD::LOAD,  MVT::v4f16, Custom);
      setOperationAction(ISD::STORE, MVT::v4f16, Custom);
 
      setOperationAction(ISD::FMINIMUM, MVT::v8f16, Custom);
      setOperationAction(ISD::FMAXIMUM, MVT::v8f16, Custom);
      setOperationAction(ISD::FMINIMUMNUM, MVT::v8f16, Custom);
      setOperationAction(ISD::FMAXIMUMNUM, MVT::v8f16, Custom);
 
      setOperationAction(ISD::FMINIMUM, MVT::v16f16, Custom);
      setOperationAction(ISD::FMAXIMUM, MVT::v16f16, Custom);
      setOperationAction(ISD::FMINIMUMNUM, MVT::v16f16, Custom);
      setOperationAction(ISD::FMAXIMUMNUM, MVT::v16f16, Custom);
      setOperationAction(ISD::LRINT, MVT::v8f16, Legal);
      setOperationAction(ISD::LRINT, MVT::v16f16, Legal);
    }
  }
 
  if (!Subtarget.useSoftFloat() &&
      (Subtarget.hasAVXNECONVERT() || Subtarget.hasBF16())) {
    addRegisterClass(MVT::v8bf16, Subtarget.hasAVX512() ? &X86::VR128XRegClass
                                                        : &X86::VR128RegClass);
    addRegisterClass(MVT::v16bf16, Subtarget.hasAVX512() ? &X86::VR256XRegClass
                                                         : &X86::VR256RegClass);
    // We set the type action of bf16 to TypeSoftPromoteHalf, but we don't
    // provide the method to promote BUILD_VECTOR and INSERT_VECTOR_ELT.
    // Set the operation action Custom to do the customization later.
    setOperationAction(ISD::BUILD_VECTOR, MVT::bf16, Custom);
    setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::bf16, Custom);
    for (auto VT : {MVT::v8bf16, MVT::v16bf16}) {
      setF16Action(VT, Expand);
      if (!Subtarget.hasBF16())
        setOperationAction(ISD::VSELECT, VT, Custom);
      setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
      setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Legal);
      setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
    }
    for (unsigned Opc : {ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FDIV}) {
      setOperationPromotedToType(Opc, MVT::v8bf16, MVT::v8f32);
      setOperationPromotedToType(Opc, MVT::v16bf16, MVT::v16f32);
    }
    setOperationAction(ISD::SETCC, MVT::v8bf16, Custom);
    setOperationAction(ISD::SETCC, MVT::v16bf16, Custom);
    setOperationAction(ISD::FP_ROUND, MVT::v8bf16, Custom);
    addLegalFPImmediate(APFloat::getZero(APFloat::BFloat()));
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasBF16() &&
      Subtarget.useAVX512Regs()) {
    addRegisterClass(MVT::v32bf16, &X86::VR512RegClass);
    setF16Action(MVT::v32bf16, Expand);
    for (unsigned Opc : {ISD::FADD, ISD::FSUB, ISD::FMUL, ISD::FDIV})
      setOperationPromotedToType(Opc, MVT::v32bf16, MVT::v32f32);
    setOperationAction(ISD::SETCC, MVT::v32bf16, Custom);
    setOperationAction(ISD::BUILD_VECTOR, MVT::v32bf16, Custom);
    setOperationAction(ISD::FP_ROUND, MVT::v16bf16, Custom);
    setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32bf16, Custom);
    setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32bf16, Legal);
    setOperationAction(ISD::CONCAT_VECTORS, MVT::v32bf16, Custom);
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasAVX10_2()) {
    setOperationAction(ISD::FADD, MVT::v32bf16, Legal);
    setOperationAction(ISD::FSUB, MVT::v32bf16, Legal);
    setOperationAction(ISD::FMUL, MVT::v32bf16, Legal);
    setOperationAction(ISD::FDIV, MVT::v32bf16, Legal);
    setOperationAction(ISD::FSQRT, MVT::v32bf16, Legal);
    setOperationAction(ISD::FMA, MVT::v32bf16, Legal);
    setOperationAction(ISD::SETCC, MVT::v32bf16, Custom);
    setOperationAction(ISD::FMINIMUM, MVT::v32bf16, Custom);
    setOperationAction(ISD::FMAXIMUM, MVT::v32bf16, Custom);
    setOperationAction(ISD::FMINIMUMNUM, MVT::v32bf16, Custom);
    setOperationAction(ISD::FMAXIMUMNUM, MVT::v32bf16, Custom);
    for (auto VT : {MVT::v8bf16, MVT::v16bf16}) {
      setOperationAction(ISD::FADD, VT, Legal);
      setOperationAction(ISD::FSUB, VT, Legal);
      setOperationAction(ISD::FMUL, VT, Legal);
      setOperationAction(ISD::FDIV, VT, Legal);
      setOperationAction(ISD::FSQRT, VT, Legal);
      setOperationAction(ISD::FMA, VT, Legal);
      setOperationAction(ISD::SETCC, VT, Custom);
      setOperationAction(ISD::FMINIMUM, VT, Custom);
      setOperationAction(ISD::FMAXIMUM, VT, Custom);
      setOperationAction(ISD::FMINIMUMNUM, VT, Custom);
      setOperationAction(ISD::FMAXIMUMNUM, VT, Custom);
    }
    for (auto VT : {MVT::f16, MVT::f32, MVT::f64}) {
      setCondCodeAction(ISD::SETOEQ, VT, Custom);
      setCondCodeAction(ISD::SETUNE, VT, Custom);
    }
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasVLX()) {
    setTruncStoreAction(MVT::v4i64, MVT::v4i8,  Legal);
    setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
    setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
    setTruncStoreAction(MVT::v8i32, MVT::v8i8,  Legal);
    setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
 
    setTruncStoreAction(MVT::v2i64, MVT::v2i8,  Legal);
    setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
    setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
    setTruncStoreAction(MVT::v4i32, MVT::v4i8,  Legal);
    setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
 
    if (Subtarget.hasBWI()) {
      setTruncStoreAction(MVT::v16i16,  MVT::v16i8, Legal);
      setTruncStoreAction(MVT::v8i16,   MVT::v8i8,  Legal);
    }
 
    if (Subtarget.hasFP16()) {
      // vcvttph2[u]dq v4f16 -> v4i32/64, v2f16 -> v2i32/64
      setOperationAction(ISD::FP_TO_SINT,        MVT::v2f16, Custom);
      setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v2f16, Custom);
      setOperationAction(ISD::FP_TO_UINT,        MVT::v2f16, Custom);
      setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v2f16, Custom);
      setOperationAction(ISD::FP_TO_SINT,        MVT::v4f16, Custom);
      setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::v4f16, Custom);
      setOperationAction(ISD::FP_TO_UINT,        MVT::v4f16, Custom);
      setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::v4f16, Custom);
      // vcvt[u]dq2ph v4i32/64 -> v4f16, v2i32/64 -> v2f16
      setOperationAction(ISD::SINT_TO_FP,        MVT::v2f16, Custom);
      setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v2f16, Custom);
      setOperationAction(ISD::UINT_TO_FP,        MVT::v2f16, Custom);
      setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v2f16, Custom);
      setOperationAction(ISD::SINT_TO_FP,        MVT::v4f16, Custom);
      setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::v4f16, Custom);
      setOperationAction(ISD::UINT_TO_FP,        MVT::v4f16, Custom);
      setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::v4f16, Custom);
      // vcvtps2phx v4f32 -> v4f16, v2f32 -> v2f16
      setOperationAction(ISD::FP_ROUND,          MVT::v2f16, Custom);
      setOperationAction(ISD::STRICT_FP_ROUND,   MVT::v2f16, Custom);
      setOperationAction(ISD::FP_ROUND,          MVT::v4f16, Custom);
      setOperationAction(ISD::STRICT_FP_ROUND,   MVT::v4f16, Custom);
      // vcvtph2psx v4f16 -> v4f32, v2f16 -> v2f32
      setOperationAction(ISD::FP_EXTEND,         MVT::v2f16, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND,  MVT::v2f16, Custom);
      setOperationAction(ISD::FP_EXTEND,         MVT::v4f16, Custom);
      setOperationAction(ISD::STRICT_FP_EXTEND,  MVT::v4f16, Custom);
    }
  }
 
  if (!Subtarget.useSoftFloat() && Subtarget.hasAMXTILE()) {
    addRegisterClass(MVT::x86amx, &X86::TILERegClass);
  }
 
  // We want to custom lower some of our intrinsics.
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
  setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
  if (!Subtarget.is64Bit()) {
    setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
  }
 
  // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
  // handle type legalization for these operations here.
  //
  // FIXME: We really should do custom legalization for addition and
  // subtraction on x86-32 once PR3203 is fixed.  We really can't do much better
  // than generic legalization for 64-bit multiplication-with-overflow, though.
  for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
    if (VT == MVT::i64 && !Subtarget.is64Bit())
      continue;
    // Add/Sub/Mul with overflow operations are custom lowered.
    setOperationAction(ISD::SADDO, VT, Custom);
    setOperationAction(ISD::UADDO, VT, Custom);
    setOperationAction(ISD::SSUBO, VT, Custom);
    setOperationAction(ISD::USUBO, VT, Custom);
    setOperationAction(ISD::SMULO, VT, Custom);
    setOperationAction(ISD::UMULO, VT, Custom);
 
    // Support carry in as value rather than glue.
    setOperationAction(ISD::UADDO_CARRY, VT, Custom);
    setOperationAction(ISD::USUBO_CARRY, VT, Custom);
    setOperationAction(ISD::SETCCCARRY, VT, Custom);
    setOperationAction(ISD::SADDO_CARRY, VT, Custom);
    setOperationAction(ISD::SSUBO_CARRY, VT, Custom);
  }
 
  // Combine sin / cos into _sincos_stret if it is available.
  setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
  setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
 
  if (Subtarget.isTargetWin64()) {
    setOperationAction(ISD::SDIV, MVT::i128, Custom);
    setOperationAction(ISD::UDIV, MVT::i128, Custom);
    setOperationAction(ISD::SREM, MVT::i128, Custom);
    setOperationAction(ISD::UREM, MVT::i128, Custom);
    setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
    setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
    setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
    setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
    setOperationAction(ISD::STRICT_FP_TO_SINT, MVT::i128, Custom);
    setOperationAction(ISD::STRICT_FP_TO_UINT, MVT::i128, Custom);
    setOperationAction(ISD::STRICT_SINT_TO_FP, MVT::i128, Custom);
    setOperationAction(ISD::STRICT_UINT_TO_FP, MVT::i128, Custom);
  }
 
  // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
  // is. We should promote the value to 64-bits to solve this.
  // This is what the CRT headers do - `fmodf` is an inline header
  // function casting to f64 and calling `fmod`.
  if (Subtarget.is32Bit() &&
      (Subtarget.isTargetWindowsMSVC() || Subtarget.isTargetWindowsItanium()))
    // clang-format off
   for (ISD::NodeType Op :
         {ISD::FACOS,  ISD::STRICT_FACOS,
          ISD::FASIN,  ISD::STRICT_FASIN,
          ISD::FATAN,  ISD::STRICT_FATAN,
          ISD::FATAN2, ISD::STRICT_FATAN2,
          ISD::FCEIL,  ISD::STRICT_FCEIL,
          ISD::FCOS,   ISD::STRICT_FCOS,
          ISD::FCOSH,  ISD::STRICT_FCOSH,
          ISD::FEXP,   ISD::STRICT_FEXP,
          ISD::FFLOOR, ISD::STRICT_FFLOOR,
          ISD::FREM,   ISD::STRICT_FREM,
          ISD::FLOG,   ISD::STRICT_FLOG,
          ISD::FLOG10, ISD::STRICT_FLOG10,
          ISD::FPOW,   ISD::STRICT_FPOW,
          ISD::FSIN,   ISD::STRICT_FSIN,
          ISD::FSINH,  ISD::STRICT_FSINH,
          ISD::FTAN,   ISD::STRICT_FTAN,
          ISD::FTANH,  ISD::STRICT_FTANH,
          // TODO: Add ISD:::STRICT_FMODF too once implemented.
          ISD::FMODF})
     if (isOperationExpandOrLibCall(Op, MVT::f32))
        setOperationAction(Op, MVT::f32, Promote);
  // clang-format on
 
  // On MSVC, both 32-bit and 64-bit, ldexpf(f32) is not defined.  MinGW has
  // it, but it's just a wrapper around ldexp.
  if (Subtarget.isOSWindows()) {
    for (ISD::NodeType Op : {ISD::FLDEXP, ISD::STRICT_FLDEXP, ISD::FFREXP})
      if (isOperationExpand(Op, MVT::f32))
        setOperationAction(Op, MVT::f32, Promote);
  }
 
  setOperationPromotedToType(ISD::ATOMIC_LOAD, MVT::f16, MVT::i16);
  setOperationPromotedToType(ISD::ATOMIC_LOAD, MVT::f32, MVT::i32);
  setOperationPromotedToType(ISD::ATOMIC_LOAD, MVT::f64, MVT::i64);
 
  setOperationPromotedToType(ISD::ATOMIC_STORE, MVT::f16, MVT::i16);
  setOperationPromotedToType(ISD::ATOMIC_STORE, MVT::f32, MVT::i32);
  setOperationPromotedToType(ISD::ATOMIC_STORE, MVT::f64, MVT::i64);
 
  // We have target-specific dag combine patterns for the following nodes:
  setTargetDAGCombine({ISD::VECTOR_SHUFFLE,
                       ISD::SCALAR_TO_VECTOR,
                       ISD::INSERT_VECTOR_ELT,
                       ISD::EXTRACT_VECTOR_ELT,
                       ISD::CONCAT_VECTORS,
                       ISD::INSERT_SUBVECTOR,
                       ISD::EXTRACT_SUBVECTOR,
                       ISD::BITCAST,
                       ISD::VSELECT,
                       ISD::SELECT,
                       ISD::SHL,
                       ISD::SRA,
                       ISD::SRL,
                       ISD::OR,
                       ISD::AND,
                       ISD::AVGCEILS,
                       ISD::AVGCEILU,
                       ISD::AVGFLOORS,
                       ISD::AVGFLOORU,
                       ISD::BITREVERSE,
                       ISD::ADD,
                       ISD::SADDSAT,
                       ISD::SSUBSAT,
                       ISD::FADD,
                       ISD::FSUB,
                       ISD::FNEG,
                       ISD::FMA,
                       ISD::STRICT_FMA,
                       ISD::FMINNUM,
                       ISD::FMAXNUM,
                       ISD::SUB,
                       ISD::LOAD,
                       ISD::LRINT,
                       ISD::LLRINT,
                       ISD::MLOAD,
                       ISD::STORE,
                       ISD::MSTORE,
                       ISD::TRUNCATE,
                       ISD::ZERO_EXTEND,
                       ISD::ANY_EXTEND,
                       ISD::SIGN_EXTEND,
                       ISD::SIGN_EXTEND_INREG,
                       ISD::ANY_EXTEND_VECTOR_INREG,
                       ISD::SIGN_EXTEND_VECTOR_INREG,
                       ISD::ZERO_EXTEND_VECTOR_INREG,
                       ISD::SINT_TO_FP,
                       ISD::UINT_TO_FP,
                       ISD::FP_TO_SINT,
                       ISD::STRICT_SINT_TO_FP,
                       ISD::STRICT_UINT_TO_FP,
                       ISD::FP_TO_SINT_SAT,
                       ISD::FP_TO_UINT_SAT,
                       ISD::SETCC,
                       ISD::MUL,
                       ISD::XOR,
                       ISD::MSCATTER,
                       ISD::MGATHER,
                       ISD::FP16_TO_FP,
                       ISD::FP_EXTEND,
                       ISD::STRICT_FP_EXTEND,
                       ISD::FP_ROUND,
                       ISD::STRICT_FP_ROUND,
                       ISD::ROTL,
                       ISD::ROTR,
                       ISD::FSHL,
                       ISD::FSHR,
                       ISD::INTRINSIC_VOID,
                       ISD::INTRINSIC_WO_CHAIN,
                       ISD::INTRINSIC_W_CHAIN});
 
  computeRegisterProperties(Subtarget.getRegisterInfo());
 
  MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
  MaxStoresPerMemsetOptSize = 8;
  MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
  MaxStoresPerMemcpyOptSize = 4;
  MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
  MaxStoresPerMemmoveOptSize = 4;
 
  // TODO: These control memcmp expansion in CGP and could be raised higher, but
  // that needs to benchmarked and balanced with the potential use of vector
  // load/store types (PR33329, PR33914).
  MaxLoadsPerMemcmp = 2;
  MaxLoadsPerMemcmpOptSize = 2;
 
  // Default loop alignment, which can be overridden by -align-loops.
  setPrefLoopAlignment(Align(16));
 
  // An out-of-order CPU can speculatively execute past a predictable branch,
  // but a conditional move could be stalled by an expensive earlier operation.
  PredictableSelectIsExpensive = Subtarget.getSchedModel().isOutOfOrder();
  EnableExtLdPromotion = true;
  setPrefFunctionAlignment(Align(16));
 
  verifyIntrinsicTables();
 
  // Default to having -disable-strictnode-mutation on
  IsStrictFPEnabled = true;
}
 
// This has so far only been implemented for 64-bit MachO.
bool X86TargetLowering::useLoadStackGuardNode(const Module &M) const {
  return Subtarget.isTargetMachO() && Subtarget.is64Bit();
}
 
bool X86TargetLowering::useStackGuardXorFP() const {
  // Currently only MSVC CRTs XOR the frame pointer into the stack guard value.
  return Subtarget.getTargetTriple().isOSMSVCRT() && !Subtarget.isTargetMachO();
}
 
SDValue X86TargetLowering::emitStackGuardXorFP(SelectionDAG &DAG, SDValue Val,
                                               const SDLoc &DL) const {
  EVT PtrTy = getPointerTy(DAG.getDataLayout());
  unsigned XorOp = Subtarget.is64Bit() ? X86::XOR64_FP : X86::XOR32_FP;
  MachineSDNode *Node = DAG.getMachineNode(XorOp, DL, PtrTy, Val);
  return SDValue(Node, 0);
}
 
TargetLoweringBase::LegalizeTypeAction
X86TargetLowering::getPreferredVectorAction(MVT VT) const {
  if ((VT == MVT::v32i1 || VT == MVT::v64i1) && Subtarget.hasAVX512() &&
      !Subtarget.hasBWI())
    return TypeSplitVector;
 
  // Since v8f16 is legal, widen anything over v4f16.
  if (!VT.isScalableVector() && VT.getVectorNumElements() != 1 &&
      VT.getVectorNumElements() <= 4 && !Subtarget.hasF16C() &&
      VT.getVectorElementType() == MVT::f16)
    return TypeSplitVector;
 
  if (!VT.isScalableVector() && VT.getVectorNumElements() != 1 &&
      VT.getVectorElementType() != MVT::i1)
    return TypeWidenVector;
 
  return TargetLoweringBase::getPreferredVectorAction(VT);
}
 
FastISel *X86TargetLowering::createFastISel(
    FunctionLoweringInfo &funcInfo, const TargetLibraryInfo *libInfo,
    const LibcallLoweringInfo *libcallLowering) const {
  return X86::createFastISel(funcInfo, libInfo, libcallLowering);
}
 
//===----------------------------------------------------------------------===//
//                           Other Lowering Hooks
//===----------------------------------------------------------------------===//
 
bool X86::mayFoldLoad(SDValue Op, const X86Subtarget &Subtarget,
                      bool AssumeSingleUse, bool IgnoreAlignment) {
  if (!AssumeSingleUse && !Op.hasOneUse())
    return false;
  if (!ISD::isNormalLoad(Op.getNode()))
    return false;
 
  // If this is an unaligned vector, make sure the target supports folding it.
  auto *Ld = cast<LoadSDNode>(Op.getNode());
  if (!IgnoreAlignment && !Subtarget.hasAVX() &&
      !Subtarget.hasSSEUnalignedMem() && Ld->getValueSizeInBits(0) == 128 &&
      Ld->getAlign() < Align(16))
    return false;
 
  // TODO: If this is a non-temporal load and the target has an instruction
  //       for it, it should not be folded. See "useNonTemporalLoad()".
 
  return true;
}
 
bool X86::mayFoldLoadIntoBroadcastFromMem(SDValue Op, MVT EltVT,
                                          const X86Subtarget &Subtarget,
                                          bool AssumeSingleUse) {
  assert(Subtarget.hasAVX() && "Expected AVX for broadcast from memory");
  if (!X86::mayFoldLoad(Op, Subtarget, AssumeSingleUse))
    return false;
 
  // We can not replace a wide volatile load with a broadcast-from-memory,
  // because that would narrow the load, which isn't legal for volatiles.
  auto *Ld = cast<LoadSDNode>(Op.getNode());
  return !Ld->isVolatile() ||
         Ld->getValueSizeInBits(0) == EltVT.getScalarSizeInBits();
}
 
bool X86::mayFoldIntoStore(SDValue Op) {
  if (!Op.hasOneUse())
    return false;
  // Peek through (oneuse) bitcast users
  SDNode *User = *Op->user_begin();
  while (User->getOpcode() == ISD::BITCAST) {
    if (!User->hasOneUse())
      return false;
    User = *User->user_begin();
  }
  return ISD::isNormalStore(User);
}
 
bool X86::mayFoldIntoZeroExtend(SDValue Op) {
  if (Op.hasOneUse()) {
    unsigned Opcode = Op.getNode()->user_begin()->getOpcode();
    return (ISD::ZERO_EXTEND == Opcode);
  }
  return false;
}
 
// Return true if its cheap to bitcast this to a vector type.
static bool mayFoldIntoVector(SDValue Op, const SelectionDAG &DAG,
                              const X86Subtarget &Subtarget) {
  if (peekThroughBitcasts(Op).getValueType().isVector())
    return true;
  if (isa<ConstantSDNode>(Op) || isa<ConstantFPSDNode>(Op))
    return true;
 
  EVT VT = Op.getValueType();
  unsigned Opcode = Op.getOpcode();
  if ((VT == MVT::i128 || VT == MVT::i256 || VT == MVT::i512) &&
      DAG.getTargetLoweringInfo().getOperationAction(Opcode, VT) ==
          TargetLowering::LegalizeAction::Custom) {
    // Check for larger than legal scalar integer ops that might have been
    // custom lowered to vector instruction.
    switch (Opcode) {
    case ISD::BITREVERSE:
      return true;
    case ISD::SHL:
    case ISD::SRL:
    case ISD::SRA:
      return mayFoldIntoVector(Op.getOperand(0), DAG, Subtarget);
    case ISD::AND:
    case ISD::OR:
    case ISD::XOR:
    case ISD::ADD:
    case ISD::SUB:
    case ISD::FSHL:
    case ISD::FSHR:
      return mayFoldIntoVector(Op.getOperand(0), DAG, Subtarget) &&
             mayFoldIntoVector(Op.getOperand(1), DAG, Subtarget);
    case ISD::SELECT:
      return mayFoldIntoVector(Op.getOperand(1), DAG, Subtarget) &&
             mayFoldIntoVector(Op.getOperand(2), DAG, Subtarget);
    }
  }
  return X86::mayFoldLoad(Op, Subtarget, /*AssumeSingleUse=*/true,
                          /*IgnoreAlignment=*/true);
}
 
static bool isLogicOp(unsigned Opcode) {
  // TODO: Add support for X86ISD::FAND/FOR/FXOR/FANDN with test coverage.
  return ISD::isBitwiseLogicOp(Opcode) || X86ISD::ANDNP == Opcode;
}
 
static bool isTargetShuffle(unsigned Opcode) {
  switch(Opcode) {
  default: return false;
  case X86ISD::BLENDI:
  case X86ISD::PSHUFB:
  case X86ISD::PSHUFD:
  case X86ISD::PSHUFHW:
  case X86ISD::PSHUFLW:
  case X86ISD::SHUFP:
  case X86ISD::INSERTPS:
  case X86ISD::EXTRQI:
  case X86ISD::INSERTQI:
  case X86ISD::VALIGN:
  case X86ISD::PALIGNR:
  case X86ISD::VSHLDQ:
  case X86ISD::VSRLDQ:
  case X86ISD::MOVLHPS:
  case X86ISD::MOVHLPS:
  case X86ISD::MOVSHDUP:
  case X86ISD::MOVSLDUP:
  case X86ISD::MOVDDUP:
  case X86ISD::MOVSS:
  case X86ISD::MOVSD:
  case X86ISD::MOVSH:
  case X86ISD::UNPCKL:
  case X86ISD::UNPCKH:
  case X86ISD::VBROADCAST:
  case X86ISD::VPERMILPI:
  case X86ISD::VPERMILPV:
  case X86ISD::VPERM2X128:
  case X86ISD::SHUF128:
  case X86ISD::VPERMIL2:
  case X86ISD::VPERMI:
  case X86ISD::VPPERM:
  case X86ISD::VPERMV:
  case X86ISD::VPERMV3:
  case X86ISD::VZEXT_MOVL:
  case X86ISD::COMPRESS:
  case X86ISD::EXPAND:
    return true;
  }
}
 
static bool isTargetShuffleVariableMask(unsigned Opcode) {
  switch (Opcode) {
  default: return false;
  // Target Shuffles.
  case X86ISD::PSHUFB:
  case X86ISD::VPERMILPV:
  case X86ISD::VPERMIL2:
  case X86ISD::VPPERM:
  case X86ISD::VPERMV:
  case X86ISD::VPERMV3:
    return true;
  // 'Faux' Target Shuffles.
  case ISD::OR:
  case ISD::AND:
  case X86ISD::ANDNP:
    return true;
  }
}
 
SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  const X86RegisterInfo *RegInfo = Subtarget.getRegisterInfo();
  X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
  int ReturnAddrIndex = FuncInfo->getRAIndex();
 
  if (ReturnAddrIndex == 0) {
    // Set up a frame object for the return address.
    unsigned SlotSize = RegInfo->getSlotSize();
    ReturnAddrIndex = MF.getFrameInfo().CreateFixedObject(SlotSize,
                                                          -(int64_t)SlotSize,
                                                          false);
    FuncInfo->setRAIndex(ReturnAddrIndex);
  }
 
  return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
}
 
bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model CM,
                                       bool HasSymbolicDisplacement) {
  // Offset should fit into 32 bit immediate field.
  if (!isInt<32>(Offset))
    return false;
 
  // If we don't have a symbolic displacement - we don't have any extra
  // restrictions.
  if (!HasSymbolicDisplacement)
    return true;
 
  // We can fold large offsets in the large code model because we always use
  // 64-bit offsets.
  if (CM == CodeModel::Large)
    return true;
 
  // For kernel code model we know that all object resist in the negative half
  // of 32bits address space. We may not accept negative offsets, since they may
  // be just off and we may accept pretty large positive ones.
  if (CM == CodeModel::Kernel)
    return Offset >= 0;
 
  // For other non-large code models we assume that latest small object is 16MB
  // before end of 31 bits boundary. We may also accept pretty large negative
  // constants knowing that all objects are in the positive half of address
  // space.
  return Offset < 16 * 1024 * 1024;
}
 
/// Return true if the condition is an signed comparison operation.
static bool isX86CCSigned(X86::CondCode X86CC) {
  switch (X86CC) {
  default:
    llvm_unreachable("Invalid integer condition!");
  case X86::COND_E:
  case X86::COND_NE:
  case X86::COND_B:
  case X86::COND_A:
  case X86::COND_BE:
  case X86::COND_AE:
    return false;
  case X86::COND_G:
  case X86::COND_GE:
  case X86::COND_L:
  case X86::COND_LE:
    return true;
  }
}
 
static X86::CondCode TranslateIntegerX86CC(ISD::CondCode SetCCOpcode) {
  switch (SetCCOpcode) {
  // clang-format off
  default: llvm_unreachable("Invalid integer condition!");
  case ISD::SETEQ:  return X86::COND_E;
  case ISD::SETGT:  return X86::COND_G;
  case ISD::SETGE:  return X86::COND_GE;
  case ISD::SETLT:  return X86::COND_L;
  case ISD::SETLE:  return X86::COND_LE;
  case ISD::SETNE:  return X86::COND_NE;
  case ISD::SETULT: return X86::COND_B;
  case ISD::SETUGT: return X86::COND_A;
  case ISD::SETULE: return X86::COND_BE;
  case ISD::SETUGE: return X86::COND_AE;
  // clang-format on
  }
}
 
/// Do a one-to-one translation of a ISD::CondCode to the X86-specific
/// condition code, returning the condition code and the LHS/RHS of the
/// comparison to make.
static X86::CondCode TranslateX86CC(ISD::CondCode SetCCOpcode, const SDLoc &DL,
                                    bool isFP, SDValue &LHS, SDValue &RHS,
                                    SelectionDAG &DAG) {
  if (!isFP) {
    if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
      if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnes()) {
        // X > -1   -> X == 0, jump !sign.
        RHS = DAG.getConstant(0, DL, RHS.getValueType());
        return X86::COND_NS;
      }
      if (SetCCOpcode == ISD::SETLT && RHSC->isZero()) {
        // X < 0   -> X == 0, jump on sign.
        return X86::COND_S;
      }
      if (SetCCOpcode == ISD::SETGE && RHSC->isZero()) {
        // X >= 0   -> X == 0, jump on !sign.
        return X86::COND_NS;
      }
      if (SetCCOpcode == ISD::SETLT && RHSC->isOne()) {
        // X < 1   -> X <= 0
        RHS = DAG.getConstant(0, DL, RHS.getValueType());
        return X86::COND_LE;
      }
    }
 
    return TranslateIntegerX86CC(SetCCOpcode);
  }
 
  // First determine if it is required or is profitable to flip the operands.
 
  // If LHS is a foldable load, but RHS is not, flip the condition.
  if (ISD::isNON_EXTLoad(LHS.getNode()) &&
      !ISD::isNON_EXTLoad(RHS.getNode())) {
    SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
    std::swap(LHS, RHS);
  }
 
  switch (SetCCOpcode) {
  default: break;
  case ISD::SETOLT:
  case ISD::SETOLE:
  case ISD::SETUGT:
  case ISD::SETUGE:
    std::swap(LHS, RHS);
    break;
  }
 
  // On a floating point condition, the flags are set as follows:
  // ZF  PF  CF   op
  //  0 | 0 | 0 | X > Y
  //  0 | 0 | 1 | X < Y
  //  1 | 0 | 0 | X == Y
  //  1 | 1 | 1 | unordered
  switch (SetCCOpcode) {
  // clang-format off
  default: llvm_unreachable("Condcode should be pre-legalized away");
  case ISD::SETUEQ:
  case ISD::SETEQ:   return X86::COND_E;
  case ISD::SETOLT:              // flipped
  case ISD::SETOGT:
  case ISD::SETGT:   return X86::COND_A;
  case ISD::SETOLE:              // flipped
  case ISD::SETOGE:
  case ISD::SETGE:   return X86::COND_AE;
  case ISD::SETUGT:              // flipped
  case ISD::SETULT:
  case ISD::SETLT:   return X86::COND_B;
  case ISD::SETUGE:              // flipped
  case ISD::SETULE:
  case ISD::SETLE:   return X86::COND_BE;
  case ISD::SETONE:
  case ISD::SETNE:   return X86::COND_NE;
  case ISD::SETUO:   return X86::COND_P;
  case ISD::SETO:    return X86::COND_NP;
  case ISD::SETOEQ:
  case ISD::SETUNE:  return X86::COND_INVALID;
  // clang-format on
  }
}
 
/// Is there a floating point cmov for the specific X86 condition code?
/// Current x86 isa includes the following FP cmov instructions:
/// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
static bool hasFPCMov(unsigned X86CC) {
  switch (X86CC) {
  default:
    return false;
  case X86::COND_B:
  case X86::COND_BE:
  case X86::COND_E:
  case X86::COND_P:
  case X86::COND_A:
  case X86::COND_AE:
  case X86::COND_NE:
  case X86::COND_NP:
    return true;
  }
}
 
static bool useVPTERNLOG(const X86Subtarget &Subtarget, MVT VT) {
  return Subtarget.hasVLX() || Subtarget.canExtendTo512DQ() ||
         VT.is512BitVector();
}
 
void X86TargetLowering::getTgtMemIntrinsic(
    SmallVectorImpl<IntrinsicInfo> &Infos, const CallBase &I,
    MachineFunction &MF, unsigned Intrinsic) const {
  IntrinsicInfo Info;
  Info.flags = MachineMemOperand::MONone;
  Info.offset = 0;
 
  const IntrinsicData* IntrData = getIntrinsicWithChain(Intrinsic);
  if (!IntrData) {
    switch (Intrinsic) {
    case Intrinsic::x86_aesenc128kl:
    case Intrinsic::x86_aesdec128kl:
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.ptrVal = I.getArgOperand(1);
      Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 48);
      Info.align = Align(1);
      Info.flags |= MachineMemOperand::MOLoad;
      Infos.push_back(Info);
      return;
    case Intrinsic::x86_aesenc256kl:
    case Intrinsic::x86_aesdec256kl:
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.ptrVal = I.getArgOperand(1);
      Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 64);
      Info.align = Align(1);
      Info.flags |= MachineMemOperand::MOLoad;
      Infos.push_back(Info);
      return;
    case Intrinsic::x86_aesencwide128kl:
    case Intrinsic::x86_aesdecwide128kl:
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.ptrVal = I.getArgOperand(0);
      Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 48);
      Info.align = Align(1);
      Info.flags |= MachineMemOperand::MOLoad;
      Infos.push_back(Info);
      return;
    case Intrinsic::x86_aesencwide256kl:
    case Intrinsic::x86_aesdecwide256kl:
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.ptrVal = I.getArgOperand(0);
      Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), 64);
      Info.align = Align(1);
      Info.flags |= MachineMemOperand::MOLoad;
      Infos.push_back(Info);
      return;
    case Intrinsic::x86_cmpccxadd32:
    case Intrinsic::x86_cmpccxadd64:
    case Intrinsic::x86_atomic_bts:
    case Intrinsic::x86_atomic_btc:
    case Intrinsic::x86_atomic_btr: {
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.ptrVal = I.getArgOperand(0);
      unsigned Size = I.getType()->getScalarSizeInBits();
      Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), Size);
      Info.align = Align(Size);
      Info.flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
                    MachineMemOperand::MOVolatile;
      Infos.push_back(Info);
      return;
    }
    case Intrinsic::x86_atomic_bts_rm:
    case Intrinsic::x86_atomic_btc_rm:
    case Intrinsic::x86_atomic_btr_rm: {
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.ptrVal = I.getArgOperand(0);
      unsigned Size = I.getArgOperand(1)->getType()->getScalarSizeInBits();
      Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), Size);
      Info.align = Align(Size);
      Info.flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
                    MachineMemOperand::MOVolatile;
      Infos.push_back(Info);
      return;
    }
    case Intrinsic::x86_aadd32:
    case Intrinsic::x86_aadd64:
    case Intrinsic::x86_aand32:
    case Intrinsic::x86_aand64:
    case Intrinsic::x86_aor32:
    case Intrinsic::x86_aor64:
    case Intrinsic::x86_axor32:
    case Intrinsic::x86_axor64:
    case Intrinsic::x86_atomic_add_cc:
    case Intrinsic::x86_atomic_sub_cc:
    case Intrinsic::x86_atomic_or_cc:
    case Intrinsic::x86_atomic_and_cc:
    case Intrinsic::x86_atomic_xor_cc: {
      Info.opc = ISD::INTRINSIC_W_CHAIN;
      Info.ptrVal = I.getArgOperand(0);
      unsigned Size = I.getArgOperand(1)->getType()->getScalarSizeInBits();
      Info.memVT = EVT::getIntegerVT(I.getType()->getContext(), Size);
      Info.align = Align(Size);
      Info.flags |= MachineMemOperand::MOLoad | MachineMemOperand::MOStore |
                    MachineMemOperand::MOVolatile;
      Infos.push_back(Info);
      return;
    }
    }
    return;
  }
 
  switch (IntrData->Type) {
  case TRUNCATE_TO_MEM_VI8:
  case TRUNCATE_TO_MEM_VI16:
  case TRUNCATE_TO_MEM_VI32: {
    Info.opc = ISD::INTRINSIC_VOID;
    Info.ptrVal = I.getArgOperand(0);
    MVT VT  = MVT::getVT(I.getArgOperand(1)->getType());
    MVT ScalarVT = MVT::INVALID_SIMPLE_VALUE_TYPE;
    if (IntrData->Type == TRUNCATE_TO_MEM_VI8)
      ScalarVT = MVT::i8;
    else if (IntrData->Type == TRUNCATE_TO_MEM_VI16)
      ScalarVT = MVT::i16;
    else if (IntrData->Type == TRUNCATE_TO_MEM_VI32)
      ScalarVT = MVT::i32;
 
    Info.memVT = MVT::getVectorVT(ScalarVT, VT.getVectorNumElements());
    Info.align = Align(1);
    Info.flags |= MachineMemOperand::MOStore;
    Infos.push_back(Info);
    return;
  }
  case GATHER:
  case GATHER_AVX2: {
    Info.opc = ISD::INTRINSIC_W_CHAIN;
    Info.ptrVal = nullptr;
    MVT DataVT = MVT::getVT(I.getType());
    MVT IndexVT = MVT::getVT(I.getArgOperand(2)->getType());
    unsigned NumElts = std::min(DataVT.getVectorNumElements(),
                                IndexVT.getVectorNumElements());
    Info.memVT = MVT::getVectorVT(DataVT.getVectorElementType(), NumElts);
    Info.align = Align(1);
    Info.flags |= MachineMemOperand::MOLoad;
    Infos.push_back(Info);
    return;
  }
  case SCATTER: {
    Info.opc = ISD::INTRINSIC_VOID;
    Info.ptrVal = nullptr;
    MVT DataVT = MVT::getVT(I.getArgOperand(3)->getType());
    MVT IndexVT = MVT::getVT(I.getArgOperand(2)->getType());
    unsigned NumElts = std::min(DataVT.getVectorNumElements(),
                                IndexVT.getVectorNumElements());
    Info.memVT = MVT::getVectorVT(DataVT.getVectorElementType(), NumElts);
    Info.align = Align(1);
    Info.flags |= MachineMemOperand::MOStore;
    Infos.push_back(Info);
    return;
  }
  default:
    return;
  }
}
 
/// Returns true if the target can instruction select the
/// specified FP immediate natively. If false, the legalizer will
/// materialize the FP immediate as a load from a constant pool.
bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT,
                                     bool ForCodeSize) const {
  for (const APFloat &FPImm : LegalFPImmediates)
    if (Imm.bitwiseIsEqual(FPImm))
      return true;
  return false;
}
 
bool X86TargetLowering::shouldReduceLoadWidth(
    SDNode *Load, ISD::LoadExtType ExtTy, EVT NewVT,
    std::optional<unsigned> ByteOffset) const {
  assert(cast<LoadSDNode>(Load)->isSimple() && "illegal to narrow");
 
  auto PeekThroughOneUserBitcasts = [](const SDNode *N) {
    while (N->getOpcode() == ISD::BITCAST && N->hasOneUse())
      N = *N->user_begin();
    return N;
  };
 
  // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
  // relocation target a movq or addq instruction: don't let the load shrink.
  SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
  if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
    if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
      return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
 
  // If this is an (1) AVX vector load with (2) multiple uses and (3) all of
  // those uses are extracted directly into a store, then the extract + store
  // can be store-folded, or (4) any use will be used by legal full width
  // instruction. Then, it's probably not worth splitting the load.
  EVT VT = Load->getValueType(0);
  if ((VT.is256BitVector() || VT.is512BitVector()) &&
      !SDValue(Load, 0).hasOneUse()) {
    bool FullWidthUse = false;
    bool AllExtractStores = true;
    for (SDUse &Use : Load->uses()) {
      // Skip uses of the chain value. Result 0 of the node is the load value.
      if (Use.getResNo() != 0)
        continue;
 
      const SDNode *User = PeekThroughOneUserBitcasts(Use.getUser());
 
      // If this use is an extract + store, it's probably not worth splitting.
      if (User->getOpcode() == ISD::EXTRACT_SUBVECTOR &&
          all_of(User->uses(), [&](const SDUse &U) {
            const SDNode *Inner = PeekThroughOneUserBitcasts(U.getUser());
            return Inner->getOpcode() == ISD::STORE;
          }))
        continue;
 
      AllExtractStores = false;
 
      // If any use is a full width legal/target bin op, then assume its legal
      // and won't split.
      if (isBinOp(User->getOpcode()) &&
          (isOperationLegal(User->getOpcode(), User->getValueType(0)) ||
           User->getOpcode() > ISD::BUILTIN_OP_END))
        FullWidthUse = true;
    }
 
    if (AllExtractStores)
      return false;
 
    // If we have an user that uses the full vector width, then this use is
    // only worth splitting if the offset isn't 0 (to avoid an
    // EXTRACT_SUBVECTOR) or we're loading a scalar integer.
    if (FullWidthUse)
      return (ByteOffset.value_or(0) > 0) || NewVT.isScalarInteger();
  }
 
  return true;
}
 
/// Returns true if it is beneficial to convert a load of a constant
/// to just the constant itself.
bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
                                                          Type *Ty) const {
  assert(Ty->isIntegerTy());
 
  unsigned BitSize = Ty->getPrimitiveSizeInBits();
  if (BitSize == 0 || BitSize > 64)
    return false;
  return true;
}
 
bool X86TargetLowering::reduceSelectOfFPConstantLoads(EVT CmpOpVT) const {
  // If we are using XMM registers in the ABI and the condition of the select is
  // a floating-point compare and we have blendv or conditional move, then it is
  // cheaper to select instead of doing a cross-register move and creating a
  // load that depends on the compare result.
  bool IsFPSetCC = CmpOpVT.isFloatingPoint() && CmpOpVT != MVT::f128;
  return !IsFPSetCC || !Subtarget.isTarget64BitLP64() || !Subtarget.hasAVX();
}
 
bool X86TargetLowering::convertSelectOfConstantsToMath(EVT VT) const {
  // TODO: It might be a win to ease or lift this restriction, but the generic
  // folds in DAGCombiner conflict with vector folds for an AVX512 target.
  if (VT.isVector() && Subtarget.hasAVX512())
    return false;
 
  return true;
}
 
bool X86TargetLowering::decomposeMulByConstant(LLVMContext &Context, EVT VT,
                                               SDValue C) const {
  // TODO: We handle scalars using custom code, but generic combining could make
  // that unnecessary.
  APInt MulC;
  if (!ISD::isConstantSplatVector(C.getNode(), MulC))
    return false;
 
  if (VT.isVector() && VT.getScalarSizeInBits() == 8) {
    // Check whether a vXi8 multiply can be decomposed into two shifts
    // (decomposing 2^m ± 2^n as 2^(a+b) ± 2^b). Similar to
    // DAGCombiner::visitMUL, consider the constant `2` decomposable as
    // (2^0 + 1).
    APInt ShiftedMulC = MulC.abs();
    unsigned TZeros = ShiftedMulC == 2 ? 0 : ShiftedMulC.countr_zero();
    ShiftedMulC.lshrInPlace(TZeros);
    if ((ShiftedMulC - 1).isPowerOf2() || (ShiftedMulC + 1).isPowerOf2())
      return true;
  }
 
  // Find the type this will be legalized too. Otherwise we might prematurely
  // convert this to shl+add/sub and then still have to type legalize those ops.
  // Another choice would be to defer the decision for illegal types until
  // after type legalization. But constant splat vectors of i64 can't make it
  // through type legalization on 32-bit targets so we would need to special
  // case vXi64.
  while (getTypeAction(Context, VT) != TypeLegal)
    VT = getTypeToTransformTo(Context, VT);
 
  // If vector multiply is legal, assume that's faster than shl + add/sub.
  // Multiply is a complex op with higher latency and lower throughput in
  // most implementations, sub-vXi32 vector multiplies are always fast,
  // vXi32 mustn't have a SlowMULLD implementation, and anything larger (vXi64)
  // is always going to be slow.
  unsigned EltSizeInBits = VT.getScalarSizeInBits();
  if (isOperationLegal(ISD::MUL, VT) && EltSizeInBits <= 32 &&
      (EltSizeInBits != 32 || !Subtarget.isPMULLDSlow()))
    return false;
 
  // shl+add, shl+sub, shl+add+neg
  return (MulC + 1).isPowerOf2() || (MulC - 1).isPowerOf2() ||
         (1 - MulC).isPowerOf2() || (-(MulC + 1)).isPowerOf2();
}
 
bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT, EVT SrcVT,
                                                unsigned Index) const {
  if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
    return false;
 
  // Mask vectors support all subregister combinations and operations that
  // extract half of vector.
  if (ResVT.getVectorElementType() == MVT::i1)
    return Index == 0 ||
           ((ResVT.getSizeInBits() * 2 == SrcVT.getSizeInBits()) &&
            (Index == ResVT.getVectorNumElements()));
 
  return (Index % ResVT.getVectorNumElements()) == 0;
}
 
bool X86TargetLowering::shouldScalarizeBinop(SDValue VecOp) const {
  unsigned Opc = VecOp.getOpcode();
 
  // Assume target opcodes can't be scalarized.
  // TODO - do we have any exceptions?
  if (Opc >= ISD::BUILTIN_OP_END || !isBinOp(Opc))
    return false;
 
  // If the vector op is not supported, try to convert to scalar.
  EVT VecVT = VecOp.getValueType();
  if (!isOperationLegalOrCustomOrPromote(Opc, VecVT))
    return true;
 
  // If the vector op is supported, but the scalar op is not, the transform may
  // not be worthwhile.
  EVT ScalarVT = VecVT.getScalarType();
  return isOperationLegalOrCustomOrPromote(Opc, ScalarVT);
}
 
bool X86TargetLowering::shouldFormOverflowOp(unsigned Opcode, EVT VT,
                                             bool) const {
  // TODO: Allow vectors?
  if (VT.isVector())
    return false;
  return VT.isSimple() || !isOperationExpand(Opcode, VT);
}
 
bool X86TargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
  // Speculate cttz only if we can directly use TZCNT/CMOV, can promote to
  // i32/i64 or can rely on BSF passthrough value.
  return Subtarget.hasBMI() || Subtarget.canUseCMOV() ||
         Subtarget.hasBitScanPassThrough() ||
         (!Ty->isVectorTy() &&
          Ty->getScalarSizeInBits() < (Subtarget.is64Bit() ? 64u : 32u));
}
 
bool X86TargetLowering::isCheapToSpeculateCtlz(Type *Ty) const {
  // Speculate ctlz only if we can directly use LZCNT/CMOV, or can rely on BSR
  // passthrough value.
  return Subtarget.hasLZCNT() || Subtarget.canUseCMOV() ||
         Subtarget.hasBitScanPassThrough();
}
 
bool X86TargetLowering::ShouldShrinkFPConstant(EVT VT) const {
  // Don't shrink FP constpool if SSE2 is available since cvtss2sd is more
  // expensive than a straight movsd. On the other hand, it's important to
  // shrink long double fp constant since fldt is very slow.
  return !Subtarget.hasSSE2() || VT == MVT::f80;
}
 
bool X86TargetLowering::isScalarFPTypeInSSEReg(EVT VT) const {
  return (VT == MVT::f64 && Subtarget.hasSSE2()) ||
         (VT == MVT::f32 && Subtarget.hasSSE1()) || VT == MVT::f16;
}
 
bool X86TargetLowering::isLoadBitCastBeneficial(EVT LoadVT, EVT BitcastVT,
                                                const SelectionDAG &DAG,
                                                const MachineMemOperand &MMO) const {
  if (!Subtarget.hasAVX512() && !LoadVT.isVector() && BitcastVT.isVector() &&
      BitcastVT.getVectorElementType() == MVT::i1)
    return false;
 
  if (!Subtarget.hasDQI() && BitcastVT == MVT::v8i1 && LoadVT == MVT::i8)
    return false;
 
  if (LoadVT.isVector() && BitcastVT.isVector()) {
    // If both types are legal vectors, it's always ok to convert them.
    // Don't convert to an illegal type.
    if (isTypeLegal(LoadVT))
      return isTypeLegal(BitcastVT);
  }
 
  // If we have a large vector type (even if illegal), don't bitcast to large
  // (illegal) scalar types. Better to load fewer vectors and extract.
  if (LoadVT.isVector() && !BitcastVT.isVector() && LoadVT.isInteger() &&
      BitcastVT.isInteger() && (LoadVT.getSizeInBits() % 128) == 0)
    return false;
 
  return TargetLowering::isLoadBitCastBeneficial(LoadVT, BitcastVT, DAG, MMO);
}
 
bool X86TargetLowering::canMergeStoresTo(unsigned AddressSpace, EVT MemVT,
                                         const MachineFunction &MF) const {
  // Do not merge to float value size (128 bytes) if no implicit
  // float attribute is set.
  bool NoFloat = MF.getFunction().hasFnAttribute(Attribute::NoImplicitFloat);
 
  if (NoFloat) {
    unsigned MaxIntSize = Subtarget.is64Bit() ? 64 : 32;
    return (MemVT.getSizeInBits() <= MaxIntSize);
  }
  // Make sure we don't merge greater than our preferred vector
  // width.
  if (MemVT.getSizeInBits() > Subtarget.getPreferVectorWidth())
    return false;
 
  return true;
}
 
bool X86TargetLowering::isCtlzFast() const {
  return Subtarget.hasFastLZCNT();
}
 
bool X86TargetLowering::isMaskAndCmp0FoldingBeneficial(
    const Instruction &AndI) const {
  return true;
}
 
bool X86TargetLowering::hasAndNotCompare(SDValue Y) const {
  // Scalar integer and-not compares are efficiently handled by NOT+TEST (or
  // BMI ANDN).
  return Y.getValueType().isScalarInteger();
}
 
bool X86TargetLowering::hasAndNot(SDValue Y) const {
  EVT VT = Y.getValueType();
 
  if (!VT.isVector()) {
    if (!Subtarget.hasBMI())
      return false;
 
    // There are only 32-bit and 64-bit forms for 'andn'.
    if (VT != MVT::i32 && VT != MVT::i64)
      return false;
    return !isa<ConstantSDNode>(Y) || cast<ConstantSDNode>(Y)->isOpaque();
  }
 
  // Vector.
  if (!Subtarget.hasSSE1() || VT.getSizeInBits() < 128)
    return false;
 
  if (VT == MVT::v4i32)
    return true;
 
  return Subtarget.hasSSE2();
}
 
bool X86TargetLowering::hasBitTest(SDValue X, SDValue Y) const {
  return X.getValueType().isScalarInteger(); // 'bt'
}
 
bool X86TargetLowering::
    shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
        SDValue X, ConstantSDNode *XC, ConstantSDNode *CC, SDValue Y,
        unsigned OldShiftOpcode, unsigned NewShiftOpcode,
        SelectionDAG &DAG) const {
  // Does baseline recommend not to perform the fold by default?
  if (!TargetLowering::shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd(
          X, XC, CC, Y, OldShiftOpcode, NewShiftOpcode, DAG))
    return false;
  // For scalars this transform is always beneficial.
  if (X.getValueType().isScalarInteger())
    return true;
  // If all the shift amounts are identical, then transform is beneficial even
  // with rudimentary SSE2 shifts.
  if (DAG.isSplatValue(Y, /*AllowUndefs=*/true))
    return true;
  // If we have AVX2 with it's powerful shift operations, then it's also good.
  if (Subtarget.hasAVX2())
    return true;
  // Pre-AVX2 vector codegen for this pattern is best for variant with 'shl'.
  return NewShiftOpcode == ISD::SHL;
}
 
unsigned X86TargetLowering::preferedOpcodeForCmpEqPiecesOfOperand(
    EVT VT, unsigned ShiftOpc, bool MayTransformRotate,
    const APInt &ShiftOrRotateAmt, const std::optional<APInt> &AndMask) const {
  if (!VT.isInteger())
    return ShiftOpc;
 
  bool PreferRotate = false;
  if (VT.isVector()) {
    // For vectors, if we have rotate instruction support, then its definetly
    // best. Otherwise its not clear what the best so just don't make changed.
    PreferRotate = Subtarget.hasAVX512() && (VT.getScalarType() == MVT::i32 ||
                                             VT.getScalarType() == MVT::i64);
  } else {
    // For scalar, if we have bmi prefer rotate for rorx. Otherwise prefer
    // rotate unless we have a zext mask+shr.
    PreferRotate = Subtarget.hasBMI2();
    if (!PreferRotate) {
      unsigned MaskBits =
          VT.getScalarSizeInBits() - ShiftOrRotateAmt.getZExtValue();
      PreferRotate = (MaskBits != 8) && (MaskBits != 16) && (MaskBits != 32);
    }
  }
 
  if (ShiftOpc == ISD::SHL || ShiftOpc == ISD::SRL) {
    assert(AndMask.has_value() && "Null andmask when querying about shift+and");
 
    if (PreferRotate && MayTransformRotate)
      return ISD::ROTL;
 
    // If vector we don't really get much benefit swapping around constants.
    // Maybe we could check if the DAG has the flipped node already in the
    // future.
    if (VT.isVector())
      return ShiftOpc;
 
    // See if the beneficial to swap shift type.
    if (ShiftOpc == ISD::SHL) {
      // If the current setup has imm64 mask, then inverse will have
      // at least imm32 mask (or be zext i32 -> i64).
      if (VT == MVT::i64)
        return AndMask->getSignificantBits() > 32 ? (unsigned)ISD::SRL
                                                  : ShiftOpc;
 
      // We can only benefit if req at least 7-bit for the mask. We
      // don't want to replace shl of 1,2,3 as they can be implemented
      // with lea/add.
      return ShiftOrRotateAmt.uge(7) ? (unsigned)ISD::SRL : ShiftOpc;
    }
 
    if (VT == MVT::i64)
      // Keep exactly 32-bit imm64, this is zext i32 -> i64 which is
      // extremely efficient.
      return AndMask->getSignificantBits() > 33 ? (unsigned)ISD::SHL : ShiftOpc;
 
    // Keep small shifts as shl so we can generate add/lea.
    return ShiftOrRotateAmt.ult(7) ? (unsigned)ISD::SHL : ShiftOpc;
  }
 
  // We prefer rotate for vectors of if we won't get a zext mask with SRL
  // (PreferRotate will be set in the latter case).
  if (PreferRotate || !MayTransformRotate || VT.isVector())
    return ShiftOpc;
 
  // Non-vector type and we have a zext mask with SRL.
  return ISD::SRL;
}
 
TargetLoweringBase::CondMergingParams
X86TargetLowering::getJumpConditionMergingParams(Instruction::BinaryOps Opc,
                                                 const Value *Lhs,
                                                 const Value *Rhs) const {
  using namespace llvm::PatternMatch;
  int BaseCost = BrMergingBaseCostThresh.getValue();
  // With CCMP, branches can be merged in a more efficient way.
  if (BaseCost >= 0 && Subtarget.hasCCMP())
    BaseCost += BrMergingCcmpBias;
  // a == b && a == c is a fast pattern on x86.
  if (BaseCost >= 0 && Opc == Instruction::And &&
      match(Lhs, m_SpecificICmp(ICmpInst::ICMP_EQ, m_Value(), m_Value())) &&
      match(Rhs, m_SpecificICmp(ICmpInst::ICMP_EQ, m_Value(), m_Value())))
    BaseCost += 1;
 
  // For OR conditions with EQ comparisons, prefer splitting into branches
  // (unless CCMP is available). OR+EQ cannot be optimized via bitwise ops,
  // unlike OR+NE which becomes (P|Q)!=0. Similarly, don't split signed
  // comparisons (SLT, SGT) that can be optimized.
  if (BaseCost >= 0 && !Subtarget.hasCCMP() && Opc == Instruction::Or &&
      match(Lhs, m_SpecificICmp(ICmpInst::ICMP_EQ, m_Value(), m_Value())) &&
      match(Rhs, m_SpecificICmp(ICmpInst::ICMP_EQ, m_Value(), m_Value())))
    return {-1, -1, -1};
 
  return {BaseCost, BrMergingLikelyBias.getValue(),
          BrMergingUnlikelyBias.getValue()};
}
 
bool X86TargetLowering::preferScalarizeSplat(SDNode *N) const {
  return N->getOpcode() != ISD::FP_EXTEND;
}
 
bool X86TargetLowering::shouldFoldConstantShiftPairToMask(
    const SDNode *N) const {
  assert(((N->getOpcode() == ISD::SHL &&
           N->getOperand(0).getOpcode() == ISD::SRL) ||
          (N->getOpcode() == ISD::SRL &&
           N->getOperand(0).getOpcode() == ISD::SHL)) &&
         "Expected shift-shift mask");
  // TODO: Should we always create i64 masks? Or only folded immediates?
  EVT VT = N->getValueType(0);
  if ((Subtarget.hasFastVectorShiftMasks() && VT.isVector()) ||
      (Subtarget.hasFastScalarShiftMasks() && !VT.isVector())) {
    // Only fold if the shift values are equal - so it folds to AND.
    // TODO - we should fold if either is a non-uniform vector but we don't do
    // the fold for non-splats yet.
    return N->getOperand(1) == N->getOperand(0).getOperand(1);
  }
  return TargetLoweringBase::shouldFoldConstantShiftPairToMask(N);
}
 
bool X86TargetLowering::shouldFoldMaskToVariableShiftPair(SDValue Y) const {
  EVT VT = Y.getValueType();
 
  // For vectors, we don't have a preference, but we probably want a mask.
  if (VT.isVector())
    return false;
 
  unsigned MaxWidth = Subtarget.is64Bit() ? 64 : 32;
  return VT.getScalarSizeInBits() <= MaxWidth;
}
 
TargetLowering::ShiftLegalizationStrategy
X86TargetLowering::preferredShiftLegalizationStrategy(
    SelectionDAG &DAG, SDNode *N, unsigned ExpansionFactor) const {
  if (DAG.getMachineFunction().getFunction().hasMinSize() &&
      !Subtarget.isOSWindows())
    return ShiftLegalizationStrategy::LowerToLibcall;
  return TargetLowering::preferredShiftLegalizationStrategy(DAG, N,
                                                            ExpansionFactor);
}
 
bool X86TargetLowering::shouldSplatInsEltVarIndex(EVT VT) const {
  // Any legal vector type can be splatted more efficiently than
  // loading/spilling from memory.
  return isTypeLegal(VT);
}
 
MVT X86TargetLowering::hasFastEqualityCompare(unsigned NumBits) const {
  MVT VT = MVT::getIntegerVT(NumBits);
  if (isTypeLegal(VT))
    return VT;
 
  // PMOVMSKB can handle this.
  if (NumBits == 128 && isTypeLegal(MVT::v16i8))
    return MVT::v16i8;
 
  // VPMOVMSKB can handle this.
  if (NumBits == 256 && isTypeLegal(MVT::v32i8))
    return MVT::v32i8;
 
  // TODO: Allow 64-bit type for 32-bit target.
  // TODO: 512-bit types should be allowed, but make sure that those
  // cases are handled in combineVectorSizedSetCCEquality().
 
  return MVT::INVALID_SIMPLE_VALUE_TYPE;
}
 
/// Val is the undef sentinel value or equal to the specified value.
static bool isUndefOrEqual(int Val, int CmpVal) {
  return ((Val == SM_SentinelUndef) || (Val == CmpVal));
}
 
/// Return true if every element in Mask is the undef sentinel value or equal to
/// the specified value.
static bool isUndefOrEqual(ArrayRef<int> Mask, int CmpVal) {
  return llvm::all_of(Mask, [CmpVal](int M) {
    return (M == SM_SentinelUndef) || (M == CmpVal);
  });
}
 
/// Return true if every element in Mask, beginning from position Pos and ending
/// in Pos+Size is the undef sentinel value or equal to the specified value.
static bool isUndefOrEqualInRange(ArrayRef<int> Mask, int CmpVal, unsigned Pos,
                                  unsigned Size) {
  return llvm::all_of(Mask.slice(Pos, Size),
                      [CmpVal](int M) { return isUndefOrEqual(M, CmpVal); });
}
 
/// Val is either the undef or zero sentinel value.
static bool isUndefOrZero(int Val) {
  return ((Val == SM_SentinelUndef) || (Val == SM_SentinelZero));
}
 
/// Return true if every element in Mask, beginning from position Pos and ending
/// in Pos+Size is the undef sentinel value.
static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
  return llvm::all_of(Mask.slice(Pos, Size), equal_to(SM_SentinelUndef));
}
 
/// Return true if the mask creates a vector whose lower half is undefined.
static bool isUndefLowerHalf(ArrayRef<int> Mask) {
  unsigned NumElts = Mask.size();
  return isUndefInRange(Mask, 0, NumElts / 2);
}
 
/// Return true if the mask creates a vector whose upper half is undefined.
static bool isUndefUpperHalf(ArrayRef<int> Mask) {
  unsigned NumElts = Mask.size();
  return isUndefInRange(Mask, NumElts / 2, NumElts / 2);
}
 
/// Return true if Val falls within the specified range (L, H].
static bool isInRange(int Val, int Low, int Hi) {
  return (Val >= Low && Val < Hi);
}
 
/// Return true if the value of any element in Mask falls within the specified
/// range (L, H].
static bool isAnyInRange(ArrayRef<int> Mask, int Low, int Hi) {
  return llvm::any_of(Mask, [Low, Hi](int M) { return isInRange(M, Low, Hi); });
}
 
/// Return true if the value of any element in Mask is the zero sentinel value.
static bool isAnyZero(ArrayRef<int> Mask) {
  return llvm::any_of(Mask, equal_to(SM_SentinelZero));
}
 
/// Return true if Val is undef or if its value falls within the
/// specified range (L, H].
static bool isUndefOrInRange(int Val, int Low, int Hi) {
  return (Val == SM_SentinelUndef) || isInRange(Val, Low, Hi);
}
 
/// Return true if every element in Mask is undef or if its value
/// falls within the specified range (L, H].
static bool isUndefOrInRange(ArrayRef<int> Mask, int Low, int Hi) {
  return llvm::all_of(
      Mask, [Low, Hi](int M) { return isUndefOrInRange(M, Low, Hi); });
}
 
/// Return true if Val is undef, zero or if its value falls within the
/// specified range (L, H].
static bool isUndefOrZeroOrInRange(int Val, int Low, int Hi) {
  return isUndefOrZero(Val) || isInRange(Val, Low, Hi);
}
 
/// Return true if every element in Mask is undef, zero or if its value
/// falls within the specified range (L, H].
static bool isUndefOrZeroOrInRange(ArrayRef<int> Mask, int Low, int Hi) {
  return llvm::all_of(
      Mask, [Low, Hi](int M) { return isUndefOrZeroOrInRange(M, Low, Hi); });
}
 
/// Return true if every element in Mask, is an in-place blend/select mask or is
/// undef.
[[maybe_unused]] static bool isBlendOrUndef(ArrayRef<int> Mask) {
  unsigned NumElts = Mask.size();
  for (auto [I, M] : enumerate(Mask))
    if (!isUndefOrEqual(M, I) && !isUndefOrEqual(M, I + NumElts))
      return false;
  return true;
}
 
/// Return true if every element in Mask, beginning
/// from position Pos and ending in Pos + Size, falls within the specified
/// sequence (Low, Low + Step, ..., Low + (Size - 1) * Step) or is undef.
static bool isSequentialOrUndefInRange(ArrayRef<int> Mask, unsigned Pos,
                                       unsigned Size, int Low, int Step = 1) {
  for (unsigned i = Pos, e = Pos + Size; i != e; ++i, Low += Step)
    if (!isUndefOrEqual(Mask[i], Low))
      return false;
  return true;
}
 
/// Return true if every element in Mask, beginning
/// from position Pos and ending in Pos+Size, falls within the specified
/// sequential range (Low, Low+Size], or is undef or is zero.
static bool isSequentialOrUndefOrZeroInRange(ArrayRef<int> Mask, unsigned Pos,
                                             unsigned Size, int Low,
                                             int Step = 1) {
  for (unsigned i = Pos, e = Pos + Size; i != e; ++i, Low += Step)
    if (!isUndefOrZero(Mask[i]) && Mask[i] != Low)
      return false;
  return true;
}
 
/// Return true if every element in Mask, beginning
/// from position Pos and ending in Pos+Size is undef or is zero.
static bool isUndefOrZeroInRange(ArrayRef<int> Mask, unsigned Pos,
                                 unsigned Size) {
  return llvm::all_of(Mask.slice(Pos, Size), isUndefOrZero);
}
 
/// Return true if every element of a single input is referenced by the shuffle
/// mask. i.e. it just permutes them all.
static bool isCompletePermute(ArrayRef<int> Mask) {
  unsigned NumElts = Mask.size();
  APInt DemandedElts = APInt::getZero(NumElts);
  for (int M : Mask)
    if (isInRange(M, 0, NumElts))
      DemandedElts.setBit(M);
  return DemandedElts.isAllOnes();
}
 
/// Helper function to test whether a shuffle mask could be
/// simplified by widening the elements being shuffled.
///
/// Appends the mask for wider elements in WidenedMask if valid. Otherwise
/// leaves it in an unspecified state.
///
/// NOTE: This must handle normal vector shuffle masks and *target* vector
/// shuffle masks. The latter have the special property of a '-2' representing
/// a zero-ed lane of a vector.
static bool canWidenShuffleElements(ArrayRef<int> Mask,
                                    SmallVectorImpl<int> &WidenedMask) {
  WidenedMask.assign(Mask.size() / 2, 0);
  for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
    int M0 = Mask[i];
    int M1 = Mask[i + 1];
 
    // If both elements are undef, its trivial.
    if (M0 == SM_SentinelUndef && M1 == SM_SentinelUndef) {
      WidenedMask[i / 2] = SM_SentinelUndef;
      continue;
    }
 
    // Check for an undef mask and a mask value properly aligned to fit with
    // a pair of values. If we find such a case, use the non-undef mask's value.
    if (M0 == SM_SentinelUndef && M1 >= 0 && (M1 % 2) == 1) {
      WidenedMask[i / 2] = M1 / 2;
      continue;
    }
    if (M1 == SM_SentinelUndef && M0 >= 0 && (M0 % 2) == 0) {
      WidenedMask[i / 2] = M0 / 2;
      continue;
    }
 
    // When zeroing, we need to spread the zeroing across both lanes to widen.
    if (M0 == SM_SentinelZero || M1 == SM_SentinelZero) {
      if ((M0 == SM_SentinelZero || M0 == SM_SentinelUndef) &&
          (M1 == SM_SentinelZero || M1 == SM_SentinelUndef)) {
        WidenedMask[i / 2] = SM_SentinelZero;
        continue;
      }
      return false;
    }
 
    // Finally check if the two mask values are adjacent and aligned with
    // a pair.
    if (M0 != SM_SentinelUndef && (M0 % 2) == 0 && (M0 + 1) == M1) {
      WidenedMask[i / 2] = M0 / 2;
      continue;
    }
 
    // Otherwise we can't safely widen the elements used in this shuffle.
    return false;
  }
  assert(WidenedMask.size() == Mask.size() / 2 &&
         "Incorrect size of mask after widening the elements!");
 
  return true;
}
 
static bool canWidenShuffleElements(ArrayRef<int> Mask,
                                    const APInt &Zeroable,
                                    bool V2IsZero,
                                    SmallVectorImpl<int> &WidenedMask) {
  // Create an alternative mask with info about zeroable elements.
  // Here we do not set undef elements as zeroable.
  SmallVector<int, 64> ZeroableMask(Mask);
  if (V2IsZero) {
    assert(!Zeroable.isZero() && "V2's non-undef elements are used?!");
    for (int i = 0, Size = Mask.size(); i != Size; ++i)
      if (Mask[i] != SM_SentinelUndef && Zeroable[i])
        ZeroableMask[i] = SM_SentinelZero;
  }
  return canWidenShuffleElements(ZeroableMask, WidenedMask);
}
 
static bool canWidenShuffleElements(ArrayRef<int> Mask) {
  SmallVector<int, 32> WidenedMask;
  return canWidenShuffleElements(Mask, WidenedMask);
}
 
// Attempt to narrow/widen shuffle mask until it matches the target number of
// elements.
static bool scaleShuffleElements(ArrayRef<int> Mask, unsigned NumDstElts,
                                 SmallVectorImpl<int> &ScaledMask) {
  unsigned NumSrcElts = Mask.size();
  assert(((NumSrcElts % NumDstElts) == 0 || (NumDstElts % NumSrcElts) == 0) &&
         "Illegal shuffle scale factor");
 
  // Narrowing is guaranteed to work.
  if (NumDstElts >= NumSrcElts) {
    int Scale = NumDstElts / NumSrcElts;
    llvm::narrowShuffleMaskElts(Scale, Mask, ScaledMask);
    return true;
  }
 
  // We have to repeat the widening until we reach the target size, but we can
  // split out the first widening as it sets up ScaledMask for us.
  if (canWidenShuffleElements(Mask, ScaledMask)) {
    while (ScaledMask.size() > NumDstElts) {
      SmallVector<int, 16> WidenedMask;
      if (!canWidenShuffleElements(ScaledMask, WidenedMask))
        return false;
      ScaledMask = std::move(WidenedMask);
    }
    return true;
  }
 
  return false;
}
 
static bool canScaleShuffleElements(ArrayRef<int> Mask, unsigned NumDstElts) {
  SmallVector<int, 32> ScaledMask;
  return scaleShuffleElements(Mask, NumDstElts, ScaledMask);
}
 
// Helper to grow the shuffle mask for a larger value type.
// NOTE: This is different to scaleShuffleElements which is a same size type.
static void growShuffleMask(ArrayRef<int> SrcMask,
                            SmallVectorImpl<int> &DstMask,
                            unsigned SrcSizeInBits, unsigned DstSizeInBits) {
  assert(DstMask.empty() && "Expected an empty shuffle mas");
  assert((DstSizeInBits % SrcSizeInBits) == 0 && "Illegal shuffle scale");
  unsigned Scale = DstSizeInBits / SrcSizeInBits;
  unsigned NumSrcElts = SrcMask.size();
  DstMask.assign(SrcMask.begin(), SrcMask.end());
  for (int &M : DstMask) {
    if (M < 0)
      continue;
    M = (M % NumSrcElts) + ((M / NumSrcElts) * Scale * NumSrcElts);
  }
  DstMask.append((Scale - 1) * NumSrcElts, SM_SentinelUndef);
}
 
/// Returns true if Elt is a constant zero or a floating point constant +0.0.
bool X86::isZeroNode(SDValue Elt) {
  return isNullConstant(Elt) || isNullFPConstant(Elt);
}
 
// Build a vector of constants.
// Use an UNDEF node if MaskElt == -1.
// Split 64-bit constants in the 32-bit mode.
static SDValue getConstVector(ArrayRef<int> Values, MVT VT, SelectionDAG &DAG,
                              const SDLoc &dl, bool IsMask = false) {
 
  SmallVector<SDValue, 32>  Ops;
  bool Split = false;
 
  MVT ConstVecVT = VT;
  unsigned NumElts = VT.getVectorNumElements();
  bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
  if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
    ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
    Split = true;
  }
 
  MVT EltVT = ConstVecVT.getVectorElementType();
  for (unsigned i = 0; i < NumElts; ++i) {
    bool IsUndef = Values[i] < 0 && IsMask;
    SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
      DAG.getConstant(Values[i], dl, EltVT);
    Ops.push_back(OpNode);
    if (Split)
      Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
                    DAG.getConstant(0, dl, EltVT));
  }
  SDValue ConstsNode = DAG.getBuildVector(ConstVecVT, dl, Ops);
  if (Split)
    ConstsNode = DAG.getBitcast(VT, ConstsNode);
  return ConstsNode;
}
 
static SDValue getConstVector(ArrayRef<APInt> Bits, const APInt &Undefs,
                              MVT VT, SelectionDAG &DAG, const SDLoc &dl) {
  assert(Bits.size() == Undefs.getBitWidth() &&
         "Unequal constant and undef arrays");
  SmallVector<SDValue, 32> Ops;
  bool Split = false;
 
  MVT ConstVecVT = VT;
  unsigned NumElts = VT.getVectorNumElements();
  bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
  if (!In64BitMode && VT.getVectorElementType() == MVT::i64) {
    ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
    Split = true;
  }
 
  MVT EltVT = ConstVecVT.getVectorElementType();
  MVT EltIntVT = EltVT.changeTypeToInteger();
  for (unsigned i = 0, e = Bits.size(); i != e; ++i) {
    if (Undefs[i]) {
      Ops.append(Split ? 2 : 1, DAG.getUNDEF(EltVT));
      continue;
    }
    const APInt &V = Bits[i];
    assert(V.getBitWidth() == VT.getScalarSizeInBits() && "Unexpected sizes");
    if (Split) {
      Ops.push_back(DAG.getConstant(V.extractBits(32, 0), dl, EltVT));
      Ops.push_back(DAG.getConstant(V.extractBits(32, 32), dl, EltVT));
    } else {
      Ops.push_back(DAG.getBitcast(EltVT, DAG.getConstant(V, dl, EltIntVT)));
    }
  }
 
  SDValue ConstsNode = DAG.getBuildVector(ConstVecVT, dl, Ops);
  return DAG.getBitcast(VT, ConstsNode);
}
 
static SDValue getConstVector(ArrayRef<APInt> Bits, MVT VT,
                              SelectionDAG &DAG, const SDLoc &dl) {
  APInt Undefs = APInt::getZero(Bits.size());
  return getConstVector(Bits, Undefs, VT, DAG, dl);
}
 
/// Returns a vector of specified type with all zero elements.
static SDValue getZeroVector(MVT VT, const X86Subtarget &Subtarget,
                             SelectionDAG &DAG, const SDLoc &dl) {
  assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector() ||
          VT.getVectorElementType() == MVT::i1) &&
         "Unexpected vector type");
 
  // Try to build SSE/AVX zero vectors as <N x i32> bitcasted to their dest
  // type. This ensures they get CSE'd. But if the integer type is not
  // available, use a floating-point +0.0 instead.
  SDValue Vec;
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  if (!Subtarget.hasSSE2() && VT.is128BitVector()) {
    Vec = DAG.getConstantFP(+0.0, dl, MVT::v4f32);
  } else if (VT.isFloatingPoint() &&
             TLI.isTypeLegal(VT.getVectorElementType())) {
    Vec = DAG.getConstantFP(+0.0, dl, VT);
  } else if (VT.getVectorElementType() == MVT::i1) {
    assert((Subtarget.hasBWI() || VT.getVectorNumElements() <= 16) &&
           "Unexpected vector type");
    Vec = DAG.getConstant(0, dl, VT);
  } else {
    unsigned Num32BitElts = VT.getSizeInBits() / 32;
    Vec = DAG.getConstant(0, dl, MVT::getVectorVT(MVT::i32, Num32BitElts));
  }
  return DAG.getBitcast(VT, Vec);
}
 
// Helper to determine if the ops are all the extracted subvectors come from a
// single source. If we allow commute they don't have to be in order (Lo/Hi).
static SDValue getSplitVectorSrc(SDValue LHS, SDValue RHS, bool AllowCommute) {
  if (LHS.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
      RHS.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
      LHS.getValueType() != RHS.getValueType() ||
      LHS.getOperand(0) != RHS.getOperand(0))
    return SDValue();
 
  SDValue Src = LHS.getOperand(0);
  if (Src.getValueSizeInBits() != (LHS.getValueSizeInBits() * 2))
    return SDValue();
 
  unsigned NumElts = LHS.getValueType().getVectorNumElements();
  if ((LHS.getConstantOperandAPInt(1) == 0 &&
       RHS.getConstantOperandAPInt(1) == NumElts) ||
      (AllowCommute && RHS.getConstantOperandAPInt(1) == 0 &&
       LHS.getConstantOperandAPInt(1) == NumElts))
    return Src;
 
  return SDValue();
}
 
static SDValue extractSubVector(SDValue Vec, unsigned IdxVal, SelectionDAG &DAG,
                                const SDLoc &dl, unsigned vectorWidth) {
  EVT VT = Vec.getValueType();
  EVT ElVT = VT.getVectorElementType();
  unsigned ResultNumElts =
      (VT.getVectorNumElements() * vectorWidth) / VT.getSizeInBits();
  EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT, ResultNumElts);
 
  assert(ResultVT.getSizeInBits() == vectorWidth &&
         "Illegal subvector extraction");
 
  // Extract the relevant vectorWidth bits.  Generate an EXTRACT_SUBVECTOR
  unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
  assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
 
  // This is the index of the first element of the vectorWidth-bit chunk
  // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
  IdxVal &= ~(ElemsPerChunk - 1);
 
  // If the input is a buildvector just emit a smaller one.
  if (Vec.getOpcode() == ISD::BUILD_VECTOR)
    return DAG.getBuildVector(ResultVT, dl,
                              Vec->ops().slice(IdxVal, ElemsPerChunk));
 
  // Check if we're extracting the upper undef of a widening pattern.
  if (Vec.getOpcode() == ISD::INSERT_SUBVECTOR && Vec.getOperand(0).isUndef() &&
      Vec.getOperand(1).getValueType().getVectorNumElements() <= IdxVal &&
      isNullConstant(Vec.getOperand(2)))
    return DAG.getUNDEF(ResultVT);
 
  return DAG.getExtractSubvector(dl, ResultVT, Vec, IdxVal);
}
 
/// Generate a DAG to grab 128-bits from a vector > 128 bits.  This
/// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
/// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
/// instructions or a simple subregister reference. Idx is an index in the
/// 128 bits we want.  It need not be aligned to a 128-bit boundary.  That makes
/// lowering EXTRACT_VECTOR_ELT operations easier.
static SDValue extract128BitVector(SDValue Vec, unsigned IdxVal,
                                   SelectionDAG &DAG, const SDLoc &dl) {
  assert((Vec.getValueType().is256BitVector() ||
          Vec.getValueType().is512BitVector()) &&
         "Unexpected vector size!");
  return extractSubVector(Vec, IdxVal, DAG, dl, 128);
}
 
/// Generate a DAG to grab 256-bits from a 512-bit vector.
static SDValue extract256BitVector(SDValue Vec, unsigned IdxVal,
                                   SelectionDAG &DAG, const SDLoc &dl) {
  assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
  return extractSubVector(Vec, IdxVal, DAG, dl, 256);
}
 
static SDValue insertSubVector(SDValue Result, SDValue Vec, unsigned IdxVal,
                               SelectionDAG &DAG, const SDLoc &dl,
                               unsigned vectorWidth) {
  assert((vectorWidth == 128 || vectorWidth == 256) &&
         "Unsupported vector width");
  // Inserting UNDEF is Result
  if (Vec.isUndef())
    return Result;
 
  // Insert the relevant vectorWidth bits.
  EVT VT = Vec.getValueType();
  unsigned ElemsPerChunk = vectorWidth / VT.getScalarSizeInBits();
  assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
 
  // This is the index of the first element of the vectorWidth-bit chunk
  // we want. Since ElemsPerChunk is a power of 2 just need to clear bits.
  IdxVal &= ~(ElemsPerChunk - 1);
  return DAG.getInsertSubvector(dl, Result, Vec, IdxVal);
}
 
/// Generate a DAG to put 128-bits into a vector > 128 bits.  This
/// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
/// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
/// simple superregister reference.  Idx is an index in the 128 bits
/// we want.  It need not be aligned to a 128-bit boundary.  That makes
/// lowering INSERT_VECTOR_ELT operations easier.
static SDValue insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
                                  SelectionDAG &DAG, const SDLoc &dl) {
  assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
  return insertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
}
 
/// Widen a vector to a larger size with the same scalar type, with the new
/// elements either zero or undef.
static SDValue widenSubVector(MVT VT, SDValue Vec, bool ZeroNewElements,
                              const X86Subtarget &Subtarget, SelectionDAG &DAG,
                              const SDLoc &dl) {
  EVT VecVT = Vec.getValueType();
  assert(VecVT.getFixedSizeInBits() <= VT.getFixedSizeInBits() &&
         VecVT.getScalarType() == VT.getScalarType() &&
         "Unsupported vector widening type");
  // If the upper 128-bits of a build vector are already undef/zero, then try to
  // widen from the lower 128-bits.
  if (Vec.getOpcode() == ISD::BUILD_VECTOR && VecVT.is256BitVector()) {
    unsigned NumSrcElts = VecVT.getVectorNumElements();
    ArrayRef<SDUse> Hi = Vec->ops().drop_front(NumSrcElts / 2);
    if (all_of(Hi, [&](SDValue V) {
          return V.isUndef() || (ZeroNewElements && X86::isZeroNode(V));
        }))
      Vec = extract128BitVector(Vec, 0, DAG, dl);
  }
  SDValue Res = ZeroNewElements ? getZeroVector(VT, Subtarget, DAG, dl)
                                : DAG.getUNDEF(VT);
  return DAG.getInsertSubvector(dl, Res, Vec, 0);
}
 
/// Widen a vector to a larger size with the same scalar type, with the new
/// elements either zero or undef.
static SDValue widenSubVector(SDValue Vec, bool ZeroNewElements,
                              const X86Subtarget &Subtarget, SelectionDAG &DAG,
                              const SDLoc &dl, unsigned WideSizeInBits) {
  assert(Vec.getValueSizeInBits() <= WideSizeInBits &&
         (WideSizeInBits % Vec.getScalarValueSizeInBits()) == 0 &&
         "Unsupported vector widening type");
  unsigned WideNumElts = WideSizeInBits / Vec.getScalarValueSizeInBits();
  MVT SVT = Vec.getSimpleValueType().getScalarType();
  MVT VT = MVT::getVectorVT(SVT, WideNumElts);
  return widenSubVector(VT, Vec, ZeroNewElements, Subtarget, DAG, dl);
}
 
/// Widen a mask vector type to a minimum of v8i1/v16i1 to allow use of KSHIFT
/// and bitcast with integer types.
static MVT widenMaskVectorType(MVT VT, const X86Subtarget &Subtarget) {
  assert(VT.getVectorElementType() == MVT::i1 && "Expected bool vector");
  unsigned NumElts = VT.getVectorNumElements();
  if ((!Subtarget.hasDQI() && NumElts == 8) || NumElts < 8)
    return Subtarget.hasDQI() ? MVT::v8i1 : MVT::v16i1;
  return VT;
}
 
/// Widen a mask vector to a minimum of v8i1/v16i1 to allow use of KSHIFT and
/// bitcast with integer types.
static SDValue widenMaskVector(SDValue Vec, bool ZeroNewElements,
                               const X86Subtarget &Subtarget, SelectionDAG &DAG,
                               const SDLoc &dl) {
  MVT VT = widenMaskVectorType(Vec.getSimpleValueType(), Subtarget);
  return widenSubVector(VT, Vec, ZeroNewElements, Subtarget, DAG, dl);
}
 
// Helper function to collect subvector ops that are concatenated together,
// either by ISD::CONCAT_VECTORS or a ISD::INSERT_SUBVECTOR series.
// The subvectors in Ops are guaranteed to be the same type.
static bool collectConcatOps(SDNode *N, SmallVectorImpl<SDValue> &Ops,
                             SelectionDAG &DAG) {
  assert(Ops.empty() && "Expected an empty ops vector");
 
  if (N->getOpcode() == ISD::CONCAT_VECTORS) {
    Ops.append(N->op_begin(), N->op_end());
    return true;
  }
 
  if (N->getOpcode() == ISD::INSERT_SUBVECTOR) {
    SDValue Src = N->getOperand(0);
    SDValue Sub = N->getOperand(1);
    const APInt &Idx = N->getConstantOperandAPInt(2);
    EVT VT = Src.getValueType();
    EVT SubVT = Sub.getValueType();
 
    if (VT.getSizeInBits() == (SubVT.getSizeInBits() * 2)) {
      // insert_subvector(undef, x, lo)
      if (Idx == 0 && Src.isUndef()) {
        Ops.push_back(Sub);
        Ops.push_back(DAG.getUNDEF(SubVT));
        return true;
      }
      if (Idx == (VT.getVectorNumElements() / 2)) {
        // insert_subvector(insert_subvector(undef, x, lo), y, hi)
        if (Src.getOpcode() == ISD::INSERT_SUBVECTOR &&
            Src.getOperand(1).getValueType() == SubVT &&
            isNullConstant(Src.getOperand(2))) {
          // Attempt to recurse into inner (matching) concats.
          SDValue Lo = Src.getOperand(1);
          SDValue Hi = Sub;
          SmallVector<SDValue, 2> LoOps, HiOps;
          if (collectConcatOps(Lo.getNode(), LoOps, DAG) &&
              collectConcatOps(Hi.getNode(), HiOps, DAG) &&
              LoOps.size() == HiOps.size()) {
            Ops.append(LoOps);
            Ops.append(HiOps);
            return true;
          }
          Ops.push_back(Lo);
          Ops.push_back(Hi);
          return true;
        }
        // insert_subvector(x, extract_subvector(x, lo), hi)
        if (Sub.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
            Sub.getOperand(0) == Src && isNullConstant(Sub.getOperand(1))) {
          Ops.append(2, Sub);
          return true;
        }
        // insert_subvector(undef, x, hi)
        if (Src.isUndef()) {
          Ops.push_back(DAG.getUNDEF(SubVT));
          Ops.push_back(Sub);
          return true;
        }
      }
    }
  }
 
  if (N->getOpcode() == ISD::EXTRACT_SUBVECTOR) {
    EVT VT = N->getValueType(0);
    SDValue Src = N->getOperand(0);
    uint64_t Idx = N->getConstantOperandVal(1);
 
    // Collect all the subvectors from the source vector and slice off the
    // extraction.
    SmallVector<SDValue, 4> SrcOps;
    if (collectConcatOps(Src.getNode(), SrcOps, DAG) &&
        VT.getSizeInBits() > SrcOps[0].getValueSizeInBits() &&
        (VT.getSizeInBits() % SrcOps[0].getValueSizeInBits()) == 0 &&
        (Idx % SrcOps[0].getValueType().getVectorNumElements()) == 0) {
      unsigned SubIdx = Idx / SrcOps[0].getValueType().getVectorNumElements();
      unsigned NumSubs = VT.getSizeInBits() / SrcOps[0].getValueSizeInBits();
      Ops.append(SrcOps.begin() + SubIdx, SrcOps.begin() + SubIdx + NumSubs);
      return true;
    }
  }
 
  assert(Ops.empty() && "Expected an empty ops vector");
  return false;
}
 
// Helper to check if \p V can be split into subvectors and the upper subvectors
// are all undef. In which case return the lower subvector.
static SDValue isUpperSubvectorUndef(SDValue V, const SDLoc &DL,
                                     SelectionDAG &DAG) {
  SmallVector<SDValue> SubOps;
  if (!collectConcatOps(V.getNode(), SubOps, DAG))
    return SDValue();
 
  unsigned NumSubOps = SubOps.size();
  unsigned HalfNumSubOps = NumSubOps / 2;
  assert((NumSubOps % 2) == 0 && "Unexpected number of subvectors");
 
  ArrayRef<SDValue> UpperOps(SubOps.begin() + HalfNumSubOps, SubOps.end());
  if (any_of(UpperOps, [](SDValue Op) { return !Op.isUndef(); }))
    return SDValue();
 
  EVT HalfVT = V.getValueType().getHalfNumVectorElementsVT(*DAG.getContext());
  ArrayRef<SDValue> LowerOps(SubOps.begin(), SubOps.begin() + HalfNumSubOps);
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, HalfVT, LowerOps);
}
 
// Helper to check if we can access all the constituent subvectors without any
// extract ops.
static bool isFreeToSplitVector(SDValue V, SelectionDAG &DAG) {
  SmallVector<SDValue> Ops;
  return collectConcatOps(V.getNode(), Ops, DAG);
}
 
static std::pair<SDValue, SDValue> splitVector(SDValue Op, SelectionDAG &DAG,
                                               const SDLoc &dl) {
  EVT VT = Op.getValueType();
  unsigned NumElems = VT.getVectorNumElements();
  unsigned SizeInBits = VT.getSizeInBits();
  assert((NumElems % 2) == 0 && (SizeInBits % 2) == 0 &&
         "Can't split odd sized vector");
 
  SmallVector<SDValue, 4> SubOps;
  if (collectConcatOps(Op.getNode(), SubOps, DAG)) {
    assert((SubOps.size() % 2) == 0 && "Can't split odd sized vector concat");
    unsigned HalfOps = SubOps.size() / 2;
    EVT HalfVT = VT.getHalfNumVectorElementsVT(*DAG.getContext());
    SmallVector<SDValue, 2> LoOps(SubOps.begin(), SubOps.begin() + HalfOps);
    SmallVector<SDValue, 2> HiOps(SubOps.begin() + HalfOps, SubOps.end());
    SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, LoOps);
    SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, HiOps);
    return std::make_pair(Lo, Hi);
  }
 
  // If this is a splat value (with no-undefs) then use the lower subvector,
  // which should be a free extraction.
  SDValue Lo = extractSubVector(Op, 0, DAG, dl, SizeInBits / 2);
  if (DAG.isSplatValue(Op, /*AllowUndefs*/ false))
    return std::make_pair(Lo, Lo);
 
  SDValue Hi = extractSubVector(Op, NumElems / 2, DAG, dl, SizeInBits / 2);
  return std::make_pair(Lo, Hi);
}
 
/// Break an operation into 2 half sized ops and then concatenate the results.
static SDValue splitVectorOp(SDValue Op, SelectionDAG &DAG, const SDLoc &dl) {
  unsigned NumOps = Op.getNumOperands();
  EVT VT = Op.getValueType();
 
  // Extract the LHS Lo/Hi vectors
  SmallVector<SDValue> LoOps(NumOps, SDValue());
  SmallVector<SDValue> HiOps(NumOps, SDValue());
  for (unsigned I = 0; I != NumOps; ++I) {
    SDValue SrcOp = Op.getOperand(I);
    if (!SrcOp.getValueType().isVector()) {
      LoOps[I] = HiOps[I] = SrcOp;
      continue;
    }
    std::tie(LoOps[I], HiOps[I]) = splitVector(SrcOp, DAG, dl);
  }
 
  EVT LoVT, HiVT;
  std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(VT);
  return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
                     DAG.getNode(Op.getOpcode(), dl, LoVT, LoOps),
                     DAG.getNode(Op.getOpcode(), dl, HiVT, HiOps));
}
 
/// Break an unary integer operation into 2 half sized ops and then
/// concatenate the result back.
static SDValue splitVectorIntUnary(SDValue Op, SelectionDAG &DAG,
                                   const SDLoc &dl) {
  // Make sure we only try to split 256/512-bit types to avoid creating
  // narrow vectors.
  [[maybe_unused]] EVT VT = Op.getValueType();
  assert((Op.getOperand(0).getValueType().is256BitVector() ||
          Op.getOperand(0).getValueType().is512BitVector()) &&
         (VT.is256BitVector() || VT.is512BitVector()) && "Unsupported VT!");
  assert(Op.getOperand(0).getValueType().getVectorNumElements() ==
             VT.getVectorNumElements() &&
         "Unexpected VTs!");
  return splitVectorOp(Op, DAG, dl);
}
 
/// Break a binary integer operation into 2 half sized ops and then
/// concatenate the result back.
static SDValue splitVectorIntBinary(SDValue Op, SelectionDAG &DAG,
                                    const SDLoc &dl) {
  // Assert that all the types match.
  [[maybe_unused]] EVT VT = Op.getValueType();
  assert(Op.getOperand(0).getValueType() == VT &&
         Op.getOperand(1).getValueType() == VT && "Unexpected VTs!");
  assert((VT.is256BitVector() || VT.is512BitVector()) && "Unsupported VT!");
  return splitVectorOp(Op, DAG, dl);
}
 
// Helper for splitting operands of an operation to legal target size and
// apply a function on each part.
// Useful for operations that are available on SSE2 in 128-bit, on AVX2 in
// 256-bit and on AVX512BW in 512-bit. The argument VT is the type used for
// deciding if/how to split Ops. Ops elements do *not* have to be of type VT.
// The argument Builder is a function that will be applied on each split part:
// SDValue Builder(SelectionDAG&G, SDLoc, ArrayRef<SDValue>)
template <typename F>
SDValue SplitOpsAndApply(SelectionDAG &DAG, const X86Subtarget &Subtarget,
                         const SDLoc &DL, EVT VT, ArrayRef<SDValue> Ops,
                         F Builder, bool CheckBWI = true,
                         bool AllowAVX512 = true) {
  assert(Subtarget.hasSSE2() && "Target assumed to support at least SSE2");
  unsigned NumSubs = 1;
  if (AllowAVX512 && ((CheckBWI && Subtarget.useBWIRegs()) ||
                      (!CheckBWI && Subtarget.useAVX512Regs()))) {
    if (VT.getSizeInBits() > 512) {
      NumSubs = VT.getSizeInBits() / 512;
      assert((VT.getSizeInBits() % 512) == 0 && "Illegal vector size");
    }
  } else if (Subtarget.hasAVX2()) {
    if (VT.getSizeInBits() > 256) {
      NumSubs = VT.getSizeInBits() / 256;
      assert((VT.getSizeInBits() % 256) == 0 && "Illegal vector size");
    }
  } else {
    if (VT.getSizeInBits() > 128) {
      NumSubs = VT.getSizeInBits() / 128;
      assert((VT.getSizeInBits() % 128) == 0 && "Illegal vector size");
    }
  }
 
  if (NumSubs == 1)
    return Builder(DAG, DL, Ops);
 
  SmallVector<SDValue, 4> Subs;
  for (unsigned i = 0; i != NumSubs; ++i) {
    SmallVector<SDValue, 2> SubOps;
    for (SDValue Op : Ops) {
      EVT OpVT = Op.getValueType();
      unsigned NumSubElts = OpVT.getVectorNumElements() / NumSubs;
      unsigned SizeSub = OpVT.getSizeInBits() / NumSubs;
      SubOps.push_back(extractSubVector(Op, i * NumSubElts, DAG, DL, SizeSub));
    }
    Subs.push_back(Builder(DAG, DL, SubOps));
  }
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Subs);
}
 
// Helper function that extends a non-512-bit vector op to 512-bits on non-VLX
// targets.
static SDValue getAVX512Node(unsigned Opcode, const SDLoc &DL, MVT VT,
                             ArrayRef<SDValue> Ops, SelectionDAG &DAG,
                             const X86Subtarget &Subtarget) {
  assert(Subtarget.hasAVX512() && "AVX512 target expected");
  MVT SVT = VT.getScalarType();
 
  // If we have a 32/64 splatted constant, splat it to DstTy to
  // encourage a foldable broadcast'd operand.
  auto MakeBroadcastOp = [&](SDValue Op, MVT OpVT, MVT DstVT) {
    unsigned OpEltSizeInBits = OpVT.getScalarSizeInBits();
    // AVX512 broadcasts 32/64-bit operands.
    // TODO: Support float once getAVX512Node is used by fp-ops.
    if (!OpVT.isInteger() || OpEltSizeInBits < 32 ||
        !DAG.getTargetLoweringInfo().isTypeLegal(SVT))
      return SDValue();
    // If we're not widening, don't bother if we're not bitcasting.
    if (OpVT == DstVT && Op.getOpcode() != ISD::BITCAST)
      return SDValue();
    if (auto *BV = dyn_cast<BuildVectorSDNode>(peekThroughBitcasts(Op))) {
      APInt SplatValue, SplatUndef;
      unsigned SplatBitSize;
      bool HasAnyUndefs;
      if (BV->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
                              HasAnyUndefs, OpEltSizeInBits) &&
          !HasAnyUndefs && SplatValue.getBitWidth() == OpEltSizeInBits)
        return DAG.getConstant(SplatValue, DL, DstVT);
    }
    return SDValue();
  };
 
  bool Widen = !(Subtarget.hasVLX() || VT.is512BitVector());
 
  MVT DstVT = VT;
  if (Widen)
    DstVT = MVT::getVectorVT(SVT, 512 / SVT.getSizeInBits());
 
  // Canonicalize src operands.
  SmallVector<SDValue> SrcOps(Ops);
  for (SDValue &Op : SrcOps) {
    MVT OpVT = Op.getSimpleValueType();
    // Just pass through scalar operands.
    if (!OpVT.isVector())
      continue;
    assert(OpVT == VT && "Vector type mismatch");
 
    if (SDValue BroadcastOp = MakeBroadcastOp(Op, OpVT, DstVT)) {
      Op = BroadcastOp;
      continue;
    }
 
    // Just widen the subvector by inserting into an undef wide vector.
    if (Widen)
      Op = widenSubVector(Op, false, Subtarget, DAG, DL, 512);
  }
 
  SDValue Res = DAG.getNode(Opcode, DL, DstVT, SrcOps);
 
  // Perform the 512-bit op then extract the bottom subvector.
  if (Widen)
    Res = extractSubVector(Res, 0, DAG, DL, VT.getSizeInBits());
  return Res;
}
 
/// Insert i1-subvector to i1-vector.
static SDValue insert1BitVector(SDValue Op, SelectionDAG &DAG,
                                const X86Subtarget &Subtarget) {
 
  SDLoc dl(Op);
  SDValue Vec = Op.getOperand(0);
  SDValue SubVec = Op.getOperand(1);
  SDValue Idx = Op.getOperand(2);
  unsigned IdxVal = Op.getConstantOperandVal(2);
 
  // Inserting undef is a nop. We can just return the original vector.
  if (SubVec.isUndef())
    return Vec;
 
  if (IdxVal == 0 && Vec.isUndef()) // the operation is legal
    return Op;
 
  MVT OpVT = Op.getSimpleValueType();
  unsigned NumElems = OpVT.getVectorNumElements();
  SDValue ZeroIdx = DAG.getVectorIdxConstant(0, dl);
 
  // Extend to natively supported kshift.
  MVT WideOpVT = widenMaskVectorType(OpVT, Subtarget);
 
  // Inserting into the lsbs of a zero vector is legal. ISel will insert shifts
  // if necessary.
  if (IdxVal == 0 && ISD::isBuildVectorAllZeros(Vec.getNode())) {
    // May need to promote to a legal type.
    Op = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT,
                     DAG.getConstant(0, dl, WideOpVT),
                     SubVec, Idx);
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx);
  }
 
  MVT SubVecVT = SubVec.getSimpleValueType();
  unsigned SubVecNumElems = SubVecVT.getVectorNumElements();
  assert(IdxVal + SubVecNumElems <= NumElems &&
         IdxVal % SubVecVT.getSizeInBits() == 0 &&
         "Unexpected index value in INSERT_SUBVECTOR");
 
  SDValue Undef = DAG.getUNDEF(WideOpVT);
 
  if (IdxVal == 0) {
    // Zero lower bits of the Vec
    SDValue ShiftBits = DAG.getTargetConstant(SubVecNumElems, dl, MVT::i8);
    Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, Undef, Vec,
                      ZeroIdx);
    Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, ShiftBits);
    Vec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec, ShiftBits);
    // Merge them together, SubVec should be zero extended.
    SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT,
                         DAG.getConstant(0, dl, WideOpVT),
                         SubVec, ZeroIdx);
    Op = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, SubVec);
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx);
  }
 
  SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT,
                       Undef, SubVec, ZeroIdx);
 
  if (Vec.isUndef()) {
    assert(IdxVal != 0 && "Unexpected index");
    SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec,
                         DAG.getTargetConstant(IdxVal, dl, MVT::i8));
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, SubVec, ZeroIdx);
  }
 
  if (ISD::isBuildVectorAllZeros(Vec.getNode())) {
    assert(IdxVal != 0 && "Unexpected index");
    // If upper elements of Vec are known undef, then just shift into place.
    if (llvm::all_of(Vec->ops().slice(IdxVal + SubVecNumElems),
                     [](SDValue V) { return V.isUndef(); })) {
      SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec,
                           DAG.getTargetConstant(IdxVal, dl, MVT::i8));
    } else {
      NumElems = WideOpVT.getVectorNumElements();
      unsigned ShiftLeft = NumElems - SubVecNumElems;
      unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
      SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec,
                           DAG.getTargetConstant(ShiftLeft, dl, MVT::i8));
      if (ShiftRight != 0)
        SubVec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, SubVec,
                             DAG.getTargetConstant(ShiftRight, dl, MVT::i8));
    }
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, SubVec, ZeroIdx);
  }
 
  // Simple case when we put subvector in the upper part
  if (IdxVal + SubVecNumElems == NumElems) {
    SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec,
                         DAG.getTargetConstant(IdxVal, dl, MVT::i8));
    if (SubVecNumElems * 2 == NumElems) {
      // Special case, use legal zero extending insert_subvector. This allows
      // isel to optimize when bits are known zero.
      Vec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVecVT, Vec, ZeroIdx);
      Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT,
                        DAG.getConstant(0, dl, WideOpVT),
                        Vec, ZeroIdx);
    } else {
      // Otherwise use explicit shifts to zero the bits.
      Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT,
                        Undef, Vec, ZeroIdx);
      NumElems = WideOpVT.getVectorNumElements();
      SDValue ShiftBits = DAG.getTargetConstant(NumElems - IdxVal, dl, MVT::i8);
      Vec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec, ShiftBits);
      Vec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec, ShiftBits);
    }
    Op = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, SubVec);
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx);
  }
 
  // Inserting into the middle is more complicated.
 
  NumElems = WideOpVT.getVectorNumElements();
 
  // Widen the vector if needed.
  Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, WideOpVT, Undef, Vec, ZeroIdx);
 
  unsigned ShiftLeft = NumElems - SubVecNumElems;
  unsigned ShiftRight = NumElems - SubVecNumElems - IdxVal;
 
  // Do an optimization for the most frequently used types.
  if (WideOpVT != MVT::v64i1 || Subtarget.is64Bit()) {
    APInt Mask0 = APInt::getBitsSet(NumElems, IdxVal, IdxVal + SubVecNumElems);
    Mask0.flipAllBits();
    SDValue CMask0 = DAG.getConstant(Mask0, dl, MVT::getIntegerVT(NumElems));
    SDValue VMask0 = DAG.getNode(ISD::BITCAST, dl, WideOpVT, CMask0);
    Vec = DAG.getNode(ISD::AND, dl, WideOpVT, Vec, VMask0);
    SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec,
                         DAG.getTargetConstant(ShiftLeft, dl, MVT::i8));
    SubVec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, SubVec,
                         DAG.getTargetConstant(ShiftRight, dl, MVT::i8));
    Op = DAG.getNode(ISD::OR, dl, WideOpVT, Vec, SubVec);
 
    // Reduce to original width if needed.
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, Op, ZeroIdx);
  }
 
  // Clear the upper bits of the subvector and move it to its insert position.
  SubVec = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, SubVec,
                       DAG.getTargetConstant(ShiftLeft, dl, MVT::i8));
  SubVec = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, SubVec,
                       DAG.getTargetConstant(ShiftRight, dl, MVT::i8));
 
  // Isolate the bits below the insertion point.
  unsigned LowShift = NumElems - IdxVal;
  SDValue Low = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, Vec,
                            DAG.getTargetConstant(LowShift, dl, MVT::i8));
  Low = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Low,
                    DAG.getTargetConstant(LowShift, dl, MVT::i8));
 
  // Isolate the bits after the last inserted bit.
  unsigned HighShift = IdxVal + SubVecNumElems;
  SDValue High = DAG.getNode(X86ISD::KSHIFTR, dl, WideOpVT, Vec,
                            DAG.getTargetConstant(HighShift, dl, MVT::i8));
  High = DAG.getNode(X86ISD::KSHIFTL, dl, WideOpVT, High,
                    DAG.getTargetConstant(HighShift, dl, MVT::i8));
 
  // Now OR all 3 pieces together.
  Vec = DAG.getNode(ISD::OR, dl, WideOpVT, Low, High);
  SubVec = DAG.getNode(ISD::OR, dl, WideOpVT, SubVec, Vec);
 
  // Reduce to original width if needed.
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, OpVT, SubVec, ZeroIdx);
}
 
static SDValue concatSubVectors(SDValue V1, SDValue V2, SelectionDAG &DAG,
                                const SDLoc &dl) {
  assert(V1.getValueType() == V2.getValueType() && "subvector type mismatch");
  EVT SubVT = V1.getValueType();
  EVT SubSVT = SubVT.getScalarType();
  unsigned SubNumElts = SubVT.getVectorNumElements();
  unsigned SubVectorWidth = SubVT.getSizeInBits();
  EVT VT = EVT::getVectorVT(*DAG.getContext(), SubSVT, 2 * SubNumElts);
  SDValue V = insertSubVector(DAG.getUNDEF(VT), V1, 0, DAG, dl, SubVectorWidth);
  return insertSubVector(V, V2, SubNumElts, DAG, dl, SubVectorWidth);
}
 
/// Returns a vector of specified type with all bits set.
/// Always build ones vectors as <4 x i32>, <8 x i32> or <16 x i32>.
/// Then bitcast to their original type, ensuring they get CSE'd.
static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, const SDLoc &dl) {
  assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
         "Expected a 128/256/512-bit vector type");
  unsigned NumElts = VT.getSizeInBits() / 32;
  SDValue Vec = DAG.getAllOnesConstant(dl, MVT::getVectorVT(MVT::i32, NumElts));
  return DAG.getBitcast(VT, Vec);
}
 
// Helper to get immediate/variable SSE shift opcode from other shift opcodes.
static unsigned getTargetVShiftUniformOpcode(unsigned Opc, bool IsVariable) {
  switch (Opc) {
  case ISD::SHL:
  case X86ISD::VSHL:
  case X86ISD::VSHLI:
    return IsVariable ? X86ISD::VSHL : X86ISD::VSHLI;
  case ISD::SRL:
  case X86ISD::VSRL:
  case X86ISD::VSRLI:
    return IsVariable ? X86ISD::VSRL : X86ISD::VSRLI;
  case ISD::SRA:
  case X86ISD::VSRA:
  case X86ISD::VSRAI:
    return IsVariable ? X86ISD::VSRA : X86ISD::VSRAI;
  }
  llvm_unreachable("Unknown target vector shift node");
}
 
/// Handle vector element shifts where the shift amount is a constant.
/// Takes immediate version of shift as input.
static SDValue getTargetVShiftByConstNode(unsigned Opc, const SDLoc &dl, MVT VT,
                                          SDValue SrcOp, uint64_t ShiftAmt,
                                          SelectionDAG &DAG) {
  assert(
      (Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI) &&
      "Unknown target vector shift-by-constant node");
 
  // Bitcast the source vector to the output type, this is mainly necessary for
  // vXi8/vXi64 shifts.
  SrcOp = DAG.getBitcast(VT, SrcOp);
 
  // Fold this packed shift into its first operand if ShiftAmt is 0.
  if (ShiftAmt == 0)
    return SrcOp;
 
  // Check for ShiftAmt >= element width
  unsigned EltSizeInBits = VT.getScalarSizeInBits();
  if (ShiftAmt >= EltSizeInBits) {
    if (Opc == X86ISD::VSRAI)
      ShiftAmt = EltSizeInBits - 1;
    else
      return DAG.getConstant(0, dl, VT);
  }
 
  // Fold this packed vector shift into a build vector if SrcOp is a
  // vector of Constants or UNDEFs.
  if (ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
    unsigned ShiftOpc;
    switch (Opc) {
    default:
      llvm_unreachable("Unknown opcode!");
    case X86ISD::VSHLI:
      ShiftOpc = ISD::SHL;
      break;
    case X86ISD::VSRLI:
      ShiftOpc = ISD::SRL;
      break;
    case X86ISD::VSRAI:
      ShiftOpc = ISD::SRA;
      break;
    }
 
    SDValue Amt = DAG.getConstant(ShiftAmt, dl, VT);
    if (SDValue C = DAG.FoldConstantArithmetic(ShiftOpc, dl, VT, {SrcOp, Amt}))
      return C;
  }
 
  return DAG.getNode(Opc, dl, VT, SrcOp,
                     DAG.getTargetConstant(ShiftAmt, dl, MVT::i8));
}
 
/// Handle vector element shifts by a splat shift amount
static SDValue getTargetVShiftNode(unsigned Opc, const SDLoc &dl, MVT VT,
                                   SDValue SrcOp, SDValue ShAmt, int ShAmtIdx,
                                   const X86Subtarget &Subtarget,
                                   SelectionDAG &DAG) {
  MVT AmtVT = ShAmt.getSimpleValueType();
  assert(AmtVT.isVector() && "Vector shift type mismatch");
  assert(0 <= ShAmtIdx && ShAmtIdx < (int)AmtVT.getVectorNumElements() &&
         "Illegal vector splat index");
 
  // Move the splat element to the bottom element.
  if (ShAmtIdx != 0) {
    SmallVector<int> Mask(AmtVT.getVectorNumElements(), -1);
    Mask[0] = ShAmtIdx;
    ShAmt = DAG.getVectorShuffle(AmtVT, dl, ShAmt, DAG.getUNDEF(AmtVT), Mask);
  }
 
  // Peek through any zext node if we can get back to a 128-bit source.
  if (AmtVT.getScalarSizeInBits() == 64 &&
      (ShAmt.getOpcode() == ISD::ZERO_EXTEND ||
       ShAmt.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) &&
      ShAmt.getOperand(0).getValueType().isSimple() &&
      ShAmt.getOperand(0).getValueType().is128BitVector()) {
    ShAmt = ShAmt.getOperand(0);
    AmtVT = ShAmt.getSimpleValueType();
  }
 
  // See if we can mask off the upper elements using the existing source node.
  // The shift uses the entire lower 64-bits of the amount vector, so no need to
  // do this for vXi64 types.
  bool IsMasked = false;
  if (AmtVT.getScalarSizeInBits() < 64) {
    if (ShAmt.getOpcode() == ISD::BUILD_VECTOR ||
        ShAmt.getOpcode() == ISD::SCALAR_TO_VECTOR) {
      // If the shift amount has come from a scalar, then zero-extend the scalar
      // before moving to the vector.
      ShAmt = DAG.getZExtOrTrunc(ShAmt.getOperand(0), dl, MVT::i32);
      ShAmt = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, ShAmt);
      ShAmt = DAG.getNode(X86ISD::VZEXT_MOVL, dl, MVT::v4i32, ShAmt);
      AmtVT = MVT::v4i32;
      IsMasked = true;
    } else if (ShAmt.getOpcode() == ISD::AND) {
      // See if the shift amount is already masked (e.g. for rotation modulo),
      // then we can zero-extend it by setting all the other mask elements to
      // zero.
      SmallVector<SDValue> MaskElts(
          AmtVT.getVectorNumElements(),
          DAG.getConstant(0, dl, AmtVT.getScalarType()));
      MaskElts[0] = DAG.getAllOnesConstant(dl, AmtVT.getScalarType());
      SDValue Mask = DAG.getBuildVector(AmtVT, dl, MaskElts);
      if ((Mask = DAG.FoldConstantArithmetic(ISD::AND, dl, AmtVT,
                                             {ShAmt.getOperand(1), Mask}))) {
        ShAmt = DAG.getNode(ISD::AND, dl, AmtVT, ShAmt.getOperand(0), Mask);
        IsMasked = true;
      }
    }
  }
 
  // Extract if the shift amount vector is larger than 128-bits.
  if (AmtVT.getSizeInBits() > 128) {
    ShAmt = extract128BitVector(ShAmt, 0, DAG, dl);
    AmtVT = ShAmt.getSimpleValueType();
  }
 
  // Zero-extend bottom element to v2i64 vector type, either by extension or
  // shuffle masking.
  if (!IsMasked && AmtVT.getScalarSizeInBits() < 64) {
    if (AmtVT == MVT::v4i32 && (ShAmt.getOpcode() == X86ISD::VBROADCAST ||
                                ShAmt.getOpcode() == X86ISD::VBROADCAST_LOAD)) {
      ShAmt = DAG.getNode(X86ISD::VZEXT_MOVL, SDLoc(ShAmt), MVT::v4i32, ShAmt);
    } else if (Subtarget.hasSSE41()) {
      ShAmt = DAG.getNode(ISD::ZERO_EXTEND_VECTOR_INREG, SDLoc(ShAmt),
                          MVT::v2i64, ShAmt);
    } else {
      SDValue ByteShift = DAG.getTargetConstant(
          (128 - AmtVT.getScalarSizeInBits()) / 8, SDLoc(ShAmt), MVT::i8);
      ShAmt = DAG.getBitcast(MVT::v16i8, ShAmt);
      ShAmt = DAG.getNode(X86ISD::VSHLDQ, SDLoc(ShAmt), MVT::v16i8, ShAmt,
                          ByteShift);
      ShAmt = DAG.getNode(X86ISD::VSRLDQ, SDLoc(ShAmt), MVT::v16i8, ShAmt,
                          ByteShift);
    }
  }
 
  // Change opcode to non-immediate version.
  Opc = getTargetVShiftUniformOpcode(Opc, true);
 
  // The return type has to be a 128-bit type with the same element
  // type as the input type.
  MVT EltVT = VT.getVectorElementType();
  MVT ShVT = MVT::getVectorVT(EltVT, 128 / EltVT.getSizeInBits());
 
  ShAmt = DAG.getBitcast(ShVT, ShAmt);
  return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
}
 
static SDValue getEXTEND_VECTOR_INREG(unsigned Opcode, const SDLoc &DL, EVT VT,
                                      SDValue In, SelectionDAG &DAG) {
  EVT InVT = In.getValueType();
  assert(VT.isVector() && InVT.isVector() && "Expected vector VTs.");
 
  // Canonicalize Opcode to general extension version.
  switch (Opcode) {
  case ISD::ANY_EXTEND:
  case ISD::ANY_EXTEND_VECTOR_INREG:
    Opcode = ISD::ANY_EXTEND;
    break;
  case ISD::SIGN_EXTEND:
  case ISD::SIGN_EXTEND_VECTOR_INREG:
    Opcode = ISD::SIGN_EXTEND;
    break;
  case ISD::ZERO_EXTEND:
  case ISD::ZERO_EXTEND_VECTOR_INREG:
    Opcode = ISD::ZERO_EXTEND;
    break;
  default:
    llvm_unreachable("Unknown extension opcode");
  }
 
  // For 256-bit vectors, we only need the lower (128-bit) input half.
  // For 512-bit vectors, we only need the lower input half or quarter.
  if (InVT.getSizeInBits() > 128) {
    assert(VT.getSizeInBits() == InVT.getSizeInBits() &&
           "Expected VTs to be the same size!");
    unsigned Scale = VT.getScalarSizeInBits() / InVT.getScalarSizeInBits();
    In = extractSubVector(In, 0, DAG, DL,
                          std::max(128U, (unsigned)VT.getSizeInBits() / Scale));
    InVT = In.getValueType();
  }
 
  if (VT.getVectorNumElements() != InVT.getVectorNumElements())
    Opcode = DAG.getOpcode_EXTEND_VECTOR_INREG(Opcode);
 
  return DAG.getNode(Opcode, DL, VT, In);
}
 
// Create OR(AND(LHS,MASK),AND(RHS,~MASK)) bit select pattern
static SDValue getBitSelect(const SDLoc &DL, MVT VT, SDValue LHS, SDValue RHS,
                            SDValue Mask, SelectionDAG &DAG) {
  LHS = DAG.getNode(ISD::AND, DL, VT, LHS, Mask);
  RHS = DAG.getNode(X86ISD::ANDNP, DL, VT, Mask, RHS);
  return DAG.getNode(ISD::OR, DL, VT, LHS, RHS);
}
 
void llvm::createUnpackShuffleMask(EVT VT, SmallVectorImpl<int> &Mask,
                                   bool Lo, bool Unary) {
  assert(VT.getScalarType().isSimple() && (VT.getSizeInBits() % 128) == 0 &&
         "Illegal vector type to unpack");
  assert(Mask.empty() && "Expected an empty shuffle mask vector");
  int NumElts = VT.getVectorNumElements();
  int NumEltsInLane = 128 / VT.getScalarSizeInBits();
  for (int i = 0; i < NumElts; ++i) {
    unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
    int Pos = (i % NumEltsInLane) / 2 + LaneStart;
    Pos += (Unary ? 0 : NumElts * (i % 2));
    Pos += (Lo ? 0 : NumEltsInLane / 2);
    Mask.push_back(Pos);
  }
}
 
/// Similar to unpacklo/unpackhi, but without the 128-bit lane limitation
/// imposed by AVX and specific to the unary pattern. Example:
/// v8iX Lo --> <0, 0, 1, 1, 2, 2, 3, 3>
/// v8iX Hi --> <4, 4, 5, 5, 6, 6, 7, 7>
void llvm::createSplat2ShuffleMask(MVT VT, SmallVectorImpl<int> &Mask,
                                   bool Lo) {
  assert(Mask.empty() && "Expected an empty shuffle mask vector");
  int NumElts = VT.getVectorNumElements();
  for (int i = 0; i < NumElts; ++i) {
    int Pos = i / 2;
    Pos += (Lo ? 0 : NumElts / 2);
    Mask.push_back(Pos);
  }
}
 
// Attempt to constant fold, else just create a VECTOR_SHUFFLE.
static SDValue getVectorShuffle(SelectionDAG &DAG, EVT VT, const SDLoc &dl,
                                SDValue V1, SDValue V2, ArrayRef<int> Mask) {
  if ((ISD::isBuildVectorOfConstantSDNodes(V1.getNode()) || V1.isUndef()) &&
      (ISD::isBuildVectorOfConstantSDNodes(V2.getNode()) || V2.isUndef())) {
    SmallVector<SDValue> Ops(Mask.size(), DAG.getUNDEF(VT.getScalarType()));
    for (int I = 0, NumElts = Mask.size(); I != NumElts; ++I) {
      int M = Mask[I];
      if (M < 0)
        continue;
      SDValue V = (M < NumElts) ? V1 : V2;
      if (V.isUndef())
        continue;
      Ops[I] = V.getOperand(M % NumElts);
    }
    return DAG.getBuildVector(VT, dl, Ops);
  }
 
  return DAG.getVectorShuffle(VT, dl, V1, V2, Mask);
}
 
/// Returns a vector_shuffle node for an unpackl operation.
static SDValue getUnpackl(SelectionDAG &DAG, const SDLoc &dl, EVT VT,
                          SDValue V1, SDValue V2) {
  SmallVector<int, 8> Mask;
  createUnpackShuffleMask(VT, Mask, /* Lo = */ true, /* Unary = */ false);
  return getVectorShuffle(DAG, VT, dl, V1, V2, Mask);
}
 
/// Returns a vector_shuffle node for an unpackh operation.
static SDValue getUnpackh(SelectionDAG &DAG, const SDLoc &dl, EVT VT,
                          SDValue V1, SDValue V2) {
  SmallVector<int, 8> Mask;
  createUnpackShuffleMask(VT, Mask, /* Lo = */ false, /* Unary = */ false);
  return getVectorShuffle(DAG, VT, dl, V1, V2, Mask);
}
 
/// Returns a node that packs the LHS + RHS nodes together at half width.
/// May return X86ISD::PACKSS/PACKUS, packing the top/bottom half.
/// TODO: Add subvector splitting if/when we have a need for it.
static SDValue getPack(SelectionDAG &DAG, const X86Subtarget &Subtarget,
                       const SDLoc &dl, MVT VT, SDValue LHS, SDValue RHS,
                       bool PackHiHalf = false) {
  MVT OpVT = LHS.getSimpleValueType();
  unsigned EltSizeInBits = VT.getScalarSizeInBits();
  bool UsePackUS = Subtarget.hasSSE41() || EltSizeInBits == 8;
  assert(OpVT == RHS.getSimpleValueType() &&
         VT.getSizeInBits() == OpVT.getSizeInBits() &&
         (EltSizeInBits * 2) == OpVT.getScalarSizeInBits() &&
         "Unexpected PACK operand types");
  assert((EltSizeInBits == 8 || EltSizeInBits == 16 || EltSizeInBits == 32) &&
         "Unexpected PACK result type");
 
  // Rely on vector shuffles for vXi64 -> vXi32 packing.
  if (EltSizeInBits == 32) {
    SmallVector<int> PackMask;
    int Offset = PackHiHalf ? 1 : 0;
    int NumElts = VT.getVectorNumElements();
    for (int I = 0; I != NumElts; I += 4) {
      PackMask.push_back(I + Offset);
      PackMask.push_back(I + Offset + 2);
      PackMask.push_back(I + Offset + NumElts);
      PackMask.push_back(I + Offset + NumElts + 2);
    }
    return DAG.getVectorShuffle(VT, dl, DAG.getBitcast(VT, LHS),
                                DAG.getBitcast(VT, RHS), PackMask);
  }
 
  // See if we already have sufficient leading bits for PACKSS/PACKUS.
  if (!PackHiHalf) {
    if (UsePackUS &&
        DAG.computeKnownBits(LHS).countMaxActiveBits() <= EltSizeInBits &&
        DAG.computeKnownBits(RHS).countMaxActiveBits() <= EltSizeInBits)
      return DAG.getNode(X86ISD::PACKUS, dl, VT, LHS, RHS);
 
    if (DAG.ComputeMaxSignificantBits(LHS) <= EltSizeInBits &&
        DAG.ComputeMaxSignificantBits(RHS) <= EltSizeInBits)
      return DAG.getNode(X86ISD::PACKSS, dl, VT, LHS, RHS);
  }
 
  // Fallback to sign/zero extending the requested half and pack.
  SDValue Amt = DAG.getTargetConstant(EltSizeInBits, dl, MVT::i8);
  if (UsePackUS) {
    if (PackHiHalf) {
      LHS = DAG.getNode(X86ISD::VSRLI, dl, OpVT, LHS, Amt);
      RHS = DAG.getNode(X86ISD::VSRLI, dl, OpVT, RHS, Amt);
    } else {
      SDValue Mask = DAG.getConstant((1ULL << EltSizeInBits) - 1, dl, OpVT);
      LHS = DAG.getNode(ISD::AND, dl, OpVT, LHS, Mask);
      RHS = DAG.getNode(ISD::AND, dl, OpVT, RHS, Mask);
    };
    return DAG.getNode(X86ISD::PACKUS, dl, VT, LHS, RHS);
  };
 
  if (!PackHiHalf) {
    LHS = DAG.getNode(X86ISD::VSHLI, dl, OpVT, LHS, Amt);
    RHS = DAG.getNode(X86ISD::VSHLI, dl, OpVT, RHS, Amt);
  }
  LHS = DAG.getNode(X86ISD::VSRAI, dl, OpVT, LHS, Amt);
  RHS = DAG.getNode(X86ISD::VSRAI, dl, OpVT, RHS, Amt);
  return DAG.getNode(X86ISD::PACKSS, dl, VT, LHS, RHS);
}
 
/// Return a vector_shuffle of the specified vector of zero or undef vector.
/// This produces a shuffle where the low element of V2 is swizzled into the
/// zero/undef vector, landing at element Idx.
/// This produces a shuffle mask like 4,1,2,3 (idx=0) or  0,1,2,4 (idx=3).
static SDValue getShuffleVectorZeroOrUndef(SDValue V2, int Idx,
                                           bool IsZero,
                                           const X86Subtarget &Subtarget,
                                           SelectionDAG &DAG) {
  MVT VT = V2.getSimpleValueType();
  SDValue V1 = IsZero
    ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
  int NumElems = VT.getVectorNumElements();
  SmallVector<int, 16> MaskVec(NumElems);
  for (int i = 0; i != NumElems; ++i)
    // If this is the insertion idx, put the low elt of V2 here.
    MaskVec[i] = (i == Idx) ? NumElems : i;
  return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, MaskVec);
}
 
static ConstantPoolSDNode *getTargetConstantPoolFromBasePtr(SDValue Ptr) {
  if (Ptr.getOpcode() == X86ISD::Wrapper ||
      Ptr.getOpcode() == X86ISD::WrapperRIP)
    Ptr = Ptr.getOperand(0);
  return dyn_cast<ConstantPoolSDNode>(Ptr);
}
 
// TODO: Add support for non-zero offsets.
static const Constant *getTargetConstantFromBasePtr(SDValue Ptr) {
  ConstantPoolSDNode *CNode = getTargetConstantPoolFromBasePtr(Ptr);
  if (!CNode || CNode->isMachineConstantPoolEntry() || CNode->getOffset() != 0)
    return nullptr;
  return CNode->getConstVal();
}
 
static const Constant *getTargetConstantFromNode(LoadSDNode *Load) {
  if (!Load || !ISD::isNormalLoad(Load))
    return nullptr;
  return getTargetConstantFromBasePtr(Load->getBasePtr());
}
 
static const Constant *getTargetConstantFromNode(SDValue Op) {
  Op = peekThroughBitcasts(Op);
  return getTargetConstantFromNode(dyn_cast<LoadSDNode>(Op));
}
 
const Constant *
X86TargetLowering::getTargetConstantFromLoad(LoadSDNode *LD) const {
  assert(LD && "Unexpected null LoadSDNode");
  return getTargetConstantFromNode(LD);
}
 
bool X86TargetLowering::isTargetCanonicalSelect(SDNode *N) const {
  // Do not fold (vselect not(C), X, 0s) to (vselect C, Os, X)
  SDValue Cond = N->getOperand(0);
  SDValue RHS = N->getOperand(2);
  EVT CondVT = Cond.getValueType();
  return N->getOpcode() == ISD::VSELECT && Subtarget.hasAVX512() &&
         CondVT.getVectorElementType() == MVT::i1 &&
         ISD::isBuildVectorAllZeros(RHS.getNode());
}
 
// Extract raw constant bits from constant pools.
static bool getTargetConstantBitsFromNode(SDValue Op, unsigned EltSizeInBits,
                                          APInt &UndefElts,
                                          SmallVectorImpl<APInt> &EltBits,
                                          bool AllowWholeUndefs = true,
                                          bool AllowPartialUndefs = false) {
  assert(EltBits.empty() && "Expected an empty EltBits vector");
 
  Op = peekThroughBitcasts(Op);
 
  EVT VT = Op.getValueType();
  unsigned SizeInBits = VT.getSizeInBits();
  unsigned NumElts = SizeInBits / EltSizeInBits;
 
  // Can't split constant.
  if ((SizeInBits % EltSizeInBits) != 0)
    return false;
 
  // Bitcast a source array of element bits to the target size.
  auto CastBitData = [&](APInt &UndefSrcElts, ArrayRef<APInt> SrcEltBits) {
    unsigned NumSrcElts = UndefSrcElts.getBitWidth();
    unsigned SrcEltSizeInBits = SrcEltBits[0].getBitWidth();
    assert((NumSrcElts * SrcEltSizeInBits) == SizeInBits &&
           "Constant bit sizes don't match");
 
    // Don't split if we don't allow undef bits.
    bool AllowUndefs = AllowWholeUndefs || AllowPartialUndefs;
    if (UndefSrcElts.getBoolValue() && !AllowUndefs)
      return false;
 
    // If we're already the right size, don't bother bitcasting.
    if (NumSrcElts == NumElts) {
      UndefElts = UndefSrcElts;
      EltBits.assign(SrcEltBits.begin(), SrcEltBits.end());
      return true;
    }
 
    // Extract all the undef/constant element data and pack into single bitsets.
    APInt UndefBits(SizeInBits, 0);
    APInt MaskBits(SizeInBits, 0);
 
    for (unsigned i = 0; i != NumSrcElts; ++i) {
      unsigned BitOffset = i * SrcEltSizeInBits;
      if (UndefSrcElts[i])
        UndefBits.setBits(BitOffset, BitOffset + SrcEltSizeInBits);
      MaskBits.insertBits(SrcEltBits[i], BitOffset);
    }
 
    // Split the undef/constant single bitset data into the target elements.
    UndefElts = APInt(NumElts, 0);
    EltBits.resize(NumElts, APInt(EltSizeInBits, 0));
 
    for (unsigned i = 0; i != NumElts; ++i) {
      unsigned BitOffset = i * EltSizeInBits;
      APInt UndefEltBits = UndefBits.extractBits(EltSizeInBits, BitOffset);
 
      // Only treat an element as UNDEF if all bits are UNDEF.
      if (UndefEltBits.isAllOnes()) {
        if (!AllowWholeUndefs)
          return false;
        UndefElts.setBit(i);
        continue;
      }
 
      // If only some bits are UNDEF then treat them as zero (or bail if not
      // supported).
      if (UndefEltBits.getBoolValue() && !AllowPartialUndefs)
        return false;
 
      EltBits[i] = MaskBits.extractBits(EltSizeInBits, BitOffset);
    }
    return true;
  };
 
  // Collect constant bits and insert into mask/undef bit masks.
  auto CollectConstantBits = [](const Constant *Cst, APInt &Mask, APInt &Undefs,
                                unsigned UndefBitIndex) {
    if (!Cst)
      return false;
    if (isa<UndefValue>(Cst)) {
      Undefs.setBit(UndefBitIndex);
      return true;
    }
    if (auto *CInt = dyn_cast<ConstantInt>(Cst)) {
      Mask = APInt::getSplat(CInt->getType()->getPrimitiveSizeInBits(),
                             CInt->getValue());
      return true;
    }
    if (auto *CFP = dyn_cast<ConstantFP>(Cst)) {
      Mask = APInt::getSplat(CFP->getType()->getPrimitiveSizeInBits(),
                             CFP->getValueAPF().bitcastToAPInt());
      return true;
    }
    if (auto *CDS = dyn_cast<ConstantDataSequential>(Cst)) {
      Type *Ty = CDS->getType();
      Mask = APInt::getZero(Ty->getPrimitiveSizeInBits());
      Type *EltTy = CDS->getElementType();
      bool IsInteger = EltTy->isIntegerTy();
      bool IsFP =
          EltTy->isHalfTy() || EltTy->isFloatTy() || EltTy->isDoubleTy();
      if (!IsInteger && !IsFP)
        return false;
      unsigned EltBits = EltTy->getPrimitiveSizeInBits();
      for (unsigned I = 0, E = CDS->getNumElements(); I != E; ++I)
        if (IsInteger)
          Mask.insertBits(CDS->getElementAsAPInt(I), I * EltBits);
        else
          Mask.insertBits(CDS->getElementAsAPFloat(I).bitcastToAPInt(),
                          I * EltBits);
      return true;
    }
    return false;
  };
 
  // Handle UNDEFs.
  if (Op.isUndef()) {
    APInt UndefSrcElts = APInt::getAllOnes(NumElts);
    SmallVector<APInt, 64> SrcEltBits(NumElts, APInt(EltSizeInBits, 0));
    return CastBitData(UndefSrcElts, SrcEltBits);
  }
 
  // Extract scalar constant bits.
  if (auto *Cst = dyn_cast<ConstantSDNode>(Op)) {
    APInt UndefSrcElts = APInt::getZero(1);
    SmallVector<APInt, 64> SrcEltBits(1, Cst->getAPIntValue());
    return CastBitData(UndefSrcElts, SrcEltBits);
  }
  if (auto *Cst = dyn_cast<ConstantFPSDNode>(Op)) {
    APInt UndefSrcElts = APInt::getZero(1);
    APInt RawBits = Cst->getValueAPF().bitcastToAPInt();
    SmallVector<APInt, 64> SrcEltBits(1, RawBits);
    return CastBitData(UndefSrcElts, SrcEltBits);
  }
 
  // Extract constant bits from build vector.
  if (auto *BV = dyn_cast<BuildVectorSDNode>(Op)) {
    BitVector Undefs;
    SmallVector<APInt> SrcEltBits;
    unsigned SrcEltSizeInBits = VT.getScalarSizeInBits();
    if (BV->getConstantRawBits(true, SrcEltSizeInBits, SrcEltBits, Undefs)) {
      APInt UndefSrcElts = APInt::getZero(SrcEltBits.size());
      for (unsigned I = 0, E = SrcEltBits.size(); I != E; ++I)
        if (Undefs[I])
          UndefSrcElts.setBit(I);
      return CastBitData(UndefSrcElts, SrcEltBits);
    }
  }
 
  // Extract constant bits from constant pool vector.
  if (auto *Cst = getTargetConstantFromNode(Op)) {
    Type *CstTy = Cst->getType();
    unsigned CstSizeInBits = CstTy->getPrimitiveSizeInBits();
    if (!CstTy->isVectorTy() || (CstSizeInBits % SizeInBits) != 0)
      return false;
 
    unsigned SrcEltSizeInBits = CstTy->getScalarSizeInBits();
    unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits;
    if ((SizeInBits % SrcEltSizeInBits) != 0)
      return false;
 
    APInt UndefSrcElts(NumSrcElts, 0);
    SmallVector<APInt, 64> SrcEltBits(NumSrcElts, APInt(SrcEltSizeInBits, 0));
    for (unsigned i = 0; i != NumSrcElts; ++i)
      if (!CollectConstantBits(Cst->getAggregateElement(i), SrcEltBits[i],
                               UndefSrcElts, i))
        return false;
 
    return CastBitData(UndefSrcElts, SrcEltBits);
  }
 
  // Extract constant bits from a broadcasted constant pool scalar.
  if (Op.getOpcode() == X86ISD::VBROADCAST_LOAD &&
      EltSizeInBits <= VT.getScalarSizeInBits()) {
    auto *MemIntr = cast<MemIntrinsicSDNode>(Op);
    if (MemIntr->getMemoryVT().getStoreSizeInBits() != VT.getScalarSizeInBits())
      return false;
 
    SDValue Ptr = MemIntr->getBasePtr();
    if (const Constant *C = getTargetConstantFromBasePtr(Ptr)) {
      unsigned SrcEltSizeInBits = VT.getScalarSizeInBits();
      unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits;
 
      APInt UndefSrcElts(NumSrcElts, 0);
      SmallVector<APInt, 64> SrcEltBits(1, APInt(SrcEltSizeInBits, 0));
      if (CollectConstantBits(C, SrcEltBits[0], UndefSrcElts, 0)) {
        if (UndefSrcElts[0])
          UndefSrcElts.setBits(0, NumSrcElts);
        if (SrcEltBits[0].getBitWidth() != SrcEltSizeInBits)
          SrcEltBits[0] = SrcEltBits[0].trunc(SrcEltSizeInBits);
        SrcEltBits.append(NumSrcElts - 1, SrcEltBits[0]);
        return CastBitData(UndefSrcElts, SrcEltBits);
      }
    }
  }
 
  // Extract constant bits from a subvector broadcast.
  if (Op.getOpcode() == X86ISD::SUBV_BROADCAST_LOAD) {
    auto *MemIntr = cast<MemIntrinsicSDNode>(Op);
    SDValue Ptr = MemIntr->getBasePtr();
    // The source constant may be larger than the subvector broadcast,
    // ensure we extract the correct subvector constants.
    if (const Constant *Cst = getTargetConstantFromBasePtr(Ptr)) {
      Type *CstTy = Cst->getType();
      unsigned CstSizeInBits = CstTy->getPrimitiveSizeInBits();
      unsigned SubVecSizeInBits = MemIntr->getMemoryVT().getStoreSizeInBits();
      if (!CstTy->isVectorTy() || (CstSizeInBits % SubVecSizeInBits) != 0 ||
          (SizeInBits % SubVecSizeInBits) != 0)
        return false;
      unsigned CstEltSizeInBits = CstTy->getScalarSizeInBits();
      unsigned NumSubElts = SubVecSizeInBits / CstEltSizeInBits;
      unsigned NumSubVecs = SizeInBits / SubVecSizeInBits;
      APInt UndefSubElts(NumSubElts, 0);
      SmallVector<APInt, 64> SubEltBits(NumSubElts * NumSubVecs,
                                        APInt(CstEltSizeInBits, 0));
      for (unsigned i = 0; i != NumSubElts; ++i) {
        if (!CollectConstantBits(Cst->getAggregateElement(i), SubEltBits[i],
                                 UndefSubElts, i))
          return false;
        for (unsigned j = 1; j != NumSubVecs; ++j)
          SubEltBits[i + (j * NumSubElts)] = SubEltBits[i];
      }
      UndefSubElts = APInt::getSplat(NumSubVecs * UndefSubElts.getBitWidth(),
                                     UndefSubElts);
      return CastBitData(UndefSubElts, SubEltBits);
    }
  }
 
  // Extract a rematerialized scalar constant insertion.
  if (Op.getOpcode() == X86ISD::VZEXT_MOVL &&
      Op.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
      isa<ConstantSDNode>(Op.getOperand(0).getOperand(0))) {
    unsigned SrcEltSizeInBits = VT.getScalarSizeInBits();
    unsigned NumSrcElts = SizeInBits / SrcEltSizeInBits;
 
    APInt UndefSrcElts(NumSrcElts, 0);
    SmallVector<APInt, 64> SrcEltBits;
    const APInt &C = Op.getOperand(0).getConstantOperandAPInt(0);
    SrcEltBits.push_back(C.zextOrTrunc(SrcEltSizeInBits));
    SrcEltBits.append(NumSrcElts - 1, APInt(SrcEltSizeInBits, 0));
    return CastBitData(UndefSrcElts, SrcEltBits);
  }
 
  // Insert constant bits from a base and sub vector sources.
  if (Op.getOpcode() == ISD::INSERT_SUBVECTOR) {
    // If bitcasts to larger elements we might lose track of undefs - don't
    // allow any to be safe.
    unsigned SrcEltSizeInBits = VT.getScalarSizeInBits();
    bool AllowUndefs = EltSizeInBits >= SrcEltSizeInBits;
 
    APInt UndefSrcElts, UndefSubElts;
    SmallVector<APInt, 32> EltSrcBits, EltSubBits;
    if (getTargetConstantBitsFromNode(Op.getOperand(1), SrcEltSizeInBits,
                                      UndefSubElts, EltSubBits,
                                      AllowWholeUndefs && AllowUndefs,
                                      AllowPartialUndefs && AllowUndefs) &&
        getTargetConstantBitsFromNode(Op.getOperand(0), SrcEltSizeInBits,
                                      UndefSrcElts, EltSrcBits,
                                      AllowWholeUndefs && AllowUndefs,
                                      AllowPartialUndefs && AllowUndefs)) {
      unsigned BaseIdx = Op.getConstantOperandVal(2);
      UndefSrcElts.insertBits(UndefSubElts, BaseIdx);
      for (unsigned i = 0, e = EltSubBits.size(); i != e; ++i)
        EltSrcBits[BaseIdx + i] = EltSubBits[i];
      return CastBitData(UndefSrcElts, EltSrcBits);
    }
  }
 
  // Extract constant bits from a subvector's source.
  if (Op.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
      getTargetConstantBitsFromNode(Op.getOperand(0), EltSizeInBits, UndefElts,
                                    EltBits, AllowWholeUndefs,
                                    AllowPartialUndefs)) {
    EVT SrcVT = Op.getOperand(0).getValueType();
    unsigned NumSrcElts = SrcVT.getSizeInBits() / EltSizeInBits;
    unsigned NumSubElts = VT.getSizeInBits() / EltSizeInBits;
    unsigned BaseOfs = Op.getConstantOperandVal(1) * VT.getScalarSizeInBits();
    unsigned BaseIdx = BaseOfs / EltSizeInBits;
    assert((SrcVT.getSizeInBits() % EltSizeInBits) == 0 &&
           (VT.getSizeInBits() % EltSizeInBits) == 0 &&
           (BaseOfs % EltSizeInBits) == 0 && "Bad subvector index");
 
    UndefElts = UndefElts.extractBits(NumSubElts, BaseIdx);
    if ((BaseIdx + NumSubElts) != NumSrcElts)
      EltBits.erase(EltBits.begin() + BaseIdx + NumSubElts, EltBits.end());
    if (BaseIdx != 0)
      EltBits.erase(EltBits.begin(), EltBits.begin() + BaseIdx);
    return true;
  }
 
  // Extract constant bits from shuffle node sources.
  if (auto *SVN = dyn_cast<ShuffleVectorSDNode>(Op)) {
    // TODO - support shuffle through bitcasts.
    if (EltSizeInBits != VT.getScalarSizeInBits())
      return false;
 
    ArrayRef<int> Mask = SVN->getMask();
    if ((!AllowWholeUndefs || !AllowPartialUndefs) &&
        llvm::any_of(Mask, [](int M) { return M < 0; }))
      return false;
 
    APInt UndefElts0, UndefElts1;
    SmallVector<APInt, 32> EltBits0, EltBits1;
    if (isAnyInRange(Mask, 0, NumElts) &&
        !getTargetConstantBitsFromNode(Op.getOperand(0), EltSizeInBits,
                                       UndefElts0, EltBits0, AllowWholeUndefs,
                                       AllowPartialUndefs))
      return false;
    if (isAnyInRange(Mask, NumElts, 2 * NumElts) &&
        !getTargetConstantBitsFromNode(Op.getOperand(1), EltSizeInBits,
                                       UndefElts1, EltBits1, AllowWholeUndefs,
                                       AllowPartialUndefs))
      return false;
 
    UndefElts = APInt::getZero(NumElts);
    for (int i = 0; i != (int)NumElts; ++i) {
      int M = Mask[i];
      if (M < 0) {
        UndefElts.setBit(i);
        EltBits.push_back(APInt::getZero(EltSizeInBits));
      } else if (M < (int)NumElts) {
        if (UndefElts0[M])
          UndefElts.setBit(i);
        EltBits.push_back(EltBits0[M]);
      } else {
        if (UndefElts1[M - NumElts])
          UndefElts.setBit(i);
        EltBits.push_back(EltBits1[M - NumElts]);
      }
    }
    return true;
  }
 
  return false;
}
 
namespace llvm {
namespace X86 {
bool isConstantSplat(SDValue Op, APInt &SplatVal, bool AllowPartialUndefs) {
  APInt UndefElts;
  SmallVector<APInt, 16> EltBits;
  if (getTargetConstantBitsFromNode(
          Op, Op.getScalarValueSizeInBits(), UndefElts, EltBits,
          /*AllowWholeUndefs*/ true, AllowPartialUndefs)) {
    int SplatIndex = -1;
    for (int i = 0, e = EltBits.size(); i != e; ++i) {
      if (UndefElts[i])
        continue;
      if (0 <= SplatIndex && EltBits[i] != EltBits[SplatIndex]) {
        SplatIndex = -1;
        break;
      }
      SplatIndex = i;
    }
    if (0 <= SplatIndex) {
      SplatVal = EltBits[SplatIndex];
      return true;
    }
  }
 
  return false;
}
 
int getRoundingModeX86(unsigned RM) {
  switch (static_cast<::llvm::RoundingMode>(RM)) {
    // clang-format off
  case ::llvm::RoundingMode::NearestTiesToEven: return X86::rmToNearest;
  case ::llvm::RoundingMode::TowardNegative:    return X86::rmDownward;
  case ::llvm::RoundingMode::TowardPositive:    return X86::rmUpward;
  case ::llvm::RoundingMode::TowardZero:        return X86::rmTowardZero;
  default:                                      return X86::rmInvalid;
    // clang-format on
  }
}
 
} // namespace X86
} // namespace llvm
 
static bool getTargetShuffleMaskIndices(SDValue MaskNode,
                                        unsigned MaskEltSizeInBits,
                                        SmallVectorImpl<uint64_t> &RawMask,
                                        APInt &UndefElts) {
  // Extract the raw target constant bits.
  SmallVector<APInt, 64> EltBits;
  if (!getTargetConstantBitsFromNode(MaskNode, MaskEltSizeInBits, UndefElts,
                                     EltBits, /* AllowWholeUndefs */ true,
                                     /* AllowPartialUndefs */ false))
    return false;
 
  // Insert the extracted elements into the mask.
  for (const APInt &Elt : EltBits)
    RawMask.push_back(Elt.getZExtValue());
 
  return true;
}
 
static bool isConstantPowerOf2(SDValue V, unsigned EltSizeInBIts,
                               bool AllowUndefs) {
  APInt UndefElts;
  SmallVector<APInt, 64> EltBits;
  if (!getTargetConstantBitsFromNode(V, EltSizeInBIts, UndefElts, EltBits,
                                     /*AllowWholeUndefs*/ AllowUndefs,
                                     /*AllowPartialUndefs*/ false))
    return false;
 
  bool IsPow2OrUndef = true;
  for (unsigned I = 0, E = EltBits.size(); I != E; ++I)
    IsPow2OrUndef &= UndefElts[I] || EltBits[I].isPowerOf2();
  return IsPow2OrUndef;
}
 
// Helper to attempt to return a cheaper, bit-inverted version of \p V.
static SDValue IsNOT(SDValue V, SelectionDAG &DAG) {
  // TODO: don't always ignore oneuse constraints.
  V = peekThroughBitcasts(V);
  EVT VT = V.getValueType();
 
  // Match not(xor X, -1) -> X.
  if (V.getOpcode() == ISD::XOR &&
      (ISD::isBuildVectorAllOnes(V.getOperand(1).getNode()) ||
       isAllOnesConstant(V.getOperand(1))))
    return V.getOperand(0);
 
  // Match not(extract_subvector(not(X)) -> extract_subvector(X).
  if (V.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
      (isNullConstant(V.getOperand(1)) || V.getOperand(0).hasOneUse())) {
    if (SDValue Not = IsNOT(V.getOperand(0), DAG)) {
      Not = DAG.getBitcast(V.getOperand(0).getValueType(), Not);
      return DAG.getNode(ISD::EXTRACT_SUBVECTOR, SDLoc(Not), VT, Not,
                         V.getOperand(1));
    }
  }
 
  // Match not(pcmpgt(C, X)) -> pcmpgt(X, C - 1).
  if (V.getOpcode() == X86ISD::PCMPGT &&
      !ISD::isBuildVectorAllZeros(V.getOperand(0).getNode()) &&
      !ISD::isBuildVectorAllOnes(V.getOperand(0).getNode()) &&
      V.getOperand(0).hasOneUse()) {
    APInt UndefElts;
    SmallVector<APInt> EltBits;
    if (getTargetConstantBitsFromNode(V.getOperand(0),
                                      V.getScalarValueSizeInBits(), UndefElts,
                                      EltBits) &&
        !ISD::isBuildVectorOfConstantSDNodes(V.getOperand(1).getNode())) {
      // Don't fold min_signed_value -> (min_signed_value - 1)
      bool MinSigned = false;
      for (APInt &Elt : EltBits) {
        MinSigned |= Elt.isMinSignedValue();
        Elt -= 1;
      }
      if (!MinSigned) {
        SDLoc DL(V);
        MVT VT = V.getSimpleValueType();
        return DAG.getNode(X86ISD::PCMPGT, DL, VT, V.getOperand(1),
                           getConstVector(EltBits, UndefElts, VT, DAG, DL));
      }
    }
  }
 
  // Match not(concat_vectors(not(X), not(Y))) -> concat_vectors(X, Y).
  SmallVector<SDValue, 2> CatOps;
  if (collectConcatOps(V.getNode(), CatOps, DAG)) {
    for (SDValue &CatOp : CatOps) {
      SDValue NotCat = IsNOT(CatOp, DAG);
      if (!NotCat)
        return SDValue();
      CatOp = DAG.getBitcast(CatOp.getValueType(), NotCat);
    }
    return DAG.getNode(ISD::CONCAT_VECTORS, SDLoc(V), VT, CatOps);
  }
 
  // Match not(or(not(X),not(Y))) -> and(X, Y).
  if (V.getOpcode() == ISD::OR && DAG.getTargetLoweringInfo().isTypeLegal(VT) &&
      V.getOperand(0).hasOneUse() && V.getOperand(1).hasOneUse()) {
    // TODO: Handle cases with single NOT operand -> ANDNP
    if (SDValue Op1 = IsNOT(V.getOperand(1), DAG))
      if (SDValue Op0 = IsNOT(V.getOperand(0), DAG))
        return DAG.getNode(ISD::AND, SDLoc(V), VT, DAG.getBitcast(VT, Op0),
                           DAG.getBitcast(VT, Op1));
  }
 
  return SDValue();
}
 
/// Create a shuffle mask that matches the PACKSS/PACKUS truncation.
/// A multi-stage pack shuffle mask is created by specifying NumStages > 1.
/// Note: This ignores saturation, so inputs must be checked first.
static void createPackShuffleMask(MVT VT, SmallVectorImpl<int> &Mask,
                                  bool Unary, unsigned NumStages = 1) {
  assert(Mask.empty() && "Expected an empty shuffle mask vector");
  unsigned NumElts = VT.getVectorNumElements();
  unsigned NumLanes = VT.getSizeInBits() / 128;
  unsigned NumEltsPerLane = 128 / VT.getScalarSizeInBits();
  unsigned Offset = Unary ? 0 : NumElts;
  unsigned Repetitions = 1u << (NumStages - 1);
  unsigned Increment = 1u << NumStages;
  assert((NumEltsPerLane >> NumStages) > 0 && "Illegal packing compaction");
 
  for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
    for (unsigned Stage = 0; Stage != Repetitions; ++Stage) {
      for (unsigned Elt = 0; Elt != NumEltsPerLane; Elt += Increment)
        Mask.push_back(Elt + (Lane * NumEltsPerLane));
      for (unsigned Elt = 0; Elt != NumEltsPerLane; Elt += Increment)
        Mask.push_back(Elt + (Lane * NumEltsPerLane) + Offset);
    }
  }
}
 
// Split the demanded elts of a PACKSS/PACKUS node between its operands.
static void getPackDemandedElts(EVT VT, const APInt &DemandedElts,
                                APInt &DemandedLHS, APInt &DemandedRHS) {
  int NumLanes = VT.getSizeInBits() / 128;
  int NumElts = DemandedElts.getBitWidth();
  int NumInnerElts = NumElts / 2;
  int NumEltsPerLane = NumElts / NumLanes;
  int NumInnerEltsPerLane = NumInnerElts / NumLanes;
 
  DemandedLHS = APInt::getZero(NumInnerElts);
  DemandedRHS = APInt::getZero(NumInnerElts);
 
  // Map DemandedElts to the packed operands.
  for (int Lane = 0; Lane != NumLanes; ++Lane) {
    for (int Elt = 0; Elt != NumInnerEltsPerLane; ++Elt) {
      int OuterIdx = (Lane * NumEltsPerLane) + Elt;
      int InnerIdx = (Lane * NumInnerEltsPerLane) + Elt;
      if (DemandedElts[OuterIdx])
        DemandedLHS.setBit(InnerIdx);
      if (DemandedElts[OuterIdx + NumInnerEltsPerLane])
        DemandedRHS.setBit(InnerIdx);
    }
  }
}
 
// Split the demanded elts of a HADD/HSUB node between its operands.
static void getHorizDemandedElts(EVT VT, const APInt &DemandedElts,
                                 APInt &DemandedLHS, APInt &DemandedRHS) {
  getHorizDemandedEltsForFirstOperand(VT.getSizeInBits(), DemandedElts,
                                      DemandedLHS, DemandedRHS);
  DemandedLHS |= DemandedLHS << 1;
  DemandedRHS |= DemandedRHS << 1;
}
 
/// Calculates the shuffle mask corresponding to the target-specific opcode.
/// If the mask could be calculated, returns it in \p Mask, returns the shuffle
/// operands in \p Ops, and returns true.
/// Sets \p IsUnary to true if only one source is used. Note that this will set
/// IsUnary for shuffles which use a single input multiple times, and in those
/// cases it will adjust the mask to only have indices within that single input.
/// It is an error to call this with non-empty Mask/Ops vectors.
static bool getTargetShuffleMask(SDValue N, bool AllowSentinelZero,
                                 SmallVectorImpl<SDValue> &Ops,
                                 SmallVectorImpl<int> &Mask, bool &IsUnary) {
  if (!isTargetShuffle(N.getOpcode()))
    return false;
 
  MVT VT = N.getSimpleValueType();
  unsigned NumElems = VT.getVectorNumElements();
  unsigned MaskEltSize = VT.getScalarSizeInBits();
  SmallVector<uint64_t, 32> RawMask;
  APInt RawUndefs;
  uint64_t ImmN;
 
  assert(Mask.empty() && "getTargetShuffleMask expects an empty Mask vector");
  assert(Ops.empty() && "getTargetShuffleMask expects an empty Ops vector");
 
  IsUnary = false;
  bool IsFakeUnary = false;
  switch (N.getOpcode()) {
  case X86ISD::BLENDI:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodeBLENDMask(NumElems, ImmN, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::SHUFP:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodeSHUFPMask(NumElems, MaskEltSize, ImmN, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::INSERTPS:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodeINSERTPSMask(ImmN, Mask, /*SrcIsMem=*/false);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::EXTRQI:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    if (isa<ConstantSDNode>(N.getOperand(1)) &&
        isa<ConstantSDNode>(N.getOperand(2))) {
      int BitLen = N.getConstantOperandVal(1);
      int BitIdx = N.getConstantOperandVal(2);
      DecodeEXTRQIMask(NumElems, MaskEltSize, BitLen, BitIdx, Mask);
      IsUnary = true;
    }
    break;
  case X86ISD::INSERTQI:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    if (isa<ConstantSDNode>(N.getOperand(2)) &&
        isa<ConstantSDNode>(N.getOperand(3))) {
      int BitLen = N.getConstantOperandVal(2);
      int BitIdx = N.getConstantOperandVal(3);
      DecodeINSERTQIMask(NumElems, MaskEltSize, BitLen, BitIdx, Mask);
      IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    }
    break;
  case X86ISD::UNPCKH:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    DecodeUNPCKHMask(NumElems, MaskEltSize, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::UNPCKL:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    DecodeUNPCKLMask(NumElems, MaskEltSize, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::MOVHLPS:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    DecodeMOVHLPSMask(NumElems, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::MOVLHPS:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    DecodeMOVLHPSMask(NumElems, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::VALIGN:
    assert((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::i64) &&
           "Only 32-bit and 64-bit elements are supported!");
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodeVALIGNMask(NumElems, ImmN, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    Ops.push_back(N.getOperand(1));
    Ops.push_back(N.getOperand(0));
    break;
  case X86ISD::PALIGNR:
    assert(VT.getScalarType() == MVT::i8 && "Byte vector expected");
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodePALIGNRMask(NumElems, ImmN, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    Ops.push_back(N.getOperand(1));
    Ops.push_back(N.getOperand(0));
    break;
  case X86ISD::VSHLDQ:
    assert(VT.getScalarType() == MVT::i8 && "Byte vector expected");
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodePSLLDQMask(NumElems, ImmN, Mask);
    IsUnary = true;
    break;
  case X86ISD::VSRLDQ:
    assert(VT.getScalarType() == MVT::i8 && "Byte vector expected");
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodePSRLDQMask(NumElems, ImmN, Mask);
    IsUnary = true;
    break;
  case X86ISD::PSHUFD:
  case X86ISD::VPERMILPI:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodePSHUFMask(NumElems, MaskEltSize, ImmN, Mask);
    IsUnary = true;
    break;
  case X86ISD::PSHUFHW:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodePSHUFHWMask(NumElems, ImmN, Mask);
    IsUnary = true;
    break;
  case X86ISD::PSHUFLW:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodePSHUFLWMask(NumElems, ImmN, Mask);
    IsUnary = true;
    break;
  case X86ISD::VZEXT_MOVL:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    DecodeZeroMoveLowMask(NumElems, Mask);
    IsUnary = true;
    break;
  case X86ISD::VBROADCAST:
    // We only decode broadcasts of same-sized vectors, peeking through to
    // extracted subvectors is likely to cause hasOneUse issues with
    // SimplifyDemandedBits etc.
    if (N.getOperand(0).getValueType() == VT) {
      DecodeVectorBroadcast(NumElems, Mask);
      IsUnary = true;
      break;
    }
    return false;
  case X86ISD::VPERMILPV: {
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    IsUnary = true;
    SDValue MaskNode = N.getOperand(1);
    if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask,
                                    RawUndefs)) {
      DecodeVPERMILPMask(NumElems, MaskEltSize, RawMask, RawUndefs, Mask);
      break;
    }
    return false;
  }
  case X86ISD::PSHUFB: {
    assert(VT.getScalarType() == MVT::i8 && "Byte vector expected");
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    IsUnary = true;
    SDValue MaskNode = N.getOperand(1);
    if (getTargetShuffleMaskIndices(MaskNode, 8, RawMask, RawUndefs)) {
      DecodePSHUFBMask(RawMask, RawUndefs, Mask);
      break;
    }
    return false;
  }
  case X86ISD::VPERMI:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodeVPERMMask(NumElems, ImmN, Mask);
    IsUnary = true;
    break;
  case X86ISD::MOVSS:
  case X86ISD::MOVSD:
  case X86ISD::MOVSH:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    DecodeScalarMoveMask(NumElems, /* IsLoad */ false, Mask);
    break;
  case X86ISD::VPERM2X128:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    DecodeVPERM2X128Mask(NumElems, ImmN, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::SHUF128:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    ImmN = N.getConstantOperandVal(N.getNumOperands() - 1);
    decodeVSHUF64x2FamilyMask(NumElems, MaskEltSize, ImmN, Mask);
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    break;
  case X86ISD::MOVSLDUP:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    DecodeMOVSLDUPMask(NumElems, Mask);
    IsUnary = true;
    break;
  case X86ISD::MOVSHDUP:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    DecodeMOVSHDUPMask(NumElems, Mask);
    IsUnary = true;
    break;
  case X86ISD::MOVDDUP:
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    DecodeMOVDDUPMask(NumElems, Mask);
    IsUnary = true;
    break;
  case X86ISD::VPERMIL2: {
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    SDValue MaskNode = N.getOperand(2);
    SDValue CtrlNode = N.getOperand(3);
    if (ConstantSDNode *CtrlOp = dyn_cast<ConstantSDNode>(CtrlNode)) {
      unsigned CtrlImm = CtrlOp->getZExtValue();
      if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask,
                                      RawUndefs)) {
        DecodeVPERMIL2PMask(NumElems, MaskEltSize, CtrlImm, RawMask, RawUndefs,
                            Mask);
        break;
      }
    }
    return false;
  }
  case X86ISD::VPPERM: {
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(1);
    SDValue MaskNode = N.getOperand(2);
    if (getTargetShuffleMaskIndices(MaskNode, 8, RawMask, RawUndefs)) {
      DecodeVPPERMMask(RawMask, RawUndefs, Mask);
      break;
    }
    return false;
  }
  case X86ISD::VPERMV: {
    assert(N.getOperand(1).getValueType() == VT && "Unexpected value type");
    IsUnary = true;
    // Unlike most shuffle nodes, VPERMV's mask operand is operand 0.
    Ops.push_back(N.getOperand(1));
    SDValue MaskNode = N.getOperand(0);
    if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask,
                                    RawUndefs)) {
      DecodeVPERMVMask(RawMask, RawUndefs, Mask);
      break;
    }
    return false;
  }
  case X86ISD::VPERMV3: {
    assert(N.getOperand(0).getValueType() == VT && "Unexpected value type");
    assert(N.getOperand(2).getValueType() == VT && "Unexpected value type");
    IsUnary = IsFakeUnary = N.getOperand(0) == N.getOperand(2);
    // Unlike most shuffle nodes, VPERMV3's mask operand is the middle one.
    Ops.push_back(N.getOperand(0));
    Ops.push_back(N.getOperand(2));
    SDValue MaskNode = N.getOperand(1);
    if (getTargetShuffleMaskIndices(MaskNode, MaskEltSize, RawMask,
                                    RawUndefs)) {
      DecodeVPERMV3Mask(RawMask, RawUndefs, Mask);
      break;
    }
    return false;
  }
  case X86ISD::COMPRESS: {
    SDValue CmpVec = N.getOperand(0);
    SDValue PassThru = N.getOperand(1);
    SDValue CmpMask = N.getOperand(2);
    APInt UndefElts;
    SmallVector<APInt> EltBits;
    if (!getTargetConstantBitsFromNode(CmpMask, 1, UndefElts, EltBits))
      return false;
    assert(UndefElts.getBitWidth() == NumElems && EltBits.size() == NumElems &&
           "Illegal compression mask");
    for (unsigned I = 0; I != NumElems; ++I) {
      if (!EltBits[I].isZero())
        Mask.push_back(I);
    }
    while (Mask.size() != NumElems) {
      Mask.push_back(NumElems + Mask.size());
    }
    Ops.push_back(CmpVec);
    Ops.push_back(PassThru);
    return true;
  }
  case X86ISD::EXPAND: {
    SDValue ExpVec = N.getOperand(0);
    SDValue PassThru = N.getOperand(1);
    SDValue ExpMask = N.getOperand(2);
    APInt UndefElts;
    SmallVector<APInt> EltBits;
    if (!getTargetConstantBitsFromNode(ExpMask, 1, UndefElts, EltBits))
      return false;
    assert(UndefElts.getBitWidth() == NumElems && EltBits.size() == NumElems &&
           "Illegal expansion mask");
    unsigned ExpIndex = 0;
    for (unsigned I = 0; I != NumElems; ++I) {
      if (EltBits[I].isZero())
        Mask.push_back(I + NumElems);
      else
        Mask.push_back(ExpIndex++);
    }
    Ops.push_back(ExpVec);
    Ops.push_back(PassThru);
    return true;
  }
  default:
    llvm_unreachable("unknown target shuffle node");
  }
 
  // Empty mask indicates the decode failed.
  if (Mask.empty())
    return false;
 
  // Check if we're getting a shuffle mask with zero'd elements.
  if (!AllowSentinelZero && isAnyZero(Mask))
    return false;
 
  // If we have a fake unary shuffle, the shuffle mask is spread across two
  // inputs that are actually the same node. Re-map the mask to always point
  // into the first input.
  if (IsFakeUnary)
    for (int &M : Mask)
      if (M >= (int)Mask.size())
        M -= Mask.size();
 
  // If we didn't already add operands in the opcode-specific code, default to
  // adding 1 or 2 operands starting at 0.
  if (Ops.empty()) {
    Ops.push_back(N.getOperand(0));
    if (!IsUnary || IsFakeUnary)
      Ops.push_back(N.getOperand(1));
  }
 
  return true;
}
 
// Wrapper for getTargetShuffleMask with InUnary;
static bool getTargetShuffleMask(SDValue N, bool AllowSentinelZero,
                                 SmallVectorImpl<SDValue> &Ops,
                                 SmallVectorImpl<int> &Mask) {
  bool IsUnary;
  return getTargetShuffleMask(N, AllowSentinelZero, Ops, Mask, IsUnary);
}
 
/// Compute whether each element of a shuffle is zeroable.
///
/// A "zeroable" vector shuffle element is one which can be lowered to zero.
/// Either it is an undef element in the shuffle mask, the element of the input
/// referenced is undef, or the element of the input referenced is known to be
/// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
/// as many lanes with this technique as possible to simplify the remaining
/// shuffle.
static void computeZeroableShuffleElements(ArrayRef<int> Mask,
                                           SDValue V1, SDValue V2,
                                           APInt &KnownUndef, APInt &KnownZero) {
  int Size = Mask.size();
  KnownUndef = KnownZero = APInt::getZero(Size);
 
  V1 = peekThroughBitcasts(V1);
  V2 = peekThroughBitcasts(V2);
 
  bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
  bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
 
  int VectorSizeInBits = V1.getValueSizeInBits();
  int ScalarSizeInBits = VectorSizeInBits / Size;
  assert(!(VectorSizeInBits % ScalarSizeInBits) && "Illegal shuffle mask size");
 
  for (int i = 0; i < Size; ++i) {
    int M = Mask[i];
    // Handle the easy cases.
    if (M < 0) {
      KnownUndef.setBit(i);
      continue;
    }
    if ((M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
      KnownZero.setBit(i);
      continue;
    }
 
    // Determine shuffle input and normalize the mask.
    SDValue V = M < Size ? V1 : V2;
    M %= Size;
 
    // Currently we can only search BUILD_VECTOR for UNDEF/ZERO elements.
    if (V.getOpcode() != ISD::BUILD_VECTOR)
      continue;
 
    // If the BUILD_VECTOR has fewer elements then the bitcasted portion of
    // the (larger) source element must be UNDEF/ZERO.
    if ((Size % V.getNumOperands()) == 0) {
      int Scale = Size / V->getNumOperands();
      SDValue Op = V.getOperand(M / Scale);
      if (Op.isUndef())
        KnownUndef.setBit(i);
      if (X86::isZeroNode(Op))
        KnownZero.setBit(i);
      else if (ConstantSDNode *Cst = dyn_cast<ConstantSDNode>(Op)) {
        APInt Val = Cst->getAPIntValue();
        Val = Val.extractBits(ScalarSizeInBits, (M % Scale) * ScalarSizeInBits);
        if (Val == 0)
          KnownZero.setBit(i);
      } else if (ConstantFPSDNode *Cst = dyn_cast<ConstantFPSDNode>(Op)) {
        APInt Val = Cst->getValueAPF().bitcastToAPInt();
        Val = Val.extractBits(ScalarSizeInBits, (M % Scale) * ScalarSizeInBits);
        if (Val == 0)
          KnownZero.setBit(i);
      }
      continue;
    }
 
    // If the BUILD_VECTOR has more elements then all the (smaller) source
    // elements must be UNDEF or ZERO.
    if ((V.getNumOperands() % Size) == 0) {
      int Scale = V->getNumOperands() / Size;
      bool AllUndef = true;
      bool AllZero = true;
      for (int j = 0; j < Scale; ++j) {
        SDValue Op = V.getOperand((M * Scale) + j);
        AllUndef &= Op.isUndef();
        AllZero &= X86::isZeroNode(Op);
      }
      if (AllUndef)
        KnownUndef.setBit(i);
      if (AllZero)
        KnownZero.setBit(i);
      continue;
    }
  }
}
 
/// Decode a target shuffle mask and inputs and see if any values are
/// known to be undef or zero from their inputs.
/// Returns true if the target shuffle mask was decoded.
/// FIXME: Merge this with computeZeroableShuffleElements?
static bool getTargetShuffleAndZeroables(SDValue N, SmallVectorImpl<int> &Mask,
                                         SmallVectorImpl<SDValue> &Ops,
                                         APInt &KnownUndef, APInt &KnownZero) {
  bool IsUnary;
  if (!isTargetShuffle(N.getOpcode()))
    return false;
 
  MVT VT = N.getSimpleValueType();
  if (!getTargetShuffleMask(N, true, Ops, Mask, IsUnary))
    return false;
 
  int Size = Mask.size();
  SDValue V1 = Ops[0];
  SDValue V2 = IsUnary ? V1 : Ops[1];
  KnownUndef = KnownZero = APInt::getZero(Size);
 
  V1 = peekThroughBitcasts(V1);
  V2 = peekThroughBitcasts(V2);
 
  assert((VT.getSizeInBits() % Size) == 0 &&
         "Illegal split of shuffle value type");
  unsigned EltSizeInBits = VT.getSizeInBits() / Size;
 
  // Extract known constant input data.
  APInt UndefSrcElts[2];
  SmallVector<APInt, 32> SrcEltBits[2];
  bool IsSrcConstant[2] = {
      getTargetConstantBitsFromNode(V1, EltSizeInBits, UndefSrcElts[0],
                                    SrcEltBits[0], /*AllowWholeUndefs*/ true,
                                    /*AllowPartialUndefs*/ false),
      getTargetConstantBitsFromNode(V2, EltSizeInBits, UndefSrcElts[1],
                                    SrcEltBits[1], /*AllowWholeUndefs*/ true,
                                    /*AllowPartialUndefs*/ false)};
 
  for (int i = 0; i < Size; ++i) {
    int M = Mask[i];
 
    // Already decoded as SM_SentinelZero / SM_SentinelUndef.
    if (M < 0) {
      assert(isUndefOrZero(M) && "Unknown shuffle sentinel value!");
      if (SM_SentinelUndef == M)
        KnownUndef.setBit(i);
      if (SM_SentinelZero == M)
        KnownZero.setBit(i);
      continue;
    }
 
    // Determine shuffle input and normalize the mask.
    unsigned SrcIdx = M / Size;
    SDValue V = M < Size ? V1 : V2;
    M %= Size;
 
    // We are referencing an UNDEF input.
    if (V.isUndef()) {
      KnownUndef.setBit(i);
      continue;
    }
 
    // SCALAR_TO_VECTOR - only the first element is defined, and the rest UNDEF.
    // TODO: We currently only set UNDEF for integer types - floats use the same
    // registers as vectors and many of the scalar folded loads rely on the
    // SCALAR_TO_VECTOR pattern.
    if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
        (Size % V.getValueType().getVectorNumElements()) == 0) {
      int Scale = Size / V.getValueType().getVectorNumElements();
      int Idx = M / Scale;
      if (Idx != 0 && !VT.isFloatingPoint())
        KnownUndef.setBit(i);
      else if (Idx == 0 && X86::isZeroNode(V.getOperand(0)))
        KnownZero.setBit(i);
      continue;
    }
 
    // INSERT_SUBVECTOR - to widen vectors we often insert them into UNDEF
    // base vectors.
    if (V.getOpcode() == ISD::INSERT_SUBVECTOR) {
      SDValue Vec = V.getOperand(0);
      int NumVecElts = Vec.getValueType().getVectorNumElements();
      if (Vec.isUndef() && Size == NumVecElts) {
        int Idx = V.getConstantOperandVal(2);
        int NumSubElts = V.getOperand(1).getValueType().getVectorNumElements();
        if (M < Idx || (Idx + NumSubElts) <= M)
          KnownUndef.setBit(i);
      }
      continue;
    }
 
    // Attempt to extract from the source's constant bits.
    if (IsSrcConstant[SrcIdx]) {
      if (UndefSrcElts[SrcIdx][M])
        KnownUndef.setBit(i);
      else if (SrcEltBits[SrcIdx][M] == 0)
        KnownZero.setBit(i);
    }
  }
 
  assert(VT.getVectorNumElements() == (unsigned)Size &&
         "Different mask size from vector size!");
  return true;
}
 
// Replace target shuffle mask elements with known undef/zero sentinels.
static void resolveTargetShuffleFromZeroables(SmallVectorImpl<int> &Mask,
                                              const APInt &KnownUndef,
                                              const APInt &KnownZero,
                                              bool ResolveKnownZeros= true) {
  unsigned NumElts = Mask.size();
  assert(KnownUndef.getBitWidth() == NumElts &&
         KnownZero.getBitWidth() == NumElts && "Shuffle mask size mismatch");
 
  for (unsigned i = 0; i != NumElts; ++i) {
    if (KnownUndef[i])
      Mask[i] = SM_SentinelUndef;
    else if (ResolveKnownZeros && KnownZero[i])
      Mask[i] = SM_SentinelZero;
  }
}
 
// Extract target shuffle mask sentinel elements to known undef/zero bitmasks.
static void resolveZeroablesFromTargetShuffle(const SmallVectorImpl<int> &Mask,
                                              APInt &KnownUndef,
                                              APInt &KnownZero) {
  unsigned NumElts = Mask.size();
  KnownUndef = KnownZero = APInt::getZero(NumElts);
 
  for (unsigned i = 0; i != NumElts; ++i) {
    int M = Mask[i];
    if (SM_SentinelUndef == M)
      KnownUndef.setBit(i);
    if (SM_SentinelZero == M)
      KnownZero.setBit(i);
  }
}
 
// Attempt to create a shuffle mask from a VSELECT/BLENDV condition mask.
static bool createShuffleMaskFromVSELECT(SmallVectorImpl<int> &Mask,
                                         SDValue Cond, bool IsBLENDV = false) {
  EVT CondVT = Cond.getValueType();
  unsigned EltSizeInBits = CondVT.getScalarSizeInBits();
  unsigned NumElts = CondVT.getVectorNumElements();
 
  APInt UndefElts;
  SmallVector<APInt, 32> EltBits;
  if (!getTargetConstantBitsFromNode(Cond, EltSizeInBits, UndefElts, EltBits,
                                     /*AllowWholeUndefs*/ true,
                                     /*AllowPartialUndefs*/ false))
    return false;
 
  Mask.resize(NumElts, SM_SentinelUndef);
 
  for (int i = 0; i != (int)NumElts; ++i) {
    Mask[i] = i;
    // Arbitrarily choose from the 2nd operand if the select condition element
    // is undef.
    // TODO: Can we do better by matching patterns such as even/odd?
    if (UndefElts[i] || (!IsBLENDV && EltBits[i].isZero()) ||
        (IsBLENDV && EltBits[i].isNonNegative()))
      Mask[i] += NumElts;
  }
 
  return true;
}
 
// Forward declaration (for getFauxShuffleMask recursive check).
static bool getTargetShuffleInputs(SDValue Op, const APInt &DemandedElts,
                                   SmallVectorImpl<SDValue> &Inputs,
                                   SmallVectorImpl<int> &Mask,
                                   const SelectionDAG &DAG, unsigned Depth,
                                   bool ResolveKnownElts);
 
// Attempt to decode ops that could be represented as a shuffle mask.
// The decoded shuffle mask may contain a different number of elements to the
// destination value type.
// TODO: Merge into getTargetShuffleInputs()
static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
                               SmallVectorImpl<int> &Mask,
                               SmallVectorImpl<SDValue> &Ops,
                               const SelectionDAG &DAG, unsigned Depth,
                               bool ResolveKnownElts) {
  Mask.clear();
  Ops.clear();
 
  MVT VT = N.getSimpleValueType();
  unsigned NumElts = VT.getVectorNumElements();
  unsigned NumSizeInBits = VT.getSizeInBits();
  unsigned NumBitsPerElt = VT.getScalarSizeInBits();
  if ((NumBitsPerElt % 8) != 0 || (NumSizeInBits % 8) != 0)
    return false;
  assert(NumElts == DemandedElts.getBitWidth() && "Unexpected vector size");
  unsigned NumSizeInBytes = NumSizeInBits / 8;
  unsigned NumBytesPerElt = NumBitsPerElt / 8;
 
  unsigned Opcode = N.getOpcode();
  switch (Opcode) {
  case ISD::VECTOR_SHUFFLE: {
    // Don't treat ISD::VECTOR_SHUFFLE as a target shuffle so decode it here.
    ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(N)->getMask();
    if (isUndefOrInRange(ShuffleMask, 0, 2 * NumElts)) {
      Mask.append(ShuffleMask.begin(), ShuffleMask.end());
      Ops.push_back(N.getOperand(0));
      Ops.push_back(N.getOperand(1));
      return true;
    }
    return false;
  }
  case ISD::AND:
  case X86ISD::ANDNP: {
    // Attempt to decode as a per-byte mask.
    APInt UndefElts;
    SmallVector<APInt, 32> EltBits;
    SDValue N0 = N.getOperand(0);
    SDValue N1 = N.getOperand(1);
    bool IsAndN = (X86ISD::ANDNP == Opcode);
    uint64_t ZeroMask = IsAndN ? 255 : 0;
    if (!getTargetConstantBitsFromNode(IsAndN ? N0 : N1, 8, UndefElts, EltBits,
                                       /*AllowWholeUndefs*/ false,
                                       /*AllowPartialUndefs*/ false))
      return false;
    // We can't assume an undef src element gives an undef dst - the other src
    // might be zero.
    assert(UndefElts.isZero() && "Unexpected UNDEF element in AND/ANDNP mask");
    for (int i = 0, e = (int)EltBits.size(); i != e; ++i) {
      const APInt &ByteBits = EltBits[i];
      if (ByteBits != 0 && ByteBits != 255)
        return false;
      Mask.push_back(ByteBits == ZeroMask ? SM_SentinelZero : i);
    }
    Ops.push_back(IsAndN ? N1 : N0);
    return true;
  }
  case ISD::OR: {
    // Handle OR(SHUFFLE,SHUFFLE) case where one source is zero and the other
    // is a valid shuffle index.
    SDValue N0 = peekThroughBitcasts(N.getOperand(0));
    SDValue N1 = peekThroughBitcasts(N.getOperand(1));
    if (!N0.getValueType().isVector() || !N1.getValueType().isVector())
      return false;
 
    SmallVector<int, 64> SrcMask0, SrcMask1;
    SmallVector<SDValue, 2> SrcInputs0, SrcInputs1;
    APInt Demand0 = APInt::getAllOnes(N0.getValueType().getVectorNumElements());
    APInt Demand1 = APInt::getAllOnes(N1.getValueType().getVectorNumElements());
    if (!getTargetShuffleInputs(N0, Demand0, SrcInputs0, SrcMask0, DAG,
                                Depth + 1, true) ||
        !getTargetShuffleInputs(N1, Demand1, SrcInputs1, SrcMask1, DAG,
                                Depth + 1, true))
      return false;
 
    size_t MaskSize = std::max(SrcMask0.size(), SrcMask1.size());
    SmallVector<int, 64> Mask0, Mask1;
    narrowShuffleMaskElts(MaskSize / SrcMask0.size(), SrcMask0, Mask0);
    narrowShuffleMaskElts(MaskSize / SrcMask1.size(), SrcMask1, Mask1);
    for (int i = 0; i != (int)MaskSize; ++i) {
      // NOTE: Don't handle demanded SM_SentinelUndef, as we can end up in
      // infinite loops converting between OR and BLEND shuffles due to
      // canWidenShuffleElements merging away undef elements, meaning we
      // fail to recognise the OR as the undef element isn't known zero.
      if (Mask0[i] == SM_SentinelZero && Mask1[i] == SM_SentinelZero)
        Mask.push_back(SM_SentinelZero);
      else if (Mask1[i] == SM_SentinelZero)
        Mask.push_back(i);
      else if (Mask0[i] == SM_SentinelZero)
        Mask.push_back(i + MaskSize);
      else if (MaskSize == NumElts && !DemandedElts[i])
        Mask.push_back(SM_SentinelUndef);
      else
        return false;
    }
    Ops.push_back(N.getOperand(0));
    Ops.push_back(N.getOperand(1));
    return true;
  }
  case ISD::CONCAT_VECTORS: {
    // Limit this to vXi64 vector cases to make the most of cross lane shuffles.
    unsigned NumSubElts = N.getOperand(0).getValueType().getVectorNumElements();
    if (NumBitsPerElt == 64) {
      for (unsigned I = 0, E = N.getNumOperands(); I != E; ++I) {
        for (unsigned M = 0; M != NumSubElts; ++M)
          Mask.push_back((I * NumElts) + M);
        Ops.push_back(N.getOperand(I));
      }
      return true;
    }
    return false;
  }
  case ISD::INSERT_SUBVECTOR: {
    SDValue Src = N.getOperand(0);
    SDValue Sub = N.getOperand(1);
    EVT SubVT = Sub.getValueType();
    unsigned NumSubElts = SubVT.getVectorNumElements();
    uint64_t InsertIdx = N.getConstantOperandVal(2);
    // Subvector isn't demanded - just return the base vector.
    if (DemandedElts.extractBits(NumSubElts, InsertIdx) == 0) {
      Mask.resize(NumElts);
      std::iota(Mask.begin(), Mask.end(), 0);
      Ops.push_back(Src);
      return true;
    }
    // Handle CONCAT(SUB0, SUB1).
    // Limit to vXi64/splat cases to make the most of cross lane shuffles.
    if (Depth > 0 && InsertIdx == NumSubElts && NumElts == (2 * NumSubElts) &&
        Src.getOpcode() == ISD::INSERT_SUBVECTOR &&
        Src.getOperand(0).isUndef() &&
        Src.getOperand(1).getValueType() == SubVT &&
        Src.getConstantOperandVal(2) == 0 &&
        (NumBitsPerElt == 64 || Src.getOperand(1) == Sub) &&
        SDNode::areOnlyUsersOf({N.getNode(), Src.getNode()}, Sub.getNode())) {
      Mask.resize(NumElts);
      std::iota(Mask.begin(), Mask.begin() + NumSubElts, 0);
      std::iota(Mask.begin() + NumSubElts, Mask.end(), NumElts);
      Ops.push_back(Src.getOperand(1));
      Ops.push_back(Sub);
      return true;
    }
    // Handle INSERT_SUBVECTOR(UNDEF, SUB, IDX) iff IDX != 0
    if (InsertIdx != 0 && Src.isUndef() &&
        peekThroughBitcasts(Sub).getOpcode() != ISD::EXTRACT_SUBVECTOR) {
      Mask.assign(NumElts, SM_SentinelUndef);
      std::iota(Mask.begin() + InsertIdx, Mask.begin() + InsertIdx + NumSubElts,
                0);
      Ops.push_back(Sub);
      return true;
    }
    if (!N->isOnlyUserOf(Sub.getNode()))
      return false;
 
    SmallVector<int, 64> SubMask;
    SmallVector<SDValue, 2> SubInputs;
    SDValue SubSrc = peekThroughOneUseBitcasts(Sub);
    EVT SubSrcVT = SubSrc.getValueType();
    if (!SubSrcVT.isVector())
      return false;
 
    // Handle INSERT_SUBVECTOR(SRC0, EXTRACT_SUBVECTOR(SRC1)).
    if (SubSrc.getOpcode() == ISD::EXTRACT_SUBVECTOR &&
        SubSrc.getOperand(0).getValueSizeInBits() == NumSizeInBits) {
      uint64_t ExtractIdx = SubSrc.getConstantOperandVal(1);
      SDValue SubSrcSrc = SubSrc.getOperand(0);
      unsigned NumSubSrcSrcElts =
          SubSrcSrc.getValueType().getVectorNumElements();
      unsigned MaxElts = std::max(NumElts, NumSubSrcSrcElts);
      assert((MaxElts % NumElts) == 0 && (MaxElts % NumSubSrcSrcElts) == 0 &&
             "Subvector valuetype mismatch");
      InsertIdx *= (MaxElts / NumElts);
      ExtractIdx *= (MaxElts / NumSubSrcSrcElts);
      NumSubElts *= (MaxElts / NumElts);
      bool SrcIsUndef = Src.isUndef();
      for (int i = 0; i != (int)MaxElts; ++i)
        Mask.push_back(SrcIsUndef ? SM_SentinelUndef : i);
      for (int i = 0; i != (int)NumSubElts; ++i)
        Mask[InsertIdx + i] = (SrcIsUndef ? 0 : MaxElts) + ExtractIdx + i;
      if (!SrcIsUndef)
        Ops.push_back(Src);
      Ops.push_back(SubSrcSrc);
      return true;
    }
 
    // Handle INSERT_SUBVECTOR(SRC0, SHUFFLE(SRC1)).
    APInt SubDemand = APInt::getAllOnes(SubSrcVT.getVectorNumElements());
    if (!getTargetShuffleInputs(SubSrc, SubDemand, SubInputs, SubMask, DAG,
                                Depth + 1, ResolveKnownElts))
      return false;
 
    // Subvector shuffle inputs must not be larger than the subvector.
    if (llvm::any_of(SubInputs, [SubVT](SDValue SubInput) {
          return SubVT.getFixedSizeInBits() <
                 SubInput.getValueSizeInBits().getFixedValue();
        }))
      return false;
 
    if (SubMask.size() != NumSubElts) {
      assert(((SubMask.size() % NumSubElts) == 0 ||
              (NumSubElts % SubMask.size()) == 0) &&
             "Illegal submask scale");
      if ((NumSubElts % SubMask.size()) == 0) {
        int Scale = NumSubElts / SubMask.size();
        SmallVector<int, 64> ScaledSubMask;
        narrowShuffleMaskElts(Scale, SubMask, ScaledSubMask);
        SubMask = ScaledSubMask;
      } else {
        int Scale = SubMask.size() / NumSubElts;
        NumSubElts = SubMask.size();
        NumElts *= Scale;
        InsertIdx *= Scale;
      }
    }
    Ops.push_back(Src);
    Ops.append(SubInputs.begin(), SubInputs.end());
    if (ISD::isBuildVectorAllZeros(Src.getNode()))
      Mask.append(NumElts, SM_SentinelZero);
    else
      for (int i = 0; i != (int)NumElts; ++i)
        Mask.push_back(i);
    for (int i = 0; i != (int)NumSubElts; ++i) {
      int M = SubMask[i];
      if (0 <= M) {
        int InputIdx = M / NumSubElts;
        M = (NumElts * (1 + InputIdx)) + (M % NumSubElts);
      }
      Mask[i + InsertIdx] = M;
    }
    return true;
  }
  case X86ISD::PINSRB:
  case X86ISD::PINSRW:
  case ISD::SCALAR_TO_VECTOR:
  case ISD::INSERT_VECTOR_ELT: {
    // Match against a insert_vector_elt/scalar_to_vector of an extract from a
    // vector, for matching src/dst vector types.
    SDValue Scl = N.getOperand(Opcode == ISD::SCALAR_TO_VECTOR ? 0 : 1);
 
    unsigned DstIdx = 0;
    if (Opcode != ISD::SCALAR_TO_VECTOR) {
      // Check we have an in-range constant insertion index.
      if (!isa<ConstantSDNode>(N.getOperand(2)) ||
          N.getConstantOperandAPInt(2).uge(NumElts))
        return false;
      DstIdx = N.getConstantOperandVal(2);
 
      // Attempt to recognise an INSERT*(VEC, 0, DstIdx) shuffle pattern.
      if (X86::isZeroNode(Scl)) {
        Ops.push_back(N.getOperand(0));
        for (unsigned i = 0; i != NumElts; ++i)
          Mask.push_back(i == DstIdx ? SM_SentinelZero : (int)i);
        return true;
      }
    }
 
    // Peek through trunc/aext/zext/bitcast.
    // TODO: aext shouldn't require SM_SentinelZero padding.
    // TODO: handle shift of scalars.
    unsigned MinBitsPerElt = Scl.getScalarValueSizeInBits();
    while (Scl.getOpcode() == ISD::TRUNCATE ||
           Scl.getOpcode() == ISD::ANY_EXTEND ||
           Scl.getOpcode() == ISD::ZERO_EXTEND ||
           (Scl.getOpcode() == ISD::BITCAST &&
            Scl.getScalarValueSizeInBits() ==
                Scl.getOperand(0).getScalarValueSizeInBits())) {
      Scl = Scl.getOperand(0);
      MinBitsPerElt =
          std::min<unsigned>(MinBitsPerElt, Scl.getScalarValueSizeInBits());
    }
    if ((MinBitsPerElt % 8) != 0)
      return false;
 
    // Attempt to find the source vector the scalar was extracted from.
    SDValue SrcExtract;
    if ((Scl.getOpcode() == ISD::EXTRACT_VECTOR_ELT ||
         Scl.getOpcode() == X86ISD::PEXTRW ||
         Scl.getOpcode() == X86ISD::PEXTRB) &&
        Scl.getOperand(0).getValueSizeInBits() == NumSizeInBits) {
      SrcExtract = Scl;
    }
    if (!SrcExtract || !isa<ConstantSDNode>(SrcExtract.getOperand(1)))
      return false;
 
    SDValue SrcVec = SrcExtract.getOperand(0);
    EVT SrcVT = SrcVec.getValueType();
    if (!SrcVT.getScalarType().isByteSized())
      return false;
    unsigned SrcIdx = SrcExtract.getConstantOperandVal(1);
    unsigned SrcByte = SrcIdx * (SrcVT.getScalarSizeInBits() / 8);
    unsigned DstByte = DstIdx * NumBytesPerElt;
    MinBitsPerElt =
        std::min<unsigned>(MinBitsPerElt, SrcVT.getScalarSizeInBits());
 
    // Create 'identity' byte level shuffle mask and then add inserted bytes.
    if (Opcode == ISD::SCALAR_TO_VECTOR) {
      Ops.push_back(SrcVec);
      Mask.append(NumSizeInBytes, SM_SentinelUndef);
    } else {
      Ops.push_back(SrcVec);
      Ops.push_back(N.getOperand(0));
      for (int i = 0; i != (int)NumSizeInBytes; ++i)
        Mask.push_back(NumSizeInBytes + i);
    }
 
    unsigned MinBytesPerElts = MinBitsPerElt / 8;
    MinBytesPerElts = std::min(MinBytesPerElts, NumBytesPerElt);
    for (unsigned i = 0; i != MinBytesPerElts; ++i)
      Mask[DstByte + i] = SrcByte + i;
    for (unsigned i = MinBytesPerElts; i < NumBytesPerElt; ++i)
      Mask[DstByte + i] = SM_SentinelZero;
    return true;
  }
  case X86ISD::PACKSS:
  case X86ISD::PACKUS: {
    SDValue N0 = N.getOperand(0);
    SDValue N1 = N.getOperand(1);
    assert(N0.getValueType().getVectorNumElements() == (NumElts / 2) &&
           N1.getValueType().getVectorNumElements() == (NumElts / 2) &&
           "Unexpected input value type");
 
    APInt EltsLHS, EltsRHS;
    getPackDemandedElts(VT, DemandedElts, EltsLHS, EltsRHS);
 
    // If we know input saturation won't happen (or we don't care for particular
    // lanes), we can treat this as a truncation shuffle.
    bool Offset0 = false, Offset1 = false;
    if (Opcode == X86ISD::PACKSS) {
      if ((!(N0.isUndef() || EltsLHS.isZero()) &&
           DAG.ComputeNumSignBits(N0, EltsLHS, Depth + 1) <= NumBitsPerElt) ||
          (!(N1.isUndef() || EltsRHS.isZero()) &&
           DAG.ComputeNumSignBits(N1, EltsRHS, Depth + 1) <= NumBitsPerElt))
        return false;
      // We can't easily fold ASHR into a shuffle, but if it was feeding a
      // PACKSS then it was likely being used for sign-extension for a
      // truncation, so just peek through and adjust the mask accordingly.
      if (N0.getOpcode() == X86ISD::VSRAI && N->isOnlyUserOf(N0.getNode()) &&
          N0.getConstantOperandAPInt(1) == NumBitsPerElt) {
        Offset0 = true;
        N0 = N0.getOperand(0);
      }
      if (N1.getOpcode() == X86ISD::VSRAI && N->isOnlyUserOf(N1.getNode()) &&
          N1.getConstantOperandAPInt(1) == NumBitsPerElt) {
        Offset1 = true;
        N1 = N1.getOperand(0);
      }
    } else {
      APInt ZeroMask = APInt::getHighBitsSet(2 * NumBitsPerElt, NumBitsPerElt);
      if ((!(N0.isUndef() || EltsLHS.isZero()) &&
           !DAG.MaskedValueIsZero(N0, ZeroMask, EltsLHS, Depth + 1)) ||
          (!(N1.isUndef() || EltsRHS.isZero()) &&
           !DAG.MaskedValueIsZero(N1, ZeroMask, EltsRHS, Depth + 1)))
        return false;
    }
 
    bool IsUnary = (N0 == N1);
 
    Ops.push_back(N0);
    if (!IsUnary)
      Ops.push_back(N1);
 
    createPackShuffleMask(VT, Mask, IsUnary);
 
    if (Offset0 || Offset1) {
      for (int &M : Mask)
        if ((Offset0 && isInRange(M, 0, NumElts)) ||
            (Offset1 && isInRange(M, NumElts, 2 * NumElts)))
          ++M;
    }
    return true;
  }
  case ISD::VSELECT:
  case X86ISD::BLENDV: {
    SDValue Cond = N.getOperand(0);
    if (createShuffleMaskFromVSELECT(Mask, Cond, Opcode == X86ISD::BLENDV)) {
      Ops.push_back(N.getOperand(1));
      Ops.push_back(N.getOperand(2));
      return true;
    }
    return false;
  }
  case X86ISD::VTRUNC: {
    SDValue Src = N.getOperand(0);
    EVT SrcVT = Src.getValueType();
    if (SrcVT.getSizeInBits() != NumSizeInBits)
      return false;
    unsigned NumSrcElts = SrcVT.getVectorNumElements();
    unsigned NumBitsPerSrcElt = SrcVT.getScalarSizeInBits();
    unsigned Scale = NumBitsPerSrcElt / NumBitsPerElt;
    assert((NumBitsPerSrcElt % NumBitsPerElt) == 0 && "Illegal truncation");
    for (unsigned i = 0; i != NumSrcElts; ++i)
      Mask.push_back(i * Scale);
    Mask.append(NumElts - NumSrcElts, SM_SentinelZero);
    Ops.push_back(Src);
    return true;
  }
  case ISD::SHL:
  case ISD::SRL: {
    APInt UndefElts;
    SmallVector<APInt, 32> EltBits;
    if (!getTargetConstantBitsFromNode(N.getOperand(1), NumBitsPerElt,
                                       UndefElts, EltBits,
                                       /*AllowWholeUndefs*/ true,
                                       /*AllowPartialUndefs*/ false))
      return false;
 
    // We can only decode 'whole byte' bit shifts as shuffles.
    for (unsigned I = 0; I != NumElts; ++I)
      if (DemandedElts[I] && !UndefElts[I] &&
          (EltBits[I].urem(8) != 0 || EltBits[I].uge(NumBitsPerElt)))
        return false;
 
    Mask.append(NumSizeInBytes, SM_SentinelUndef);
    Ops.push_back(N.getOperand(0));
 
    for (unsigned I = 0; I != NumElts; ++I) {
      if (!DemandedElts[I] || UndefElts[I])
        continue;
      unsigned ByteShift = EltBits[I].getZExtValue() / 8;
      unsigned Lo = I * NumBytesPerElt;
      unsigned Hi = Lo + NumBytesPerElt;
      // Clear mask to all zeros and insert the shifted byte indices.
      std::fill(Mask.begin() + Lo, Mask.begin() + Hi, SM_SentinelZero);
      if (ISD::SHL == Opcode)
        std::iota(Mask.begin() + Lo + ByteShift, Mask.begin() + Hi, Lo);
      else
        std::iota(Mask.begin() + Lo, Mask.begin() + Hi - ByteShift,
                  Lo + ByteShift);
    }
    return true;
  }
  case X86ISD::VSHLI:
  case X86ISD::VSRLI: {
    uint64_t ShiftVal = N.getConstantOperandVal(1);
    // Out of range bit shifts are guaranteed to be zero.
    if (NumBitsPerElt <= ShiftVal) {
      Mask.append(NumElts, SM_SentinelZero);
      return true;
    }
 
    // We can only decode 'whole byte' bit shifts as shuffles.
    if ((ShiftVal % 8) != 0)
      break;
 
    uint64_t ByteShift = ShiftVal / 8;
    Ops.push_back(N.getOperand(0));
 
    // Clear mask to all zeros and insert the shifted byte indices.
    Mask.append(NumSizeInBytes, SM_SentinelZero);
 
    if (X86ISD::VSHLI == Opcode) {
      for (unsigned i = 0; i != NumSizeInBytes; i += NumBytesPerElt)
        for (unsigned j = ByteShift; j != NumBytesPerElt; ++j)
          Mask[i + j] = i + j - ByteShift;
    } else {
      for (unsigned i = 0; i != NumSizeInBytes; i += NumBytesPerElt)
        for (unsigned j = ByteShift; j != NumBytesPerElt; ++j)
          Mask[i + j - ByteShift] = i + j;
    }
    return true;
  }
  case ISD::ROTL:
  case ISD::ROTR: {
    APInt UndefElts;
    SmallVector<APInt, 32> EltBits;
    if (!getTargetConstantBitsFromNode(N.getOperand(1), NumBitsPerElt,
                                       UndefElts, EltBits,
                                       /*AllowWholeUndefs*/ true,
                                       /*AllowPartialUndefs*/ false))
      return false;
 
    // We can only decode 'whole byte' bit rotates as shuffles.
    for (unsigned I = 0; I != NumElts; ++I)
      if (DemandedElts[I] && !UndefElts[I] &&
          (EltBits[I].urem(NumBitsPerElt) % 8) != 0)
        return false;
 
    Ops.push_back(N.getOperand(0));
    for (unsigned I = 0; I != NumElts; ++I) {
      if (!DemandedElts[I] || UndefElts[I]) {
        Mask.append(NumBytesPerElt, SM_SentinelUndef);
        continue;
      }
      int Offset = EltBits[I].urem(NumBitsPerElt) / 8;
      Offset = (ISD::ROTL == Opcode ? NumBytesPerElt - Offset : Offset);
      int BaseIdx = I * NumBytesPerElt;
      for (int J = 0; J != (int)NumBytesPerElt; ++J) {
        Mask.push_back(BaseIdx + ((Offset + J) % NumBytesPerElt));
      }
    }
    return true;
  }
  case X86ISD::VROTLI:
  case X86ISD::VROTRI: {
    // We can only decode 'whole byte' bit rotates as shuffles.
    uint64_t RotateVal = N.getConstantOperandAPInt(1).urem(NumBitsPerElt);
    if ((RotateVal % 8) != 0)
      return false;
    Ops.push_back(N.getOperand(0));
    int Offset = RotateVal / 8;
    Offset = (X86ISD::VROTLI == Opcode ? NumBytesPerElt - Offset : Offset);
    for (int i = 0; i != (int)NumElts; ++i) {
      int BaseIdx = i * NumBytesPerElt;
      for (int j = 0; j != (int)NumBytesPerElt; ++j) {
        Mask.push_back(BaseIdx + ((Offset + j) % NumBytesPerElt));
      }
    }
    return true;
  }
  case X86ISD::VBROADCAST: {
    SDValue Src = N.getOperand(0);
    if (!Src.getSimpleValueType().isVector()) {
      if (Src.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
          !isNullConstant(Src.getOperand(1)) ||
          Src.getOperand(0).getValueType().getScalarType() !=
              VT.getScalarType())
        return false;
      Src = Src.getOperand(0);
    }
    Ops.push_back(Src);
    Mask.append(NumElts, 0);
    return true;
  }
  case ISD::SIGN_EXTEND_VECTOR_INREG: {
    SDValue Src = N.getOperand(0);
    EVT SrcVT = Src.getValueType();
    unsigned NumBitsPerSrcElt = SrcVT.getScalarSizeInBits();
 
    // Extended source must be a simple vector.
    if (!SrcVT.isSimple() || (SrcVT.getSizeInBits() % 128) != 0 ||
        (NumBitsPerSrcElt % 8) != 0)
      return false;
 
    // We can only handle all-signbits extensions.
    APInt DemandedSrcElts =
        DemandedElts.zextOrTrunc(SrcVT.getVectorNumElements());
    if (DAG.ComputeNumSignBits(Src, DemandedSrcElts) != NumBitsPerSrcElt)
      return false;
 
    assert((NumBitsPerElt % NumBitsPerSrcElt) == 0 && "Unexpected extension");
    unsigned Scale = NumBitsPerElt / NumBitsPerSrcElt;
    for (unsigned I = 0; I != NumElts; ++I)
      Mask.append(Scale, I);
    Ops.push_back(Src);
    return true;
  }
  case ISD::ZERO_EXTEND:
  case ISD::ANY_EXTEND:
  case ISD::ZERO_EXTEND_VECTOR_INREG:
  case ISD::ANY_EXTEND_VECTOR_INREG: {
    SDValue Src = N.getOperand(0);
    EVT SrcVT = Src.getValueType();
 
    // Extended source must be a simple vector.
    if (!SrcVT.isSimple() || (SrcVT.getSizeInBits() % 128) != 0 ||
        (SrcVT.getScalarSizeInBits() % 8) != 0)
      return false;
 
    bool IsAnyExtend =
        (ISD::ANY_EXTEND == Opcode || ISD::ANY_EXTEND_VECTOR_INREG == Opcode);
    DecodeZeroExtendMask(SrcVT.getScalarSizeInBits(), NumBitsPerElt, NumElts,
                         IsAnyExtend, Mask);
    Ops.push_back(Src);
    return true;
  }
  }
 
  return false;
}
 
/// Removes unused/repeated shuffle source inputs and adjusts the shuffle mask.
static void resolveTargetShuffleInputsAndMask(SmallVectorImpl<SDValue> &Inputs,
                                              SmallVectorImpl<int> &Mask) {
  int MaskWidth = Mask.size();
  SmallVector<SDValue, 16> UsedInputs;
  for (int i = 0, e = Inputs.size(); i < e; ++i) {
    int lo = UsedInputs.size() * MaskWidth;
    int hi = lo + MaskWidth;
 
    // Strip UNDEF input usage.
    if (Inputs[i].isUndef())
      for (int &M : Mask)
        if ((lo <= M) && (M < hi))
          M = SM_SentinelUndef;
 
    // Check for unused inputs.
    if (none_of(Mask, [lo, hi](int i) { return (lo <= i) && (i < hi); })) {
      for (int &M : Mask)
        if (lo <= M)
          M -= MaskWidth;
      continue;
    }
 
    // Check for repeated inputs.
    bool IsRepeat = false;
    for (int j = 0, ue = UsedInputs.size(); j != ue; ++j) {
      if (peekThroughBitcasts(UsedInputs[j]) != peekThroughBitcasts(Inputs[i]))
        continue;
      for (int &M : Mask)
        if (lo <= M)
          M = (M < hi) ? ((M - lo) + (j * MaskWidth)) : (M - MaskWidth);
      IsRepeat = true;
      break;
    }
    if (IsRepeat)
      continue;
 
    UsedInputs.push_back(Inputs[i]);
  }
  Inputs = std::move(UsedInputs);
}
 
/// Calls getTargetShuffleAndZeroables to resolve a target shuffle mask's inputs
/// and then sets the SM_SentinelUndef and SM_SentinelZero values.
/// Returns true if the target shuffle mask was decoded.
static bool getTargetShuffleInputs(SDValue Op, const APInt &DemandedElts,
                                   SmallVectorImpl<SDValue> &Inputs,
                                   SmallVectorImpl<int> &Mask,
                                   APInt &KnownUndef, APInt &KnownZero,
                                   const SelectionDAG &DAG, unsigned Depth,
                                   bool ResolveKnownElts) {
  if (Depth >= SelectionDAG::MaxRecursionDepth)
    return false; // Limit search depth.
 
  EVT VT = Op.getValueType();
  if (!VT.isSimple() || !VT.isVector())
    return false;
 
  if (getTargetShuffleAndZeroables(Op, Mask, Inputs, KnownUndef, KnownZero)) {
    if (ResolveKnownElts)
      resolveTargetShuffleFromZeroables(Mask, KnownUndef, KnownZero);
    return true;
  }
  if (getFauxShuffleMask(Op, DemandedElts, Mask, Inputs, DAG, Depth,
                         ResolveKnownElts)) {
    resolveZeroablesFromTargetShuffle(Mask, KnownUndef, KnownZero);
    return true;
  }
  return false;
}
 
static bool getTargetShuffleInputs(SDValue Op, const APInt &DemandedElts,
                                   SmallVectorImpl<SDValue> &Inputs,
                                   SmallVectorImpl<int> &Mask,
                                   const SelectionDAG &DAG, unsigned Depth,
                                   bool ResolveKnownElts) {
  APInt KnownUndef, KnownZero;
  return getTargetShuffleInputs(Op, DemandedElts, Inputs, Mask, KnownUndef,
                                KnownZero, DAG, Depth, ResolveKnownElts);
}
 
static bool getTargetShuffleInputs(SDValue Op, SmallVectorImpl<SDValue> &Inputs,
                                   SmallVectorImpl<int> &Mask,
                                   const SelectionDAG &DAG, unsigned Depth = 0,
                                   bool ResolveKnownElts = true) {
  EVT VT = Op.getValueType();
  if (!VT.isSimple() || !VT.isVector())
    return false;
 
  unsigned NumElts = Op.getValueType().getVectorNumElements();
  APInt DemandedElts = APInt::getAllOnes(NumElts);
  return getTargetShuffleInputs(Op, DemandedElts, Inputs, Mask, DAG, Depth,
                                ResolveKnownElts);
}
 
// Attempt to create a scalar/subvector broadcast from the base MemSDNode.
static SDValue getBROADCAST_LOAD(unsigned Opcode, const SDLoc &DL, EVT VT,
                                 EVT MemVT, MemSDNode *Mem, unsigned Offset,
                                 SelectionDAG &DAG) {
  assert((Opcode == X86ISD::VBROADCAST_LOAD ||
          Opcode == X86ISD::SUBV_BROADCAST_LOAD) &&
         "Unknown broadcast load type");
 
  // Ensure this is a simple (non-atomic, non-voltile), temporal read memop.
  if (!Mem || !Mem->readMem() || !Mem->isSimple() || Mem->isNonTemporal())
    return SDValue();
 
  SDValue Ptr = DAG.getMemBasePlusOffset(Mem->getBasePtr(),
                                         TypeSize::getFixed(Offset), DL);
  SDVTList Tys = DAG.getVTList(VT, MVT::Other);
  SDValue Ops[] = {Mem->getChain(), Ptr};
  SDValue BcstLd = DAG.getMemIntrinsicNode(
      Opcode, DL, Tys, Ops, MemVT,
      DAG.getMachineFunction().getMachineMemOperand(
          Mem->getMemOperand(), Offset, MemVT.getStoreSize()));
  DAG.makeEquivalentMemoryOrdering(SDValue(Mem, 1), BcstLd.getValue(1));
  return BcstLd;
}
 
/// Returns the scalar element that will make up the i'th
/// element of the result of the vector shuffle.
static SDValue getShuffleScalarElt(SDValue Op, unsigned Index,
                                   SelectionDAG &DAG, unsigned Depth) {
  if (Depth >= SelectionDAG::MaxRecursionDepth)
    return SDValue(); // Limit search depth.
 
  EVT VT = Op.getValueType();
  unsigned Opcode = Op.getOpcode();
  unsigned NumElems = VT.getVectorNumElements();
 
  // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
  if (auto *SV = dyn_cast<ShuffleVectorSDNode>(Op)) {
    int Elt = SV->getMaskElt(Index);
 
    if (Elt < 0)
      return DAG.getUNDEF(VT.getVectorElementType());
 
    SDValue Src = (Elt < (int)NumElems) ? SV->getOperand(0) : SV->getOperand(1);
    return getShuffleScalarElt(Src, Elt % NumElems, DAG, Depth + 1);
  }
 
  // Recurse into target specific vector shuffles to find scalars.
  if (isTargetShuffle(Opcode)) {
    MVT ShufVT = VT.getSimpleVT();
    MVT ShufSVT = ShufVT.getVectorElementType();
    int NumElems = (int)ShufVT.getVectorNumElements();
    SmallVector<int, 16> ShuffleMask;
    SmallVector<SDValue, 16> ShuffleOps;
    if (!getTargetShuffleMask(Op, true, ShuffleOps, ShuffleMask))
      return SDValue();
 
    int Elt = ShuffleMask[Index];
    if (Elt == SM_SentinelZero)
      return ShufSVT.isInteger() ? DAG.getConstant(0, SDLoc(Op), ShufSVT)
                                 : DAG.getConstantFP(+0.0, SDLoc(Op), ShufSVT);
    if (Elt == SM_SentinelUndef)
      return DAG.getUNDEF(ShufSVT);
 
    assert(0 <= Elt && Elt < (2 * NumElems) && "Shuffle index out of range");
    SDValue Src = (Elt < NumElems) ? ShuffleOps[0] : ShuffleOps[1];
    return getShuffleScalarElt(Src, Elt % NumElems, DAG, Depth + 1);
  }
 
  // Recurse into insert_subvector base/sub vector to find scalars.
  if (Opcode == ISD::INSERT_SUBVECTOR) {
    SDValue Vec = Op.getOperand(0);
    SDValue Sub = Op.getOperand(1);
    uint64_t SubIdx = Op.getConstantOperandVal(2);
    unsigned NumSubElts = Sub.getValueType().getVectorNumElements();
 
    if (SubIdx <= Index && Index < (SubIdx + NumSubElts))
      return getShuffleScalarElt(Sub, Index - SubIdx, DAG, Depth + 1);
    return getShuffleScalarElt(Vec, Index, DAG, Depth + 1);
  }
 
  // Recurse into concat_vectors sub vector to find scalars.
  if (Opcode == ISD::CONCAT_VECTORS) {
    EVT SubVT = Op.getOperand(0).getValueType();
    unsigned NumSubElts = SubVT.getVectorNumElements();
    uint64_t SubIdx = Index / NumSubElts;
    uint64_t SubElt = Index % NumSubElts;
    return getShuffleScalarElt(Op.getOperand(SubIdx), SubElt, DAG, Depth + 1);
  }
 
  // Recurse into extract_subvector src vector to find scalars.
  if (Opcode == ISD::EXTRACT_SUBVECTOR) {
    SDValue Src = Op.getOperand(0);
    uint64_t SrcIdx = Op.getConstantOperandVal(1);
    return getShuffleScalarElt(Src, Index + SrcIdx, DAG, Depth + 1);
  }
 
  // We only peek through bitcasts of the same vector width.
  if (Opcode == ISD::BITCAST) {
    SDValue Src = Op.getOperand(0);
    EVT SrcVT = Src.getValueType();
    if (SrcVT.isVector() && SrcVT.getVectorNumElements() == NumElems)
      return getShuffleScalarElt(Src, Index, DAG, Depth + 1);
    return SDValue();
  }
 
  // Actual nodes that may contain scalar elements
 
  // For insert_vector_elt - either return the index matching scalar or recurse
  // into the base vector.
  if (Opcode == ISD::INSERT_VECTOR_ELT &&
      isa<ConstantSDNode>(Op.getOperand(2))) {
    if (Op.getConstantOperandAPInt(2) == Index)
      return Op.getOperand(1);
    return getShuffleScalarElt(Op.getOperand(0), Index, DAG, Depth + 1);
  }
 
  if (Opcode == ISD::SCALAR_TO_VECTOR)
    return (Index == 0) ? Op.getOperand(0)
                        : DAG.getUNDEF(VT.getVectorElementType());
 
  if (Opcode == ISD::BUILD_VECTOR)
    return Op.getOperand(Index);
 
  return SDValue();
}
 
// Use PINSRB/PINSRW/PINSRD to create a build vector.
static SDValue LowerBuildVectorAsInsert(SDValue Op, const SDLoc &DL,
                                        const APInt &NonZeroMask,
                                        unsigned NumNonZero, unsigned NumZero,
                                        SelectionDAG &DAG,
                                        const X86Subtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  unsigned NumElts = VT.getVectorNumElements();
  assert(((VT == MVT::v8i16 && Subtarget.hasSSE2()) ||
          ((VT == MVT::v16i8 || VT == MVT::v4i32) && Subtarget.hasSSE41())) &&
         "Illegal vector insertion");
 
  SDValue V;
  bool First = true;
 
  for (unsigned i = 0; i < NumElts; ++i) {
    bool IsNonZero = NonZeroMask[i];
    if (!IsNonZero)
      continue;
 
    // If the build vector contains zeros or our first insertion is not the
    // first index then insert into zero vector to break any register
    // dependency else use SCALAR_TO_VECTOR.
    if (First) {
      First = false;
      if (NumZero || 0 != i)
        V = getZeroVector(VT, Subtarget, DAG, DL);
      else {
        assert(0 == i && "Expected insertion into zero-index");
        V = DAG.getAnyExtOrTrunc(Op.getOperand(i), DL, MVT::i32);
        V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4i32, V);
        V = DAG.getBitcast(VT, V);
        continue;
      }
    }
    V = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, V, Op.getOperand(i),
                    DAG.getVectorIdxConstant(i, DL));
  }
 
  return V;
}
 
/// Custom lower build_vector of v16i8.
static SDValue LowerBuildVectorv16i8(SDValue Op, const SDLoc &DL,
                                     const APInt &NonZeroMask,
                                     unsigned NumNonZero, unsigned NumZero,
                                     SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
  if (NumNonZero > 8 && !Subtarget.hasSSE41())
    return SDValue();
 
  // SSE4.1 - use PINSRB to insert each byte directly.
  if (Subtarget.hasSSE41())
    return LowerBuildVectorAsInsert(Op, DL, NonZeroMask, NumNonZero, NumZero,
                                    DAG, Subtarget);
 
  SDValue V;
 
  // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
  // If both the lowest 16-bits are non-zero, then convert to MOVD.
  if (!NonZeroMask.extractBits(2, 0).isZero() &&
      !NonZeroMask.extractBits(2, 2).isZero()) {
    for (unsigned I = 0; I != 4; ++I) {
      if (!NonZeroMask[I])
        continue;
      SDValue Elt = DAG.getZExtOrTrunc(Op.getOperand(I), DL, MVT::i32);
      if (I != 0)
        Elt = DAG.getNode(ISD::SHL, DL, MVT::i32, Elt,
                          DAG.getConstant(I * 8, DL, MVT::i8));
      V = V ? DAG.getNode(ISD::OR, DL, MVT::i32, V, Elt) : Elt;
    }
    assert(V && "Failed to fold v16i8 vector to zero");
    V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4i32, V);
    V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v4i32, V);
    V = DAG.getBitcast(MVT::v8i16, V);
  }
  for (unsigned i = V ? 4 : 0; i < 16; i += 2) {
    bool ThisIsNonZero = NonZeroMask[i];
    bool NextIsNonZero = NonZeroMask[i + 1];
    if (!ThisIsNonZero && !NextIsNonZero)
      continue;
 
    SDValue Elt;
    if (ThisIsNonZero) {
      if (NumZero || NextIsNonZero)
        Elt = DAG.getZExtOrTrunc(Op.getOperand(i), DL, MVT::i32);
      else
        Elt = DAG.getAnyExtOrTrunc(Op.getOperand(i), DL, MVT::i32);
    }
 
    if (NextIsNonZero) {
      SDValue NextElt = Op.getOperand(i + 1);
      if (i == 0 && NumZero)
        NextElt = DAG.getZExtOrTrunc(NextElt, DL, MVT::i32);
      else
        NextElt = DAG.getAnyExtOrTrunc(NextElt, DL, MVT::i32);
      NextElt = DAG.getNode(ISD::SHL, DL, MVT::i32, NextElt,
                            DAG.getConstant(8, DL, MVT::i8));
      if (ThisIsNonZero)
        Elt = DAG.getNode(ISD::OR, DL, MVT::i32, NextElt, Elt);
      else
        Elt = NextElt;
    }
 
    // If our first insertion is not the first index or zeros are needed, then
    // insert into zero vector. Otherwise, use SCALAR_TO_VECTOR (leaves high
    // elements undefined).
    if (!V) {
      if (i != 0 || NumZero)
        V = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
      else {
        V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4i32, Elt);
        V = DAG.getBitcast(MVT::v8i16, V);
        continue;
      }
    }
    Elt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i16, Elt);
    V = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, MVT::v8i16, V, Elt,
                    DAG.getVectorIdxConstant(i / 2, DL));
  }
 
  return DAG.getBitcast(MVT::v16i8, V);
}
 
/// Custom lower build_vector of v8i16.
static SDValue LowerBuildVectorv8i16(SDValue Op, const SDLoc &DL,
                                     const APInt &NonZeroMask,
                                     unsigned NumNonZero, unsigned NumZero,
                                     SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
  if (NumNonZero > 4 && !Subtarget.hasSSE41())
    return SDValue();
 
  // Use PINSRW to insert each byte directly.
  return LowerBuildVectorAsInsert(Op, DL, NonZeroMask, NumNonZero, NumZero, DAG,
                                  Subtarget);
}
 
/// Custom lower build_vector of v4i32 or v4f32.
static SDValue LowerBuildVectorv4x32(SDValue Op, const SDLoc &DL,
                                     SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
  // If this is a splat of a pair of elements, use MOVDDUP (unless the target
  // has XOP; in that case defer lowering to potentially use VPERMIL2PS).
  // Because we're creating a less complicated build vector here, we may enable
  // further folding of the MOVDDUP via shuffle transforms.
  if (Subtarget.hasSSE3() && !Subtarget.hasXOP() &&
      Op.getOperand(0) == Op.getOperand(2) &&
      Op.getOperand(1) == Op.getOperand(3) &&
      Op.getOperand(0) != Op.getOperand(1)) {
    MVT VT = Op.getSimpleValueType();
    MVT EltVT = VT.getVectorElementType();
    // Create a new build vector with the first 2 elements followed by undef
    // padding, bitcast to v2f64, duplicate, and bitcast back.
    SDValue Ops[4] = { Op.getOperand(0), Op.getOperand(1),
                       DAG.getUNDEF(EltVT), DAG.getUNDEF(EltVT) };
    SDValue NewBV = DAG.getBitcast(MVT::v2f64, DAG.getBuildVector(VT, DL, Ops));
    SDValue Dup = DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, NewBV);
    return DAG.getBitcast(VT, Dup);
  }
 
  // Find all zeroable elements.
  std::bitset<4> Zeroable, Undefs;
  for (int i = 0; i < 4; ++i) {
    SDValue Elt = Op.getOperand(i);
    Undefs[i] = Elt.isUndef();
    Zeroable[i] = (Elt.isUndef() || X86::isZeroNode(Elt));
  }
  assert(Zeroable.size() - Zeroable.count() > 1 &&
         "We expect at least two non-zero elements!");
 
  // We only know how to deal with build_vector nodes where elements are either
  // zeroable or extract_vector_elt with constant index.
  SDValue FirstNonZero;
  unsigned FirstNonZeroIdx;
  for (unsigned i = 0; i < 4; ++i) {
    if (Zeroable[i])
      continue;
    SDValue Elt = Op.getOperand(i);
    if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
        !isa<ConstantSDNode>(Elt.getOperand(1)))
      return SDValue();
    // Make sure that this node is extracting from a 128-bit vector.
    MVT VT = Elt.getOperand(0).getSimpleValueType();
    if (!VT.is128BitVector())
      return SDValue();
    if (!FirstNonZero.getNode()) {
      FirstNonZero = Elt;
      FirstNonZeroIdx = i;
    }
  }
 
  assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
  SDValue V1 = FirstNonZero.getOperand(0);
  MVT VT = V1.getSimpleValueType();
 
  // See if this build_vector can be lowered as a blend with zero.
  SDValue Elt;
  unsigned EltMaskIdx, EltIdx;
  int Mask[4];
  for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
    if (Zeroable[EltIdx]) {
      // The zero vector will be on the right hand side.
      Mask[EltIdx] = EltIdx+4;
      continue;
    }
 
    Elt = Op->getOperand(EltIdx);
    // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
    EltMaskIdx = Elt.getConstantOperandVal(1);
    if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
      break;
    Mask[EltIdx] = EltIdx;
  }
 
  if (EltIdx == 4) {
    // Let the shuffle legalizer deal with blend operations.
    SDValue VZeroOrUndef = (Zeroable == Undefs)
                               ? DAG.getUNDEF(VT)
                               : getZeroVector(VT, Subtarget, DAG, DL);
    if (V1.getSimpleValueType() != VT)
      V1 = DAG.getBitcast(VT, V1);
    return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZeroOrUndef, Mask);
  }
 
  // See if we can lower this build_vector to a INSERTPS.
  if (!Subtarget.hasSSE41())
    return SDValue();
 
  SDValue V2 = Elt.getOperand(0);
  if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
    V1 = SDValue();
 
  bool CanFold = true;
  for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
    if (Zeroable[i])
      continue;
 
    SDValue Current = Op->getOperand(i);
    SDValue SrcVector = Current->getOperand(0);
    if (!V1.getNode())
      V1 = SrcVector;
    CanFold = (SrcVector == V1) && (Current.getConstantOperandAPInt(1) == i);
  }
 
  if (!CanFold)
    return SDValue();
 
  assert(V1.getNode() && "Expected at least two non-zero elements!");
  if (V1.getSimpleValueType() != MVT::v4f32)
    V1 = DAG.getBitcast(MVT::v4f32, V1);
  if (V2.getSimpleValueType() != MVT::v4f32)
    V2 = DAG.getBitcast(MVT::v4f32, V2);
 
  // Ok, we can emit an INSERTPS instruction.
  unsigned ZMask = Zeroable.to_ulong();
 
  unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
  assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
  SDValue Result =
      DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
                  DAG.getTargetConstant(InsertPSMask, DL, MVT::i8));
  return DAG.getBitcast(VT, Result);
}
 
/// Return a vector logical shift node.
static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp, unsigned NumBits,
                         SelectionDAG &DAG, const TargetLowering &TLI,
                         const SDLoc &dl) {
  assert(VT.is128BitVector() && "Unknown type for VShift");
  MVT ShVT = MVT::v16i8;
  unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
  SrcOp = DAG.getBitcast(ShVT, SrcOp);
  assert(NumBits % 8 == 0 && "Only support byte sized shifts");
  SDValue ShiftVal = DAG.getTargetConstant(NumBits / 8, dl, MVT::i8);
  return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
}
 
static SDValue LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, const SDLoc &dl,
                                      SelectionDAG &DAG) {
 
  // Check if the scalar load can be widened into a vector load. And if
  // the address is "base + cst" see if the cst can be "absorbed" into
  // the shuffle mask.
  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
    SDValue Ptr = LD->getBasePtr();
    if (!ISD::isNormalLoad(LD) || !LD->isSimple())
      return SDValue();
    EVT PVT = LD->getValueType(0);
    if (PVT != MVT::i32 && PVT != MVT::f32)
      return SDValue();
 
    int FI = -1;
    int64_t Offset = 0;
    if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
      FI = FINode->getIndex();
      Offset = 0;
    } else if (DAG.isBaseWithConstantOffset(Ptr) &&
               isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
      FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
      Offset = Ptr.getConstantOperandVal(1);
      Ptr = Ptr.getOperand(0);
    } else {
      return SDValue();
    }
 
    // FIXME: 256-bit vector instructions don't require a strict alignment,
    // improve this code to support it better.
    Align RequiredAlign(VT.getSizeInBits() / 8);
    SDValue Chain = LD->getChain();
    // Make sure the stack object alignment is at least 16 or 32.
    MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
    MaybeAlign InferredAlign = DAG.InferPtrAlign(Ptr);
    if (!InferredAlign || *InferredAlign < RequiredAlign) {
      if (MFI.isFixedObjectIndex(FI)) {
        // Can't change the alignment. FIXME: It's possible to compute
        // the exact stack offset and reference FI + adjust offset instead.
        // If someone *really* cares about this. That's the way to implement it.
        return SDValue();
      } else {
        MFI.setObjectAlignment(FI, RequiredAlign);
      }
    }
 
    // (Offset % 16 or 32) must be multiple of 4. Then address is then
    // Ptr + (Offset & ~15).
    if (Offset < 0)
      return SDValue();
    if ((Offset % RequiredAlign.value()) & 3)
      return SDValue();
    int64_t StartOffset = Offset & ~int64_t(RequiredAlign.value() - 1);
    if (StartOffset) {
      SDLoc DL(Ptr);
      Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
                        DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
    }
 
    int EltNo = (Offset - StartOffset) >> 2;
    unsigned NumElems = VT.getVectorNumElements();
 
    EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
    SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
                             LD->getPointerInfo().getWithOffset(StartOffset));
 
    SmallVector<int, 8> Mask(NumElems, EltNo);
 
    return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), Mask);
  }
 
  return SDValue();
}
 
// Recurse to find a LoadSDNode source and the accumulated ByteOffest.
static bool findEltLoadSrc(SDValue Elt, LoadSDNode *&Ld, int64_t &ByteOffset) {
  if (ISD::isNON_EXTLoad(Elt.getNode())) {
    auto *BaseLd = cast<LoadSDNode>(Elt);
    if (!BaseLd->isSimple())
      return false;
    Ld = BaseLd;
    ByteOffset = 0;
    return true;
  }
 
  switch (Elt.getOpcode()) {
  case ISD::BITCAST:
  case ISD::TRUNCATE:
  case ISD::SCALAR_TO_VECTOR:
    return findEltLoadSrc(Elt.getOperand(0), Ld, ByteOffset);
  case ISD::SRL:
    if (auto *AmtC = dyn_cast<ConstantSDNode>(Elt.getOperand(1))) {
      uint64_t Amt = AmtC->getZExtValue();
      if ((Amt % 8) == 0 && findEltLoadSrc(Elt.getOperand(0), Ld, ByteOffset)) {
        ByteOffset += Amt / 8;
        return true;
      }
    }
    break;
  case ISD::EXTRACT_VECTOR_ELT:
    if (auto *IdxC = dyn_cast<ConstantSDNode>(Elt.getOperand(1))) {
      SDValue Src = Elt.getOperand(0);
      unsigned SrcSizeInBits = Src.getScalarValueSizeInBits();
      unsigned DstSizeInBits = Elt.getScalarValueSizeInBits();
      if (DstSizeInBits == SrcSizeInBits && (SrcSizeInBits % 8) == 0 &&
          findEltLoadSrc(Src, Ld, ByteOffset)) {
        uint64_t Idx = IdxC->getZExtValue();
        ByteOffset += Idx * (SrcSizeInBits / 8);
        return true;
      }
    }
    break;
  }
 
  return false;
}
 
/// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
/// elements can be replaced by a single large load which has the same value as
/// a build_vector or insert_subvector whose loaded operands are 'Elts'.
///
/// Example: <load i32 *a, load i32 *a+4, zero, undef> -> zextload a
static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
                                        const SDLoc &DL, SelectionDAG &DAG,
                                        const X86Subtarget &Subtarget,
                                        bool IsAfterLegalize,
                                        unsigned Depth = 0) {
  if (Depth >= SelectionDAG::MaxRecursionDepth)
    return SDValue(); // Limit search depth.
  if ((VT.getScalarSizeInBits() % 8) != 0)
    return SDValue();
 
  unsigned NumElems = Elts.size();
 
  int LastLoadedElt = -1;
  APInt LoadMask = APInt::getZero(NumElems);
  APInt ZeroMask = APInt::getZero(NumElems);
  APInt UndefMask = APInt::getZero(NumElems);
 
  SmallVector<LoadSDNode*, 8> Loads(NumElems, nullptr);
  SmallVector<int64_t, 8> ByteOffsets(NumElems, 0);
 
  // For each element in the initializer, see if we've found a load, zero or an
  // undef.
  for (unsigned i = 0; i < NumElems; ++i) {
    SDValue Elt = peekThroughBitcasts(Elts[i]);
    if (!Elt.getNode())
      return SDValue();
    if (Elt.isUndef()) {
      UndefMask.setBit(i);
      continue;
    }
    if (X86::isZeroNode(Elt) || ISD::isBuildVectorAllZeros(Elt.getNode())) {
      ZeroMask.setBit(i);
      continue;
    }
 
    // Each loaded element must be the correct fractional portion of the
    // requested vector load.
    unsigned EltSizeInBits = Elt.getValueSizeInBits();
    if ((NumElems * EltSizeInBits) != VT.getSizeInBits())
      return SDValue();
 
    if (!findEltLoadSrc(Elt, Loads[i], ByteOffsets[i]) || ByteOffsets[i] < 0)
      return SDValue();
    unsigned LoadSizeInBits = Loads[i]->getValueSizeInBits(0);
    if (((ByteOffsets[i] * 8) + EltSizeInBits) > LoadSizeInBits)
      return SDValue();
 
    LoadMask.setBit(i);
    LastLoadedElt = i;
  }
  assert((ZeroMask.popcount() + UndefMask.popcount() + LoadMask.popcount()) ==
             NumElems &&
         "Incomplete element masks");
 
  // Handle Special Cases - all undef or undef/zero.
  if (UndefMask.popcount() == NumElems)
    return DAG.getUNDEF(VT);
  if ((ZeroMask.popcount() + UndefMask.popcount()) == NumElems)
    return VT.isInteger() ? DAG.getConstant(0, DL, VT)
                          : DAG.getConstantFP(0.0, DL, VT);
 
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  int FirstLoadedElt = LoadMask.countr_zero();
  SDValue EltBase = peekThroughBitcasts(Elts[FirstLoadedElt]);
  EVT EltBaseVT = EltBase.getValueType();
  assert(EltBaseVT.getSizeInBits() == EltBaseVT.getStoreSizeInBits() &&
         "Register/Memory size mismatch");
  LoadSDNode *LDBase = Loads[FirstLoadedElt];
  assert(LDBase && "Did not find base load for merging consecutive loads");
  unsigned BaseSizeInBits = EltBaseVT.getStoreSizeInBits();
  unsigned BaseSizeInBytes = BaseSizeInBits / 8;
  int NumLoadedElts = (1 + LastLoadedElt - FirstLoadedElt);
  int LoadSizeInBits = NumLoadedElts * BaseSizeInBits;
  assert((BaseSizeInBits % 8) == 0 && "Sub-byte element loads detected");
 
  // TODO: Support offsetting the base load.
  if (ByteOffsets[FirstLoadedElt] != 0)
    return SDValue();
 
  // Check to see if the element's load is consecutive to the base load
  // or offset from a previous (already checked) load.
  auto CheckConsecutiveLoad = [&](LoadSDNode *Base, int EltIdx) {
    LoadSDNode *Ld = Loads[EltIdx];
    int64_t ByteOffset = ByteOffsets[EltIdx];
    if (ByteOffset && (ByteOffset % BaseSizeInBytes) == 0) {
      int64_t BaseIdx = EltIdx - (ByteOffset / BaseSizeInBytes);
      return (0 <= BaseIdx && BaseIdx < (int)NumElems && LoadMask[BaseIdx] &&
              Loads[BaseIdx] == Ld && ByteOffsets[BaseIdx] == 0);
    }
    int Stride = EltIdx - FirstLoadedElt;
    if (DAG.areNonVolatileConsecutiveLoads(Ld, Base, BaseSizeInBytes, Stride))
      return true;
    // Try again using the memory load size (we might have broken a large load
    // into smaller elements), ensure the stride is the full memory load size
    // apart and a whole number of elements fit in each memory load.
    unsigned BaseMemSizeInBits = Base->getMemoryVT().getSizeInBits();
    if (((Stride * BaseSizeInBits) % BaseMemSizeInBits) == 0 &&
        (BaseMemSizeInBits % BaseSizeInBits) == 0) {
      unsigned Scale = BaseMemSizeInBits / BaseSizeInBits;
      return DAG.areNonVolatileConsecutiveLoads(Ld, Base, BaseMemSizeInBits / 8,
                                                Stride / Scale);
    }
    return false;
  };
 
  // Consecutive loads can contain UNDEFS but not ZERO elements.
  // Consecutive loads with UNDEFs and ZEROs elements require a
  // an additional shuffle stage to clear the ZERO elements.
  bool IsConsecutiveLoad = true;
  bool IsConsecutiveLoadWithZeros = true;
  for (int i = FirstLoadedElt + 1; i <= LastLoadedElt; ++i) {
    if (LoadMask[i]) {
      if (!CheckConsecutiveLoad(LDBase, i)) {
        IsConsecutiveLoad = false;
        IsConsecutiveLoadWithZeros = false;
        break;
      }
    } else if (ZeroMask[i]) {
      IsConsecutiveLoad = false;
    }
  }
 
  auto CreateLoad = [&DAG, &DL, &Loads](EVT VT, LoadSDNode *LDBase) {
    auto MMOFlags = LDBase->getMemOperand()->getFlags();
    assert(LDBase->isSimple() &&
           "Cannot merge volatile or atomic loads.");
    SDValue NewLd =
        DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
                    LDBase->getPointerInfo(), LDBase->getBaseAlign(), MMOFlags);
    for (auto *LD : Loads)
      if (LD)
        DAG.makeEquivalentMemoryOrdering(LD, NewLd);
    return NewLd;
  };
 
  // Check if the base load is entirely dereferenceable.
  bool IsDereferenceable = LDBase->getPointerInfo().isDereferenceable(
      VT.getSizeInBits() / 8, *DAG.getContext(), DAG.getDataLayout());
 
  // LOAD - all consecutive load/undefs (must start/end with a load or be
  // entirely dereferenceable). If we have found an entire vector of loads and
  // undefs, then return a large load of the entire vector width starting at the
  // base pointer. If the vector contains zeros, then attempt to shuffle those
  // elements.
  if (FirstLoadedElt == 0 &&
      (NumLoadedElts == (int)NumElems || IsDereferenceable) &&
      (IsConsecutiveLoad || IsConsecutiveLoadWithZeros)) {
    if (IsAfterLegalize && !TLI.isOperationLegal(ISD::LOAD, VT))
      return SDValue();
 
    // Don't create 256-bit non-temporal aligned loads without AVX2 as these
    // will lower to regular temporal loads and use the cache.
    if (LDBase->isNonTemporal() && LDBase->getAlign() >= Align(32) &&
        VT.is256BitVector() && !Subtarget.hasInt256())
      return SDValue();
 
    if (NumElems == 1)
      return DAG.getBitcast(VT, Elts[FirstLoadedElt]);
 
    if (!ZeroMask)
      return CreateLoad(VT, LDBase);
 
    // IsConsecutiveLoadWithZeros - we need to create a shuffle of the loaded
    // vector and a zero vector to clear out the zero elements.
    if (!IsAfterLegalize && VT.isVector()) {
      unsigned NumMaskElts = VT.getVectorNumElements();
      if ((NumMaskElts % NumElems) == 0) {
        unsigned Scale = NumMaskElts / NumElems;
        SmallVector<int, 4> ClearMask(NumMaskElts, -1);
        for (unsigned i = 0; i < NumElems; ++i) {
          if (UndefMask[i])
            continue;
          int Offset = ZeroMask[i] ? NumMaskElts : 0;
          for (unsigned j = 0; j != Scale; ++j)
            ClearMask[(i * Scale) + j] = (i * Scale) + j + Offset;
        }
        SDValue V = CreateLoad(VT, LDBase);
        SDValue Z = VT.isInteger() ? DAG.getConstant(0, DL, VT)
                                   : DAG.getConstantFP(0.0, DL, VT);
        return DAG.getVectorShuffle(VT, DL, V, Z, ClearMask);
      }
    }
  }
 
  // If the upper half of a ymm/zmm load is undef then just load the lower half.
  if (VT.is256BitVector() || VT.is512BitVector()) {
    unsigned HalfNumElems = NumElems / 2;
    if (UndefMask.extractBits(HalfNumElems, HalfNumElems).isAllOnes()) {
      EVT HalfVT =
          EVT::getVectorVT(*DAG.getContext(), VT.getScalarType(), HalfNumElems);
      SDValue HalfLD =
          EltsFromConsecutiveLoads(HalfVT, Elts.drop_back(HalfNumElems), DL,
                                   DAG, Subtarget, IsAfterLegalize, Depth + 1);
      if (HalfLD)
        return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT),
                           HalfLD, DAG.getVectorIdxConstant(0, DL));
    }
  }
 
  // VZEXT_LOAD - consecutive 32/64-bit load/undefs followed by zeros/undefs.
  if (IsConsecutiveLoad && FirstLoadedElt == 0 &&
      ((LoadSizeInBits == 16 && Subtarget.hasFP16()) || LoadSizeInBits == 32 ||
       LoadSizeInBits == 64) &&
      ((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()))) {
    MVT VecSVT = VT.isFloatingPoint() ? MVT::getFloatingPointVT(LoadSizeInBits)
                                      : MVT::getIntegerVT(LoadSizeInBits);
    MVT VecVT = MVT::getVectorVT(VecSVT, VT.getSizeInBits() / LoadSizeInBits);
    // Allow v4f32 on SSE1 only targets.
    // FIXME: Add more isel patterns so we can just use VT directly.
    if (!Subtarget.hasSSE2() && VT == MVT::v4f32)
      VecVT = MVT::v4f32;
    if (TLI.isTypeLegal(VecVT)) {
      SDVTList Tys = DAG.getVTList(VecVT, MVT::Other);
      SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
      SDValue ResNode = DAG.getMemIntrinsicNode(
          X86ISD::VZEXT_LOAD, DL, Tys, Ops, VecSVT, LDBase->getPointerInfo(),
          LDBase->getBaseAlign(), MachineMemOperand::MOLoad);
      for (auto *LD : Loads)
        if (LD)
          DAG.makeEquivalentMemoryOrdering(LD, ResNode);
      return DAG.getBitcast(VT, ResNode);
    }
  }
 
  // BROADCAST - match the smallest possible repetition pattern, load that
  // scalar/subvector element and then broadcast to the entire vector.
  if (ZeroMask.isZero() && isPowerOf2_32(NumElems) && Subtarget.hasAVX() &&
      (VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector())) {
    for (unsigned SubElems = 1; SubElems < NumElems; SubElems *= 2) {
      unsigned RepeatSize = SubElems * BaseSizeInBits;
      unsigned ScalarSize = std::min(RepeatSize, 64u);
      if (!Subtarget.hasAVX2() && ScalarSize < 32)
        continue;
 
      // Don't attempt a 1:N subvector broadcast - it should be caught by
      // combineConcatVectorOps, else will cause infinite loops.
      if (RepeatSize > ScalarSize && SubElems == 1)
        continue;
 
      bool Match = true;
      SmallVector<SDValue, 8> RepeatedLoads(SubElems, DAG.getUNDEF(EltBaseVT));
      for (unsigned i = 0; i != NumElems && Match; ++i) {
        if (!LoadMask[i])
          continue;
        SDValue Elt = peekThroughBitcasts(Elts[i]);
        if (RepeatedLoads[i % SubElems].isUndef())
          RepeatedLoads[i % SubElems] = Elt;
        else
          Match &= (RepeatedLoads[i % SubElems] == Elt);
      }
 
      // We must have loads at both ends of the repetition.
      Match &= !RepeatedLoads.front().isUndef();
      Match &= !RepeatedLoads.back().isUndef();
      if (!Match)
        continue;
 
      EVT RepeatVT =
          VT.isInteger() && (RepeatSize != 64 || TLI.isTypeLegal(MVT::i64))
              ? EVT::getIntegerVT(*DAG.getContext(), ScalarSize)
              : EVT::getFloatingPointVT(ScalarSize);
      if (RepeatSize > ScalarSize)
        RepeatVT = EVT::getVectorVT(*DAG.getContext(), RepeatVT,
                                    RepeatSize / ScalarSize);
      EVT BroadcastVT =
          EVT::getVectorVT(*DAG.getContext(), RepeatVT.getScalarType(),
                           VT.getSizeInBits() / ScalarSize);
      if (TLI.isTypeLegal(BroadcastVT)) {
        if (SDValue RepeatLoad = EltsFromConsecutiveLoads(
                RepeatVT, RepeatedLoads, DL, DAG, Subtarget, IsAfterLegalize,
                Depth + 1)) {
          SDValue Broadcast = RepeatLoad;
          if (RepeatSize > ScalarSize) {
            while (Broadcast.getValueSizeInBits() < VT.getSizeInBits())
              Broadcast = concatSubVectors(Broadcast, Broadcast, DAG, DL);
          } else {
            if (!Subtarget.hasAVX2() &&
                !X86::mayFoldLoadIntoBroadcastFromMem(
                    RepeatLoad, RepeatVT.getScalarType().getSimpleVT(),
                    Subtarget,
                    /*AssumeSingleUse=*/true))
              return SDValue();
            Broadcast =
                DAG.getNode(X86ISD::VBROADCAST, DL, BroadcastVT, RepeatLoad);
          }
          return DAG.getBitcast(VT, Broadcast);
        }
      }
    }
  }
 
  // REVERSE - attempt to match the loads in reverse and then shuffle back.
  // TODO: Do this for any permute or mismatching element counts.
  if (Depth == 0 && ZeroMask.isZero() && UndefMask.isZero() &&
      TLI.isTypeLegal(VT) && VT.isVector() &&
      NumElems == VT.getVectorNumElements()) {
    SmallVector<SDValue, 16> ReverseElts(Elts.rbegin(), Elts.rend());
    if (SDValue RevLd = EltsFromConsecutiveLoads(
            VT, ReverseElts, DL, DAG, Subtarget, IsAfterLegalize, Depth + 1)) {
      SmallVector<int, 16> ReverseMask(NumElems);
      std::iota(ReverseMask.rbegin(), ReverseMask.rend(), 0);
      return DAG.getVectorShuffle(VT, DL, RevLd, DAG.getUNDEF(VT), ReverseMask);
    }
  }
 
  return SDValue();
}
 
// Combine a vector ops (shuffles etc.) that is equal to build_vector load1,
// load2, load3, load4, <0, 1, 2, 3> into a vector load if the load addresses
// are consecutive, non-overlapping, and in the right order.
static SDValue combineToConsecutiveLoads(EVT VT, SDValue Op, const SDLoc &DL,
                                         SelectionDAG &DAG,
                                         const X86Subtarget &Subtarget,
                                         bool IsAfterLegalize) {
  SmallVector<SDValue, 64> Elts;
  for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
    if (SDValue Elt = getShuffleScalarElt(Op, i, DAG, 0)) {
      Elts.push_back(Elt);
      continue;
    }
    return SDValue();
  }
  assert(Elts.size() == VT.getVectorNumElements());
  return EltsFromConsecutiveLoads(VT, Elts, DL, DAG, Subtarget,
                                  IsAfterLegalize);
}
 
static Constant *getConstantVector(MVT VT, ArrayRef<APInt> Bits,
                                   const APInt &Undefs, LLVMContext &C) {
  unsigned ScalarSize = VT.getScalarSizeInBits();
  Type *Ty = EVT(VT.getScalarType()).getTypeForEVT(C);
 
  auto getConstantScalar = [&](const APInt &Val) -> Constant * {
    if (VT.isFloatingPoint()) {
      if (ScalarSize == 16)
        return ConstantFP::get(C, APFloat(APFloat::IEEEhalf(), Val));
      if (ScalarSize == 32)
        return ConstantFP::get(C, APFloat(APFloat::IEEEsingle(), Val));
      assert(ScalarSize == 64 && "Unsupported floating point scalar size");
      return ConstantFP::get(C, APFloat(APFloat::IEEEdouble(), Val));
    }
    return Constant::getIntegerValue(Ty, Val);
  };
 
  SmallVector<Constant *, 32> ConstantVec;
  for (unsigned I = 0, E = Bits.size(); I != E; ++I)
    ConstantVec.push_back(Undefs[I] ? UndefValue::get(Ty)
                                    : getConstantScalar(Bits[I]));
 
  return ConstantVector::get(ArrayRef<Constant *>(ConstantVec));
}
 
static Constant *getConstantVector(MVT VT, const APInt &SplatValue,
                                   unsigned SplatBitSize, LLVMContext &C) {
  unsigned ScalarSize = VT.getScalarSizeInBits();
 
  auto getConstantScalar = [&](const APInt &Val) -> Constant * {
    if (VT.isFloatingPoint()) {
      if (ScalarSize == 16)
        return ConstantFP::get(C, APFloat(APFloat::IEEEhalf(), Val));
      if (ScalarSize == 32)
        return ConstantFP::get(C, APFloat(APFloat::IEEEsingle(), Val));
      assert(ScalarSize == 64 && "Unsupported floating point scalar size");
      return ConstantFP::get(C, APFloat(APFloat::IEEEdouble(), Val));
    }
    return Constant::getIntegerValue(Type::getIntNTy(C, ScalarSize), Val);
  };
 
  if (ScalarSize == SplatBitSize)
    return getConstantScalar(SplatValue);
 
  unsigned NumElm = SplatBitSize / ScalarSize;
  SmallVector<Constant *, 32> ConstantVec;
  for (unsigned I = 0; I != NumElm; ++I) {
    APInt Val = SplatValue.extractBits(ScalarSize, ScalarSize * I);
    ConstantVec.push_back(getConstantScalar(Val));
  }
  return ConstantVector::get(ArrayRef<Constant *>(ConstantVec));
}
 
static bool isFoldableUseOfShuffle(SDNode *N) {
  for (auto *U : N->users()) {
    unsigned Opc = U->getOpcode();
    // VPERMV/VPERMV3 shuffles can never fold their index operands.
    if (Opc == X86ISD::VPERMV && U->getOperand(0).getNode() == N)
      return false;
    if (Opc == X86ISD::VPERMV3 && U->getOperand(1).getNode() == N)
      return false;
    if (isTargetShuffle(Opc))
      return true;
    if (Opc == ISD::BITCAST) // Ignore bitcasts
      return isFoldableUseOfShuffle(U);
    if (N->hasOneUse()) {
      // TODO, there may be some general way to know if a SDNode can
      // be folded. We now only know whether an MI is foldable.
      if (Opc == X86ISD::VPDPBUSD && U->getOperand(2).getNode() != N)
        return false;
      return true;
    }
  }
  return false;
}
 
// If the node has a single use by a VSELECT then AVX512 targets may be able to
// fold as a predicated instruction.
static bool isMaskableNode(SDValue V, const X86Subtarget &Subtarget) {
  unsigned SizeInBits = V.getValueSizeInBits();
  if ((SizeInBits == 512 && Subtarget.hasAVX512()) ||
      (SizeInBits >= 128 && Subtarget.hasVLX())) {
    if (V.hasOneUse() && V->user_begin()->getOpcode() == ISD::VSELECT &&
        V->user_begin()->getOperand(0).getScalarValueSizeInBits() == 1) {
      return true;
    }
  }
  return false;
}
 
/// Attempt to use the vbroadcast instruction to generate a splat value
/// from a splat BUILD_VECTOR which uses:
///  a. A single scalar load, or a constant.
///  b. Repeated pattern of constants (e.g. <0,1,0,1> or <0,1,2,3,0,1,2,3>).
///
/// The VBROADCAST node is returned when a pattern is found,
/// or SDValue() otherwise.
static SDValue lowerBuildVectorAsBroadcast(BuildVectorSDNode *BVOp,
                                           const SDLoc &dl,
                                           const X86Subtarget &Subtarget,
                                           SelectionDAG &DAG) {
  // VBROADCAST requires AVX.
  // TODO: Splats could be generated for non-AVX CPUs using SSE
  // instructions, but there's less potential gain for only 128-bit vectors.
  if (!Subtarget.hasAVX())
    return SDValue();
 
  MVT VT = BVOp->getSimpleValueType(0);
  unsigned NumElts = VT.getVectorNumElements();
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
         "Unsupported vector type for broadcast.");
 
  // See if the build vector is a repeating sequence of scalars (inc. splat).
  SDValue Ld;
  BitVector UndefElements;
  SmallVector<SDValue, 16> Sequence;
  if (BVOp->getRepeatedSequence(Sequence, &UndefElements)) {
    assert((NumElts % Sequence.size()) == 0 && "Sequence doesn't fit.");
    if (Sequence.size() == 1)
      Ld = Sequence[0];
  }
 
  // Attempt to use VBROADCASTM
  // From this pattern:
  // a. t0 = (zext_i64 (bitcast_i8 v2i1 X))
  // b. t1 = (build_vector t0 t0)
  //
  // Create (VBROADCASTM v2i1 X)
  if (!Sequence.empty() && Subtarget.hasCDI()) {
    // If not a splat, are the upper sequence values zeroable?
    unsigned SeqLen = Sequence.size();
    bool UpperZeroOrUndef =
        SeqLen == 1 ||
        llvm::all_of(ArrayRef(Sequence).drop_front(),
                     [](SDValue V) { return !V || isNullConstantOrUndef(V); });
    SDValue Op0 = Sequence[0];
    if (UpperZeroOrUndef && ((Op0.getOpcode() == ISD::BITCAST) ||
                             (Op0.getOpcode() == ISD::ZERO_EXTEND &&
                              Op0.getOperand(0).getOpcode() == ISD::BITCAST))) {
      SDValue BOperand = Op0.getOpcode() == ISD::BITCAST
                             ? Op0.getOperand(0)
                             : Op0.getOperand(0).getOperand(0);
      MVT MaskVT = BOperand.getSimpleValueType();
      MVT EltType = MVT::getIntegerVT(VT.getScalarSizeInBits() * SeqLen);
      if ((EltType == MVT::i64 && MaskVT == MVT::v8i1) ||  // for broadcastmb2q
          (EltType == MVT::i32 && MaskVT == MVT::v16i1)) { // for broadcastmw2d
        MVT BcstVT = MVT::getVectorVT(EltType, NumElts / SeqLen);
        if (!VT.is512BitVector() && !Subtarget.hasVLX()) {
          unsigned Scale = 512 / VT.getSizeInBits();
          BcstVT = MVT::getVectorVT(EltType, Scale * (NumElts / SeqLen));
        }
        SDValue Bcst = DAG.getNode(X86ISD::VBROADCASTM, dl, BcstVT, BOperand);
        if (BcstVT.getSizeInBits() != VT.getSizeInBits())
          Bcst = extractSubVector(Bcst, 0, DAG, dl, VT.getSizeInBits());
        return DAG.getBitcast(VT, Bcst);
      }
    }
  }
 
  unsigned NumUndefElts = UndefElements.count();
  if (!Ld || (NumElts - NumUndefElts) <= 1) {
    APInt SplatValue, Undef;
    unsigned SplatBitSize;
    bool HasUndef;
    // Check if this is a repeated constant pattern suitable for broadcasting.
    if (BVOp->isConstantSplat(SplatValue, Undef, SplatBitSize, HasUndef) &&
        SplatBitSize > VT.getScalarSizeInBits() &&
        SplatBitSize < VT.getSizeInBits()) {
      // Avoid replacing with broadcast when it's a use of a shuffle
      // instruction to preserve the present custom lowering of shuffles.
      if (isFoldableUseOfShuffle(BVOp))
        return SDValue();
      // replace BUILD_VECTOR with broadcast of the repeated constants.
      LLVMContext *Ctx = DAG.getContext();
      MVT PVT = TLI.getPointerTy(DAG.getDataLayout());
      if (SplatBitSize == 32 || SplatBitSize == 64 ||
          (SplatBitSize < 32 && Subtarget.hasAVX2())) {
        // Load the constant scalar/subvector and broadcast it.
        MVT CVT = MVT::getIntegerVT(SplatBitSize);
        Constant *C = getConstantVector(VT, SplatValue, SplatBitSize, *Ctx);
        SDValue CP = DAG.getConstantPool(C, PVT);
        unsigned Repeat = VT.getSizeInBits() / SplatBitSize;
 
        Align Alignment = cast<ConstantPoolSDNode>(CP)->getAlign();
        SDVTList Tys = DAG.getVTList(MVT::getVectorVT(CVT, Repeat), MVT::Other);
        SDValue Ops[] = {DAG.getEntryNode(), CP};
        MachinePointerInfo MPI =
            MachinePointerInfo::getConstantPool(DAG.getMachineFunction());
        SDValue Brdcst =
            DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, CVT,
                                    MPI, Alignment, MachineMemOperand::MOLoad);
        return DAG.getBitcast(VT, Brdcst);
      }
      if (SplatBitSize > 64) {
        // Load the vector of constants and broadcast it.
        Constant *VecC = getConstantVector(VT, SplatValue, SplatBitSize, *Ctx);
        SDValue VCP = DAG.getConstantPool(VecC, PVT);
        unsigned NumElm = SplatBitSize / VT.getScalarSizeInBits();
        MVT VVT = MVT::getVectorVT(VT.getScalarType(), NumElm);
        Align Alignment = cast<ConstantPoolSDNode>(VCP)->getAlign();
        SDVTList Tys = DAG.getVTList(VT, MVT::Other);
        SDValue Ops[] = {DAG.getEntryNode(), VCP};
        MachinePointerInfo MPI =
            MachinePointerInfo::getConstantPool(DAG.getMachineFunction());
        return DAG.getMemIntrinsicNode(X86ISD::SUBV_BROADCAST_LOAD, dl, Tys,
                                       Ops, VVT, MPI, Alignment,
                                       MachineMemOperand::MOLoad);
      }
    }
 
    // If we are moving a scalar into a vector (Ld must be set and all elements
    // but 1 are undef) and that operation is not obviously supported by
    // vmovd/vmovq/vmovss/vmovsd, then keep trying to form a broadcast.
    // That's better than general shuffling and may eliminate a load to GPR and
    // move from scalar to vector register.
    if (!Ld || NumElts - NumUndefElts != 1)
      return SDValue();
    unsigned ScalarSize = Ld.getValueSizeInBits();
    if (!(UndefElements[0] || (ScalarSize != 32 && ScalarSize != 64)))
      return SDValue();
  }
 
  bool ConstSplatVal =
      (Ld.getOpcode() == ISD::Constant || Ld.getOpcode() == ISD::ConstantFP);
  bool IsLoad = ISD::isNormalLoad(Ld.getNode());
 
  // TODO: Handle broadcasts of non-constant sequences.
 
  // Make sure that all of the users of a non-constant load are from the
  // BUILD_VECTOR node.
  // FIXME: Is the use count needed for non-constant, non-load case?
  if (!ConstSplatVal && !IsLoad && !BVOp->isOnlyUserOf(Ld.getNode()))
    return SDValue();
 
  unsigned ScalarSize = Ld.getValueSizeInBits();
  bool IsGE256 = (VT.getSizeInBits() >= 256);
 
  // When optimizing for size, generate up to 5 extra bytes for a broadcast
  // instruction to save 8 or more bytes of constant pool data.
  // TODO: If multiple splats are generated to load the same constant,
  // it may be detrimental to overall size. There needs to be a way to detect
  // that condition to know if this is truly a size win.
  bool OptForSize = DAG.shouldOptForSize();
 
  // Handle broadcasting a single constant scalar from the constant pool
  // into a vector.
  // On Sandybridge (no AVX2), it is still better to load a constant vector
  // from the constant pool and not to broadcast it from a scalar.
  // But override that restriction when optimizing for size.
  // TODO: Check if splatting is recommended for other AVX-capable CPUs.
  if (ConstSplatVal && (Subtarget.hasAVX2() || OptForSize)) {
    EVT CVT = Ld.getValueType();
    assert(!CVT.isVector() && "Must not broadcast a vector type");
 
    // Splat f16, f32, i32, v4f64, v4i64 in all cases with AVX2.
    // For size optimization, also splat v2f64 and v2i64, and for size opt
    // with AVX2, also splat i8 and i16.
    // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
    if (ScalarSize == 32 ||
        (ScalarSize == 64 && (IsGE256 || Subtarget.hasVLX())) ||
        (CVT == MVT::f16 && Subtarget.hasAVX2()) ||
        (OptForSize && (ScalarSize == 64 || Subtarget.hasAVX2()))) {
      const Constant *C = nullptr;
      if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
        C = CI->getConstantIntValue();
      else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
        C = CF->getConstantFPValue();
 
      assert(C && "Invalid constant type");
 
      SDValue CP =
          DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
      Align Alignment = cast<ConstantPoolSDNode>(CP)->getAlign();
 
      SDVTList Tys = DAG.getVTList(VT, MVT::Other);
      SDValue Ops[] = {DAG.getEntryNode(), CP};
      MachinePointerInfo MPI =
          MachinePointerInfo::getConstantPool(DAG.getMachineFunction());
      return DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, CVT,
                                     MPI, Alignment, MachineMemOperand::MOLoad);
    }
  }
 
  // Handle AVX2 in-register broadcasts.
  if (!IsLoad && Subtarget.hasInt256() &&
      (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
    return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
 
  // The scalar source must be a normal load.
  if (!IsLoad)
    return SDValue();
 
  // Make sure the non-chain result is only used by this build vector.
  if (!Ld->hasNUsesOfValue(NumElts - NumUndefElts, 0))
    return SDValue();
 
  if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
      (Subtarget.hasVLX() && ScalarSize == 64)) {
    auto *LN = cast<LoadSDNode>(Ld);
    SDVTList Tys = DAG.getVTList(VT, MVT::Other);
    SDValue Ops[] = {LN->getChain(), LN->getBasePtr()};
    SDValue BCast =
        DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops,
                                LN->getMemoryVT(), LN->getMemOperand());
    DAG.ReplaceAllUsesOfValueWith(SDValue(LN, 1), BCast.getValue(1));
    return BCast;
  }
 
  // The integer check is needed for the 64-bit into 128-bit so it doesn't match
  // double since there is no vbroadcastsd xmm
  if (Subtarget.hasInt256() && Ld.getValueType().isInteger() &&
      (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)) {
    auto *LN = cast<LoadSDNode>(Ld);
    SDVTList Tys = DAG.getVTList(VT, MVT::Other);
    SDValue Ops[] = {LN->getChain(), LN->getBasePtr()};
    SDValue BCast =
        DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops,
                                LN->getMemoryVT(), LN->getMemOperand());
    DAG.ReplaceAllUsesOfValueWith(SDValue(LN, 1), BCast.getValue(1));
    return BCast;
  }
 
  if (ScalarSize == 16 && Subtarget.hasFP16() && IsGE256)
    return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
 
  // Unsupported broadcast.
  return SDValue();
}
 
/// For an EXTRACT_VECTOR_ELT with a constant index return the real
/// underlying vector and index.
///
/// Modifies \p ExtractedFromVec to the real vector and returns the real
/// index.
static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
                                         SDValue ExtIdx) {
  int Idx = ExtIdx->getAsZExtVal();
  if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
    return Idx;
 
  // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
  // lowered this:
  //   (extract_vector_elt (v8f32 %1), Constant<6>)
  // to:
  //   (extract_vector_elt (vector_shuffle<2,u,u,u>
  //                           (extract_subvector (v8f32 %0), Constant<4>),
  //                           undef)
  //                       Constant<0>)
  // In this case the vector is the extract_subvector expression and the index
  // is 2, as specified by the shuffle.
  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
  SDValue ShuffleVec = SVOp->getOperand(0);
  MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
  assert(ShuffleVecVT.getVectorElementType() ==
         ExtractedFromVec.getSimpleValueType().getVectorElementType());
 
  int ShuffleIdx = SVOp->getMaskElt(Idx);
  if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
    ExtractedFromVec = ShuffleVec;
    return ShuffleIdx;
  }
  return Idx;
}
 
static SDValue buildFromShuffleMostly(SDValue Op, const SDLoc &DL,
                                      SelectionDAG &DAG) {
  MVT VT = Op.getSimpleValueType();
 
  // Skip if insert_vec_elt is not supported.
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
    return SDValue();
 
  unsigned NumElems = Op.getNumOperands();
  SDValue VecIn1;
  SDValue VecIn2;
  SmallVector<unsigned, 4> InsertIndices;
  SmallVector<int, 8> Mask(NumElems, -1);
 
  for (unsigned i = 0; i != NumElems; ++i) {
    unsigned Opc = Op.getOperand(i).getOpcode();
 
    if (Opc == ISD::POISON || Opc == ISD::UNDEF)
      continue;
 
    if (Opc != ISD::EXTRACT_VECTOR_ELT) {
      // Quit if more than 1 elements need inserting.
      if (InsertIndices.size() > 1)
        return SDValue();
 
      InsertIndices.push_back(i);
      continue;
    }
 
    SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
    SDValue ExtIdx = Op.getOperand(i).getOperand(1);
 
    // Quit if non-constant index.
    if (!isa<ConstantSDNode>(ExtIdx))
      return SDValue();
    int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
 
    // Quit if extracted from vector of different type.
    if (ExtractedFromVec.getValueType() != VT)
      return SDValue();
 
    if (!VecIn1.getNode())
      VecIn1 = ExtractedFromVec;
    else if (VecIn1 != ExtractedFromVec) {
      if (!VecIn2.getNode())
        VecIn2 = ExtractedFromVec;
      else if (VecIn2 != ExtractedFromVec)
        // Quit if more than 2 vectors to shuffle
        return SDValue();
    }
 
    if (ExtractedFromVec == VecIn1)
      Mask[i] = Idx;
    else if (ExtractedFromVec == VecIn2)
      Mask[i] = Idx + NumElems;
  }
 
  if (!VecIn1.getNode())
    return SDValue();
 
  VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getPOISON(VT);
  SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, Mask);
 
  for (unsigned Idx : InsertIndices)
    NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
                     DAG.getVectorIdxConstant(Idx, DL));
 
  return NV;
}
 
// Lower BUILD_VECTOR operation for v8bf16, v16bf16 and v32bf16 types.
static SDValue LowerBUILD_VECTORvXbf16(SDValue Op, SelectionDAG &DAG,
                                       const X86Subtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
  MVT SVT = Subtarget.hasFP16() ? MVT::f16 : MVT::i16;
  MVT IVT = VT.changeVectorElementType(SVT);
  SmallVector<SDValue, 16> NewOps;
  for (unsigned I = 0, E = Op.getNumOperands(); I != E; ++I)
    NewOps.push_back(DAG.getBitcast(SVT, Op.getOperand(I)));
  SDValue Res = DAG.getNode(ISD::BUILD_VECTOR, SDLoc(), IVT, NewOps);
  return DAG.getBitcast(VT, Res);
}
 
// Lower BUILD_VECTOR operation for vXi1 types.
static SDValue LowerBUILD_VECTORvXi1(SDValue Op, const SDLoc &dl,
                                     SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
 
  MVT VT = Op.getSimpleValueType();
  assert((VT.getVectorElementType() == MVT::i1) &&
         "Unexpected type in LowerBUILD_VECTORvXi1!");
  if (ISD::isBuildVectorAllZeros(Op.getNode()) ||
      ISD::isBuildVectorAllOnes(Op.getNode()))
    return Op;
 
  uint64_t Immediate = 0;
  SmallVector<unsigned, 16> NonConstIdx;
  bool IsSplat = true;
  bool HasConstElts = false;
  int SplatIdx = -1;
  for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
    SDValue In = Op.getOperand(idx);
    if (In.isUndef())
      continue;
    if (auto *InC = dyn_cast<ConstantSDNode>(In)) {
      Immediate |= (InC->getZExtValue() & 0x1) << idx;
      HasConstElts = true;
    } else {
      NonConstIdx.push_back(idx);
    }
    if (SplatIdx < 0)
      SplatIdx = idx;
    else if (In != Op.getOperand(SplatIdx))
      IsSplat = false;
  }
 
  // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
  if (IsSplat) {
    // The build_vector allows the scalar element to be larger than the vector
    // element type. We need to mask it to use as a condition unless we know
    // the upper bits are zero.
    // FIXME: Use computeKnownBits instead of checking specific opcode?
    SDValue Cond = Op.getOperand(SplatIdx);
    assert(Cond.getValueType() == MVT::i8 && "Unexpected VT!");
    if (Cond.getOpcode() != ISD::SETCC)
      Cond = DAG.getNode(ISD::AND, dl, MVT::i8, Cond,
                         DAG.getConstant(1, dl, MVT::i8));
 
    // Perform the select in the scalar domain so we can use cmov.
    if (VT == MVT::v64i1 && !Subtarget.is64Bit()) {
      SDValue Select = DAG.getSelect(dl, MVT::i32, Cond,
                                     DAG.getAllOnesConstant(dl, MVT::i32),
                                     DAG.getConstant(0, dl, MVT::i32));
      Select = DAG.getBitcast(MVT::v32i1, Select);
      return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Select, Select);
    } else {
      MVT ImmVT = MVT::getIntegerVT(std::max((unsigned)VT.getSizeInBits(), 8U));
      SDValue Select = DAG.getSelect(dl, ImmVT, Cond,
                                     DAG.getAllOnesConstant(dl, ImmVT),
                                     DAG.getConstant(0, dl, ImmVT));
      MVT VecVT = VT.getSizeInBits() >= 8 ? VT : MVT::v8i1;
      Select = DAG.getBitcast(VecVT, Select);
      return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Select,
                         DAG.getVectorIdxConstant(0, dl));
    }
  }
 
  // See if we can cheaply generate a vXi8 vector and convert to vXi1.
  MVT OpVT = Op.getOperand(0).getSimpleValueType();
  if (NonConstIdx.size() > 1 && OpVT == MVT::i8) {
    // On pre-BWI targets, we must extend to vXi32 instead.
    MVT ByteVT = VT.changeVectorElementType(MVT::i8);
    MVT WideSVT = Subtarget.hasBWI() ? MVT::i8 : MVT::i32;
    if (ByteVT.getSizeInBits() < 128) {
      WideSVT = ByteVT == MVT::v4i8 ? MVT::i32 : MVT::i64;
      ByteVT = MVT::v16i8;
    }
    MVT WideVT = VT.changeVectorElementType(WideSVT);
    if (DAG.getTargetLoweringInfo().isTypeLegal(ByteVT) &&
        DAG.getTargetLoweringInfo().isTypeLegal(WideVT)) {
      SmallVector<SDValue, 16> Elts(Op->op_values());
      Elts.append(ByteVT.getVectorNumElements() - Elts.size(),
                  DAG.getPOISON(OpVT));
      SDValue ByteBV = DAG.getBuildVector(ByteVT, dl, Elts);
      SDValue WideBV =
          getEXTEND_VECTOR_INREG(ISD::ANY_EXTEND, dl, WideVT, ByteBV, DAG);
      WideBV = DAG.getNode(ISD::AND, dl, WideVT, WideBV,
                           DAG.getConstant(1, dl, WideVT));
      return DAG.getSetCC(dl, VT, WideBV, DAG.getConstant(0, dl, WideVT),
                          ISD::SETNE);
    }
  }
 
  // insert elements one by one
  SDValue DstVec;
  if (HasConstElts) {
    if (VT == MVT::v64i1 && !Subtarget.is64Bit()) {
      SDValue ImmL = DAG.getConstant(Lo_32(Immediate), dl, MVT::i32);
      SDValue ImmH = DAG.getConstant(Hi_32(Immediate), dl, MVT::i32);
      ImmL = DAG.getBitcast(MVT::v32i1, ImmL);
      ImmH = DAG.getBitcast(MVT::v32i1, ImmH);
      DstVec = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, ImmL, ImmH);
    } else {
      MVT ImmVT = MVT::getIntegerVT(std::max((unsigned)VT.getSizeInBits(), 8U));
      SDValue Imm = DAG.getConstant(Immediate, dl, ImmVT);
      MVT VecVT = VT.getSizeInBits() >= 8 ? VT : MVT::v8i1;
      DstVec = DAG.getBitcast(VecVT, Imm);
      DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, DstVec,
                           DAG.getVectorIdxConstant(0, dl));
    }
  } else
    DstVec = DAG.getUNDEF(VT);
 
  for (unsigned InsertIdx : NonConstIdx) {
    DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
                         Op.getOperand(InsertIdx),
                         DAG.getVectorIdxConstant(InsertIdx, dl));
  }
  return DstVec;
}
 
[[maybe_unused]] static bool isHorizOp(unsigned Opcode) {
  switch (Opcode) {
  case X86ISD::PACKSS:
  case X86ISD::PACKUS:
  case X86ISD::FHADD:
  case X86ISD::FHSUB:
  case X86ISD::HADD:
  case X86ISD::HSUB:
  case X86ISD::HADDS:
  case X86ISD::HSUBS:
    return true;
  }
  return false;
}
 
/// This is a helper function of LowerToHorizontalOp().
/// This function checks that the build_vector \p N in input implements a
/// 128-bit partial horizontal operation on a 256-bit vector, but that operation
/// may not match the layout of an x86 256-bit horizontal instruction.
/// In other words, if this returns true, then some extraction/insertion will
/// be required to produce a valid horizontal instruction.
///
/// Parameter \p Opcode defines the kind of horizontal operation to match.
/// For example, if \p Opcode is equal to ISD::ADD, then this function
/// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
/// is equal to ISD::SUB, then this function checks if this is a horizontal
/// arithmetic sub.
///
/// This function only analyzes elements of \p N whose indices are
/// in range [BaseIdx, LastIdx).
///
/// TODO: This function was originally used to match both real and fake partial
/// horizontal operations, but the index-matching logic is incorrect for that.
/// See the corrected implementation in isHopBuildVector(). Can we reduce this
/// code because it is only used for partial h-op matching now?
static bool isHorizontalBinOpPart(const BuildVectorSDNode *N, unsigned Opcode,
                                  const SDLoc &DL, SelectionDAG &DAG,
                                  unsigned BaseIdx, unsigned LastIdx,
                                  SDValue &V0, SDValue &V1) {
  EVT VT = N->getValueType(0);
  assert(VT.is256BitVector() && "Only use for matching partial 256-bit h-ops");
  assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
  assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
         "Invalid Vector in input!");
 
  bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
  bool CanFold = true;
  unsigned ExpectedVExtractIdx = BaseIdx;
  unsigned NumElts = LastIdx - BaseIdx;
  V0 = DAG.getUNDEF(VT);
  V1 = DAG.getUNDEF(VT);
 
  // Check if N implements a horizontal binop.
  for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
    SDValue Op = N->getOperand(i + BaseIdx);
 
    // Skip UNDEFs.
    if (Op->isUndef()) {
      // Update the expected vector extract index.
      if (i * 2 == NumElts)
        ExpectedVExtractIdx = BaseIdx;
      ExpectedVExtractIdx += 2;
      continue;
    }
 
    CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
 
    if (!CanFold)
      break;
 
    SDValue Op0 = Op.getOperand(0);
    SDValue Op1 = Op.getOperand(1);
 
    // Try to match the following pattern:
    // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
    CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
        Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
        Op0.getOperand(0) == Op1.getOperand(0) &&
        isa<ConstantSDNode>(Op0.getOperand(1)) &&
        isa<ConstantSDNode>(Op1.getOperand(1)));
    if (!CanFold)
      break;
 
    unsigned I0 = Op0.getConstantOperandVal(1);
    unsigned I1 = Op1.getConstantOperandVal(1);
 
    if (i * 2 < NumElts) {
      if (V0.isUndef()) {
        V0 = Op0.getOperand(0);
        if (V0.getValueType() != VT)
          return false;
      }
    } else {
      if (V1.isUndef()) {
        V1 = Op0.getOperand(0);
        if (V1.getValueType() != VT)
          return false;
      }
      if (i * 2 == NumElts)
        ExpectedVExtractIdx = BaseIdx;
    }
 
    SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
    if (I0 == ExpectedVExtractIdx)
      CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
    else if (IsCommutable && I1 == ExpectedVExtractIdx) {
      // Try to match the following dag sequence:
      // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
      CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
    } else
      CanFold = false;
 
    ExpectedVExtractIdx += 2;
  }
 
  return CanFold;
}
 
/// Emit a sequence of two 128-bit horizontal add/sub followed by
/// a concat_vector.
///
/// This is a helper function of LowerToHorizontalOp().
/// This function expects two 256-bit vectors called V0 and V1.
/// At first, each vector is split into two separate 128-bit vectors.
/// Then, the resulting 128-bit vectors are used to implement two
/// horizontal binary operations.
///
/// The kind of horizontal binary operation is defined by \p X86Opcode.
///
/// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
/// the two new horizontal binop.
/// When Mode is set, the first horizontal binop dag node would take as input
/// the lower 128-bit of V0 and the upper 128-bit of V0. The second
/// horizontal binop dag node would take as input the lower 128-bit of V1
/// and the upper 128-bit of V1.
///   Example:
///     HADD V0_LO, V0_HI
///     HADD V1_LO, V1_HI
///
/// Otherwise, the first horizontal binop dag node takes as input the lower
/// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
/// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
///   Example:
///     HADD V0_LO, V1_LO
///     HADD V0_HI, V1_HI
///
/// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
/// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
/// the upper 128-bits of the result.
static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
                                     const SDLoc &DL, SelectionDAG &DAG,
                                     unsigned X86Opcode, bool Mode,
                                     bool isUndefLO, bool isUndefHI) {
  MVT VT = V0.getSimpleValueType();
  assert(VT.is256BitVector() && VT == V1.getSimpleValueType() &&
         "Invalid nodes in input!");
 
  unsigned NumElts = VT.getVectorNumElements();
  SDValue V0_LO = extract128BitVector(V0, 0, DAG, DL);
  SDValue V0_HI = extract128BitVector(V0, NumElts/2, DAG, DL);
  SDValue V1_LO = extract128BitVector(V1, 0, DAG, DL);
  SDValue V1_HI = extract128BitVector(V1, NumElts/2, DAG, DL);
  MVT NewVT = V0_LO.getSimpleValueType();
 
  SDValue LO = DAG.getUNDEF(NewVT);
  SDValue HI = DAG.getUNDEF(NewVT);
 
  if (Mode) {
    // Don't emit a horizontal binop if the result is expected to be UNDEF.
    if (!isUndefLO && !V0->isUndef())
      LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
    if (!isUndefHI && !V1->isUndef())
      HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
  } else {
    // Don't emit a horizontal binop if the result is expected to be UNDEF.
    if (!isUndefLO && (!V0_LO->isUndef() || !V1_LO->isUndef()))
      LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
 
    if (!isUndefHI && (!V0_HI->isUndef() || !V1_HI->isUndef()))
      HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
  }
 
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
}
 
/// Returns true iff \p BV builds a vector with the result equivalent to
/// the result of ADDSUB/SUBADD operation.
/// If true is returned then the operands of ADDSUB = Opnd0 +- Opnd1
/// (SUBADD = Opnd0 -+ Opnd1) operation are written to the parameters
/// \p Opnd0 and \p Opnd1.
static bool isAddSubOrSubAdd(const BuildVectorSDNode *BV,
                             const X86Subtarget &Subtarget, SelectionDAG &DAG,
                             SDValue &Opnd0, SDValue &Opnd1,
                             unsigned &NumExtracts, bool &IsSubAdd,
                             bool &HasAllowContract) {
  using namespace SDPatternMatch;
 
  MVT VT = BV->getSimpleValueType(0);
  if (!Subtarget.hasSSE3() || !VT.isFloatingPoint())
    return false;
 
  unsigned NumElts = VT.getVectorNumElements();
  SDValue InVec0 = DAG.getUNDEF(VT);
  SDValue InVec1 = DAG.getUNDEF(VT);
 
  NumExtracts = 0;
  HasAllowContract = NumElts != 0;
 
  // Odd-numbered elements in the input build vector are obtained from
  // adding/subtracting two integer/float elements.
  // Even-numbered elements in the input build vector are obtained from
  // subtracting/adding two integer/float elements.
  unsigned Opc[2] = {0, 0};
  for (unsigned i = 0, e = NumElts; i != e; ++i) {
    SDValue Op = BV->getOperand(i);
 
    // Skip 'undef' values.
    unsigned Opcode = Op.getOpcode();
    if (Opcode == ISD::UNDEF)
      continue;
 
    // Early exit if we found an unexpected opcode.
    if (Opcode != ISD::FADD && Opcode != ISD::FSUB)
      return false;
 
    SDValue Op0 = Op.getOperand(0);
    SDValue Op1 = Op.getOperand(1);
 
    // Try to match the following pattern:
    // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
    // Early exit if we cannot match that sequence.
    if (!sd_match(Op0, m_ExtractElt(m_SpecificVT(VT), m_SpecificInt(i))) ||
        !sd_match(Op1, m_ExtractElt(m_SpecificVT(VT), m_SpecificInt(i))))
      return false;
 
    // We found a valid add/sub node, make sure its the same opcode as previous
    // elements for this parity.
    if (Opc[i % 2] != 0 && Opc[i % 2] != Opcode)
      return false;
    Opc[i % 2] = Opcode;
 
    // Update InVec0 and InVec1.
    if (InVec0.isUndef())
      InVec0 = Op0.getOperand(0);
    if (InVec1.isUndef())
      InVec1 = Op1.getOperand(0);
 
    // Make sure that operands in input to each add/sub node always
    // come from a same pair of vectors.
    if (InVec0 != Op0.getOperand(0)) {
      if (Opcode == ISD::FSUB)
        return false;
 
      // FADD is commutable. Try to commute the operands
      // and then test again.
      std::swap(Op0, Op1);
      if (InVec0 != Op0.getOperand(0))
        return false;
    }
 
    if (InVec1 != Op1.getOperand(0))
      return false;
 
    // Increment the number of extractions done.
    ++NumExtracts;
    HasAllowContract &= Op->getFlags().hasAllowContract();
  }
 
  // Ensure we have found an opcode for both parities and that they are
  // different. Don't try to fold this build_vector into an ADDSUB/SUBADD if the
  // inputs are undef.
  if (!Opc[0] || !Opc[1] || Opc[0] == Opc[1] ||
      InVec0.isUndef() || InVec1.isUndef())
    return false;
 
  IsSubAdd = Opc[0] == ISD::FADD;
 
  Opnd0 = InVec0;
  Opnd1 = InVec1;
  return true;
}
 
/// Returns true if is possible to fold MUL and an idiom that has already been
/// recognized as ADDSUB/SUBADD(\p Opnd0, \p Opnd1) into
/// FMADDSUB/FMSUBADD(x, y, \p Opnd1). If (and only if) true is returned, the
/// operands of FMADDSUB/FMSUBADD are written to parameters \p Opnd0, \p Opnd1, \p Opnd2.
///
/// Prior to calling this function it should be known that there is some
/// SDNode that potentially can be replaced with an X86ISD::ADDSUB operation
/// using \p Opnd0 and \p Opnd1 as operands. Also, this method is called
/// before replacement of such SDNode with ADDSUB operation. Thus the number
/// of \p Opnd0 uses is expected to be equal to 2.
/// For example, this function may be called for the following IR:
///    %AB = fmul fast <2 x double> %A, %B
///    %Sub = fsub fast <2 x double> %AB, %C
///    %Add = fadd fast <2 x double> %AB, %C
///    %Addsub = shufflevector <2 x double> %Sub, <2 x double> %Add,
///                            <2 x i32> <i32 0, i32 3>
/// There is a def for %Addsub here, which potentially can be replaced by
/// X86ISD::ADDSUB operation:
///    %Addsub = X86ISD::ADDSUB %AB, %C
/// and such ADDSUB can further be replaced with FMADDSUB:
///    %Addsub = FMADDSUB %A, %B, %C.
///
/// The main reason why this method is called before the replacement of the
/// recognized ADDSUB idiom with ADDSUB operation is that such replacement
/// is illegal sometimes. E.g. 512-bit ADDSUB is not available, while 512-bit
/// FMADDSUB is.
static bool isFMAddSubOrFMSubAdd(const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG, SDValue &Opnd0,
                                 SDValue &Opnd1, SDValue &Opnd2,
                                 unsigned ExpectedUses,
                                 bool AllowSubAddOrAddSubContract) {
  if (Opnd0.getOpcode() != ISD::FMUL ||
      !Opnd0->hasNUsesOfValue(ExpectedUses, 0) || !Subtarget.hasAnyFMA())
    return false;
 
  // FIXME: These checks must match the similar ones in
  // DAGCombiner::visitFADDForFMACombine. It would be good to have one
  // function that would answer if it is Ok to fuse MUL + ADD to FMADD
  // or MUL + ADDSUB to FMADDSUB.
  bool AllowFusion =
      (AllowSubAddOrAddSubContract && Opnd0->getFlags().hasAllowContract());
  if (!AllowFusion)
    return false;
 
  Opnd2 = Opnd1;
  Opnd1 = Opnd0.getOperand(1);
  Opnd0 = Opnd0.getOperand(0);
 
  return true;
}
 
/// Try to fold a build_vector that performs an 'addsub' or 'fmaddsub' or
/// 'fsubadd' operation accordingly to X86ISD::ADDSUB or X86ISD::FMADDSUB or
/// X86ISD::FMSUBADD node.
static SDValue lowerToAddSubOrFMAddSub(const BuildVectorSDNode *BV,
                                       const SDLoc &DL,
                                       const X86Subtarget &Subtarget,
                                       SelectionDAG &DAG) {
  SDValue Opnd0, Opnd1;
  unsigned NumExtracts;
  bool IsSubAdd;
  bool HasAllowContract;
  if (!isAddSubOrSubAdd(BV, Subtarget, DAG, Opnd0, Opnd1, NumExtracts, IsSubAdd,
                        HasAllowContract))
    return SDValue();
 
  MVT VT = BV->getSimpleValueType(0);
 
  // Try to generate X86ISD::FMADDSUB node here.
  SDValue Opnd2;
  if (isFMAddSubOrFMSubAdd(Subtarget, DAG, Opnd0, Opnd1, Opnd2, NumExtracts,
                           HasAllowContract)) {
    unsigned Opc = IsSubAdd ? X86ISD::FMSUBADD : X86ISD::FMADDSUB;
    return DAG.getNode(Opc, DL, VT, Opnd0, Opnd1, Opnd2);
  }
 
  // We only support ADDSUB.
  if (IsSubAdd)
    return SDValue();
 
  // There are no known X86 targets with 512-bit ADDSUB instructions!
  // Convert to blend(fsub,fadd).
  if (VT.is512BitVector()) {
    SmallVector<int> Mask;
    for (int I = 0, E = VT.getVectorNumElements(); I != E; I += 2) {
        Mask.push_back(I);
        Mask.push_back(I + E + 1);
    }
    SDValue Sub = DAG.getNode(ISD::FSUB, DL, VT, Opnd0, Opnd1);
    SDValue Add = DAG.getNode(ISD::FADD, DL, VT, Opnd0, Opnd1);
    return DAG.getVectorShuffle(VT, DL, Sub, Add, Mask);
  }
 
  return DAG.getNode(X86ISD::ADDSUB, DL, VT, Opnd0, Opnd1);
}
 
static bool isHopBuildVector(const BuildVectorSDNode *BV, SelectionDAG &DAG,
                             unsigned &HOpcode, SDValue &V0, SDValue &V1) {
  // Initialize outputs to known values.
  MVT VT = BV->getSimpleValueType(0);
  HOpcode = ISD::DELETED_NODE;
  V0 = DAG.getUNDEF(VT);
  V1 = DAG.getUNDEF(VT);
 
  // x86 256-bit horizontal ops are defined in a non-obvious way. Each 128-bit
  // half of the result is calculated independently from the 128-bit halves of
  // the inputs, so that makes the index-checking logic below more complicated.
  unsigned NumElts = VT.getVectorNumElements();
  unsigned GenericOpcode = ISD::DELETED_NODE;
  unsigned Num128BitChunks = VT.is256BitVector() ? 2 : 1;
  unsigned NumEltsIn128Bits = NumElts / Num128BitChunks;
  unsigned NumEltsIn64Bits = NumEltsIn128Bits / 2;
  for (unsigned i = 0; i != Num128BitChunks; ++i) {
    for (unsigned j = 0; j != NumEltsIn128Bits; ++j) {
      // Ignore undef elements.
      SDValue Op = BV->getOperand(i * NumEltsIn128Bits + j);
      if (Op.isUndef())
        continue;
 
      // If there's an opcode mismatch, we're done.
      if (HOpcode != ISD::DELETED_NODE && Op.getOpcode() != GenericOpcode)
        return false;
 
      // Initialize horizontal opcode.
      if (HOpcode == ISD::DELETED_NODE) {
        GenericOpcode = Op.getOpcode();
        switch (GenericOpcode) {
        // clang-format off
        case ISD::ADD: HOpcode = X86ISD::HADD; break;
        case ISD::SUB: HOpcode = X86ISD::HSUB; break;
        case ISD::FADD: HOpcode = X86ISD::FHADD; break;
        case ISD::FSUB: HOpcode = X86ISD::FHSUB; break;
        default: return false;
        // clang-format on
        }
      }
 
      SDValue Op0 = Op.getOperand(0);
      SDValue Op1 = Op.getOperand(1);
      if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
          Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
          Op0.getOperand(0) != Op1.getOperand(0) ||
          !isa<ConstantSDNode>(Op0.getOperand(1)) ||
          !isa<ConstantSDNode>(Op1.getOperand(1)) || !Op.hasOneUse())
        return false;
 
      // The source vector is chosen based on which 64-bit half of the
      // destination vector is being calculated.
      if (j < NumEltsIn64Bits) {
        if (V0.isUndef())
          V0 = Op0.getOperand(0);
      } else {
        if (V1.isUndef())
          V1 = Op0.getOperand(0);
      }
 
      SDValue SourceVec = (j < NumEltsIn64Bits) ? V0 : V1;
      if (SourceVec != Op0.getOperand(0))
        return false;
 
      // op (extract_vector_elt A, I), (extract_vector_elt A, I+1)
      unsigned ExtIndex0 = Op0.getConstantOperandVal(1);
      unsigned ExtIndex1 = Op1.getConstantOperandVal(1);
      unsigned ExpectedIndex = i * NumEltsIn128Bits +
                               (j % NumEltsIn64Bits) * 2;
      if (ExpectedIndex == ExtIndex0 && ExtIndex1 == ExtIndex0 + 1)
        continue;
 
      // If this is not a commutative op, this does not match.
      if (GenericOpcode != ISD::ADD && GenericOpcode != ISD::FADD)
        return false;
 
      // Addition is commutative, so try swapping the extract indexes.
      // op (extract_vector_elt A, I+1), (extract_vector_elt A, I)
      if (ExpectedIndex == ExtIndex1 && ExtIndex0 == ExtIndex1 + 1)
        continue;
 
      // Extract indexes do not match horizontal requirement.
      return false;
    }
  }
  // We matched. Opcode and operands are returned by reference as arguments.
  return true;
}
 
static SDValue getHopForBuildVector(const BuildVectorSDNode *BV,
                                    const SDLoc &DL, SelectionDAG &DAG,
                                    unsigned HOpcode, SDValue V0, SDValue V1) {
  // If either input vector is not the same size as the build vector,
  // extract/insert the low bits to the correct size.
  // This is free (examples: zmm --> xmm, xmm --> ymm).
  MVT VT = BV->getSimpleValueType(0);
  unsigned Width = VT.getSizeInBits();
  if (V0.getValueSizeInBits() > Width)
    V0 = extractSubVector(V0, 0, DAG, DL, Width);
  else if (V0.getValueSizeInBits() < Width)
    V0 = insertSubVector(DAG.getUNDEF(VT), V0, 0, DAG, DL, Width);
 
  if (V1.getValueSizeInBits() > Width)
    V1 = extractSubVector(V1, 0, DAG, DL, Width);
  else if (V1.getValueSizeInBits() < Width)
    V1 = insertSubVector(DAG.getUNDEF(VT), V1, 0, DAG, DL, Width);
 
  unsigned NumElts = VT.getVectorNumElements();
  APInt DemandedElts = APInt::getAllOnes(NumElts);
  for (unsigned i = 0; i != NumElts; ++i)
    if (BV->getOperand(i).isUndef())
      DemandedElts.clearBit(i);
 
  // If we don't need the upper xmm, then perform as a xmm hop.
  unsigned HalfNumElts = NumElts / 2;
  if (VT.is256BitVector() && DemandedElts.lshr(HalfNumElts) == 0) {
    MVT HalfVT = VT.getHalfNumVectorElementsVT();
    V0 = extractSubVector(V0, 0, DAG, DL, 128);
    V1 = extractSubVector(V1, 0, DAG, DL, 128);
    SDValue Half = DAG.getNode(HOpcode, DL, HalfVT, V0, V1);
    return insertSubVector(DAG.getUNDEF(VT), Half, 0, DAG, DL, 256);
  }
 
  return DAG.getNode(HOpcode, DL, VT, V0, V1);
}
 
/// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV, const SDLoc &DL,
                                   const X86Subtarget &Subtarget,
                                   SelectionDAG &DAG) {
  // We need at least 2 non-undef elements to make this worthwhile by default.
  unsigned NumNonUndefs =
      count_if(BV->op_values(), [](SDValue V) { return !V.isUndef(); });
  if (NumNonUndefs < 2)
    return SDValue();
 
  // There are 4 sets of horizontal math operations distinguished by type:
  // int/FP at 128-bit/256-bit. Each type was introduced with a different
  // subtarget feature. Try to match those "native" patterns first.
  MVT VT = BV->getSimpleValueType(0);
  if (((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget.hasSSE3()) ||
      ((VT == MVT::v8i16 || VT == MVT::v4i32) && Subtarget.hasSSSE3()) ||
      ((VT == MVT::v8f32 || VT == MVT::v4f64) && Subtarget.hasAVX()) ||
      ((VT == MVT::v16i16 || VT == MVT::v8i32) && Subtarget.hasAVX2())) {
    unsigned HOpcode;
    SDValue V0, V1;
    if (isHopBuildVector(BV, DAG, HOpcode, V0, V1))
      return getHopForBuildVector(BV, DL, DAG, HOpcode, V0, V1);
  }
 
  // Try harder to match 256-bit ops by using extract/concat.
  if (!Subtarget.hasAVX() || !VT.is256BitVector())
    return SDValue();
 
  // Count the number of UNDEF operands in the build_vector in input.
  unsigned NumElts = VT.getVectorNumElements();
  unsigned Half = NumElts / 2;
  unsigned NumUndefsLO = 0;
  unsigned NumUndefsHI = 0;
  for (unsigned i = 0, e = Half; i != e; ++i)
    if (BV->getOperand(i)->isUndef())
      NumUndefsLO++;
 
  for (unsigned i = Half, e = NumElts; i != e; ++i)
    if (BV->getOperand(i)->isUndef())
      NumUndefsHI++;
 
  SDValue InVec0, InVec1;
  if (VT == MVT::v8i32 || VT == MVT::v16i16) {
    SDValue InVec2, InVec3;
    unsigned X86Opcode;
    bool CanFold = true;
 
    if (isHorizontalBinOpPart(BV, ISD::ADD, DL, DAG, 0, Half, InVec0, InVec1) &&
        isHorizontalBinOpPart(BV, ISD::ADD, DL, DAG, Half, NumElts, InVec2,
                              InVec3) &&
        ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) &&
        ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3))
      X86Opcode = X86ISD::HADD;
    else if (isHorizontalBinOpPart(BV, ISD::SUB, DL, DAG, 0, Half, InVec0,
                                   InVec1) &&
             isHorizontalBinOpPart(BV, ISD::SUB, DL, DAG, Half, NumElts, InVec2,
                                   InVec3) &&
             ((InVec0.isUndef() || InVec2.isUndef()) || InVec0 == InVec2) &&
             ((InVec1.isUndef() || InVec3.isUndef()) || InVec1 == InVec3))
      X86Opcode = X86ISD::HSUB;
    else
      CanFold = false;
 
    if (CanFold) {
      // Do not try to expand this build_vector into a pair of horizontal
      // add/sub if we can emit a pair of scalar add/sub.
      if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
        return SDValue();
 
      // Convert this build_vector into a pair of horizontal binops followed by
      // a concat vector. We must adjust the outputs from the partial horizontal
      // matching calls above to account for undefined vector halves.
      SDValue V0 = InVec0.isUndef() ? InVec2 : InVec0;
      SDValue V1 = InVec1.isUndef() ? InVec3 : InVec1;
      assert((!V0.isUndef() || !V1.isUndef()) && "Horizontal-op of undefs?");
      bool isUndefLO = NumUndefsLO == Half;
      bool isUndefHI = NumUndefsHI == Half;
      return ExpandHorizontalBinOp(V0, V1, DL, DAG, X86Opcode, false, isUndefLO,
                                   isUndefHI);
    }
  }
 
  if (VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
      VT == MVT::v16i16) {
    unsigned X86Opcode;
    if (isHorizontalBinOpPart(BV, ISD::ADD, DL, DAG, 0, NumElts, InVec0,
                              InVec1))
      X86Opcode = X86ISD::HADD;
    else if (isHorizontalBinOpPart(BV, ISD::SUB, DL, DAG, 0, NumElts, InVec0,
                                   InVec1))
      X86Opcode = X86ISD::HSUB;
    else if (isHorizontalBinOpPart(BV, ISD::FADD, DL, DAG, 0, NumElts, InVec0,
                                   InVec1))
      X86Opcode = X86ISD::FHADD;
    else if (isHorizontalBinOpPart(BV, ISD::FSUB, DL, DAG, 0, NumElts, InVec0,
                                   InVec1))
      X86Opcode = X86ISD::FHSUB;
    else
      return SDValue();
 
    // Don't try to expand this build_vector into a pair of horizontal add/sub
    // if we can simply emit a pair of scalar add/sub.
    if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
      return SDValue();
 
    // Convert this build_vector into two horizontal add/sub followed by
    // a concat vector.
    bool isUndefLO = NumUndefsLO == Half;
    bool isUndefHI = NumUndefsHI == Half;
    return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
                                 isUndefLO, isUndefHI);
  }
 
  return SDValue();
}
 
static SDValue LowerShift(SDValue Op, const X86Subtarget &Subtarget,
                          SelectionDAG &DAG);
 
/// If a BUILD_VECTOR's source elements all apply the same bit operation and
/// one of their operands is constant, lower to a pair of BUILD_VECTOR and
/// just apply the bit to the vectors.
/// NOTE: Its not in our interest to start make a general purpose vectorizer
/// from this, but enough scalar bit operations are created from the later
/// legalization + scalarization stages to need basic support.
static SDValue lowerBuildVectorToBitOp(BuildVectorSDNode *Op, const SDLoc &DL,
                                       const X86Subtarget &Subtarget,
                                       SelectionDAG &DAG) {
  MVT VT = Op->getSimpleValueType(0);
  unsigned NumElems = VT.getVectorNumElements();
  unsigned ElemSize = VT.getScalarSizeInBits();
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
  // Check that all elements have the same opcode.
  // TODO: Should we allow UNDEFS and if so how many?
  unsigned Opcode = Op->getOperand(0).getOpcode();
  for (unsigned i = 1; i < NumElems; ++i)
    if (Opcode != Op->getOperand(i).getOpcode())
      return SDValue();
 
  // TODO: We may be able to add support for other Ops (e.g. ADD/SUB).
  bool IsShift = false;
  switch (Opcode) {
  default:
    return SDValue();
  case ISD::SHL:
  case ISD::SRL:
  case ISD::SRA:
    IsShift = true;
    break;
  case ISD::AND:
  case ISD::XOR:
  case ISD::OR:
    // Don't do this if the buildvector is a splat - we'd replace one
    // constant with an entire vector.
    if (Op->getSplatValue())
      return SDValue();
    if (!TLI.isOperationLegalOrPromote(Opcode, VT))
      return SDValue();
    break;
  }
 
  // Collect elements.
  bool RHSAllConst = true;
  SmallVector<SDValue, 4> LHSElts, RHSElts;
  for (SDValue Elt : Op->ops()) {
    SDValue LHS = Elt.getOperand(0);
    SDValue RHS = Elt.getOperand(1);
    RHSAllConst &= isa<ConstantSDNode>(RHS);
    LHSElts.push_back(LHS);
    RHSElts.push_back(RHS);
  }
 
  // Canonicalize shift amounts.
  if (IsShift) {
    // We expect the canonicalized RHS operand to be the constant.
    // TODO: Permit non-constant XOP/AVX2 cases?
    if (!RHSAllConst)
      return SDValue();
 
    // Extend shift amounts.
    for (SDValue &Op1 : RHSElts)
      if (Op1.getValueSizeInBits() != ElemSize)
        Op1 = DAG.getZExtOrTrunc(Op1, DL, VT.getScalarType());
 
    // Limit to shifts by uniform immediates.
    // TODO: Only accept vXi8/vXi64 special cases?
    // TODO: Permit non-uniform XOP/AVX2/MULLO cases?
    if (any_of(RHSElts, [&](SDValue V) { return RHSElts[0] != V; }))
      return SDValue();
  }
  assert(all_of(llvm::concat<SDValue>(LHSElts, RHSElts),
                [ElemSize](SDValue V) {
                  return V.getValueSizeInBits() == ElemSize;
                }) &&
         "Element size mismatch");
 
  // To avoid an increase in GPR->FPU instructions, LHS/RHS must be foldable as
  // a load or RHS must be constant.
  SDValue LHS = EltsFromConsecutiveLoads(VT, LHSElts, DL, DAG, Subtarget,
                                         /*IsAfterLegalize=*/true);
  SDValue RHS = EltsFromConsecutiveLoads(VT, RHSElts, DL, DAG, Subtarget,
                                         /*IsAfterLegalize=*/true);
  if (!LHS && !RHS && !RHSAllConst)
    return SDValue();
 
  if (!LHS)
    LHS = DAG.getBuildVector(VT, DL, LHSElts);
  if (!RHS)
    RHS = DAG.getBuildVector(VT, DL, RHSElts);
  SDValue Res = DAG.getNode(Opcode, DL, VT, LHS, RHS);
 
  if (!IsShift)
    return Res;
 
  // Immediately lower the shift to ensure the constant build vector doesn't
  // get converted to a constant pool before the shift is lowered.
  return LowerShift(Res, Subtarget, DAG);
}
 
static bool isShuffleFoldableLoad(SDValue);
 
/// Attempt to lower a BUILD_VECTOR of scalar values to a shuffle of splats
/// representing a blend.
static SDValue lowerBuildVectorAsBlend(BuildVectorSDNode *BVOp, SDLoc const &DL,
                                       X86Subtarget const &Subtarget,
                                       SelectionDAG &DAG) {
  MVT VT = BVOp->getSimpleValueType(0u);
 
  if (VT != MVT::v4f64)
    return SDValue();
 
  // Collect unique operands.
  auto UniqueOps = SmallSet<SDValue, 16u>();
  for (SDValue Op : BVOp->ops()) {
    if (isIntOrFPConstant(Op) || Op.isUndef())
      return SDValue();
    UniqueOps.insert(Op);
  }
 
  // Candidate BUILD_VECTOR must have 2 unique operands.
  if (UniqueOps.size() != 2u)
    return SDValue();
 
  SDValue Op0 = BVOp->getOperand(0u);
  UniqueOps.erase(Op0);
  SDValue Op1 = *UniqueOps.begin();
 
  if (Subtarget.hasAVX2() || isShuffleFoldableLoad(Op0) ||
      isShuffleFoldableLoad(Op1)) {
    // Create shuffle mask.
    auto const NumElems = VT.getVectorNumElements();
    SmallVector<int, 16u> Mask(NumElems);
    for (auto I = 0u; I < NumElems; ++I) {
      SDValue Op = BVOp->getOperand(I);
      Mask[I] = Op == Op0 ? I : I + NumElems;
    }
    // Create shuffle of splats.
    SDValue NewOp0 = DAG.getSplatBuildVector(VT, DL, Op0);
    SDValue NewOp1 = DAG.getSplatBuildVector(VT, DL, Op1);
    return DAG.getVectorShuffle(VT, DL, NewOp0, NewOp1, Mask);
  }
 
  return SDValue();
}
 
/// Widen a BUILD_VECTOR if the scalar operands are freely mergeable.
static SDValue widenBuildVector(BuildVectorSDNode *BVOp, SDLoc const &DL,
                                X86Subtarget const &Subtarget,
                                SelectionDAG &DAG) {
  using namespace SDPatternMatch;
  MVT VT = BVOp->getSimpleValueType(0);
  MVT SVT = VT.getScalarType();
  unsigned NumElts = VT.getVectorNumElements();
  unsigned EltBits = SVT.getSizeInBits();
 
  if (SVT != MVT::i8 && SVT != MVT::i16 && SVT != MVT::i32)
    return SDValue();
 
  unsigned WideBits = 2 * EltBits;
  MVT WideSVT = MVT::getIntegerVT(WideBits);
  MVT WideVT = MVT::getVectorVT(WideSVT, NumElts / 2);
  if (!DAG.getTargetLoweringInfo().isTypeLegal(WideSVT))
    return SDValue();
 
  SmallVector<SDValue, 8> WideOps;
  for (unsigned I = 0; I != NumElts; I += 2) {
    SDValue Op0 = BVOp->getOperand(I + 0);
    SDValue Op1 = BVOp->getOperand(I + 1);
 
    if (Op0.isUndef() && Op1.isUndef()) {
      WideOps.push_back(DAG.getUNDEF(WideSVT));
      continue;
    }
 
    // TODO: Constant repacking?
 
    // Merge scalars that have been split from the same source.
    SDValue X, Y;
    if (sd_match(Op0, m_Trunc(m_Value(X))) &&
        sd_match(Op1, m_Trunc(m_Srl(m_Value(Y), m_SpecificInt(EltBits)))) &&
        peekThroughTruncates(X) == peekThroughTruncates(Y) &&
        X.getValueType().bitsGE(WideSVT)) {
      if (X.getValueType().bitsGT(WideSVT))
        X = DAG.getNode(ISD::TRUNCATE, DL, WideSVT, X);
      WideOps.push_back(X);
      continue;
    }
 
    return SDValue();
  }
 
  assert(WideOps.size() == (NumElts / 2) && "Failed to widen build vector");
  return DAG.getBitcast(VT, DAG.getBuildVector(WideVT, DL, WideOps));
}
 
/// Create a vector constant without a load. SSE/AVX provide the bare minimum
/// functionality to do this, so it's all zeros, all ones, or some derivation
/// that is cheap to calculate.
static SDValue materializeVectorConstant(SDValue Op, const SDLoc &DL,
                                         SelectionDAG &DAG,
                                         const X86Subtarget &Subtarget) {
  MVT VT = Op.getSimpleValueType();
 
  // Vectors containing all zeros can be matched by pxor and xorps.
  if (ISD::isBuildVectorAllZeros(Op.getNode()))
    return Op;
 
  // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
  // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
  // vpcmpeqd on 256-bit vectors.
  if (Subtarget.hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
    if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
      return Op;
 
    return getOnesVector(VT, DAG, DL);
  }
 
  return SDValue();
}
 
/// Look for opportunities to create a VPERMV/VPERMILPV/PSHUFB variable permute
/// from a vector of source values and a vector of extraction indices.
/// The vectors might be manipulated to match the type of the permute op.
static SDValue createVariablePermute(MVT VT, SDValue SrcVec, SDValue IndicesVec,
                                     const SDLoc &DL, SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
  MVT ShuffleVT = VT;
  EVT IndicesVT = EVT(VT).changeVectorElementTypeToInteger();
  unsigned NumElts = VT.getVectorNumElements();
  unsigned SizeInBits = VT.getSizeInBits();
 
  // Adjust IndicesVec to match VT size.
  assert(IndicesVec.getValueType().getVectorNumElements() >= NumElts &&
         "Illegal variable permute mask size");
  if (IndicesVec.getValueType().getVectorNumElements() > NumElts) {
    // Narrow/widen the indices vector to the correct size.
    if (IndicesVec.getValueSizeInBits() > SizeInBits)
      IndicesVec = extractSubVector(IndicesVec, 0, DAG, SDLoc(IndicesVec),
                                    NumElts * VT.getScalarSizeInBits());
    else if (IndicesVec.getValueSizeInBits() < SizeInBits)
      IndicesVec = widenSubVector(IndicesVec, false, Subtarget, DAG,
                                  SDLoc(IndicesVec), SizeInBits);
    // Zero-extend the index elements within the vector.
    if (IndicesVec.getValueType().getVectorNumElements() > NumElts)
      IndicesVec = DAG.getNode(ISD::ZERO_EXTEND_VECTOR_INREG, SDLoc(IndicesVec),
                               IndicesVT, IndicesVec);
  }
  IndicesVec = DAG.getZExtOrTrunc(IndicesVec, SDLoc(IndicesVec), IndicesVT);
 
  // Handle SrcVec that don't match VT type.
  if (SrcVec.getValueSizeInBits() != SizeInBits) {
    if ((SrcVec.getValueSizeInBits() % SizeInBits) == 0) {
      // Handle larger SrcVec by treating it as a larger permute.
      unsigned Scale = SrcVec.getValueSizeInBits() / SizeInBits;
      VT = MVT::getVectorVT(VT.getScalarType(), Scale * NumElts);
      IndicesVT = EVT(VT).changeVectorElementTypeToInteger();
      IndicesVec = widenSubVector(IndicesVT.getSimpleVT(), IndicesVec, false,
                                  Subtarget, DAG, SDLoc(IndicesVec));
      SDValue NewSrcVec =
          createVariablePermute(VT, SrcVec, IndicesVec, DL, DAG, Subtarget);
      if (NewSrcVec)
        return extractSubVector(NewSrcVec, 0, DAG, DL, SizeInBits);
      return SDValue();
    } else if (SrcVec.getValueSizeInBits() < SizeInBits) {
      // Widen smaller SrcVec to match VT.
      SrcVec = widenSubVector(VT, SrcVec, false, Subtarget, DAG, SDLoc(SrcVec));
    } else
      return SDValue();
  }
 
  auto ScaleIndices = [&DAG](SDValue Idx, uint64_t Scale) {
    assert(isPowerOf2_64(Scale) && "Illegal variable permute shuffle scale");
    EVT SrcVT = Idx.getValueType();
    unsigned NumDstBits = SrcVT.getScalarSizeInBits() / Scale;
    uint64_t IndexScale = 0;
    uint64_t IndexOffset = 0;
 
    // If we're scaling a smaller permute op, then we need to repeat the
    // indices, scaling and offsetting them as well.
    // e.g. v4i32 -> v16i8 (Scale = 4)
    // IndexScale = v4i32 Splat(4 << 24 | 4 << 16 | 4 << 8 | 4)
    // IndexOffset = v4i32 Splat(3 << 24 | 2 << 16 | 1 << 8 | 0)
    for (uint64_t i = 0; i != Scale; ++i) {
      IndexScale |= Scale << (i * NumDstBits);
      IndexOffset |= i << (i * NumDstBits);
    }
 
    Idx = DAG.getNode(ISD::MUL, SDLoc(Idx), SrcVT, Idx,
                      DAG.getConstant(IndexScale, SDLoc(Idx), SrcVT));
    Idx = DAG.getNode(ISD::ADD, SDLoc(Idx), SrcVT, Idx,
                      DAG.getConstant(IndexOffset, SDLoc(Idx), SrcVT));
    return Idx;
  };
 
  unsigned Opcode = 0;
  switch (VT.SimpleTy) {
  default:
    break;
  case MVT::v16i8:
    if (Subtarget.hasSSSE3())
      Opcode = X86ISD::PSHUFB;
    break;
  case MVT::v8i16:
    if (Subtarget.hasVLX() && Subtarget.hasBWI())
      Opcode = X86ISD::VPERMV;
    else if (Subtarget.hasSSSE3()) {
      Opcode = X86ISD::PSHUFB;
      ShuffleVT = MVT::v16i8;
    }
    break;
  case MVT::v4f32:
  case MVT::v4i32:
    if (Subtarget.hasAVX()) {
      Opcode = X86ISD::VPERMILPV;
      ShuffleVT = MVT::v4f32;
    } else if (Subtarget.hasSSSE3()) {
      Opcode = X86ISD::PSHUFB;
      ShuffleVT = MVT::v16i8;
    }
    break;
  case MVT::v2f64:
  case MVT::v2i64:
    if (Subtarget.hasAVX()) {
      // VPERMILPD selects using bit#1 of the index vector, so scale IndicesVec.
      IndicesVec = DAG.getNode(ISD::ADD, DL, IndicesVT, IndicesVec, IndicesVec);
      Opcode = X86ISD::VPERMILPV;
      ShuffleVT = MVT::v2f64;
    } else if (Subtarget.hasSSE41()) {
      // SSE41 can compare v2i64 - select between indices 0 and 1.
      return DAG.getSelectCC(
          DL, IndicesVec,
          getZeroVector(IndicesVT.getSimpleVT(), Subtarget, DAG, DL),
          DAG.getVectorShuffle(VT, DL, SrcVec, SrcVec, {0, 0}),
          DAG.getVectorShuffle(VT, DL, SrcVec, SrcVec, {1, 1}),
          ISD::CondCode::SETEQ);
    }
    break;
  case MVT::v32i8:
    if (Subtarget.hasVLX() && Subtarget.hasVBMI())
      Opcode = X86ISD::VPERMV;
    else if (Subtarget.hasXOP()) {
      SDValue LoSrc = extract128BitVector(SrcVec, 0, DAG, DL);
      SDValue HiSrc = extract128BitVector(SrcVec, 16, DAG, DL);
      SDValue LoIdx = extract128BitVector(IndicesVec, 0, DAG, DL);
      SDValue HiIdx = extract128BitVector(IndicesVec, 16, DAG, DL);
      return DAG.getNode(
          ISD::CONCAT_VECTORS, DL, VT,
          DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, LoSrc, HiSrc, LoIdx),
          DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, LoSrc, HiSrc, HiIdx));
    } else if (Subtarget.hasAVX()) {
      SDValue Lo = extract128BitVector(SrcVec, 0, DAG, DL);
      SDValue Hi = extract128BitVector(SrcVec, 16, DAG, DL);
      SDValue LoLo = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Lo);
      SDValue HiHi = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Hi, Hi);
      auto PSHUFBBuilder = [](SelectionDAG &DAG, const SDLoc &DL,
                              ArrayRef<SDValue> Ops) {
        // Permute Lo and Hi and then select based on index range.
        // This works as SHUFB uses bits[3:0] to permute elements and we don't
        // care about the bit[7] as its just an index vector.
        SDValue Idx = Ops[2];
        EVT VT = Idx.getValueType();
        return DAG.getSelectCC(DL, Idx, DAG.getConstant(15, DL, VT),
                               DAG.getNode(X86ISD::PSHUFB, DL, VT, Ops[1], Idx),
                               DAG.getNode(X86ISD::PSHUFB, DL, VT, Ops[0], Idx),
                               ISD::CondCode::SETGT);
      };
      SDValue Ops[] = {LoLo, HiHi, IndicesVec};
      return SplitOpsAndApply(DAG, Subtarget, DL, MVT::v32i8, Ops,
                              PSHUFBBuilder);
    }
    break;
  case MVT::v16i16:
    if (Subtarget.hasVLX() && Subtarget.hasBWI())
      Opcode = X86ISD::VPERMV;
    else if (Subtarget.hasAVX()) {
      // Scale to v32i8 and perform as v32i8.
      IndicesVec = ScaleIndices(IndicesVec, 2);
      return DAG.getBitcast(
          VT, createVariablePermute(
                  MVT::v32i8, DAG.getBitcast(MVT::v32i8, SrcVec),
                  DAG.getBitcast(MVT::v32i8, IndicesVec), DL, DAG, Subtarget));
    }
    break;
  case MVT::v8f32:
  case MVT::v8i32:
    if (Subtarget.hasAVX2())
      Opcode = X86ISD::VPERMV;
    else if (Subtarget.hasAVX()) {
      SrcVec = DAG.getBitcast(MVT::v8f32, SrcVec);
      SDValue LoLo = DAG.getVectorShuffle(MVT::v8f32, DL, SrcVec, SrcVec,
                                          {0, 1, 2, 3, 0, 1, 2, 3});
      SDValue HiHi = DAG.getVectorShuffle(MVT::v8f32, DL, SrcVec, SrcVec,
                                          {4, 5, 6, 7, 4, 5, 6, 7});
      if (Subtarget.hasXOP())
        return DAG.getBitcast(
            VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v8f32, LoLo, HiHi,
                            IndicesVec, DAG.getTargetConstant(0, DL, MVT::i8)));
      // Permute Lo and Hi and then select based on index range.
      // This works as VPERMILPS only uses index bits[0:1] to permute elements.
      SDValue Res = DAG.getSelectCC(
          DL, IndicesVec, DAG.getConstant(3, DL, MVT::v8i32),
          DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, HiHi, IndicesVec),
          DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, LoLo, IndicesVec),
          ISD::CondCode::SETGT);
      return DAG.getBitcast(VT, Res);
    }
    break;
  case MVT::v4i64:
  case MVT::v4f64:
    if (Subtarget.hasAVX512()) {
      if (!Subtarget.hasVLX()) {
        MVT WidenSrcVT = MVT::getVectorVT(VT.getScalarType(), 8);
        SrcVec = widenSubVector(WidenSrcVT, SrcVec, false, Subtarget, DAG,
                                SDLoc(SrcVec));
        IndicesVec = widenSubVector(MVT::v8i64, IndicesVec, false, Subtarget,
                                    DAG, SDLoc(IndicesVec));
        SDValue Res = createVariablePermute(WidenSrcVT, SrcVec, IndicesVec, DL,
                                            DAG, Subtarget);
        return extract256BitVector(Res, 0, DAG, DL);
      }
      Opcode = X86ISD::VPERMV;
    } else if (Subtarget.hasAVX()) {
      SrcVec = DAG.getBitcast(MVT::v4f64, SrcVec);
      SDValue LoLo =
          DAG.getVectorShuffle(MVT::v4f64, DL, SrcVec, SrcVec, {0, 1, 0, 1});
      SDValue HiHi =
          DAG.getVectorShuffle(MVT::v4f64, DL, SrcVec, SrcVec, {2, 3, 2, 3});
      // VPERMIL2PD selects with bit#1 of the index vector, so scale IndicesVec.
      IndicesVec = DAG.getNode(ISD::ADD, DL, IndicesVT, IndicesVec, IndicesVec);
      if (Subtarget.hasXOP())
        return DAG.getBitcast(
            VT, DAG.getNode(X86ISD::VPERMIL2, DL, MVT::v4f64, LoLo, HiHi,
                            IndicesVec, DAG.getTargetConstant(0, DL, MVT::i8)));
      // Permute Lo and Hi and then select based on index range.
      // This works as VPERMILPD only uses index bit[1] to permute elements.
      SDValue Res = DAG.getSelectCC(
          DL, IndicesVec, DAG.getConstant(2, DL, MVT::v4i64),
          DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v4f64, HiHi, IndicesVec),
          DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v4f64, LoLo, IndicesVec),
          ISD::CondCode::SETGT);
      return DAG.getBitcast(VT, Res);
    }
    break;
  case MVT::v64i8:
    if (Subtarget.hasVBMI())
      Opcode = X86ISD::VPERMV;
    break;
  case MVT::v32i16:
    if (Subtarget.hasBWI())
      Opcode = X86ISD::VPERMV;
    break;
  case MVT::v16f32:
  case MVT::v16i32:
  case MVT::v8f64:
  case MVT::v8i64:
    if (Subtarget.hasAVX512())
      Opcode = X86ISD::VPERMV;
    break;
  }
  if (!Opcode)
    return SDValue();
 
  assert((VT.getSizeInBits() == ShuffleVT.getSizeInBits()) &&
         (VT.getScalarSizeInBits() % ShuffleVT.getScalarSizeInBits()) == 0 &&
         "Illegal variable permute shuffle type");
 
  uint64_t Scale = VT.getScalarSizeInBits() / ShuffleVT.getScalarSizeInBits();
  if (Scale > 1)
    IndicesVec = ScaleIndices(IndicesVec, Scale);
 
  EVT ShuffleIdxVT = EVT(ShuffleVT).changeVectorElementTypeToInteger();
  IndicesVec = DAG.getBitcast(ShuffleIdxVT, IndicesVec);
 
  SrcVec = DAG.getBitcast(ShuffleVT, SrcVec);
  SDValue Res = Opcode == X86ISD::VPERMV
                    ? DAG.getNode(Opcode, DL, ShuffleVT, IndicesVec, SrcVec)
                    : DAG.getNode(Opcode, DL, ShuffleVT, SrcVec, IndicesVec);
  return DAG.getBitcast(VT, Res);
}
 
// Tries to lower a BUILD_VECTOR composed of extract-extract chains that can be
// reasoned to be a permutation of a vector by indices in a non-constant vector.
// (build_vector (extract_elt V, (extract_elt I, 0)),
//               (extract_elt V, (extract_elt I, 1)),
//                    ...
// ->
// (vpermv I, V)
//
// TODO: Handle undefs
// TODO: Utilize pshufb and zero mask blending to support more efficient
// construction of vectors with constant-0 elements.
static SDValue
LowerBUILD_VECTORAsVariablePermute(SDValue V, const SDLoc &DL,
                                   SelectionDAG &DAG,
                                   const X86Subtarget &Subtarget) {
  SDValue SrcVec, IndicesVec;
 
  // Check for a match of the permute source vector and permute index elements.
  // This is done by checking that the i-th build_vector operand is of the form:
  // (extract_elt SrcVec, (extract_elt IndicesVec, i)).
  for (unsigned Idx = 0, E = V.getNumOperands(); Idx != E; ++Idx) {
    SDValue Op = peekThroughOneUseFreeze(V.getOperand(Idx));
    if (Op.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      return SDValue();
 
    // If this is the first extract encountered in V, set the source vector,
    // otherwise verify the extract is from the previously defined source
    // vector.
    if (!SrcVec)
      SrcVec = Op.getOperand(0);
    else if (SrcVec != Op.getOperand(0))
      return SDValue();
    SDValue ExtractedIndex = Op->getOperand(1);
    // Peek through extends.
    if (ExtractedIndex.getOpcode() == ISD::ZERO_EXTEND ||
        ExtractedIndex.getOpcode() == ISD::SIGN_EXTEND)
      ExtractedIndex = ExtractedIndex.getOperand(0);
    if (ExtractedIndex.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
      return SDValue();
 
    // If this is the first extract from the index vector candidate, set the
    // indices vector, otherwise verify the extract is from the previously
    // defined indices vector.
    if (!IndicesVec)
      IndicesVec = ExtractedIndex.getOperand(0);
    else if (IndicesVec != ExtractedIndex.getOperand(0))
      return SDValue();
 
    auto *PermIdx = dyn_cast<ConstantSDNode>(ExtractedIndex.getOperand(1));
    if (!PermIdx || PermIdx->getAPIntValue() != Idx)
      return SDValue();
  }
 
  MVT VT = V.getSimpleValueType();
  return createVariablePermute(VT, SrcVec, IndicesVec, DL, DAG, Subtarget);
}
 
SDValue
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
  SDLoc dl(Op);
 
  MVT VT = Op.getSimpleValueType();
  MVT EltVT = VT.getVectorElementType();
  MVT OpEltVT = Op.getOperand(0).getSimpleValueType();
  unsigned NumElems = Op.getNumOperands();
 
  // Generate vectors for predicate vectors.
  if (VT.getVectorElementType() == MVT::i1 && Subtarget.hasAVX512())
    return LowerBUILD_VECTORvXi1(Op, dl, DAG, Subtarget);
 
  if (VT.getVectorElementType() == MVT::bf16 &&
      (Subtarget.hasAVXNECONVERT() || Subtarget.hasBF16()))
    return LowerBUILD_VECTORvXbf16(Op, DAG, Subtarget);
 
  if (SDValue VectorCst = materializeVectorConstant(Op, dl, DAG, Subtarget))
    return VectorCst;
 
  unsigned EVTBits = EltVT.getSizeInBits();
  APInt UndefMask = APInt::getZero(NumElems);
  APInt FrozenUndefMask = APInt::getZero(NumElems);
  APInt ZeroMask = APInt::getZero(NumElems);
  APInt NonZeroMask = APInt::getZero(NumElems);
  bool IsAllConstants = true;
  bool OneUseFrozenUndefs = true;
  SmallSet<SDValue, 8> Values;
  unsigned NumConstants = NumElems;
  for (unsigned i = 0; i < NumElems; ++i) {
    SDValue Elt = Op.getOperand(i);
    if (Elt.isUndef()) {
      UndefMask.setBit(i);
      continue;
    }
    if (ISD::isFreezeUndef(Elt.getNode())) {
      OneUseFrozenUndefs = OneUseFrozenUndefs && Elt->hasOneUse();
      FrozenUndefMask.setBit(i);
      continue;
    }
    Values.insert(Elt);
    if (!isIntOrFPConstant(Elt)) {
      IsAllConstants = false;
      NumConstants--;
    }
    if (X86::isZeroNode(Elt)) {
      ZeroMask.setBit(i);
    } else {
      NonZeroMask.setBit(i);
    }
  }
 
  // All undef vector. Return an UNDEF.
  if (UndefMask.isAllOnes())
    return DAG.getUNDEF(VT);
 
  // All undef/freeze(undef) vector. Return a FREEZE UNDEF.
  if (OneUseFrozenUndefs && (UndefMask | FrozenUndefMask).isAllOnes())
    return DAG.getFreeze(DAG.getUNDEF(VT));
 
  // All undef/freeze(undef)/zero vector. Return a zero vector.
  if ((UndefMask | FrozenUndefMask | ZeroMask).isAllOnes())
    return getZeroVector(VT, Subtarget, DAG, dl);
 
  // If we have multiple FREEZE-UNDEF operands, we are likely going to end up
  // lowering into a suboptimal insertion sequence. Instead, thaw the UNDEF in
  // our source BUILD_VECTOR, create another FREEZE-UNDEF splat BUILD_VECTOR,
  // and blend the FREEZE-UNDEF operands back in.
  // FIXME: is this worthwhile even for a single FREEZE-UNDEF operand?
  if (unsigned NumFrozenUndefElts = FrozenUndefMask.popcount();
      NumFrozenUndefElts >= 2 && NumFrozenUndefElts < NumElems) {
    SmallVector<int, 16> BlendMask(NumElems, -1);
    SmallVector<SDValue, 16> Elts(NumElems, DAG.getUNDEF(OpEltVT));
    for (unsigned i = 0; i < NumElems; ++i) {
      if (UndefMask[i]) {
        BlendMask[i] = -1;
        continue;
      }
      BlendMask[i] = i;
      if (!FrozenUndefMask[i])
        Elts[i] = Op.getOperand(i);
      else
        BlendMask[i] += NumElems;
    }
    SDValue EltsBV = DAG.getBuildVector(VT, dl, Elts);
    SDValue FrozenUndefElt = DAG.getFreeze(DAG.getUNDEF(OpEltVT));
    SDValue FrozenUndefBV = DAG.getSplatBuildVector(VT, dl, FrozenUndefElt);
    return DAG.getVectorShuffle(VT, dl, EltsBV, FrozenUndefBV, BlendMask);
  }
 
  BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
 
  // If the upper elts of a ymm/zmm are undef/freeze(undef)/zero then we might
  // be better off lowering to a smaller build vector and padding with
  // undef/zero.
  if ((VT.is256BitVector() || VT.is512BitVector()) &&
      !isFoldableUseOfShuffle(BV)) {
    unsigned UpperElems = NumElems / 2;
    APInt UndefOrZeroMask = FrozenUndefMask | UndefMask | ZeroMask;
    unsigned NumUpperUndefsOrZeros = UndefOrZeroMask.countl_one();
    if (NumUpperUndefsOrZeros >= UpperElems) {
      if (VT.is512BitVector() &&
          NumUpperUndefsOrZeros >= (NumElems - (NumElems / 4)))
        UpperElems = NumElems - (NumElems / 4);
      // If freeze(undef) is in any upper elements, force to zero.
      bool UndefUpper = UndefMask.countl_one() >= UpperElems;
      MVT LowerVT = MVT::getVectorVT(EltVT, NumElems - UpperElems);
      SDValue NewBV =
          DAG.getBuildVector(LowerVT, dl, Op->ops().drop_back(UpperElems));
      return widenSubVector(VT, NewBV, !UndefUpper, Subtarget, DAG, dl);
    }
  }
 
  if (SDValue AddSub = lowerToAddSubOrFMAddSub(BV, dl, Subtarget, DAG))
    return AddSub;
  if (SDValue HorizontalOp = LowerToHorizontalOp(BV, dl, Subtarget, DAG))
    return HorizontalOp;
  if (SDValue Broadcast = lowerBuildVectorAsBroadcast(BV, dl, Subtarget, DAG))
    return Broadcast;
  if (SDValue BitOp = lowerBuildVectorToBitOp(BV, dl, Subtarget, DAG))
    return BitOp;
  if (SDValue Blend = lowerBuildVectorAsBlend(BV, dl, Subtarget, DAG))
    return Blend;
  if (SDValue WideBV = widenBuildVector(BV, dl, Subtarget, DAG))
    return WideBV;
 
  unsigned NumZero = ZeroMask.popcount();
  unsigned NumNonZero = NonZeroMask.popcount();
 
  // If we are inserting one variable into a vector of non-zero constants, try
  // to avoid loading each constant element as a scalar. Load the constants as a
  // vector and then insert the variable scalar element. If insertion is not
  // supported, fall back to a shuffle to get the scalar blended with the
  // constants. Insertion into a zero vector is handled as a special-case
  // somewhere below here.
  if (NumConstants == NumElems - 1 && NumNonZero != 1 &&
      FrozenUndefMask.isZero() &&
      (isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT) ||
       isOperationLegalOrCustom(ISD::VECTOR_SHUFFLE, VT))) {
    // Create an all-constant vector. The variable element in the old
    // build vector is replaced by undef in the constant vector. Save the
    // variable scalar element and its index for use in the insertelement.
    LLVMContext &Context = *DAG.getContext();
    Type *EltType = Op.getValueType().getScalarType().getTypeForEVT(Context);
    SmallVector<Constant *, 16> ConstVecOps(NumElems, UndefValue::get(EltType));
    SDValue VarElt;
    SDValue InsIndex;
    for (unsigned i = 0; i != NumElems; ++i) {
      SDValue Elt = Op.getOperand(i);
      if (auto *C = dyn_cast<ConstantSDNode>(Elt))
        ConstVecOps[i] = ConstantInt::get(Context, C->getAPIntValue());
      else if (auto *C = dyn_cast<ConstantFPSDNode>(Elt))
        ConstVecOps[i] = ConstantFP::get(Context, C->getValueAPF());
      else if (!Elt.isUndef()) {
        assert(!VarElt.getNode() && !InsIndex.getNode() &&
               "Expected one variable element in this vector");
        VarElt = Elt;
        InsIndex = DAG.getVectorIdxConstant(i, dl);
      }
    }
    Constant *CV = ConstantVector::get(ConstVecOps);
    SDValue DAGConstVec = DAG.getConstantPool(CV, VT);
 
    // The constants we just created may not be legal (eg, floating point). We
    // must lower the vector right here because we can not guarantee that we'll
    // legalize it before loading it. This is also why we could not just create
    // a new build vector here. If the build vector contains illegal constants,
    // it could get split back up into a series of insert elements.
    // TODO: Improve this by using shorter loads with broadcast/VZEXT_LOAD.
    SDValue LegalDAGConstVec = LowerConstantPool(DAGConstVec, DAG);
    MachineFunction &MF = DAG.getMachineFunction();
    MachinePointerInfo MPI = MachinePointerInfo::getConstantPool(MF);
    SDValue Ld = DAG.getLoad(VT, dl, DAG.getEntryNode(), LegalDAGConstVec, MPI);
    unsigned InsertC = InsIndex->getAsZExtVal();
    unsigned NumEltsInLow128Bits = 128 / VT.getScalarSizeInBits();
    if (InsertC < NumEltsInLow128Bits)
      return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Ld, VarElt, InsIndex);
 
    // There's no good way to insert into the high elements of a >128-bit
    // vector, so use shuffles to avoid an extract/insert sequence.
    assert(VT.getSizeInBits() > 128 && "Invalid insertion index?");
    assert(Subtarget.hasAVX() && "Must have AVX with >16-byte vector");
    SmallVector<int, 8> ShuffleMask;
    unsigned NumElts = VT.getVectorNumElements();
    for (unsigned i = 0; i != NumElts; ++i)
      ShuffleMask.push_back(i == InsertC ? NumElts : i);
    SDValue S2V = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, VarElt);
    return DAG.getVectorShuffle(VT, dl, Ld, S2V, ShuffleMask);
  }
 
  // Special case for single non-zero, non-undef, element.
  if (NumNonZero == 1) {
    unsigned Idx = NonZeroMask.countr_zero();
    SDValue Item = Op.getOperand(Idx);
 
    // If we have a constant or non-constant insertion into the low element of
    // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
    // the rest of the elements.  This will be matched as movd/movq/movss/movsd
    // depending on what the source datatype is.
    if (Idx == 0) {
      if (NumZero == 0)
        return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
 
      if (EltVT == MVT::i32 || EltVT == MVT::f16 || EltVT == MVT::f32 ||
          EltVT == MVT::f64 || (EltVT == MVT::i64 && Subtarget.is64Bit()) ||
          (EltVT == MVT::i16 && Subtarget.hasFP16())) {
        assert((VT.is128BitVector() || VT.is256BitVector() ||
                VT.is512BitVector()) &&
               "Expected an SSE value type!");
        Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
        // Turn it into a MOVL (i.e. movsh, movss, movsd, movw or movd) to a
        // zero vector.
        return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
      }
 
      // We can't directly insert an i8 or i16 into a vector, so zero extend
      // it to i32 first.
      if (EltVT == MVT::i16 || EltVT == MVT::i8) {
        Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
        MVT ShufVT = MVT::getVectorVT(MVT::i32, VT.getSizeInBits() / 32);
        Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShufVT, Item);
        Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
        return DAG.getBitcast(VT, Item);
      }
    }
 
    // Is it a vector logical left shift?
    if (NumElems == 2 && Idx == 1 &&
        X86::isZeroNode(Op.getOperand(0)) &&
        !X86::isZeroNode(Op.getOperand(1))) {
      unsigned NumBits = VT.getSizeInBits();
      return getVShift(true, VT,
                       DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
                                   VT, Op.getOperand(1)),
                       NumBits/2, DAG, *this, dl);
    }
 
    if (IsAllConstants) // Otherwise, it's better to do a constpool load.
      return SDValue();
 
    // Otherwise, if this is a vector with i32 or f32 elements, and the element
    // is a non-constant being inserted into an element other than the low one,
    // we can't use a constant pool load.  Instead, use SCALAR_TO_VECTOR (aka
    // movd/movss) to move this into the low element, then shuffle it into
    // place.
    if (EVTBits == 32) {
      Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
      return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
    }
  }
 
  // Splat is obviously ok. Let legalizer expand it to a shuffle.
  if (Values.size() == 1) {
    if (EVTBits == 32) {
      // Instead of a shuffle like this:
      // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
      // Check if it's possible to issue this instead.
      // shuffle (vload ptr)), undef, <1, 1, 1, 1>
      unsigned Idx = NonZeroMask.countr_zero();
      SDValue Item = Op.getOperand(Idx);
      if (Op.getNode()->isOnlyUserOf(Item.getNode()))
        return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
    }
    return SDValue();
  }
 
  // A vector full of immediates; various special cases are already
  // handled, so this is best done with a single constant-pool load.
  if (IsAllConstants)
    return SDValue();
 
  if (SDValue V = LowerBUILD_VECTORAsVariablePermute(Op, dl, DAG, Subtarget))
    return V;
 
  // See if we can use a vector load to get all of the elements.
  {
    SmallVector<SDValue, 64> Ops(Op->ops().take_front(NumElems));
    if (SDValue LD =
            EltsFromConsecutiveLoads(VT, Ops, dl, DAG, Subtarget, false))
      return LD;
  }
 
  // If this is a splat of pairs of 32-bit elements, we can use a narrower
  // build_vector and broadcast it.
  // TODO: We could probably generalize this more.
  if (Subtarget.hasAVX2() && EVTBits == 32 && Values.size() == 2) {
    SDValue Ops[4] = { Op.getOperand(0), Op.getOperand(1),
                       DAG.getUNDEF(EltVT), DAG.getUNDEF(EltVT) };
    auto CanSplat = [](SDValue Op, unsigned NumElems, ArrayRef<SDValue> Ops) {
      // Make sure all the even/odd operands match.
      for (unsigned i = 2; i != NumElems; ++i)
        if (Ops[i % 2] != Op.getOperand(i))
          return false;
      return true;
    };
    if (CanSplat(Op, NumElems, Ops)) {
      MVT WideEltVT = VT.isFloatingPoint() ? MVT::f64 : MVT::i64;
      MVT NarrowVT = MVT::getVectorVT(EltVT, 4);
      // Create a new build vector and cast to v2i64/v2f64.
      SDValue NewBV = DAG.getBitcast(MVT::getVectorVT(WideEltVT, 2),
                                     DAG.getBuildVector(NarrowVT, dl, Ops));
      // Broadcast from v2i64/v2f64 and cast to final VT.
      MVT BcastVT = MVT::getVectorVT(WideEltVT, NumElems / 2);
      return DAG.getBitcast(VT, DAG.getNode(X86ISD::VBROADCAST, dl, BcastVT,
                                            NewBV));
    }
  }
 
  // For AVX-length vectors, build the individual 128-bit pieces and use
  // shuffles to put them in place.
  if (VT.getSizeInBits() > 128) {
    MVT HVT = MVT::getVectorVT(EltVT, NumElems / 2);
 
    // Build both the lower and upper subvector.
    SDValue Lower =
        DAG.getBuildVector(HVT, dl, Op->ops().slice(0, NumElems / 2));
    SDValue Upper = DAG.getBuildVector(
        HVT, dl, Op->ops().slice(NumElems / 2, NumElems /2));
 
    // Recreate the wider vector with the lower and upper part.
    return concatSubVectors(Lower, Upper, DAG, dl);
  }
 
  // Let legalizer expand 2-wide build_vectors.
  if (EVTBits == 64) {
    if (NumNonZero == 1) {
      // One half is zero or undef.
      unsigned Idx = NonZeroMask.countr_zero();
      SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
                               Op.getOperand(Idx));
      return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
    }
    return SDValue();
  }
 
  // If element VT is < 32 bits, convert it to inserts into a zero vector.
  if (EVTBits == 8 && NumElems == 16)
    if (SDValue V = LowerBuildVectorv16i8(Op, dl, NonZeroMask, NumNonZero,
                                          NumZero, DAG, Subtarget))
      return V;
 
  if (EltVT == MVT::i16 && NumElems == 8)
    if (SDValue V = LowerBuildVectorv8i16(Op, dl, NonZeroMask, NumNonZero,
                                          NumZero, DAG, Subtarget))
      return V;
 
  // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
  if (EVTBits == 32 && NumElems == 4)
    if (SDValue V = LowerBuildVectorv4x32(Op, dl, DAG, Subtarget))
      return V;
 
  // If element VT is == 32 bits, turn it into a number of shuffles.
  if (NumElems == 4 && NumZero > 0) {
    SmallVector<SDValue, 8> Ops(NumElems);
    for (unsigned i = 0; i < 4; ++i) {
      bool isZero = !NonZeroMask[i];
      if (isZero)
        Ops[i] = getZeroVector(VT, Subtarget, DAG, dl);
      else
        Ops[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
    }
 
    for (unsigned i = 0; i < 2; ++i) {
      switch (NonZeroMask.extractBitsAsZExtValue(2, i * 2)) {
        default: llvm_unreachable("Unexpected NonZero count");
        case 0:
          Ops[i] = Ops[i*2];  // Must be a zero vector.
          break;
        case 1:
          Ops[i] = getMOVL(DAG, dl, VT, Ops[i*2+1], Ops[i*2]);
          break;
        case 2:
          Ops[i] = getMOVL(DAG, dl, VT, Ops[i*2], Ops[i*2+1]);
          break;
        case 3:
          Ops[i] = getUnpackl(DAG, dl, VT, Ops[i*2], Ops[i*2+1]);
          break;
      }
    }
 
    bool Reverse1 = NonZeroMask.extractBitsAsZExtValue(2, 0) == 2;
    bool Reverse2 = NonZeroMask.extractBitsAsZExtValue(2, 2) == 2;
    int MaskVec[] = {
      Reverse1 ? 1 : 0,
      Reverse1 ? 0 : 1,
      static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
      static_cast<int>(Reverse2 ? NumElems   : NumElems+1)
    };
    return DAG.getVectorShuffle(VT, dl, Ops[0], Ops[1], MaskVec);
  }
 
  assert(Values.size() > 1 && "Expected non-undef and non-splat vector");
 
  // Check for a build vector from mostly shuffle plus few inserting.
  if (SDValue Sh = buildFromShuffleMostly(Op, dl, DAG))
    return Sh;
 
  // For SSE 4.1, use insertps to put the high elements into the low element.
  if (Subtarget.hasSSE41() && EltVT != MVT::f16) {
    SDValue Result;
    if (!Op.getOperand(0).isUndef())
      Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
    else
      Result = DAG.getUNDEF(VT);
 
    for (unsigned i = 1; i < NumElems; ++i) {
      if (Op.getOperand(i).isUndef()) continue;
      Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
                           Op.getOperand(i), DAG.getVectorIdxConstant(i, dl));
    }
    return Result;
  }
 
  // Otherwise, expand into a number of unpckl*, start by extending each of
  // our (non-undef) elements to the full vector width with the element in the
  // bottom slot of the vector (which generates no code for SSE).
  SmallVector<SDValue, 8> Ops(NumElems);
  for (unsigned i = 0; i < NumElems; ++i) {
    if (!Op.getOperand(i).isUndef())
      Ops[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
    else
      Ops[i] = DAG.getUNDEF(VT);
  }
 
  // Next, we iteratively mix elements, e.g. for v4f32:
  //   Step 1: unpcklps 0, 1 ==> X: <?, ?, 1, 0>
  //         : unpcklps 2, 3 ==> Y: <?, ?, 3, 2>
  //   Step 2: unpcklpd X, Y ==>    <3, 2, 1, 0>
  for (unsigned Scale = 1; Scale < NumElems; Scale *= 2) {
    // Generate scaled UNPCKL shuffle mask.
    SmallVector<int, 16> Mask;
    for(unsigned i = 0; i != Scale; ++i)
      Mask.push_back(i);
    for (unsigned i = 0; i != Scale; ++i)
      Mask.push_back(NumElems+i);
    Mask.append(NumElems - Mask.size(), SM_SentinelUndef);
 
    for (unsigned i = 0, e = NumElems / (2 * Scale); i != e; ++i)
      Ops[i] = DAG.getVectorShuffle(VT, dl, Ops[2*i], Ops[(2*i)+1], Mask);
  }
  return Ops[0];
}
 
// 256-bit AVX can use the vinsertf128 instruction
// to create 256-bit vectors from two other 128-bit ones.
// TODO: Detect subvector broadcast here instead of DAG combine?
static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, const SDLoc &dl,
                                      SelectionDAG &DAG,
                                      const X86Subtarget &Subtarget) {
  MVT ResVT = Op.getSimpleValueType();
  assert((ResVT.is256BitVector() || ResVT.is512BitVector()) &&
         "Value type must be 256-/512-bit wide");
 
  unsigned NumOperands = Op.getNumOperands();
  unsigned NumFreezeUndef = 0;
  unsigned NumZero = 0;
  unsigned NumNonZero = 0;
  unsigned NonZeros = 0;
  SmallSet<SDValue, 4> Undefs;
  for (unsigned i = 0; i != NumOperands; ++i) {
    SDValue SubVec = Op.getOperand(i);
    if (SubVec.isUndef())
      continue;
    if (ISD::isFreezeUndef(SubVec.getNode())) {
        // If the freeze(undef) has multiple uses then we must fold to zero.
        if (SubVec.hasOneUse()) {
          ++NumFreezeUndef;
        } else {
          ++NumZero;
          Undefs.insert(SubVec);
        }
    }
    else if (ISD::isBuildVectorAllZeros(SubVec.getNode()))
      ++NumZero;
    else {
      assert(i < sizeof(NonZeros) * CHAR_BIT); // Ensure the shift is in range.
      NonZeros |= 1 << i;
      ++NumNonZero;
    }
  }
 
  // If we have more than 2 non-zeros, build each half separately.
  if (NumNonZero > 2) {
    MVT HalfVT = ResVT.getHalfNumVectorElementsVT();
    ArrayRef<SDUse> Ops = Op->ops();
    SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT,
                             Ops.slice(0, NumOperands/2));
    SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT,
                             Ops.slice(NumOperands/2));
    return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
  }
 
  // Otherwise, build it up through insert_subvectors.
  SDValue Vec = NumZero ? getZeroVector(ResVT, Subtarget, DAG, dl)
                        : (NumFreezeUndef ? DAG.getFreeze(DAG.getUNDEF(ResVT))
                                          : DAG.getUNDEF(ResVT));
 
  // Replace Undef operands with ZeroVector.
  for (SDValue U : Undefs)
    DAG.ReplaceAllUsesWith(
        U, getZeroVector(U.getSimpleValueType(), Subtarget, DAG, dl));
 
  MVT SubVT = Op.getOperand(0).getSimpleValueType();
  unsigned NumSubElems = SubVT.getVectorNumElements();
  for (unsigned i = 0; i != NumOperands; ++i) {
    if ((NonZeros & (1 << i)) == 0)
      continue;
 
    Vec = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Vec, Op.getOperand(i),
                      DAG.getVectorIdxConstant(i * NumSubElems, dl));
  }
 
  return Vec;
}
 
// Returns true if the given node is a type promotion (by concatenating i1
// zeros) of the result of a node that already zeros all upper bits of
// k-register.
// TODO: Merge this with LowerAVXCONCAT_VECTORS?
static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op, const SDLoc &dl,
                                       const X86Subtarget &Subtarget,
                                       SelectionDAG & DAG) {
  MVT ResVT = Op.getSimpleValueType();
  unsigned NumOperands = Op.getNumOperands();
  assert(NumOperands > 1 && isPowerOf2_32(NumOperands) &&
         "Unexpected number of operands in CONCAT_VECTORS");
 
  uint64_t Zeros = 0;
  uint64_t NonZeros = 0;
  for (unsigned i = 0; i != NumOperands; ++i) {
    SDValue SubVec = Op.getOperand(i);
    if (SubVec.isUndef())
      continue;
    assert(i < sizeof(NonZeros) * CHAR_BIT); // Ensure the shift is in range.
    if (ISD::isBuildVectorAllZeros(SubVec.getNode()))
      Zeros |= (uint64_t)1 << i;
    else
      NonZeros |= (uint64_t)1 << i;
  }
 
  unsigned NumElems = ResVT.getVectorNumElements();
 
  // If we are inserting non-zero vector and there are zeros in LSBs and undef
  // in the MSBs we need to emit a KSHIFTL. The generic lowering to
  // insert_subvector will give us two kshifts.
  if (isPowerOf2_64(NonZeros) && Zeros != 0 && NonZeros > Zeros &&
      Log2_64(NonZeros) != NumOperands - 1) {
    unsigned Idx = Log2_64(NonZeros);
    SDValue SubVec = Op.getOperand(Idx);
    unsigned SubVecNumElts = SubVec.getSimpleValueType().getVectorNumElements();
    MVT ShiftVT = widenMaskVectorType(ResVT, Subtarget);
    Op = widenSubVector(ShiftVT, SubVec, false, Subtarget, DAG, dl);
    Op = DAG.getNode(X86ISD::KSHIFTL, dl, ShiftVT, Op,
                     DAG.getTargetConstant(Idx * SubVecNumElts, dl, MVT::i8));
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResVT, Op,
                       DAG.getVectorIdxConstant(0, dl));
  }
 
  // If there are zero or one non-zeros we can handle this very simply.
  if (NonZeros == 0 || isPowerOf2_64(NonZeros)) {
    SDValue Vec = Zeros ? DAG.getConstant(0, dl, ResVT) : DAG.getUNDEF(ResVT);
    if (!NonZeros)
      return Vec;
    unsigned Idx = Log2_64(NonZeros);
    SDValue SubVec = Op.getOperand(Idx);
    unsigned SubVecNumElts = SubVec.getSimpleValueType().getVectorNumElements();
    return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Vec, SubVec,
                       DAG.getVectorIdxConstant(Idx * SubVecNumElts, dl));
  }
 
  if (NumOperands > 2) {
    MVT HalfVT = ResVT.getHalfNumVectorElementsVT();
    ArrayRef<SDUse> Ops = Op->ops();
    SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT,
                             Ops.slice(0, NumOperands / 2));
    SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT,
                             Ops.slice(NumOperands / 2));
    return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
  }
 
  assert(llvm::popcount(NonZeros) == 2 && "Simple cases not handled?");
 
  if (ResVT.getVectorNumElements() >= 16)
    return Op; // The operation is legal with KUNPCK
 
  SDValue Vec =
      DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, DAG.getUNDEF(ResVT),
                  Op.getOperand(0), DAG.getVectorIdxConstant(0, dl));
  return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Vec, Op.getOperand(1),
                     DAG.getVectorIdxConstant(NumElems / 2, dl));
}
 
static SDValue LowerCONCAT_VECTORS(SDValue Op,
                                   const X86Subtarget &Subtarget,
                                   SelectionDAG &DAG) {
  SDLoc DL(Op);
  MVT VT = Op.getSimpleValueType();
  if (VT.getVectorElementType() == MVT::i1)
    return LowerCONCAT_VECTORSvXi1(Op, DL, Subtarget, DAG);
 
  // AVX can use the vinsertf128 instruction to create 256-bit vectors
  // from two other 128-bit ones.
  // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
  assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
         (VT.is512BitVector() &&
          (Op.getNumOperands() == 2 || Op.getNumOperands() == 4)));
  return LowerAVXCONCAT_VECTORS(Op, DL, DAG, Subtarget);
}
 
//===----------------------------------------------------------------------===//
// Vector shuffle lowering
//
// This is an experimental code path for lowering vector shuffles on x86. It is
// designed to handle arbitrary vector shuffles and blends, gracefully
// degrading performance as necessary. It works hard to recognize idiomatic
// shuffles and lower them to optimal instruction patterns without leaving
// a framework that allows reasonably efficient handling of all vector shuffle
// patterns.
//===----------------------------------------------------------------------===//
 
/// Checks whether the vector elements referenced by two shuffle masks are
/// equivalent.
static bool IsElementEquivalent(int MaskSize, SDValue Op, SDValue ExpectedOp,
                                int Idx, int ExpectedIdx) {
  assert(0 <= Idx && Idx < MaskSize && 0 <= ExpectedIdx &&
         ExpectedIdx < MaskSize && "Out of range element index");
  if (!Op || !ExpectedOp || Op.getOpcode() != ExpectedOp.getOpcode())
    return false;
 
  EVT VT = Op.getValueType();
  EVT ExpectedVT = ExpectedOp.getValueType();
 
  // Sources must be vectors and match the mask's element count.
  if (!VT.isVector() || !ExpectedVT.isVector() ||
      (int)VT.getVectorNumElements() != MaskSize ||
      (int)ExpectedVT.getVectorNumElements() != MaskSize)
    return false;
 
  // Exact match.
  if (Idx == ExpectedIdx && Op == ExpectedOp)
    return true;
 
  switch (Op.getOpcode()) {
  case ISD::BUILD_VECTOR:
    // If the values are build vectors, we can look through them to find
    // equivalent inputs that make the shuffles equivalent.
    return Op.getOperand(Idx) == ExpectedOp.getOperand(ExpectedIdx);
  case ISD::BITCAST: {
    SDValue Src = peekThroughBitcasts(Op);
    EVT SrcVT = Src.getValueType();
    if (Op == ExpectedOp && SrcVT.isVector()) {
      if ((SrcVT.getScalarSizeInBits() % VT.getScalarSizeInBits()) == 0) {
        unsigned Scale = SrcVT.getScalarSizeInBits() / VT.getScalarSizeInBits();
        return (Idx % Scale) == (ExpectedIdx % Scale) &&
               IsElementEquivalent(SrcVT.getVectorNumElements(), Src, Src,
                                   Idx / Scale, ExpectedIdx / Scale);
      }
      if ((VT.getScalarSizeInBits() % SrcVT.getScalarSizeInBits()) == 0) {
        unsigned Scale = VT.getScalarSizeInBits() / SrcVT.getScalarSizeInBits();
        for (unsigned I = 0; I != Scale; ++I)
          if (!IsElementEquivalent(SrcVT.getVectorNumElements(), Src, Src,
                                   (Idx * Scale) + I,
                                   (ExpectedIdx * Scale) + I))
            return false;
        return true;
      }
    }
    break;
  }
  case ISD::VECTOR_SHUFFLE: {
    auto *SVN = cast<ShuffleVectorSDNode>(Op);
    return Op == ExpectedOp &&
           SVN->getMaskElt(Idx) == SVN->getMaskElt(ExpectedIdx);
  }
  case X86ISD::VBROADCAST:
  case X86ISD::VBROADCAST_LOAD:
    return Op == ExpectedOp;
  case X86ISD::SUBV_BROADCAST_LOAD:
    if (Op == ExpectedOp) {
      auto *MemOp = cast<MemSDNode>(Op);
      unsigned NumMemElts = MemOp->getMemoryVT().getVectorNumElements();
      return (Idx % NumMemElts) == (ExpectedIdx % NumMemElts);
    }
    break;
  case X86ISD::VPERMI: {
    if (Op == ExpectedOp) {
      SmallVector<int, 8> Mask;
      DecodeVPERMMask(MaskSize, Op.getConstantOperandVal(1), Mask);
      SDValue Src = Op.getOperand(0);
      return IsElementEquivalent(MaskSize, Src, Src, Mask[Idx],
                                 Mask[ExpectedIdx]);
    }
    break;
  }
  case X86ISD::HADD:
  case X86ISD::HSUB:
  case X86ISD::FHADD:
  case X86ISD::FHSUB:
  case X86ISD::PACKSS:
  case X86ISD::PACKUS:
    // HOP(X,X) can refer to the elt from the lower/upper half of a lane.
    // TODO: Handle HOP(X,Y) vs HOP(Y,X) equivalence cases.
    if (Op == ExpectedOp && Op.getOperand(0) == Op.getOperand(1)) {
      int NumElts = VT.getVectorNumElements();
      int NumLanes = VT.getSizeInBits() / 128;
      int NumEltsPerLane = NumElts / NumLanes;
      int NumHalfEltsPerLane = NumEltsPerLane / 2;
      bool SameLane = (Idx / NumEltsPerLane) == (ExpectedIdx / NumEltsPerLane);
      bool SameElt =
          (Idx % NumHalfEltsPerLane) == (ExpectedIdx % NumHalfEltsPerLane);
      return SameLane && SameElt;
    }
    break;
  }
 
  return false;
}
 
/// Tiny helper function to identify a no-op mask.
///
/// This is a somewhat boring predicate function. It checks whether the mask
/// array input, which is assumed to be a single-input shuffle mask of the kind
/// used by the X86 shuffle instructions (not a fully general
/// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
/// in-place shuffle are 'no-op's.
static bool isNoopShuffleMask(ArrayRef<int> Mask) {
  for (int i = 0, Size = Mask.size(); i < Size; ++i) {
    assert(Mask[i] >= -1 && "Out of bound mask element!");
    if (Mask[i] >= 0 && Mask[i] != i)
      return false;
  }
  return true;
}
 
/// Test whether there are elements crossing LaneSizeInBits lanes in this
/// shuffle mask.
///
/// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
/// and we routinely test for these.
static bool isLaneCrossingShuffleMask(unsigned LaneSizeInBits,
                                      unsigned ScalarSizeInBits,
                                      ArrayRef<int> Mask) {
  assert(LaneSizeInBits && ScalarSizeInBits &&
         (LaneSizeInBits % ScalarSizeInBits) == 0 &&
         "Illegal shuffle lane size");
  int LaneSize = LaneSizeInBits / ScalarSizeInBits;
  int Size = Mask.size();
  for (int i = 0; i < Size; ++i)
    if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
      return true;
  return false;
}
 
/// Test whether there are elements crossing 128-bit lanes in this
/// shuffle mask.
static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
  return isLaneCrossingShuffleMask(128, VT.getScalarSizeInBits(), Mask);
}
 
/// Test whether elements in each LaneSizeInBits lane in this shuffle mask come
/// from multiple lanes - this is different to isLaneCrossingShuffleMask to
/// better support 'repeated mask + lane permute' style shuffles.
static bool isMultiLaneShuffleMask(unsigned LaneSizeInBits,
                                   unsigned ScalarSizeInBits,
                                   ArrayRef<int> Mask) {
  assert(LaneSizeInBits && ScalarSizeInBits &&
         (LaneSizeInBits % ScalarSizeInBits) == 0 &&
         "Illegal shuffle lane size");
  int NumElts = Mask.size();
  int NumEltsPerLane = LaneSizeInBits / ScalarSizeInBits;
  int NumLanes = NumElts / NumEltsPerLane;
  if (NumLanes > 1) {
    for (int i = 0; i != NumLanes; ++i) {
      int SrcLane = -1;
      for (int j = 0; j != NumEltsPerLane; ++j) {
        int M = Mask[(i * NumEltsPerLane) + j];
        if (M < 0)
          continue;
        int Lane = (M % NumElts) / NumEltsPerLane;
        if (SrcLane >= 0 && SrcLane != Lane)
          return true;
        SrcLane = Lane;
      }
    }
  }
  return false;
}
 
/// Test whether a shuffle mask is equivalent within each sub-lane.
///
/// This checks a shuffle mask to see if it is performing the same
/// lane-relative shuffle in each sub-lane. This trivially implies
/// that it is also not lane-crossing. It may however involve a blend from the
/// same lane of a second vector.
///
/// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
/// non-trivial to compute in the face of undef lanes. The representation is
/// suitable for use with existing 128-bit shuffles as entries from the second
/// vector have been remapped to [LaneSize, 2*LaneSize).
static bool isRepeatedShuffleMask(unsigned LaneSizeInBits, MVT VT,
                                  ArrayRef<int> Mask,
                                  SmallVectorImpl<int> &RepeatedMask) {
  auto LaneSize = LaneSizeInBits / VT.getScalarSizeInBits();
  RepeatedMask.assign(LaneSize, -1);
  int Size = Mask.size();
  for (int i = 0; i < Size; ++i) {
    assert(Mask[i] == SM_SentinelUndef || Mask[i] >= 0);
    if (Mask[i] < 0)
      continue;
    if ((Mask[i] % Size) / LaneSize != i / LaneSize)
      // This entry crosses lanes, so there is no way to model this shuffle.
      return false;
 
    // Ok, handle the in-lane shuffles by detecting if and when they repeat.
    // Adjust second vector indices to start at LaneSize instead of Size.
    int LocalM = Mask[i] < Size ? Mask[i] % LaneSize
                                : Mask[i] % LaneSize + LaneSize;
    if (RepeatedMask[i % LaneSize] < 0)
      // This is the first non-undef entry in this slot of a 128-bit lane.
      RepeatedMask[i % LaneSize] = LocalM;
    else if (RepeatedMask[i % LaneSize] != LocalM)
      // Found a mismatch with the repeated mask.
      return false;
  }
  return true;
}
 
/// Test whether a shuffle mask is equivalent within each 128-bit lane.
static bool
is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
                                SmallVectorImpl<int> &RepeatedMask) {
  return isRepeatedShuffleMask(128, VT, Mask, RepeatedMask);
}
 
static bool
is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask) {
  SmallVector<int, 32> RepeatedMask;
  return isRepeatedShuffleMask(128, VT, Mask, RepeatedMask);
}
 
/// Test whether a shuffle mask is equivalent within each 256-bit lane.
static bool
is256BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
                                SmallVectorImpl<int> &RepeatedMask) {
  return isRepeatedShuffleMask(256, VT, Mask, RepeatedMask);
}
 
/// Test whether a target shuffle mask is equivalent within each sub-lane.
/// Unlike isRepeatedShuffleMask we must respect SM_SentinelZero.
static bool isRepeatedTargetShuffleMask(unsigned LaneSizeInBits,
                                        unsigned EltSizeInBits,
                                        ArrayRef<int> Mask,
                                        SmallVectorImpl<int> &RepeatedMask) {
  int LaneSize = LaneSizeInBits / EltSizeInBits;
  RepeatedMask.assign(LaneSize, SM_SentinelUndef);
  int Size = Mask.size();
  for (int i = 0; i < Size; ++i) {
    assert(isUndefOrZero(Mask[i]) || (Mask[i] >= 0));
    if (Mask[i] == SM_SentinelUndef)
      continue;
    if (Mask[i] == SM_SentinelZero) {
      if (!isUndefOrZero(RepeatedMask[i % LaneSize]))
        return false;
      RepeatedMask[i % LaneSize] = SM_SentinelZero;
      continue;
    }
    if ((Mask[i] % Size) / LaneSize != i / LaneSize)
      // This entry crosses lanes, so there is no way to model this shuffle.
      return false;
 
    // Handle the in-lane shuffles by detecting if and when they repeat. Adjust
    // later vector indices to start at multiples of LaneSize instead of Size.
    int LaneM = Mask[i] / Size;
    int LocalM = (Mask[i] % LaneSize) + (LaneM * LaneSize);
    if (RepeatedMask[i % LaneSize] == SM_SentinelUndef)
      // This is the first non-undef entry in this slot of a 128-bit lane.
      RepeatedMask[i % LaneSize] = LocalM;
    else if (RepeatedMask[i % LaneSize] != LocalM)
      // Found a mismatch with the repeated mask.
      return false;
  }
  return true;
}
 
/// Test whether a target shuffle mask is equivalent within each sub-lane.
/// Unlike isRepeatedShuffleMask we must respect SM_SentinelZero.
static bool isRepeatedTargetShuffleMask(unsigned LaneSizeInBits, MVT VT,
                                        ArrayRef<int> Mask,
                                        SmallVectorImpl<int> &RepeatedMask) {
  return isRepeatedTargetShuffleMask(LaneSizeInBits, VT.getScalarSizeInBits(),
                                     Mask, RepeatedMask);
}
 
/// Checks whether a shuffle mask is equivalent to an explicit list of
/// arguments.
///
/// This is a fast way to test a shuffle mask against a fixed pattern:
///
///   if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
///
/// It returns true if the mask is exactly as wide as the argument list, and
/// each element of the mask is either -1 (signifying undef) or the value given
/// in the argument.
static bool isShuffleEquivalent(ArrayRef<int> Mask, ArrayRef<int> ExpectedMask,
                                SDValue V1 = SDValue(),
                                SDValue V2 = SDValue()) {
  int Size = Mask.size();
  if (Size != (int)ExpectedMask.size())
    return false;
 
  for (int i = 0; i < Size; ++i) {
    assert(Mask[i] >= -1 && "Out of bound mask element!");
    int MaskIdx = Mask[i];
    int ExpectedIdx = ExpectedMask[i];
    if (0 <= MaskIdx && MaskIdx != ExpectedIdx) {
      SDValue MaskV = MaskIdx < Size ? V1 : V2;
      SDValue ExpectedV = ExpectedIdx < Size ? V1 : V2;
      MaskIdx = MaskIdx < Size ? MaskIdx : (MaskIdx - Size);
      ExpectedIdx = ExpectedIdx < Size ? ExpectedIdx : (ExpectedIdx - Size);
      if (!IsElementEquivalent(Size, MaskV, ExpectedV, MaskIdx, ExpectedIdx))
        return false;
    }
  }
  return true;
}
 
/// Checks whether a target shuffle mask is equivalent to an explicit pattern.
///
/// The masks must be exactly the same width.
///
/// If an element in Mask matches SM_SentinelUndef (-1) then the corresponding
/// value in ExpectedMask is always accepted. Otherwise the indices must match.
///
/// SM_SentinelZero is accepted as a valid negative index but must match in
/// both, or via a known bits test.
static bool isTargetShuffleEquivalent(MVT VT, ArrayRef<int> Mask,
                                      ArrayRef<int> ExpectedMask,
                                      const SelectionDAG &DAG,
                                      SDValue V1 = SDValue(),
                                      SDValue V2 = SDValue()) {
  int Size = Mask.size();
  if (Size != (int)ExpectedMask.size())
    return false;
  assert(llvm::all_of(ExpectedMask,
                      [Size](int M) {
                        return M == SM_SentinelZero ||
                               isInRange(M, 0, 2 * Size);
                      }) &&
         "Illegal target shuffle mask");
 
  // Check for out-of-range target shuffle mask indices.
  if (!isUndefOrZeroOrInRange(Mask, 0, 2 * Size))
    return false;
 
  // Don't use V1/V2 if they're not the same size as the shuffle mask type.
  if (V1 && (V1.getValueSizeInBits() != VT.getSizeInBits() ||
             !V1.getValueType().isVector()))
    V1 = SDValue();
  if (V2 && (V2.getValueSizeInBits() != VT.getSizeInBits() ||
             !V2.getValueType().isVector()))
    V2 = SDValue();
 
  APInt ZeroV1 = APInt::getZero(Size);
  APInt ZeroV2 = APInt::getZero(Size);
 
  for (int i = 0; i < Size; ++i) {
    int MaskIdx = Mask[i];
    int ExpectedIdx = ExpectedMask[i];
    if (MaskIdx == SM_SentinelUndef || MaskIdx == ExpectedIdx)
      continue;
    // If we failed to match an expected SM_SentinelZero then early out.
    if (ExpectedIdx < 0)
      return false;
    if (MaskIdx == SM_SentinelZero) {
      // If we need this expected index to be a zero element, then update the
      // relevant zero mask and perform the known bits at the end to minimize
      // repeated computes.
      SDValue ExpectedV = ExpectedIdx < Size ? V1 : V2;
      if (ExpectedV &&
          Size == (int)ExpectedV.getValueType().getVectorNumElements()) {
        int BitIdx = ExpectedIdx < Size ? ExpectedIdx : (ExpectedIdx - Size);
        APInt &ZeroMask = ExpectedIdx < Size ? ZeroV1 : ZeroV2;
        ZeroMask.setBit(BitIdx);
        continue;
      }
    }
    if (MaskIdx >= 0) {
      SDValue MaskV = MaskIdx < Size ? V1 : V2;
      SDValue ExpectedV = ExpectedIdx < Size ? V1 : V2;
      MaskIdx = MaskIdx < Size ? MaskIdx : (MaskIdx - Size);
      ExpectedIdx = ExpectedIdx < Size ? ExpectedIdx : (ExpectedIdx - Size);
      if (IsElementEquivalent(Size, MaskV, ExpectedV, MaskIdx, ExpectedIdx))
        continue;
    }
    return false;
  }
  return (ZeroV1.isZero() || DAG.MaskedVectorIsZero(V1, ZeroV1)) &&
         (ZeroV2.isZero() || DAG.MaskedVectorIsZero(V2, ZeroV2));
}
 
// Check if the shuffle mask is suitable for the AVX vpunpcklwd or vpunpckhwd
// instructions.
static bool isUnpackWdShuffleMask(ArrayRef<int> Mask, MVT VT,
                                  const SelectionDAG &DAG) {
  if (VT != MVT::v8i32 && VT != MVT::v8f32)
    return false;
 
  SmallVector<int, 8> Unpcklwd;
  createUnpackShuffleMask(MVT::v8i16, Unpcklwd, /* Lo = */ true,
                          /* Unary = */ false);
  SmallVector<int, 8> Unpckhwd;
  createUnpackShuffleMask(MVT::v8i16, Unpckhwd, /* Lo = */ false,
                          /* Unary = */ false);
  bool IsUnpackwdMask = (isTargetShuffleEquivalent(VT, Mask, Unpcklwd, DAG) ||
                         isTargetShuffleEquivalent(VT, Mask, Unpckhwd, DAG));
  return IsUnpackwdMask;
}
 
static bool is128BitUnpackShuffleMask(ArrayRef<int> Mask,
                                      const SelectionDAG &DAG) {
  // Create 128-bit vector type based on mask size.
  MVT EltVT = MVT::getIntegerVT(128 / Mask.size());
  MVT VT = MVT::getVectorVT(EltVT, Mask.size());
 
  // We can't assume a canonical shuffle mask, so try the commuted version too.
  SmallVector<int, 4> CommutedMask(Mask);
  ShuffleVectorSDNode::commuteMask(CommutedMask);
 
  // Match any of unary/binary or low/high.
  for (unsigned i = 0; i != 4; ++i) {
    SmallVector<int, 16> UnpackMask;
    createUnpackShuffleMask(VT, UnpackMask, (i >> 1) % 2, i % 2);
    if (isTargetShuffleEquivalent(VT, Mask, UnpackMask, DAG) ||
        isTargetShuffleEquivalent(VT, CommutedMask, UnpackMask, DAG))
      return true;
  }
  return false;
}
 
/// Return true if a shuffle mask chooses elements identically in its top and
/// bottom halves. For example, any splat mask has the same top and bottom
/// halves. If an element is undefined in only one half of the mask, the halves
/// are not considered identical.
static bool hasIdenticalHalvesShuffleMask(ArrayRef<int> Mask) {
  assert(Mask.size() % 2 == 0 && "Expecting even number of elements in mask");
  unsigned HalfSize = Mask.size() / 2;
  for (unsigned i = 0; i != HalfSize; ++i) {
    if (Mask[i] != Mask[i + HalfSize])
      return false;
  }
  return true;
}
 
/// Get a 4-lane 8-bit shuffle immediate for a mask.
///
/// This helper function produces an 8-bit shuffle immediate corresponding to
/// the ubiquitous shuffle encoding scheme used in x86 instructions for
/// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
/// example.
///
/// NB: We rely heavily on "undef" masks preserving the input lane.
static unsigned getV4X86ShuffleImm(ArrayRef<int> Mask) {
  assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
  assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
  assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
  assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
  assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
 
  // If the mask only uses one non-undef element, then fully 'splat' it to
  // improve later broadcast matching.
  int FirstIndex = find_if(Mask, [](int M) { return M >= 0; }) - Mask.begin();
  assert(0 <= FirstIndex && FirstIndex < 4 && "All undef shuffle mask");
 
  int FirstElt = Mask[FirstIndex];
  if (all_of(Mask, [FirstElt](int M) { return M < 0 || M == FirstElt; }))
    return (FirstElt << 6) | (FirstElt << 4) | (FirstElt << 2) | FirstElt;
 
  unsigned Imm = 0;
  Imm |= (Mask[0] < 0 ? 0 : Mask[0]) << 0;
  Imm |= (Mask[1] < 0 ? 1 : Mask[1]) << 2;
  Imm |= (Mask[2] < 0 ? 2 : Mask[2]) << 4;
  Imm |= (Mask[3] < 0 ? 3 : Mask[3]) << 6;
  return Imm;
}
 
static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, const SDLoc &DL,
                                          SelectionDAG &DAG) {
  return DAG.getTargetConstant(getV4X86ShuffleImm(Mask), DL, MVT::i8);
}
 
// Canonicalize SHUFPD mask to improve chances of further folding.
// Mask elements are assumed to be -1, 0 or 1 to match the SHUFPD lo/hi pattern.
static unsigned getSHUFPDImm(ArrayRef<int> Mask) {
  assert((Mask.size() == 2 || Mask.size() == 4 || Mask.size() == 8) &&
         "Unexpected SHUFPD mask size");
  assert(all_of(Mask, [](int M) { return -1 <= M && M <= 1; }) &&
         "Unexpected SHUFPD mask elements");
 
  // If the mask only uses one non-undef element, then fully 'splat' it to
  // improve later broadcast matching.
  int FirstIndex = find_if(Mask, [](int M) { return M >= 0; }) - Mask.begin();
  assert(0 <= FirstIndex && FirstIndex < (int)Mask.size() &&
         "All undef shuffle mask");
 
  int FirstElt = Mask[FirstIndex];
  if (all_of(Mask, [FirstElt](int M) { return M < 0 || M == FirstElt; }) &&
      count_if(Mask, [FirstElt](int M) { return M == FirstElt; }) > 1) {
    unsigned Imm = 0;
    for (unsigned I = 0, E = Mask.size(); I != E; ++I)
      Imm |= FirstElt << I;
    return Imm;
  }
 
  // Attempt to keep any undef elements in place to improve chances of the
  // shuffle becoming a (commutative) blend.
  unsigned Imm = 0;
  for (unsigned I = 0, E = Mask.size(); I != E; ++I)
    Imm |= (Mask[I] < 0 ? (I & 1) : Mask[I]) << I;
 
  return Imm;
}
 
static SDValue getSHUFPDImmForMask(ArrayRef<int> Mask, const SDLoc &DL,
                                   SelectionDAG &DAG) {
  return DAG.getTargetConstant(getSHUFPDImm(Mask), DL, MVT::i8);
}
 
// The Shuffle result is as follow:
// 0*a[0]0*a[1]...0*a[n] , n >=0 where a[] elements in a ascending order.
// Each Zeroable's element correspond to a particular Mask's element.
// As described in computeZeroableShuffleElements function.
//
// The function looks for a sub-mask that the nonzero elements are in
// increasing order. If such sub-mask exist. The function returns true.
static bool isNonZeroElementsInOrder(const APInt &Zeroable,
                                     ArrayRef<int> Mask, const EVT &VectorType,
                                     bool &IsZeroSideLeft) {
  int NextElement = -1;
  // Check if the Mask's nonzero elements are in increasing order.
  for (int i = 0, e = Mask.size(); i < e; i++) {
    // Checks if the mask's zeros elements are built from only zeros.
    assert(Mask[i] >= -1 && "Out of bound mask element!");
    if (Mask[i] < 0)
      return false;
    if (Zeroable[i])
      continue;
    // Find the lowest non zero element
    if (NextElement < 0) {
      NextElement = Mask[i] != 0 ? VectorType.getVectorNumElements() : 0;
      IsZeroSideLeft = NextElement != 0;
    }
    // Exit if the mask's non zero elements are not in increasing order.
    if (NextElement != Mask[i])
      return false;
    NextElement++;
  }
  return true;
}
 
static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
                                      ArrayRef<SDValue> Ops, SelectionDAG &DAG,
                                      const X86Subtarget &Subtarget,
                                      unsigned Depth = 0);
 
/// Try to lower a shuffle with a single PSHUFB of V1 or V2.
static SDValue lowerShuffleWithPSHUFB(const SDLoc &DL, MVT VT,
                                      ArrayRef<int> Mask, SDValue V1,
                                      SDValue V2, const APInt &Zeroable,
                                      const X86Subtarget &Subtarget,
                                      SelectionDAG &DAG) {
  int Size = Mask.size();
  int LaneSize = 128 / VT.getScalarSizeInBits();
  const int NumBytes = VT.getSizeInBits() / 8;
  const int NumEltBytes = VT.getScalarSizeInBits() / 8;
 
  assert((Subtarget.hasSSSE3() && VT.is128BitVector()) ||
         (Subtarget.hasAVX2() && VT.is256BitVector()) ||
         (Subtarget.hasBWI() && VT.is512BitVector()));
 
  SmallVector<SDValue, 64> PSHUFBMask(NumBytes);
  // Sign bit set in i8 mask means zero element.
  SDValue ZeroMask = DAG.getConstant(0x80, DL, MVT::i8);
 
  SDValue V;
  for (int i = 0; i < NumBytes; ++i) {
    int M = Mask[i / NumEltBytes];
    if (M < 0) {
      PSHUFBMask[i] = DAG.getUNDEF(MVT::i8);
      continue;
    }
    if (Zeroable[i / NumEltBytes]) {
      PSHUFBMask[i] = ZeroMask;
      continue;
    }
 
    // We can only use a single input of V1 or V2.
    SDValue SrcV = (M >= Size ? V2 : V1);
    if (V && V != SrcV)
      return SDValue();
    V = SrcV;
    M %= Size;
 
    // PSHUFB can't cross lanes, ensure this doesn't happen.
    if ((M / LaneSize) != ((i / NumEltBytes) / LaneSize))
      return SDValue();
 
    M = M % LaneSize;
    M = M * NumEltBytes + (i % NumEltBytes);
    PSHUFBMask[i] = DAG.getConstant(M, DL, MVT::i8);
  }
  assert(V && "Failed to find a source input");
 
  MVT I8VT = MVT::getVectorVT(MVT::i8, NumBytes);
  return DAG.getBitcast(
      VT, DAG.getNode(X86ISD::PSHUFB, DL, I8VT, DAG.getBitcast(I8VT, V),
                      DAG.getBuildVector(I8VT, DL, PSHUFBMask)));
}
 
/// Return Mask with the necessary casting or extending
/// for \p Mask according to \p MaskVT when lowering masking intrinsics
static SDValue getMaskNode(SDValue Mask, MVT MaskVT,
                           const X86Subtarget &Subtarget, SelectionDAG &DAG,
                           const SDLoc &dl) {
  MVT SrcVT = Mask.getSimpleValueType();
  assert(SrcVT.isScalarInteger() && "Expected scalar integer mask source!");
  assert(MaskVT.bitsLE(SrcVT) && "Unexpected mask size!");
  assert(MaskVT.getVectorElementType() == MVT::i1 && "Bool vector expected!");
 
  if (isAllOnesConstant(Mask))
    return DAG.getConstant(1, dl, MaskVT);
  if (X86::isZeroNode(Mask))
    return DAG.getConstant(0, dl, MaskVT);
 
  // Attempt to pre-truncate the mask source (to a minimum of i8).
  if (SrcVT.getSizeInBits() > MaskVT.getVectorNumElements()) {
    SrcVT = MVT::getIntegerVT(std::max((int)MaskVT.getVectorNumElements(), 8));
    Mask = DAG.getNode(ISD::TRUNCATE, dl, SrcVT, Mask);
  }
 
  if (SrcVT == MVT::i64 && Subtarget.is32Bit()) {
    assert(MaskVT == MVT::v64i1 && "Expected v64i1 mask!");
    assert(Subtarget.hasBWI() && "Expected AVX512BW target!");
    // In case 32bit mode, bitcast i64 is illegal, extend/split it.
    SDValue Lo, Hi;
    std::tie(Lo, Hi) = DAG.SplitScalar(Mask, dl, MVT::i32, MVT::i32);
    Lo = DAG.getBitcast(MVT::v32i1, Lo);
    Hi = DAG.getBitcast(MVT::v32i1, Hi);
    return DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v64i1, Lo, Hi);
  }
 
  MVT BitcastVT = MVT::getVectorVT(MVT::i1, SrcVT.getSizeInBits());
  // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
  // are extracted by EXTRACT_SUBVECTOR.
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
                     DAG.getBitcast(BitcastVT, Mask),
                     DAG.getVectorIdxConstant(0, dl));
}
 
// X86 has dedicated shuffle that can be lowered to VEXPAND
static SDValue lowerShuffleWithEXPAND(const SDLoc &DL, MVT VT, SDValue V1,
                                      SDValue V2, ArrayRef<int> Mask,
                                      const APInt &Zeroable,
                                      const X86Subtarget &Subtarget,
                                      SelectionDAG &DAG) {
  bool IsLeftZeroSide = true;
  if (!isNonZeroElementsInOrder(Zeroable, Mask, V1.getValueType(),
                                IsLeftZeroSide))
    return SDValue();
  unsigned VEXPANDMask = (~Zeroable).getZExtValue();
  MVT IntegerType =
      MVT::getIntegerVT(std::max((int)VT.getVectorNumElements(), 8));
  SDValue MaskNode = DAG.getConstant(VEXPANDMask, DL, IntegerType);
  unsigned NumElts = VT.getVectorNumElements();
  assert((NumElts == 4 || NumElts == 8 || NumElts == 16) &&
         "Unexpected number of vector elements");
  SDValue VMask = getMaskNode(MaskNode, MVT::getVectorVT(MVT::i1, NumElts),
                              Subtarget, DAG, DL);
  SDValue ZeroVector = getZeroVector(VT, Subtarget, DAG, DL);
  SDValue ExpandedVector = IsLeftZeroSide ? V2 : V1;
  return DAG.getNode(X86ISD::EXPAND, DL, VT, ExpandedVector, ZeroVector, VMask);
}
 
static bool matchShuffleWithUNPCK(MVT VT, SDValue &V1, SDValue &V2,
                                  unsigned &UnpackOpcode, bool IsUnary,
                                  ArrayRef<int> TargetMask, const SDLoc &DL,
                                  SelectionDAG &DAG,
                                  const X86Subtarget &Subtarget) {
  int NumElts = VT.getVectorNumElements();
 
  bool Undef1 = true, Undef2 = true, Zero1 = true, Zero2 = true;
  for (int i = 0; i != NumElts; i += 2) {
    int M1 = TargetMask[i + 0];
    int M2 = TargetMask[i + 1];
    Undef1 &= (SM_SentinelUndef == M1);
    Undef2 &= (SM_SentinelUndef == M2);
    Zero1 &= isUndefOrZero(M1);
    Zero2 &= isUndefOrZero(M2);
  }
  assert(!((Undef1 || Zero1) && (Undef2 || Zero2)) &&
         "Zeroable shuffle detected");
 
  // Attempt to match the target mask against the unpack lo/hi mask patterns.
  SmallVector<int, 64> Unpckl, Unpckh;
  createUnpackShuffleMask(VT, Unpckl, /* Lo = */ true, IsUnary);
  if (isTargetShuffleEquivalent(VT, TargetMask, Unpckl, DAG, V1,
                                (IsUnary ? V1 : V2))) {
    UnpackOpcode = X86ISD::UNPCKL;
    V2 = (Undef2 ? DAG.getUNDEF(VT) : (IsUnary ? V1 : V2));
    V1 = (Undef1 ? DAG.getUNDEF(VT) : V1);
    return true;
  }
 
  createUnpackShuffleMask(VT, Unpckh, /* Lo = */ false, IsUnary);
  if (isTargetShuffleEquivalent(VT, TargetMask, Unpckh, DAG, V1,
                                (IsUnary ? V1 : V2))) {
    UnpackOpcode = X86ISD::UNPCKH;
    V2 = (Undef2 ? DAG.getUNDEF(VT) : (IsUnary ? V1 : V2));
    V1 = (Undef1 ? DAG.getUNDEF(VT) : V1);
    return true;
  }
 
  // If an unary shuffle, attempt to match as an unpack lo/hi with zero.
  if (IsUnary && (Zero1 || Zero2)) {
    // Don't bother if we can blend instead.
    if ((Subtarget.hasSSE41() || VT == MVT::v2i64 || VT == MVT::v2f64) &&
        isSequentialOrUndefOrZeroInRange(TargetMask, 0, NumElts, 0))
      return false;
 
    bool MatchLo = true, MatchHi = true;
    for (int i = 0; (i != NumElts) && (MatchLo || MatchHi); ++i) {
      int M = TargetMask[i];
 
      // Ignore if the input is known to be zero or the index is undef.
      if ((((i & 1) == 0) && Zero1) || (((i & 1) == 1) && Zero2) ||
          (M == SM_SentinelUndef))
        continue;
 
      MatchLo &= (M == Unpckl[i]);
      MatchHi &= (M == Unpckh[i]);
    }
 
    if (MatchLo || MatchHi) {
      UnpackOpcode = MatchLo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
      V2 = Zero2 ? getZeroVector(VT, Subtarget, DAG, DL) : V1;
      V1 = Zero1 ? getZeroVector(VT, Subtarget, DAG, DL) : V1;
      return true;
    }
  }
 
  // If a binary shuffle, commute and try again.
  if (!IsUnary) {
    ShuffleVectorSDNode::commuteMask(Unpckl);
    if (isTargetShuffleEquivalent(VT, TargetMask, Unpckl, DAG)) {
      UnpackOpcode = X86ISD::UNPCKL;
      std::swap(V1, V2);
      return true;
    }
 
    ShuffleVectorSDNode::commuteMask(Unpckh);
    if (isTargetShuffleEquivalent(VT, TargetMask, Unpckh, DAG)) {
      UnpackOpcode = X86ISD::UNPCKH;
      std::swap(V1, V2);
      return true;
    }
  }
 
  return false;
}
 
// X86 has dedicated unpack instructions that can handle specific blend
// operations: UNPCKH and UNPCKL.
static SDValue lowerShuffleWithUNPCK(const SDLoc &DL, MVT VT, SDValue V1,
                                     SDValue V2, ArrayRef<int> Mask,
                                     SelectionDAG &DAG) {
  SmallVector<int, 8> Unpckl;
  createUnpackShuffleMask(VT, Unpckl, /* Lo = */ true, /* Unary = */ false);
  if (isShuffleEquivalent(Mask, Unpckl, V1, V2))
    return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
 
  SmallVector<int, 8> Unpckh;
  createUnpackShuffleMask(VT, Unpckh, /* Lo = */ false, /* Unary = */ false);
  if (isShuffleEquivalent(Mask, Unpckh, V1, V2))
    return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
 
  // Commute and try again.
  ShuffleVectorSDNode::commuteMask(Unpckl);
  if (isShuffleEquivalent(Mask, Unpckl, V1, V2))
    return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
 
  ShuffleVectorSDNode::commuteMask(Unpckh);
  if (isShuffleEquivalent(Mask, Unpckh, V1, V2))
    return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
 
  return SDValue();
}
 
/// Check if the mask can be mapped to a preliminary shuffle (vperm 64-bit)
/// followed by unpack 256-bit.
static SDValue lowerShuffleWithUNPCK256(const SDLoc &DL, MVT VT, SDValue V1,
                                        SDValue V2, ArrayRef<int> Mask,
                                        SelectionDAG &DAG) {
  SmallVector<int, 32> Unpckl, Unpckh;
  createSplat2ShuffleMask(VT, Unpckl, /* Lo */ true);
  createSplat2ShuffleMask(VT, Unpckh, /* Lo */ false);
 
  unsigned UnpackOpcode;
  if (isShuffleEquivalent(Mask, Unpckl, V1, V2))
    UnpackOpcode = X86ISD::UNPCKL;
  else if (isShuffleEquivalent(Mask, Unpckh, V1, V2))
    UnpackOpcode = X86ISD::UNPCKH;
  else
    return SDValue();
 
  // This is a "natural" unpack operation (rather than the 128-bit sectored
  // operation implemented by AVX). We need to rearrange 64-bit chunks of the
  // input in order to use the x86 instruction.
  V1 = DAG.getVectorShuffle(MVT::v4f64, DL, DAG.getBitcast(MVT::v4f64, V1),
                            DAG.getUNDEF(MVT::v4f64), {0, 2, 1, 3});
  V1 = DAG.getBitcast(VT, V1);
  return DAG.getNode(UnpackOpcode, DL, VT, V1, V1);
}
 
// Check if the mask can be mapped to a TRUNCATE or VTRUNC, truncating the
// source into the lower elements and zeroing the upper elements.
static bool matchShuffleAsVTRUNC(MVT &SrcVT, MVT &DstVT, MVT VT,
                                 ArrayRef<int> Mask, const APInt &Zeroable,
                                 const X86Subtarget &Subtarget) {
  if (!VT.is512BitVector() && !Subtarget.hasVLX())
    return false;
 
  unsigned NumElts = Mask.size();
  unsigned EltSizeInBits = VT.getScalarSizeInBits();
  unsigned MaxScale = 64 / EltSizeInBits;
 
  for (unsigned Scale = 2; Scale <= MaxScale; Scale += Scale) {
    unsigned SrcEltBits = EltSizeInBits * Scale;
    if (SrcEltBits < 32 && !Subtarget.hasBWI())
      continue;
    unsigned NumSrcElts = NumElts / Scale;
    if (!isSequentialOrUndefInRange(Mask, 0, NumSrcElts, 0, Scale))
      continue;
    unsigned UpperElts = NumElts - NumSrcElts;
    if (!Zeroable.extractBits(UpperElts, NumSrcElts).isAllOnes())
      continue;
    SrcVT = MVT::getIntegerVT(EltSizeInBits * Scale);
    SrcVT = MVT::getVectorVT(SrcVT, NumSrcElts);
    DstVT = MVT::getIntegerVT(EltSizeInBits);
    if ((NumSrcElts * EltSizeInBits) >= 128) {
      // ISD::TRUNCATE
      DstVT = MVT::getVectorVT(DstVT, NumSrcElts);
    } else {
      // X86ISD::VTRUNC
      DstVT = MVT::getVectorVT(DstVT, 128 / EltSizeInBits);
    }
    return true;
  }
 
  return false;
}
 
// Helper to create TRUNCATE/VTRUNC nodes, optionally with zero/undef upper
// element padding to the final DstVT.
static SDValue getAVX512TruncNode(const SDLoc &DL, MVT DstVT, SDValue Src,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG, bool ZeroUppers) {
  MVT SrcVT = Src.getSimpleValueType();
  MVT DstSVT = DstVT.getScalarType();
  unsigned NumDstElts = DstVT.getVectorNumElements();
  unsigned NumSrcElts = SrcVT.getVectorNumElements();
  unsigned DstEltSizeInBits = DstVT.getScalarSizeInBits();
 
  if (!DAG.getTargetLoweringInfo().isTypeLegal(SrcVT))
    return SDValue();
 
  // Perform a direct ISD::TRUNCATE if possible.
  if (NumSrcElts == NumDstElts)
    return DAG.getNode(ISD::TRUNCATE, DL, DstVT, Src);
 
  if (NumSrcElts > NumDstElts) {
    MVT TruncVT = MVT::getVectorVT(DstSVT, NumSrcElts);
    SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Src);
    return extractSubVector(Trunc, 0, DAG, DL, DstVT.getSizeInBits());
  }
 
  if ((NumSrcElts * DstEltSizeInBits) >= 128) {
    MVT TruncVT = MVT::getVectorVT(DstSVT, NumSrcElts);
    SDValue Trunc = DAG.getNode(ISD::TRUNCATE, DL, TruncVT, Src);
    return widenSubVector(Trunc, ZeroUppers, Subtarget, DAG, DL,
                          DstVT.getSizeInBits());
  }
 
  // Non-VLX targets must truncate from a 512-bit type, so we need to
  // widen, truncate and then possibly extract the original subvector.
  if (!Subtarget.hasVLX() && !SrcVT.is512BitVector()) {
    SDValue NewSrc = widenSubVector(Src, ZeroUppers, Subtarget, DAG, DL, 512);
    return getAVX512TruncNode(DL, DstVT, NewSrc, Subtarget, DAG, ZeroUppers);
  }
 
  // Fallback to a X86ISD::VTRUNC, padding if necessary.
  MVT TruncVT = MVT::getVectorVT(DstSVT, 128 / DstEltSizeInBits);
  SDValue Trunc = DAG.getNode(X86ISD::VTRUNC, DL, TruncVT, Src);
  if (DstVT != TruncVT)
    Trunc = widenSubVector(Trunc, ZeroUppers, Subtarget, DAG, DL,
                           DstVT.getSizeInBits());
  return Trunc;
}
 
// Try to lower trunc+vector_shuffle to a vpmovdb or a vpmovdw instruction.
//
// An example is the following:
//
// t0: ch = EntryToken
//           t2: v4i64,ch = CopyFromReg t0, Register:v4i64 %0
//         t25: v4i32 = truncate t2
//       t41: v8i16 = bitcast t25
//       t21: v8i16 = BUILD_VECTOR undef:i16, undef:i16, undef:i16, undef:i16,
//       Constant:i16<0>, Constant:i16<0>, Constant:i16<0>, Constant:i16<0>
//     t51: v8i16 = vector_shuffle<0,2,4,6,12,13,14,15> t41, t21
//   t18: v2i64 = bitcast t51
//
// One can just use a single vpmovdw instruction, without avx512vl we need to
// use the zmm variant and extract the lower subvector, padding with zeroes.
// TODO: Merge with lowerShuffleAsVTRUNC.
static SDValue lowerShuffleWithVPMOV(const SDLoc &DL, MVT VT, SDValue V1,
                                     SDValue V2, ArrayRef<int> Mask,
                                     const APInt &Zeroable,
                                     const X86Subtarget &Subtarget,
                                     SelectionDAG &DAG) {
  assert((VT == MVT::v16i8 || VT == MVT::v8i16) && "Unexpected VTRUNC type");
  if (!Subtarget.hasAVX512())
    return SDValue();
 
  unsigned NumElts = VT.getVectorNumElements();
  unsigned EltSizeInBits = VT.getScalarSizeInBits();
  unsigned MaxScale = 64 / EltSizeInBits;
  for (unsigned Scale = 2; Scale <= MaxScale; Scale += Scale) {
    unsigned SrcEltBits = EltSizeInBits * Scale;
    unsigned NumSrcElts = NumElts / Scale;
    unsigned UpperElts = NumElts - NumSrcElts;
    if (!isSequentialOrUndefInRange(Mask, 0, NumSrcElts, 0, Scale) ||
        !Zeroable.extractBits(UpperElts, NumSrcElts).isAllOnes())
      continue;
 
    // Attempt to find a matching source truncation, but as a fall back VLX
    // cases can use the VPMOV directly.
    SDValue Src = peekThroughBitcasts(V1);
    if (Src.getOpcode() == ISD::TRUNCATE &&
        Src.getScalarValueSizeInBits() == SrcEltBits) {
      Src = Src.getOperand(0);
    } else if (Subtarget.hasVLX()) {
      MVT SrcSVT = MVT::getIntegerVT(SrcEltBits);
      MVT SrcVT = MVT::getVectorVT(SrcSVT, NumSrcElts);
      Src = DAG.getBitcast(SrcVT, Src);
      // Don't do this if PACKSS/PACKUS could perform it cheaper.
      if (Scale == 2 &&
          ((DAG.ComputeNumSignBits(Src) > EltSizeInBits) ||
           (DAG.computeKnownBits(Src).countMinLeadingZeros() >= EltSizeInBits)))
        return SDValue();
    } else
      return SDValue();
 
    // VPMOVWB is only available with avx512bw.
    if (!Subtarget.hasBWI() && Src.getScalarValueSizeInBits() < 32)
      return SDValue();
 
    bool UndefUppers = isUndefInRange(Mask, NumSrcElts, UpperElts);
    return getAVX512TruncNode(DL, VT, Src, Subtarget, DAG, !UndefUppers);
  }
 
  return SDValue();
}
 
// Attempt to match binary shuffle patterns as a truncate.
static SDValue lowerShuffleAsVTRUNC(const SDLoc &DL, MVT VT, SDValue V1,
                                    SDValue V2, ArrayRef<int> Mask,
                                    const APInt &Zeroable,
                                    const X86Subtarget &Subtarget,
                                    SelectionDAG &DAG) {
  assert((VT.is128BitVector() || VT.is256BitVector()) &&
         "Unexpected VTRUNC type");
  if (!Subtarget.hasAVX512() ||
      (VT.is256BitVector() && !Subtarget.useAVX512Regs()))
    return SDValue();
 
  unsigned NumElts = VT.getVectorNumElements();
  unsigned EltSizeInBits = VT.getScalarSizeInBits();
  unsigned MaxScale = 64 / EltSizeInBits;
  for (unsigned Scale = 2; Scale <= MaxScale; Scale += Scale) {
    // TODO: Support non-BWI VPMOVWB truncations?
    unsigned SrcEltBits = EltSizeInBits * Scale;
    if (SrcEltBits < 32 && !Subtarget.hasBWI())
      continue;
 
    // Match shuffle <Ofs,Ofs+Scale,Ofs+2*Scale,..,undef_or_zero,undef_or_zero>
    // Bail if the V2 elements are undef.
    unsigned NumHalfSrcElts = NumElts / Scale;
    unsigned NumSrcElts = 2 * NumHalfSrcElts;
    for (unsigned Offset = 0; Offset != Scale; ++Offset) {
      if (!isSequentialOrUndefInRange(Mask, 0, NumSrcElts, Offset, Scale) ||
          isUndefInRange(Mask, NumHalfSrcElts, NumHalfSrcElts))
        continue;
 
      // The elements beyond the truncation must be undef/zero.
      unsigned UpperElts = NumElts - NumSrcElts;
      if (UpperElts > 0 &&
          !Zeroable.extractBits(UpperElts, NumSrcElts).isAllOnes())
        continue;
      bool UndefUppers =
          UpperElts > 0 && isUndefInRange(Mask, NumSrcElts, UpperElts);
 
      // As we're using both sources then we need to concat them together
      // and truncate from the double-sized src.
      MVT ConcatVT = VT.getDoubleNumVectorElementsVT();
 
      // For offset truncations, ensure that the concat is cheap.
      SDValue Src =
          combineConcatVectorOps(DL, ConcatVT, {V1, V2}, DAG, Subtarget);
      if (!Src) {
        if (Offset)
          continue;
        Src = DAG.getNode(ISD::CONCAT_VECTORS, DL, ConcatVT, V1, V2);
      }
 
      MVT SrcSVT = MVT::getIntegerVT(SrcEltBits);
      MVT SrcVT = MVT::getVectorVT(SrcSVT, NumSrcElts);
      Src = getTargetVShiftByConstNode(X86ISD::VSRLI, DL, SrcVT, Src,
                                       Offset * EltSizeInBits, DAG);
      return getAVX512TruncNode(DL, VT, Src, Subtarget, DAG, !UndefUppers);
    }
  }
 
  return SDValue();
}
 
/// Check whether a compaction lowering can be done by dropping even/odd
/// elements and compute how many times even/odd elements must be dropped.
///
/// This handles shuffles which take every Nth element where N is a power of
/// two. Example shuffle masks:
///
/// (even)
///  N = 1:  0,  2,  4,  6,  8, 10, 12, 14,  0,  2,  4,  6,  8, 10, 12, 14
///  N = 1:  0,  2,  4,  6,  8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
///  N = 2:  0,  4,  8, 12,  0,  4,  8, 12,  0,  4,  8, 12,  0,  4,  8, 12
///  N = 2:  0,  4,  8, 12, 16, 20, 24, 28,  0,  4,  8, 12, 16, 20, 24, 28
///  N = 3:  0,  8,  0,  8,  0,  8,  0,  8,  0,  8,  0,  8,  0,  8,  0,  8
///  N = 3:  0,  8, 16, 24,  0,  8, 16, 24,  0,  8, 16, 24,  0,  8, 16, 24
///
/// (odd)
///  N = 1:  1,  3,  5,  7,  9, 11, 13, 15,  0,  2,  4,  6,  8, 10, 12, 14
///  N = 1:  1,  3,  5,  7,  9, 11, 13, 15, 17, 19, 21, 23, 25, 27, 29, 31
///
/// Any of these lanes can of course be undef.
///
/// This routine only supports N <= 3.
/// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
/// for larger N.
///
/// \returns N above, or the number of times even/odd elements must be dropped
/// if there is such a number. Otherwise returns zero.
static int canLowerByDroppingElements(ArrayRef<int> Mask, bool MatchEven,
                                      bool IsSingleInput) {
  // The modulus for the shuffle vector entries is based on whether this is
  // a single input or not.
  int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
  assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
         "We should only be called with masks with a power-of-2 size!");
 
  uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
  int Offset = MatchEven ? 0 : 1;
 
  // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
  // and 2^3 simultaneously. This is because we may have ambiguity with
  // partially undef inputs.
  bool ViableForN[3] = {true, true, true};
 
  for (int i = 0, e = Mask.size(); i < e; ++i) {
    // Ignore undef lanes, we'll optimistically collapse them to the pattern we
    // want.
    if (Mask[i] < 0)
      continue;
 
    bool IsAnyViable = false;
    for (unsigned j = 0; j != std::size(ViableForN); ++j)
      if (ViableForN[j]) {
        uint64_t N = j + 1;
 
        // The shuffle mask must be equal to (i * 2^N) % M.
        if ((uint64_t)(Mask[i] - Offset) == (((uint64_t)i << N) & ModMask))
          IsAnyViable = true;
        else
          ViableForN[j] = false;
      }
    // Early exit if we exhaust the possible powers of two.
    if (!IsAnyViable)
      break;
  }
 
  for (unsigned j = 0; j != std::size(ViableForN); ++j)
    if (ViableForN[j])
      return j + 1;
 
  // Return 0 as there is no viable power of two.
  return 0;
}
 
// X86 has dedicated pack instructions that can handle specific truncation
// operations: PACKSS and PACKUS.
// Checks for compaction shuffle masks if MaxStages > 1.
// TODO: Add support for matching multiple PACKSS/PACKUS stages.
static bool matchShuffleWithPACK(MVT VT, MVT &SrcVT, SDValue &V1, SDValue &V2,
                                 unsigned &PackOpcode, ArrayRef<int> TargetMask,
                                 const SelectionDAG &DAG,
                                 const X86Subtarget &Subtarget,
                                 unsigned MaxStages = 1) {
  unsigned NumElts = VT.getVectorNumElements();
  unsigned BitSize = VT.getScalarSizeInBits();
  assert(0 < MaxStages && MaxStages <= 3 && (BitSize << MaxStages) <= 64 &&
         "Illegal maximum compaction");
 
  auto MatchPACK = [&](SDValue N1, SDValue N2, MVT PackVT) {
    unsigned NumSrcBits = PackVT.getScalarSizeInBits();
    unsigned NumPackedBits = NumSrcBits - BitSize;
    N1 = peekThroughBitcasts(N1);
    N2 = peekThroughBitcasts(N2);
    unsigned NumBits1 = N1.getScalarValueSizeInBits();
    unsigned NumBits2 = N2.getScalarValueSizeInBits();
    bool IsZero1 = llvm::isNullOrNullSplat(N1, /*AllowUndefs*/ false);
    bool IsZero2 = llvm::isNullOrNullSplat(N2, /*AllowUndefs*/ false);
    if ((!N1.isUndef() && !IsZero1 && NumBits1 != NumSrcBits) ||
        (!N2.isUndef() && !IsZero2 && NumBits2 != NumSrcBits))
      return false;
    if (Subtarget.hasSSE41() || BitSize == 8) {
      APInt ZeroMask = APInt::getHighBitsSet(NumSrcBits, NumPackedBits);
      if ((N1.isUndef() || IsZero1 || DAG.MaskedValueIsZero(N1, ZeroMask)) &&
          (N2.isUndef() || IsZero2 || DAG.MaskedValueIsZero(N2, ZeroMask))) {
        V1 = N1;
        V2 = N2;
        SrcVT = PackVT;
        PackOpcode = X86ISD::PACKUS;
        return true;
      }
    }
    bool IsAllOnes1 = llvm::isAllOnesOrAllOnesSplat(N1, /*AllowUndefs*/ false);
    bool IsAllOnes2 = llvm::isAllOnesOrAllOnesSplat(N2, /*AllowUndefs*/ false);
    if ((N1.isUndef() || IsZero1 || IsAllOnes1 ||
         DAG.ComputeNumSignBits(N1) > NumPackedBits) &&
        (N2.isUndef() || IsZero2 || IsAllOnes2 ||
         DAG.ComputeNumSignBits(N2) > NumPackedBits)) {
      V1 = N1;
      V2 = N2;
      SrcVT = PackVT;
      PackOpcode = X86ISD::PACKSS;
      return true;
    }
    return false;
  };
 
  // Attempt to match against wider and wider compaction patterns.
  for (unsigned NumStages = 1; NumStages <= MaxStages; ++NumStages) {
    MVT PackSVT = MVT::getIntegerVT(BitSize << NumStages);
    MVT PackVT = MVT::getVectorVT(PackSVT, NumElts >> NumStages);
 
    // Try binary shuffle.
    SmallVector<int, 32> BinaryMask;
    createPackShuffleMask(VT, BinaryMask, false, NumStages);
    if (isTargetShuffleEquivalent(VT, TargetMask, BinaryMask, DAG, V1, V2))
      if (MatchPACK(V1, V2, PackVT))
        return true;
 
    // Try unary shuffle.
    SmallVector<int, 32> UnaryMask;
    createPackShuffleMask(VT, UnaryMask, true, NumStages);
    if (isTargetShuffleEquivalent(VT, TargetMask, UnaryMask, DAG, V1))
      if (MatchPACK(V1, V1, PackVT))
        return true;
  }
 
  return false;
}
 
static SDValue lowerShuffleWithPACK(const SDLoc &DL, MVT VT, SDValue V1,
                                    SDValue V2, ArrayRef<int> Mask,
                                    const X86Subtarget &Subtarget,
                                    SelectionDAG &DAG) {
  MVT PackVT;
  unsigned PackOpcode;
  unsigned SizeBits = VT.getSizeInBits();
  unsigned EltBits = VT.getScalarSizeInBits();
  unsigned MaxStages = Log2_32(64 / EltBits);
  if (!matchShuffleWithPACK(VT, PackVT, V1, V2, PackOpcode, Mask, DAG,
                            Subtarget, MaxStages))
    return SDValue();
 
  unsigned CurrentEltBits = PackVT.getScalarSizeInBits();
  unsigned NumStages = Log2_32(CurrentEltBits / EltBits);
 
  // Don't lower multi-stage packs on AVX512, truncation is better.
  if (NumStages != 1 && SizeBits == 128 && Subtarget.hasVLX())
    return SDValue();
 
  // Pack to the largest type possible:
  // vXi64/vXi32 -> PACK*SDW and vXi16 -> PACK*SWB.
  unsigned MaxPackBits = 16;
  if (CurrentEltBits > 16 &&
      (PackOpcode == X86ISD::PACKSS || Subtarget.hasSSE41()))
    MaxPackBits = 32;
 
  // Repeatedly pack down to the target size.
  SDValue Res;
  for (unsigned i = 0; i != NumStages; ++i) {
    unsigned SrcEltBits = std::min(MaxPackBits, CurrentEltBits);
    unsigned NumSrcElts = SizeBits / SrcEltBits;
    MVT SrcSVT = MVT::getIntegerVT(SrcEltBits);
    MVT DstSVT = MVT::getIntegerVT(SrcEltBits / 2);
    MVT SrcVT = MVT::getVectorVT(SrcSVT, NumSrcElts);
    MVT DstVT = MVT::getVectorVT(DstSVT, NumSrcElts * 2);
    Res = DAG.getNode(PackOpcode, DL, DstVT, DAG.getBitcast(SrcVT, V1),
                      DAG.getBitcast(SrcVT, V2));
    V1 = V2 = Res;
    CurrentEltBits /= 2;
  }
  assert(Res && Res.getValueType() == VT &&
         "Failed to lower compaction shuffle");
  return Res;
}
 
/// Try to emit a bitmask instruction for a shuffle.
///
/// This handles cases where we can model a blend exactly as a bitmask due to
/// one of the inputs being zeroable.
static SDValue lowerShuffleAsBitMask(const SDLoc &DL, MVT VT, SDValue V1,
                                     SDValue V2, ArrayRef<int> Mask,
                                     const APInt &Zeroable,
                                     const X86Subtarget &Subtarget,
                                     SelectionDAG &DAG) {
  MVT MaskVT = VT;
  MVT EltVT = VT.getVectorElementType();
  SDValue Zero, AllOnes;
  // Use f64 if i64 isn't legal.
  if (EltVT == MVT::i64 && !Subtarget.is64Bit()) {
    EltVT = MVT::f64;
    MaskVT = MVT::getVectorVT(EltVT, Mask.size());
  }
 
  MVT LogicVT = VT;
  if (EltVT.isFloatingPoint()) {
    Zero = DAG.getConstantFP(0.0, DL, EltVT);
    APFloat AllOnesValue = APFloat::getAllOnesValue(EltVT.getFltSemantics());
    AllOnes = DAG.getConstantFP(AllOnesValue, DL, EltVT);
    LogicVT = MVT::getVectorVT(EltVT.changeTypeToInteger(), Mask.size());
  } else {
    Zero = DAG.getConstant(0, DL, EltVT);
    AllOnes = DAG.getAllOnesConstant(DL, EltVT);
  }
 
  SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
  SDValue V;
  for (int i = 0, Size = Mask.size(); i < Size; ++i) {
    if (Zeroable[i])
      continue;
    if (Mask[i] % Size != i)
      return SDValue(); // Not a blend.
    if (!V)
      V = Mask[i] < Size ? V1 : V2;
    else if (V != (Mask[i] < Size ? V1 : V2))
      return SDValue(); // Can only let one input through the mask.
 
    VMaskOps[i] = AllOnes;
  }
  if (!V)
    return SDValue(); // No non-zeroable elements!
 
  SDValue VMask = DAG.getBuildVector(MaskVT, DL, VMaskOps);
  VMask = DAG.getBitcast(LogicVT, VMask);
  V = DAG.getBitcast(LogicVT, V);
  SDValue And = DAG.getNode(ISD::AND, DL, LogicVT, V, VMask);
  return DAG.getBitcast(VT, And);
}
 
/// Try to emit a blend instruction for a shuffle using bit math.
///
/// This is used as a fallback approach when first class blend instructions are
/// unavailable. Currently it is only suitable for integer vectors, but could
/// be generalized for floating point vectors if desirable.
static SDValue lowerShuffleAsBitBlend(const SDLoc &DL, MVT VT, SDValue V1,
                                      SDValue V2, ArrayRef<int> Mask,
                                      SelectionDAG &DAG) {
  assert(VT.isInteger() && "Only supports integer vector types!");
  MVT EltVT = VT.getVectorElementType();
  SDValue Zero = DAG.getConstant(0, DL, EltVT);
  SDValue AllOnes = DAG.getAllOnesConstant(DL, EltVT);
  SmallVector<SDValue, 16> MaskOps;
  for (int i = 0, Size = Mask.size(); i < Size; ++i) {
    if (Mask[i] >= 0 && Mask[i] != i && Mask[i] != i + Size)
      return SDValue(); // Shuffled input!
    MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
  }
 
  SDValue V1Mask = DAG.getBuildVector(VT, DL, MaskOps);
  return getBitSelect(DL, VT, V1, V2, V1Mask, DAG);
}
 
static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
                                    SDValue PreservedSrc,
                                    const X86Subtarget &Subtarget,
                                    SelectionDAG &DAG);
 
static bool matchShuffleAsBlend(MVT VT, SDValue V1, SDValue V2,
                                MutableArrayRef<int> Mask,
                                const APInt &Zeroable, bool &ForceV1Zero,
                                bool &ForceV2Zero, uint64_t &BlendMask) {
  bool V1IsZeroOrUndef =
      V1.isUndef() || ISD::isBuildVectorAllZeros(V1.getNode());
  bool V2IsZeroOrUndef =
      V2.isUndef() || ISD::isBuildVectorAllZeros(V2.getNode());
 
  BlendMask = 0;
  ForceV1Zero = false, ForceV2Zero = false;
  assert(Mask.size() <= 64 && "Shuffle mask too big for blend mask");
 
  int NumElts = Mask.size();
  int NumLanes = VT.getSizeInBits() / 128;
  int NumEltsPerLane = NumElts / NumLanes;
  assert((NumLanes * NumEltsPerLane) == NumElts && "Value type mismatch");
 
  // For 32/64-bit elements, if we only reference one input (plus any undefs),
  // then ensure the blend mask part for that lane just references that input.
  bool ForceWholeLaneMasks =
      VT.is256BitVector() && VT.getScalarSizeInBits() >= 32;
 
  // Attempt to generate the binary blend mask. If an input is zero then
  // we can use any lane.
  for (int Lane = 0; Lane != NumLanes; ++Lane) {
    // Keep track of the inputs used per lane.
    bool LaneV1InUse = false;
    bool LaneV2InUse = false;
    uint64_t LaneBlendMask = 0;
    for (int LaneElt = 0; LaneElt != NumEltsPerLane; ++LaneElt) {
      int Elt = (Lane * NumEltsPerLane) + LaneElt;
      int M = Mask[Elt];
      if (M == SM_SentinelUndef)
        continue;
      if (M == Elt || (0 <= M && M < NumElts &&
                     IsElementEquivalent(NumElts, V1, V1, M, Elt))) {
        Mask[Elt] = Elt;
        LaneV1InUse = true;
        continue;
      }
      if (M == (Elt + NumElts) ||
          (NumElts <= M &&
           IsElementEquivalent(NumElts, V2, V2, M - NumElts, Elt))) {
        LaneBlendMask |= 1ull << LaneElt;
        Mask[Elt] = Elt + NumElts;
        LaneV2InUse = true;
        continue;
      }
      if (Zeroable[Elt]) {
        if (V1IsZeroOrUndef) {
          ForceV1Zero = true;
          Mask[Elt] = Elt;
          LaneV1InUse = true;
          continue;
        }
        if (V2IsZeroOrUndef) {
          ForceV2Zero = true;
          LaneBlendMask |= 1ull << LaneElt;
          Mask[Elt] = Elt + NumElts;
          LaneV2InUse = true;
          continue;
        }
      }
      return false;
    }
 
    // If we only used V2 then splat the lane blend mask to avoid any demanded
    // elts from V1 in this lane (the V1 equivalent is implicit with a zero
    // blend mask bit).
    if (ForceWholeLaneMasks && LaneV2InUse && !LaneV1InUse)
      LaneBlendMask = (1ull << NumEltsPerLane) - 1;
 
    BlendMask |= LaneBlendMask << (Lane * NumEltsPerLane);
  }
  return true;
}
 
/// Try to emit a blend instruction for a shuffle.
///
/// This doesn't do any checks for the availability of instructions for blending
/// these values. It relies on the availability of the X86ISD::BLENDI pattern to
/// be matched in the backend with the type given. What it does check for is
/// that the shuffle mask is a blend, or convertible into a blend with zero.
static SDValue lowerShuffleAsBlend(const SDLoc &DL, MVT VT, SDValue V1,
                                   SDValue V2, ArrayRef<int> Original,
                                   const APInt &Zeroable,
                                   const X86Subtarget &Subtarget,
                                   SelectionDAG &DAG) {
  uint64_t BlendMask = 0;
  bool ForceV1Zero = false, ForceV2Zero = false;
  SmallVector<int, 64> Mask(Original);
  if (!matchShuffleAsBlend(VT, V1, V2, Mask, Zeroable, ForceV1Zero, ForceV2Zero,
                           BlendMask))
    return SDValue();
 
  // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
  if (ForceV1Zero)
    V1 = getZeroVector(VT, Subtarget, DAG, DL);
  if (ForceV2Zero)
    V2 = getZeroVector(VT, Subtarget, DAG, DL);
 
  unsigned NumElts = VT.getVectorNumElements();
 
  switch (VT.SimpleTy) {
  case MVT::v4i64:
  case MVT::v8i32:
    assert(Subtarget.hasAVX2() && "256-bit integer blends require AVX2!");
    [[fallthrough]];
  case MVT::v4f64:
  case MVT::v8f32:
    assert(Subtarget.hasAVX() && "256-bit float blends require AVX!");
    [[fallthrough]];
  case MVT::v2f64:
  case MVT::v2i64:
  case MVT::v4f32:
  case MVT::v4i32:
  case MVT::v8i16:
    assert(Subtarget.hasSSE41() && "128-bit blends require SSE41!");
    return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
                       DAG.getTargetConstant(BlendMask, DL, MVT::i8));
  case MVT::v16i16: {
    assert(Subtarget.hasAVX2() && "v16i16 blends require AVX2!");
    SmallVector<int, 8> RepeatedMask;
    if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
      // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
      assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
      BlendMask = 0;
      for (int i = 0; i < 8; ++i)
        if (RepeatedMask[i] >= 8)
          BlendMask |= 1ull << i;
      return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
                         DAG.getTargetConstant(BlendMask, DL, MVT::i8));
    }
    // Use PBLENDW for lower/upper lanes and then blend lanes.
    // TODO - we should allow 2 PBLENDW here and leave shuffle combine to
    // merge to VSELECT where useful.
    uint64_t LoMask = BlendMask & 0xFF;
    uint64_t HiMask = (BlendMask >> 8) & 0xFF;
    if (LoMask == 0 || LoMask == 255 || HiMask == 0 || HiMask == 255) {
      SDValue Lo = DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
                               DAG.getTargetConstant(LoMask, DL, MVT::i8));
      SDValue Hi = DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
                               DAG.getTargetConstant(HiMask, DL, MVT::i8));
      return DAG.getVectorShuffle(
          MVT::v16i16, DL, Lo, Hi,
          {0, 1, 2, 3, 4, 5, 6, 7, 24, 25, 26, 27, 28, 29, 30, 31});
    }
    [[fallthrough]];
  }
  case MVT::v32i8:
    assert(Subtarget.hasAVX2() && "256-bit byte-blends require AVX2!");
    [[fallthrough]];
  case MVT::v16i8: {
    assert(Subtarget.hasSSE41() && "128-bit byte-blends require SSE41!");
 
    // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
    if (SDValue Masked = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable,
                                               Subtarget, DAG))
      return Masked;
 
    if (Subtarget.hasBWI() && Subtarget.hasVLX()) {
      MVT IntegerType = MVT::getIntegerVT(std::max<unsigned>(NumElts, 8));
      SDValue MaskNode = DAG.getConstant(BlendMask, DL, IntegerType);
      return getVectorMaskingNode(V2, MaskNode, V1, Subtarget, DAG);
    }
 
    // If we have VPTERNLOG, we can use that as a bit blend.
    if (Subtarget.hasVLX())
      if (SDValue BitBlend =
              lowerShuffleAsBitBlend(DL, VT, V1, V2, Mask, DAG))
        return BitBlend;
 
    // Scale the blend by the number of bytes per element.
    int Scale = VT.getScalarSizeInBits() / 8;
 
    // This form of blend is always done on bytes. Compute the byte vector
    // type.
    MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
 
    // x86 allows load folding with blendvb from the 2nd source operand. But
    // we are still using LLVM select here (see comment below), so that's V1.
    // If V2 can be load-folded and V1 cannot be load-folded, then commute to
    // allow that load-folding possibility.
    if (!ISD::isNormalLoad(V1.getNode()) && ISD::isNormalLoad(V2.getNode())) {
      ShuffleVectorSDNode::commuteMask(Mask);
      std::swap(V1, V2);
    }
 
    // Compute the VSELECT mask. Note that VSELECT is really confusing in the
    // mix of LLVM's code generator and the x86 backend. We tell the code
    // generator that boolean values in the elements of an x86 vector register
    // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
    // mapping a select to operand #1, and 'false' mapping to operand #2. The
    // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
    // of the element (the remaining are ignored) and 0 in that high bit would
    // mean operand #1 while 1 in the high bit would mean operand #2. So while
    // the LLVM model for boolean values in vector elements gets the relevant
    // bit set, it is set backwards and over constrained relative to x86's
    // actual model.
    SmallVector<SDValue, 32> VSELECTMask;
    for (int i = 0, Size = Mask.size(); i < Size; ++i)
      for (int j = 0; j < Scale; ++j)
        VSELECTMask.push_back(
            Mask[i] < 0
                ? DAG.getUNDEF(MVT::i8)
                : DAG.getSignedConstant(Mask[i] < Size ? -1 : 0, DL, MVT::i8));
 
    V1 = DAG.getBitcast(BlendVT, V1);
    V2 = DAG.getBitcast(BlendVT, V2);
    return DAG.getBitcast(
        VT,
        DAG.getSelect(DL, BlendVT, DAG.getBuildVector(BlendVT, DL, VSELECTMask),
                      V1, V2));
  }
  case MVT::v16f32:
  case MVT::v8f64:
  case MVT::v8i64:
  case MVT::v16i32:
  case MVT::v32i16:
  case MVT::v64i8: {
    // Attempt to lower to a bitmask if we can. Only if not optimizing for size.
    bool OptForSize = DAG.shouldOptForSize();
    if (!OptForSize) {
      if (SDValue Masked = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable,
                                                 Subtarget, DAG))
        return Masked;
    }
 
    // Otherwise load an immediate into a GPR, cast to k-register, and use a
    // masked move.
    MVT IntegerType = MVT::getIntegerVT(std::max<unsigned>(NumElts, 8));
    SDValue MaskNode = DAG.getConstant(BlendMask, DL, IntegerType);
    return getVectorMaskingNode(V2, MaskNode, V1, Subtarget, DAG);
  }
  default:
    llvm_unreachable("Not a supported integer vector type!");
  }
}
 
/// Try to lower as a blend of elements from two inputs followed by
/// a single-input permutation.
///
/// This matches the pattern where we can blend elements from two inputs and
/// then reduce the shuffle to a single-input permutation.
static SDValue lowerShuffleAsBlendAndPermute(const SDLoc &DL, MVT VT,
                                             SDValue V1, SDValue V2,
                                             ArrayRef<int> Mask,
                                             SelectionDAG &DAG,
                                             bool ImmBlends = false) {
  // We build up the blend mask while checking whether a blend is a viable way
  // to reduce the shuffle.
  SmallVector<int, 32> BlendMask(Mask.size(), -1);
  SmallVector<int, 32> PermuteMask(Mask.size(), -1);
 
  for (int i = 0, Size = Mask.size(); i < Size; ++i) {
    if (Mask[i] < 0)
      continue;
 
    assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
 
    if (BlendMask[Mask[i] % Size] < 0)
      BlendMask[Mask[i] % Size] = Mask[i];
    else if (BlendMask[Mask[i] % Size] != Mask[i])
      return SDValue(); // Can't blend in the needed input!
 
    PermuteMask[i] = Mask[i] % Size;
  }
 
  // If only immediate blends, then bail if the blend mask can't be widened to
  // i16.
  unsigned EltSize = VT.getScalarSizeInBits();
  if (ImmBlends && EltSize == 8 && !canWidenShuffleElements(BlendMask))
    return SDValue();
 
  SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
  return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
}
 
/// Try to lower as an unpack of elements from two inputs followed by
/// a single-input permutation.
///
/// This matches the pattern where we can unpack elements from two inputs and
/// then reduce the shuffle to a single-input (wider) permutation.
static SDValue lowerShuffleAsUNPCKAndPermute(const SDLoc &DL, MVT VT,
                                             SDValue V1, SDValue V2,
                                             ArrayRef<int> Mask,
                                             SelectionDAG &DAG) {
  int NumElts = Mask.size();
  int NumLanes = VT.getSizeInBits() / 128;
  int NumLaneElts = NumElts / NumLanes;
  int NumHalfLaneElts = NumLaneElts / 2;
 
  bool MatchLo = true, MatchHi = true;
  SDValue Ops[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT)};
 
  // Determine UNPCKL/UNPCKH type and operand order.
  for (int Elt = 0; Elt != NumElts; ++Elt) {
    int M = Mask[Elt];
    if (M < 0)
      continue;
 
    // Normalize the mask value depending on whether it's V1 or V2.
    int NormM = M;
    SDValue &Op = Ops[Elt & 1];
    if (M < NumElts && (Op.isUndef() || Op == V1))
      Op = V1;
    else if (NumElts <= M && (Op.isUndef() || Op == V2)) {
      Op = V2;
      NormM -= NumElts;
    } else
      return SDValue();
 
    bool MatchLoAnyLane = false, MatchHiAnyLane = false;
    for (int Lane = 0; Lane != NumElts; Lane += NumLaneElts) {
      int Lo = Lane, Mid = Lane + NumHalfLaneElts, Hi = Lane + NumLaneElts;
      MatchLoAnyLane |= isUndefOrInRange(NormM, Lo, Mid);
      MatchHiAnyLane |= isUndefOrInRange(NormM, Mid, Hi);
      if (MatchLoAnyLane || MatchHiAnyLane) {
        assert((MatchLoAnyLane ^ MatchHiAnyLane) &&
               "Failed to match UNPCKLO/UNPCKHI");
        break;
      }
    }
    MatchLo &= MatchLoAnyLane;
    MatchHi &= MatchHiAnyLane;
    if (!MatchLo && !MatchHi)
      return SDValue();
  }
  assert((MatchLo ^ MatchHi) && "Failed to match UNPCKLO/UNPCKHI");
 
  // Element indices have changed after unpacking. Calculate permute mask
  // so that they will be put back to the position as dictated by the
  // original shuffle mask indices.
  SmallVector<int, 32> PermuteMask(NumElts, -1);
  for (int Elt = 0; Elt != NumElts; ++Elt) {
    int M = Mask[Elt];
    if (M < 0)
      continue;
    int NormM = M;
    if (NumElts <= M)
      NormM -= NumElts;
    bool IsFirstOp = M < NumElts;
    int BaseMaskElt =
        NumLaneElts * (NormM / NumLaneElts) + (2 * (NormM % NumHalfLaneElts));
    if ((IsFirstOp && V1 == Ops[0]) || (!IsFirstOp && V2 == Ops[0]))
      PermuteMask[Elt] = BaseMaskElt;
    else if ((IsFirstOp && V1 == Ops[1]) || (!IsFirstOp && V2 == Ops[1]))
      PermuteMask[Elt] = BaseMaskElt + 1;
    assert(PermuteMask[Elt] != -1 &&
           "Input mask element is defined but failed to assign permute mask");
  }
 
  unsigned UnpckOp = MatchLo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
  SDValue Unpck = DAG.getNode(UnpckOp, DL, VT, Ops);
  return DAG.getVectorShuffle(VT, DL, Unpck, DAG.getUNDEF(VT), PermuteMask);
}
 
/// Try to lower a shuffle as a permute of the inputs followed by an
/// UNPCK instruction.
///
/// This specifically targets cases where we end up with alternating between
/// the two inputs, and so can permute them into something that feeds a single
/// UNPCK instruction. Note that this routine only targets integer vectors
/// because for floating point vectors we have a generalized SHUFPS lowering
/// strategy that handles everything that doesn't *exactly* match an unpack,
/// making this clever lowering unnecessary.
static SDValue lowerShuffleAsPermuteAndUnpack(const SDLoc &DL, MVT VT,
                                              SDValue V1, SDValue V2,
                                              ArrayRef<int> Mask,
                                              const X86Subtarget &Subtarget,
                                              SelectionDAG &DAG) {
  int Size = Mask.size();
  assert(Mask.size() >= 2 && "Single element masks are invalid.");
 
  // This routine only supports 128-bit integer dual input vectors.
  if (VT.isFloatingPoint() || !VT.is128BitVector() || V2.isUndef())
    return SDValue();
 
  int NumLoInputs =
      count_if(Mask, [Size](int M) { return M >= 0 && M % Size < Size / 2; });
  int NumHiInputs =
      count_if(Mask, [Size](int M) { return M % Size >= Size / 2; });
 
  bool UnpackLo = NumLoInputs >= NumHiInputs;
 
  auto TryUnpack = [&](int ScalarSize, int Scale) {
    SmallVector<int, 16> V1Mask((unsigned)Size, -1);
    SmallVector<int, 16> V2Mask((unsigned)Size, -1);
 
    for (int i = 0; i < Size; ++i) {
      if (Mask[i] < 0)
        continue;
 
      // Each element of the unpack contains Scale elements from this mask.
      int UnpackIdx = i / Scale;
 
      // We only handle the case where V1 feeds the first slots of the unpack.
      // We rely on canonicalization to ensure this is the case.
      if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
        return SDValue();
 
      // Setup the mask for this input. The indexing is tricky as we have to
      // handle the unpack stride.
      SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
      VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
          Mask[i] % Size;
    }
 
    // If we will have to shuffle both inputs to use the unpack, check whether
    // we can just unpack first and shuffle the result. If so, skip this unpack.
    if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
        !isNoopShuffleMask(V2Mask))
      return SDValue();
 
    // Shuffle the inputs into place.
    V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
    V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
 
    // Cast the inputs to the type we will use to unpack them.
    MVT UnpackVT =
        MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), Size / Scale);
    V1 = DAG.getBitcast(UnpackVT, V1);
    V2 = DAG.getBitcast(UnpackVT, V2);
 
    // Unpack the inputs and cast the result back to the desired type.
    return DAG.getBitcast(
        VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
                        UnpackVT, V1, V2));
  };
 
  // We try each unpack from the largest to the smallest to try and find one
  // that fits this mask.
  int OrigScalarSize = VT.getScalarSizeInBits();
  for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2)
    if (SDValue Unpack = TryUnpack(ScalarSize, ScalarSize / OrigScalarSize))
      return Unpack;
 
  // If we're shuffling with a zero vector then we're better off not doing
  // VECTOR_SHUFFLE(UNPCK()) as we lose track of those zero elements.
  if (ISD::isBuildVectorAllZeros(V1.getNode()) ||
      ISD::isBuildVectorAllZeros(V2.getNode()))
    return SDValue();
 
  // If none of the unpack-rooted lowerings worked (or were profitable) try an
  // initial unpack.
  if (NumLoInputs == 0 || NumHiInputs == 0) {
    assert((NumLoInputs > 0 || NumHiInputs > 0) &&
           "We have to have *some* inputs!");
    int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
 
    // FIXME: We could consider the total complexity of the permute of each
    // possible unpacking. Or at the least we should consider how many
    // half-crossings are created.
    // FIXME: We could consider commuting the unpacks.
 
    SmallVector<int, 32> PermMask((unsigned)Size, -1);
    for (int i = 0; i < Size; ++i) {
      if (Mask[i] < 0)
        continue;
 
      assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
 
      PermMask[i] =
          2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
    }
    return DAG.getVectorShuffle(
        VT, DL,
        DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL, DL, VT,
                    V1, V2),
        DAG.getUNDEF(VT), PermMask);
  }
 
  return SDValue();
}
 
/// Helper to form a PALIGNR-based rotate+permute, merging 2 inputs and then
/// permuting the elements of the result in place.
static SDValue lowerShuffleAsByteRotateAndPermute(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    const X86Subtarget &Subtarget, SelectionDAG &DAG) {
  if ((VT.is128BitVector() && !Subtarget.hasSSSE3()) ||
      (VT.is256BitVector() && !Subtarget.hasAVX2()) ||
      (VT.is512BitVector() && !Subtarget.hasBWI()))
    return SDValue();
 
  // We don't currently support lane crossing permutes.
  if (is128BitLaneCrossingShuffleMask(VT, Mask))
    return SDValue();
 
  int Scale = VT.getScalarSizeInBits() / 8;
  int NumLanes = VT.getSizeInBits() / 128;
  int NumElts = VT.getVectorNumElements();
  int NumEltsPerLane = NumElts / NumLanes;
 
  // Determine range of mask elts.
  bool Blend1 = true;
  bool Blend2 = true;
  std::pair<int, int> Range1 = std::make_pair(INT_MAX, INT_MIN);
  std::pair<int, int> Range2 = std::make_pair(INT_MAX, INT_MIN);
  for (int Lane = 0; Lane != NumElts; Lane += NumEltsPerLane) {
    for (int Elt = 0; Elt != NumEltsPerLane; ++Elt) {
      int M = Mask[Lane + Elt];
      if (M < 0)
        continue;
      if (M < NumElts) {
        Blend1 &= (M == (Lane + Elt));
        assert(Lane <= M && M < (Lane + NumEltsPerLane) && "Out of range mask");
        M = M % NumEltsPerLane;
        Range1.first = std::min(Range1.first, M);
        Range1.second = std::max(Range1.second, M);
      } else {
        M -= NumElts;
        Blend2 &= (M == (Lane + Elt));
        assert(Lane <= M && M < (Lane + NumEltsPerLane) && "Out of range mask");
        M = M % NumEltsPerLane;
        Range2.first = std::min(Range2.first, M);
        Range2.second = std::max(Range2.second, M);
      }
    }
  }
 
  // Bail if we don't need both elements.
  // TODO - it might be worth doing this for unary shuffles if the permute
  // can be widened.
  if (!(0 <= Range1.first && Range1.second < NumEltsPerLane) ||
      !(0 <= Range2.first && Range2.second < NumEltsPerLane))
    return SDValue();
 
  if (VT.getSizeInBits() > 128 && (Blend1 || Blend2))
    return SDValue();
 
  // Rotate the 2 ops so we can access both ranges, then permute the result.
  auto RotateAndPermute = [&](SDValue Lo, SDValue Hi, int RotAmt, int Ofs) {
    MVT ByteVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
    SDValue Rotate = DAG.getBitcast(
        VT, DAG.getNode(X86ISD::PALIGNR, DL, ByteVT, DAG.getBitcast(ByteVT, Hi),
                        DAG.getBitcast(ByteVT, Lo),
                        DAG.getTargetConstant(Scale * RotAmt, DL, MVT::i8)));
    SmallVector<int, 64> PermMask(NumElts, SM_SentinelUndef);
    for (int Lane = 0; Lane != NumElts; Lane += NumEltsPerLane) {
      for (int Elt = 0; Elt != NumEltsPerLane; ++Elt) {
        int M = Mask[Lane + Elt];
        if (M < 0)
          continue;
        if (M < NumElts)
          PermMask[Lane + Elt] = Lane + ((M + Ofs - RotAmt) % NumEltsPerLane);
        else
          PermMask[Lane + Elt] = Lane + ((M - Ofs - RotAmt) % NumEltsPerLane);
      }
    }
    return DAG.getVectorShuffle(VT, DL, Rotate, DAG.getUNDEF(VT), PermMask);
  };
 
  // Check if the ranges are small enough to rotate from either direction.
  if (Range2.second < Range1.first)
    return RotateAndPermute(V1, V2, Range1.first, 0);
  if (Range1.second < Range2.first)
    return RotateAndPermute(V2, V1, Range2.first, NumElts);
  return SDValue();
}
 
static bool isBroadcastShuffleMask(ArrayRef<int> Mask) {
  return isUndefOrEqual(Mask, 0);
}
 
static bool isNoopOrBroadcastShuffleMask(ArrayRef<int> Mask) {
  return isNoopShuffleMask(Mask) || isBroadcastShuffleMask(Mask);
}
 
/// Check if the Mask consists of the same element repeated multiple times.
static bool isSingleElementRepeatedMask(ArrayRef<int> Mask) {
  size_t NumUndefs = 0;
  std::optional<int> UniqueElt;
  for (int Elt : Mask) {
    if (Elt == SM_SentinelUndef) {
      NumUndefs++;
      continue;
    }
    if (UniqueElt.has_value() && UniqueElt.value() != Elt)
      return false;
    UniqueElt = Elt;
  }
  // Make sure the element is repeated enough times by checking the number of
  // undefs is small.
  return NumUndefs <= Mask.size() / 2 && UniqueElt.has_value();
}
 
/// Generic routine to decompose a shuffle and blend into independent
/// blends and permutes.
///
/// This matches the extremely common pattern for handling combined
/// shuffle+blend operations on newer X86 ISAs where we have very fast blend
/// operations. It will try to pick the best arrangement of shuffles and
/// blends. For vXi8/vXi16 shuffles we may use unpack instead of blend.
static SDValue lowerShuffleAsDecomposedShuffleMerge(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    const APInt &Zeroable, const X86Subtarget &Subtarget, SelectionDAG &DAG) {
  int NumElts = Mask.size();
  int NumLanes = VT.getSizeInBits() / 128;
  int NumEltsPerLane = NumElts / NumLanes;
 
  // Shuffle the input elements into the desired positions in V1 and V2 and
  // unpack/blend them together.
  bool IsAlternating = true;
  bool V1Zero = true, V2Zero = true;
  SmallVector<int, 32> V1Mask(NumElts, -1);
  SmallVector<int, 32> V2Mask(NumElts, -1);
  SmallVector<int, 32> FinalMask(NumElts, -1);
  for (int i = 0; i < NumElts; ++i) {
    int M = Mask[i];
    if (M >= 0 && M < NumElts) {
      V1Mask[i] = M;
      FinalMask[i] = i;
      V1Zero &= Zeroable[i];
      IsAlternating &= (i & 1) == 0;
    } else if (M >= NumElts) {
      V2Mask[i] = M - NumElts;
      FinalMask[i] = i + NumElts;
      V2Zero &= Zeroable[i];
      IsAlternating &= (i & 1) == 1;
    }
  }
 
  // If we effectively only demand the 0'th element of \p Input, and not only
  // as 0'th element, then broadcast said input,
  // and change \p InputMask to be a no-op (identity) mask.
  auto canonicalizeBroadcastableInput = [DL, VT, &Subtarget,
                                         &DAG](SDValue &Input,
                                               MutableArrayRef<int> InputMask) {
    unsigned EltSizeInBits = Input.getScalarValueSizeInBits();
    if (!Subtarget.hasAVX2() && (!Subtarget.hasAVX() || EltSizeInBits < 32 ||
                                 !X86::mayFoldLoad(Input, Subtarget)))
      return;
    if (isNoopShuffleMask(InputMask))
      return;
    assert(isBroadcastShuffleMask(InputMask) &&
           "Expected to demand only the 0'th element.");
    Input = DAG.getNode(X86ISD::VBROADCAST, DL, VT, Input);
    for (auto I : enumerate(InputMask)) {
      int &InputMaskElt = I.value();
      if (InputMaskElt >= 0)
        InputMaskElt = I.index();
    }
  };
 
  // Currently, we may need to produce one shuffle per input, and blend results.
  // It is possible that the shuffle for one of the inputs is already a no-op.
  // See if we can simplify non-no-op shuffles into broadcasts,
  // which we consider to be strictly better than an arbitrary shuffle.
  if (isNoopOrBroadcastShuffleMask(V1Mask) &&
      isNoopOrBroadcastShuffleMask(V2Mask)) {
    canonicalizeBroadcastableInput(V1, V1Mask);
    canonicalizeBroadcastableInput(V2, V2Mask);
  }
 
  // Try to lower with the simpler initial blend/unpack/rotate strategies unless
  // one of the input shuffles would be a no-op. We prefer to shuffle inputs as
  // the shuffle may be able to fold with a load or other benefit. However, when
  // we'll have to do 2x as many shuffles in order to achieve this, a 2-input
  // pre-shuffle first is a better strategy.
  if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask)) {
    // If we don't have blends, see if we can create a cheap unpack.
    if (!Subtarget.hasSSE41() && VT.is128BitVector() &&
        (is128BitUnpackShuffleMask(V1Mask, DAG) ||
         is128BitUnpackShuffleMask(V2Mask, DAG)))
      if (SDValue PermUnpack = lowerShuffleAsPermuteAndUnpack(
              DL, VT, V1, V2, Mask, Subtarget, DAG))
        return PermUnpack;
 
    // Only prefer immediate blends to unpack/rotate.
    if (SDValue BlendPerm =
            lowerShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG, true))
      return BlendPerm;
 
    // If either input vector provides only a single element which is repeated
    // multiple times, unpacking from both input vectors would generate worse
    // code. e.g. for
    // t5: v16i8 = vector_shuffle<16,0,16,1,16,2,16,3,16,4,16,5,16,6,16,7> t2, t4
    // it is better to process t4 first to create a vector of t4[0], then unpack
    // that vector with t2.
    if (!V1Zero && !V2Zero && !isSingleElementRepeatedMask(V1Mask) &&
        !isSingleElementRepeatedMask(V2Mask))
      if (SDValue UnpackPerm =
              lowerShuffleAsUNPCKAndPermute(DL, VT, V1, V2, Mask, DAG))
        return UnpackPerm;
 
    if (SDValue RotatePerm = lowerShuffleAsByteRotateAndPermute(
            DL, VT, V1, V2, Mask, Subtarget, DAG))
      return RotatePerm;
 
    // Unpack/rotate failed - try again with variable blends.
    if (SDValue BlendPerm = lowerShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask,
                                                          DAG))
      return BlendPerm;
 
    if (VT.getScalarSizeInBits() >= 32)
      if (SDValue PermUnpack = lowerShuffleAsPermuteAndUnpack(
              DL, VT, V1, V2, Mask, Subtarget, DAG))
        return PermUnpack;
  }
 
  // If the final mask is an alternating blend of vXi8/vXi16, convert to an
  // UNPCKL(SHUFFLE, SHUFFLE) pattern.
  // TODO: It doesn't have to be alternating - but each lane mustn't have more
  // than half the elements coming from each source.
  if (IsAlternating && VT.getScalarSizeInBits() < 32) {
    V1Mask.assign(NumElts, -1);
    V2Mask.assign(NumElts, -1);
    FinalMask.assign(NumElts, -1);
    for (int i = 0; i != NumElts; i += NumEltsPerLane)
      for (int j = 0; j != NumEltsPerLane; ++j) {
        int M = Mask[i + j];
        if (M >= 0 && M < NumElts) {
          V1Mask[i + (j / 2)] = M;
          FinalMask[i + j] = i + (j / 2);
        } else if (M >= NumElts) {
          V2Mask[i + (j / 2)] = M - NumElts;
          FinalMask[i + j] = i + (j / 2) + NumElts;
        }
      }
  }
 
  V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
  V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
  return DAG.getVectorShuffle(VT, DL, V1, V2, FinalMask);
}
 
static int matchShuffleAsBitRotate(MVT &RotateVT, int EltSizeInBits,
                                   const X86Subtarget &Subtarget,
                                   ArrayRef<int> Mask) {
  assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
  assert(EltSizeInBits < 64 && "Can't rotate 64-bit integers");
 
  // AVX512 only has vXi32/vXi64 rotates, so limit the rotation sub group size.
  int MinSubElts = Subtarget.hasAVX512() ? std::max(32 / EltSizeInBits, 2) : 2;
  int MaxSubElts = 64 / EltSizeInBits;
  unsigned RotateAmt, NumSubElts;
  if (!ShuffleVectorInst::isBitRotateMask(Mask, EltSizeInBits, MinSubElts,
                                          MaxSubElts, NumSubElts, RotateAmt))
    return -1;
  unsigned NumElts = Mask.size();
  MVT RotateSVT = MVT::getIntegerVT(EltSizeInBits * NumSubElts);
  RotateVT = MVT::getVectorVT(RotateSVT, NumElts / NumSubElts);
  return RotateAmt;
}
 
/// Lower shuffle using X86ISD::VROTLI rotations.
static SDValue lowerShuffleAsBitRotate(const SDLoc &DL, MVT VT, SDValue V1,
                                       ArrayRef<int> Mask,
                                       const X86Subtarget &Subtarget,
                                       SelectionDAG &DAG) {
  // Only XOP + AVX512 targets have bit rotation instructions.
  // If we at least have SSSE3 (PSHUFB) then we shouldn't attempt to use this.
  bool IsLegal =
      (VT.is128BitVector() && Subtarget.hasXOP()) || Subtarget.hasAVX512();
  if (!IsLegal && Subtarget.hasSSE3())
    return SDValue();
 
  MVT RotateVT;
  int RotateAmt = matchShuffleAsBitRotate(RotateVT, VT.getScalarSizeInBits(),
                                          Subtarget, Mask);
  if (RotateAmt < 0)
    return SDValue();
 
  // For pre-SSSE3 targets, if we are shuffling vXi8 elts then ISD::ROTL,
  // expanded to OR(SRL,SHL), will be more efficient, but if they can
  // widen to vXi16 or more then existing lowering should will be better.
  if (!IsLegal) {
    if ((RotateAmt % 16) == 0)
      return SDValue();
    unsigned ShlAmt = RotateAmt;
    unsigned SrlAmt = RotateVT.getScalarSizeInBits() - RotateAmt;
    SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, DL, RotateVT, V1,
                                             ShlAmt, DAG);
    SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, DL, RotateVT, V1,
                                             SrlAmt, DAG);
    SDValue Rot = DAG.getNode(ISD::OR, DL, RotateVT, SHL, SRL);
    return DAG.getBitcast(VT, Rot);
  }
 
  SDValue Rot =
      DAG.getNode(X86ISD::VROTLI, DL, RotateVT, DAG.getBitcast(RotateVT, V1),
                  DAG.getTargetConstant(RotateAmt, DL, MVT::i8));
  return DAG.getBitcast(VT, Rot);
}
 
/// Try to match a vector shuffle as an element rotation.
///
/// This is used for support PALIGNR for SSSE3 or VALIGND/Q for AVX512.
static int matchShuffleAsElementRotate(SDValue &V1, SDValue &V2,
                                       ArrayRef<int> Mask) {
  int NumElts = Mask.size();
 
  // We need to detect various ways of spelling a rotation:
  //   [11, 12, 13, 14, 15,  0,  1,  2]
  //   [-1, 12, 13, 14, -1, -1,  1, -1]
  //   [-1, -1, -1, -1, -1, -1,  1,  2]
  //   [ 3,  4,  5,  6,  7,  8,  9, 10]
  //   [-1,  4,  5,  6, -1, -1,  9, -1]
  //   [-1,  4,  5,  6, -1, -1, -1, -1]
  int Rotation = 0;
  SDValue Lo, Hi;
  for (int i = 0; i < NumElts; ++i) {
    int M = Mask[i];
    assert((M == SM_SentinelUndef || (0 <= M && M < (2*NumElts))) &&
           "Unexpected mask index.");
    if (M < 0)
      continue;
 
    // Determine where a rotated vector would have started.
    int StartIdx = i - (M % NumElts);
    if (StartIdx == 0)
      // The identity rotation isn't interesting, stop.
      return -1;
 
    // If we found the tail of a vector the rotation must be the missing
    // front. If we found the head of a vector, it must be how much of the
    // head.
    int CandidateRotation = StartIdx < 0 ? -StartIdx : NumElts - StartIdx;
 
    if (Rotation == 0)
      Rotation = CandidateRotation;
    else if (Rotation != CandidateRotation)
      // The rotations don't match, so we can't match this mask.
      return -1;
 
    // Compute which value this mask is pointing at.
    SDValue MaskV = M < NumElts ? V1 : V2;
 
    // Compute which of the two target values this index should be assigned
    // to. This reflects whether the high elements are remaining or the low
    // elements are remaining.
    SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
 
    // Either set up this value if we've not encountered it before, or check
    // that it remains consistent.
    if (!TargetV)
      TargetV = MaskV;
    else if (TargetV != MaskV)
      // This may be a rotation, but it pulls from the inputs in some
      // unsupported interleaving.
      return -1;
  }
 
  // Check that we successfully analyzed the mask, and normalize the results.
  assert(Rotation != 0 && "Failed to locate a viable rotation!");
  assert((Lo || Hi) && "Failed to find a rotated input vector!");
  if (!Lo)
    Lo = Hi;
  else if (!Hi)
    Hi = Lo;
 
  V1 = Lo;
  V2 = Hi;
 
  return Rotation;
}
 
/// Try to lower a vector shuffle as a byte rotation.
///
/// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
/// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
/// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
/// try to generically lower a vector shuffle through such an pattern. It
/// does not check for the profitability of lowering either as PALIGNR or
/// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
/// This matches shuffle vectors that look like:
///
///   v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
///
/// Essentially it concatenates V1 and V2, shifts right by some number of
/// elements, and takes the low elements as the result. Note that while this is
/// specified as a *right shift* because x86 is little-endian, it is a *left
/// rotate* of the vector lanes.
static int matchShuffleAsByteRotate(MVT VT, SDValue &V1, SDValue &V2,
                                    ArrayRef<int> Mask) {
  // Don't accept any shuffles with zero elements.
  if (isAnyZero(Mask))
    return -1;
 
  // PALIGNR works on 128-bit lanes.
  SmallVector<int, 16> RepeatedMask;
  if (!is128BitLaneRepeatedShuffleMask(VT, Mask, RepeatedMask))
    return -1;
 
  int Rotation = matchShuffleAsElementRotate(V1, V2, RepeatedMask);
  if (Rotation <= 0)
    return -1;
 
  // PALIGNR rotates bytes, so we need to scale the
  // rotation based on how many bytes are in the vector lane.
  int NumElts = RepeatedMask.size();
  int Scale = 16 / NumElts;
  return Rotation * Scale;
}
 
static SDValue lowerShuffleAsByteRotate(const SDLoc &DL, MVT VT, SDValue V1,
                                        SDValue V2, ArrayRef<int> Mask,
                                        const X86Subtarget &Subtarget,
                                        SelectionDAG &DAG) {
  assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
 
  SDValue Lo = V1, Hi = V2;
  int ByteRotation = matchShuffleAsByteRotate(VT, Lo, Hi, Mask);
  if (ByteRotation <= 0)
    return SDValue();
 
  // Cast the inputs to i8 vector of correct length to match PALIGNR or
  // PSLLDQ/PSRLDQ.
  MVT ByteVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
  Lo = DAG.getBitcast(ByteVT, Lo);
  Hi = DAG.getBitcast(ByteVT, Hi);
 
  // SSSE3 targets can use the palignr instruction.
  if (Subtarget.hasSSSE3()) {
    assert((!VT.is512BitVector() || Subtarget.hasBWI()) &&
           "512-bit PALIGNR requires BWI instructions");
    return DAG.getBitcast(
        VT, DAG.getNode(X86ISD::PALIGNR, DL, ByteVT, Lo, Hi,
                        DAG.getTargetConstant(ByteRotation, DL, MVT::i8)));
  }
 
  assert(VT.is128BitVector() &&
         "Rotate-based lowering only supports 128-bit lowering!");
  assert(Mask.size() <= 16 &&
         "Can shuffle at most 16 bytes in a 128-bit vector!");
  assert(ByteVT == MVT::v16i8 &&
         "SSE2 rotate lowering only needed for v16i8!");
 
  // Default SSE2 implementation
  int LoByteShift = 16 - ByteRotation;
  int HiByteShift = ByteRotation;
 
  SDValue LoShift =
      DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Lo,
                  DAG.getTargetConstant(LoByteShift, DL, MVT::i8));
  SDValue HiShift =
      DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Hi,
                  DAG.getTargetConstant(HiByteShift, DL, MVT::i8));
  return DAG.getBitcast(VT,
                        DAG.getNode(ISD::OR, DL, MVT::v16i8, LoShift, HiShift));
}
 
/// Try to lower a vector shuffle as a dword/qword rotation.
///
/// AVX512 has a VALIGND/VALIGNQ instructions that will do an arbitrary
/// rotation of the concatenation of two vectors; This routine will
/// try to generically lower a vector shuffle through such an pattern.
///
/// Essentially it concatenates V1 and V2, shifts right by some number of
/// elements, and takes the low elements as the result. Note that while this is
/// specified as a *right shift* because x86 is little-endian, it is a *left
/// rotate* of the vector lanes.
static SDValue lowerShuffleAsVALIGN(const SDLoc &DL, MVT VT, SDValue V1,
                                    SDValue V2, ArrayRef<int> Mask,
                                    const APInt &Zeroable,
                                    const X86Subtarget &Subtarget,
                                    SelectionDAG &DAG) {
  assert((VT.getScalarType() == MVT::i32 || VT.getScalarType() == MVT::i64) &&
         "Only 32-bit and 64-bit elements are supported!");
 
  // 128/256-bit vectors are only supported with VLX.
  assert((Subtarget.hasVLX() || (!VT.is128BitVector() && !VT.is256BitVector()))
         && "VLX required for 128/256-bit vectors");
 
  SDValue Lo = V1, Hi = V2;
  int Rotation = matchShuffleAsElementRotate(Lo, Hi, Mask);
  if (0 < Rotation)
    return DAG.getNode(X86ISD::VALIGN, DL, VT, Lo, Hi,
                       DAG.getTargetConstant(Rotation, DL, MVT::i8));
 
  // See if we can use VALIGN as a cross-lane version of VSHLDQ/VSRLDQ.
  // TODO: Pull this out as a matchShuffleAsElementShift helper?
  // TODO: We can probably make this more aggressive and use shift-pairs like
  // lowerShuffleAsByteShiftMask.
  unsigned NumElts = Mask.size();
  unsigned ZeroLo = Zeroable.countr_one();
  unsigned ZeroHi = Zeroable.countl_one();
  assert((ZeroLo + ZeroHi) < NumElts && "Zeroable shuffle detected");
  if (!ZeroLo && !ZeroHi)
    return SDValue();
 
  if (ZeroLo) {
    SDValue Src = Mask[ZeroLo] < (int)NumElts ? V1 : V2;
    int Low = Mask[ZeroLo] < (int)NumElts ? 0 : NumElts;
    if (isSequentialOrUndefInRange(Mask, ZeroLo, NumElts - ZeroLo, Low))
      return DAG.getNode(X86ISD::VALIGN, DL, VT, Src,
                         getZeroVector(VT, Subtarget, DAG, DL),
                         DAG.getTargetConstant(NumElts - ZeroLo, DL, MVT::i8));
  }
 
  if (ZeroHi) {
    SDValue Src = Mask[0] < (int)NumElts ? V1 : V2;
    int Low = Mask[0] < (int)NumElts ? 0 : NumElts;
    if (isSequentialOrUndefInRange(Mask, 0, NumElts - ZeroHi, Low + ZeroHi))
      return DAG.getNode(X86ISD::VALIGN, DL, VT,
                         getZeroVector(VT, Subtarget, DAG, DL), Src,
                         DAG.getTargetConstant(ZeroHi, DL, MVT::i8));
  }
 
  return SDValue();
}
 
/// Try to lower a vector shuffle as a byte shift sequence.
static SDValue lowerShuffleAsByteShiftMask(const SDLoc &DL, MVT VT, SDValue V1,
                                           SDValue V2, ArrayRef<int> Mask,
                                           const APInt &Zeroable,
                                           const X86Subtarget &Subtarget,
                                           SelectionDAG &DAG) {
  assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
  assert(VT.is128BitVector() && "Only 128-bit vectors supported");
 
  // We need a shuffle that has zeros at one/both ends and a sequential
  // shuffle from one source within.
  unsigned ZeroLo = Zeroable.countr_one();
  unsigned ZeroHi = Zeroable.countl_one();
  if (!ZeroLo && !ZeroHi)
    return SDValue();
 
  unsigned NumElts = Mask.size();
  unsigned Len = NumElts - (ZeroLo + ZeroHi);
  if (!isSequentialOrUndefInRange(Mask, ZeroLo, Len, Mask[ZeroLo]))
    return SDValue();
 
  unsigned Scale = VT.getScalarSizeInBits() / 8;
  ArrayRef<int> StubMask = Mask.slice(ZeroLo, Len);
  if (!isUndefOrInRange(StubMask, 0, NumElts) &&
      !isUndefOrInRange(StubMask, NumElts, 2 * NumElts))
    return SDValue();
 
  SDValue Res = Mask[ZeroLo] < (int)NumElts ? V1 : V2;
  Res = DAG.getBitcast(MVT::v16i8, Res);
 
  // Use VSHLDQ/VSRLDQ ops to zero the ends of a vector and leave an
  // inner sequential set of elements, possibly offset:
  // 01234567 --> zzzzzz01 --> 1zzzzzzz
  // 01234567 --> 4567zzzz --> zzzzz456
  // 01234567 --> z0123456 --> 3456zzzz --> zz3456zz
  if (ZeroLo == 0) {
    unsigned Shift = (NumElts - 1) - (Mask[ZeroLo + Len - 1] % NumElts);
    Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res,
                      DAG.getTargetConstant(Scale * Shift, DL, MVT::i8));
    Res = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Res,
                      DAG.getTargetConstant(Scale * ZeroHi, DL, MVT::i8));
  } else if (ZeroHi == 0) {
    unsigned Shift = Mask[ZeroLo] % NumElts;
    Res = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Res,
                      DAG.getTargetConstant(Scale * Shift, DL, MVT::i8));
    Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res,
                      DAG.getTargetConstant(Scale * ZeroLo, DL, MVT::i8));
  } else if (!Subtarget.hasSSSE3()) {
    // If we don't have PSHUFB then its worth avoiding an AND constant mask
    // by performing 3 byte shifts. Shuffle combining can kick in above that.
    // TODO: There may be some cases where VSH{LR}DQ+PAND is still better.
    unsigned Shift = (NumElts - 1) - (Mask[ZeroLo + Len - 1] % NumElts);
    Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res,
                      DAG.getTargetConstant(Scale * Shift, DL, MVT::i8));
    Shift += Mask[ZeroLo] % NumElts;
    Res = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v16i8, Res,
                      DAG.getTargetConstant(Scale * Shift, DL, MVT::i8));
    Res = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v16i8, Res,
                      DAG.getTargetConstant(Scale * ZeroLo, DL, MVT::i8));
  } else
    return SDValue();
 
  return DAG.getBitcast(VT, Res);
}
 
/// Try to lower a vector shuffle as a bit shift (shifts in zeros).
///
/// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
/// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
/// matches elements from one of the input vectors shuffled to the left or
/// right with zeroable elements 'shifted in'. It handles both the strictly
/// bit-wise element shifts and the byte shift across an entire 128-bit double
/// quad word lane.
///
/// PSHL : (little-endian) left bit shift.
/// [ zz, 0, zz,  2 ]
/// [ -1, 4, zz, -1 ]
/// PSRL : (little-endian) right bit shift.
/// [  1, zz,  3, zz]
/// [ -1, -1,  7, zz]
/// PSLLDQ : (little-endian) left byte shift
/// [ zz,  0,  1,  2,  3,  4,  5,  6]
/// [ zz, zz, -1, -1,  2,  3,  4, -1]
/// [ zz, zz, zz, zz, zz, zz, -1,  1]
/// PSRLDQ : (little-endian) right byte shift
/// [  5, 6,  7, zz, zz, zz, zz, zz]
/// [ -1, 5,  6,  7, zz, zz, zz, zz]
/// [  1, 2, -1, -1, -1, -1, zz, zz]
static int matchShuffleAsShift(MVT &ShiftVT, unsigned &Opcode,
                               unsigned ScalarSizeInBits, ArrayRef<int> Mask,
                               int MaskOffset, const APInt &Zeroable,
                               const X86Subtarget &Subtarget) {
  int Size = Mask.size();
  unsigned SizeInBits = Size * ScalarSizeInBits;
 
  auto CheckZeros = [&](int Shift, int Scale, bool Left) {
    for (int i = 0; i < Size; i += Scale)
      for (int j = 0; j < Shift; ++j)
        if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
          return false;
 
    return true;
  };
 
  auto MatchShift = [&](int Shift, int Scale, bool Left) {
    for (int i = 0; i != Size; i += Scale) {
      unsigned Pos = Left ? i + Shift : i;
      unsigned Low = Left ? i : i + Shift;
      unsigned Len = Scale - Shift;
      if (!isSequentialOrUndefInRange(Mask, Pos, Len, Low + MaskOffset))
        return -1;
    }
 
    int ShiftEltBits = ScalarSizeInBits * Scale;
    bool ByteShift = ShiftEltBits > 64;
    Opcode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
                  : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
    int ShiftAmt = Shift * ScalarSizeInBits / (ByteShift ? 8 : 1);
 
    // Normalize the scale for byte shifts to still produce an i64 element
    // type.
    Scale = ByteShift ? Scale / 2 : Scale;
 
    // We need to round trip through the appropriate type for the shift.
    MVT ShiftSVT = MVT::getIntegerVT(ScalarSizeInBits * Scale);
    ShiftVT = ByteShift ? MVT::getVectorVT(MVT::i8, SizeInBits / 8)
                        : MVT::getVectorVT(ShiftSVT, Size / Scale);
    return ShiftAmt;
  };
 
  // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
  // keep doubling the size of the integer elements up to that. We can
  // then shift the elements of the integer vector by whole multiples of
  // their width within the elements of the larger integer vector. Test each
  // multiple to see if we can find a match with the moved element indices
  // and that the shifted in elements are all zeroable.
  unsigned MaxWidth = ((SizeInBits == 512) && !Subtarget.hasBWI() ? 64 : 128);
  for (int Scale = 2; Scale * ScalarSizeInBits <= MaxWidth; Scale *= 2)
    for (int Shift = 1; Shift != Scale; ++Shift)
      for (bool Left : {true, false})
        if (CheckZeros(Shift, Scale, Left)) {
          int ShiftAmt = MatchShift(Shift, Scale, Left);
          if (0 < ShiftAmt)
            return ShiftAmt;
        }
 
  // no match
  return -1;
}
 
static SDValue lowerShuffleAsShift(const SDLoc &DL, MVT VT, SDValue V1,
                                   SDValue V2, ArrayRef<int> Mask,
                                   const APInt &Zeroable,
                                   const X86Subtarget &Subtarget,
                                   SelectionDAG &DAG, bool BitwiseOnly) {
  int Size = Mask.size();
  assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
 
  MVT ShiftVT;
  SDValue V = V1;
  unsigned Opcode;
 
  // Try to match shuffle against V1 shift.
  int ShiftAmt = matchShuffleAsShift(ShiftVT, Opcode, VT.getScalarSizeInBits(),
                                     Mask, 0, Zeroable, Subtarget);
 
  // If V1 failed, try to match shuffle against V2 shift.
  if (ShiftAmt < 0) {
    ShiftAmt = matchShuffleAsShift(ShiftVT, Opcode, VT.getScalarSizeInBits(),
                                   Mask, Size, Zeroable, Subtarget);
    V = V2;
  }
 
  if (ShiftAmt < 0)
    return SDValue();
 
  if (BitwiseOnly && (Opcode == X86ISD::VSHLDQ || Opcode == X86ISD::VSRLDQ))
    return SDValue();
 
  assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
         "Illegal integer vector type");
  V = DAG.getBitcast(ShiftVT, V);
  V = DAG.getNode(Opcode, DL, ShiftVT, V,
                  DAG.getTargetConstant(ShiftAmt, DL, MVT::i8));
  return DAG.getBitcast(VT, V);
}
 
// EXTRQ: Extract Len elements from lower half of source, starting at Idx.
// Remainder of lower half result is zero and upper half is all undef.
static bool matchShuffleAsEXTRQ(MVT VT, SDValue &V1, SDValue &V2,
                                ArrayRef<int> Mask, uint64_t &BitLen,
                                uint64_t &BitIdx, const APInt &Zeroable) {
  int Size = Mask.size();
  int HalfSize = Size / 2;
  assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
  assert(!Zeroable.isAllOnes() && "Fully zeroable shuffle mask");
 
  // Upper half must be undefined.
  if (!isUndefUpperHalf(Mask))
    return false;
 
  // Determine the extraction length from the part of the
  // lower half that isn't zeroable.
  int Len = HalfSize;
  for (; Len > 0; --Len)
    if (!Zeroable[Len - 1])
      break;
  assert(Len > 0 && "Zeroable shuffle mask");
 
  // Attempt to match first Len sequential elements from the lower half.
  SDValue Src;
  int Idx = -1;
  for (int i = 0; i != Len; ++i) {
    int M = Mask[i];
    if (M == SM_SentinelUndef)
      continue;
    SDValue &V = (M < Size ? V1 : V2);
    M = M % Size;
 
    // The extracted elements must start at a valid index and all mask
    // elements must be in the lower half.
    if (i > M || M >= HalfSize)
      return false;
 
    if (Idx < 0 || (Src == V && Idx == (M - i))) {
      Src = V;
      Idx = M - i;
      continue;
    }
    return false;
  }
 
  if (!Src || Idx < 0)
    return false;
 
  assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
  BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
  BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
  V1 = Src;
  return true;
}
 
// INSERTQ: Extract lowest Len elements from lower half of second source and
// insert over first source, starting at Idx.
// { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
static bool matchShuffleAsINSERTQ(MVT VT, SDValue &V1, SDValue &V2,
                                  ArrayRef<int> Mask, uint64_t &BitLen,
                                  uint64_t &BitIdx) {
  int Size = Mask.size();
  int HalfSize = Size / 2;
  assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
 
  // Upper half must be undefined.
  if (!isUndefUpperHalf(Mask))
    return false;
 
  for (int Idx = 0; Idx != HalfSize; ++Idx) {
    SDValue Base;
 
    // Attempt to match first source from mask before insertion point.
    if (isUndefInRange(Mask, 0, Idx)) {
      /* EMPTY */
    } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
      Base = V1;
    } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
      Base = V2;
    } else {
      continue;
    }
 
    // Extend the extraction length looking to match both the insertion of
    // the second source and the remaining elements of the first.
    for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
      SDValue Insert;
      int Len = Hi - Idx;
 
      // Match insertion.
      if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
        Insert = V1;
      } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
        Insert = V2;
      } else {
        continue;
      }
 
      // Match the remaining elements of the lower half.
      if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
        /* EMPTY */
      } else if ((!Base || (Base == V1)) &&
                 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
        Base = V1;
      } else if ((!Base || (Base == V2)) &&
                 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
                                            Size + Hi)) {
        Base = V2;
      } else {
        continue;
      }
 
      BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
      BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
      V1 = Base;
      V2 = Insert;
      return true;
    }
  }
 
  return false;
}
 
/// Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
static SDValue lowerShuffleWithSSE4A(const SDLoc &DL, MVT VT, SDValue V1,
                                     SDValue V2, ArrayRef<int> Mask,
                                     const APInt &Zeroable, SelectionDAG &DAG) {
  uint64_t BitLen, BitIdx;
  if (matchShuffleAsEXTRQ(VT, V1, V2, Mask, BitLen, BitIdx, Zeroable))
    return DAG.getNode(X86ISD::EXTRQI, DL, VT, V1,
                       DAG.getTargetConstant(BitLen, DL, MVT::i8),
                       DAG.getTargetConstant(BitIdx, DL, MVT::i8));
 
  if (matchShuffleAsINSERTQ(VT, V1, V2, Mask, BitLen, BitIdx))
    return DAG.getNode(X86ISD::INSERTQI, DL, VT, V1 ? V1 : DAG.getUNDEF(VT),
                       V2 ? V2 : DAG.getUNDEF(VT),
                       DAG.getTargetConstant(BitLen, DL, MVT::i8),
                       DAG.getTargetConstant(BitIdx, DL, MVT::i8));
 
  return SDValue();
}
 
/// Lower a vector shuffle as an any/signed/zero extension.
///
/// Given a specific number of elements, element bit width, and extension
/// stride, produce either an extension based on the available
/// features of the subtarget. The extended elements are consecutive and
/// begin and can start from an offsetted element index in the input; to
/// avoid excess shuffling the offset must either being in the bottom lane
/// or at the start of a higher lane. All extended elements must be from
/// the same lane.
static SDValue lowerShuffleAsSpecificExtension(const SDLoc &DL, MVT VT,
                                               int Scale, int Offset,
                                               unsigned ExtOpc, SDValue InputV,
                                               ArrayRef<int> Mask,
                                               const X86Subtarget &Subtarget,
                                               SelectionDAG &DAG) {
  assert(Scale > 1 && "Need a scale to extend.");
  assert(ISD::isExtOpcode(ExtOpc) && "Unsupported extension");
  int EltBits = VT.getScalarSizeInBits();
  int NumElements = VT.getVectorNumElements();
  int NumEltsPerLane = 128 / EltBits;
  int OffsetLane = Offset / NumEltsPerLane;
  assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
         "Only 8, 16, and 32 bit elements can be extended.");
  assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
  assert(0 <= Offset && "Extension offset must be positive.");
  assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
         "Extension offset must be in the first lane or start an upper lane.");
 
  // Check that an index is in same lane as the base offset.
  auto SafeOffset = [&](int Idx) {
    return OffsetLane == (Idx / NumEltsPerLane);
  };
 
  // Shift along an input so that the offset base moves to the first element.
  auto ShuffleOffset = [&](SDValue V) {
    if (!Offset)
      return V;
 
    SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
    for (int i = 0; i * Scale < NumElements; ++i) {
      int SrcIdx = i + Offset;
      ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
    }
    return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
  };
 
  // Found a valid a/zext mask! Try various lowering strategies based on the
  // input type and available ISA extensions.
  if (Subtarget.hasSSE41()) {
    // Not worth offsetting 128-bit vectors if scale == 2, a pattern using
    // PUNPCK will catch this in a later shuffle match.
    if (Offset && Scale == 2 && VT.is128BitVector())
      return SDValue();
    MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
                                 NumElements / Scale);
    InputV = DAG.getBitcast(VT, InputV);
    InputV = ShuffleOffset(InputV);
    InputV = getEXTEND_VECTOR_INREG(ExtOpc, DL, ExtVT, InputV, DAG);
    return DAG.getBitcast(VT, InputV);
  }
 
  assert(VT.is128BitVector() && "Only 128-bit vectors can be extended.");
  InputV = DAG.getBitcast(VT, InputV);
  bool AnyExt = ExtOpc == ISD::ANY_EXTEND;
 
  // TODO: Add pre-SSE41 SIGN_EXTEND_VECTOR_INREG handling.
  if (ExtOpc == ISD::SIGN_EXTEND)
    return SDValue();
 
  // For any extends we can cheat for larger element sizes and use shuffle
  // instructions that can fold with a load and/or copy.
  if (AnyExt && EltBits == 32) {
    int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
                         -1};
    return DAG.getBitcast(
        VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
                        DAG.getBitcast(MVT::v4i32, InputV),
                        getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
  }
  if (AnyExt && EltBits == 16 && Scale > 2) {
    int PSHUFDMask[4] = {Offset / 2, -1,
                         SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
    InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
                         DAG.getBitcast(MVT::v4i32, InputV),
                         getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
    int PSHUFWMask[4] = {1, -1, -1, -1};
    unsigned OddEvenOp = (Offset & 1) ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
    return DAG.getBitcast(
        VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
                        DAG.getBitcast(MVT::v8i16, InputV),
                        getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
  }
 
  // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
  // to 64-bits.
  if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget.hasSSE4A()) {
    assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
    assert(VT.is128BitVector() && "Unexpected vector width!");
 
    int LoIdx = Offset * EltBits;
    SDValue Lo = DAG.getBitcast(
        MVT::v2i64, DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
                                DAG.getTargetConstant(EltBits, DL, MVT::i8),
                                DAG.getTargetConstant(LoIdx, DL, MVT::i8)));
 
    if (isUndefUpperHalf(Mask) || !SafeOffset(Offset + 1))
      return DAG.getBitcast(VT, Lo);
 
    int HiIdx = (Offset + 1) * EltBits;
    SDValue Hi = DAG.getBitcast(
        MVT::v2i64, DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
                                DAG.getTargetConstant(EltBits, DL, MVT::i8),
                                DAG.getTargetConstant(HiIdx, DL, MVT::i8)));
    return DAG.getBitcast(VT,
                          DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
  }
 
  // If this would require more than 2 unpack instructions to expand, use
  // pshufb when available. We can only use more than 2 unpack instructions
  // when zero extending i8 elements which also makes it easier to use pshufb.
  if (Scale > 4 && EltBits == 8 && Subtarget.hasSSSE3()) {
    assert(NumElements == 16 && "Unexpected byte vector width!");
    SDValue PSHUFBMask[16];
    for (int i = 0; i < 16; ++i) {
      int Idx = Offset + (i / Scale);
      if ((i % Scale == 0 && SafeOffset(Idx))) {
        PSHUFBMask[i] = DAG.getConstant(Idx, DL, MVT::i8);
        continue;
      }
      PSHUFBMask[i] =
          AnyExt ? DAG.getUNDEF(MVT::i8) : DAG.getConstant(0x80, DL, MVT::i8);
    }
    InputV = DAG.getBitcast(MVT::v16i8, InputV);
    return DAG.getBitcast(
        VT, DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
                        DAG.getBuildVector(MVT::v16i8, DL, PSHUFBMask)));
  }
 
  // If we are extending from an offset, ensure we start on a boundary that
  // we can unpack from.
  int AlignToUnpack = Offset % (NumElements / Scale);
  if (AlignToUnpack) {
    SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
    for (int i = AlignToUnpack; i < NumElements; ++i)
      ShMask[i - AlignToUnpack] = i;
    InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
    Offset -= AlignToUnpack;
  }
 
  // Otherwise emit a sequence of unpacks.
  do {
    unsigned UnpackLoHi = X86ISD::UNPCKL;
    if (Offset >= (NumElements / 2)) {
      UnpackLoHi = X86ISD::UNPCKH;
      Offset -= (NumElements / 2);
    }
 
    MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
    SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
                         : getZeroVector(InputVT, Subtarget, DAG, DL);
    InputV = DAG.getBitcast(InputVT, InputV);
    InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
    Scale /= 2;
    EltBits *= 2;
    NumElements /= 2;
  } while (Scale > 1);
  return DAG.getBitcast(VT, InputV);
}
 
/// Try to lower a vector shuffle as a zero extension on any microarch.
///
/// This routine will try to do everything in its power to cleverly lower
/// a shuffle which happens to match the pattern of a zero extend. It doesn't
/// check for the profitability of this lowering,  it tries to aggressively
/// match this pattern. It will use all of the micro-architectural details it
/// can to emit an efficient lowering. It handles both blends with all-zero
/// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
/// masking out later).
///
/// The reason we have dedicated lowering for zext-style shuffles is that they
/// are both incredibly common and often quite performance sensitive.
static SDValue lowerShuffleAsZeroOrAnyExtend(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    const APInt &Zeroable, const X86Subtarget &Subtarget,
    SelectionDAG &DAG) {
  int Bits = VT.getSizeInBits();
  int NumLanes = Bits / 128;
  int NumElements = VT.getVectorNumElements();
  int NumEltsPerLane = NumElements / NumLanes;
  assert(VT.getScalarSizeInBits() <= 32 &&
         "Exceeds 32-bit integer zero extension limit");
  assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
 
  // Define a helper function to check a particular ext-scale and lower to it if
  // valid.
  auto Lower = [&](int Scale) -> SDValue {
    SDValue InputV;
    bool AnyExt = true;
    int Offset = 0;
    int Matches = 0;
    for (int i = 0; i < NumElements; ++i) {
      int M = Mask[i];
      if (M < 0)
        continue; // Valid anywhere but doesn't tell us anything.
      if (i % Scale != 0) {
        // Each of the extended elements need to be zeroable.
        if (!Zeroable[i])
          return SDValue();
 
        // We no longer are in the anyext case.
        AnyExt = false;
        continue;
      }
 
      // Each of the base elements needs to be consecutive indices into the
      // same input vector.
      SDValue V = M < NumElements ? V1 : V2;
      M = M % NumElements;
      if (!InputV) {
        InputV = V;
        Offset = M - (i / Scale);
      } else if (InputV != V)
        return SDValue(); // Flip-flopping inputs.
 
      // Offset must start in the lowest 128-bit lane or at the start of an
      // upper lane.
      // FIXME: Is it ever worth allowing a negative base offset?
      if (!((0 <= Offset && Offset < NumEltsPerLane) ||
            (Offset % NumEltsPerLane) == 0))
        return SDValue();
 
      // If we are offsetting, all referenced entries must come from the same
      // lane.
      if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
        return SDValue();
 
      if ((M % NumElements) != (Offset + (i / Scale)))
        return SDValue(); // Non-consecutive strided elements.
      Matches++;
    }
 
    // If we fail to find an input, we have a zero-shuffle which should always
    // have already been handled.
    // FIXME: Maybe handle this here in case during blending we end up with one?
    if (!InputV)
      return SDValue();
 
    // If we are offsetting, don't extend if we only match a single input, we
    // can always do better by using a basic PSHUF or PUNPCK.
    if (Offset != 0 && Matches < 2)
      return SDValue();
 
    unsigned ExtOpc = AnyExt ? ISD::ANY_EXTEND : ISD::ZERO_EXTEND;
    return lowerShuffleAsSpecificExtension(DL, VT, Scale, Offset, ExtOpc,
                                           InputV, Mask, Subtarget, DAG);
  };
 
  // The widest scale possible for extending is to a 64-bit integer.
  assert(Bits % 64 == 0 &&
         "The number of bits in a vector must be divisible by 64 on x86!");
  int NumExtElements = Bits / 64;
 
  // Each iteration, try extending the elements half as much, but into twice as
  // many elements.
  for (; NumExtElements < NumElements; NumExtElements *= 2) {
    assert(NumElements % NumExtElements == 0 &&
           "The input vector size must be divisible by the extended size.");
    if (SDValue V = Lower(NumElements / NumExtElements))
      return V;
  }
 
  // General extends failed, but 128-bit vectors may be able to use MOVQ.
  if (Bits != 128)
    return SDValue();
 
  // Returns one of the source operands if the shuffle can be reduced to a
  // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
  auto CanZExtLowHalf = [&]() {
    for (int i = NumElements / 2; i != NumElements; ++i)
      if (!Zeroable[i])
        return SDValue();
    if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
      return V1;
    if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
      return V2;
    return SDValue();
  };
 
  if (SDValue V = CanZExtLowHalf()) {
    V = DAG.getBitcast(MVT::v2i64, V);
    V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
    return DAG.getBitcast(VT, V);
  }
 
  // No viable ext lowering found.
  return SDValue();
}
 
/// Try to get a scalar value for a specific element of a vector.
///
/// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
                                              SelectionDAG &DAG) {
  MVT VT = V.getSimpleValueType();
  MVT EltVT = VT.getVectorElementType();
  V = peekThroughBitcasts(V);
 
  // If the bitcasts shift the element size, we can't extract an equivalent
  // element from it.
  MVT NewVT = V.getSimpleValueType();
  if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
    return SDValue();
 
  if (V.getOpcode() == ISD::BUILD_VECTOR ||
      (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
    // Ensure the scalar operand is the same size as the destination.
    // FIXME: Add support for scalar truncation where possible.
    SDValue S = V.getOperand(Idx);
    if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
      return DAG.getBitcast(EltVT, S);
  }
 
  return SDValue();
}
 
/// Helper to test for a load that can be folded with x86 shuffles.
///
/// This is particularly important because the set of instructions varies
/// significantly based on whether the operand is a load or not.
static bool isShuffleFoldableLoad(SDValue V) {
  return V.hasOneUse() &&
         ISD::isNON_EXTLoad(peekThroughOneUseBitcasts(V).getNode());
}
 
template<typename T>
static bool isSoftF16(T VT, const X86Subtarget &Subtarget) {
  T EltVT = VT.getScalarType();
  return (EltVT == MVT::bf16 && !Subtarget.hasAVX10_2()) ||
         (EltVT == MVT::f16 && !Subtarget.hasFP16());
}
 
template<typename T>
static bool isBF16orSoftF16(T VT, const X86Subtarget &Subtarget) {
  T EltVT = VT.getScalarType();
  return EltVT == MVT::bf16 || (EltVT == MVT::f16 && !Subtarget.hasFP16());
}
 
/// Try to lower insertion of a single element into a zero vector.
///
/// This is a common pattern that we have especially efficient patterns to lower
/// across all subtarget feature sets.
static SDValue lowerShuffleAsElementInsertion(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    const APInt &Zeroable, const X86Subtarget &Subtarget,
    SelectionDAG &DAG) {
  MVT ExtVT = VT;
  MVT EltVT = VT.getVectorElementType();
  unsigned NumElts = VT.getVectorNumElements();
  unsigned EltBits = VT.getScalarSizeInBits();
 
  if (isSoftF16(EltVT, Subtarget))
    return SDValue();
 
  int V2Index =
      find_if(Mask, [&Mask](int M) { return M >= (int)Mask.size(); }) -
      Mask.begin();
  bool IsV1Constant = getTargetConstantFromNode(V1) != nullptr;
  bool IsV1Zeroable = true;
  for (int i = 0, Size = Mask.size(); i < Size; ++i)
    if (i != V2Index && !Zeroable[i]) {
      IsV1Zeroable = false;
      break;
    }
 
  // Bail if a non-zero V1 isn't used in place.
  if (!IsV1Zeroable) {
    SmallVector<int, 8> V1Mask(Mask);
    V1Mask[V2Index] = -1;
    if (!isNoopShuffleMask(V1Mask))
      return SDValue();
  }
 
  // Check for a single input from a SCALAR_TO_VECTOR node.
  // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
  // all the smarts here sunk into that routine. However, the current
  // lowering of BUILD_VECTOR makes that nearly impossible until the old
  // vector shuffle lowering is dead.
  SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
                                               DAG);
  if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
    // We need to zext the scalar if it is smaller than an i32.
    V2S = DAG.getBitcast(EltVT, V2S);
    if (EltVT == MVT::i8 || (EltVT == MVT::i16 && !Subtarget.hasFP16())) {
      // Using zext to expand a narrow element won't work for non-zero
      // insertions. But we can use a masked constant vector if we're
      // inserting V2 into the bottom of V1.
      if (!IsV1Zeroable && !(IsV1Constant && V2Index == 0))
        return SDValue();
 
      // Zero-extend directly to i32.
      ExtVT = MVT::getVectorVT(MVT::i32, ExtVT.getSizeInBits() / 32);
      V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
 
      // If we're inserting into a constant, mask off the inserted index
      // and OR with the zero-extended scalar.
      if (!IsV1Zeroable) {
        SmallVector<APInt> Bits(NumElts, APInt::getAllOnes(EltBits));
        Bits[V2Index] = APInt::getZero(EltBits);
        SDValue BitMask = getConstVector(Bits, VT, DAG, DL);
        V1 = DAG.getNode(ISD::AND, DL, VT, V1, BitMask);
        V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
        V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2));
        return DAG.getNode(ISD::OR, DL, VT, V1, V2);
      }
    }
    V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
  } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
             (EltVT == MVT::i16 && !Subtarget.hasAVX10_2())) {
    // Either not inserting from the low element of the input or the input
    // element size is too small to use VZEXT_MOVL to clear the high bits.
    return SDValue();
  }
 
  if (!IsV1Zeroable) {
    // If V1 can't be treated as a zero vector we have fewer options to lower
    // this. We can't support integer vectors or non-zero targets cheaply.
    assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
    if (!VT.isFloatingPoint() || V2Index != 0)
      return SDValue();
    if (!VT.is128BitVector())
      return SDValue();
 
    // Otherwise, use MOVSD, MOVSS or MOVSH.
    unsigned MovOpc = 0;
    if (EltVT == MVT::f16)
      MovOpc = X86ISD::MOVSH;
    else if (EltVT == MVT::f32)
      MovOpc = X86ISD::MOVSS;
    else if (EltVT == MVT::f64)
      MovOpc = X86ISD::MOVSD;
    else
      llvm_unreachable("Unsupported floating point element type to handle!");
    return DAG.getNode(MovOpc, DL, ExtVT, V1, V2);
  }
 
  // This lowering only works for the low element with floating point vectors.
  if (VT.isFloatingPoint() && V2Index != 0)
    return SDValue();
 
  V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
  if (ExtVT != VT)
    V2 = DAG.getBitcast(VT, V2);
 
  if (V2Index != 0) {
    // If we have 4 or fewer lanes we can cheaply shuffle the element into
    // the desired position. Otherwise it is more efficient to do a vector
    // shift left. We know that we can do a vector shift left because all
    // the inputs are zero.
    if (VT.isFloatingPoint() || NumElts <= 4) {
      SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
      V2Shuffle[V2Index] = 0;
      V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
    } else {
      V2 = DAG.getBitcast(MVT::v16i8, V2);
      V2 = DAG.getNode(
          X86ISD::VSHLDQ, DL, MVT::v16i8, V2,
          DAG.getTargetConstant(V2Index * EltBits / 8, DL, MVT::i8));
      V2 = DAG.getBitcast(VT, V2);
    }
  }
  return V2;
}
 
/// Try to lower broadcast of a single - truncated - integer element,
/// coming from a scalar_to_vector/build_vector node \p V0 with larger elements.
///
/// This assumes we have AVX2.
static SDValue lowerShuffleAsTruncBroadcast(const SDLoc &DL, MVT VT, SDValue V0,
                                            int BroadcastIdx,
                                            const X86Subtarget &Subtarget,
                                            SelectionDAG &DAG) {
  assert(Subtarget.hasAVX2() &&
         "We can only lower integer broadcasts with AVX2!");
 
  MVT EltVT = VT.getVectorElementType();
  MVT V0VT = V0.getSimpleValueType();
 
  assert(VT.isInteger() && "Unexpected non-integer trunc broadcast!");
  assert(V0VT.isVector() && "Unexpected non-vector vector-sized value!");
 
  MVT V0EltVT = V0VT.getVectorElementType();
  if (!V0EltVT.isInteger())
    return SDValue();
 
  const unsigned EltSize = EltVT.getSizeInBits();
  const unsigned V0EltSize = V0EltVT.getSizeInBits();
 
  // This is only a truncation if the original element type is larger.
  if (V0EltSize <= EltSize)
    return SDValue();
 
  assert(((V0EltSize % EltSize) == 0) &&
         "Scalar type sizes must all be powers of 2 on x86!");
 
  const unsigned V0Opc = V0.getOpcode();
  const unsigned Scale = V0EltSize / EltSize;
  const unsigned V0BroadcastIdx = BroadcastIdx / Scale;
 
  if ((V0Opc != ISD::SCALAR_TO_VECTOR || V0BroadcastIdx != 0) &&
      V0Opc != ISD::BUILD_VECTOR)
    return SDValue();
 
  SDValue Scalar = V0.getOperand(V0BroadcastIdx);
 
  // If we're extracting non-least-significant bits, shift so we can truncate.
  // Hopefully, we can fold away the trunc/srl/load into the broadcast.
  // Even if we can't (and !isShuffleFoldableLoad(Scalar)), prefer
  // vpbroadcast+vmovd+shr to vpshufb(m)+vmovd.
  if (const int OffsetIdx = BroadcastIdx % Scale)
    Scalar = DAG.getNode(ISD::SRL, DL, Scalar.getValueType(), Scalar,
                         DAG.getConstant(OffsetIdx * EltSize, DL, MVT::i8));
 
  return DAG.getNode(X86ISD::VBROADCAST, DL, VT,
                     DAG.getNode(ISD::TRUNCATE, DL, EltVT, Scalar));
}
 
/// Test whether this can be lowered with a single SHUFPS instruction.
///
/// This is used to disable more specialized lowerings when the shufps lowering
/// will happen to be efficient.
static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
  // This routine only handles 128-bit shufps.
  assert(Mask.size() == 4 && "Unsupported mask size!");
  assert(Mask[0] >= -1 && Mask[0] < 8 && "Out of bound mask element!");
  assert(Mask[1] >= -1 && Mask[1] < 8 && "Out of bound mask element!");
  assert(Mask[2] >= -1 && Mask[2] < 8 && "Out of bound mask element!");
  assert(Mask[3] >= -1 && Mask[3] < 8 && "Out of bound mask element!");
 
  // To lower with a single SHUFPS we need to have the low half and high half
  // each requiring a single input.
  if (Mask[0] >= 0 && Mask[1] >= 0 && (Mask[0] < 4) != (Mask[1] < 4))
    return false;
  if (Mask[2] >= 0 && Mask[3] >= 0 && (Mask[2] < 4) != (Mask[3] < 4))
    return false;
 
  return true;
}
 
/// Test whether the specified input (0 or 1) is in-place blended by the
/// given mask.
///
/// This returns true if the elements from a particular input are already in the
/// slot required by the given mask and require no permutation.
static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
  assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
  int Size = Mask.size();
  for (int i = 0; i < Size; ++i)
    if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
      return false;
 
  return true;
}
 
/// Test whether the specified input (0 or 1) is a broadcast/splat blended by
/// the given mask.
///
static bool isShuffleMaskInputBroadcastable(int Input, ArrayRef<int> Mask,
                                            int BroadcastableElement = 0) {
  assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
  int Size = Mask.size();
  for (int i = 0; i < Size; ++i)
    if (Mask[i] >= 0 && Mask[i] / Size == Input &&
        Mask[i] % Size != BroadcastableElement)
      return false;
  return true;
}
 
/// If we are extracting two 128-bit halves of a vector and shuffling the
/// result, match that to a 256-bit AVX2 vperm* instruction to avoid a
/// multi-shuffle lowering.
static SDValue lowerShuffleOfExtractsAsVperm(const SDLoc &DL, SDValue N0,
                                             SDValue N1, ArrayRef<int> Mask,
                                             SelectionDAG &DAG) {
  MVT VT = N0.getSimpleValueType();
  assert((VT.is128BitVector() &&
          (VT.getScalarSizeInBits() == 32 || VT.getScalarSizeInBits() == 64)) &&
         "VPERM* family of shuffles requires 32-bit or 64-bit elements");
 
  // Check that both sources are extracts of the same source vector.
  if (N0.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
      N1.getOpcode() != ISD::EXTRACT_SUBVECTOR ||
      N0.getOperand(0) != N1.getOperand(0) ||
      !N0.hasOneUse() || !N1.hasOneUse())
    return SDValue();
 
  SDValue WideVec = N0.getOperand(0);
  MVT WideVT = WideVec.getSimpleValueType();
  if (!WideVT.is256BitVector())
    return SDValue();
 
  // Match extracts of each half of the wide source vector. Commute the shuffle
  // if the extract of the low half is N1.
  unsigned NumElts = VT.getVectorNumElements();
  SmallVector<int, 4> NewMask(Mask);
  const APInt &ExtIndex0 = N0.getConstantOperandAPInt(1);
  const APInt &ExtIndex1 = N1.getConstantOperandAPInt(1);
  if (ExtIndex1 == 0 && ExtIndex0 == NumElts)
    ShuffleVectorSDNode::commuteMask(NewMask);
  else if (ExtIndex0 != 0 || ExtIndex1 != NumElts)
    return SDValue();
 
  // Final bailout: if the mask is simple, we are better off using an extract
  // and a simple narrow shuffle. Prefer extract+unpack(h/l)ps to vpermps
  // because that avoids a constant load from memory.
  if (NumElts == 4 &&
      (isSingleSHUFPSMask(NewMask) || is128BitUnpackShuffleMask(NewMask, DAG)))
    return SDValue();
 
  // Extend the shuffle mask with undef elements.
  NewMask.append(NumElts, -1);
 
  // shuf (extract X, 0), (extract X, 4), M --> extract (shuf X, undef, M'), 0
  SDValue Shuf = DAG.getVectorShuffle(WideVT, DL, WideVec, DAG.getUNDEF(WideVT),
                                      NewMask);
  // This is free: ymm -> xmm.
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Shuf,
                     DAG.getVectorIdxConstant(0, DL));
}
 
/// Try to lower broadcast of a single element.
///
/// For convenience, this code also bundles all of the subtarget feature set
/// filtering. While a little annoying to re-dispatch on type here, there isn't
/// a convenient way to factor it out.
static SDValue lowerShuffleAsBroadcast(const SDLoc &DL, MVT VT, SDValue V1,
                                       SDValue V2, ArrayRef<int> Mask,
                                       const X86Subtarget &Subtarget,
                                       SelectionDAG &DAG) {
  MVT EltVT = VT.getVectorElementType();
  if (!((Subtarget.hasSSE3() && VT == MVT::v2f64) ||
        (Subtarget.hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
        (Subtarget.hasAVX2() && (VT.isInteger() || EltVT == MVT::f16))))
    return SDValue();
 
  // With MOVDDUP (v2f64) we can broadcast from a register or a load, otherwise
  // we can only broadcast from a register with AVX2.
  unsigned NumEltBits = VT.getScalarSizeInBits();
  unsigned Opcode = (VT == MVT::v2f64 && !Subtarget.hasAVX2())
                        ? X86ISD::MOVDDUP
                        : X86ISD::VBROADCAST;
  bool BroadcastFromReg = (Opcode == X86ISD::MOVDDUP) || Subtarget.hasAVX2();
 
  // Check that the mask is a broadcast.
  int BroadcastIdx = getSplatIndex(Mask);
  if (BroadcastIdx < 0) {
    // Check for hidden broadcast.
    SmallVector<int, 16> BroadcastMask(VT.getVectorNumElements(), 0);
    if (!isShuffleEquivalent(Mask, BroadcastMask, V1, V2))
      return SDValue();
    BroadcastIdx = 0;
  }
  assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
                                            "a sorted mask where the broadcast "
                                            "comes from V1.");
  int NumActiveElts = count_if(Mask, [](int M) { return M >= 0; });
 
  // Go up the chain of (vector) values to find a scalar load that we can
  // combine with the broadcast.
  // TODO: Combine this logic with findEltLoadSrc() used by
  //       EltsFromConsecutiveLoads().
  int BitOffset = BroadcastIdx * NumEltBits;
  SDValue V = V1;
  for (;;) {
    switch (V.getOpcode()) {
    case ISD::BITCAST: {
      V = V.getOperand(0);
      continue;
    }
    case ISD::CONCAT_VECTORS: {
      int OpBitWidth = V.getOperand(0).getValueSizeInBits();
      int OpIdx = BitOffset / OpBitWidth;
      V = V.getOperand(OpIdx);
      BitOffset %= OpBitWidth;
      continue;
    }
    case ISD::EXTRACT_SUBVECTOR: {
      // The extraction index adds to the existing offset.
      unsigned EltBitWidth = V.getScalarValueSizeInBits();
      unsigned Idx = V.getConstantOperandVal(1);
      unsigned BeginOffset = Idx * EltBitWidth;
      BitOffset += BeginOffset;
      V = V.getOperand(0);
      continue;
    }
    case ISD::INSERT_SUBVECTOR: {
      SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
      int EltBitWidth = VOuter.getScalarValueSizeInBits();
      int Idx = (int)V.getConstantOperandVal(2);
      int NumSubElts = (int)VInner.getSimpleValueType().getVectorNumElements();
      int BeginOffset = Idx * EltBitWidth;
      int EndOffset = BeginOffset + NumSubElts * EltBitWidth;
      if (BeginOffset <= BitOffset && BitOffset < EndOffset) {
        BitOffset -= BeginOffset;
        V = VInner;
      } else {
        V = VOuter;
      }
      continue;
    }
    }
    break;
  }
  assert((BitOffset % NumEltBits) == 0 && "Illegal bit-offset");
  BroadcastIdx = BitOffset / NumEltBits;
 
  // Do we need to bitcast the source to retrieve the original broadcast index?
  bool BitCastSrc = V.getScalarValueSizeInBits() != NumEltBits;
 
  // Check if this is a broadcast of a scalar. We special case lowering
  // for scalars so that we can more effectively fold with loads.
  // If the original value has a larger element type than the shuffle, the
  // broadcast element is in essence truncated. Make that explicit to ease
  // folding.
  if (BitCastSrc && VT.isInteger())
    if (SDValue TruncBroadcast = lowerShuffleAsTruncBroadcast(
            DL, VT, V, BroadcastIdx, Subtarget, DAG))
      return TruncBroadcast;
 
  // Also check the simpler case, where we can directly reuse the scalar.
  if (!BitCastSrc &&
      ((V.getOpcode() == ISD::BUILD_VECTOR && V.hasOneUse()) ||
       (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0))) {
    V = V.getOperand(BroadcastIdx);
 
    // If we can't broadcast from a register, check that the input is a load.
    if (!BroadcastFromReg && !isShuffleFoldableLoad(V))
      return SDValue();
  } else if (ISD::isNormalLoad(V.getNode()) &&
             cast<LoadSDNode>(V)->isSimple()) {
    // We do not check for one-use of the vector load because a broadcast load
    // is expected to be a win for code size, register pressure, and possibly
    // uops even if the original vector load is not eliminated.
 
    // Reduce the vector load and shuffle to a broadcasted scalar load.
    auto *Ld = cast<LoadSDNode>(V);
    SDValue BaseAddr = Ld->getBasePtr();
    MVT SVT = VT.getScalarType();
    unsigned Offset = BroadcastIdx * SVT.getStoreSize();
    assert((int)(Offset * 8) == BitOffset && "Unexpected bit-offset");
    SDValue NewAddr =
        DAG.getMemBasePlusOffset(BaseAddr, TypeSize::getFixed(Offset), DL);
 
    // Directly form VBROADCAST_LOAD if we're using VBROADCAST opcode rather
    // than MOVDDUP.
    // FIXME: Should we add VBROADCAST_LOAD isel patterns for pre-AVX?
    if (Opcode == X86ISD::VBROADCAST) {
      SDVTList Tys = DAG.getVTList(VT, MVT::Other);
      SDValue Ops[] = {Ld->getChain(), NewAddr};
      V = DAG.getMemIntrinsicNode(
          X86ISD::VBROADCAST_LOAD, DL, Tys, Ops, SVT,
          DAG.getMachineFunction().getMachineMemOperand(
              Ld->getMemOperand(), Offset, SVT.getStoreSize()));
      DAG.makeEquivalentMemoryOrdering(Ld, V);
      return DAG.getBitcast(VT, V);
    }
    assert(SVT == MVT::f64 && "Unexpected VT!");
    V = DAG.getLoad(SVT, DL, Ld->getChain(), NewAddr,
                    DAG.getMachineFunction().getMachineMemOperand(
                        Ld->getMemOperand(), Offset, SVT.getStoreSize()));
    DAG.makeEquivalentMemoryOrdering(Ld, V);
  } else if (!BroadcastFromReg) {
    // We can't broadcast from a vector register.
    return SDValue();
  } else if (BitOffset != 0) {
    // We can only broadcast from the zero-element of a vector register,
    // but it can be advantageous to broadcast from the zero-element of a
    // subvector.
    if (!VT.is256BitVector() && !VT.is512BitVector())
      return SDValue();
 
    // VPERMQ/VPERMPD can perform the cross-lane shuffle directly.
    if (VT == MVT::v4f64 || VT == MVT::v4i64)
      return SDValue();
 
    // If we are broadcasting an element from the lowest 128-bit subvector, try
    // to move the element in position.
    if (BitOffset < 128 && NumActiveElts > 1 &&
        V.getScalarValueSizeInBits() == NumEltBits) {
      assert((BitOffset % V.getScalarValueSizeInBits()) == 0 &&
             "Unexpected bit-offset");
      SmallVector<int, 16> ExtractMask(128 / NumEltBits, SM_SentinelUndef);
      ExtractMask[0] = BitOffset / V.getScalarValueSizeInBits();
      V = extractSubVector(V, 0, DAG, DL, 128);
      V = DAG.getVectorShuffle(V.getValueType(), DL, V, V, ExtractMask);
    } else {
      // Only broadcast the zero-element of a 128-bit subvector.
      if ((BitOffset % 128) != 0)
        return SDValue();
 
      assert((BitOffset % V.getScalarValueSizeInBits()) == 0 &&
             "Unexpected bit-offset");
      assert((V.getValueSizeInBits() == 256 || V.getValueSizeInBits() == 512) &&
             "Unexpected vector size");
      unsigned ExtractIdx = BitOffset / V.getScalarValueSizeInBits();
      V = extract128BitVector(V, ExtractIdx, DAG, DL);
    }
  }
 
  // On AVX we can use VBROADCAST directly for scalar sources.
  if (Opcode == X86ISD::MOVDDUP && !V.getValueType().isVector()) {
    V = DAG.getBitcast(MVT::f64, V);
    if (Subtarget.hasAVX()) {
      V = DAG.getNode(X86ISD::VBROADCAST, DL, MVT::v2f64, V);
      return DAG.getBitcast(VT, V);
    }
    V = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V);
  }
 
  // If this is a scalar, do the broadcast on this type and bitcast.
  if (!V.getValueType().isVector()) {
    assert(V.getScalarValueSizeInBits() == NumEltBits &&
           "Unexpected scalar size");
    MVT BroadcastVT = MVT::getVectorVT(V.getSimpleValueType(),
                                       VT.getVectorNumElements());
    return DAG.getBitcast(VT, DAG.getNode(Opcode, DL, BroadcastVT, V));
  }
 
  // We only support broadcasting from 128-bit vectors to minimize the
  // number of patterns we need to deal with in isel. So extract down to
  // 128-bits, removing as many bitcasts as possible.
  if (V.getValueSizeInBits() > 128)
    V = extract128BitVector(peekThroughBitcasts(V), 0, DAG, DL);
 
  // Otherwise cast V to a vector with the same element type as VT, but
  // possibly narrower than VT. Then perform the broadcast.
  unsigned NumSrcElts = V.getValueSizeInBits() / NumEltBits;
  MVT CastVT = MVT::getVectorVT(VT.getVectorElementType(), NumSrcElts);
  return DAG.getNode(Opcode, DL, VT, DAG.getBitcast(CastVT, V));
}
 
// Check for whether we can use INSERTPS to perform the shuffle. We only use
// INSERTPS when the V1 elements are already in the correct locations
// because otherwise we can just always use two SHUFPS instructions which
// are much smaller to encode than a SHUFPS and an INSERTPS. We can also
// perform INSERTPS if a single V1 element is out of place and all V2
// elements are zeroable.
static bool matchShuffleAsInsertPS(SDValue &V1, SDValue &V2,
                                   unsigned &InsertPSMask,
                                   const APInt &Zeroable,
                                   ArrayRef<int> Mask, SelectionDAG &DAG) {
  assert(V1.getSimpleValueType().is128BitVector() && "Bad operand type!");
  assert(V2.getSimpleValueType().is128BitVector() && "Bad operand type!");
  assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
 
  // Attempt to match INSERTPS with one element from VA or VB being
  // inserted into VA (or undef). If successful, V1, V2 and InsertPSMask
  // are updated.
  auto matchAsInsertPS = [&](SDValue VA, SDValue VB,
                             ArrayRef<int> CandidateMask) {
    unsigned ZMask = 0;
    int VADstIndex = -1;
    int VBDstIndex = -1;
    bool VAUsedInPlace = false;
 
    for (int i = 0; i < 4; ++i) {
      // Synthesize a zero mask from the zeroable elements (includes undefs).
      if (Zeroable[i]) {
        ZMask |= 1 << i;
        continue;
      }
 
      // Flag if we use any VA inputs in place.
      if (i == CandidateMask[i]) {
        VAUsedInPlace = true;
        continue;
      }
 
      // We can only insert a single non-zeroable element.
      if (VADstIndex >= 0 || VBDstIndex >= 0)
        return false;
 
      if (CandidateMask[i] < 4) {
        // VA input out of place for insertion.
        VADstIndex = i;
      } else {
        // VB input for insertion.
        VBDstIndex = i;
      }
    }
 
    // Don't bother if we have no (non-zeroable) element for insertion.
    if (VADstIndex < 0 && VBDstIndex < 0)
      return false;
 
    // Determine element insertion src/dst indices. The src index is from the
    // start of the inserted vector, not the start of the concatenated vector.
    unsigned VBSrcIndex = 0;
    if (VADstIndex >= 0) {
      // If we have a VA input out of place, we use VA as the V2 element
      // insertion and don't use the original V2 at all.
      VBSrcIndex = CandidateMask[VADstIndex];
      VBDstIndex = VADstIndex;
      VB = VA;
    } else {
      VBSrcIndex = CandidateMask[VBDstIndex] - 4;
    }
 
    // If no V1 inputs are used in place, then the result is created only from
    // the zero mask and the V2 insertion - so remove V1 dependency.
    if (!VAUsedInPlace)
      VA = DAG.getUNDEF(MVT::v4f32);
 
    // Update V1, V2 and InsertPSMask accordingly.
    V1 = VA;
    V2 = VB;
 
    // Insert the V2 element into the desired position.
    InsertPSMask = VBSrcIndex << 6 | VBDstIndex << 4 | ZMask;
    assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
    return true;
  };
 
  if (matchAsInsertPS(V1, V2, Mask))
    return true;
 
  // Commute and try again.
  SmallVector<int, 4> CommutedMask(Mask);
  ShuffleVectorSDNode::commuteMask(CommutedMask);
  if (matchAsInsertPS(V2, V1, CommutedMask))
    return true;
 
  return false;
}
 
static SDValue lowerShuffleAsInsertPS(const SDLoc &DL, SDValue V1, SDValue V2,
                                      ArrayRef<int> Mask, const APInt &Zeroable,
                                      SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
 
  // Attempt to match the insertps pattern.
  unsigned InsertPSMask = 0;
  if (!matchShuffleAsInsertPS(V1, V2, InsertPSMask, Zeroable, Mask, DAG))
    return SDValue();
 
  // Insert the V2 element into the desired position.
  return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
                     DAG.getTargetConstant(InsertPSMask, DL, MVT::i8));
}
 
/// Handle lowering of 2-lane 64-bit floating point shuffles.
///
/// This is the basis function for the 2-lane 64-bit shuffles as we have full
/// support for floating point shuffles but not integer shuffles. These
/// instructions will incur a domain crossing penalty on some chips though so
/// it is better to avoid lowering through this for integer vectors where
/// possible.
static SDValue lowerV2F64Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
  assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
 
  if (V2.isUndef()) {
    // Check for being able to broadcast a single element.
    if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v2f64, V1, V2,
                                                    Mask, Subtarget, DAG))
      return Broadcast;
 
    // Straight shuffle of a single input vector. Simulate this by using the
    // single input as both of the "inputs" to this instruction..
    unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
 
    if (Subtarget.hasAVX()) {
      // If we have AVX, we can use VPERMILPS which will allow folding a load
      // into the shuffle.
      return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
                         DAG.getTargetConstant(SHUFPDMask, DL, MVT::i8));
    }
 
    return DAG.getNode(
        X86ISD::SHUFP, DL, MVT::v2f64,
        Mask[0] == SM_SentinelUndef ? DAG.getUNDEF(MVT::v2f64) : V1,
        Mask[1] == SM_SentinelUndef ? DAG.getUNDEF(MVT::v2f64) : V1,
        DAG.getTargetConstant(SHUFPDMask, DL, MVT::i8));
  }
  assert(Mask[0] >= 0 && "No undef lanes in multi-input v2 shuffles!");
  assert(Mask[1] >= 0 && "No undef lanes in multi-input v2 shuffles!");
  assert(Mask[0] < 2 && "We sort V1 to be the first input.");
  assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
 
  if (Subtarget.hasAVX2())
    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
      return Extract;
 
  // When loading a scalar and then shuffling it into a vector we can often do
  // the insertion cheaply.
  if (SDValue Insertion = lowerShuffleAsElementInsertion(
          DL, MVT::v2f64, V1, V2, Mask, Zeroable, Subtarget, DAG))
    return Insertion;
  // Try inverting the insertion since for v2 masks it is easy to do and we
  // can't reliably sort the mask one way or the other.
  int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
                        Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
  if (SDValue Insertion = lowerShuffleAsElementInsertion(
          DL, MVT::v2f64, V2, V1, InverseMask, Zeroable, Subtarget, DAG))
    return Insertion;
 
  // Try to use one of the special instruction patterns to handle two common
  // blend patterns if a zero-blend above didn't work.
  if (isShuffleEquivalent(Mask, {0, 3}, V1, V2) ||
      isShuffleEquivalent(Mask, {1, 3}, V1, V2))
    if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
      // We can either use a special instruction to load over the low double or
      // to move just the low double.
      return DAG.getNode(
          X86ISD::MOVSD, DL, MVT::v2f64, V2,
          DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
 
  if (Subtarget.hasSSE41())
    if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
      return Blend;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v2f64, V1, V2, Mask, DAG))
    return V;
 
  unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
  return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
                     DAG.getTargetConstant(SHUFPDMask, DL, MVT::i8));
}
 
/// Handle lowering of 2-lane 64-bit integer shuffles.
///
/// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
/// the integer unit to minimize domain crossing penalties. However, for blends
/// it falls back to the floating point shuffle operation with appropriate bit
/// casting.
static SDValue lowerV2I64Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
  assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
 
  if (V2.isUndef()) {
    // Check for being able to broadcast a single element.
    if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v2i64, V1, V2,
                                                    Mask, Subtarget, DAG))
      return Broadcast;
 
    // Straight shuffle of a single input vector. For everything from SSE2
    // onward this has a single fast instruction with no scary immediates.
    // We have to map the mask as it is actually a v4i32 shuffle instruction.
    V1 = DAG.getBitcast(MVT::v4i32, V1);
    int WidenedMask[4] = {Mask[0] < 0 ? -1 : (Mask[0] * 2),
                          Mask[0] < 0 ? -1 : ((Mask[0] * 2) + 1),
                          Mask[1] < 0 ? -1 : (Mask[1] * 2),
                          Mask[1] < 0 ? -1 : ((Mask[1] * 2) + 1)};
    return DAG.getBitcast(
        MVT::v2i64,
        DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
                    getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
  }
  assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
  assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
  assert(Mask[0] < 2 && "We sort V1 to be the first input.");
  assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
 
  if (Subtarget.hasAVX2())
    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
      return Extract;
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // When loading a scalar and then shuffling it into a vector we can often do
  // the insertion cheaply.
  if (SDValue Insertion = lowerShuffleAsElementInsertion(
          DL, MVT::v2i64, V1, V2, Mask, Zeroable, Subtarget, DAG))
    return Insertion;
  // Try inverting the insertion since for v2 masks it is easy to do and we
  // can't reliably sort the mask one way or the other.
  int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
  if (SDValue Insertion = lowerShuffleAsElementInsertion(
          DL, MVT::v2i64, V2, V1, InverseMask, Zeroable, Subtarget, DAG))
    return Insertion;
 
  // We have different paths for blend lowering, but they all must use the
  // *exact* same predicate.
  bool IsBlendSupported = Subtarget.hasSSE41();
  if (IsBlendSupported)
    if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
      return Blend;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v2i64, V1, V2, Mask, DAG))
    return V;
 
  // Try to use byte rotation instructions.
  // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
  if (Subtarget.hasSSSE3()) {
    if (Subtarget.hasVLX())
      if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v2i64, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG))
        return Rotate;
 
    if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v2i64, V1, V2, Mask,
                                                  Subtarget, DAG))
      return Rotate;
  }
 
  // If we have direct support for blends, we should lower by decomposing into
  // a permute. That will be faster than the domain cross.
  if (IsBlendSupported)
    return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v2i64, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG);
 
  // We implement this with SHUFPD which is pretty lame because it will likely
  // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
  // However, all the alternatives are still more cycles and newer chips don't
  // have this problem. It would be really nice if x86 had better shuffles here.
  V1 = DAG.getBitcast(MVT::v2f64, V1);
  V2 = DAG.getBitcast(MVT::v2f64, V2);
  return DAG.getBitcast(MVT::v2i64,
                        DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
}
 
/// Lower a vector shuffle using the SHUFPS instruction.
///
/// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
/// It makes no assumptions about whether this is the *best* lowering, it simply
/// uses it.
static SDValue lowerShuffleWithSHUFPS(const SDLoc &DL, MVT VT,
                                      ArrayRef<int> Mask, SDValue V1,
                                      SDValue V2, SelectionDAG &DAG) {
  SDValue LowV = V1, HighV = V2;
  SmallVector<int, 4> NewMask(Mask);
  int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; });
 
  if (NumV2Elements == 1) {
    int V2Index = find_if(Mask, [](int M) { return M >= 4; }) - Mask.begin();
 
    // Compute the index adjacent to V2Index and in the same half by toggling
    // the low bit.
    int V2AdjIndex = V2Index ^ 1;
 
    if (Mask[V2AdjIndex] < 0) {
      // Handles all the cases where we have a single V2 element and an undef.
      // This will only ever happen in the high lanes because we commute the
      // vector otherwise.
      if (V2Index < 2)
        std::swap(LowV, HighV);
      NewMask[V2Index] -= 4;
    } else {
      // Handle the case where the V2 element ends up adjacent to a V1 element.
      // To make this work, blend them together as the first step.
      int V1Index = V2AdjIndex;
      int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
      V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
                       getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
 
      // Now proceed to reconstruct the final blend as we have the necessary
      // high or low half formed.
      if (V2Index < 2) {
        LowV = V2;
        HighV = V1;
      } else {
        HighV = V2;
      }
      NewMask[V1Index] = 2; // We put the V1 element in V2[2].
      NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
    }
  } else if (NumV2Elements == 2) {
    if (Mask[0] < 4 && Mask[1] < 4) {
      // Handle the easy case where we have V1 in the low lanes and V2 in the
      // high lanes.
      NewMask[2] -= 4;
      NewMask[3] -= 4;
    } else if (Mask[2] < 4 && Mask[3] < 4) {
      // We also handle the reversed case because this utility may get called
      // when we detect a SHUFPS pattern but can't easily commute the shuffle to
      // arrange things in the right direction.
      NewMask[0] -= 4;
      NewMask[1] -= 4;
      HighV = V1;
      LowV = V2;
    } else {
      // We have a mixture of V1 and V2 in both low and high lanes. Rather than
      // trying to place elements directly, just blend them and set up the final
      // shuffle to place them.
 
      // The first two blend mask elements are for V1, the second two are for
      // V2.
      int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
                          Mask[2] < 4 ? Mask[2] : Mask[3],
                          (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
                          (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
      V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
                       getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
 
      // Now we do a normal shuffle of V1 by giving V1 as both operands to
      // a blend.
      LowV = HighV = V1;
      NewMask[0] = Mask[0] < 4 ? 0 : 2;
      NewMask[1] = Mask[0] < 4 ? 2 : 0;
      NewMask[2] = Mask[2] < 4 ? 1 : 3;
      NewMask[3] = Mask[2] < 4 ? 3 : 1;
    }
  } else if (NumV2Elements == 3) {
    // Ideally canonicalizeShuffleMaskWithCommute should have caught this, but
    // we can get here due to other paths (e.g repeated mask matching) that we
    // don't want to do another round of lowerVECTOR_SHUFFLE.
    ShuffleVectorSDNode::commuteMask(NewMask);
    return lowerShuffleWithSHUFPS(DL, VT, NewMask, V2, V1, DAG);
  }
  return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
                     getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
}
 
/// Lower 4-lane 32-bit floating point shuffles.
///
/// Uses instructions exclusively from the floating point unit to minimize
/// domain crossing penalties, as these are sufficient to implement all v4f32
/// shuffles.
static SDValue lowerV4F32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
  assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
 
  if (Subtarget.hasSSE41())
    if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
      return Blend;
 
  int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; });
 
  if (NumV2Elements == 0) {
    // Check for being able to broadcast a single element.
    if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4f32, V1, V2,
                                                    Mask, Subtarget, DAG))
      return Broadcast;
 
    // Use even/odd duplicate instructions for masks that match their pattern.
    if (Subtarget.hasSSE3()) {
      if (isShuffleEquivalent(Mask, {0, 0, 2, 2}, V1, V2))
        return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
      if (isShuffleEquivalent(Mask, {1, 1, 3, 3}, V1, V2))
        return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
    }
 
    if (Subtarget.hasAVX()) {
      // If we have AVX, we can use VPERMILPS which will allow folding a load
      // into the shuffle.
      return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
                         getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
    }
 
    // Use MOVLHPS/MOVHLPS to simulate unary shuffles. These are only valid
    // in SSE1 because otherwise they are widened to v2f64 and never get here.
    if (!Subtarget.hasSSE2()) {
      if (isShuffleEquivalent(Mask, {0, 1, 0, 1}, V1, V2))
        return DAG.getNode(X86ISD::MOVLHPS, DL, MVT::v4f32, V1, V1);
      if (isShuffleEquivalent(Mask, {2, 3, 2, 3}, V1, V2))
        return DAG.getNode(X86ISD::MOVHLPS, DL, MVT::v4f32, V1, V1);
    }
 
    // Otherwise, use a straight shuffle of a single input vector. We pass the
    // input vector to both operands to simulate this with a SHUFPS.
    return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
                       getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
  }
 
  if (Subtarget.hasSSE2())
    if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(
            DL, MVT::v4i32, V1, V2, Mask, Zeroable, Subtarget, DAG)) {
      ZExt = DAG.getBitcast(MVT::v4f32, ZExt);
      return ZExt;
    }
 
  if (Subtarget.hasAVX2())
    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
      return Extract;
 
  // There are special ways we can lower some single-element blends. However, we
  // have custom ways we can lower more complex single-element blends below that
  // we defer to if both this and BLENDPS fail to match, so restrict this to
  // when the V2 input is targeting element 0 of the mask -- that is the fast
  // case here.
  if (NumV2Elements == 1 && Mask[0] >= 4)
    if (SDValue V = lowerShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2, Mask,
                                                   Zeroable, Subtarget, DAG))
      return V;
 
  if (Subtarget.hasSSE41()) {
    bool MatchesShufPS = isSingleSHUFPSMask(Mask);
 
    // Use INSERTPS if we can complete the shuffle efficiently.
    if (!MatchesShufPS || Zeroable == 0x3 || Zeroable == 0xC)
      if (SDValue V = lowerShuffleAsInsertPS(DL, V1, V2, Mask, Zeroable, DAG))
        return V;
 
    if (!MatchesShufPS)
      if (SDValue BlendPerm =
              lowerShuffleAsBlendAndPermute(DL, MVT::v4f32, V1, V2, Mask, DAG))
        return BlendPerm;
  }
 
  // Use low/high mov instructions. These are only valid in SSE1 because
  // otherwise they are widened to v2f64 and never get here.
  if (!Subtarget.hasSSE2()) {
    if (isShuffleEquivalent(Mask, {0, 1, 4, 5}, V1, V2))
      return DAG.getNode(X86ISD::MOVLHPS, DL, MVT::v4f32, V1, V2);
    if (isShuffleEquivalent(Mask, {2, 3, 6, 7}, V1, V2))
      return DAG.getNode(X86ISD::MOVHLPS, DL, MVT::v4f32, V2, V1);
  }
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4f32, V1, V2, Mask, DAG))
    return V;
 
  // Otherwise fall back to a SHUFPS lowering strategy.
  return lowerShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
}
 
/// Lower 4-lane i32 vector shuffles.
///
/// We try to handle these with integer-domain shuffles where we can, but for
/// blends we use the floating point domain blend instructions.
static SDValue lowerV4I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
  assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative. It also allows us to fold memory operands into the
  // shuffle in many cases.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2, Mask,
                                                   Zeroable, Subtarget, DAG))
    return ZExt;
 
  int NumV2Elements = count_if(Mask, [](int M) { return M >= 4; });
 
  // Try to use shift instructions if fast.
  if (Subtarget.preferLowerShuffleAsShift()) {
    if (SDValue Shift =
            lowerShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, Zeroable,
                                Subtarget, DAG, /*BitwiseOnly*/ true))
      return Shift;
    if (NumV2Elements == 0)
      if (SDValue Rotate =
              lowerShuffleAsBitRotate(DL, MVT::v4i32, V1, Mask, Subtarget, DAG))
        return Rotate;
  }
 
  if (NumV2Elements == 0) {
    // Try to use broadcast unless the mask only has one non-undef element.
    if (count_if(Mask, [](int M) { return M >= 0 && M < 4; }) > 1) {
      if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4i32, V1, V2,
                                                      Mask, Subtarget, DAG))
        return Broadcast;
    }
 
    // Straight shuffle of a single input vector. For everything from SSE2
    // onward this has a single fast instruction with no scary immediates.
    // We coerce the shuffle pattern to be compatible with UNPCK instructions
    // but we aren't actually going to use the UNPCK instruction because doing
    // so prevents folding a load into this instruction or making a copy.
    const int UnpackLoMask[] = {0, 0, 1, 1};
    const int UnpackHiMask[] = {2, 2, 3, 3};
    if (!isSingleElementRepeatedMask(Mask)) {
      if (isShuffleEquivalent(Mask, {0, 0, 1, 1}, V1, V2))
        Mask = UnpackLoMask;
      else if (isShuffleEquivalent(Mask, {2, 2, 3, 3}, V1, V2))
        Mask = UnpackHiMask;
    }
 
    return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
                       getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
  }
 
  if (Subtarget.hasAVX2())
    if (SDValue Extract = lowerShuffleOfExtractsAsVperm(DL, V1, V2, Mask, DAG))
      return Extract;
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // There are special ways we can lower some single-element blends.
  if (NumV2Elements == 1)
    if (SDValue V = lowerShuffleAsElementInsertion(
            DL, MVT::v4i32, V1, V2, Mask, Zeroable, Subtarget, DAG))
      return V;
 
  // We have different paths for blend lowering, but they all must use the
  // *exact* same predicate.
  bool IsBlendSupported = Subtarget.hasSSE41();
  if (IsBlendSupported)
    if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
      return Blend;
 
  if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask,
                                             Zeroable, Subtarget, DAG))
    return Masked;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4i32, V1, V2, Mask, DAG))
    return V;
 
  // Try to use byte rotation instructions.
  // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
  if (Subtarget.hasSSSE3()) {
    if (Subtarget.hasVLX())
      if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v4i32, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG))
        return Rotate;
 
    if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v4i32, V1, V2, Mask,
                                                  Subtarget, DAG))
      return Rotate;
  }
 
  // Assume that a single SHUFPS is faster than an alternative sequence of
  // multiple instructions (even if the CPU has a domain penalty).
  // If some CPU is harmed by the domain switch, we can fix it in a later pass.
  if (!isSingleSHUFPSMask(Mask)) {
    // If we have direct support for blends, we should lower by decomposing into
    // a permute. That will be faster than the domain cross.
    if (IsBlendSupported)
      return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4i32, V1, V2, Mask,
                                                  Zeroable, Subtarget, DAG);
 
    // Try to lower by permuting the inputs into an unpack instruction.
    if (SDValue Unpack = lowerShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1, V2,
                                                        Mask, Subtarget, DAG))
      return Unpack;
  }
 
  // We implement this with SHUFPS because it can blend from two vectors.
  // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
  // up the inputs, bypassing domain shift penalties that we would incur if we
  // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
  // relevant.
  SDValue CastV1 = DAG.getBitcast(MVT::v4f32, V1);
  SDValue CastV2 = DAG.getBitcast(MVT::v4f32, V2);
  SDValue ShufPS = DAG.getVectorShuffle(MVT::v4f32, DL, CastV1, CastV2, Mask);
  return DAG.getBitcast(MVT::v4i32, ShufPS);
}
 
/// Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
/// shuffle lowering, and the most complex part.
///
/// The lowering strategy is to try to form pairs of input lanes which are
/// targeted at the same half of the final vector, and then use a dword shuffle
/// to place them onto the right half, and finally unpack the paired lanes into
/// their final position.
///
/// The exact breakdown of how to form these dword pairs and align them on the
/// correct sides is really tricky. See the comments within the function for
/// more of the details.
///
/// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
/// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
/// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
/// vector, form the analogous 128-bit 8-element Mask.
static SDValue lowerV8I16GeneralSingleInputShuffle(
    const SDLoc &DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
    const X86Subtarget &Subtarget, SelectionDAG &DAG) {
  assert(VT.getVectorElementType() == MVT::i16 && "Bad input type!");
  MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
 
  assert(Mask.size() == 8 && "Shuffle mask length doesn't match!");
  MutableArrayRef<int> LoMask = Mask.slice(0, 4);
  MutableArrayRef<int> HiMask = Mask.slice(4, 4);
 
  // Attempt to directly match PSHUFLW or PSHUFHW.
  if (isUndefOrInRange(LoMask, 0, 4) &&
      isSequentialOrUndefInRange(HiMask, 0, 4, 4)) {
    return DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
                       getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
  }
  if (isUndefOrInRange(HiMask, 4, 8) &&
      isSequentialOrUndefInRange(LoMask, 0, 4, 0)) {
    for (int i = 0; i != 4; ++i)
      HiMask[i] = (HiMask[i] < 0 ? HiMask[i] : (HiMask[i] - 4));
    return DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
                       getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
  }
 
  SmallVector<int, 4> LoInputs;
  copy_if(LoMask, std::back_inserter(LoInputs), [](int M) { return M >= 0; });
  array_pod_sort(LoInputs.begin(), LoInputs.end());
  LoInputs.erase(llvm::unique(LoInputs), LoInputs.end());
  SmallVector<int, 4> HiInputs;
  copy_if(HiMask, std::back_inserter(HiInputs), [](int M) { return M >= 0; });
  array_pod_sort(HiInputs.begin(), HiInputs.end());
  HiInputs.erase(llvm::unique(HiInputs), HiInputs.end());
  int NumLToL = llvm::lower_bound(LoInputs, 4) - LoInputs.begin();
  int NumHToL = LoInputs.size() - NumLToL;
  int NumLToH = llvm::lower_bound(HiInputs, 4) - HiInputs.begin();
  int NumHToH = HiInputs.size() - NumLToH;
  MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
  MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
  MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
  MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
 
  // If we are shuffling values from one half - check how many different DWORD
  // pairs we need to create. If only 1 or 2 then we can perform this as a
  // PSHUFLW/PSHUFHW + PSHUFD instead of the PSHUFD+PSHUFLW+PSHUFHW chain below.
  auto ShuffleDWordPairs = [&](ArrayRef<int> PSHUFHalfMask,
                               ArrayRef<int> PSHUFDMask, unsigned ShufWOp) {
    V = DAG.getNode(ShufWOp, DL, VT, V,
                    getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
    V = DAG.getBitcast(PSHUFDVT, V);
    V = DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, V,
                    getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
    return DAG.getBitcast(VT, V);
  };
 
  if ((NumHToL + NumHToH) == 0 || (NumLToL + NumLToH) == 0) {
    int PSHUFDMask[4] = { -1, -1, -1, -1 };
    SmallVector<std::pair<int, int>, 4> DWordPairs;
    int DOffset = ((NumHToL + NumHToH) == 0 ? 0 : 2);
 
    // Collect the different DWORD pairs.
    for (int DWord = 0; DWord != 4; ++DWord) {
      int M0 = Mask[2 * DWord + 0];
      int M1 = Mask[2 * DWord + 1];
      M0 = (M0 >= 0 ? M0 % 4 : M0);
      M1 = (M1 >= 0 ? M1 % 4 : M1);
      if (M0 < 0 && M1 < 0)
        continue;
 
      bool Match = false;
      for (int j = 0, e = DWordPairs.size(); j < e; ++j) {
        auto &DWordPair = DWordPairs[j];
        if ((M0 < 0 || isUndefOrEqual(DWordPair.first, M0)) &&
            (M1 < 0 || isUndefOrEqual(DWordPair.second, M1))) {
          DWordPair.first = (M0 >= 0 ? M0 : DWordPair.first);
          DWordPair.second = (M1 >= 0 ? M1 : DWordPair.second);
          PSHUFDMask[DWord] = DOffset + j;
          Match = true;
          break;
        }
      }
      if (!Match) {
        PSHUFDMask[DWord] = DOffset + DWordPairs.size();
        DWordPairs.push_back(std::make_pair(M0, M1));
      }
    }
 
    if (DWordPairs.size() <= 2) {
      DWordPairs.resize(2, std::make_pair(-1, -1));
      int PSHUFHalfMask[4] = {DWordPairs[0].first, DWordPairs[0].second,
                              DWordPairs[1].first, DWordPairs[1].second};
      // For splat, ensure we widen the PSHUFDMask to allow vXi64 folds.
      if (ShuffleVectorSDNode::isSplatMask(PSHUFDMask) &&
          ShuffleVectorSDNode::isSplatMask(PSHUFHalfMask)) {
        int SplatIdx = ShuffleVectorSDNode::getSplatMaskIndex(PSHUFHalfMask);
        std::fill(PSHUFHalfMask, PSHUFHalfMask + 4, SplatIdx);
        PSHUFDMask[0] = PSHUFDMask[2] = DOffset + 0;
        PSHUFDMask[1] = PSHUFDMask[3] = DOffset + 1;
      }
      if ((NumHToL + NumHToH) == 0)
        return ShuffleDWordPairs(PSHUFHalfMask, PSHUFDMask, X86ISD::PSHUFLW);
      if ((NumLToL + NumLToH) == 0)
        return ShuffleDWordPairs(PSHUFHalfMask, PSHUFDMask, X86ISD::PSHUFHW);
    }
  }
 
  // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
  // such inputs we can swap two of the dwords across the half mark and end up
  // with <=2 inputs to each half in each half. Once there, we can fall through
  // to the generic code below. For example:
  //
  // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
  // Mask:  [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
  //
  // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
  // and an existing 2-into-2 on the other half. In this case we may have to
  // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
  // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
  // Fortunately, we don't have to handle anything but a 2-into-2 pattern
  // because any other situation (including a 3-into-1 or 1-into-3 in the other
  // half than the one we target for fixing) will be fixed when we re-enter this
  // path. We will also combine away any sequence of PSHUFD instructions that
  // result into a single instruction. Here is an example of the tricky case:
  //
  // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
  // Mask:  [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
  //
  // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
  //
  // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
  // Mask:  [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
  //
  // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
  // Mask:  [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
  //
  // The result is fine to be handled by the generic logic.
  auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
                          ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
                          int AOffset, int BOffset) {
    assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
           "Must call this with A having 3 or 1 inputs from the A half.");
    assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
           "Must call this with B having 1 or 3 inputs from the B half.");
    assert(AToAInputs.size() + BToAInputs.size() == 4 &&
           "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
 
    bool ThreeAInputs = AToAInputs.size() == 3;
 
    // Compute the index of dword with only one word among the three inputs in
    // a half by taking the sum of the half with three inputs and subtracting
    // the sum of the actual three inputs. The difference is the remaining
    // slot.
    int ADWord = 0, BDWord = 0;
    int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
    int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
    int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
    ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
    int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
    int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
    int TripleNonInputIdx =
        TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
    TripleDWord = TripleNonInputIdx / 2;
 
    // We use xor with one to compute the adjacent DWord to whichever one the
    // OneInput is in.
    OneInputDWord = (OneInput / 2) ^ 1;
 
    // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
    // and BToA inputs. If there is also such a problem with the BToB and AToB
    // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
    // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
    // is essential that we don't *create* a 3<-1 as then we might oscillate.
    if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
      // Compute how many inputs will be flipped by swapping these DWords. We
      // need
      // to balance this to ensure we don't form a 3-1 shuffle in the other
      // half.
      int NumFlippedAToBInputs = llvm::count(AToBInputs, 2 * ADWord) +
                                 llvm::count(AToBInputs, 2 * ADWord + 1);
      int NumFlippedBToBInputs = llvm::count(BToBInputs, 2 * BDWord) +
                                 llvm::count(BToBInputs, 2 * BDWord + 1);
      if ((NumFlippedAToBInputs == 1 &&
           (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
          (NumFlippedBToBInputs == 1 &&
           (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
        // We choose whether to fix the A half or B half based on whether that
        // half has zero flipped inputs. At zero, we may not be able to fix it
        // with that half. We also bias towards fixing the B half because that
        // will more commonly be the high half, and we have to bias one way.
        auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
                                                       ArrayRef<int> Inputs) {
          int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
          bool IsFixIdxInput = is_contained(Inputs, PinnedIdx ^ 1);
          // Determine whether the free index is in the flipped dword or the
          // unflipped dword based on where the pinned index is. We use this bit
          // in an xor to conditionally select the adjacent dword.
          int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
          bool IsFixFreeIdxInput = is_contained(Inputs, FixFreeIdx);
          if (IsFixIdxInput == IsFixFreeIdxInput)
            FixFreeIdx += 1;
          IsFixFreeIdxInput = is_contained(Inputs, FixFreeIdx);
          assert(IsFixIdxInput != IsFixFreeIdxInput &&
                 "We need to be changing the number of flipped inputs!");
          int PSHUFHalfMask[] = {0, 1, 2, 3};
          std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
          V = DAG.getNode(
              FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
              MVT::getVectorVT(MVT::i16, V.getValueSizeInBits() / 16), V,
              getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
 
          for (int &M : Mask)
            if (M >= 0 && M == FixIdx)
              M = FixFreeIdx;
            else if (M >= 0 && M == FixFreeIdx)
              M = FixIdx;
        };
        if (NumFlippedBToBInputs != 0) {
          int BPinnedIdx =
              BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
          FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
        } else {
          assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
          int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
          FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
        }
      }
    }
 
    int PSHUFDMask[] = {0, 1, 2, 3};
    PSHUFDMask[ADWord] = BDWord;
    PSHUFDMask[BDWord] = ADWord;
    V = DAG.getBitcast(
        VT,
        DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
                    getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
 
    // Adjust the mask to match the new locations of A and B.
    for (int &M : Mask)
      if (M >= 0 && M/2 == ADWord)
        M = 2 * BDWord + M % 2;
      else if (M >= 0 && M/2 == BDWord)
        M = 2 * ADWord + M % 2;
 
    // Recurse back into this routine to re-compute state now that this isn't
    // a 3 and 1 problem.
    return lowerV8I16GeneralSingleInputShuffle(DL, VT, V, Mask, Subtarget, DAG);
  };
  if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
    return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
  if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
    return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
 
  // At this point there are at most two inputs to the low and high halves from
  // each half. That means the inputs can always be grouped into dwords and
  // those dwords can then be moved to the correct half with a dword shuffle.
  // We use at most one low and one high word shuffle to collect these paired
  // inputs into dwords, and finally a dword shuffle to place them.
  int PSHUFLMask[4] = {-1, -1, -1, -1};
  int PSHUFHMask[4] = {-1, -1, -1, -1};
  int PSHUFDMask[4] = {-1, -1, -1, -1};
 
  // First fix the masks for all the inputs that are staying in their
  // original halves. This will then dictate the targets of the cross-half
  // shuffles.
  auto fixInPlaceInputs =
      [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
                    MutableArrayRef<int> SourceHalfMask,
                    MutableArrayRef<int> HalfMask, int HalfOffset) {
    if (InPlaceInputs.empty())
      return;
    if (InPlaceInputs.size() == 1) {
      SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
          InPlaceInputs[0] - HalfOffset;
      PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
      return;
    }
    if (IncomingInputs.empty()) {
      // Just fix all of the in place inputs.
      for (int Input : InPlaceInputs) {
        SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
        PSHUFDMask[Input / 2] = Input / 2;
      }
      return;
    }
 
    assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
    SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
        InPlaceInputs[0] - HalfOffset;
    // Put the second input next to the first so that they are packed into
    // a dword. We find the adjacent index by toggling the low bit.
    int AdjIndex = InPlaceInputs[0] ^ 1;
    SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
    llvm::replace(HalfMask, InPlaceInputs[1], AdjIndex);
    PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
  };
  fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
  fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
 
  // Now gather the cross-half inputs and place them into a free dword of
  // their target half.
  // FIXME: This operation could almost certainly be simplified dramatically to
  // look more like the 3-1 fixing operation.
  auto moveInputsToRightHalf = [&PSHUFDMask](
      MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
      MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
      MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
      int DestOffset) {
    auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
      return SourceHalfMask[Word] >= 0 && SourceHalfMask[Word] != Word;
    };
    auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
                                               int Word) {
      int LowWord = Word & ~1;
      int HighWord = Word | 1;
      return isWordClobbered(SourceHalfMask, LowWord) ||
             isWordClobbered(SourceHalfMask, HighWord);
    };
 
    if (IncomingInputs.empty())
      return;
 
    if (ExistingInputs.empty()) {
      // Map any dwords with inputs from them into the right half.
      for (int Input : IncomingInputs) {
        // If the source half mask maps over the inputs, turn those into
        // swaps and use the swapped lane.
        if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
          if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] < 0) {
            SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
                Input - SourceOffset;
            // We have to swap the uses in our half mask in one sweep.
            for (int &M : HalfMask)
              if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
                M = Input;
              else if (M == Input)
                M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
          } else {
            assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
                       Input - SourceOffset &&
                   "Previous placement doesn't match!");
          }
          // Note that this correctly re-maps both when we do a swap and when
          // we observe the other side of the swap above. We rely on that to
          // avoid swapping the members of the input list directly.
          Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
        }
 
        // Map the input's dword into the correct half.
        if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] < 0)
          PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
        else
          assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
                     Input / 2 &&
                 "Previous placement doesn't match!");
      }
 
      // And just directly shift any other-half mask elements to be same-half
      // as we will have mirrored the dword containing the element into the
      // same position within that half.
      for (int &M : HalfMask)
        if (M >= SourceOffset && M < SourceOffset + 4) {
          M = M - SourceOffset + DestOffset;
          assert(M >= 0 && "This should never wrap below zero!");
        }
      return;
    }
 
    // Ensure we have the input in a viable dword of its current half. This
    // is particularly tricky because the original position may be clobbered
    // by inputs being moved and *staying* in that half.
    if (IncomingInputs.size() == 1) {
      if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
        int InputFixed = find(SourceHalfMask, -1) - std::begin(SourceHalfMask) +
                         SourceOffset;
        SourceHalfMask[InputFixed - SourceOffset] =
            IncomingInputs[0] - SourceOffset;
        llvm::replace(HalfMask, IncomingInputs[0], InputFixed);
        IncomingInputs[0] = InputFixed;
      }
    } else if (IncomingInputs.size() == 2) {
      if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
          isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
        // We have two non-adjacent or clobbered inputs we need to extract from
        // the source half. To do this, we need to map them into some adjacent
        // dword slot in the source mask.
        int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
                              IncomingInputs[1] - SourceOffset};
 
        // If there is a free slot in the source half mask adjacent to one of
        // the inputs, place the other input in it. We use (Index XOR 1) to
        // compute an adjacent index.
        if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
            SourceHalfMask[InputsFixed[0] ^ 1] < 0) {
          SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
          SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
          InputsFixed[1] = InputsFixed[0] ^ 1;
        } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
                   SourceHalfMask[InputsFixed[1] ^ 1] < 0) {
          SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
          SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
          InputsFixed[0] = InputsFixed[1] ^ 1;
        } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] < 0 &&
                   SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] < 0) {
          // The two inputs are in the same DWord but it is clobbered and the
          // adjacent DWord isn't used at all. Move both inputs to the free
          // slot.
          SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
          SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
          InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
          InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
        } else {
          // The only way we hit this point is if there is no clobbering
          // (because there are no off-half inputs to this half) and there is no
          // free slot adjacent to one of the inputs. In this case, we have to
          // swap an input with a non-input.
          for (int i = 0; i < 4; ++i)
            assert((SourceHalfMask[i] < 0 || SourceHalfMask[i] == i) &&
                   "We can't handle any clobbers here!");
          assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
                 "Cannot have adjacent inputs here!");
 
          SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
          SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
 
          // We also have to update the final source mask in this case because
          // it may need to undo the above swap.
          for (int &M : FinalSourceHalfMask)
            if (M == (InputsFixed[0] ^ 1) + SourceOffset)
              M = InputsFixed[1] + SourceOffset;
            else if (M == InputsFixed[1] + SourceOffset)
              M = (InputsFixed[0] ^ 1) + SourceOffset;
 
          InputsFixed[1] = InputsFixed[0] ^ 1;
        }
 
        // Point everything at the fixed inputs.
        for (int &M : HalfMask)
          if (M == IncomingInputs[0])
            M = InputsFixed[0] + SourceOffset;
          else if (M == IncomingInputs[1])
            M = InputsFixed[1] + SourceOffset;
 
        IncomingInputs[0] = InputsFixed[0] + SourceOffset;
        IncomingInputs[1] = InputsFixed[1] + SourceOffset;
      }
    } else {
      llvm_unreachable("Unhandled input size!");
    }
 
    // Now hoist the DWord down to the right half.
    int FreeDWord = (PSHUFDMask[DestOffset / 2] < 0 ? 0 : 1) + DestOffset / 2;
    assert(PSHUFDMask[FreeDWord] < 0 && "DWord not free");
    PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
    for (int &M : HalfMask)
      for (int Input : IncomingInputs)
        if (M == Input)
          M = FreeDWord * 2 + Input % 2;
  };
  moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
                        /*SourceOffset*/ 4, /*DestOffset*/ 0);
  moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
                        /*SourceOffset*/ 0, /*DestOffset*/ 4);
 
  // Now enact all the shuffles we've computed to move the inputs into their
  // target half.
  if (!isNoopShuffleMask(PSHUFLMask))
    V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
                    getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
  if (!isNoopShuffleMask(PSHUFHMask))
    V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
                    getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
  if (!isNoopShuffleMask(PSHUFDMask))
    V = DAG.getBitcast(
        VT,
        DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
                    getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
 
  // At this point, each half should contain all its inputs, and we can then
  // just shuffle them into their final position.
  assert(none_of(LoMask, [](int M) { return M >= 4; }) &&
         "Failed to lift all the high half inputs to the low mask!");
  assert(none_of(HiMask, [](int M) { return M >= 0 && M < 4; }) &&
         "Failed to lift all the low half inputs to the high mask!");
 
  // Do a half shuffle for the low mask.
  if (!isNoopShuffleMask(LoMask))
    V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
                    getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
 
  // Do a half shuffle with the high mask after shifting its values down.
  for (int &M : HiMask)
    if (M >= 0)
      M -= 4;
  if (!isNoopShuffleMask(HiMask))
    V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
                    getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
 
  return V;
}
 
/// Helper to form a PSHUFB-based shuffle+blend, opportunistically avoiding the
/// blend if only one input is used.
static SDValue lowerShuffleAsBlendOfPSHUFBs(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    const APInt &Zeroable, SelectionDAG &DAG, bool &V1InUse, bool &V2InUse) {
  assert(!is128BitLaneCrossingShuffleMask(VT, Mask) &&
         "Lane crossing shuffle masks not supported");
 
  int NumBytes = VT.getSizeInBits() / 8;
  int Size = Mask.size();
  int Scale = NumBytes / Size;
 
  SmallVector<SDValue, 64> V1Mask(NumBytes, DAG.getUNDEF(MVT::i8));
  SmallVector<SDValue, 64> V2Mask(NumBytes, DAG.getUNDEF(MVT::i8));
  V1InUse = false;
  V2InUse = false;
 
  for (int i = 0; i < NumBytes; ++i) {
    int M = Mask[i / Scale];
    if (M < 0)
      continue;
 
    const int ZeroMask = 0x80;
    int V1Idx = M < Size ? M * Scale + i % Scale : ZeroMask;
    int V2Idx = M < Size ? ZeroMask : (M - Size) * Scale + i % Scale;
    if (Zeroable[i / Scale])
      V1Idx = V2Idx = ZeroMask;
 
    V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
    V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
    V1InUse |= (ZeroMask != V1Idx);
    V2InUse |= (ZeroMask != V2Idx);
  }
 
  MVT ShufVT = MVT::getVectorVT(MVT::i8, NumBytes);
  if (V1InUse)
    V1 = DAG.getNode(X86ISD::PSHUFB, DL, ShufVT, DAG.getBitcast(ShufVT, V1),
                     DAG.getBuildVector(ShufVT, DL, V1Mask));
  if (V2InUse)
    V2 = DAG.getNode(X86ISD::PSHUFB, DL, ShufVT, DAG.getBitcast(ShufVT, V2),
                     DAG.getBuildVector(ShufVT, DL, V2Mask));
 
  // If we need shuffled inputs from both, blend the two.
  SDValue V;
  if (V1InUse && V2InUse)
    V = DAG.getNode(ISD::OR, DL, ShufVT, V1, V2);
  else
    V = V1InUse ? V1 : V2;
 
  // Cast the result back to the correct type.
  return DAG.getBitcast(VT, V);
}
 
/// Generic lowering of 8-lane i16 shuffles.
///
/// This handles both single-input shuffles and combined shuffle/blends with
/// two inputs. The single input shuffles are immediately delegated to
/// a dedicated lowering routine.
///
/// The blends are lowered in one of three fundamental ways. If there are few
/// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
/// of the input is significantly cheaper when lowered as an interleaving of
/// the two inputs, try to interleave them. Otherwise, blend the low and high
/// halves of the inputs separately (making them have relatively few inputs)
/// and then concatenate them.
static SDValue lowerV8I16Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
  assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v8i16, V1, V2, Mask,
                                                   Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Try to use lower using a truncation.
  if (SDValue V = lowerShuffleWithVPMOV(DL, MVT::v8i16, V1, V2, Mask, Zeroable,
                                        Subtarget, DAG))
    return V;
 
  int NumV2Inputs = count_if(Mask, [](int M) { return M >= 8; });
 
  if (NumV2Inputs == 0) {
    // Try to use shift instructions.
    if (SDValue Shift =
            lowerShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, Zeroable,
                                Subtarget, DAG, /*BitwiseOnly*/ false))
      return Shift;
 
    // Check for being able to broadcast a single element.
    if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8i16, V1, V2,
                                                    Mask, Subtarget, DAG))
      return Broadcast;
 
    // Try to use bit rotation instructions.
    if (SDValue Rotate = lowerShuffleAsBitRotate(DL, MVT::v8i16, V1, Mask,
                                                 Subtarget, DAG))
      return Rotate;
 
    // Use dedicated unpack instructions for masks that match their pattern.
    if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8i16, V1, V2, Mask, DAG))
      return V;
 
    // Use dedicated pack instructions for masks that match their pattern.
    if (SDValue V =
            lowerShuffleWithPACK(DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
      return V;
 
    // Try to use byte rotation instructions.
    if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i16, V1, V1, Mask,
                                                  Subtarget, DAG))
      return Rotate;
 
    // Make a copy of the mask so it can be modified.
    SmallVector<int, 8> MutableMask(Mask);
    return lowerV8I16GeneralSingleInputShuffle(DL, MVT::v8i16, V1, MutableMask,
                                               Subtarget, DAG);
  }
 
  assert(llvm::any_of(Mask, [](int M) { return M >= 0 && M < 8; }) &&
         "All single-input shuffles should be canonicalized to be V1-input "
         "shuffles.");
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // See if we can use SSE4A Extraction / Insertion.
  if (Subtarget.hasSSE4A())
    if (SDValue V = lowerShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask,
                                          Zeroable, DAG))
      return V;
 
  // There are special ways we can lower some single-element blends.
  if (NumV2Inputs == 1)
    if (SDValue V = lowerShuffleAsElementInsertion(
            DL, MVT::v8i16, V1, V2, Mask, Zeroable, Subtarget, DAG))
      return V;
 
  // We have different paths for blend lowering, but they all must use the
  // *exact* same predicate.
  bool IsBlendSupported = Subtarget.hasSSE41();
  if (IsBlendSupported)
    if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
      return Blend;
 
  if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask,
                                             Zeroable, Subtarget, DAG))
    return Masked;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8i16, V1, V2, Mask, DAG))
    return V;
 
  // Use dedicated pack instructions for masks that match their pattern.
  if (SDValue V =
          lowerShuffleWithPACK(DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to use lower using a truncation.
  if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v8i16, V1, V2, Mask, Zeroable,
                                       Subtarget, DAG))
    return V;
 
  // Try to use byte rotation instructions.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i16, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  if (SDValue BitBlend =
          lowerShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
    return BitBlend;
 
  // Try to use byte shift instructions to mask.
  if (SDValue V = lowerShuffleAsByteShiftMask(DL, MVT::v8i16, V1, V2, Mask,
                                              Zeroable, Subtarget, DAG))
    return V;
 
  // Attempt to lower using compaction, SSE41 is necessary for PACKUSDW.
  int NumEvenDrops = canLowerByDroppingElements(Mask, true, false);
  if ((NumEvenDrops == 1 || (NumEvenDrops == 2 && Subtarget.hasSSE41())) &&
      !Subtarget.hasVLX()) {
    // Check if this is part of a 256-bit vector truncation.
    unsigned PackOpc = 0;
    if (NumEvenDrops == 2 && Subtarget.hasAVX2() &&
        peekThroughBitcasts(V1).getOpcode() == ISD::EXTRACT_SUBVECTOR &&
        peekThroughBitcasts(V2).getOpcode() == ISD::EXTRACT_SUBVECTOR) {
      SDValue V1V2 = concatSubVectors(V1, V2, DAG, DL);
      V1V2 = DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1V2,
                         getZeroVector(MVT::v16i16, Subtarget, DAG, DL),
                         DAG.getTargetConstant(0xEE, DL, MVT::i8));
      V1V2 = DAG.getBitcast(MVT::v8i32, V1V2);
      V1 = extract128BitVector(V1V2, 0, DAG, DL);
      V2 = extract128BitVector(V1V2, 4, DAG, DL);
      PackOpc = X86ISD::PACKUS;
    } else if (Subtarget.hasSSE41()) {
      SmallVector<SDValue, 4> DWordClearOps(4,
                                            DAG.getConstant(0, DL, MVT::i32));
      for (unsigned i = 0; i != 4; i += 1 << (NumEvenDrops - 1))
        DWordClearOps[i] = DAG.getConstant(0xFFFF, DL, MVT::i32);
      SDValue DWordClearMask =
          DAG.getBuildVector(MVT::v4i32, DL, DWordClearOps);
      V1 = DAG.getNode(ISD::AND, DL, MVT::v4i32, DAG.getBitcast(MVT::v4i32, V1),
                       DWordClearMask);
      V2 = DAG.getNode(ISD::AND, DL, MVT::v4i32, DAG.getBitcast(MVT::v4i32, V2),
                       DWordClearMask);
      PackOpc = X86ISD::PACKUS;
    } else if (!Subtarget.hasSSSE3()) {
      SDValue ShAmt = DAG.getTargetConstant(16, DL, MVT::i8);
      V1 = DAG.getBitcast(MVT::v4i32, V1);
      V2 = DAG.getBitcast(MVT::v4i32, V2);
      V1 = DAG.getNode(X86ISD::VSHLI, DL, MVT::v4i32, V1, ShAmt);
      V2 = DAG.getNode(X86ISD::VSHLI, DL, MVT::v4i32, V2, ShAmt);
      V1 = DAG.getNode(X86ISD::VSRAI, DL, MVT::v4i32, V1, ShAmt);
      V2 = DAG.getNode(X86ISD::VSRAI, DL, MVT::v4i32, V2, ShAmt);
      PackOpc = X86ISD::PACKSS;
    }
    if (PackOpc) {
      // Now pack things back together.
      SDValue Result = DAG.getNode(PackOpc, DL, MVT::v8i16, V1, V2);
      if (NumEvenDrops == 2) {
        Result = DAG.getBitcast(MVT::v4i32, Result);
        Result = DAG.getNode(PackOpc, DL, MVT::v8i16, Result, Result);
      }
      return Result;
    }
  }
 
  // When compacting odd (upper) elements, use PACKSS pre-SSE41.
  int NumOddDrops = canLowerByDroppingElements(Mask, false, false);
  if (NumOddDrops == 1) {
    bool HasSSE41 = Subtarget.hasSSE41();
    V1 = DAG.getNode(HasSSE41 ? X86ISD::VSRLI : X86ISD::VSRAI, DL, MVT::v4i32,
                     DAG.getBitcast(MVT::v4i32, V1),
                     DAG.getTargetConstant(16, DL, MVT::i8));
    V2 = DAG.getNode(HasSSE41 ? X86ISD::VSRLI : X86ISD::VSRAI, DL, MVT::v4i32,
                     DAG.getBitcast(MVT::v4i32, V2),
                     DAG.getTargetConstant(16, DL, MVT::i8));
    return DAG.getNode(HasSSE41 ? X86ISD::PACKUS : X86ISD::PACKSS, DL,
                       MVT::v8i16, V1, V2);
  }
 
  // Try to lower by permuting the inputs into an unpack instruction.
  if (SDValue Unpack = lowerShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1, V2,
                                                      Mask, Subtarget, DAG))
    return Unpack;
 
  // If we can't directly blend but can use PSHUFB, that will be better as it
  // can both shuffle and set up the inefficient blend.
  if (!IsBlendSupported && Subtarget.hasSSSE3()) {
    bool V1InUse, V2InUse;
    return lowerShuffleAsBlendOfPSHUFBs(DL, MVT::v8i16, V1, V2, Mask,
                                        Zeroable, DAG, V1InUse, V2InUse);
  }
 
  // We can always bit-blend if we have to so the fallback strategy is to
  // decompose into single-input permutes and blends/unpacks.
  return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v8i16, V1, V2, Mask,
                                              Zeroable, Subtarget, DAG);
}
 
/// Lower 8-lane 16-bit floating point shuffles.
static SDValue lowerV8F16Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v8f16 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v8f16 && "Bad operand type!");
  assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
  int NumV2Elements = count_if(Mask, [](int M) { return M >= 8; });
 
  if (Subtarget.hasFP16()) {
    if (NumV2Elements == 0) {
      // Check for being able to broadcast a single element.
      if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8f16, V1, V2,
                                                      Mask, Subtarget, DAG))
        return Broadcast;
    }
    if (NumV2Elements == 1 && Mask[0] >= 8)
      if (SDValue V = lowerShuffleAsElementInsertion(
              DL, MVT::v8f16, V1, V2, Mask, Zeroable, Subtarget, DAG))
        return V;
  }
 
  V1 = DAG.getBitcast(MVT::v8i16, V1);
  V2 = DAG.getBitcast(MVT::v8i16, V2);
  return DAG.getBitcast(MVT::v8f16,
                        DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, Mask));
}
 
// Lowers unary/binary shuffle as VPERMV/VPERMV3, for non-VLX targets,
// sub-512-bit shuffles are padded to 512-bits for the shuffle and then
// the active subvector is extracted.
static SDValue lowerShuffleWithPERMV(const SDLoc &DL, MVT VT,
                                     ArrayRef<int> OriginalMask, SDValue V1,
                                     SDValue V2, const X86Subtarget &Subtarget,
                                     SelectionDAG &DAG) {
  // Commute binary inputs so V2 is a load to simplify VPERMI2/T2 folds.
  SmallVector<int, 32> Mask(OriginalMask);
  if (!V2.isUndef() && isShuffleFoldableLoad(V1) &&
      !isShuffleFoldableLoad(V2)) {
    ShuffleVectorSDNode::commuteMask(Mask);
    std::swap(V1, V2);
  }
 
  MVT MaskVT = VT.changeTypeToInteger();
  SDValue MaskNode;
  MVT ShuffleVT = VT;
  if (!VT.is512BitVector() && !Subtarget.hasVLX()) {
    V1 = widenSubVector(V1, false, Subtarget, DAG, DL, 512);
    V2 = widenSubVector(V2, false, Subtarget, DAG, DL, 512);
    ShuffleVT = V1.getSimpleValueType();
 
    // Adjust mask to correct indices for the second input.
    int NumElts = VT.getVectorNumElements();
    unsigned Scale = 512 / VT.getSizeInBits();
    SmallVector<int, 32> AdjustedMask(Mask);
    for (int &M : AdjustedMask)
      if (NumElts <= M)
        M += (Scale - 1) * NumElts;
    MaskNode = getConstVector(AdjustedMask, MaskVT, DAG, DL, true);
    MaskNode = widenSubVector(MaskNode, false, Subtarget, DAG, DL, 512);
  } else {
    MaskNode = getConstVector(Mask, MaskVT, DAG, DL, true);
  }
 
  SDValue Result;
  if (V2.isUndef())
    Result = DAG.getNode(X86ISD::VPERMV, DL, ShuffleVT, MaskNode, V1);
  else
    Result = DAG.getNode(X86ISD::VPERMV3, DL, ShuffleVT, V1, MaskNode, V2);
 
  if (VT != ShuffleVT)
    Result = extractSubVector(Result, 0, DAG, DL, VT.getSizeInBits());
 
  return Result;
}
 
/// Generic lowering of v16i8 shuffles.
///
/// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
/// detect any complexity reducing interleaving. If that doesn't help, it uses
/// UNPCK to spread the i8 elements across two i16-element vectors, and uses
/// the existing lowering for v8i16 blends on each half, finally PACK-ing them
/// back together.
static SDValue lowerV16I8Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
  assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // Try to use byte rotation instructions.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v16i8, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  // Use dedicated pack instructions for masks that match their pattern.
  if (SDValue V =
          lowerShuffleWithPACK(DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to use a zext lowering.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v16i8, V1, V2, Mask,
                                                   Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Try to use lower using a truncation.
  if (SDValue V = lowerShuffleWithVPMOV(DL, MVT::v16i8, V1, V2, Mask, Zeroable,
                                        Subtarget, DAG))
    return V;
 
  if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v16i8, V1, V2, Mask, Zeroable,
                                       Subtarget, DAG))
    return V;
 
  // See if we can use SSE4A Extraction / Insertion.
  if (Subtarget.hasSSE4A())
    if (SDValue V = lowerShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask,
                                          Zeroable, DAG))
      return V;
 
  int NumV2Elements = count_if(Mask, [](int M) { return M >= 16; });
 
  // For single-input shuffles, there are some nicer lowering tricks we can use.
  if (NumV2Elements == 0) {
    // Check for being able to broadcast a single element.
    if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v16i8, V1, V2,
                                                    Mask, Subtarget, DAG))
      return Broadcast;
 
    // Try to use bit rotation instructions.
    if (SDValue Rotate = lowerShuffleAsBitRotate(DL, MVT::v16i8, V1, Mask,
                                                 Subtarget, DAG))
      return Rotate;
 
    if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i8, V1, V2, Mask, DAG))
      return V;
 
    // Check whether we can widen this to an i16 shuffle by duplicating bytes.
    // Notably, this handles splat and partial-splat shuffles more efficiently.
    // However, it only makes sense if the pre-duplication shuffle simplifies
    // things significantly. Currently, this means we need to be able to
    // express the pre-duplication shuffle as an i16 shuffle.
    //
    // FIXME: We should check for other patterns which can be widened into an
    // i16 shuffle as well.
    auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
      for (int i = 0; i < 16; i += 2)
        if (Mask[i] >= 0 && Mask[i + 1] >= 0 && Mask[i] != Mask[i + 1])
          return false;
 
      return true;
    };
    auto tryToWidenViaDuplication = [&]() -> SDValue {
      if (!canWidenViaDuplication(Mask))
        return SDValue();
      SmallVector<int, 4> LoInputs;
      copy_if(Mask, std::back_inserter(LoInputs),
              [](int M) { return M >= 0 && M < 8; });
      array_pod_sort(LoInputs.begin(), LoInputs.end());
      LoInputs.erase(llvm::unique(LoInputs), LoInputs.end());
      SmallVector<int, 4> HiInputs;
      copy_if(Mask, std::back_inserter(HiInputs), [](int M) { return M >= 8; });
      array_pod_sort(HiInputs.begin(), HiInputs.end());
      HiInputs.erase(llvm::unique(HiInputs), HiInputs.end());
 
      bool TargetLo = LoInputs.size() >= HiInputs.size();
      ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
      ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
 
      int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
      SmallDenseMap<int, int, 8> LaneMap;
      for (int I : InPlaceInputs) {
        PreDupI16Shuffle[I/2] = I/2;
        LaneMap[I] = I;
      }
      int j = TargetLo ? 0 : 4, je = j + 4;
      for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
        // Check if j is already a shuffle of this input. This happens when
        // there are two adjacent bytes after we move the low one.
        if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
          // If we haven't yet mapped the input, search for a slot into which
          // we can map it.
          while (j < je && PreDupI16Shuffle[j] >= 0)
            ++j;
 
          if (j == je)
            // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
            return SDValue();
 
          // Map this input with the i16 shuffle.
          PreDupI16Shuffle[j] = MovingInputs[i] / 2;
        }
 
        // Update the lane map based on the mapping we ended up with.
        LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
      }
      V1 = DAG.getBitcast(
          MVT::v16i8,
          DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
                               DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
 
      // Unpack the bytes to form the i16s that will be shuffled into place.
      bool EvenInUse = false, OddInUse = false;
      for (int i = 0; i < 16; i += 2) {
        EvenInUse |= (Mask[i + 0] >= 0);
        OddInUse |= (Mask[i + 1] >= 0);
        if (EvenInUse && OddInUse)
          break;
      }
      V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
                       MVT::v16i8, EvenInUse ? V1 : DAG.getUNDEF(MVT::v16i8),
                       OddInUse ? V1 : DAG.getUNDEF(MVT::v16i8));
 
      int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
      for (int i = 0; i < 16; ++i)
        if (Mask[i] >= 0) {
          int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
          assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
          if (PostDupI16Shuffle[i / 2] < 0)
            PostDupI16Shuffle[i / 2] = MappedMask;
          else
            assert(PostDupI16Shuffle[i / 2] == MappedMask &&
                   "Conflicting entries in the original shuffle!");
        }
      return DAG.getBitcast(
          MVT::v16i8,
          DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
                               DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
    };
    if (SDValue V = tryToWidenViaDuplication())
      return V;
  }
 
  if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask,
                                             Zeroable, Subtarget, DAG))
    return Masked;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i8, V1, V2, Mask, DAG))
    return V;
 
  // Try to use byte shift instructions to mask.
  if (SDValue V = lowerShuffleAsByteShiftMask(DL, MVT::v16i8, V1, V2, Mask,
                                              Zeroable, Subtarget, DAG))
    return V;
 
  // Check for compaction patterns.
  bool IsSingleInput = V2.isUndef();
  int NumEvenDrops = canLowerByDroppingElements(Mask, true, IsSingleInput);
 
  // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
  // with PSHUFB. It is important to do this before we attempt to generate any
  // blends but after all of the single-input lowerings. If the single input
  // lowerings can find an instruction sequence that is faster than a PSHUFB, we
  // want to preserve that and we can DAG combine any longer sequences into
  // a PSHUFB in the end. But once we start blending from multiple inputs,
  // the complexity of DAG combining bad patterns back into PSHUFB is too high,
  // and there are *very* few patterns that would actually be faster than the
  // PSHUFB approach because of its ability to zero lanes.
  //
  // If the mask is a binary compaction, we can more efficiently perform this
  // as a PACKUS(AND(),AND()) - which is quicker than UNPACK(PSHUFB(),PSHUFB()).
  //
  // FIXME: The only exceptions to the above are blends which are exact
  // interleavings with direct instructions supporting them. We currently don't
  // handle those well here.
  if (Subtarget.hasSSSE3() && (IsSingleInput || NumEvenDrops != 1)) {
    bool V1InUse = false;
    bool V2InUse = false;
 
    SDValue PSHUFB = lowerShuffleAsBlendOfPSHUFBs(
        DL, MVT::v16i8, V1, V2, Mask, Zeroable, DAG, V1InUse, V2InUse);
 
    // If both V1 and V2 are in use and we can use a direct blend or an unpack,
    // do so. This avoids using them to handle blends-with-zero which is
    // important as a single pshufb is significantly faster for that.
    if (V1InUse && V2InUse) {
      if (Subtarget.hasSSE41())
        if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16i8, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG))
          return Blend;
 
      // We can use an unpack to do the blending rather than an or in some
      // cases. Even though the or may be (very minorly) more efficient, we
      // preference this lowering because there are common cases where part of
      // the complexity of the shuffles goes away when we do the final blend as
      // an unpack.
      // FIXME: It might be worth trying to detect if the unpack-feeding
      // shuffles will both be pshufb, in which case we shouldn't bother with
      // this.
      if (SDValue Unpack = lowerShuffleAsPermuteAndUnpack(
              DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
        return Unpack;
 
      // AVX512VBMI can lower to VPERMB (non-VLX will pad to v64i8).
      if (Subtarget.hasVBMI())
        return lowerShuffleWithPERMV(DL, MVT::v16i8, Mask, V1, V2, Subtarget,
                                     DAG);
 
      // If we have XOP we can use one VPPERM instead of multiple PSHUFBs.
      if (Subtarget.hasXOP()) {
        SDValue MaskNode = getConstVector(Mask, MVT::v16i8, DAG, DL, true);
        return DAG.getNode(X86ISD::VPPERM, DL, MVT::v16i8, V1, V2, MaskNode);
      }
 
      // Use PALIGNR+Permute if possible - permute might become PSHUFB but the
      // PALIGNR will be cheaper than the second PSHUFB+OR.
      if (SDValue V = lowerShuffleAsByteRotateAndPermute(
              DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
        return V;
    }
 
    return PSHUFB;
  }
 
  // There are special ways we can lower some single-element blends.
  if (NumV2Elements == 1)
    if (SDValue V = lowerShuffleAsElementInsertion(
            DL, MVT::v16i8, V1, V2, Mask, Zeroable, Subtarget, DAG))
      return V;
 
  if (SDValue Blend = lowerShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
    return Blend;
 
  // Check whether a compaction lowering can be done. This handles shuffles
  // which take every Nth element for some even N. See the helper function for
  // details.
  //
  // We special case these as they can be particularly efficiently handled with
  // the PACKUSB instruction on x86 and they show up in common patterns of
  // rearranging bytes to truncate wide elements.
  if (NumEvenDrops) {
    // NumEvenDrops is the power of two stride of the elements. Another way of
    // thinking about it is that we need to drop the even elements this many
    // times to get the original input.
 
    // First we need to zero all the dropped bytes.
    assert(NumEvenDrops <= 3 &&
           "No support for dropping even elements more than 3 times.");
    SmallVector<SDValue, 8> WordClearOps(8, DAG.getConstant(0, DL, MVT::i16));
    for (unsigned i = 0; i != 8; i += 1 << (NumEvenDrops - 1))
      WordClearOps[i] = DAG.getConstant(0xFF, DL, MVT::i16);
    SDValue WordClearMask = DAG.getBuildVector(MVT::v8i16, DL, WordClearOps);
    V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, DAG.getBitcast(MVT::v8i16, V1),
                     WordClearMask);
    if (!IsSingleInput)
      V2 = DAG.getNode(ISD::AND, DL, MVT::v8i16, DAG.getBitcast(MVT::v8i16, V2),
                       WordClearMask);
 
    // Now pack things back together.
    SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1,
                                 IsSingleInput ? V1 : V2);
    for (int i = 1; i < NumEvenDrops; ++i) {
      Result = DAG.getBitcast(MVT::v8i16, Result);
      Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
    }
    return Result;
  }
 
  int NumOddDrops = canLowerByDroppingElements(Mask, false, IsSingleInput);
  if (NumOddDrops == 1) {
    V1 = DAG.getNode(X86ISD::VSRLI, DL, MVT::v8i16,
                     DAG.getBitcast(MVT::v8i16, V1),
                     DAG.getTargetConstant(8, DL, MVT::i8));
    if (!IsSingleInput)
      V2 = DAG.getNode(X86ISD::VSRLI, DL, MVT::v8i16,
                       DAG.getBitcast(MVT::v8i16, V2),
                       DAG.getTargetConstant(8, DL, MVT::i8));
    return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1,
                       IsSingleInput ? V1 : V2);
  }
 
  // Handle multi-input cases by blending/unpacking single-input shuffles.
  if (NumV2Elements > 0)
    return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v16i8, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG);
 
  // The fallback path for single-input shuffles widens this into two v8i16
  // vectors with unpacks, shuffles those, and then pulls them back together
  // with a pack.
  SDValue V = V1;
 
  std::array<int, 8> LoBlendMask = {{-1, -1, -1, -1, -1, -1, -1, -1}};
  std::array<int, 8> HiBlendMask = {{-1, -1, -1, -1, -1, -1, -1, -1}};
  for (int i = 0; i < 16; ++i)
    if (Mask[i] >= 0)
      (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
 
  SDValue VLoHalf, VHiHalf;
  // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
  // them out and avoid using UNPCK{L,H} to extract the elements of V as
  // i16s.
  if (none_of(LoBlendMask, [](int M) { return M >= 0 && M % 2 == 1; }) &&
      none_of(HiBlendMask, [](int M) { return M >= 0 && M % 2 == 1; })) {
    // Use a mask to drop the high bytes.
    VLoHalf = DAG.getBitcast(MVT::v8i16, V);
    VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
                          DAG.getConstant(0x00FF, DL, MVT::v8i16));
 
    // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
    VHiHalf = DAG.getUNDEF(MVT::v8i16);
 
    // Squash the masks to point directly into VLoHalf.
    for (int &M : LoBlendMask)
      if (M >= 0)
        M /= 2;
    for (int &M : HiBlendMask)
      if (M >= 0)
        M /= 2;
  } else {
    // Otherwise just unpack the low half of V into VLoHalf and the high half into
    // VHiHalf so that we can blend them as i16s.
    SDValue Zero = getZeroVector(MVT::v16i8, Subtarget, DAG, DL);
 
    VLoHalf = DAG.getBitcast(
        MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
    VHiHalf = DAG.getBitcast(
        MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
  }
 
  SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
  SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
 
  return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
}
 
/// Dispatching routine to lower various 128-bit x86 vector shuffles.
///
/// This routine breaks down the specific type of 128-bit shuffle and
/// dispatches to the lowering routines accordingly.
static SDValue lower128BitShuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                  MVT VT, SDValue V1, SDValue V2,
                                  const APInt &Zeroable,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  if (VT == MVT::v8bf16) {
    V1 = DAG.getBitcast(MVT::v8i16, V1);
    V2 = DAG.getBitcast(MVT::v8i16, V2);
    return DAG.getBitcast(VT,
                          DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, Mask));
  }
 
  switch (VT.SimpleTy) {
  case MVT::v2i64:
    return lowerV2I64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v2f64:
    return lowerV2F64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v4i32:
    return lowerV4I32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v4f32:
    return lowerV4F32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v8i16:
    return lowerV8I16Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v8f16:
    return lowerV8F16Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v16i8:
    return lowerV16I8Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
 
  default:
    llvm_unreachable("Unimplemented!");
  }
}
 
/// Generic routine to split vector shuffle into half-sized shuffles.
///
/// This routine just extracts two subvectors, shuffles them independently, and
/// then concatenates them back together. This should work effectively with all
/// AVX vector shuffle types.
static SDValue splitAndLowerShuffle(const SDLoc &DL, MVT VT, SDValue V1,
                                    SDValue V2, ArrayRef<int> Mask,
                                    SelectionDAG &DAG, bool SimpleOnly) {
  assert(VT.getSizeInBits() >= 256 &&
         "Only for 256-bit or wider vector shuffles!");
  assert(V1.getSimpleValueType() == VT && "Bad operand type!");
  assert(V2.getSimpleValueType() == VT && "Bad operand type!");
 
  // If this came from the AVX1 v8i32 -> v8f32 bitcast, split using v4i32.
  if (VT == MVT::v8f32) {
    SDValue BC1 = peekThroughBitcasts(V1);
    SDValue BC2 = peekThroughBitcasts(V2);
    if (BC1.getValueType() == MVT::v8i32 && BC2.getValueType() == MVT::v8i32) {
      if (SDValue Split = splitAndLowerShuffle(DL, MVT::v8i32, BC1, BC2, Mask,
                                               DAG, SimpleOnly))
        return DAG.getBitcast(VT, Split);
    }
  }
 
  ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
  ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
 
  int NumElements = VT.getVectorNumElements();
  int SplitNumElements = NumElements / 2;
  MVT ScalarVT = VT.getVectorElementType();
  MVT SplitVT = MVT::getVectorVT(ScalarVT, SplitNumElements);
 
  // Use splitVector/extractSubVector so that split build-vectors just build two
  // narrower build vectors. This helps shuffling with splats and zeros.
  auto SplitVector = [&](SDValue V) {
    SDValue LoV, HiV;
    std::tie(LoV, HiV) = splitVector(peekThroughBitcasts(V), DAG, DL);
    return std::make_pair(DAG.getBitcast(SplitVT, LoV),
                          DAG.getBitcast(SplitVT, HiV));
  };
 
  SDValue LoV1, HiV1, LoV2, HiV2;
  std::tie(LoV1, HiV1) = SplitVector(V1);
  std::tie(LoV2, HiV2) = SplitVector(V2);
 
  // Now create two 4-way blends of these half-width vectors.
  auto GetHalfBlendPiecesReq = [&](const ArrayRef<int> &HalfMask, bool &UseLoV1,
                                   bool &UseHiV1, bool &UseLoV2,
                                   bool &UseHiV2) {
    UseLoV1 = UseHiV1 = UseLoV2 = UseHiV2 = false;
    for (int i = 0; i < SplitNumElements; ++i) {
      int M = HalfMask[i];
      if (M >= NumElements) {
        if (M >= NumElements + SplitNumElements)
          UseHiV2 = true;
        else
          UseLoV2 = true;
      } else if (M >= 0) {
        if (M >= SplitNumElements)
          UseHiV1 = true;
        else
          UseLoV1 = true;
      }
    }
  };
 
  auto CheckHalfBlendUsable = [&](const ArrayRef<int> &HalfMask) -> bool {
    if (!SimpleOnly)
      return true;
 
    bool UseLoV1, UseHiV1, UseLoV2, UseHiV2;
    GetHalfBlendPiecesReq(HalfMask, UseLoV1, UseHiV1, UseLoV2, UseHiV2);
 
    return !(UseHiV1 || UseHiV2);
  };
 
  auto HalfBlend = [&](ArrayRef<int> HalfMask) {
    SmallVector<int, 32> V1BlendMask((unsigned)SplitNumElements, -1);
    SmallVector<int, 32> V2BlendMask((unsigned)SplitNumElements, -1);
    SmallVector<int, 32> BlendMask((unsigned)SplitNumElements, -1);
    for (int i = 0; i < SplitNumElements; ++i) {
      int M = HalfMask[i];
      if (M >= NumElements) {
        V2BlendMask[i] = M - NumElements;
        BlendMask[i] = SplitNumElements + i;
      } else if (M >= 0) {
        V1BlendMask[i] = M;
        BlendMask[i] = i;
      }
    }
 
    bool UseLoV1, UseHiV1, UseLoV2, UseHiV2;
    GetHalfBlendPiecesReq(HalfMask, UseLoV1, UseHiV1, UseLoV2, UseHiV2);
 
    // Because the lowering happens after all combining takes place, we need to
    // manually combine these blend masks as much as possible so that we create
    // a minimal number of high-level vector shuffle nodes.
    assert((!SimpleOnly || (!UseHiV1 && !UseHiV2)) && "Shuffle isn't simple");
 
    // First try just blending the halves of V1 or V2.
    if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
      return DAG.getUNDEF(SplitVT);
    if (!UseLoV2 && !UseHiV2)
      return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
    if (!UseLoV1 && !UseHiV1)
      return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
 
    SDValue V1Blend, V2Blend;
    if (UseLoV1 && UseHiV1) {
      V1Blend = DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
    } else {
      // We only use half of V1 so map the usage down into the final blend mask.
      V1Blend = UseLoV1 ? LoV1 : HiV1;
      for (int i = 0; i < SplitNumElements; ++i)
        if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
          BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
    }
    if (UseLoV2 && UseHiV2) {
      V2Blend = DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
    } else {
      // We only use half of V2 so map the usage down into the final blend mask.
      V2Blend = UseLoV2 ? LoV2 : HiV2;
      for (int i = 0; i < SplitNumElements; ++i)
        if (BlendMask[i] >= SplitNumElements)
          BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
    }
    return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
  };
 
  if (!CheckHalfBlendUsable(LoMask) || !CheckHalfBlendUsable(HiMask))
    return SDValue();
 
  SDValue Lo = HalfBlend(LoMask);
  SDValue Hi = HalfBlend(HiMask);
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
}
 
/// Either split a vector in halves or decompose the shuffles and the
/// blend/unpack.
///
/// This is provided as a good fallback for many lowerings of non-single-input
/// shuffles with more than one 128-bit lane. In those cases, we want to select
/// between splitting the shuffle into 128-bit components and stitching those
/// back together vs. extracting the single-input shuffles and blending those
/// results.
static SDValue lowerShuffleAsSplitOrBlend(const SDLoc &DL, MVT VT, SDValue V1,
                                          SDValue V2, ArrayRef<int> Mask,
                                          const APInt &Zeroable,
                                          const X86Subtarget &Subtarget,
                                          SelectionDAG &DAG) {
  assert(!V2.isUndef() && "This routine must not be used to lower single-input "
         "shuffles as it could then recurse on itself.");
  int Size = Mask.size();
 
  // If this can be modeled as a broadcast of two elements followed by a blend,
  // prefer that lowering. This is especially important because broadcasts can
  // often fold with memory operands.
  auto DoBothBroadcast = [&] {
    int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
    for (int M : Mask)
      if (M >= Size) {
        if (V2BroadcastIdx < 0)
          V2BroadcastIdx = M - Size;
        else if ((M - Size) != V2BroadcastIdx &&
                 !IsElementEquivalent(Size, V2, V2, M - Size, V2BroadcastIdx))
          return false;
      } else if (M >= 0) {
        if (V1BroadcastIdx < 0)
          V1BroadcastIdx = M;
        else if (M != V1BroadcastIdx &&
                 !IsElementEquivalent(Size, V1, V1, M, V1BroadcastIdx))
          return false;
      }
    return true;
  };
  if (DoBothBroadcast())
    return lowerShuffleAsDecomposedShuffleMerge(DL, VT, V1, V2, Mask, Zeroable,
                                                Subtarget, DAG);
 
  // If the inputs all stem from a single 128-bit lane of each input, then we
  // split them rather than blending because the split will decompose to
  // unusually few instructions.
  int LaneCount = VT.getSizeInBits() / 128;
  int LaneSize = Size / LaneCount;
  SmallBitVector LaneInputs[2];
  LaneInputs[0].resize(LaneCount, false);
  LaneInputs[1].resize(LaneCount, false);
  for (int i = 0; i < Size; ++i)
    if (Mask[i] >= 0)
      LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
  if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
    return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG,
                                /*SimpleOnly*/ false);
 
  // Without AVX2, if we can freely split the subvectors then we're better off
  // performing half width shuffles.
  if (!Subtarget.hasAVX2()) {
    SDValue BC1 = peekThroughBitcasts(V1);
    SDValue BC2 = peekThroughBitcasts(V2);
    bool SplatOrSplitV1 = isFreeToSplitVector(BC1, DAG) ||
                          DAG.isSplatValue(BC1, /*AllowUndefs=*/true);
    bool SplatOrSplitV2 = isFreeToSplitVector(BC2, DAG) ||
                          DAG.isSplatValue(BC2, /*AllowUndefs=*/true);
    if (SplatOrSplitV1 && SplatOrSplitV2)
      return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG,
                                  /*SimpleOnly*/ false);
  }
 
  // Otherwise, just fall back to decomposed shuffles and a blend/unpack. This
  // requires that the decomposed single-input shuffles don't end up here.
  return lowerShuffleAsDecomposedShuffleMerge(DL, VT, V1, V2, Mask, Zeroable,
                                              Subtarget, DAG);
}
 
// Lower as SHUFPD(VPERM2F128(V1, V2), VPERM2F128(V1, V2)).
// TODO: Extend to support v8f32 (+ 512-bit shuffles).
static SDValue lowerShuffleAsLanePermuteAndSHUFP(const SDLoc &DL, MVT VT,
                                                 SDValue V1, SDValue V2,
                                                 ArrayRef<int> Mask,
                                                 SelectionDAG &DAG) {
  assert(VT == MVT::v4f64 && "Only for v4f64 shuffles");
 
  int LHSMask[4] = {-1, -1, -1, -1};
  int RHSMask[4] = {-1, -1, -1, -1};
  int SHUFPDMask[4] = {-1, -1, -1, -1};
 
  // As SHUFPD uses a single LHS/RHS element per lane, we can always
  // perform the shuffle once the lanes have been shuffled in place.
  for (int i = 0; i != 4; ++i) {
    int M = Mask[i];
    if (M < 0)
      continue;
    int LaneBase = i & ~1;
    auto &LaneMask = (i & 1) ? RHSMask : LHSMask;
    LaneMask[LaneBase + (M & 1)] = M;
    SHUFPDMask[i] = M & 1;
  }
 
  SDValue LHS = DAG.getVectorShuffle(VT, DL, V1, V2, LHSMask);
  SDValue RHS = DAG.getVectorShuffle(VT, DL, V1, V2, RHSMask);
  return DAG.getNode(X86ISD::SHUFP, DL, VT, LHS, RHS,
                     getSHUFPDImmForMask(SHUFPDMask, DL, DAG));
}
 
/// Lower a vector shuffle crossing multiple 128-bit lanes as
/// a lane permutation followed by a per-lane permutation.
///
/// This is mainly for cases where we can have non-repeating permutes
/// in each lane.
///
/// TODO: This is very similar to lowerShuffleAsLanePermuteAndRepeatedMask,
/// we should investigate merging them.
static SDValue lowerShuffleAsLanePermuteAndPermute(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    SelectionDAG &DAG, const X86Subtarget &Subtarget) {
  int NumElts = VT.getVectorNumElements();
  int NumLanes = VT.getSizeInBits() / 128;
  int NumEltsPerLane = NumElts / NumLanes;
  bool CanUseSublanes = Subtarget.hasAVX2() && V2.isUndef();
 
  /// Attempts to find a sublane permute with the given size
  /// that gets all elements into their target lanes.
  ///
  /// If successful, fills CrossLaneMask and InLaneMask and returns true.
  /// If unsuccessful, returns false and may overwrite InLaneMask.
  auto getSublanePermute = [&](int NumSublanes) -> SDValue {
    int NumSublanesPerLane = NumSublanes / NumLanes;
    int NumEltsPerSublane = NumElts / NumSublanes;
 
    SmallVector<int, 16> CrossLaneMask;
    SmallVector<int, 16> InLaneMask(NumElts, SM_SentinelUndef);
    // CrossLaneMask but one entry == one sublane.
    SmallVector<int, 16> CrossLaneMaskLarge(NumSublanes, SM_SentinelUndef);
    APInt DemandedCrossLane = APInt::getZero(NumElts);
 
    for (int i = 0; i != NumElts; ++i) {
      int M = Mask[i];
      if (M < 0)
        continue;
 
      int SrcSublane = M / NumEltsPerSublane;
      int DstLane = i / NumEltsPerLane;
 
      // We only need to get the elements into the right lane, not sublane.
      // So search all sublanes that make up the destination lane.
      bool Found = false;
      int DstSubStart = DstLane * NumSublanesPerLane;
      int DstSubEnd = DstSubStart + NumSublanesPerLane;
      for (int DstSublane = DstSubStart; DstSublane < DstSubEnd; ++DstSublane) {
        if (!isUndefOrEqual(CrossLaneMaskLarge[DstSublane], SrcSublane))
          continue;
 
        Found = true;
        CrossLaneMaskLarge[DstSublane] = SrcSublane;
        int DstSublaneOffset = DstSublane * NumEltsPerSublane;
        InLaneMask[i] = DstSublaneOffset + M % NumEltsPerSublane;
        DemandedCrossLane.setBit(InLaneMask[i]);
        break;
      }
      if (!Found)
        return SDValue();
    }
 
    // Fill CrossLaneMask using CrossLaneMaskLarge.
    narrowShuffleMaskElts(NumEltsPerSublane, CrossLaneMaskLarge, CrossLaneMask);
 
    if (!CanUseSublanes) {
      // If we're only shuffling a single lowest lane and the rest are identity
      // then don't bother.
      // TODO - isShuffleMaskInputInPlace could be extended to something like
      // this.
      int NumIdentityLanes = 0;
      bool OnlyShuffleLowestLane = true;
      for (int i = 0; i != NumLanes; ++i) {
        int LaneOffset = i * NumEltsPerLane;
        if (isSequentialOrUndefInRange(InLaneMask, LaneOffset, NumEltsPerLane,
                                       i * NumEltsPerLane))
          NumIdentityLanes++;
        else if (CrossLaneMask[LaneOffset] != 0)
          OnlyShuffleLowestLane = false;
      }
      if (OnlyShuffleLowestLane && NumIdentityLanes == (NumLanes - 1))
        return SDValue();
    }
 
    // Simplify CrossLaneMask based on the actual demanded elements.
    if (V1.hasOneUse())
      for (int i = 0; i != NumElts; ++i)
        if (!DemandedCrossLane[i])
          CrossLaneMask[i] = SM_SentinelUndef;
 
    // Avoid returning the same shuffle operation. For example,
    // t7: v16i16 = vector_shuffle<8,9,10,11,4,5,6,7,0,1,2,3,12,13,14,15> t5,
    //                             undef:v16i16
    if (CrossLaneMask == Mask || InLaneMask == Mask)
      return SDValue();
 
    SDValue CrossLane = DAG.getVectorShuffle(VT, DL, V1, V2, CrossLaneMask);
    return DAG.getVectorShuffle(VT, DL, CrossLane, DAG.getUNDEF(VT),
                                InLaneMask);
  };
 
  // First attempt a solution with full lanes.
  if (SDValue V = getSublanePermute(/*NumSublanes=*/NumLanes))
    return V;
 
  // The rest of the solutions use sublanes.
  if (!CanUseSublanes)
    return SDValue();
 
  // Then attempt a solution with 64-bit sublanes (vpermq).
  if (SDValue V = getSublanePermute(/*NumSublanes=*/NumLanes * 2))
    return V;
 
  // If that doesn't work and we have fast variable cross-lane shuffle,
  // attempt 32-bit sublanes (vpermd).
  if (!Subtarget.hasFastVariableCrossLaneShuffle())
    return SDValue();
 
  return getSublanePermute(/*NumSublanes=*/NumLanes * 4);
}
 
/// Helper to get compute inlane shuffle mask for a complete shuffle mask.
static void computeInLaneShuffleMask(const ArrayRef<int> &Mask, int LaneSize,
                                     SmallVector<int> &InLaneMask) {
  int Size = Mask.size();
  InLaneMask.assign(Mask.begin(), Mask.end());
  for (int i = 0; i < Size; ++i) {
    int &M = InLaneMask[i];
    if (M < 0)
      continue;
    if (((M % Size) / LaneSize) != (i / LaneSize))
      M = (M % LaneSize) + ((i / LaneSize) * LaneSize) + Size;
  }
}
 
/// Lower a vector shuffle crossing multiple 128-bit lanes by shuffling one
/// source with a lane permutation.
///
/// This lowering strategy results in four instructions in the worst case for a
/// single-input cross lane shuffle which is lower than any other fully general
/// cross-lane shuffle strategy I'm aware of. Special cases for each particular
/// shuffle pattern should be handled prior to trying this lowering.
static SDValue lowerShuffleAsLanePermuteAndShuffle(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    SelectionDAG &DAG, const X86Subtarget &Subtarget) {
  // FIXME: This should probably be generalized for 512-bit vectors as well.
  assert(VT.is256BitVector() && "Only for 256-bit vector shuffles!");
  int Size = Mask.size();
  int LaneSize = Size / 2;
 
  // Fold to SHUFPD(VPERM2F128(V1, V2), VPERM2F128(V1, V2)).
  // Only do this if the elements aren't all from the lower lane,
  // otherwise we're (probably) better off doing a split.
  if (VT == MVT::v4f64 &&
      !all_of(Mask, [LaneSize](int M) { return M < LaneSize; }))
    return lowerShuffleAsLanePermuteAndSHUFP(DL, VT, V1, V2, Mask, DAG);
 
  // If there are only inputs from one 128-bit lane, splitting will in fact be
  // less expensive. The flags track whether the given lane contains an element
  // that crosses to another lane.
  bool AllLanes;
  if (!Subtarget.hasAVX2()) {
    bool LaneCrossing[2] = {false, false};
    for (int i = 0; i < Size; ++i)
      if (Mask[i] >= 0 && ((Mask[i] % Size) / LaneSize) != (i / LaneSize))
        LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
    AllLanes = LaneCrossing[0] && LaneCrossing[1];
  } else {
    bool LaneUsed[2] = {false, false};
    for (int i = 0; i < Size; ++i)
      if (Mask[i] >= 0)
        LaneUsed[(Mask[i] % Size) / LaneSize] = true;
    AllLanes = LaneUsed[0] && LaneUsed[1];
  }
 
  // TODO - we could support shuffling V2 in the Flipped input.
  assert(V2.isUndef() &&
         "This last part of this routine only works on single input shuffles");
 
  SmallVector<int> InLaneMask;
  computeInLaneShuffleMask(Mask, Mask.size() / 2, InLaneMask);
 
  assert(!is128BitLaneCrossingShuffleMask(VT, InLaneMask) &&
         "In-lane shuffle mask expected");
 
  // If we're not using both lanes in each lane and the inlane mask is not
  // repeating, then we're better off splitting.
  if (!AllLanes && !is128BitLaneRepeatedShuffleMask(VT, InLaneMask))
    return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG,
                                /*SimpleOnly*/ false);
 
  // Flip the lanes, and shuffle the results which should now be in-lane.
  MVT PVT = VT.isFloatingPoint() ? MVT::v4f64 : MVT::v4i64;
  SDValue Flipped = DAG.getBitcast(PVT, V1);
  Flipped =
      DAG.getVectorShuffle(PVT, DL, Flipped, DAG.getUNDEF(PVT), {2, 3, 0, 1});
  Flipped = DAG.getBitcast(VT, Flipped);
  return DAG.getVectorShuffle(VT, DL, V1, Flipped, InLaneMask);
}
 
/// Handle lowering 2-lane 128-bit shuffles.
static SDValue lowerV2X128Shuffle(const SDLoc &DL, MVT VT, SDValue V1,
                                  SDValue V2, ArrayRef<int> Mask,
                                  const APInt &Zeroable,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  if (V2.isUndef()) {
    // Attempt to match VBROADCAST*128 subvector broadcast load.
    bool SplatLo = isShuffleEquivalent(Mask, {0, 1, 0, 1}, V1);
    bool SplatHi = isShuffleEquivalent(Mask, {2, 3, 2, 3}, V1);
    if ((SplatLo || SplatHi) && !Subtarget.hasAVX512() && V1.hasOneUse() &&
        X86::mayFoldLoad(peekThroughOneUseBitcasts(V1), Subtarget)) {
      MVT MemVT = VT.getHalfNumVectorElementsVT();
      unsigned Ofs = SplatLo ? 0 : MemVT.getStoreSize();
      auto *Ld = cast<LoadSDNode>(peekThroughOneUseBitcasts(V1));
      if (SDValue BcstLd = getBROADCAST_LOAD(X86ISD::SUBV_BROADCAST_LOAD, DL,
                                             VT, MemVT, Ld, Ofs, DAG))
        return BcstLd;
    }
 
    // With AVX2, use VPERMQ/VPERMPD for unary shuffles to allow memory folding.
    if (Subtarget.hasAVX2())
      return SDValue();
  }
 
  bool V2IsZero = !V2.isUndef() && ISD::isBuildVectorAllZeros(V2.getNode());
 
  SmallVector<int, 4> WidenedMask;
  if (!canWidenShuffleElements(Mask, Zeroable, V2IsZero, WidenedMask))
    return SDValue();
 
  bool IsLowZero = (Zeroable & 0x3) == 0x3;
  bool IsHighZero = (Zeroable & 0xc) == 0xc;
 
  // Try to use an insert into a zero vector.
  if (WidenedMask[0] == 0 && IsHighZero) {
    MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 2);
    SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
                              DAG.getVectorIdxConstant(0, DL));
    return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
                       getZeroVector(VT, Subtarget, DAG, DL), LoV,
                       DAG.getVectorIdxConstant(0, DL));
  }
 
  // TODO: If minimizing size and one of the inputs is a zero vector and the
  // the zero vector has only one use, we could use a VPERM2X128 to save the
  // instruction bytes needed to explicitly generate the zero vector.
 
  // Blends are faster and handle all the non-lane-crossing cases.
  if (SDValue Blend = lowerShuffleAsBlend(DL, VT, V1, V2, Mask, Zeroable,
                                          Subtarget, DAG))
    return Blend;
 
  // If either input operand is a zero vector, use VPERM2X128 because its mask
  // allows us to replace the zero input with an implicit zero.
  if (!IsLowZero && !IsHighZero) {
    // Check for patterns which can be matched with a single insert of a 128-bit
    // subvector.
    bool OnlyUsesV1 = isShuffleEquivalent(Mask, {0, 1, 0, 1}, V1, V2);
    if (OnlyUsesV1 || isShuffleEquivalent(Mask, {0, 1, 4, 5}, V1, V2)) {
 
      // With AVX1, use vperm2f128 (below) to allow load folding. Otherwise,
      // this will likely become vinsertf128 which can't fold a 256-bit memop.
      if (!isa<LoadSDNode>(peekThroughBitcasts(V1))) {
        MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 2);
        SDValue SubVec =
            DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, OnlyUsesV1 ? V1 : V2,
                        DAG.getVectorIdxConstant(0, DL));
        return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, V1, SubVec,
                           DAG.getVectorIdxConstant(2, DL));
      }
    }
 
    // Try to use SHUF128 if possible.
    if (Subtarget.hasVLX()) {
      if (WidenedMask[0] < 2 && WidenedMask[1] >= 2) {
        unsigned PermMask = ((WidenedMask[0] % 2) << 0) |
                            ((WidenedMask[1] % 2) << 1);
        return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
                           DAG.getTargetConstant(PermMask, DL, MVT::i8));
      }
    }
  }
 
  // Otherwise form a 128-bit permutation. After accounting for undefs,
  // convert the 64-bit shuffle mask selection values into 128-bit
  // selection bits by dividing the indexes by 2 and shifting into positions
  // defined by a vperm2*128 instruction's immediate control byte.
 
  // The immediate permute control byte looks like this:
  //    [1:0] - select 128 bits from sources for low half of destination
  //    [2]   - ignore
  //    [3]   - zero low half of destination
  //    [5:4] - select 128 bits from sources for high half of destination
  //    [6]   - ignore
  //    [7]   - zero high half of destination
 
  assert((WidenedMask[0] >= 0 || IsLowZero) &&
         (WidenedMask[1] >= 0 || IsHighZero) && "Undef half?");
 
  unsigned PermMask = 0;
  PermMask |= IsLowZero  ? 0x08 : (WidenedMask[0] << 0);
  PermMask |= IsHighZero ? 0x80 : (WidenedMask[1] << 4);
 
  // Check the immediate mask and replace unused sources with undef.
  if ((PermMask & 0x0a) != 0x00 && (PermMask & 0xa0) != 0x00)
    V1 = DAG.getUNDEF(VT);
  if ((PermMask & 0x0a) != 0x02 && (PermMask & 0xa0) != 0x20)
    V2 = DAG.getUNDEF(VT);
 
  return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
                     DAG.getTargetConstant(PermMask, DL, MVT::i8));
}
 
/// Lower a vector shuffle by first fixing the 128-bit lanes and then
/// shuffling each lane.
///
/// This attempts to create a repeated lane shuffle where each lane uses one
/// or two of the lanes of the inputs. The lanes of the input vectors are
/// shuffled in one or two independent shuffles to get the lanes into the
/// position needed by the final shuffle.
static SDValue lowerShuffleAsLanePermuteAndRepeatedMask(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    const X86Subtarget &Subtarget, SelectionDAG &DAG) {
  // This is only useful for binary shuffle with a non-repeating mask.
  if (V2.isUndef() || is128BitLaneRepeatedShuffleMask(VT, Mask))
    return SDValue();
 
  int NumElts = Mask.size();
  int NumLanes = VT.getSizeInBits() / 128;
  int NumLaneElts = 128 / VT.getScalarSizeInBits();
  SmallVector<int, 16> RepeatMask(NumLaneElts, -1);
  SmallVector<std::array<int, 2>, 2> LaneSrcs(NumLanes, {{-1, -1}});
 
  // First pass will try to fill in the RepeatMask from lanes that need two
  // sources.
  for (int Lane = 0; Lane != NumLanes; ++Lane) {
    int Srcs[2] = {-1, -1};
    SmallVector<int, 16> InLaneMask(NumLaneElts, -1);
    for (int i = 0; i != NumLaneElts; ++i) {
      int M = Mask[(Lane * NumLaneElts) + i];
      if (M < 0)
        continue;
      // Determine which of the possible input lanes (NumLanes from each source)
      // this element comes from. Assign that as one of the sources for this
      // lane. We can assign up to 2 sources for this lane. If we run out
      // sources we can't do anything.
      int LaneSrc = M / NumLaneElts;
      int Src;
      if (Srcs[0] < 0 || Srcs[0] == LaneSrc)
        Src = 0;
      else if (Srcs[1] < 0 || Srcs[1] == LaneSrc)
        Src = 1;
      else
        return SDValue();
 
      Srcs[Src] = LaneSrc;
      InLaneMask[i] = (M % NumLaneElts) + Src * NumElts;
    }
 
    // If this lane has two sources, see if it fits with the repeat mask so far.
    if (Srcs[1] < 0)
      continue;
 
    LaneSrcs[Lane][0] = Srcs[0];
    LaneSrcs[Lane][1] = Srcs[1];
 
    auto MatchMasks = [](ArrayRef<int> M1, ArrayRef<int> M2) {
      assert(M1.size() == M2.size() && "Unexpected mask size");
      for (int i = 0, e = M1.size(); i != e; ++i)
        if (M1[i] >= 0 && M2[i] >= 0 && M1[i] != M2[i])
          return false;
      return true;
    };
 
    auto MergeMasks = [](ArrayRef<int> Mask, MutableArrayRef<int> MergedMask) {
      assert(Mask.size() == MergedMask.size() && "Unexpected mask size");
      for (int i = 0, e = MergedMask.size(); i != e; ++i) {
        int M = Mask[i];
        if (M < 0)
          continue;
        assert((MergedMask[i] < 0 || MergedMask[i] == M) &&
               "Unexpected mask element");
        MergedMask[i] = M;
      }
    };
 
    if (MatchMasks(InLaneMask, RepeatMask)) {
      // Merge this lane mask into the final repeat mask.
      MergeMasks(InLaneMask, RepeatMask);
      continue;
    }
 
    // Didn't find a match. Swap the operands and try again.
    std::swap(LaneSrcs[Lane][0], LaneSrcs[Lane][1]);
    ShuffleVectorSDNode::commuteMask(InLaneMask);
 
    if (MatchMasks(InLaneMask, RepeatMask)) {
      // Merge this lane mask into the final repeat mask.
      MergeMasks(InLaneMask, RepeatMask);
      continue;
    }
 
    // Couldn't find a match with the operands in either order.
    return SDValue();
  }
 
  // Now handle any lanes with only one source.
  for (int Lane = 0; Lane != NumLanes; ++Lane) {
    // If this lane has already been processed, skip it.
    if (LaneSrcs[Lane][0] >= 0)
      continue;
 
    for (int i = 0; i != NumLaneElts; ++i) {
      int M = Mask[(Lane * NumLaneElts) + i];
      if (M < 0)
        continue;
 
      // If RepeatMask isn't defined yet we can define it ourself.
      if (RepeatMask[i] < 0)
        RepeatMask[i] = M % NumLaneElts;
 
      if (RepeatMask[i] < NumElts) {
        if (RepeatMask[i] != M % NumLaneElts)
          return SDValue();
        LaneSrcs[Lane][0] = M / NumLaneElts;
      } else {
        if (RepeatMask[i] != ((M % NumLaneElts) + NumElts))
          return SDValue();
        LaneSrcs[Lane][1] = M / NumLaneElts;
      }
    }
 
    if (LaneSrcs[Lane][0] < 0 && LaneSrcs[Lane][1] < 0)
      return SDValue();
  }
 
  SmallVector<int, 16> NewMask(NumElts, -1);
  for (int Lane = 0; Lane != NumLanes; ++Lane) {
    int Src = LaneSrcs[Lane][0];
    for (int i = 0; i != NumLaneElts; ++i) {
      int M = -1;
      if (Src >= 0)
        M = Src * NumLaneElts + i;
      NewMask[Lane * NumLaneElts + i] = M;
    }
  }
  SDValue NewV1 = DAG.getVectorShuffle(VT, DL, V1, V2, NewMask);
  // Ensure we didn't get back the shuffle we started with.
  // FIXME: This is a hack to make up for some splat handling code in
  // getVectorShuffle.
  if (isa<ShuffleVectorSDNode>(NewV1) &&
      cast<ShuffleVectorSDNode>(NewV1)->getMask() == Mask)
    return SDValue();
 
  for (int Lane = 0; Lane != NumLanes; ++Lane) {
    int Src = LaneSrcs[Lane][1];
    for (int i = 0; i != NumLaneElts; ++i) {
      int M = -1;
      if (Src >= 0)
        M = Src * NumLaneElts + i;
      NewMask[Lane * NumLaneElts + i] = M;
    }
  }
  SDValue NewV2 = DAG.getVectorShuffle(VT, DL, V1, V2, NewMask);
  // Ensure we didn't get back the shuffle we started with.
  // FIXME: This is a hack to make up for some splat handling code in
  // getVectorShuffle.
  if (isa<ShuffleVectorSDNode>(NewV2) &&
      cast<ShuffleVectorSDNode>(NewV2)->getMask() == Mask)
    return SDValue();
 
  for (int i = 0; i != NumElts; ++i) {
    if (Mask[i] < 0) {
      NewMask[i] = -1;
      continue;
    }
    NewMask[i] = RepeatMask[i % NumLaneElts];
    if (NewMask[i] < 0)
      continue;
 
    NewMask[i] += (i / NumLaneElts) * NumLaneElts;
  }
  return DAG.getVectorShuffle(VT, DL, NewV1, NewV2, NewMask);
}
 
/// If the input shuffle mask results in a vector that is undefined in all upper
/// or lower half elements and that mask accesses only 2 halves of the
/// shuffle's operands, return true. A mask of half the width with mask indexes
/// adjusted to access the extracted halves of the original shuffle operands is
/// returned in HalfMask. HalfIdx1 and HalfIdx2 return whether the upper or
/// lower half of each input operand is accessed.
static bool
getHalfShuffleMask(ArrayRef<int> Mask, MutableArrayRef<int> HalfMask,
                   int &HalfIdx1, int &HalfIdx2) {
  assert((Mask.size() == HalfMask.size() * 2) &&
         "Expected input mask to be twice as long as output");
 
  // Exactly one half of the result must be undef to allow narrowing.
  bool UndefLower = isUndefLowerHalf(Mask);
  bool UndefUpper = isUndefUpperHalf(Mask);
  if (UndefLower == UndefUpper)
    return false;
 
  unsigned HalfNumElts = HalfMask.size();
  unsigned MaskIndexOffset = UndefLower ? HalfNumElts : 0;
  HalfIdx1 = -1;
  HalfIdx2 = -1;
  for (unsigned i = 0; i != HalfNumElts; ++i) {
    int M = Mask[i + MaskIndexOffset];
    if (M < 0) {
      HalfMask[i] = M;
      continue;
    }
 
    // Determine which of the 4 half vectors this element is from.
    // i.e. 0 = Lower V1, 1 = Upper V1, 2 = Lower V2, 3 = Upper V2.
    int HalfIdx = M / HalfNumElts;
 
    // Determine the element index into its half vector source.
    int HalfElt = M % HalfNumElts;
 
    // We can shuffle with up to 2 half vectors, set the new 'half'
    // shuffle mask accordingly.
    if (HalfIdx1 < 0 || HalfIdx1 == HalfIdx) {
      HalfMask[i] = HalfElt;
      HalfIdx1 = HalfIdx;
      continue;
    }
    if (HalfIdx2 < 0 || HalfIdx2 == HalfIdx) {
      HalfMask[i] = HalfElt + HalfNumElts;
      HalfIdx2 = HalfIdx;
      continue;
    }
 
    // Too many half vectors referenced.
    return false;
  }
 
  return true;
}
 
/// Given the output values from getHalfShuffleMask(), create a half width
/// shuffle of extracted vectors followed by an insert back to full width.
static SDValue getShuffleHalfVectors(const SDLoc &DL, SDValue V1, SDValue V2,
                                     ArrayRef<int> HalfMask, int HalfIdx1,
                                     int HalfIdx2, bool UndefLower,
                                     SelectionDAG &DAG, bool UseConcat = false) {
  assert(V1.getValueType() == V2.getValueType() && "Different sized vectors?");
  assert(V1.getValueType().isSimple() && "Expecting only simple types");
 
  MVT VT = V1.getSimpleValueType();
  MVT HalfVT = VT.getHalfNumVectorElementsVT();
  unsigned HalfNumElts = HalfVT.getVectorNumElements();
 
  auto getHalfVector = [&](int HalfIdx) {
    if (HalfIdx < 0)
      return DAG.getUNDEF(HalfVT);
    SDValue V = (HalfIdx < 2 ? V1 : V2);
    HalfIdx = (HalfIdx % 2) * HalfNumElts;
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V,
                       DAG.getVectorIdxConstant(HalfIdx, DL));
  };
 
  // ins undef, (shuf (ext V1, HalfIdx1), (ext V2, HalfIdx2), HalfMask), Offset
  SDValue Half1 = getHalfVector(HalfIdx1);
  SDValue Half2 = getHalfVector(HalfIdx2);
  SDValue V = DAG.getVectorShuffle(HalfVT, DL, Half1, Half2, HalfMask);
  if (UseConcat) {
    SDValue Op0 = V;
    SDValue Op1 = DAG.getUNDEF(HalfVT);
    if (UndefLower)
      std::swap(Op0, Op1);
    return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Op0, Op1);
  }
 
  unsigned Offset = UndefLower ? HalfNumElts : 0;
  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), V,
                     DAG.getVectorIdxConstant(Offset, DL));
}
 
/// Lower shuffles where an entire half of a 256 or 512-bit vector is UNDEF.
/// This allows for fast cases such as subvector extraction/insertion
/// or shuffling smaller vector types which can lower more efficiently.
static SDValue lowerShuffleWithUndefHalf(const SDLoc &DL, MVT VT, SDValue V1,
                                         SDValue V2, ArrayRef<int> Mask,
                                         const X86Subtarget &Subtarget,
                                         SelectionDAG &DAG) {
  assert((VT.is256BitVector() || VT.is512BitVector()) &&
         "Expected 256-bit or 512-bit vector");
 
  bool UndefLower = isUndefLowerHalf(Mask);
  if (!UndefLower && !isUndefUpperHalf(Mask))
    return SDValue();
 
  assert((!UndefLower || !isUndefUpperHalf(Mask)) &&
         "Completely undef shuffle mask should have been simplified already");
 
  // Upper half is undef and lower half is whole upper subvector.
  // e.g. vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
  MVT HalfVT = VT.getHalfNumVectorElementsVT();
  unsigned HalfNumElts = HalfVT.getVectorNumElements();
  if (!UndefLower &&
      isSequentialOrUndefInRange(Mask, 0, HalfNumElts, HalfNumElts)) {
    SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
                             DAG.getVectorIdxConstant(HalfNumElts, DL));
    return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
                       DAG.getVectorIdxConstant(0, DL));
  }
 
  // Lower half is undef and upper half is whole lower subvector.
  // e.g. vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
  if (UndefLower &&
      isSequentialOrUndefInRange(Mask, HalfNumElts, HalfNumElts, 0)) {
    SDValue Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, HalfVT, V1,
                             DAG.getVectorIdxConstant(0, DL));
    return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, DAG.getUNDEF(VT), Hi,
                       DAG.getVectorIdxConstant(HalfNumElts, DL));
  }
 
  int HalfIdx1, HalfIdx2;
  SmallVector<int, 8> HalfMask(HalfNumElts);
  if (!getHalfShuffleMask(Mask, HalfMask, HalfIdx1, HalfIdx2))
    return SDValue();
 
  assert(HalfMask.size() == HalfNumElts && "Unexpected shuffle mask length");
 
  // Only shuffle the halves of the inputs when useful.
  unsigned NumLowerHalves =
      (HalfIdx1 == 0 || HalfIdx1 == 2) + (HalfIdx2 == 0 || HalfIdx2 == 2);
  unsigned NumUpperHalves =
      (HalfIdx1 == 1 || HalfIdx1 == 3) + (HalfIdx2 == 1 || HalfIdx2 == 3);
  assert(NumLowerHalves + NumUpperHalves <= 2 && "Only 1 or 2 halves allowed");
 
  // Determine the larger pattern of undef/halves, then decide if it's worth
  // splitting the shuffle based on subtarget capabilities and types.
  unsigned EltWidth = VT.getVectorElementType().getSizeInBits();
  if (!UndefLower) {
    // XXXXuuuu: no insert is needed.
    // Always extract lowers when setting lower - these are all free subreg ops.
    if (NumUpperHalves == 0)
      return getShuffleHalfVectors(DL, V1, V2, HalfMask, HalfIdx1, HalfIdx2,
                                   UndefLower, DAG);
 
    if (NumUpperHalves == 1) {
      // AVX2 has efficient 32/64-bit element cross-lane shuffles.
      if (Subtarget.hasAVX2()) {
        // extract128 + vunpckhps/vshufps, is better than vblend + vpermps.
        if (EltWidth == 32 && NumLowerHalves && HalfVT.is128BitVector() &&
            !is128BitUnpackShuffleMask(HalfMask, DAG) &&
            (!isSingleSHUFPSMask(HalfMask) ||
             Subtarget.hasFastVariableCrossLaneShuffle()))
          return SDValue();
        // If this is an unary shuffle (assume that the 2nd operand is
        // canonicalized to undef), then we can use vpermpd. Otherwise, we
        // are better off extracting the upper half of 1 operand and using a
        // narrow shuffle.
        if (EltWidth == 64 && V2.isUndef())
          return SDValue();
        // If this is an unary vXi8 shuffle with inplace halves, then perform as
        // full width pshufb, and then merge.
        if (EltWidth == 8 && HalfIdx1 == 0 && HalfIdx2 == 1)
          return SDValue();
      }
      // AVX512 has efficient cross-lane shuffles for all legal 512-bit types.
      if (Subtarget.hasAVX512() && VT.is512BitVector())
        return SDValue();
      // Extract + narrow shuffle is better than the wide alternative.
      return getShuffleHalfVectors(DL, V1, V2, HalfMask, HalfIdx1, HalfIdx2,
                                   UndefLower, DAG);
    }
 
    // Don't extract both uppers, instead shuffle and then extract.
    assert(NumUpperHalves == 2 && "Half vector count went wrong");
    return SDValue();
  }
 
  // UndefLower - uuuuXXXX: an insert to high half is required if we split this.
  if (NumUpperHalves == 0) {
    // AVX2 has efficient 64-bit element cross-lane shuffles.
    // TODO: Refine to account for unary shuffle, splat, and other masks?
    if (Subtarget.hasAVX2() && EltWidth == 64)
      return SDValue();
    // AVX512 has efficient cross-lane shuffles for all legal 512-bit types.
    if (Subtarget.hasAVX512() && VT.is512BitVector())
      return SDValue();
    // Narrow shuffle + insert is better than the wide alternative.
    return getShuffleHalfVectors(DL, V1, V2, HalfMask, HalfIdx1, HalfIdx2,
                                 UndefLower, DAG);
  }
 
  // NumUpperHalves != 0: don't bother with extract, shuffle, and then insert.
  return SDValue();
}
 
/// Handle case where shuffle sources are coming from the same 128-bit lane and
/// every lane can be represented as the same repeating mask - allowing us to
/// shuffle the sources with the repeating shuffle and then permute the result
/// to the destination lanes.
static SDValue lowerShuffleAsRepeatedMaskAndLanePermute(
    const SDLoc &DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
    const X86Subtarget &Subtarget, SelectionDAG &DAG) {
  int NumElts = VT.getVectorNumElements();
  int NumLanes = VT.getSizeInBits() / 128;
  int NumLaneElts = NumElts / NumLanes;
 
  // On AVX2 we may be able to just shuffle the lowest elements and then
  // broadcast the result.
  if (Subtarget.hasAVX2()) {
    for (unsigned BroadcastSize : {16, 32, 64}) {
      if (BroadcastSize <= VT.getScalarSizeInBits())
        continue;
      int NumBroadcastElts = BroadcastSize / VT.getScalarSizeInBits();
 
      // Attempt to match a repeating pattern every NumBroadcastElts,
      // accounting for UNDEFs but only references the lowest 128-bit
      // lane of the inputs.
      auto FindRepeatingBroadcastMask = [&](SmallVectorImpl<int> &RepeatMask) {
        for (int i = 0; i != NumElts; i += NumBroadcastElts)
          for (int j = 0; j != NumBroadcastElts; ++j) {
            int M = Mask[i + j];
            if (M < 0)
              continue;
            int &R = RepeatMask[j];
            if (0 != ((M % NumElts) / NumLaneElts))
              return false;
            if (0 <= R && R != M)
              return false;
            R = M;
          }
        return true;
      };
 
      SmallVector<int, 8> RepeatMask((unsigned)NumElts, -1);
      if (!FindRepeatingBroadcastMask(RepeatMask))
        continue;
 
      // Shuffle the (lowest) repeated elements in place for broadcast.
      SDValue RepeatShuf = DAG.getVectorShuffle(VT, DL, V1, V2, RepeatMask);
 
      // Shuffle the actual broadcast.
      SmallVector<int, 8> BroadcastMask((unsigned)NumElts, -1);
      for (int i = 0; i != NumElts; i += NumBroadcastElts)
        for (int j = 0; j != NumBroadcastElts; ++j)
          BroadcastMask[i + j] = j;
 
      // Avoid returning the same shuffle operation. For example,
      // v8i32 = vector_shuffle<0,1,0,1,0,1,0,1> t5, undef:v8i32
      if (BroadcastMask == Mask)
        return SDValue();
 
      return DAG.getVectorShuffle(VT, DL, RepeatShuf, DAG.getUNDEF(VT),
                                  BroadcastMask);
    }
  }
 
  // Bail if the shuffle mask doesn't cross 128-bit lanes.
  if (!is128BitLaneCrossingShuffleMask(VT, Mask))
    return SDValue();
 
  // Bail if we already have a repeated lane shuffle mask.
  if (is128BitLaneRepeatedShuffleMask(VT, Mask))
    return SDValue();
 
  // Helper to look for repeated mask in each split sublane, and that those
  // sublanes can then be permuted into place.
  auto ShuffleSubLanes = [&](int SubLaneScale) {
    int NumSubLanes = NumLanes * SubLaneScale;
    int NumSubLaneElts = NumLaneElts / SubLaneScale;
 
    // Check that all the sources are coming from the same lane and see if we
    // can form a repeating shuffle mask (local to each sub-lane). At the same
    // time, determine the source sub-lane for each destination sub-lane.
    int TopSrcSubLane = -1;
    SmallVector<int, 8> Dst2SrcSubLanes((unsigned)NumSubLanes, -1);
    SmallVector<SmallVector<int, 8>> RepeatedSubLaneMasks(
        SubLaneScale,
        SmallVector<int, 8>((unsigned)NumSubLaneElts, SM_SentinelUndef));
 
    for (int DstSubLane = 0; DstSubLane != NumSubLanes; ++DstSubLane) {
      // Extract the sub-lane mask, check that it all comes from the same lane
      // and normalize the mask entries to come from the first lane.
      int SrcLane = -1;
      SmallVector<int, 8> SubLaneMask((unsigned)NumSubLaneElts, -1);
      for (int Elt = 0; Elt != NumSubLaneElts; ++Elt) {
        int M = Mask[(DstSubLane * NumSubLaneElts) + Elt];
        if (M < 0)
          continue;
        int Lane = (M % NumElts) / NumLaneElts;
        if ((0 <= SrcLane) && (SrcLane != Lane))
          return SDValue();
        SrcLane = Lane;
        int LocalM = (M % NumLaneElts) + (M < NumElts ? 0 : NumElts);
        SubLaneMask[Elt] = LocalM;
      }
 
      // Whole sub-lane is UNDEF.
      if (SrcLane < 0)
        continue;
 
      // Attempt to match against the candidate repeated sub-lane masks.
      for (int SubLane = 0; SubLane != SubLaneScale; ++SubLane) {
        auto MatchMasks = [NumSubLaneElts](ArrayRef<int> M1, ArrayRef<int> M2) {
          for (int i = 0; i != NumSubLaneElts; ++i) {
            if (M1[i] < 0 || M2[i] < 0)
              continue;
            if (M1[i] != M2[i])
              return false;
          }
          return true;
        };
 
        auto &RepeatedSubLaneMask = RepeatedSubLaneMasks[SubLane];
        if (!MatchMasks(SubLaneMask, RepeatedSubLaneMask))
          continue;
 
        // Merge the sub-lane mask into the matching repeated sub-lane mask.
        for (int i = 0; i != NumSubLaneElts; ++i) {
          int M = SubLaneMask[i];
          if (M < 0)
            continue;
          assert((RepeatedSubLaneMask[i] < 0 || RepeatedSubLaneMask[i] == M) &&
                 "Unexpected mask element");
          RepeatedSubLaneMask[i] = M;
        }
 
        // Track the top most source sub-lane - by setting the remaining to
        // UNDEF we can greatly simplify shuffle matching.
        int SrcSubLane = (SrcLane * SubLaneScale) + SubLane;
        TopSrcSubLane = std::max(TopSrcSubLane, SrcSubLane);
        Dst2SrcSubLanes[DstSubLane] = SrcSubLane;
        break;
      }
 
      // Bail if we failed to find a matching repeated sub-lane mask.
      if (Dst2SrcSubLanes[DstSubLane] < 0)
        return SDValue();
    }
    assert(0 <= TopSrcSubLane && TopSrcSubLane < NumSubLanes &&
           "Unexpected source lane");
 
    // Create a repeating shuffle mask for the entire vector.
    SmallVector<int, 8> RepeatedMask((unsigned)NumElts, -1);
    for (int SubLane = 0; SubLane <= TopSrcSubLane; ++SubLane) {
      int Lane = SubLane / SubLaneScale;
      auto &RepeatedSubLaneMask = RepeatedSubLaneMasks[SubLane % SubLaneScale];
      for (int Elt = 0; Elt != NumSubLaneElts; ++Elt) {
        int M = RepeatedSubLaneMask[Elt];
        if (M < 0)
          continue;
        int Idx = (SubLane * NumSubLaneElts) + Elt;
        RepeatedMask[Idx] = M + (Lane * NumLaneElts);
      }
    }
 
    // Shuffle each source sub-lane to its destination.
    SmallVector<int, 8> SubLaneMask((unsigned)NumElts, -1);
    for (int i = 0; i != NumElts; i += NumSubLaneElts) {
      int SrcSubLane = Dst2SrcSubLanes[i / NumSubLaneElts];
      if (SrcSubLane < 0)
        continue;
      for (int j = 0; j != NumSubLaneElts; ++j)
        SubLaneMask[i + j] = j + (SrcSubLane * NumSubLaneElts);
    }
 
    // Avoid returning the same shuffle operation.
    // v8i32 = vector_shuffle<0,1,4,5,2,3,6,7> t5, undef:v8i32
    if (RepeatedMask == Mask || SubLaneMask == Mask)
      return SDValue();
 
    SDValue RepeatedShuffle =
        DAG.getVectorShuffle(VT, DL, V1, V2, RepeatedMask);
 
    return DAG.getVectorShuffle(VT, DL, RepeatedShuffle, DAG.getUNDEF(VT),
                                SubLaneMask);
  };
 
  // On AVX2 targets we can permute 256-bit vectors as 64-bit sub-lanes
  // (with PERMQ/PERMPD). On AVX2/AVX512BW targets, permuting 32-bit sub-lanes,
  // even with a variable shuffle, can be worth it for v32i8/v64i8 vectors.
  // Otherwise we can only permute whole 128-bit lanes.
  int MinSubLaneScale = 1, MaxSubLaneScale = 1;
  if (Subtarget.hasAVX2() && VT.is256BitVector()) {
    bool OnlyLowestElts = isUndefOrInRange(Mask, 0, NumLaneElts);
    MinSubLaneScale = 2;
    MaxSubLaneScale =
        (!OnlyLowestElts && V2.isUndef() && VT == MVT::v32i8) ? 4 : 2;
  }
  if (Subtarget.hasBWI() && VT == MVT::v64i8)
    MinSubLaneScale = MaxSubLaneScale = 4;
 
  for (int Scale = MinSubLaneScale; Scale <= MaxSubLaneScale; Scale *= 2)
    if (SDValue Shuffle = ShuffleSubLanes(Scale))
      return Shuffle;
 
  return SDValue();
}
 
static bool matchShuffleWithSHUFPD(MVT VT, SDValue &V1, SDValue &V2,
                                   bool &ForceV1Zero, bool &ForceV2Zero,
                                   unsigned &ShuffleImm, ArrayRef<int> Mask,
                                   const APInt &Zeroable) {
  int NumElts = VT.getVectorNumElements();
  assert(VT.getScalarSizeInBits() == 64 &&
         (NumElts == 2 || NumElts == 4 || NumElts == 8) &&
         "Unexpected data type for VSHUFPD");
  assert(isUndefOrZeroOrInRange(Mask, 0, 2 * NumElts) &&
         "Illegal shuffle mask");
 
  bool ZeroLane[2] = { true, true };
  for (int i = 0; i < NumElts; ++i)
    ZeroLane[i & 1] &= Zeroable[i];
 
  // Mask for V8F64: 0/1,  8/9,  2/3,  10/11, 4/5, ..
  // Mask for V4F64; 0/1,  4/5,  2/3,  6/7..
  bool IsSHUFPD = true;
  bool IsCommutable = true;
  SmallVector<int, 8> SHUFPDMask(NumElts, -1);
  for (int i = 0; i < NumElts; ++i) {
    if (Mask[i] == SM_SentinelUndef || ZeroLane[i & 1])
      continue;
    if (Mask[i] < 0)
      return false;
    int Val = (i & 6) + NumElts * (i & 1);
    int CommutVal = (i & 0xe) + NumElts * ((i & 1) ^ 1);
    if (Mask[i] < Val || Mask[i] > Val + 1)
      IsSHUFPD = false;
    if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
      IsCommutable = false;
    SHUFPDMask[i] = Mask[i] % 2;
  }
 
  if (!IsSHUFPD && !IsCommutable)
    return false;
 
  if (!IsSHUFPD && IsCommutable)
    std::swap(V1, V2);
 
  ForceV1Zero = ZeroLane[0];
  ForceV2Zero = ZeroLane[1];
  ShuffleImm = getSHUFPDImm(SHUFPDMask);
  return true;
}
 
static SDValue lowerShuffleWithSHUFPD(const SDLoc &DL, MVT VT, SDValue V1,
                                      SDValue V2, ArrayRef<int> Mask,
                                      const APInt &Zeroable,
                                      const X86Subtarget &Subtarget,
                                      SelectionDAG &DAG) {
  assert((VT == MVT::v2f64 || VT == MVT::v4f64 || VT == MVT::v8f64) &&
         "Unexpected data type for VSHUFPD");
 
  unsigned Immediate = 0;
  bool ForceV1Zero = false, ForceV2Zero = false;
  if (!matchShuffleWithSHUFPD(VT, V1, V2, ForceV1Zero, ForceV2Zero, Immediate,
                              Mask, Zeroable))
    return SDValue();
 
  // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
  if (ForceV1Zero)
    V1 = getZeroVector(VT, Subtarget, DAG, DL);
  if (ForceV2Zero)
    V2 = getZeroVector(VT, Subtarget, DAG, DL);
 
  return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
                     DAG.getTargetConstant(Immediate, DL, MVT::i8));
}
 
// Look for {0, 8, 16, 24, 32, 40, 48, 56 } in the first 8 elements. Followed
// by zeroable elements in the remaining 24 elements. Turn this into two
// vmovqb instructions shuffled together.
static SDValue lowerShuffleAsVTRUNCAndUnpack(const SDLoc &DL, MVT VT,
                                             SDValue V1, SDValue V2,
                                             ArrayRef<int> Mask,
                                             const APInt &Zeroable,
                                             SelectionDAG &DAG) {
  assert(VT == MVT::v32i8 && "Unexpected type!");
 
  // The first 8 indices should be every 8th element.
  if (!isSequentialOrUndefInRange(Mask, 0, 8, 0, 8))
    return SDValue();
 
  // Remaining elements need to be zeroable.
  if (Zeroable.countl_one() < (Mask.size() - 8))
    return SDValue();
 
  V1 = DAG.getBitcast(MVT::v4i64, V1);
  V2 = DAG.getBitcast(MVT::v4i64, V2);
 
  V1 = DAG.getNode(X86ISD::VTRUNC, DL, MVT::v16i8, V1);
  V2 = DAG.getNode(X86ISD::VTRUNC, DL, MVT::v16i8, V2);
 
  // The VTRUNCs will put 0s in the upper 12 bytes. Use them to put zeroes in
  // the upper bits of the result using an unpckldq.
  SDValue Unpack = DAG.getVectorShuffle(MVT::v16i8, DL, V1, V2,
                                        { 0, 1, 2, 3, 16, 17, 18, 19,
                                          4, 5, 6, 7, 20, 21, 22, 23 });
  // Insert the unpckldq into a zero vector to widen to v32i8.
  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v32i8,
                     DAG.getConstant(0, DL, MVT::v32i8), Unpack,
                     DAG.getVectorIdxConstant(0, DL));
}
 
// a = shuffle v1, v2, mask1    ; interleaving lower lanes of v1 and v2
// b = shuffle v1, v2, mask2    ; interleaving higher lanes of v1 and v2
//     =>
// ul = unpckl v1, v2
// uh = unpckh v1, v2
// a = vperm ul, uh
// b = vperm ul, uh
//
// Pattern-match interleave(256b v1, 256b v2) -> 512b v3 and lower it into unpck
// and permute. We cannot directly match v3 because it is split into two
// 256-bit vectors in earlier isel stages. Therefore, this function matches a
// pair of 256-bit shuffles and makes sure the masks are consecutive.
//
// Once unpck and permute nodes are created, the permute corresponding to this
// shuffle is returned, while the other permute replaces the other half of the
// shuffle in the selection dag.
static SDValue lowerShufflePairAsUNPCKAndPermute(const SDLoc &DL, MVT VT,
                                                 SDValue V1, SDValue V2,
                                                 ArrayRef<int> Mask,
                                                 SelectionDAG &DAG) {
  if (VT != MVT::v8f32 && VT != MVT::v8i32 && VT != MVT::v16i16 &&
      VT != MVT::v32i8)
    return SDValue();
  // <B0, B1, B0+1, B1+1, ..., >
  auto IsInterleavingPattern = [&](ArrayRef<int> Mask, unsigned Begin0,
                                   unsigned Begin1) {
    size_t Size = Mask.size();
    assert(Size % 2 == 0 && "Expected even mask size");
    for (unsigned I = 0; I < Size; I += 2) {
      if (Mask[I] != (int)(Begin0 + I / 2) ||
          Mask[I + 1] != (int)(Begin1 + I / 2))
        return false;
    }
    return true;
  };
  // Check which half is this shuffle node
  int NumElts = VT.getVectorNumElements();
  size_t FirstQtr = NumElts / 2;
  size_t ThirdQtr = NumElts + NumElts / 2;
  bool IsFirstHalf = IsInterleavingPattern(Mask, 0, NumElts);
  bool IsSecondHalf = IsInterleavingPattern(Mask, FirstQtr, ThirdQtr);
  if (!IsFirstHalf && !IsSecondHalf)
    return SDValue();
 
  // Find the intersection between shuffle users of V1 and V2.
  SmallVector<SDNode *, 2> Shuffles;
  for (SDNode *User : V1->users())
    if (User->getOpcode() == ISD::VECTOR_SHUFFLE && User->getOperand(0) == V1 &&
        User->getOperand(1) == V2)
      Shuffles.push_back(User);
  // Limit user size to two for now.
  if (Shuffles.size() != 2)
    return SDValue();
  // Find out which half of the 512-bit shuffles is each smaller shuffle
  auto *SVN1 = cast<ShuffleVectorSDNode>(Shuffles[0]);
  auto *SVN2 = cast<ShuffleVectorSDNode>(Shuffles[1]);
  SDNode *FirstHalf;
  SDNode *SecondHalf;
  if (IsInterleavingPattern(SVN1->getMask(), 0, NumElts) &&
      IsInterleavingPattern(SVN2->getMask(), FirstQtr, ThirdQtr)) {
    FirstHalf = Shuffles[0];
    SecondHalf = Shuffles[1];
  } else if (IsInterleavingPattern(SVN1->getMask(), FirstQtr, ThirdQtr) &&
             IsInterleavingPattern(SVN2->getMask(), 0, NumElts)) {
    FirstHalf = Shuffles[1];
    SecondHalf = Shuffles[0];
  } else {
    return SDValue();
  }
  // Lower into unpck and perm. Return the perm of this shuffle and replace
  // the other.
  SDValue Unpckl = DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
  SDValue Unpckh = DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
  SDValue Perm1 = DAG.getNode(X86ISD::VPERM2X128, DL, VT, Unpckl, Unpckh,
                              DAG.getTargetConstant(0x20, DL, MVT::i8));
  SDValue Perm2 = DAG.getNode(X86ISD::VPERM2X128, DL, VT, Unpckl, Unpckh,
                              DAG.getTargetConstant(0x31, DL, MVT::i8));
  if (IsFirstHalf) {
    DAG.ReplaceAllUsesWith(SecondHalf, &Perm2);
    return Perm1;
  }
  DAG.ReplaceAllUsesWith(FirstHalf, &Perm1);
  return Perm2;
}
 
/// Handle lowering of 4-lane 64-bit floating point shuffles.
///
/// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
/// isn't available.
static SDValue lowerV4F64Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
  assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
 
  if (SDValue V = lowerV2X128Shuffle(DL, MVT::v4f64, V1, V2, Mask, Zeroable,
                                     Subtarget, DAG))
    return V;
 
  if (V2.isUndef()) {
    // Check for being able to broadcast a single element.
    if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4f64, V1, V2,
                                                    Mask, Subtarget, DAG))
      return Broadcast;
 
    // Use low duplicate instructions for masks that match their pattern.
    if (isShuffleEquivalent(Mask, {0, 0, 2, 2}, V1, V2))
      return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
 
    if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
      // Non-half-crossing single input shuffles can be lowered with an
      // interleaved permutation.
      unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
                              ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
      return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
                         DAG.getTargetConstant(VPERMILPMask, DL, MVT::i8));
    }
 
    // With AVX2 we have direct support for this permutation.
    if (Subtarget.hasAVX2())
      return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
                         getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
 
    // Try to create an in-lane repeating shuffle mask and then shuffle the
    // results into the target lanes.
    if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
            DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
      return V;
 
    // Try to permute the lanes and then use a per-lane permute.
    if (SDValue V = lowerShuffleAsLanePermuteAndPermute(DL, MVT::v4f64, V1, V2,
                                                        Mask, DAG, Subtarget))
      return V;
 
    // Otherwise, fall back.
    return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v4f64, V1, V2, Mask,
                                               DAG, Subtarget);
  }
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4f64, V1, V2, Mask, DAG))
    return V;
 
  if (SDValue Op = lowerShuffleWithSHUFPD(DL, MVT::v4f64, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Op;
 
  bool V1IsInPlace = isShuffleMaskInputInPlace(0, Mask);
  bool V2IsInPlace = isShuffleMaskInputInPlace(1, Mask);
  bool V1IsSplat = isShuffleMaskInputBroadcastable(0, Mask);
  bool V2IsSplat = isShuffleMaskInputBroadcastable(1, Mask);
 
  // If we have lane crossing shuffles AND they don't all come from the lower
  // lane elements, lower to SHUFPD(VPERM2F128(V1, V2), VPERM2F128(V1, V2)).
  // TODO: Handle BUILD_VECTOR sources which getVectorShuffle currently
  // canonicalize to a blend of splat which isn't necessary for this combine.
  if (is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask) &&
      !all_of(Mask, [](int M) { return M < 2 || (4 <= M && M < 6); }) &&
      (V1.getOpcode() != ISD::BUILD_VECTOR) &&
      (V2.getOpcode() != ISD::BUILD_VECTOR) &&
      (!Subtarget.hasAVX2() ||
       !((V1IsInPlace || V1IsSplat) && (V2IsInPlace || V2IsSplat))))
    return lowerShuffleAsLanePermuteAndSHUFP(DL, MVT::v4f64, V1, V2, Mask, DAG);
 
  // If we have one input in place, then we can permute the other input and
  // blend the result.
  if (V1IsInPlace || V2IsInPlace)
    return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4f64, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG);
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle. However, if we have AVX2 and either inputs are already in place,
  // we will be able to shuffle even across lanes the other input in a single
  // instruction so skip this pattern.
  if (!(Subtarget.hasAVX2() && (V1IsInPlace || V2IsInPlace)))
    if (SDValue V = lowerShuffleAsLanePermuteAndRepeatedMask(
            DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
      return V;
 
  // If we have VLX support, we can use VEXPAND.
  if (Subtarget.hasVLX())
    if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v4f64, V1, V2, Mask,
                                           Zeroable, Subtarget, DAG))
      return V;
 
  // If we have AVX2 then we always want to lower with a blend because an v4 we
  // can fully permute the elements.
  if (Subtarget.hasAVX2())
    return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4f64, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG);
 
  // Otherwise fall back on generic lowering.
  return lowerShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, Zeroable,
                                    Subtarget, DAG);
}
 
/// Handle lowering of 4-lane 64-bit integer shuffles.
///
/// This routine is only called when we have AVX2 and thus a reasonable
/// instruction set for v4i64 shuffling..
static SDValue lowerV4I64Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
  assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
  assert(Subtarget.hasAVX2() && "We can only lower v4i64 with AVX2!");
 
  if (SDValue V = lowerV2X128Shuffle(DL, MVT::v4i64, V1, V2, Mask, Zeroable,
                                     Subtarget, DAG))
    return V;
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  // Check for being able to broadcast a single element.
  if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v4i64, V1, V2, Mask,
                                                  Subtarget, DAG))
    return Broadcast;
 
  // Try to use shift instructions if fast.
  if (Subtarget.preferLowerShuffleAsShift())
    if (SDValue Shift =
            lowerShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, Zeroable,
                                Subtarget, DAG, /*BitwiseOnly*/ true))
      return Shift;
 
  if (V2.isUndef()) {
    // When the shuffle is mirrored between the 128-bit lanes of the unit, we
    // can use lower latency instructions that will operate on both lanes.
    SmallVector<int, 2> RepeatedMask;
    if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
      SmallVector<int, 4> PSHUFDMask;
      narrowShuffleMaskElts(2, RepeatedMask, PSHUFDMask);
      return DAG.getBitcast(
          MVT::v4i64,
          DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
                      DAG.getBitcast(MVT::v8i32, V1),
                      getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
    }
 
    // AVX2 provides a direct instruction for permuting a single input across
    // lanes.
    return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
                       getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
  }
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // If we have VLX support, we can use VALIGN or VEXPAND.
  if (Subtarget.hasVLX()) {
    if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v4i64, V1, V2, Mask,
                                              Zeroable, Subtarget, DAG))
      return Rotate;
 
    if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v4i64, V1, V2, Mask,
                                           Zeroable, Subtarget, DAG))
      return V;
  }
 
  // Try to use PALIGNR.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v4i64, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v4i64, V1, V2, Mask, DAG))
    return V;
 
  bool V1IsInPlace = isShuffleMaskInputInPlace(0, Mask);
  bool V2IsInPlace = isShuffleMaskInputInPlace(1, Mask);
 
  // If we have one input in place, then we can permute the other input and
  // blend the result.
  if (V1IsInPlace || V2IsInPlace)
    return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4i64, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG);
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to lower to PERMQ(BLENDD(V1,V2)).
  if (SDValue V =
          lowerShuffleAsBlendAndPermute(DL, MVT::v4i64, V1, V2, Mask, DAG))
    return V;
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle. However, if we have AVX2 and either inputs are already in place,
  // we will be able to shuffle even across lanes the other input in a single
  // instruction so skip this pattern.
  if (!V1IsInPlace && !V2IsInPlace)
    if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask(
            DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
      return Result;
 
  // Otherwise fall back on generic blend lowering.
  return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v4i64, V1, V2, Mask,
                                              Zeroable, Subtarget, DAG);
}
 
/// Handle lowering of 8-lane 32-bit floating point shuffles.
///
/// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
/// isn't available.
static SDValue lowerV8F32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
  assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  // Check for being able to broadcast a single element.
  if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8f32, V1, V2, Mask,
                                                  Subtarget, DAG))
    return Broadcast;
 
  if (!Subtarget.hasAVX2()) {
    SmallVector<int> InLaneMask;
    computeInLaneShuffleMask(Mask, Mask.size() / 2, InLaneMask);
 
    if (!is128BitLaneRepeatedShuffleMask(MVT::v8f32, InLaneMask))
      if (SDValue R = splitAndLowerShuffle(DL, MVT::v8f32, V1, V2, Mask, DAG,
                                           /*SimpleOnly*/ true))
        return R;
  }
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2, Mask,
                                                   Zeroable, Subtarget, DAG))
    return DAG.getBitcast(MVT::v8f32, ZExt);
 
  // If the shuffle mask is repeated in each 128-bit lane, we have many more
  // options to efficiently lower the shuffle.
  SmallVector<int, 4> RepeatedMask;
  if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
    assert(RepeatedMask.size() == 4 &&
           "Repeated masks must be half the mask width!");
 
    // Use even/odd duplicate instructions for masks that match their pattern.
    if (isShuffleEquivalent(RepeatedMask, {0, 0, 2, 2}, V1, V2))
      return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
    if (isShuffleEquivalent(RepeatedMask, {1, 1, 3, 3}, V1, V2))
      return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
 
    if (V2.isUndef())
      return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
                         getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
 
    // Use dedicated unpack instructions for masks that match their pattern.
    if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8f32, V1, V2, Mask, DAG))
      return V;
 
    // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
    // have already handled any direct blends.
    return lowerShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
  }
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // If we have a single input shuffle with different shuffle patterns in the
  // two 128-bit lanes use the variable mask to VPERMILPS.
  if (V2.isUndef()) {
    if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask)) {
      SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true);
      return DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v8f32, V1, VPermMask);
    }
    if (Subtarget.hasAVX2()) {
      SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true);
      return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8f32, VPermMask, V1);
    }
    // Otherwise, fall back.
    return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v8f32, V1, V2, Mask,
                                               DAG, Subtarget);
  }
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle.
  if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask(
          DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
    return Result;
 
  // If we have VLX support, we can use VEXPAND.
  if (Subtarget.hasVLX())
    if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v8f32, V1, V2, Mask,
                                           Zeroable, Subtarget, DAG))
      return V;
 
  // Try to match an interleave of two v8f32s and lower them as unpck and
  // permutes using ymms. This needs to go before we try to split the vectors.
  // Don't attempt on AVX1 if we're likely to split vectors anyway.
  if ((Subtarget.hasAVX2() ||
       !(isFreeToSplitVector(peekThroughBitcasts(V1), DAG) ||
         isFreeToSplitVector(peekThroughBitcasts(V2), DAG))) &&
      !Subtarget.hasAVX512())
    if (SDValue V = lowerShufflePairAsUNPCKAndPermute(DL, MVT::v8f32, V1, V2,
                                                      Mask, DAG))
      return V;
 
  // For non-AVX512 if the Mask is of 16bit elements in lane then try to split
  // since after split we get a more efficient code using vpunpcklwd and
  // vpunpckhwd instrs than vblend.
  if (!Subtarget.hasAVX512() && isUnpackWdShuffleMask(Mask, MVT::v8f32, DAG))
    return lowerShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, Zeroable,
                                      Subtarget, DAG);
 
  // If we have AVX2 then we always want to lower with a blend because at v8 we
  // can fully permute the elements.
  if (Subtarget.hasAVX2())
    return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v8f32, V1, V2, Mask,
                                                Zeroable, Subtarget, DAG);
 
  // Otherwise fall back on generic lowering.
  return lowerShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, Zeroable,
                                    Subtarget, DAG);
}
 
/// Handle lowering of 8-lane 32-bit integer shuffles.
///
/// This routine is only called when we have AVX2 and thus a reasonable
/// instruction set for v8i32 shuffling..
static SDValue lowerV8I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
  assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
  assert(Subtarget.hasAVX2() && "We can only lower v8i32 with AVX2!");
 
  int NumV2Elements = count_if(Mask, [](int M) { return M >= 8; });
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative. It also allows us to fold memory operands into the
  // shuffle in many cases.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2, Mask,
                                                   Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Try to match an interleave of two v8i32s and lower them as unpck and
  // permutes using ymms. This needs to go before we try to split the vectors.
  if (!Subtarget.hasAVX512())
    if (SDValue V = lowerShufflePairAsUNPCKAndPermute(DL, MVT::v8i32, V1, V2,
                                                      Mask, DAG))
      return V;
 
  // For non-AVX512 if the Mask is of 16bit elements in lane then try to split
  // since after split we get a more efficient code than vblend by using
  // vpunpcklwd and vpunpckhwd instrs.
  if (isUnpackWdShuffleMask(Mask, MVT::v8i32, DAG) && !V2.isUndef() &&
      !Subtarget.hasAVX512())
    return lowerShuffleAsSplitOrBlend(DL, MVT::v8i32, V1, V2, Mask, Zeroable,
                                      Subtarget, DAG);
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  // Check for being able to broadcast a single element.
  if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v8i32, V1, V2, Mask,
                                                  Subtarget, DAG))
    return Broadcast;
 
  // Try to use shift instructions if fast.
  if (Subtarget.preferLowerShuffleAsShift()) {
    if (SDValue Shift =
            lowerShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, Zeroable,
                                Subtarget, DAG, /*BitwiseOnly*/ true))
      return Shift;
    if (NumV2Elements == 0)
      if (SDValue Rotate =
              lowerShuffleAsBitRotate(DL, MVT::v8i32, V1, Mask, Subtarget, DAG))
        return Rotate;
  }
 
  // If the shuffle mask is repeated in each 128-bit lane we can use more
  // efficient instructions that mirror the shuffles across the two 128-bit
  // lanes.
  SmallVector<int, 4> RepeatedMask;
  bool Is128BitLaneRepeatedShuffle =
      is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask);
  if (Is128BitLaneRepeatedShuffle) {
    assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
    if (V2.isUndef())
      return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
                         getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
 
    // Use dedicated unpack instructions for masks that match their pattern.
    if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v8i32, V1, V2, Mask, DAG))
      return V;
  }
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  if (!Subtarget.preferLowerShuffleAsShift() && NumV2Elements == 0)
    if (SDValue Rotate =
            lowerShuffleAsBitRotate(DL, MVT::v8i32, V1, Mask, Subtarget, DAG))
      return Rotate;
 
  // If we have VLX support, we can use VALIGN or EXPAND.
  if (Subtarget.hasVLX()) {
    if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v8i32, V1, V2, Mask,
                                              Zeroable, Subtarget, DAG))
      return Rotate;
 
    if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v8i32, V1, V2, Mask,
                                           Zeroable, Subtarget, DAG))
      return V;
  }
 
  // Try to use byte rotation instructions.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i32, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  if (V2.isUndef()) {
    // Try to produce a fixed cross-128-bit lane permute followed by unpack
    // because that should be faster than the variable permute alternatives.
    if (SDValue V = lowerShuffleWithUNPCK256(DL, MVT::v8i32, V1, V2, Mask, DAG))
      return V;
 
    // If the shuffle patterns aren't repeated but it's a single input, directly
    // generate a cross-lane VPERMD instruction.
    SDValue VPermMask = getConstVector(Mask, MVT::v8i32, DAG, DL, true);
    return DAG.getNode(X86ISD::VPERMV, DL, MVT::v8i32, VPermMask, V1);
  }
 
  // Assume that a single SHUFPS is faster than an alternative sequence of
  // multiple instructions (even if the CPU has a domain penalty).
  // If some CPU is harmed by the domain switch, we can fix it in a later pass.
  if (Is128BitLaneRepeatedShuffle && isSingleSHUFPSMask(RepeatedMask)) {
    SDValue CastV1 = DAG.getBitcast(MVT::v8f32, V1);
    SDValue CastV2 = DAG.getBitcast(MVT::v8f32, V2);
    SDValue ShufPS = lowerShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask,
                                            CastV1, CastV2, DAG);
    return DAG.getBitcast(MVT::v8i32, ShufPS);
  }
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle.
  if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask(
          DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
    return Result;
 
  // Otherwise fall back on generic blend lowering.
  return lowerShuffleAsDecomposedShuffleMerge(DL, MVT::v8i32, V1, V2, Mask,
                                              Zeroable, Subtarget, DAG);
}
 
/// Handle lowering of 16-lane 16-bit integer shuffles.
///
/// This routine is only called when we have AVX2 and thus a reasonable
/// instruction set for v16i16 shuffling..
static SDValue lowerV16I16Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                  const APInt &Zeroable, SDValue V1, SDValue V2,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
  assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
  assert(Subtarget.hasAVX2() && "We can only lower v16i16 with AVX2!");
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative. It also allows us to fold memory operands into the
  // shuffle in many cases.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(
          DL, MVT::v16i16, V1, V2, Mask, Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Check for being able to broadcast a single element.
  if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v16i16, V1, V2, Mask,
                                                  Subtarget, DAG))
    return Broadcast;
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i16, V1, V2, Mask, DAG))
    return V;
 
  // Use dedicated pack instructions for masks that match their pattern.
  if (SDValue V =
          lowerShuffleWithPACK(DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to use lower using a truncation.
  if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v16i16, V1, V2, Mask, Zeroable,
                                       Subtarget, DAG))
    return V;
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, Zeroable,
                              Subtarget, DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // Try to use byte rotation instructions.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v16i16, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  if (V2.isUndef()) {
    // Try to use bit rotation instructions.
    if (SDValue Rotate =
            lowerShuffleAsBitRotate(DL, MVT::v16i16, V1, Mask, Subtarget, DAG))
      return Rotate;
 
    // Try to produce a fixed cross-128-bit lane permute followed by unpack
    // because that should be faster than the variable permute alternatives.
    if (SDValue V = lowerShuffleWithUNPCK256(DL, MVT::v16i16, V1, V2, Mask, DAG))
      return V;
 
    // There are no generalized cross-lane shuffle operations available on i16
    // element types.
    if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask)) {
      if (SDValue V = lowerShuffleAsLanePermuteAndPermute(
              DL, MVT::v16i16, V1, V2, Mask, DAG, Subtarget))
        return V;
 
      return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v16i16, V1, V2, Mask,
                                                 DAG, Subtarget);
    }
 
    SmallVector<int, 8> RepeatedMask;
    if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
      // As this is a single-input shuffle, the repeated mask should be
      // a strictly valid v8i16 mask that we can pass through to the v8i16
      // lowering to handle even the v16 case.
      return lowerV8I16GeneralSingleInputShuffle(
          DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
    }
  }
 
  if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v16i16, Mask, V1, V2,
                                              Zeroable, Subtarget, DAG))
    return PSHUFB;
 
  // AVX512BW can lower to VPERMW (non-VLX will pad to v32i16).
  if (Subtarget.hasBWI())
    return lowerShuffleWithPERMV(DL, MVT::v16i16, Mask, V1, V2, Subtarget, DAG);
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle.
  if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask(
          DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
    return Result;
 
  // Try to permute the lanes and then use a per-lane permute.
  if (SDValue V = lowerShuffleAsLanePermuteAndPermute(
          DL, MVT::v16i16, V1, V2, Mask, DAG, Subtarget))
    return V;
 
  // Try to match an interleave of two v16i16s and lower them as unpck and
  // permutes using ymms.
  if (!Subtarget.hasAVX512())
    if (SDValue V = lowerShufflePairAsUNPCKAndPermute(DL, MVT::v16i16, V1, V2,
                                                      Mask, DAG))
      return V;
 
  // Otherwise fall back on generic lowering.
  return lowerShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, Zeroable,
                                    Subtarget, DAG);
}
 
/// Handle lowering of 32-lane 8-bit integer shuffles.
///
/// This routine is only called when we have AVX2 and thus a reasonable
/// instruction set for v32i8 shuffling..
static SDValue lowerV32I8Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
  assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
  assert(Subtarget.hasAVX2() && "We can only lower v32i8 with AVX2!");
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative. It also allows us to fold memory operands into the
  // shuffle in many cases.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2, Mask,
                                                   Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Check for being able to broadcast a single element.
  if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, MVT::v32i8, V1, V2, Mask,
                                                  Subtarget, DAG))
    return Broadcast;
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v32i8, V1, V2, Mask, DAG))
    return V;
 
  // Use dedicated pack instructions for masks that match their pattern.
  if (SDValue V =
          lowerShuffleWithPACK(DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to use lower using a truncation.
  if (SDValue V = lowerShuffleAsVTRUNC(DL, MVT::v32i8, V1, V2, Mask, Zeroable,
                                       Subtarget, DAG))
    return V;
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // Try to use byte rotation instructions.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v32i8, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  // Try to use bit rotation instructions.
  if (V2.isUndef())
    if (SDValue Rotate =
            lowerShuffleAsBitRotate(DL, MVT::v32i8, V1, Mask, Subtarget, DAG))
      return Rotate;
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // There are no generalized cross-lane shuffle operations available on i8
  // element types.
  if (V2.isUndef() && is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask)) {
    // Try to produce a fixed cross-128-bit lane permute followed by unpack
    // because that should be faster than the variable permute alternatives.
    if (SDValue V = lowerShuffleWithUNPCK256(DL, MVT::v32i8, V1, V2, Mask, DAG))
      return V;
 
    if (SDValue V = lowerShuffleAsLanePermuteAndPermute(
            DL, MVT::v32i8, V1, V2, Mask, DAG, Subtarget))
      return V;
 
    return lowerShuffleAsLanePermuteAndShuffle(DL, MVT::v32i8, V1, V2, Mask,
                                               DAG, Subtarget);
  }
 
  if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v32i8, Mask, V1, V2,
                                              Zeroable, Subtarget, DAG))
    return PSHUFB;
 
  // AVX512VBMI can lower to VPERMB (non-VLX will pad to v64i8).
  if (Subtarget.hasVBMI())
    return lowerShuffleWithPERMV(DL, MVT::v32i8, Mask, V1, V2, Subtarget, DAG);
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle.
  if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask(
          DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
    return Result;
 
  // Try to permute the lanes and then use a per-lane permute.
  if (SDValue V = lowerShuffleAsLanePermuteAndPermute(
          DL, MVT::v32i8, V1, V2, Mask, DAG, Subtarget))
    return V;
 
  // Look for {0, 8, 16, 24, 32, 40, 48, 56 } in the first 8 elements. Followed
  // by zeroable elements in the remaining 24 elements. Turn this into two
  // vmovqb instructions shuffled together.
  if (Subtarget.hasVLX())
    if (SDValue V = lowerShuffleAsVTRUNCAndUnpack(DL, MVT::v32i8, V1, V2,
                                                  Mask, Zeroable, DAG))
      return V;
 
  // Try to match an interleave of two v32i8s and lower them as unpck and
  // permutes using ymms.
  if (!Subtarget.hasAVX512())
    if (SDValue V = lowerShufflePairAsUNPCKAndPermute(DL, MVT::v32i8, V1, V2,
                                                      Mask, DAG))
      return V;
 
  // Otherwise fall back on generic lowering.
  return lowerShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, Zeroable,
                                    Subtarget, DAG);
}
 
/// High-level routine to lower various 256-bit x86 vector shuffles.
///
/// This routine either breaks down the specific type of a 256-bit x86 vector
/// shuffle or splits it into two 128-bit shuffles and fuses the results back
/// together based on the available instructions.
static SDValue lower256BitShuffle(const SDLoc &DL, ArrayRef<int> Mask, MVT VT,
                                  SDValue V1, SDValue V2, const APInt &Zeroable,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  // If we have a single input to the zero element, insert that into V1 if we
  // can do so cheaply.
  int NumElts = VT.getVectorNumElements();
  int NumV2Elements = count_if(Mask, [NumElts](int M) { return M >= NumElts; });
 
  if (NumV2Elements == 1 && Mask[0] >= NumElts)
    if (SDValue Insertion = lowerShuffleAsElementInsertion(
            DL, VT, V1, V2, Mask, Zeroable, Subtarget, DAG))
      return Insertion;
 
  // Handle special cases where the lower or upper half is UNDEF.
  if (SDValue V =
          lowerShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // There is a really nice hard cut-over between AVX1 and AVX2 that means we
  // can check for those subtargets here and avoid much of the subtarget
  // querying in the per-vector-type lowering routines. With AVX1 we have
  // essentially *zero* ability to manipulate a 256-bit vector with integer
  // types. Since we'll use floating point types there eventually, just
  // immediately cast everything to a float and operate entirely in that domain.
  if (VT.isInteger() && !Subtarget.hasAVX2()) {
    int ElementBits = VT.getScalarSizeInBits();
    if (ElementBits < 32) {
      // No floating point type available, if we can't use the bit operations
      // for masking/blending then decompose into 128-bit vectors.
      if (SDValue V = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable,
                                            Subtarget, DAG))
        return V;
      if (SDValue V = lowerShuffleAsBitBlend(DL, VT, V1, V2, Mask, DAG))
        return V;
      return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG, /*SimpleOnly*/ false);
    }
 
    MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
                                VT.getVectorNumElements());
    V1 = DAG.getBitcast(FpVT, V1);
    V2 = DAG.getBitcast(FpVT, V2);
    return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
  }
 
  if (VT == MVT::v16f16 || VT == MVT::v16bf16) {
    V1 = DAG.getBitcast(MVT::v16i16, V1);
    V2 = DAG.getBitcast(MVT::v16i16, V2);
    return DAG.getBitcast(VT,
                          DAG.getVectorShuffle(MVT::v16i16, DL, V1, V2, Mask));
  }
 
  switch (VT.SimpleTy) {
  case MVT::v4f64:
    return lowerV4F64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v4i64:
    return lowerV4I64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v8f32:
    return lowerV8F32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v8i32:
    return lowerV8I32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v16i16:
    return lowerV16I16Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v32i8:
    return lowerV32I8Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
 
  default:
    llvm_unreachable("Not a valid 256-bit x86 vector type!");
  }
}
 
/// Try to lower a vector shuffle as a 128-bit shuffles.
static SDValue lowerV4X128Shuffle(const SDLoc &DL, MVT VT, ArrayRef<int> Mask,
                                  const APInt &Zeroable, SDValue V1, SDValue V2,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  assert(VT.getScalarSizeInBits() == 64 &&
         "Unexpected element type size for 128bit shuffle.");
 
  // To handle 256 bit vector requires VLX and most probably
  // function lowerV2X128VectorShuffle() is better solution.
  assert(VT.is512BitVector() && "Unexpected vector size for 512bit shuffle.");
 
  // TODO - use Zeroable like we do for lowerV2X128VectorShuffle?
  SmallVector<int, 4> Widened128Mask;
  if (!canWidenShuffleElements(Mask, Widened128Mask))
    return SDValue();
  assert(Widened128Mask.size() == 4 && "Shuffle widening mismatch");
 
  // Try to use an insert into a zero vector.
  if (Widened128Mask[0] == 0 && (Zeroable & 0xf0) == 0xf0 &&
      (Widened128Mask[1] == 1 || (Zeroable & 0x0c) == 0x0c)) {
    unsigned NumElts = ((Zeroable & 0x0c) == 0x0c) ? 2 : 4;
    MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), NumElts);
    SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
                              DAG.getVectorIdxConstant(0, DL));
    return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
                       getZeroVector(VT, Subtarget, DAG, DL), LoV,
                       DAG.getVectorIdxConstant(0, DL));
  }
 
  // Check for patterns which can be matched with a single insert of a 256-bit
  // subvector.
  bool OnlyUsesV1 = isShuffleEquivalent(Mask, {0, 1, 2, 3, 0, 1, 2, 3}, V1, V2);
  if (OnlyUsesV1 ||
      isShuffleEquivalent(Mask, {0, 1, 2, 3, 8, 9, 10, 11}, V1, V2)) {
    MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 4);
    SDValue SubVec =
        DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, OnlyUsesV1 ? V1 : V2,
                    DAG.getVectorIdxConstant(0, DL));
    return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, V1, SubVec,
                       DAG.getVectorIdxConstant(4, DL));
  }
 
  // See if this is an insertion of the lower 128-bits of V2 into V1.
  bool IsInsert = true;
  int V2Index = -1;
  for (int i = 0; i < 4; ++i) {
    assert(Widened128Mask[i] >= -1 && "Illegal shuffle sentinel value");
    if (Widened128Mask[i] < 0)
      continue;
 
    // Make sure all V1 subvectors are in place.
    if (Widened128Mask[i] < 4) {
      if (Widened128Mask[i] != i) {
        IsInsert = false;
        break;
      }
    } else {
      // Make sure we only have a single V2 index and its the lowest 128-bits.
      if (V2Index >= 0 || Widened128Mask[i] != 4) {
        IsInsert = false;
        break;
      }
      V2Index = i;
    }
  }
  if (IsInsert && V2Index >= 0) {
    MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(), 2);
    SDValue Subvec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V2,
                                 DAG.getVectorIdxConstant(0, DL));
    return insert128BitVector(V1, Subvec, V2Index * 2, DAG, DL);
  }
 
  // See if we can widen to a 256-bit lane shuffle, we're going to lose 128-lane
  // UNDEF info by lowering to X86ISD::SHUF128 anyway, so by widening where
  // possible we at least ensure the lanes stay sequential to help later
  // combines.
  SmallVector<int, 2> Widened256Mask;
  if (canWidenShuffleElements(Widened128Mask, Widened256Mask)) {
    Widened128Mask.clear();
    narrowShuffleMaskElts(2, Widened256Mask, Widened128Mask);
  }
 
  // Try to lower to vshuf64x2/vshuf32x4.
  SDValue Ops[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT)};
  int PermMask[4] = {-1, -1, -1, -1};
  // Ensure elements came from the same Op.
  for (int i = 0; i < 4; ++i) {
    assert(Widened128Mask[i] >= -1 && "Illegal shuffle sentinel value");
    if (Widened128Mask[i] < 0)
      continue;
 
    SDValue Op = Widened128Mask[i] >= 4 ? V2 : V1;
    unsigned OpIndex = i / 2;
    if (Ops[OpIndex].isUndef())
      Ops[OpIndex] = Op;
    else if (Ops[OpIndex] != Op)
      return SDValue();
 
    PermMask[i] = Widened128Mask[i] % 4;
  }
 
  return DAG.getNode(X86ISD::SHUF128, DL, VT, Ops[0], Ops[1],
                     getV4X86ShuffleImm8ForMask(PermMask, DL, DAG));
}
 
/// Handle lowering of 8-lane 64-bit floating point shuffles.
static SDValue lowerV8F64Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
  assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
 
  if (V2.isUndef()) {
    // Use low duplicate instructions for masks that match their pattern.
    if (isShuffleEquivalent(Mask, {0, 0, 2, 2, 4, 4, 6, 6}, V1, V2))
      return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v8f64, V1);
 
    if (!is128BitLaneCrossingShuffleMask(MVT::v8f64, Mask)) {
      // Non-half-crossing single input shuffles can be lowered with an
      // interleaved permutation.
      unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
                              ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3) |
                              ((Mask[4] == 5) << 4) | ((Mask[5] == 5) << 5) |
                              ((Mask[6] == 7) << 6) | ((Mask[7] == 7) << 7);
      return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f64, V1,
                         DAG.getTargetConstant(VPERMILPMask, DL, MVT::i8));
    }
 
    SmallVector<int, 4> RepeatedMask;
    if (is256BitLaneRepeatedShuffleMask(MVT::v8f64, Mask, RepeatedMask))
      return DAG.getNode(X86ISD::VPERMI, DL, MVT::v8f64, V1,
                         getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
  }
 
  if (SDValue Shuf128 = lowerV4X128Shuffle(DL, MVT::v8f64, Mask, Zeroable, V1,
                                           V2, Subtarget, DAG))
    return Shuf128;
 
  if (SDValue Unpck = lowerShuffleWithUNPCK(DL, MVT::v8f64, V1, V2, Mask, DAG))
    return Unpck;
 
  // Check if the blend happens to exactly fit that of SHUFPD.
  if (SDValue Op = lowerShuffleWithSHUFPD(DL, MVT::v8f64, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Op;
 
  if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v8f64, V1, V2, Mask, Zeroable,
                                         Subtarget, DAG))
    return V;
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8f64, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  return lowerShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, Subtarget, DAG);
}
 
/// Handle lowering of 16-lane 32-bit floating point shuffles.
static SDValue lowerV16F32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                  const APInt &Zeroable, SDValue V1, SDValue V2,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
  assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
 
  // If the shuffle mask is repeated in each 128-bit lane, we have many more
  // options to efficiently lower the shuffle.
  SmallVector<int, 4> RepeatedMask;
  if (is128BitLaneRepeatedShuffleMask(MVT::v16f32, Mask, RepeatedMask)) {
    assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
 
    // Use even/odd duplicate instructions for masks that match their pattern.
    if (isShuffleEquivalent(RepeatedMask, {0, 0, 2, 2}, V1, V2))
      return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v16f32, V1);
    if (isShuffleEquivalent(RepeatedMask, {1, 1, 3, 3}, V1, V2))
      return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v16f32, V1);
 
    if (V2.isUndef())
      return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v16f32, V1,
                         getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
 
    // Use dedicated unpack instructions for masks that match their pattern.
    if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16f32, V1, V2, Mask, DAG))
      return V;
 
    if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16f32, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
      return Blend;
 
    // Otherwise, fall back to a SHUFPS sequence.
    return lowerShuffleWithSHUFPS(DL, MVT::v16f32, RepeatedMask, V1, V2, DAG);
  }
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16f32, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(
          DL, MVT::v16i32, V1, V2, Mask, Zeroable, Subtarget, DAG))
    return DAG.getBitcast(MVT::v16f32, ZExt);
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v16f32, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // If we have a single input shuffle with different shuffle patterns in the
  // 128-bit lanes and don't lane cross, use variable mask VPERMILPS.
  if (V2.isUndef() &&
      !is128BitLaneCrossingShuffleMask(MVT::v16f32, Mask)) {
    SDValue VPermMask = getConstVector(Mask, MVT::v16i32, DAG, DL, true);
    return DAG.getNode(X86ISD::VPERMILPV, DL, MVT::v16f32, V1, VPermMask);
  }
 
  // If we have AVX512F support, we can use VEXPAND.
  if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v16f32, V1, V2, Mask,
                                         Zeroable, Subtarget, DAG))
    return V;
 
  return lowerShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, Subtarget, DAG);
}
 
/// Handle lowering of 8-lane 64-bit integer shuffles.
static SDValue lowerV8I64Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
  assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
 
  // Try to use shift instructions if fast.
  if (Subtarget.preferLowerShuffleAsShift())
    if (SDValue Shift =
            lowerShuffleAsShift(DL, MVT::v8i64, V1, V2, Mask, Zeroable,
                                Subtarget, DAG, /*BitwiseOnly*/ true))
      return Shift;
 
  if (V2.isUndef()) {
    // When the shuffle is mirrored between the 128-bit lanes of the unit, we
    // can use lower latency instructions that will operate on all four
    // 128-bit lanes.
    SmallVector<int, 2> Repeated128Mask;
    if (is128BitLaneRepeatedShuffleMask(MVT::v8i64, Mask, Repeated128Mask)) {
      SmallVector<int, 4> PSHUFDMask;
      narrowShuffleMaskElts(2, Repeated128Mask, PSHUFDMask);
      return DAG.getBitcast(
          MVT::v8i64,
          DAG.getNode(X86ISD::PSHUFD, DL, MVT::v16i32,
                      DAG.getBitcast(MVT::v16i32, V1),
                      getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
    }
 
    SmallVector<int, 4> Repeated256Mask;
    if (is256BitLaneRepeatedShuffleMask(MVT::v8i64, Mask, Repeated256Mask))
      return DAG.getNode(X86ISD::VPERMI, DL, MVT::v8i64, V1,
                         getV4X86ShuffleImm8ForMask(Repeated256Mask, DL, DAG));
  }
 
  if (SDValue Shuf128 = lowerV4X128Shuffle(DL, MVT::v8i64, Mask, Zeroable, V1,
                                           V2, Subtarget, DAG))
    return Shuf128;
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v8i64, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // Try to use VALIGN.
  if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v8i64, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
    return Rotate;
 
  // Try to use PALIGNR.
  if (Subtarget.hasBWI())
    if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v8i64, V1, V2, Mask,
                                                  Subtarget, DAG))
      return Rotate;
 
  if (SDValue Unpck = lowerShuffleWithUNPCK(DL, MVT::v8i64, V1, V2, Mask, DAG))
    return Unpck;
 
  // If we have AVX512F support, we can use VEXPAND.
  if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v8i64, V1, V2, Mask, Zeroable,
                                         Subtarget, DAG))
    return V;
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v8i64, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  return lowerShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, Subtarget, DAG);
}
 
/// Handle lowering of 16-lane 32-bit integer shuffles.
static SDValue lowerV16I32Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                  const APInt &Zeroable, SDValue V1, SDValue V2,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
  assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
 
  int NumV2Elements = count_if(Mask, [](int M) { return M >= 16; });
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative. It also allows us to fold memory operands into the
  // shuffle in many cases.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(
          DL, MVT::v16i32, V1, V2, Mask, Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Try to use shift instructions if fast.
  if (Subtarget.preferLowerShuffleAsShift()) {
    if (SDValue Shift =
            lowerShuffleAsShift(DL, MVT::v16i32, V1, V2, Mask, Zeroable,
                                Subtarget, DAG, /*BitwiseOnly*/ true))
      return Shift;
    if (NumV2Elements == 0)
      if (SDValue Rotate = lowerShuffleAsBitRotate(DL, MVT::v16i32, V1, Mask,
                                                   Subtarget, DAG))
        return Rotate;
  }
 
  // If the shuffle mask is repeated in each 128-bit lane we can use more
  // efficient instructions that mirror the shuffles across the four 128-bit
  // lanes.
  SmallVector<int, 4> RepeatedMask;
  bool Is128BitLaneRepeatedShuffle =
      is128BitLaneRepeatedShuffleMask(MVT::v16i32, Mask, RepeatedMask);
  if (Is128BitLaneRepeatedShuffle) {
    assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
    if (V2.isUndef())
      return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v16i32, V1,
                         getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
 
    // Use dedicated unpack instructions for masks that match their pattern.
    if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v16i32, V1, V2, Mask, DAG))
      return V;
  }
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v16i32, V1, V2, Mask, Zeroable,
                              Subtarget, DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  if (!Subtarget.preferLowerShuffleAsShift() && NumV2Elements != 0)
    if (SDValue Rotate =
            lowerShuffleAsBitRotate(DL, MVT::v16i32, V1, Mask, Subtarget, DAG))
      return Rotate;
 
  // Try to use VALIGN.
  if (SDValue Rotate = lowerShuffleAsVALIGN(DL, MVT::v16i32, V1, V2, Mask,
                                            Zeroable, Subtarget, DAG))
    return Rotate;
 
  // Try to use byte rotation instructions.
  if (Subtarget.hasBWI())
    if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v16i32, V1, V2, Mask,
                                                  Subtarget, DAG))
      return Rotate;
 
  // Assume that a single SHUFPS is faster than using a permv shuffle.
  // If some CPU is harmed by the domain switch, we can fix it in a later pass.
  if (Is128BitLaneRepeatedShuffle && isSingleSHUFPSMask(RepeatedMask)) {
    SDValue CastV1 = DAG.getBitcast(MVT::v16f32, V1);
    SDValue CastV2 = DAG.getBitcast(MVT::v16f32, V2);
    SDValue ShufPS = lowerShuffleWithSHUFPS(DL, MVT::v16f32, RepeatedMask,
                                            CastV1, CastV2, DAG);
    return DAG.getBitcast(MVT::v16i32, ShufPS);
  }
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
          DL, MVT::v16i32, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // If we have AVX512F support, we can use VEXPAND.
  if (SDValue V = lowerShuffleWithEXPAND(DL, MVT::v16i32, V1, V2, Mask,
                                         Zeroable, Subtarget, DAG))
    return V;
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v16i32, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  return lowerShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, Subtarget, DAG);
}
 
/// Handle lowering of 32-lane 16-bit integer shuffles.
static SDValue lowerV32I16Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                  const APInt &Zeroable, SDValue V1, SDValue V2,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
  assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
  assert(Subtarget.hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative. It also allows us to fold memory operands into the
  // shuffle in many cases.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(
          DL, MVT::v32i16, V1, V2, Mask, Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v32i16, V1, V2, Mask, DAG))
    return V;
 
  // Use dedicated pack instructions for masks that match their pattern.
  if (SDValue V =
          lowerShuffleWithPACK(DL, MVT::v32i16, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v32i16, V1, V2, Mask, Zeroable,
                              Subtarget, DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // Try to use byte rotation instructions.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v32i16, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  if (V2.isUndef()) {
    // Try to use bit rotation instructions.
    if (SDValue Rotate =
            lowerShuffleAsBitRotate(DL, MVT::v32i16, V1, Mask, Subtarget, DAG))
      return Rotate;
 
    SmallVector<int, 8> RepeatedMask;
    if (is128BitLaneRepeatedShuffleMask(MVT::v32i16, Mask, RepeatedMask)) {
      // As this is a single-input shuffle, the repeated mask should be
      // a strictly valid v8i16 mask that we can pass through to the v8i16
      // lowering to handle even the v32 case.
      return lowerV8I16GeneralSingleInputShuffle(DL, MVT::v32i16, V1,
                                                 RepeatedMask, Subtarget, DAG);
    }
  }
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v32i16, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v32i16, Mask, V1, V2,
                                              Zeroable, Subtarget, DAG))
    return PSHUFB;
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle.
  if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask(
          DL, MVT::v32i16, V1, V2, Mask, Subtarget, DAG))
    return Result;
 
  return lowerShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, Subtarget, DAG);
}
 
/// Handle lowering of 64-lane 8-bit integer shuffles.
static SDValue lowerV64I8Shuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                 const APInt &Zeroable, SDValue V1, SDValue V2,
                                 const X86Subtarget &Subtarget,
                                 SelectionDAG &DAG) {
  assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
  assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
  assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
  assert(Subtarget.hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
 
  // Whenever we can lower this as a zext, that instruction is strictly faster
  // than any alternative. It also allows us to fold memory operands into the
  // shuffle in many cases.
  if (SDValue ZExt = lowerShuffleAsZeroOrAnyExtend(
          DL, MVT::v64i8, V1, V2, Mask, Zeroable, Subtarget, DAG))
    return ZExt;
 
  // Use dedicated unpack instructions for masks that match their pattern.
  if (SDValue V = lowerShuffleWithUNPCK(DL, MVT::v64i8, V1, V2, Mask, DAG))
    return V;
 
  // Use dedicated pack instructions for masks that match their pattern.
  if (SDValue V =
          lowerShuffleWithPACK(DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Try to use shift instructions.
  if (SDValue Shift =
          lowerShuffleAsShift(DL, MVT::v64i8, V1, V2, Mask, Zeroable, Subtarget,
                              DAG, /*BitwiseOnly*/ false))
    return Shift;
 
  // Try to use byte rotation instructions.
  if (SDValue Rotate = lowerShuffleAsByteRotate(DL, MVT::v64i8, V1, V2, Mask,
                                                Subtarget, DAG))
    return Rotate;
 
  // Try to use bit rotation instructions.
  if (V2.isUndef())
    if (SDValue Rotate =
            lowerShuffleAsBitRotate(DL, MVT::v64i8, V1, Mask, Subtarget, DAG))
      return Rotate;
 
  // Lower as AND if possible.
  if (SDValue Masked = lowerShuffleAsBitMask(DL, MVT::v64i8, V1, V2, Mask,
                                             Zeroable, Subtarget, DAG))
    return Masked;
 
  if (SDValue PSHUFB = lowerShuffleWithPSHUFB(DL, MVT::v64i8, Mask, V1, V2,
                                              Zeroable, Subtarget, DAG))
    return PSHUFB;
 
  // Try to create an in-lane repeating shuffle mask and then shuffle the
  // results into the target lanes.
  // FIXME: Avoid on VBMI targets as the post lane permute often interferes
  // with shuffle combining (should be fixed by topological DAG sorting).
  if (!Subtarget.hasVBMI())
    if (SDValue V = lowerShuffleAsRepeatedMaskAndLanePermute(
            DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG))
      return V;
 
  if (SDValue Result = lowerShuffleAsLanePermuteAndPermute(
          DL, MVT::v64i8, V1, V2, Mask, DAG, Subtarget))
    return Result;
 
  if (SDValue Blend = lowerShuffleAsBlend(DL, MVT::v64i8, V1, V2, Mask,
                                          Zeroable, Subtarget, DAG))
    return Blend;
 
  if (!is128BitLaneCrossingShuffleMask(MVT::v64i8, Mask)) {
    // Use PALIGNR+Permute if possible - permute might become PSHUFB but the
    // PALIGNR will be cheaper than the second PSHUFB+OR.
    if (SDValue V = lowerShuffleAsByteRotateAndPermute(DL, MVT::v64i8, V1, V2,
                                                       Mask, Subtarget, DAG))
      return V;
 
    // VBMI can use VPERMV/VPERMV3 byte shuffles more efficiently than
    // OR(PSHUFB,PSHUFB).
    if (Subtarget.hasVBMI())
      return lowerShuffleWithPERMV(DL, MVT::v64i8, Mask, V1, V2, Subtarget,
                                   DAG);
 
    // If we can't directly blend but can use PSHUFB, that will be better as it
    // can both shuffle and set up the inefficient blend.
    bool V1InUse, V2InUse;
    return lowerShuffleAsBlendOfPSHUFBs(DL, MVT::v64i8, V1, V2, Mask, Zeroable,
                                        DAG, V1InUse, V2InUse);
  }
 
  // Try to simplify this by merging 128-bit lanes to enable a lane-based
  // shuffle.
  if (SDValue Result = lowerShuffleAsLanePermuteAndRepeatedMask(
          DL, MVT::v64i8, V1, V2, Mask, Subtarget, DAG))
    return Result;
 
  // VBMI can use VPERMV/VPERMV3 byte shuffles.
  if (Subtarget.hasVBMI())
    return lowerShuffleWithPERMV(DL, MVT::v64i8, Mask, V1, V2, Subtarget, DAG);
 
  return splitAndLowerShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG,
                              /*SimpleOnly*/ false);
}
 
/// High-level routine to lower various 512-bit x86 vector shuffles.
///
/// This routine either breaks down the specific type of a 512-bit x86 vector
/// shuffle or splits it into two 256-bit shuffles and fuses the results back
/// together based on the available instructions.
static SDValue lower512BitShuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                  MVT VT, SDValue V1, SDValue V2,
                                  const APInt &Zeroable,
                                  const X86Subtarget &Subtarget,
                                  SelectionDAG &DAG) {
  assert(Subtarget.hasAVX512() &&
         "Cannot lower 512-bit vectors w/ basic ISA!");
 
  // If we have a single input to the zero element, insert that into V1 if we
  // can do so cheaply.
  int NumElts = Mask.size();
  int NumV2Elements = count_if(Mask, [NumElts](int M) { return M >= NumElts; });
 
  if (NumV2Elements == 1 && Mask[0] >= NumElts)
    if (SDValue Insertion = lowerShuffleAsElementInsertion(
            DL, VT, V1, V2, Mask, Zeroable, Subtarget, DAG))
      return Insertion;
 
  // Handle special cases where the lower or upper half is UNDEF.
  if (SDValue V =
          lowerShuffleWithUndefHalf(DL, VT, V1, V2, Mask, Subtarget, DAG))
    return V;
 
  // Check for being able to broadcast a single element.
  if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, VT, V1, V2, Mask,
                                                  Subtarget, DAG))
    return Broadcast;
 
  if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI()) {
    // Try using bit ops for masking and blending before falling back to
    // splitting.
    if (SDValue V = lowerShuffleAsBitMask(DL, VT, V1, V2, Mask, Zeroable,
                                          Subtarget, DAG))
      return V;
    if (SDValue V = lowerShuffleAsBitBlend(DL, VT, V1, V2, Mask, DAG))
      return V;
 
    return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG, /*SimpleOnly*/ false);
  }
 
  if (VT == MVT::v32f16 || VT == MVT::v32bf16) {
    if (!Subtarget.hasBWI())
      return splitAndLowerShuffle(DL, VT, V1, V2, Mask, DAG,
                                  /*SimpleOnly*/ false);
 
    V1 = DAG.getBitcast(MVT::v32i16, V1);
    V2 = DAG.getBitcast(MVT::v32i16, V2);
    return DAG.getBitcast(VT,
                          DAG.getVectorShuffle(MVT::v32i16, DL, V1, V2, Mask));
  }
 
  // Dispatch to each element type for lowering. If we don't have support for
  // specific element type shuffles at 512 bits, immediately split them and
  // lower them. Each lowering routine of a given type is allowed to assume that
  // the requisite ISA extensions for that element type are available.
  switch (VT.SimpleTy) {
  case MVT::v8f64:
    return lowerV8F64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v16f32:
    return lowerV16F32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v8i64:
    return lowerV8I64Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v16i32:
    return lowerV16I32Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v32i16:
    return lowerV32I16Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
  case MVT::v64i8:
    return lowerV64I8Shuffle(DL, Mask, Zeroable, V1, V2, Subtarget, DAG);
 
  default:
    llvm_unreachable("Not a valid 512-bit x86 vector type!");
  }
}
 
static SDValue lower1BitShuffleAsKSHIFTR(const SDLoc &DL, ArrayRef<int> Mask,
                                         MVT VT, SDValue V1, SDValue V2,
                                         const X86Subtarget &Subtarget,
                                         SelectionDAG &DAG) {
  // Shuffle should be unary.
  if (!V2.isUndef())
    return SDValue();
 
  int ShiftAmt = -1;
  int NumElts = Mask.size();
  for (int i = 0; i != NumElts; ++i) {
    int M = Mask[i];
    assert((M == SM_SentinelUndef || (0 <= M && M < NumElts)) &&
           "Unexpected mask index.");
    if (M < 0)
      continue;
 
    // The first non-undef element determines our shift amount.
    if (ShiftAmt < 0) {
      ShiftAmt = M - i;
      // Need to be shifting right.
      if (ShiftAmt <= 0)
        return SDValue();
    }
    // All non-undef elements must shift by the same amount.
    if (ShiftAmt != M - i)
      return SDValue();
  }
  assert(ShiftAmt >= 0 && "All undef?");
 
  // Great we found a shift right.
  SDValue Res = widenMaskVector(V1, false, Subtarget, DAG, DL);
  Res = DAG.getNode(X86ISD::KSHIFTR, DL, Res.getValueType(), Res,
                    DAG.getTargetConstant(ShiftAmt, DL, MVT::i8));
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res,
                     DAG.getVectorIdxConstant(0, DL));
}
 
// Determine if this shuffle can be implemented with a KSHIFT instruction.
// Returns the shift amount if possible or -1 if not. This is a simplified
// version of matchShuffleAsShift.
static int match1BitShuffleAsKSHIFT(unsigned &Opcode, ArrayRef<int> Mask,
                                    int MaskOffset, const APInt &Zeroable) {
  int Size = Mask.size();
 
  auto CheckZeros = [&](int Shift, bool Left) {
    for (int j = 0; j < Shift; ++j)
      if (!Zeroable[j + (Left ? 0 : (Size - Shift))])
        return false;
 
    return true;
  };
 
  auto MatchShift = [&](int Shift, bool Left) {
    unsigned Pos = Left ? Shift : 0;
    unsigned Low = Left ? 0 : Shift;
    unsigned Len = Size - Shift;
    return isSequentialOrUndefInRange(Mask, Pos, Len, Low + MaskOffset);
  };
 
  for (int Shift = 1; Shift != Size; ++Shift)
    for (bool Left : {true, false})
      if (CheckZeros(Shift, Left) && MatchShift(Shift, Left)) {
        Opcode = Left ? X86ISD::KSHIFTL : X86ISD::KSHIFTR;
        return Shift;
      }
 
  return -1;
}
 
 
// Lower vXi1 vector shuffles.
// There is no a dedicated instruction on AVX-512 that shuffles the masks.
// The only way to shuffle bits is to sign-extend the mask vector to SIMD
// vector, shuffle and then truncate it back.
static SDValue lower1BitShuffle(const SDLoc &DL, ArrayRef<int> Mask,
                                MVT VT, SDValue V1, SDValue V2,
                                const APInt &Zeroable,
                                const X86Subtarget &Subtarget,
                                SelectionDAG &DAG) {
  assert(Subtarget.hasAVX512() &&
         "Cannot lower 512-bit vectors w/o basic ISA!");
 
  int NumElts = Mask.size();
  int NumV2Elements = count_if(Mask, [NumElts](int M) { return M >= NumElts; });
 
  // Try to recognize shuffles that are just padding a subvector with zeros.
  int SubvecElts = 0;
  int Src = -1;
  for (int i = 0; i != NumElts; ++i) {
    if (Mask[i] >= 0) {
      // Grab the source from the first valid mask. All subsequent elements need
      // to use this same source.
      if (Src < 0)
        Src = Mask[i] / NumElts;
      if (Src != (Mask[i] / NumElts) || (Mask[i] % NumElts) != i)
        break;
    }
 
    ++SubvecElts;
  }
  assert(SubvecElts != NumElts && "Identity shuffle?");
 
  // Clip to a power 2.
  SubvecElts = llvm::bit_floor<uint32_t>(SubvecElts);
 
  // Make sure the number of zeroable bits in the top at least covers the bits
  // not covered by the subvector.
  if ((int)Zeroable.countl_one() >= (NumElts - SubvecElts)) {
    assert(Src >= 0 && "Expected a source!");
    MVT ExtractVT = MVT::getVectorVT(MVT::i1, SubvecElts);
    SDValue Extract =
        DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ExtractVT, Src == 0 ? V1 : V2,
                    DAG.getVectorIdxConstant(0, DL));
    return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT,
                       DAG.getConstant(0, DL, VT), Extract,
                       DAG.getVectorIdxConstant(0, DL));
  }
 
  // Try a simple shift right with undef elements. Later we'll try with zeros.
  if (SDValue Shift =
          lower1BitShuffleAsKSHIFTR(DL, Mask, VT, V1, V2, Subtarget, DAG))
    return Shift;
 
  // Try to match KSHIFTs.
  unsigned Offset = 0;
  for (SDValue V : {V1, V2}) {
    unsigned Opcode;
    int ShiftAmt = match1BitShuffleAsKSHIFT(Opcode, Mask, Offset, Zeroable);
    if (ShiftAmt >= 0) {
      SDValue Res = widenMaskVector(V, false, Subtarget, DAG, DL);
      MVT WideVT = Res.getSimpleValueType();
      // Widened right shifts need two shifts to ensure we shift in zeroes.
      if (Opcode == X86ISD::KSHIFTR && WideVT != VT) {
        int WideElts = WideVT.getVectorNumElements();
        // Shift left to put the original vector in the MSBs of the new size.
        Res =
            DAG.getNode(X86ISD::KSHIFTL, DL, WideVT, Res,
                        DAG.getTargetConstant(WideElts - NumElts, DL, MVT::i8));
        // Increase the shift amount to account for the left shift.
        ShiftAmt += WideElts - NumElts;
      }
 
      Res = DAG.getNode(Opcode, DL, WideVT, Res,
                        DAG.getTargetConstant(ShiftAmt, DL, MVT::i8));
      return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res,
                         DAG.getVectorIdxConstant(0, DL));
    }
    Offset += NumElts; // Increment for next iteration.
  }
 
  // If we're performing an unary shuffle on a SETCC result, try to shuffle the
  // ops instead.
  // TODO: What other unary shuffles would benefit from this?
  if (NumV2Elements == 0 && V1.getOpcode() == ISD::SETCC && V1->hasOneUse()) {
    SDValue Op0 = V1.getOperand(0);
    SDValue Op1 = V1.getOperand(1);
    ISD::CondCode CC = cast<CondCodeSDNode>(V1.getOperand(2))->get();
    EVT OpVT = Op0.getValueType();
    if (OpVT.getScalarSizeInBits() >= 32 || isBroadcastShuffleMask(Mask))
      return DAG.getSetCC(
          DL, VT, DAG.getVectorShuffle(OpVT, DL, Op0, DAG.getUNDEF(OpVT), Mask),
          DAG.getVectorShuffle(OpVT, DL, Op1, DAG.getUNDEF(OpVT), Mask), CC);
  }
 
  // If this is a sequential shuffle with zero'd elements - then lower to AND.
  bool IsBlendWithZero = all_of(enumerate(Mask), [&Zeroable](auto M) {
    return Zeroable[M.index()] || (M.value() == (int)M.index());
  });
  if (IsBlendWithZero) {
    const unsigned Width = std::max<unsigned>(NumElts, 8u);
    MVT IntVT = MVT::getIntegerVT(Width);
 
    APInt MaskValue = (~Zeroable).zextOrTrunc(Width);
    SDValue MaskNode = DAG.getConstant(MaskValue, DL, IntVT);
 
    MVT MaskVecVT = MVT::getVectorVT(MVT::i1, Width);
    SDValue MaskVecNode = DAG.getBitcast(MaskVecVT, MaskNode);
 
    SDValue MaskVec = DAG.getExtractSubvector(DL, VT, MaskVecNode, 0);
    return DAG.getNode(ISD::AND, DL, VT, V1, MaskVec);
  }
 
  MVT ExtVT;
  switch (VT.SimpleTy) {
  default:
    llvm_unreachable("Expected a vector of i1 elements");
  case MVT::v2i1:
    ExtVT = MVT::v2i64;
    break;
  case MVT::v4i1:
    ExtVT = MVT::v4i32;
    break;
  case MVT::v8i1:
    // Take 512-bit type, more shuffles on KNL. If we have VLX use a 256-bit
    // shuffle.
    ExtVT = Subtarget.hasVLX() ? MVT::v8i32 : MVT::v8i64;
    break;
  case MVT::v16i1:
    // Take 512-bit type, unless we are avoiding 512-bit types and have the
    // 256-bit operation available.
    ExtVT = Subtarget.canExtendTo512DQ() ? MVT::v16i32 : MVT::v16i16;
    break;
  case MVT::v32i1:
    // Take 512-bit type, unless we are avoiding 512-bit types and have the
    // 256-bit operation available.
    assert(Subtarget.hasBWI() && "Expected AVX512BW support");
    ExtVT = Subtarget.canExtendTo512BW() ? MVT::v32i16 : MVT::v32i8;
    break;
  case MVT::v64i1:
    // Fall back to scalarization. FIXME: We can do better if the shuffle
    // can be partitioned cleanly.
    if (!Subtarget.useBWIRegs())
      return SDValue();
    ExtVT = MVT::v64i8;
    break;
  }
 
  V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
  V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
 
  SDValue Shuffle = DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask);
  // i1 was sign extended we can use X86ISD::CVT2MASK.
  int NumElems = VT.getVectorNumElements();
  if ((Subtarget.hasBWI() && (NumElems >= 32)) ||
      (Subtarget.hasDQI() && (NumElems < 32)))
    return DAG.getSetCC(DL, VT, DAG.getConstant(0, DL, ExtVT),
                       Shuffle, ISD::SETGT);
 
  return DAG.getNode(ISD::TRUNCATE, DL, VT, Shuffle);
}
 
/// Helper function that returns true if the shuffle mask should be
/// commuted to improve canonicalization.
static bool canonicalizeShuffleMaskWithCommute(ArrayRef<int> Mask) {
  int NumElements = Mask.size();
 
  int NumV1Elements = 0, NumV2Elements = 0;
  for (int M : Mask)
    if (M < 0)
      continue;
    else if (M < NumElements)
      ++NumV1Elements;
    else
      ++NumV2Elements;
 
  // Commute the shuffle as needed such that more elements come from V1 than
  // V2. This allows us to match the shuffle pattern strictly on how many
  // elements come from V1 without handling the symmetric cases.
  if (NumV2Elements > NumV1Elements)
    return true;
 
  assert(NumV1Elements > 0 && "No V1 indices");
 
  if (NumV2Elements == 0)
    return false;
 
  // When the number of V1 and V2 elements are the same, try to minimize the
  // number of uses of V2 in the low half of the vector. When that is tied,
  // ensure that the sum of indices for V1 is equal to or lower than the sum
  // indices for V2. When those are equal, try to ensure that the number of odd
  // indices for V1 is lower than the number of odd indices for V2.
  if (NumV1Elements == NumV2Elements) {
    int LowV1Elements = 0, LowV2Elements = 0;
    for (int M : Mask.slice(0, NumElements / 2))
      if (M >= NumElements)
        ++LowV2Elements;
      else if (M >= 0)
        ++LowV1Elements;
    if (LowV2Elements > LowV1Elements)
      return true;
    if (LowV2Elements == LowV1Elements) {
      int SumV1Indices = 0, SumV2Indices = 0;
      for (int i = 0, Size = Mask.size(); i < Size; ++i)
        if (Mask[i] >= NumElements)
          SumV2Indices += i;
        else if (Mask[i] >= 0)
          SumV1Indices += i;
      if (SumV2Indices < SumV1Indices)
        return true;
      if (SumV2Indices == SumV1Indices) {
        int NumV1OddIndices = 0, NumV2OddIndices = 0;
        for (int i = 0, Size = Mask.size(); i < Size; ++i)
          if (Mask[i] >= NumElements)
            NumV2OddIndices += i % 2;
          else if (Mask[i] >= 0)
            NumV1OddIndices += i % 2;
        if (NumV2OddIndices < NumV1OddIndices)
          return true;
      }
    }
  }
 
  return false;
}
 
static bool canCombineAsMaskOperation(SDValue V,
                                      const X86Subtarget &Subtarget) {
  if (!Subtarget.hasAVX512())
    return false;
 
  if (!V.getValueType().isSimple())
    return false;
 
  MVT VT = V.getSimpleValueType().getScalarType();
  if ((VT == MVT::i16 || VT == MVT::i8) && !Subtarget.hasBWI())
    return false;
 
  // If vec width < 512, widen i8/i16 even with BWI as blendd/blendps/blendpd
  // are preferable to blendw/blendvb/masked-mov.
  if ((VT == MVT::i16 || VT == MVT::i8) &&
      V.getSimpleValueType().getSizeInBits() < 512)
    return false;
 
  auto HasMaskOperation = [&](SDValue V) {
    // TODO: Currently we only check limited opcode. We probably extend
    // it to all binary operation by checking TLI.isBinOp().
    switch (V->getOpcode()) {
    default:
      return false;
    case ISD::ADD:
    case ISD::SUB:
    case ISD::AND:
    case ISD::XOR:
    case ISD::OR:
    case ISD::SMAX:
    case ISD::SMIN:
    case ISD::UMAX:
    case ISD::UMIN:
    case ISD::ABS:
    case ISD::SHL:
    case ISD::SRL:
    case ISD::SRA:
    case ISD::MUL:
      break;
    }
    if (!V->hasOneUse())
      return false;
 
    return true;
  };
 
  if (HasMaskOperation(V))
    return true;
 
  return false;
}
 
// Forward declaration.
static SDValue canonicalizeShuffleMaskWithHorizOp(
    MutableArrayRef<SDValue> Ops, MutableArrayRef<int> Mask,
    unsigned RootSizeInBits, const SDLoc &DL, SelectionDAG &DAG,
    const X86Subtarget &Subtarget);
 
    /// Top-level lowering for x86 vector shuffles.
///
/// This handles decomposition, canonicalization, and lowering of all x86
/// vector shuffles. Most of the specific lowering strategies are encapsulated
/// above in helper routines. The canonicalization attempts to widen shuffles
/// to involve fewer lanes of wider elements, consolidate symmetric patterns
/// s.t. only one of the two inputs needs to be tested, etc.
static SDValue lowerVECTOR_SHUFFLE(SDValue Op, const X86Subtarget &Subtarget,
                                   SelectionDAG &DAG) {
  ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
  ArrayRef<int> OrigMask = SVOp->getMask();
  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);
  MVT VT = Op.getSimpleValueType();
  int NumElements = VT.getVectorNumElements();
  SDLoc DL(Op);
  bool Is1BitVector = (VT.getVectorElementType() == MVT::i1);
 
  assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
         "Can't lower MMX shuffles");
 
  bool V1IsUndef = V1.isUndef();
  bool V2IsUndef = V2.isUndef();
  if (V1IsUndef && V2IsUndef)
    return DAG.getUNDEF(VT);
 
  // When we create a shuffle node we put the UNDEF node to second operand,
  // but in some cases the first operand may be transformed to UNDEF.
  // In this case we should just commute the node.
  if (V1IsUndef)
    return DAG.getCommutedVectorShuffle(*SVOp);
 
  // Check for non-undef masks pointing at an undef vector and make the masks
  // undef as well. This makes it easier to match the shuffle based solely on
  // the mask.
  if (V2IsUndef &&
      any_of(OrigMask, [NumElements](int M) { return M >= NumElements; })) {
    SmallVector<int, 8> NewMask(OrigMask);
    for (int &M : NewMask)
      if (M >= NumElements)
        M = -1;
    return DAG.getVectorShuffle(VT, DL, V1, V2, NewMask);
  }
 
  // Check for illegal shuffle mask element index values.
  int MaskUpperLimit = OrigMask.size() * (V2IsUndef ? 1 : 2);
  (void)MaskUpperLimit;
  assert(llvm::all_of(OrigMask,
                      [&](int M) { return -1 <= M && M < MaskUpperLimit; }) &&
         "Out of bounds shuffle index");
 
  // We actually see shuffles that are entirely re-arrangements of a set of
  // zero inputs. This mostly happens while decomposing complex shuffles into
  // simple ones. Directly lower these as a buildvector of zeros.
  APInt KnownUndef, KnownZero;
  computeZeroableShuffleElements(OrigMask, V1, V2, KnownUndef, KnownZero);
 
  APInt Zeroable = KnownUndef | KnownZero;
  if (Zeroable.isAllOnes())
    return getZeroVector(VT, Subtarget, DAG, DL);
 
  bool V2IsZero = !V2IsUndef && ISD::isBuildVectorAllZeros(V2.getNode());
 
  // Try to collapse shuffles into using a vector type with fewer elements but
  // wider element types. We cap this to not form integers or floating point
  // elements wider than 64 bits. It does not seem beneficial to form i128
  // integers to handle flipping the low and high halves of AVX 256-bit vectors.
  SmallVector<int, 16> WidenedMask;
  if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
      !canCombineAsMaskOperation(V1, Subtarget) &&
      !canCombineAsMaskOperation(V2, Subtarget) &&
      canWidenShuffleElements(OrigMask, Zeroable, V2IsZero, WidenedMask)) {
    // Shuffle mask widening should not interfere with a broadcast opportunity
    // by obfuscating the operands with bitcasts.
    // TODO: Avoid lowering directly from this top-level function: make this
    // a query (canLowerAsBroadcast) and defer lowering to the type-based calls.
    if (SDValue Broadcast = lowerShuffleAsBroadcast(DL, VT, V1, V2, OrigMask,
                                                    Subtarget, DAG))
      return Broadcast;
 
    MVT NewEltVT = VT.isFloatingPoint()
                       ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
                       : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
    int NewNumElts = NumElements / 2;
    MVT NewVT = MVT::getVectorVT(NewEltVT, NewNumElts);
    // Make sure that the new vector type is legal. For example, v2f64 isn't
    // legal on SSE1.
    if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
      if (V2IsZero) {
        // Modify the new Mask to take all zeros from the all-zero vector.
        // Choose indices that are blend-friendly.
        bool UsedZeroVector = false;
        assert(is_contained(WidenedMask, SM_SentinelZero) &&
               "V2's non-undef elements are used?!");
        for (int i = 0; i != NewNumElts; ++i)
          if (WidenedMask[i] == SM_SentinelZero) {
            WidenedMask[i] = i + NewNumElts;
            UsedZeroVector = true;
          }
        // Ensure all elements of V2 are zero - isBuildVectorAllZeros permits
        // some elements to be undef.
        if (UsedZeroVector)
          V2 = getZeroVector(NewVT, Subtarget, DAG, DL);
      }
      V1 = DAG.getBitcast(NewVT, V1);
      V2 = DAG.getBitcast(NewVT, V2);
      return DAG.getBitcast(
          VT, DAG.getVectorShuffle(NewVT, DL, V1, V2, WidenedMask));
    }
  }
 
  SmallVector<SDValue> Ops = {V1, V2};
  SmallVector<int> Mask(OrigMask);
 
  // Canonicalize the shuffle with any horizontal ops inputs.
  // Don't attempt this if the shuffle can still be widened as we may lose
  // whole lane shuffle patterns.
  // NOTE: This may update Ops and Mask.
  if (!canWidenShuffleElements(Mask)) {
    if (SDValue HOp = canonicalizeShuffleMaskWithHorizOp(
            Ops, Mask, VT.getSizeInBits(), DL, DAG, Subtarget))
      return DAG.getBitcast(VT, HOp);
 
    V1 = DAG.getBitcast(VT, Ops[0]);
    V2 = DAG.getBitcast(VT, Ops[1]);
    assert(NumElements == (int)Mask.size() &&
           "canonicalizeShuffleMaskWithHorizOp "
           "shouldn't alter the shuffle mask size");
  }
 
  // Canonicalize zeros/ones/fp splat constants to ensure no undefs.
  // These will be materialized uniformly anyway, so make splat matching easier.
  // TODO: Allow all int constants?
  auto CanonicalizeConstant = [VT, &DL, &DAG](SDValue V) {
    if (auto *BV = dyn_cast<BuildVectorSDNode>(V)) {
      BitVector Undefs;
      if (SDValue Splat = BV->getSplatValue(&Undefs)) {
        if (Undefs.any() &&
            (isNullConstant(Splat) || isAllOnesConstant(Splat) ||
             isa<ConstantFPSDNode>(Splat))) {
          V = DAG.getBitcast(VT, DAG.getSplat(BV->getValueType(0), DL, Splat));
        }
      }
    }
    return V;
  };
  V1 = CanonicalizeConstant(V1);
  V2 = CanonicalizeConstant(V2);
 
  // Commute the shuffle if it will improve canonicalization.
  if (canonicalizeShuffleMaskWithCommute(Mask)) {
    ShuffleVectorSDNode::commuteMask(Mask);
    std::swap(V1, V2);
  }
 
  // For each vector width, delegate to a specialized lowering routine.
  if (VT.is128BitVector())
    return lower128BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG);
 
  if (VT.is256BitVector())
    return lower256BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG);
 
  if (VT.is512BitVector())
    return lower512BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG);
 
  if (Is1BitVector)
    return lower1BitShuffle(DL, Mask, VT, V1, V2, Zeroable, Subtarget, DAG);
 
  llvm_unreachable("Unimplemented!");
}
 
// As legal vpcompress instructions depend on various AVX512 extensions, try to
// convert illegal vector sizes to legal ones to avoid expansion.
static SDValue lowerVECTOR_COMPRESS(SDValue Op, const X86Subtarget &Subtarget,
                                    SelectionDAG &DAG) {
  assert(Subtarget.hasAVX512() &&
         "Need AVX512 for custom VECTOR_COMPRESS lowering.");
 
  SDLoc DL(Op);
  SDValue Vec = Op.getOperand(0);
  SDValue Mask = Op.getOperand(1);
  SDValue Passthru = Op.getOperand(2);
 
  EVT VecVT = Vec.getValueType();
  EVT ElementVT = VecVT.getVectorElementType();
  unsigned NumElements = VecVT.getVectorNumElements();
  unsigned NumVecBits = VecVT.getFixedSizeInBits();
  unsigned NumElementBits = ElementVT.getFixedSizeInBits();
 
  // 128- and 256-bit vectors with <= 16 elements can be converted to and
  // compressed as 512-bit vectors in AVX512F.
  if (NumVecBits != 128 && NumVecBits != 256)
    return SDValue();
 
  if (NumElementBits == 32 || NumElementBits == 64) {
    unsigned NumLargeElements = 512 / NumElementBits;
    MVT LargeVecVT =
        MVT::getVectorVT(ElementVT.getSimpleVT(), NumLargeElements);
    MVT LargeMaskVT = MVT::getVectorVT(MVT::i1, NumLargeElements);
 
    Vec = widenSubVector(LargeVecVT, Vec, /*ZeroNewElements=*/false, Subtarget,
                         DAG, DL);
    Mask = widenSubVector(LargeMaskVT, Mask, /*ZeroNewElements=*/true,
                          Subtarget, DAG, DL);
    Passthru = Passthru.isUndef() ? DAG.getUNDEF(LargeVecVT)
                                  : widenSubVector(LargeVecVT, Passthru,
                                                   /*ZeroNewElements=*/false,
                                                   Subtarget, DAG, DL);
 
    SDValue Compressed =
        DAG.getNode(ISD::VECTOR_COMPRESS, DL, LargeVecVT, Vec, Mask, Passthru);
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VecVT, Compressed,
                       DAG.getConstant(0, DL, MVT::i64));
  }
 
  if (VecVT == MVT::v8i16 || VecVT == MVT::v8i8 || VecVT == MVT::v16i8 ||
      VecVT == MVT::v16i16) {
    MVT LageElementVT = MVT::getIntegerVT(512 / NumElements);
    EVT LargeVecVT = MVT::getVectorVT(LageElementVT, NumElements);
 
    Vec = DAG.getNode(ISD::ANY_EXTEND, DL, LargeVecVT, Vec);
    Passthru = Passthru.isUndef()
                   ? DAG.getUNDEF(LargeVecVT)
                   : DAG.getNode(ISD::ANY_EXTEND, DL, LargeVecVT, Passthru);
 
    SDValue Compressed =
        DAG.getNode(ISD::VECTOR_COMPRESS, DL, LargeVecVT, Vec, Mask, Passthru);
    return DAG.getNode(ISD::TRUNCATE, DL, VecVT, Compressed);
  }
 
  return SDValue();
}
 
/// Try to lower a VSELECT instruction to a vector shuffle.
static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
                                           const X86Subtarget &Subtarget,
                                           SelectionDAG &DAG) {
  SDValue Cond = Op.getOperand(0);
  SDValue LHS = Op.getOperand(1);
  SDValue RHS = Op.getOperand(2);
  MVT VT = Op.getSimpleValueType();
 
  // Only non-legal VSELECTs reach this lowering, convert those into generic
  // shuffles and re-use the shuffle lowering path for blends.
  if (ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
    SmallVector<int, 32> Mask;
    if (createShuffleMaskFromVSELECT(Mask, Cond))
      return DAG.getVectorShuffle(VT, SDLoc(Op), LHS, RHS, Mask);
  }
 
  return SDValue();
}
 
SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
  SDValue Cond = Op.getOperand(0);
  SDValue LHS = Op.getOperand(1);
  SDValue RHS = Op.getOperand(2);
 
  SDLoc dl(Op);
  MVT VT = Op.getSimpleValueType();
  if (isSoftF16(VT, Subtarget)) {
    MVT NVT = VT.changeVectorElementTypeToInteger();
    return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, dl, NVT, Cond,
                                          DAG.getBitcast(NVT, LHS),
                                          DAG.getBitcast(NVT, RHS)));
  }
 
  // A vselect where all conditions and data are constants can be optimized into
  // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
  if (ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()) &&
      ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
      ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
    return SDValue();
 
  // Try to lower this to a blend-style vector shuffle. This can handle all
  // constant condition cases.
  if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
    return BlendOp;
 
  // If this VSELECT has a vector if i1 as a mask, it will be directly matched
  // with patterns on the mask registers on AVX-512.
  MVT CondVT = Cond.getSimpleValueType();
  unsigned CondEltSize = Cond.getScalarValueSizeInBits();
  if (CondEltSize == 1)
    return Op;
 
  // Variable blends are only legal from SSE4.1 onward.
  if (!Subtarget.hasSSE41())
    return SDValue();
 
  unsigned EltSize = VT.getScalarSizeInBits();
  unsigned NumElts = VT.getVectorNumElements();
 
  // Expand v32i16/v64i8 without BWI.
  if ((VT == MVT::v32i16 || VT == MVT::v64i8) && !Subtarget.hasBWI())
    return SDValue();
 
  // If the VSELECT is on a 512-bit type, we have to convert a non-i1 condition
  // into an i1 condition so that we can use the mask-based 512-bit blend
  // instructions.
  if (VT.getSizeInBits() == 512) {
    // Build a mask by testing the condition against zero.
    MVT MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
    SDValue Mask = DAG.getSetCC(dl, MaskVT, Cond,
                                DAG.getConstant(0, dl, CondVT),
                                ISD::SETNE);
    // Now return a new VSELECT using the mask.
    return DAG.getSelect(dl, VT, Mask, LHS, RHS);
  }
 
  // SEXT/TRUNC cases where the mask doesn't match the destination size.
  if (CondEltSize != EltSize) {
    // If we don't have a sign splat, rely on the expansion.
    if (CondEltSize != DAG.ComputeNumSignBits(Cond))
      return SDValue();
 
    MVT NewCondSVT = MVT::getIntegerVT(EltSize);
    MVT NewCondVT = MVT::getVectorVT(NewCondSVT, NumElts);
    Cond = DAG.getSExtOrTrunc(Cond, dl, NewCondVT);
    return DAG.getNode(ISD::VSELECT, dl, VT, Cond, LHS, RHS);
  }
 
  // v16i16/v32i8 selects without AVX2, if the condition and another operand
  // are free to split, then better to split before expanding the
  // select. Don't bother with XOP as it has the fast VPCMOV instruction.
  // TODO: This is very similar to narrowVectorSelect.
  // TODO: Add Load splitting to isFreeToSplitVector ?
  if (EltSize < 32 && VT.is256BitVector() && !Subtarget.hasAVX2() &&
      !Subtarget.hasXOP()) {
    bool FreeCond = isFreeToSplitVector(Cond, DAG);
    bool FreeLHS = isFreeToSplitVector(LHS, DAG) ||
                   (ISD::isNormalLoad(LHS.getNode()) && LHS.hasOneUse());
    bool FreeRHS = isFreeToSplitVector(RHS, DAG) ||
                   (ISD::isNormalLoad(RHS.getNode()) && RHS.hasOneUse());
    if (FreeCond && (FreeLHS || FreeRHS))
      return splitVectorOp(Op, DAG, dl);
  }
 
  // Only some types will be legal on some subtargets. If we can emit a legal
  // VSELECT-matching blend, return Op, and but if we need to expand, return
  // a null value.
  switch (VT.SimpleTy) {
  default:
    // Most of the vector types have blends past SSE4.1.
    return Op;
 
  case MVT::v32i8:
    // The byte blends for AVX vectors were introduced only in AVX2.
    if (Subtarget.hasAVX2())
      return Op;
 
    return SDValue();
 
  case MVT::v8i16:
  case MVT::v16i16:
  case MVT::v8f16:
  case MVT::v16f16: {
    // Bitcast everything to the vXi8 type and use a vXi8 vselect.
    MVT CastVT = MVT::getVectorVT(MVT::i8, NumElts * 2);
    Cond = DAG.getBitcast(CastVT, Cond);
    LHS = DAG.getBitcast(CastVT, LHS);
    RHS = DAG.getBitcast(CastVT, RHS);
    SDValue Select = DAG.getNode(ISD::VSELECT, dl, CastVT, Cond, LHS, RHS);
    return DAG.getBitcast(VT, Select);
  }
  }
}
 
static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
  MVT VT = Op.getSimpleValueType();
  SDValue Vec = Op.getOperand(0);
  SDValue Idx = Op.getOperand(1);
  assert(isa<ConstantSDNode>(Idx) && "Constant index expected");
  SDLoc dl(Op);
 
  if (!Vec.getSimpleValueType().is128BitVector())
    return SDValue();
 
  if (VT.getSizeInBits() == 8) {
    // If IdxVal is 0, it's cheaper to do a move instead of a pextrb, unless
    // we're going to zero extend the register or fold the store.
    if (llvm::isNullConstant(Idx) && !X86::mayFoldIntoZeroExtend(Op) &&
        !X86::mayFoldIntoStore(Op))
      return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8,
                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
                                     DAG.getBitcast(MVT::v4i32, Vec), Idx));
 
    unsigned IdxVal = Idx->getAsZExtVal();
    SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32, Vec,
                                  DAG.getTargetConstant(IdxVal, dl, MVT::i8));
    return DAG.getNode(ISD::TRUNCATE, dl, VT, Extract);
  }
 
  if (VT == MVT::f32) {
    // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
    // the result back to FR32 register. It's only worth matching if the
    // result has a single use which is a store or a bitcast to i32.  And in
    // the case of a store, it's not worth it if the index is a constant 0,
    // because a MOVSSmr can be used instead, which is smaller and faster.
    if (!Op.hasOneUse())
      return SDValue();
    SDNode *User = *Op.getNode()->user_begin();
    if ((User->getOpcode() != ISD::STORE || isNullConstant(Idx)) &&
        (User->getOpcode() != ISD::BITCAST ||
         User->getValueType(0) != MVT::i32))
      return SDValue();
    SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
                                  DAG.getBitcast(MVT::v4i32, Vec), Idx);
    return DAG.getBitcast(MVT::f32, Extract);
  }
 
  if (VT == MVT::i32 || VT == MVT::i64)
      return Op;
 
  return SDValue();
}
 
/// Extract one bit from mask vector, like v16i1 or v8i1.
/// AVX-512 feature.
static SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG,
                                        const X86Subtarget &Subtarget) {
  SDValue Vec = Op.getOperand(0);
  SDLoc dl(Vec);
  MVT VecVT = Vec.getSimpleValueType();
  SDValue Idx = Op.getOperand(1);
  auto* IdxC = dyn_cast<ConstantSDNode>(Idx);
  MVT EltVT = Op.getSimpleValueType();
 
  assert((VecVT.getVectorNumElements() <= 16 || Subtarget.hasBWI()) &&
         "Unexpected vector type in ExtractBitFromMaskVector");
 
  // variable index can't be handled in mask registers,
  // extend vector to VR512/128
  if (!IdxC) {
    unsigned NumElts = VecVT.getVectorNumElements();
    // Extending v8i1/v16i1 to 512-bit get better performance on KNL
    // than extending to 128/256bit.
    if (NumElts == 1) {
      Vec = widenMaskVector(Vec, false, Subtarget, DAG, dl);
      MVT IntVT = MVT::getIntegerVT(Vec.getValueType().getVectorNumElements());
      return DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, DAG.getBitcast(IntVT, Vec));
    }
    MVT ExtEltVT = (NumElts <= 8) ? MVT::getIntegerVT(128 / NumElts) : MVT::i8;
    MVT ExtVecVT = MVT::getVectorVT(ExtEltVT, NumElts);
    SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, dl, ExtVecVT, Vec);
    SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ExtEltVT, Ext, Idx);
    return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
  }
 
  unsigned IdxVal = IdxC->getZExtValue();
  if (IdxVal == 0) // the operation is legal
    return Op;
 
  // Extend to natively supported kshift.
  Vec = widenMaskVector(Vec, false, Subtarget, DAG, dl);
 
  // Use kshiftr instruction to move to the lower element.
  Vec = DAG.getNode(X86ISD::KSHIFTR, dl, Vec.getSimpleValueType(), Vec,
                    DAG.getTargetConstant(IdxVal, dl, MVT::i8));
 
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
                     DAG.getVectorIdxConstant(0, dl));
}
 
// Helper to find all the extracted elements from a vector.
static APInt getExtractedDemandedElts(SDNode *N) {
  MVT VT = N->getSimpleValueType(0);
  unsigned NumElts = VT.getVectorNumElements();
  APInt DemandedElts = APInt::getZero(NumElts);
  for (SDNode *User : N->users()) {
    switch (User->getOpcode()) {
    case X86ISD::PEXTRB:
    case X86ISD::PEXTRW:
    case ISD::EXTRACT_VECTOR_ELT:
      if (!isa<ConstantSDNode>(User->getOperand(1))) {
        DemandedElts.setAllBits();
        return DemandedElts;
      }
      DemandedElts.setBit(User->getConstantOperandVal(1));
      break;
    case ISD::BITCAST: {
      if (!User->getValueType(0).isSimple() ||
          !User->getValueType(0).isVector()) {
        DemandedElts.setAllBits();
        return DemandedElts;
      }
      APInt DemandedSrcElts = getExtractedDemandedElts(User);
      DemandedElts |= APIntOps::ScaleBitMask(DemandedSrcElts, NumElts);
      break;
    }
    default:
      DemandedElts.setAllBits();
      return DemandedElts;
    }
  }
  return DemandedElts;
}
 
SDValue
X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
                                           SelectionDAG &DAG) const {
  SDLoc dl(Op);
  SDValue Vec = Op.getOperand(0);
  MVT VecVT = Vec.getSimpleValueType();
  SDValue Idx = Op.getOperand(1);
  auto* IdxC = dyn_cast<ConstantSDNode>(Idx);
 
  if (VecVT.getVectorElementType() == MVT::i1)
    return ExtractBitFromMaskVector(Op, DAG, Subtarget);
 
  if (!IdxC) {
    // Its more profitable to go through memory (1 cycles throughput)
    // than using VMOVD + VPERMV/PSHUFB sequence (2/3 cycles throughput)
    // IACA tool was used to get performance estimation
    // (https://software.intel.com/en-us/articles/intel-architecture-code-analyzer)
    //
    // example : extractelement <16 x i8> %a, i32 %i
    //
    // Block Throughput: 3.00 Cycles
    // Throughput Bottleneck: Port5
    //
    // | Num Of |   Ports pressure in cycles  |    |
    // |  Uops  |  0  - DV  |  5  |  6  |  7  |    |
    // ---------------------------------------------
    // |   1    |           | 1.0 |     |     | CP | vmovd xmm1, edi
    // |   1    |           | 1.0 |     |     | CP | vpshufb xmm0, xmm0, xmm1
    // |   2    | 1.0       | 1.0 |     |     | CP | vpextrb eax, xmm0, 0x0
    // Total Num Of Uops: 4
    //
    //
    // Block Throughput: 1.00 Cycles
    // Throughput Bottleneck: PORT2_AGU, PORT3_AGU, Port4
    //
    // |    |  Ports pressure in cycles   |  |
    // |Uops| 1 | 2 - D  |3 -  D  | 4 | 5 |  |
    // ---------------------------------------------------------
    // |2^  |   | 0.5    | 0.5    |1.0|   |CP| vmovaps xmmword ptr [rsp-0x18], xmm0
    // |1   |0.5|        |        |   |0.5|  | lea rax, ptr [rsp-0x18]
    // |1   |   |0.5, 0.5|0.5, 0.5|   |   |CP| mov al, byte ptr [rdi+rax*1]
    // Total Num Of Uops: 4
 
    return SDValue();
  }
 
  unsigned IdxVal = IdxC->getZExtValue();
 
  // If this is a 256-bit vector result, first extract the 128-bit vector and
  // then extract the element from the 128-bit vector.
  if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
    // Get the 128-bit vector.
    Vec = extract128BitVector(Vec, IdxVal, DAG, dl);
    MVT EltVT = VecVT.getVectorElementType();
 
    unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
    assert(isPowerOf2_32(ElemsPerChunk) && "Elements per chunk not power of 2");
 
    // Find IdxVal modulo ElemsPerChunk. Since ElemsPerChunk is a power of 2
    // this can be done with a mask.
    IdxVal &= ElemsPerChunk - 1;
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
                       DAG.getVectorIdxConstant(IdxVal, dl));
  }
 
  assert(VecVT.is128BitVector() && "Unexpected vector length");
 
  MVT VT = Op.getSimpleValueType();
 
  if (VT == MVT::i16) {
    // If IdxVal is 0, it's cheaper to do a move instead of a pextrw, unless
    // we're going to zero extend the register or fold the store (SSE41 only).
    if (IdxVal == 0 && !X86::mayFoldIntoZeroExtend(Op) &&
        !(Subtarget.hasSSE41() && X86::mayFoldIntoStore(Op))) {
      if (Subtarget.hasFP16())
        return Op;
 
      return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
                                     DAG.getBitcast(MVT::v4i32, Vec), Idx));
    }
 
    SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32, Vec,
                                  DAG.getTargetConstant(IdxVal, dl, MVT::i8));
    return DAG.getNode(ISD::TRUNCATE, dl, VT, Extract);
  }
 
  if (Subtarget.hasSSE41())
    if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
      return Res;
 
  // Only extract a single element from a v16i8 source - determine the common
  // DWORD/WORD that all extractions share, and extract the sub-byte.
  // TODO: Add QWORD MOVQ extraction?
  if (VT == MVT::i8) {
    APInt DemandedElts = getExtractedDemandedElts(Vec.getNode());
    assert(DemandedElts.getBitWidth() == 16 && "Vector width mismatch");
 
    // Extract either the lowest i32 or any i16, and extract the sub-byte.
    int DWordIdx = IdxVal / 4;
    if (DWordIdx == 0 && DemandedElts == (DemandedElts & 15)) {
      SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
                                DAG.getBitcast(MVT::v4i32, Vec),
                                DAG.getVectorIdxConstant(DWordIdx, dl));
      int ShiftVal = (IdxVal % 4) * 8;
      if (ShiftVal != 0)
        Res = DAG.getNode(ISD::SRL, dl, MVT::i32, Res,
                          DAG.getConstant(ShiftVal, dl, MVT::i8));
      return DAG.getNode(ISD::TRUNCATE, dl, VT, Res);
    }
 
    int WordIdx = IdxVal / 2;
    if (DemandedElts == (DemandedElts & (3 << (WordIdx * 2)))) {
      SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
                                DAG.getBitcast(MVT::v8i16, Vec),
                                DAG.getVectorIdxConstant(WordIdx, dl));
      int ShiftVal = (IdxVal % 2) * 8;
      if (ShiftVal != 0)
        Res = DAG.getNode(ISD::SRL, dl, MVT::i16, Res,
                          DAG.getConstant(ShiftVal, dl, MVT::i8));
      return DAG.getNode(ISD::TRUNCATE, dl, VT, Res);
    }
  }
 
  if (VT == MVT::f16 || VT.getSizeInBits() == 32) {
    if (IdxVal == 0)
      return Op;
 
    // Shuffle the element to the lowest element, then movss or movsh.
    SmallVector<int, 8> Mask(VecVT.getVectorNumElements(), -1);
    Mask[0] = static_cast<int>(IdxVal);
    Vec = DAG.getVectorShuffle(VecVT, dl, Vec, DAG.getUNDEF(VecVT), Mask);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
                       DAG.getVectorIdxConstant(0, dl));
  }
 
  if (VT.getSizeInBits() == 64) {
    // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
    // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
    //        to match extract_elt for f64.
    if (IdxVal == 0)
      return Op;
 
    // UNPCKHPD the element to the lowest double word, then movsd.
    // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
    // to a f64mem, the whole operation is folded into a single MOVHPDmr.
    int Mask[2] = { 1, -1 };
    Vec = DAG.getVectorShuffle(VecVT, dl, Vec, DAG.getUNDEF(VecVT), Mask);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
                       DAG.getVectorIdxConstant(0, dl));
  }
 
  return SDValue();
}
 
/// Insert one bit to mask vector, like v16i1 or v8i1.
/// AVX-512 feature.
static SDValue InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
  SDLoc dl(Op);
  SDValue Vec = Op.getOperand(0);
  SDValue Elt = Op.getOperand(1);
  SDValue Idx = Op.getOperand(2);
  MVT VecVT = Vec.getSimpleValueType();
 
  if (!isa<ConstantSDNode>(Idx)) {
    // Non constant index. Extend source and destination,
    // insert element and then truncate the result.
    unsigned NumElts = VecVT.getVectorNumElements();
    MVT ExtEltVT = (NumElts <= 8) ? MVT::getIntegerVT(128 / NumElts) : MVT::i8;
    MVT ExtVecVT = MVT::getVectorVT(ExtEltVT, NumElts);
    SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
      DAG.getNode(ISD::SIGN_EXTEND, dl, ExtVecVT, Vec),
      DAG.getNode(ISD::SIGN_EXTEND, dl, ExtEltVT, Elt), Idx);
    return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
  }
 
  // Copy into a k-register, extract to v1i1 and insert_subvector.
  SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i1, Elt);
  return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, VecVT, Vec, EltInVec, Idx);
}
 
SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
                                                  SelectionDAG &DAG) const {
  MVT VT = Op.getSimpleValueType();
  MVT EltVT = VT.getVectorElementType();
  unsigned NumElts = VT.getVectorNumElements();
  unsigned EltSizeInBits = EltVT.getScalarSizeInBits();
 
  if (EltVT == MVT::i1)
    return InsertBitToMaskVector(Op, DAG, Subtarget);
 
  SDLoc dl(Op);
  SDValue N0 = Op.getOperand(0);
  SDValue N1 = Op.getOperand(1);
  SDValue N2 = Op.getOperand(2);
  auto *N2C = dyn_cast<ConstantSDNode>(N2);
 
  if (EltVT == MVT::bf16) {
    MVT IVT = VT.changeVectorElementTypeToInteger();
    SDValue Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, IVT,
                              DAG.getBitcast(IVT, N0),
                              DAG.getBitcast(MVT::i16, N1), N2);
    return DAG.getBitcast(VT, Res);
  }
 
  if (!N2C) {
    // Variable insertion indices, usually we're better off spilling to stack,
    // but AVX512 can use a variable compare+select by comparing against all
    // possible vector indices, and FP insertion has less gpr->simd traffic.
    if (!(Subtarget.hasBWI() ||
          (Subtarget.hasAVX512() && EltSizeInBits >= 32) ||
          (Subtarget.hasSSE41() && (EltVT == MVT::f32 || EltVT == MVT::f64))))
      return SDValue();
 
    MVT IdxSVT = MVT::getIntegerVT(EltSizeInBits);
    MVT IdxVT = MVT::getVectorVT(IdxSVT, NumElts);
    if (!isTypeLegal(IdxSVT) || !isTypeLegal(IdxVT))
      return SDValue();
 
    SDValue IdxExt = DAG.getZExtOrTrunc(N2, dl, IdxSVT);
    SDValue IdxSplat = DAG.getSplatBuildVector(IdxVT, dl, IdxExt);
    SDValue EltSplat = DAG.getSplatBuildVector(VT, dl, N1);
 
    SmallVector<SDValue, 16> RawIndices;
    for (unsigned I = 0; I != NumElts; ++I)
      RawIndices.push_back(DAG.getConstant(I, dl, IdxSVT));
    SDValue Indices = DAG.getBuildVector(IdxVT, dl, RawIndices);
 
    // inselt N0, N1, N2 --> select (SplatN2 == {0,1,2...}) ? SplatN1 : N0.
    return DAG.getSelectCC(dl, IdxSplat, Indices, EltSplat, N0,
                           ISD::CondCode::SETEQ);
  }
 
  if (N2C->getAPIntValue().uge(NumElts))
    return SDValue();
  uint64_t IdxVal = N2C->getZExtValue();
 
  bool IsZeroElt = X86::isZeroNode(N1);
  bool IsAllOnesElt = VT.isInteger() && llvm::isAllOnesConstant(N1);
 
  if (IsZeroElt || IsAllOnesElt) {
    // Lower insertion of v16i8/v32i8/v64i16 -1 elts as an 'OR' blend.
    // We don't deal with i8 0 since it appears to be handled elsewhere.
    if (IsAllOnesElt &&
        ((VT == MVT::v16i8 && !Subtarget.hasSSE41()) ||
         ((VT == MVT::v32i8 || VT == MVT::v16i16) && !Subtarget.hasInt256()))) {
      SDValue ZeroCst = DAG.getConstant(0, dl, VT.getScalarType());
      SDValue OnesCst = DAG.getAllOnesConstant(dl, VT.getScalarType());
      SmallVector<SDValue, 8> CstVectorElts(NumElts, ZeroCst);
      CstVectorElts[IdxVal] = OnesCst;
      SDValue CstVector = DAG.getBuildVector(VT, dl, CstVectorElts);
      return DAG.getNode(ISD::OR, dl, VT, N0, CstVector);
    }
    // See if we can do this more efficiently with a blend shuffle with a
    // rematerializable vector.
    if (Subtarget.hasSSE41() &&
        (EltSizeInBits >= 16 || (IsZeroElt && !VT.is128BitVector()))) {
      SmallVector<int, 8> BlendMask;
      for (unsigned i = 0; i != NumElts; ++i)
        BlendMask.push_back(i == IdxVal ? i + NumElts : i);
      SDValue CstVector = IsZeroElt ? getZeroVector(VT, Subtarget, DAG, dl)
                                    : getOnesVector(VT, DAG, dl);
      return DAG.getVectorShuffle(VT, dl, N0, CstVector, BlendMask);
    }
  }
 
  // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
  // into that, and then insert the subvector back into the result.
  if (VT.is256BitVector() || VT.is512BitVector()) {
    // With a 256-bit vector, we can insert into the zero element efficiently
    // using a blend if we have AVX or AVX2 and the right data type.
    if (VT.is256BitVector() && IdxVal == 0) {
      // TODO: It is worthwhile to cast integer to floating point and back
      // and incur a domain crossing penalty if that's what we'll end up
      // doing anyway after extracting to a 128-bit vector.
      if ((Subtarget.hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
          (Subtarget.hasAVX2() && (EltVT == MVT::i32 || EltVT == MVT::i64))) {
        SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
        return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec,
                           DAG.getTargetConstant(1, dl, MVT::i8));
      }
    }
 
    unsigned NumEltsIn128 = 128 / EltSizeInBits;
    assert(isPowerOf2_32(NumEltsIn128) &&
           "Vectors will always have power-of-two number of elements.");
 
    // If we are not inserting into the low 128-bit vector chunk,
    // then prefer the broadcast+blend sequence.
    // FIXME: relax the profitability check iff all N1 uses are insertions.
    if (IdxVal >= NumEltsIn128 &&
        ((Subtarget.hasAVX2() && EltSizeInBits != 8) ||
         (Subtarget.hasAVX() && (EltSizeInBits >= 32) &&
          X86::mayFoldLoad(N1, Subtarget)))) {
      SDValue N1SplatVec = DAG.getSplatBuildVector(VT, dl, N1);
      SmallVector<int, 8> BlendMask;
      for (unsigned i = 0; i != NumElts; ++i)
        BlendMask.push_back(i == IdxVal ? i + NumElts : i);
      return DAG.getVectorShuffle(VT, dl, N0, N1SplatVec, BlendMask);
    }
 
    // Get the desired 128-bit vector chunk.
    SDValue V = extract128BitVector(N0, IdxVal, DAG, dl);
 
    // Insert the element into the desired chunk.
    // Since NumEltsIn128 is a power of 2 we can use mask instead of modulo.
    unsigned IdxIn128 = IdxVal & (NumEltsIn128 - 1);
 
    V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
                    DAG.getVectorIdxConstant(IdxIn128, dl));
 
    // Insert the changed part back into the bigger vector
    return insert128BitVector(N0, V, IdxVal, DAG, dl);
  }
  assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
 
  // This will be just movw/movd/movq/movsh/movss/movsd.
  if (IdxVal == 0 && ISD::isBuildVectorAllZeros(N0.getNode())) {
    if (EltVT == MVT::i32 || EltVT == MVT::f32 || EltVT == MVT::f64 ||
        EltVT == MVT::f16 || EltVT == MVT::i64) {
      N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
      return getShuffleVectorZeroOrUndef(N1, 0, true, Subtarget, DAG);
    }
 
    // We can't directly insert an i8 or i16 into a vector, so zero extend
    // it to i32 first.
    if (EltVT == MVT::i16 || EltVT == MVT::i8) {
      N1 = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, N1);
      MVT ShufVT = MVT::getVectorVT(MVT::i32, VT.getSizeInBits() / 32);
      N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShufVT, N1);
      N1 = getShuffleVectorZeroOrUndef(N1, 0, true, Subtarget, DAG);
      return DAG.getBitcast(VT, N1);
    }
  }
 
  // Transform it so it match pinsr{b,w} which expects a GR32 as its second
  // argument. SSE41 required for pinsrb.
  if (VT == MVT::v8i16 || (VT == MVT::v16i8 && Subtarget.hasSSE41())) {
    unsigned Opc;
    if (VT == MVT::v8i16) {
      assert(Subtarget.hasSSE2() && "SSE2 required for PINSRW");
      Opc = X86ISD::PINSRW;
    } else {
      assert(VT == MVT::v16i8 && "PINSRB requires v16i8 vector");
      assert(Subtarget.hasSSE41() && "SSE41 required for PINSRB");
      Opc = X86ISD::PINSRB;
    }
 
    assert(N1.getValueType() != MVT::i32 && "Unexpected VT");
    N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
    N2 = DAG.getTargetConstant(IdxVal, dl, MVT::i8);
    return DAG.getNode(Opc, dl, VT, N0, N1, N2);
  }
 
  if (Subtarget.hasSSE41()) {
    if (EltVT == MVT::f32) {
      // Bits [7:6] of the constant are the source select. This will always be
      //   zero here. The DAG Combiner may combine an extract_elt index into
      //   these bits. For example (insert (extract, 3), 2) could be matched by
      //   putting the '3' into bits [7:6] of X86ISD::INSERTPS.
      // Bits [5:4] of the constant are the destination select. This is the
      //   value of the incoming immediate.
      // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
      //   combine either bitwise AND or insert of float 0.0 to set these bits.
 
      bool MinSize = DAG.getMachineFunction().getFunction().hasMinSize();
      if (IdxVal == 0 && (!MinSize || !X86::mayFoldLoad(N1, Subtarget))) {
        // If this is an insertion of 32-bits into the low 32-bits of
        // a vector, we prefer to generate a blend with immediate rather
        // than an insertps. Blends are simpler operations in hardware and so
        // will always have equal or better performance than insertps.
        // But if optimizing for size and there's a load folding opportunity,
        // generate insertps because blendps does not have a 32-bit memory
        // operand form.
        N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
        return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1,
                           DAG.getTargetConstant(1, dl, MVT::i8));
      }
      // Create this as a scalar to vector..
      N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
      return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1,
                         DAG.getTargetConstant(IdxVal << 4, dl, MVT::i8));
    }
 
    // PINSR* works with constant index.
    if (EltVT == MVT::i32 || EltVT == MVT::i64)
      return Op;
  }
 
  return SDValue();
}
 
static SDValue LowerFLDEXP(SDValue Op, const X86Subtarget &Subtarget,
                           SelectionDAG &DAG) {
  SDLoc DL(Op);
  SDValue X = Op.getOperand(0);
  MVT XTy = X.getSimpleValueType();
  SDValue Exp = Op.getOperand(1);
 
  switch (XTy.SimpleTy) {
  default:
    return SDValue();
  case MVT::f16:
    if (!Subtarget.hasFP16())
      X = DAG.getFPExtendOrRound(X, DL, MVT::f32);
    [[fallthrough]];
  case MVT::f32:
  case MVT::f64: {
    MVT VT = MVT::getVectorVT(X.getSimpleValueType(),
                              128 / X.getSimpleValueType().getSizeInBits());
    Exp = DAG.getNode(ISD::SINT_TO_FP, DL, X.getValueType(), Exp);
    SDValue VX = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, X);
    SDValue VExp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VT, Exp);
    SDValue Scalefs = DAG.getNode(X86ISD::SCALEFS, DL, VT, VX, VExp);
    SDValue Final = DAG.getExtractVectorElt(DL, X.getValueType(), Scalefs, 0);
    return DAG.getFPExtendOrRound(Final, DL, XTy);
  }
  case MVT::v4f32:
  case MVT::v2f64:
  case MVT::v8f32:
  case MVT::v4f64:
  case MVT::v16f32:
  case MVT::v8f64:
    if (XTy.getSizeInBits() == 512 || Subtarget.hasVLX()) {
      Exp = DAG.getNode(ISD::SINT_TO_FP, DL, XTy, Exp);
      return DAG.getNode(X86ISD::SCALEF, DL, XTy, X, Exp);
    }
    break;
  case MVT::v8f16:
  case MVT::v16f16:
    if (Subtarget.hasFP16()) {
      if (Subtarget.hasVLX()) {
        Exp = DAG.getNode(ISD::SINT_TO_FP, DL, XTy, Exp);
        return DAG.getNode(X86ISD::SCALEF, DL, XTy, X, Exp);
      }
      break;
    }
    X = DAG.getFPExtendOrRound(X, DL, XTy.changeVectorElementType(MVT::f32));
    Exp = DAG.getSExtOrTrunc(Exp, DL,
                             X.getSimpleValueType().changeTypeToInteger());
    break;
  case MVT::v32f16:
    if (Subtarget.hasFP16()) {
      Exp = DAG.getNode(ISD::SINT_TO_FP, DL, XTy, Exp);
      return DAG.getNode(X86ISD::SCALEF, DL, XTy, X, Exp);
    }
    return splitVectorOp(Op, DAG, DL);
  }
  SDValue WideX = widenSubVector(X, true, Subtarget, DAG, DL, 512);
  SDValue WideExp = widenSubVector(Exp, true, Subtarget, DAG, DL, 512);
  Exp = DAG.getNode(ISD::SINT_TO_FP, DL, WideExp.getSimpleValueType(), Exp);
  SDValue Scalef =
      DAG.getNode(X86ISD::SCALEF, DL, WideX.getValueType(), WideX, WideExp);
  SDValue Final =
      DAG.getExtractSubvector(DL, X.getSimpleValueType(), Scalef, 0);
  return DAG.getFPExtendOrRound(Final, DL, XTy);
}
 
static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, const X86Subtarget &Subtarget,
                                     SelectionDAG &DAG) {
  SDLoc dl(Op);
  MVT OpVT = Op.getSimpleValueType();
 
  // It's always cheaper to replace a xor+movd with xorps and simplifies further
  // combines.
  if (X86::isZeroNode(Op.getOperand(0)))
    return getZeroVector(OpVT, Subtarget, DAG, dl);
 
  // If this is a 256-bit vector result, first insert into a 128-bit
  // vector and then insert into the 256-bit vector.
  if (!OpVT.is128BitVector()) {
    // Insert into a 128-bit vector.
    unsigned SizeFactor = OpVT.getSizeInBits() / 128;
    MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
                                 OpVT.getVectorNumElements() / SizeFactor);
 
    Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
 
    // Insert the 128-bit vector.
    return insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
  }
  assert(OpVT.is128BitVector() && OpVT.isInteger() && OpVT != MVT::v2i64 &&
         "Expected an SSE type!");
 
  // Pass through a v4i32 or V8i16 SCALAR_TO_VECTOR as that's what we use in
  // tblgen.
  if (OpVT == MVT::v4i32 || (OpVT == MVT::v8i16 && Subtarget.hasFP16()))
    return Op;
 
  SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
  return DAG.getBitcast(
      OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
}
 
// Lower a node with an INSERT_SUBVECTOR opcode.  This may result in a
// simple superregister reference or explicit instructions to insert
// the upper bits of a vector.
static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget,
                                     SelectionDAG &DAG) {
  assert(Op.getSimpleValueType().getVectorElementType() == MVT::i1);
 
  return insert1BitVector(Op, DAG, Subtarget);
}
 
static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget &Subtarget,
                                      SelectionDAG &DAG) {
  assert(Op.getSimpleValueType().getVectorElementType() == MVT::i1 &&
         "Only vXi1 extract_subvectors need custom lowering");
 
  SDLoc dl(Op);
  SDValue Vec = Op.getOperand(0);
  uint64_t IdxVal = Op.getConstantOperandVal(1);
 
  if (IdxVal == 0) // the operation is legal
    return Op;
 
  // Extend to natively supported kshift.
  Vec = widenMaskVector(Vec, false, Subtarget, DAG, dl);
 
  // Shift to the LSB.
  Vec = DAG.getNode(X86ISD::KSHIFTR, dl, Vec.getSimpleValueType(), Vec,
                    DAG.getTargetConstant(IdxVal, dl, MVT::i8));
 
  return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, Op.getValueType(), Vec,
                     DAG.getVectorIdxConstant(0, dl));
}
 
// Returns the appropriate wrapper opcode for a global reference.
unsigned X86TargetLowering::getGlobalWrapperKind(
    const GlobalValue *GV, const unsigned char OpFlags) const {
  // References to absolute symbols are never PC-relative.
  if (GV && GV->isAbsoluteSymbolRef())
    return X86ISD::Wrapper;
 
  // The following OpFlags under RIP-rel PIC use RIP.
  if (Subtarget.isPICStyleRIPRel() &&
      (OpFlags == X86II::MO_NO_FLAG || OpFlags == X86II::MO_COFFSTUB ||
       OpFlags == X86II::MO_DLLIMPORT))
    return X86ISD::WrapperRIP;
 
  // GOTPCREL references must always use RIP.
  if (OpFlags == X86II::MO_GOTPCREL || OpFlags == X86II::MO_GOTPCREL_NORELAX)
    return X86ISD::WrapperRIP;
 
  return X86ISD::Wrapper;
}
 
// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
// their target counterpart wrapped in the X86ISD::Wrapper node. Suppose N is
// one of the above mentioned nodes. It has to be wrapped because otherwise
// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
// be used to form addressing mode. These wrapped nodes will be selected
// into MOV32ri.
SDValue
X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
 
  // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
  // global base reg.
  unsigned char OpFlag = Subtarget.classifyLocalReference(nullptr);
 
  auto PtrVT = getPointerTy(DAG.getDataLayout());
  SDValue Result = DAG.getTargetConstantPool(
      CP->getConstVal(), PtrVT, CP->getAlign(), CP->getOffset(), OpFlag);
  SDLoc DL(CP);
  Result =
      DAG.getNode(getGlobalWrapperKind(nullptr, OpFlag), DL, PtrVT, Result);
  // With PIC, the address is actually $g + Offset.
  if (OpFlag) {
    Result =
        DAG.getNode(ISD::ADD, DL, PtrVT,
                    DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
  }
 
  return Result;
}
 
SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
  JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
 
  // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
  // global base reg.
  unsigned char OpFlag = Subtarget.classifyLocalReference(nullptr);
 
  EVT PtrVT = Op.getValueType();
  SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
  SDLoc DL(JT);
  Result =
      DAG.getNode(getGlobalWrapperKind(nullptr, OpFlag), DL, PtrVT, Result);
 
  // With PIC, the address is actually $g + Offset.
  if (OpFlag)
    Result =
        DAG.getNode(ISD::ADD, DL, PtrVT,
                    DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
 
  return Result;
}
 
SDValue X86TargetLowering::LowerExternalSymbol(SDValue Op,
                                               SelectionDAG &DAG) const {
  return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false, nullptr);
}
 
SDValue
X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
  // Create the TargetBlockAddressAddress node.
  unsigned char OpFlags =
    Subtarget.classifyBlockAddressReference();
  const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
  int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
  SDLoc dl(Op);
  EVT PtrVT = Op.getValueType();
  SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
  Result =
      DAG.getNode(getGlobalWrapperKind(nullptr, OpFlags), dl, PtrVT, Result);
 
  // With PIC, the address is actually $g + Offset.
  if (isGlobalRelativeToPICBase(OpFlags)) {
    Result = DAG.getNode(ISD::ADD, dl, PtrVT,
                         DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
  }
 
  return Result;
}
 
/// Creates target global address or external symbol nodes for calls or
/// other uses.
SDValue X86TargetLowering::LowerGlobalOrExternal(SDValue Op, SelectionDAG &DAG,
                                                 bool ForCall,
                                                 bool *IsImpCall) const {
  // Unpack the global address or external symbol.
  SDLoc dl(Op);
  const GlobalValue *GV = nullptr;
  int64_t Offset = 0;
  const char *ExternalSym = nullptr;
  if (const auto *G = dyn_cast<GlobalAddressSDNode>(Op)) {
    GV = G->getGlobal();
    Offset = G->getOffset();
  } else {
    const auto *ES = cast<ExternalSymbolSDNode>(Op);
    ExternalSym = ES->getSymbol();
  }
 
  // Calculate some flags for address lowering.
  const Module &Mod = *DAG.getMachineFunction().getFunction().getParent();
  unsigned char OpFlags;
  if (ForCall)
    OpFlags = Subtarget.classifyGlobalFunctionReference(GV, Mod);
  else
    OpFlags = Subtarget.classifyGlobalReference(GV, Mod);
  bool HasPICReg = isGlobalRelativeToPICBase(OpFlags);
  bool NeedsLoad = isGlobalStubReference(OpFlags);
 
  CodeModel::Model M = DAG.getTarget().getCodeModel();
  EVT PtrVT = Op.getValueType();
  SDValue Result;
 
  if (GV) {
    // Create a target global address if this is a global. If possible, fold the
    // offset into the global address reference. Otherwise, ADD it on later.
    // Suppress the folding if Offset is negative: movl foo-1, %eax is not
    // allowed because if the address of foo is 0, the ELF R_X86_64_32
    // relocation will compute to a negative value, which is invalid.
    int64_t GlobalOffset = 0;
    if (OpFlags == X86II::MO_NO_FLAG && Offset >= 0 &&
        X86::isOffsetSuitableForCodeModel(Offset, M, true)) {
      std::swap(GlobalOffset, Offset);
    }
    Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, GlobalOffset, OpFlags);
  } else {
    // If this is not a global address, this must be an external symbol.
    Result = DAG.getTargetExternalSymbol(ExternalSym, PtrVT, OpFlags);
  }
 
  // If this is a direct call, avoid the wrapper if we don't need to do any
  // loads or adds. This allows SDAG ISel to match direct calls.
  if (ForCall && !NeedsLoad && !HasPICReg && Offset == 0)
    return Result;
 
  // If Import Call Optimization is enabled and this is an imported function
  // then make a note of it and return the global address without wrapping.
  if (IsImpCall && (OpFlags == X86II::MO_DLLIMPORT) &&
      Mod.getModuleFlag("import-call-optimization")) {
    assert(ForCall && "Should only enable import call optimization if we are "
                      "lowering a call");
    *IsImpCall = true;
    return Result;
  }
 
  Result = DAG.getNode(getGlobalWrapperKind(GV, OpFlags), dl, PtrVT, Result);
 
  // With PIC, the address is actually $g + Offset.
  if (HasPICReg) {
    Result = DAG.getNode(ISD::ADD, dl, PtrVT,
                         DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
  }
 
  // For globals that require a load from a stub to get the address, emit the
  // load.
  if (NeedsLoad)
    Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
                         MachinePointerInfo::getGOT(DAG.getMachineFunction()));
 
  // If there was a non-zero offset that we didn't fold, create an explicit
  // addition for it.
  if (Offset != 0)
    Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
                         DAG.getSignedConstant(Offset, dl, PtrVT));
 
  return Result;
}
 
SDValue
X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
  return LowerGlobalOrExternal(Op, DAG, /*ForCall=*/false, nullptr);
}
 
static SDValue GetTLSADDR(SelectionDAG &DAG, GlobalAddressSDNode *GA,
                          const EVT PtrVT, unsigned ReturnReg,
                          unsigned char OperandFlags,
                          bool LoadGlobalBaseReg = false,
                          bool LocalDynamic = false) {
  MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
  SDLoc dl(GA);
  SDValue TGA;
  bool UseTLSDESC = DAG.getTarget().useTLSDESC();
  SDValue Chain = DAG.getEntryNode();
  SDValue Ret;
  if (LocalDynamic && UseTLSDESC) {
    TGA = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT, OperandFlags);
    // Reuse existing GetTLSADDR node if we can find it.
    if (TGA->hasOneUse()) {
      // TLSDESC uses TGA.
      SDNode *TLSDescOp = *TGA->user_begin();
      assert(TLSDescOp->getOpcode() == X86ISD::TLSDESC &&
             "Unexpected TLSDESC DAG");
      // CALLSEQ_END uses TGA via a chain and glue.
      auto *CallSeqEndOp = TLSDescOp->getGluedUser();
      assert(CallSeqEndOp && CallSeqEndOp->getOpcode() == ISD::CALLSEQ_END &&
             "Unexpected TLSDESC DAG");
      // CopyFromReg uses CALLSEQ_END via a chain and glue.
      auto *CopyFromRegOp = CallSeqEndOp->getGluedUser();
      assert(CopyFromRegOp && CopyFromRegOp->getOpcode() == ISD::CopyFromReg &&
             "Unexpected TLSDESC DAG");
      Ret = SDValue(CopyFromRegOp, 0);
    }
  } else {
    TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
                                     GA->getOffset(), OperandFlags);
  }
 
  if (!Ret) {
    unsigned CallType = UseTLSDESC     ? X86ISD::TLSDESC
                        : LocalDynamic ? X86ISD::TLSBASEADDR
                                       : X86ISD::TLSADDR;
 
    Chain = DAG.getCALLSEQ_START(Chain, 0, 0, dl);
    if (LoadGlobalBaseReg) {
      SDValue InGlue;
      Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
                               DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT),
                               InGlue);
      InGlue = Chain.getValue(1);
      Chain = DAG.getNode(CallType, dl, NodeTys, {Chain, TGA, InGlue});
    } else {
      Chain = DAG.getNode(CallType, dl, NodeTys, {Chain, TGA});
    }
    Chain = DAG.getCALLSEQ_END(Chain, 0, 0, Chain.getValue(1), dl);
 
    // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
    MFI.setHasCalls(true);
 
    SDValue Glue = Chain.getValue(1);
    Ret = DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Glue);
  }
 
  if (!UseTLSDESC)
    return Ret;
 
  const X86Subtarget &Subtarget = DAG.getSubtarget<X86Subtarget>();
  unsigned Seg = Subtarget.is64Bit() ? X86AS::FS : X86AS::GS;
 
  Value *Ptr = Constant::getNullValue(PointerType::get(*DAG.getContext(), Seg));
  SDValue Offset =
      DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
                  MachinePointerInfo(Ptr));
  return DAG.getNode(ISD::ADD, dl, PtrVT, Ret, Offset);
}
 
// Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
static SDValue
LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
                                const EVT PtrVT) {
  return GetTLSADDR(DAG, GA, PtrVT, X86::EAX, X86II::MO_TLSGD,
                    /*LoadGlobalBaseReg=*/true);
}
 
// Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit LP64
static SDValue
LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
                                const EVT PtrVT) {
  return GetTLSADDR(DAG, GA, PtrVT, X86::RAX, X86II::MO_TLSGD);
}
 
// Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit ILP32
static SDValue
LowerToTLSGeneralDynamicModelX32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
                                 const EVT PtrVT) {
  return GetTLSADDR(DAG, GA, PtrVT, X86::EAX, X86II::MO_TLSGD);
}
 
static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
                                           SelectionDAG &DAG, const EVT PtrVT,
                                           bool Is64Bit, bool Is64BitLP64) {
  SDLoc dl(GA);
 
  // Get the start address of the TLS block for this module.
  X86MachineFunctionInfo *MFI =
      DAG.getMachineFunction().getInfo<X86MachineFunctionInfo>();
  MFI->incNumLocalDynamicTLSAccesses();
 
  SDValue Base;
  if (Is64Bit) {
    unsigned ReturnReg = Is64BitLP64 ? X86::RAX : X86::EAX;
    Base = GetTLSADDR(DAG, GA, PtrVT, ReturnReg, X86II::MO_TLSLD,
                      /*LoadGlobalBaseReg=*/false,
                      /*LocalDynamic=*/true);
  } else {
    Base = GetTLSADDR(DAG, GA, PtrVT, X86::EAX, X86II::MO_TLSLDM,
                      /*LoadGlobalBaseReg=*/true,
                      /*LocalDynamic=*/true);
  }
 
  // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
  // of Base.
 
  // Build x@dtpoff.
  unsigned char OperandFlags = X86II::MO_DTPOFF;
  unsigned WrapperKind = X86ISD::Wrapper;
  SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
                                           GA->getValueType(0),
                                           GA->getOffset(), OperandFlags);
  SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
 
  // Add x@dtpoff with the base.
  return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
}
 
// Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
                                   const EVT PtrVT, TLSModel::Model model,
                                   bool is64Bit, bool isPIC) {
  SDLoc dl(GA);
 
  // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
  Value *Ptr = Constant::getNullValue(
      PointerType::get(*DAG.getContext(), is64Bit ? X86AS::FS : X86AS::GS));
 
  SDValue ThreadPointer =
      DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
                  MachinePointerInfo(Ptr));
 
  unsigned char OperandFlags = 0;
  // Most TLS accesses are not RIP relative, even on x86-64.  One exception is
  // initialexec.
  unsigned WrapperKind = X86ISD::Wrapper;
  if (model == TLSModel::LocalExec) {
    OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
  } else if (model == TLSModel::InitialExec) {
    if (is64Bit) {
      OperandFlags = X86II::MO_GOTTPOFF;
      WrapperKind = X86ISD::WrapperRIP;
    } else {
      OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
    }
  } else {
    llvm_unreachable("Unexpected model");
  }
 
  // emit "addl x@ntpoff,%eax" (local exec)
  // or "addl x@indntpoff,%eax" (initial exec)
  // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
  SDValue TGA =
      DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
                                 GA->getOffset(), OperandFlags);
  SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
 
  if (model == TLSModel::InitialExec) {
    if (isPIC && !is64Bit) {
      Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
                           DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
                           Offset);
    }
 
    Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
                         MachinePointerInfo::getGOT(DAG.getMachineFunction()));
  }
 
  // The address of the thread local variable is the add of the thread
  // pointer with the offset of the variable.
  return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
}
 
SDValue
X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
 
  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
 
  if (DAG.getTarget().useEmulatedTLS())
    return LowerToTLSEmulatedModel(GA, DAG);
 
  const GlobalValue *GV = GA->getGlobal();
  EVT PtrVT = Op.getValueType();
  bool PositionIndependent = isPositionIndependent();
 
  if (Subtarget.isTargetELF()) {
    TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
    switch (model) {
      case TLSModel::GeneralDynamic:
        if (Subtarget.is64Bit()) {
          if (Subtarget.isTarget64BitLP64())
            return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
          return LowerToTLSGeneralDynamicModelX32(GA, DAG, PtrVT);
        }
        return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
      case TLSModel::LocalDynamic:
        return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT, Subtarget.is64Bit(),
                                           Subtarget.isTarget64BitLP64());
      case TLSModel::InitialExec:
      case TLSModel::LocalExec:
        return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget.is64Bit(),
                                   PositionIndependent);
    }
    llvm_unreachable("Unknown TLS model.");
  }
 
  if (Subtarget.isTargetDarwin()) {
    // Darwin only has one model of TLS.  Lower to that.
    unsigned char OpFlag = 0;
    unsigned WrapperKind = 0;
 
    // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
    // global base reg.
    bool PIC32 = PositionIndependent && !Subtarget.is64Bit();
    if (PIC32) {
      OpFlag = X86II::MO_TLVP_PIC_BASE;
      WrapperKind = X86ISD::Wrapper;
    } else {
      OpFlag = X86II::MO_TLVP;
      WrapperKind = X86ISD::WrapperRIP;
    }
    SDLoc DL(Op);
    SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
                                                GA->getValueType(0),
                                                GA->getOffset(), OpFlag);
    SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
 
    // With PIC32, the address is actually $g + Offset.
    if (PIC32)
      Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
                           DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
                           Offset);
 
    // Lowering the machine isd will make sure everything is in the right
    // location.
    SDValue Chain = DAG.getEntryNode();
    SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
    Chain = DAG.getCALLSEQ_START(Chain, 0, 0, DL);
    SDValue Args[] = { Chain, Offset };
    Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
    Chain = DAG.getCALLSEQ_END(Chain, 0, 0, Chain.getValue(1), DL);
 
    // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
    MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
    MFI.setAdjustsStack(true);
 
    // And our return value (tls address) is in the standard call return value
    // location.
    unsigned Reg = Subtarget.is64Bit() ? X86::RAX : X86::EAX;
    return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
  }
 
  if (Subtarget.isOSWindows()) {
    // Just use the implicit TLS architecture
    // Need to generate something similar to:
    //   mov     rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
    //                                  ; from TEB
    //   mov     ecx, dword [rel _tls_index]: Load index (from C runtime)
    //   mov     rcx, qword [rdx+rcx*8]
    //   mov     eax, .tls$:tlsvar
    //   [rax+rcx] contains the address
    // Windows 64bit: gs:0x58
    // Windows 32bit: fs:__tls_array
 
    SDLoc dl(GA);
    SDValue Chain = DAG.getEntryNode();
 
    // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
    // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
    // use its literal value of 0x2C.
    Value *Ptr = Constant::getNullValue(
        Subtarget.is64Bit() ? PointerType::get(*DAG.getContext(), X86AS::GS)
                            : PointerType::get(*DAG.getContext(), X86AS::FS));
 
    SDValue TlsArray = Subtarget.is64Bit()
                           ? DAG.getIntPtrConstant(0x58, dl)
                           : (Subtarget.isTargetWindowsGNU()
                                  ? DAG.getIntPtrConstant(0x2C, dl)
                                  : DAG.getExternalSymbol("_tls_array", PtrVT));
 
    SDValue ThreadPointer =
        DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr));
 
    SDValue res;
    if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
      res = ThreadPointer;
    } else {
      // Load the _tls_index variable
      SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
      if (Subtarget.is64Bit())
        IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
                             MachinePointerInfo(), MVT::i32);
      else
        IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo());
 
      const DataLayout &DL = DAG.getDataLayout();
      SDValue Scale =
          DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, MVT::i8);
      IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
 
      res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
    }
 
    res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo());
 
    // Get the offset of start of .tls section
    SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
                                             GA->getValueType(0),
                                             GA->getOffset(), X86II::MO_SECREL);
    SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
 
    // The address of the thread local variable is the add of the thread
    // pointer with the offset of the variable.
    return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
  }
 
  llvm_unreachable("TLS not implemented for this target.");
}
 
bool X86TargetLowering::addressingModeSupportsTLS(const GlobalValue &GV) const {
  if (Subtarget.is64Bit() && Subtarget.isTargetELF()) {
    const TargetMachine &TM = getTargetMachine();
    TLSModel::Model Model = TM.getTLSModel(&GV);
    switch (Model) {
    case TLSModel::LocalExec:
    case TLSModel::InitialExec:
      // We can include the %fs segment register in addressing modes.
      return true;
    case TLSModel::LocalDynamic:
    case TLSModel::GeneralDynamic:
      // These models do not result in %fs relative addresses unless
      // TLS descriptior are used.
      //
      // Even in the case of TLS descriptors we currently have no way to model
      // the difference between %fs access and the computations needed for the
      // offset and returning `true` for TLS-desc currently duplicates both
      // which is detrimental :-/
      return false;
    }
  }
  return false;
}
 
/// Lower SRA_PARTS and friends, which return two i32 values
/// and take a 2 x i32 value to shift plus a shift amount.
/// TODO: Can this be moved to general expansion code?
static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
  SDValue Lo, Hi;
  DAG.getTargetLoweringInfo().expandShiftParts(Op.getNode(), Lo, Hi, DAG);
  return DAG.getMergeValues({Lo, Hi}, SDLoc(Op));
}
 
// Try to use a packed vector operation to handle i64 on 32-bit targets when
// AVX512DQ is enabled.
static SDValue LowerI64IntToFP_AVX512DQ(SDValue Op, const SDLoc &dl,
                                        SelectionDAG &DAG,
                                        const X86Subtarget &Subtarget) {
  assert((Op.getOpcode() == ISD::SINT_TO_FP ||
          Op.getOpcode() == ISD::STRICT_SINT_TO_FP ||
          Op.getOpcode() == ISD::STRICT_UINT_TO_FP ||
          Op.getOpcode() == ISD::UINT_TO_FP) &&
         "Unexpected opcode!");
  bool IsStrict = Op->isStrictFPOpcode();
  unsigned OpNo = IsStrict ? 1 : 0;
  SDValue Src = Op.getOperand(OpNo);
  MVT SrcVT = Src.getSimpleValueType();
  MVT VT = Op.getSimpleValueType();
 
   if (!Subtarget.hasDQI() || SrcVT != MVT::i64 || Subtarget.is64Bit() ||
       (VT != MVT::f32 && VT != MVT::f64))
    return SDValue();
 
  // Pack the i64 into a vector, do the operation and extract.
 
  // Using 256-bit to ensure result is 128-bits for f32 case.
  unsigned NumElts = Subtarget.hasVLX() ? 4 : 8;
  MVT VecInVT = MVT::getVectorVT(MVT::i64, NumElts);
  MVT VecVT = MVT::getVectorVT(VT, NumElts);
 
  SDValue InVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecInVT, Src);
  if (IsStrict) {
    SDValue CvtVec = DAG.getNode(Op.getOpcode(), dl, {VecVT, MVT::Other},
                                 {Op.getOperand(0), InVec});
    SDValue Chain = CvtVec.getValue(1);
    SDValue Value = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, CvtVec,
                                DAG.getVectorIdxConstant(0, dl));
    return DAG.getMergeValues({Value, Chain}, dl);
  }
 
  SDValue CvtVec = DAG.getNode(Op.getOpcode(), dl, VecVT, InVec);
 
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, CvtVec,
                     DAG.getVectorIdxConstant(0, dl));
}
 
// Try to use a packed vector operation to handle i64 on 32-bit targets.
static SDValue LowerI64IntToFP16(SDValue Op, const SDLoc &dl, SelectionDAG &DAG,
                                 const X86Subtarget &Subtarget) {
  assert((Op.getOpcode() == ISD::SINT_TO_FP ||
          Op.getOpcode() == ISD::STRICT_SINT_TO_FP ||
          Op.getOpcode() == ISD::STRICT_UINT_TO_FP ||
          Op.getOpcode() == ISD::UINT_TO_FP) &&
         "Unexpected opcode!");
  bool IsStrict = Op->isStrictFPOpcode();
  SDValue Src = Op.getOperand(IsStrict ? 1 : 0);
  MVT SrcVT = Src.getSimpleValueType();
  MVT VT = Op.getSimpleValueType();
 
  if (SrcVT != MVT::i64 || Subtarget.is64Bit() || VT != MVT::f16)
    return SDValue();
 
  // Pack the i64 into a vector, do the operation and extract.
 
  assert(Subtarget.hasFP16() && "Expected FP16");
 
  SDValue InVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Src);
  if (IsStrict) {
    SDValue CvtVec = DAG.getNode(Op.getOpcode(), dl, {MVT::v2f16, MVT::Other},
                                 {Op.getOperand(0), InVec});
    SDValue Chain = CvtVec.getValue(1);
    SDValue Value = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, CvtVec,
                                DAG.getVectorIdxConstant(0, dl));
    return DAG.getMergeValues({Value, Chain}, dl);
  }
 
  SDValue CvtVec = DAG.getNode(Op.getOpcode(), dl, MVT::v2f16, InVec);
 
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, CvtVec,
                     DAG.getVectorIdxConstant(0, dl));
}
 
static bool useVectorCast(unsigned Opcode, MVT FromVT, MVT ToVT,
                          const X86Subtarget &Subtarget) {
  switch (Opcode) {
    case ISD::SINT_TO_FP:
      // TODO: Handle wider types with AVX/AVX512.
      if (!Subtarget.hasSSE2() || FromVT != MVT::v4i32)
        return false;
      // CVTDQ2PS or (V)CVTDQ2PD
      return ToVT == MVT::v4f32 || (Subtarget.hasAVX() && ToVT == MVT::v4f64);
 
    case ISD::UINT_TO_FP:
      // TODO: Handle wider types and i64 elements.
      if (!Subtarget.hasAVX512() || FromVT != MVT::v4i32)
        return false;
      // VCVTUDQ2PS or VCVTUDQ2PD
      return ToVT == MVT::v4f32 || ToVT == MVT::v4f64;
 
    default:
      return false;
  }
}
 
/// Given a scalar cast operation that is extracted from a vector, try to
/// vectorize the cast op followed by extraction. This will avoid an expensive
/// round-trip between XMM and GPR.
static SDValue vectorizeExtractedCast(SDValue Cast, const SDLoc &DL,
                                      SelectionDAG &DAG,
                                      const X86Subtarget &Subtarget) {
  // TODO: This could be enhanced to handle smaller integer types by peeking
  // through an extend.
  SDValue Extract = Cast.getOperand(0);
  MVT DestVT = Cast.getSimpleValueType();
  if (Extract.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
      !isa<ConstantSDNode>(Extract.getOperand(1)))
    return SDValue();
 
  // See if we have a 128-bit vector cast op for this type of cast.
  SDValue VecOp = Extract.getOperand(0);
  EVT FromVT = VecOp.getValueType();
  unsigned NumEltsInXMM = 128 / FromVT.getScalarSizeInBits();
  MVT Vec128VT =
      MVT::getVectorVT(FromVT.getScalarType().getSimpleVT(), NumEltsInXMM);
  MVT ToVT = MVT::getVectorVT(DestVT, NumEltsInXMM);
  if (!useVectorCast(Cast.getOpcode(), Vec128VT, ToVT, Subtarget))
    return SDValue();
 
  // If we are extracting from a non-zero element, first shuffle the source
  // vector to allow extracting from element zero.
  if (!isNullConstant(Extract.getOperand(1))) {
    SmallVector<int, 16> Mask(FromVT.getVectorNumElements(), -1);
    Mask[0] = Extract.getConstantOperandVal(1);
    VecOp = DAG.getVectorShuffle(FromVT, DL, VecOp, DAG.getUNDEF(FromVT), Mask);
  }
  // If the source vector is wider than 128-bits, extract the low part. Do not
  // create an unnecessarily wide vector cast op.
  if (FromVT != Vec128VT)
    VecOp = extract128BitVector(VecOp, 0, DAG, DL);
 
  // cast (extelt V, 0) --> extelt (cast (extract_subv V)), 0
  // cast (extelt V, C) --> extelt (cast (extract_subv (shuffle V, [C...]))), 0
  SDValue VCast = DAG.getNode(Cast.getOpcode(), DL, ToVT, VecOp);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, DestVT, VCast,
                     DAG.getVectorIdxConstant(0, DL));
}
 
/// Given a scalar cast to FP with a cast to integer operand (almost an ftrunc),
/// try to vectorize the cast ops. This will avoid an expensive round-trip
/// between XMM and GPR.
static SDValue lowerFPToIntToFP(SDValue CastToFP, const SDLoc &DL,
                                SelectionDAG &DAG,
                                const X86Subtarget &Subtarget) {
  SDValue CastToInt = CastToFP.getOperand(0);
  MVT VT = CastToFP.getSimpleValueType();
  if ((CastToInt.getOpcode() != ISD::FP_TO_SINT &&
       CastToInt.getOpcode() != ISD::FP_TO_UINT) ||
      VT.isVector())
    return SDValue();
 
  MVT IntVT = CastToInt.getSimpleValueType();
  SDValue X = CastToInt.getOperand(0);
  MVT SrcVT = X.getSimpleValueType();
  if (SrcVT != MVT::f32 && SrcVT != MVT::f64)
    return SDValue();
 
  // See if we have 128-bit vector cast instructions for this type of cast.
  // We need cvttps2dq/cvttpd2dq and cvtdq2ps/cvtdq2pd.
  if (!Subtarget.hasSSE2() || (VT != MVT::f32 && VT != MVT::f64) ||
      (IntVT != MVT::i32 && IntVT != MVT::i64))
    return SDValue();
 
  unsigned SrcSize = SrcVT.getSizeInBits();
  unsigned IntSize = IntVT.getSizeInBits();
  unsigned VTSize = VT.getSizeInBits();
  bool IsUnsigned = CastToInt.getOpcode() == ISD::FP_TO_UINT;
  unsigned ToIntOpcode =
      SrcSize != IntSize ? X86ISD::CVTTP2SI : (unsigned)ISD::FP_TO_SINT;
  unsigned ToFPOpcode =
      IntSize != VTSize ? X86ISD::CVTSI2P : (unsigned)ISD::SINT_TO_FP;
  unsigned Width = 128;
 
  if (Subtarget.hasVLX() && Subtarget.hasDQI()) {
    // AVX512DQ+VLX
    if (IsUnsigned) {
      ToIntOpcode =
          SrcSize != IntSize ? X86ISD::CVTTP2UI : (unsigned)ISD::FP_TO_UINT;
      ToFPOpcode =
          IntSize != VTSize ? X86ISD::CVTUI2P : (unsigned)ISD::UINT_TO_FP;
    }
  } else {
    if (IsUnsigned || IntVT == MVT::i64) {
      // SSE2 can only perform f64/f32 <-> i32 signed.
      if (!Subtarget.useAVX512Regs() || !Subtarget.hasDQI())
        return SDValue();
 
      // Need to extend width for AVX512DQ without AVX512VL.
      Width = 512;
      ToIntOpcode = CastToInt.getOpcode();
      ToFPOpcode = IsUnsigned ? ISD::UINT_TO_FP : ISD::SINT_TO_FP;
    }
  }
 
  MVT VecSrcVT, VecIntVT, VecVT;
  unsigned NumElts;
  unsigned SrcElts, VTElts;
  // Some conversions are only legal with uniform vector sizes on AVX512DQ.
  if (Width == 512) {
    NumElts = std::min(Width / IntSize, Width / SrcSize);
    SrcElts = NumElts;
    VTElts = NumElts;
  } else {
    NumElts = Width / IntSize;
    SrcElts = Width / SrcSize;
    VTElts = Width / VTSize;
  }
  VecIntVT = MVT::getVectorVT(IntVT, NumElts);
  VecSrcVT = MVT::getVectorVT(SrcVT, SrcElts);
  VecVT = MVT::getVectorVT(VT, VTElts);
  // sint_to_fp (fp_to_sint X) --> extelt (sint_to_fp (fp_to_sint (s2v X))), 0
  //
  // We are not defining the high elements (for example, zero them) because
  // that could nullify any performance advantage that we hoped to gain from
  // this vector op hack. We do not expect any adverse effects (like denorm
  // penalties) with cast ops.
  SDValue ZeroIdx = DAG.getVectorIdxConstant(0, DL);
  SDValue VecX = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecSrcVT, X);
  SDValue VCastToInt = DAG.getNode(ToIntOpcode, DL, VecIntVT, VecX);
  SDValue VCastToFP = DAG.getNode(ToFPOpcode, DL, VecVT, VCastToInt);
  return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT, VCastToFP, ZeroIdx);
}
 
static SDValue lowerINT_TO_FP_vXi64(SDValue Op, const SDLoc &DL,
                                    SelectionDAG &DAG,
                                    const X86Subtarget &Subtarget) {
  bool IsStrict = Op->isStrictFPOpcode();
  MVT VT = Op->getSimpleValueType(0);
  SDValue Src = Op->getOperand(IsStrict ? 1 : 0);
 
  if (Subtarget.hasDQI()) {
    assert(!Subtarget.hasVLX() && "Unexpected features");
 
    assert((Src.getSimpleValueType() == MVT::v2i64 ||
            Src.getSimpleValueType() == MVT::v4i64) &&
           "Unsupported custom type");
 
    // With AVX512DQ, but not VLX we need to widen to get a 512-bit result type.
    assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v4f64) &&
           "Unexpected VT!");
    MVT WideVT = VT == MVT::v4f32 ? MVT::v8f32 : MVT::v8f64;
 
    // Need to concat with zero vector for strict fp to avoid spurious
    // exceptions.
    SDValue Tmp = IsStrict ? DAG.getConstant(0, DL, MVT::v8i64)
                           : DAG.getUNDEF(MVT::v8i64);
    Src = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i64, Tmp, Src,
                      DAG.getVectorIdxConstant(0, DL));
    SDValue Res, Chain;
    if (IsStrict) {
      Res = DAG.getNode(Op.getOpcode(), DL, {WideVT, MVT::Other},
                        {Op->getOperand(0), Src});
      Chain = Res.getValue(1);
    } else {
      Res = DAG.getNode(Op.getOpcode(), DL, WideVT, Src);
    }
 
    Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res,
                      DAG.getVectorIdxConstant(0, DL));
 
    if (IsStrict)
      return DAG.getMergeValues({Res, Chain}, DL);
    return Res;
  }
 
  bool IsSigned = Op->getOpcode() == ISD::SINT_TO_FP ||
                  Op->getOpcode() == ISD::STRICT_SINT_TO_FP;
  if (VT != MVT::v4f32 || IsSigned)
    return SDValue();
 
  SDValue Zero = DAG.getConstant(0, DL, MVT::v4i64);
  SDValue One  = DAG.getConstant(1, DL, MVT::v4i64);
  SDValue Sign = DAG.getNode(ISD::OR, DL, MVT::v4i64,
                             DAG.getNode(ISD::SRL, DL, MVT::v4i64, Src, One),
                             DAG.getNode(ISD::AND, DL, MVT::v4i64, Src, One));
  SDValue IsNeg = DAG.getSetCC(DL, MVT::v4i64, Src, Zero, ISD::SETLT);
  SDValue SignSrc = DAG.getSelect(DL, MVT::v4i64, IsNeg, Sign, Src);
  SmallVector<SDValue, 4> SignCvts(4);
  SmallVector<SDValue, 4> Chains(4);
  for (int i = 0; i != 4; ++i) {
    SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::i64, SignSrc,
                              DAG.getVectorIdxConstant(i, DL));
    if (IsStrict) {
      SignCvts[i] =
          DAG.getNode(ISD::STRICT_SINT_TO_FP, DL, {MVT::f32, MVT::Other},
                      {Op.getOperand(0), Elt});
      Chains[i] = SignCvts[i].getValue(1);
    } else {
      SignCvts[i] = DAG.getNode(ISD::SINT_TO_FP, DL, MVT::f32, Elt);
    }
  }
  SDValue SignCvt = DAG.getBuildVector(VT, DL, SignCvts);
 
  SDValue Slow, Chain;
  if (IsStrict) {
    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Chains);
    Slow = DAG.getNode(ISD::STRICT_FADD, DL, {MVT::v4f32, MVT::Other},
                       {Chain, SignCvt, SignCvt});
    Chain = Slow.getValue(1);
  } else {
    Slow = DAG.getNode(ISD::FADD, DL, MVT::v4f32, SignCvt, SignCvt);
  }
 
  IsNeg = DAG.getNode(ISD::TRUNCATE, DL, MVT::v4i32, IsNeg);
  SDValue Cvt = DAG.getSelect(DL, MVT::v4f32, IsNeg, Slow, SignCvt);
 
  if (IsStrict)
    return DAG.getMergeValues({Cvt, Chain}, DL);
 
  return Cvt;
}
 
static SDValue promoteXINT_TO_FP(SDValue Op, const SDLoc &dl,
                                 SelectionDAG &DAG) {
  bool IsStrict = Op->isStrictFPOpcode();
  SDValue Src = Op.getOperand(IsStrict ? 1 : 0);
  SDValue Chain = IsStrict ? Op->getOperand(0) : DAG.getEntryNode();
  MVT VT = Op.getSimpleValueType();
  MVT NVT = VT.isVector() ? VT.changeVectorElementType(MVT::f32) : MVT::f32;
 
  SDValue Rnd = DAG.getIntPtrConstant(0, dl, /*isTarget=*/true);
  if (IsStrict)
    return DAG.getNode(
        ISD::STRICT_FP_ROUND, dl, {VT, MVT::Other},
        {Chain,
         DAG.getNode(Op.getOpcode(), dl, {NVT, MVT::Other}, {Chain, Src}),
         Rnd});
  return DAG.getNode(ISD::FP_ROUND, dl, VT,
                     DAG.getNode(Op.getOpcode(), dl, NVT, Src), Rnd);
}
 
static bool isLegalConversion(MVT VT, MVT FloatVT, bool IsSigned,
                              const X86Subtarget &Subtarget) {
  if (FloatVT.getScalarType() != MVT::f16 || Subtarget.hasVLX()) {
    if (VT == MVT::v4i32 && Subtarget.hasSSE2() && IsSigned)
      return true;
    if (VT == MVT::v8i32 && Subtarget.hasAVX() && IsSigned)
      return true;
  }
  if (Subtarget.hasVLX() && (VT == MVT::v4i32 || VT == MVT::v8i32))
    return true;
  if (Subtarget.useAVX512Regs()) {
    if (VT == MVT::v16i32)
      return true;
    if (VT == MVT::v8i64 && FloatVT == MVT::v8f16 && Subtarget.hasFP16())
      return true;
    if (VT == MVT::v8i64 && Subtarget.hasDQI())
      return true;
  }
  if (Subtarget.hasDQI() && Subtarget.hasVLX() &&
      (VT == MVT::v2i64 || VT == MVT::v4i64))
    return true;
  return false;
}
 
SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
                                           SelectionDAG &DAG) const {
  bool IsStrict = Op->isStrictFPOpcode();
  unsigned OpNo = IsStrict ? 1 : 0;
  SDValue Src = Op.getOperand(OpNo);
  SDValue Chain = IsStrict ? Op->getOperand(0) : DAG.getEntryNode();
  MVT SrcVT = Src.getSimpleValueType();
  MVT VT = Op.getSimpleValueType();
  SDLoc dl(Op);
 
  if (isBF16orSoftF16(VT, Subtarget))
    return promoteXINT_TO_FP(Op, dl, DAG);
  else if (isLegalConversion(SrcVT, VT, true, Subtarget))
    return Op;
 
  if (Subtarget.isTargetWin64() && SrcVT == MVT::i128)
    return LowerWin64_INT128_TO_FP(Op, DAG);
 
  if (SDValue Extract = vectorizeExtractedCast(Op, dl, DAG, Subtarget))
    return Extract;
 
  if (SDValue R = lowerFPToIntToFP(Op, dl, DAG, Subtarget))
    return R;
 
  if (SrcVT.isVector()) {
    if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
      // Note: Since v2f64 is a legal type. We don't need to zero extend the
      // source for strict FP.
      if (IsStrict)
        return DAG.getNode(
            X86ISD::STRICT_CVTSI2P, dl, {VT, MVT::Other},
            {Chain, DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
                                DAG.getUNDEF(SrcVT))});
      return DAG.getNode(X86ISD::CVTSI2P, dl, VT,
                         DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
                                     DAG.getUNDEF(SrcVT)));
    }
    if (SrcVT == MVT::v2i64 || SrcVT == MVT::v4i64)
      return lowerINT_TO_FP_vXi64(Op, dl, DAG, Subtarget);
 
    return SDValue();
  }
 
  assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
         "Unknown SINT_TO_FP to lower!");
 
  bool UseSSEReg = isScalarFPTypeInSSEReg(VT);
 
  // These are really Legal; return the operand so the caller accepts it as
  // Legal.
  if (SrcVT == MVT::i32 && UseSSEReg)
    return Op;
  if (SrcVT == MVT::i64 && UseSSEReg && Subtarget.is64Bit())
    return Op;
 
  if (SDValue V = LowerI64IntToFP_AVX512DQ(Op, dl, DAG, Subtarget))
    return V;
  if (SDValue V = LowerI64IntToFP16(Op, dl, DAG, Subtarget))
    return V;
 
  // SSE doesn't have an i16 conversion so we need to promote.
  if (SrcVT == MVT::i16 && (UseSSEReg || VT == MVT::f128)) {
    SDValue Ext = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i32, Src);
    if (IsStrict)
      return DAG.getNode(ISD::STRICT_SINT_TO_FP, dl, {VT, MVT::Other},
                         {Chain, Ext});
 
    return DAG.getNode(ISD::SINT_TO_FP, dl, VT, Ext);
  }
 
  if (VT == MVT::f128 || !Subtarget.hasX87())
    return SDValue();
 
  SDValue ValueToStore = Src;
  if (SrcVT == MVT::i64 && Subtarget.hasSSE2() && !Subtarget.is64Bit())
    // Bitcasting to f64 here allows us to do a single 64-bit store from
    // an SSE register, avoiding the store forwarding penalty that would come
    // with two 32-bit stores.
    ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore);
 
  unsigned Size = SrcVT.getStoreSize();
  Align Alignment(Size);
  MachineFunction &MF = DAG.getMachineFunction();
  auto PtrVT = getPointerTy(MF.getDataLayout());
  int SSFI = MF.getFrameInfo().CreateStackObject(Size, Alignment, false);
  MachinePointerInfo MPI =
      MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI);
  SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
  Chain = DAG.getStore(Chain, dl, ValueToStore, StackSlot, MPI, Alignment);
  std::pair<SDValue, SDValue> Tmp =
      BuildFILD(VT, SrcVT, dl, Chain, StackSlot, MPI, Alignment, DAG);
 
  if (IsStrict)
    return DAG.getMergeValues({Tmp.first, Tmp.second}, dl);
 
  return Tmp.first;
}
 
std::pair<SDValue, SDValue> X86TargetLowering::BuildFILD(
    EVT DstVT, EVT SrcVT, const SDLoc &DL, SDValue Chain, SDValue Pointer,
    MachinePointerInfo PtrInfo, Align Alignment, SelectionDAG &DAG) const {
  // Build the FILD
  SDVTList Tys;
  bool useSSE = isScalarFPTypeInSSEReg(DstVT);
  if (useSSE)
    Tys = DAG.getVTList(MVT::f80, MVT::Other);
  else
    Tys = DAG.getVTList(DstVT, MVT::Other);
 
  SDValue FILDOps[] = {Chain, Pointer};
  SDValue Result =
      DAG.getMemIntrinsicNode(X86ISD::FILD, DL, Tys, FILDOps, SrcVT, PtrInfo,
                              Alignment, MachineMemOperand::MOLoad);
  Chain = Result.getValue(1);
 
  if (useSSE) {
    MachineFunction &MF = DAG.getMachineFunction();
    unsigned SSFISize = DstVT.getStoreSize();
    int SSFI =
        MF.getFrameInfo().CreateStackObject(SSFISize, Align(SSFISize), false);
    auto PtrVT = getPointerTy(MF.getDataLayout());
    SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
    Tys = DAG.getVTList(MVT::Other);
    SDValue FSTOps[] = {Chain, Result, StackSlot};
    MachineMemOperand *StoreMMO = DAG.getMachineFunction().getMachineMemOperand(
        MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
        MachineMemOperand::MOStore, SSFISize, Align(SSFISize));
 
    Chain =
        DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys, FSTOps, DstVT, StoreMMO);
    Result = DAG.getLoad(
        DstVT, DL, Chain, StackSlot,
        MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI));
    Chain = Result.getValue(1);
  }
 
  return { Result, Chain };
}
 
/// Horizontal vector math instructions may be slower than normal math with
/// shuffles. Limit horizontal op codegen based on size/speed trade-offs, uarch
/// implementation, and likely shuffle complexity of the alternate sequence.
static bool shouldUseHorizontalOp(bool IsSingleSource, const SelectionDAG &DAG,
                                  const X86Subtarget &Subtarget) {
  bool IsOptimizingSize = DAG.shouldOptForSize();
  bool HasFastHOps = Subtarget.hasFastHorizontalOps();
  return !IsSingleSource || IsOptimizingSize || HasFastHOps;
}
 
/// 64-bit unsigned integer to double expansion.
static SDValue LowerUINT_TO_FP_i64(SDValue Op, const SDLoc &dl,
                                   SelectionDAG &DAG,
                                   const X86Subtarget &Subtarget) {
  // We can't use this algorithm for strict fp. It produces -0.0 instead of +0.0
  // when converting 0 when rounding toward negative infinity. Caller will
  // fall back to Expand for when i64 or is legal or use FILD in 32-bit mode.
  assert(!Op->isStrictFPOpcode() && "Expected non-strict uint_to_fp!");
  // This algorithm is not obvious. Here it is what we're trying to output:
  /*
     movq       %rax,  %xmm0
     punpckldq  (c0),  %xmm0  // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
     subpd      (c1),  %xmm0  // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
     #ifdef __SSE3__
       haddpd   %xmm0, %xmm0
     #else
       pshufd   $0x4e, %xmm0, %xmm1
       addpd    %xmm1, %xmm0
     #endif
  */
 
  LLVMContext *Context = DAG.getContext();
 
  // Build some magic constants.
  static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
  Constant *C0 = ConstantDataVector::get(*Context, CV0);
  auto PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
  SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, Align(16));
 
  SmallVector<Constant*,2> CV1;
  CV1.push_back(
    ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble(),
                                      APInt(64, 0x4330000000000000ULL))));
  CV1.push_back(
    ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble(),
                                      APInt(64, 0x4530000000000000ULL))));
  Constant *C1 = ConstantVector::get(CV1);
  SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, Align(16));
 
  // Load the 64-bit value into an XMM register.
  SDValue XR1 =
      DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Op.getOperand(0));
  SDValue CLod0 = DAG.getLoad(
      MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
      MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Align(16));
  SDValue Unpck1 =
      getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
 
  SDValue CLod1 = DAG.getLoad(
      MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
      MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Align(16));
  SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
  // TODO: Are there any fast-math-flags to propagate here?
  SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
  SDValue Result;
 
  if (Subtarget.hasSSE3() &&
      shouldUseHorizontalOp(true, DAG, Subtarget)) {
    Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
  } else {
    SDValue Shuffle = DAG.getVectorShuffle(MVT::v2f64, dl, Sub, Sub, {1,-1});
    Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuffle, Sub);
  }
  Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
                       DAG.getVectorIdxConstant(0, dl));
  return Result;
}
 
/// 32-bit unsigned integer to float expansion.
static SDValue LowerUINT_TO_FP_i32(SDValue Op, const SDLoc &dl,
                                   SelectionDAG &DAG,
                                   const X86Subtarget &Subtarget) {
  unsigned OpNo = Op.getNode()->isStrictFPOpcode() ? 1 : 0;
  // FP constant to bias correct the final result.
  SDValue Bias = DAG.getConstantFP(
      llvm::bit_cast<double>(0x4330000000000000ULL), dl, MVT::f64);
 
  // Load the 32-bit value into an XMM register.
  SDValue Load =
      DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Op.getOperand(OpNo));
 
  // Zero out the upper parts of the register.
  Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
 
  // Or the load with the bias.
  SDValue Or = DAG.getNode(
      ISD::OR, dl, MVT::v2i64,
      DAG.getBitcast(MVT::v2i64, Load),
      DAG.getBitcast(MVT::v2i64,
                     DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
  Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
                   DAG.getBitcast(MVT::v2f64, Or),
                   DAG.getVectorIdxConstant(0, dl));
 
  if (Op.getNode()->isStrictFPOpcode()) {
    // Subtract the bias.
    // TODO: Are there any fast-math-flags to propagate here?
    SDValue Chain = Op.getOperand(0);
    SDValue Sub = DAG.getNode(ISD::STRICT_FSUB, dl, {MVT::f64, MVT::Other},
                              {Chain, Or, Bias});
 
    if (Op.getValueType() == Sub.getValueType())
      return Sub;
 
    // Handle final rounding.
    std::pair<SDValue, SDValue> ResultPair = DAG.getStrictFPExtendOrRound(
        Sub, Sub.getValue(1), dl, Op.getSimpleValueType());
 
    return DAG.getMergeValues({ResultPair.first, ResultPair.second}, dl);
  }
 
  // Subtract the bias.
  // TODO: Are there any fast-math-flags to propagate here?
  SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
 
  // Handle final rounding.
  return DAG.getFPExtendOrRound(Sub, dl, Op.getSimpleValueType());
}
 
static SDValue lowerUINT_TO_FP_v2i32(SDValue Op, const SDLoc &DL,
                                     SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
  if (Op.getSimpleValueType() != MVT::v2f64)
    return SDValue();
 
  bool IsStrict = Op->isStrictFPOpcode();
 
  SDValue N0 = Op.getOperand(IsStrict ? 1 : 0);
  assert(N0.getSimpleValueType() == MVT::v2i32 && "Unexpected input type");
 
  if (Subtarget.hasAVX512()) {
    if (!Subtarget.hasVLX()) {
      // Let generic type legalization widen this.
      if (!IsStrict)
        return SDValue();
      // Otherwise pad the integer input with 0s and widen the operation.
      N0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4i32, N0,
                       DAG.getConstant(0, DL, MVT::v2i32));
      SDValue Res = DAG.getNode(Op->getOpcode(), DL, {MVT::v4f64, MVT::Other},
                                {Op.getOperand(0), N0});
      SDValue Chain = Res.getValue(1);
      Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2f64, Res,
                        DAG.getVectorIdxConstant(0, DL));
      return DAG.getMergeValues({Res, Chain}, DL);
    }
 
    // Legalize to v4i32 type.
    N0 = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4i32, N0,
                     DAG.getUNDEF(MVT::v2i32));
    if (IsStrict)
      return DAG.getNode(X86ISD::STRICT_CVTUI2P, DL, {MVT::v2f64, MVT::Other},
                         {Op.getOperand(0), N0});
    return DAG.getNode(X86ISD::CVTUI2P, DL, MVT::v2f64, N0);
  }
 
  // Zero extend to 2i64, OR with the floating point representation of 2^52.
  // This gives us the floating point equivalent of 2^52 + the i32 integer
  // since double has 52-bits of mantissa. Then subtract 2^52 in floating
  // point leaving just our i32 integers in double format.
  SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v2i64, N0);
  SDValue VBias = DAG.getConstantFP(
      llvm::bit_cast<double>(0x4330000000000000ULL), DL, MVT::v2f64);
  SDValue Or = DAG.getNode(ISD::OR, DL, MVT::v2i64, ZExtIn,
                           DAG.getBitcast(MVT::v2i64, VBias));
  Or = DAG.getBitcast(MVT::v2f64, Or);
 
  if (IsStrict)
    return DAG.getNode(ISD::STRICT_FSUB, DL, {MVT::v2f64, MVT::Other},
                       {Op.getOperand(0), Or, VBias});
  return DAG.getNode(ISD::FSUB, DL, MVT::v2f64, Or, VBias);
}
 
static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, const SDLoc &DL,
                                     SelectionDAG &DAG,
                                     const X86Subtarget &Subtarget) {
  bool IsStrict = Op->isStrictFPOpcode();
  SDValue V = Op->getOperand(IsStrict ? 1 : 0);
  MVT VecIntVT = V.getSimpleValueType();
  assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
         "Unsupported custom type");
 
  if (Subtarget.hasAVX512()) {
    // With AVX512, but not VLX we need to widen to get a 512-bit result type.
    assert(!Subtarget.hasVLX() && "Unexpected features");
    MVT VT = Op->getSimpleValueType(0);
 
    // v8i32->v8f64 is legal with AVX512 so just return it.
    if (VT == MVT::v8f64)
      return Op;
 
    assert((VT == MVT::v4f32 || VT == MVT::v8f32 || VT == MVT::v4f64 ||
            VT == MVT::v8f16) &&
           "Unexpected VT!");
    MVT WideVT = VT == MVT::v8f16 ? MVT::v16f16 : MVT::v16f32;
    MVT WideIntVT = MVT::v16i32;
    if (VT == MVT::v4f64) {
      WideVT = MVT::v8f64;
      WideIntVT = MVT::v8i32;
    }
 
    // Need to concat with zero vector for strict fp to avoid spurious
    // exceptions.
    SDValue Tmp =
        IsStrict ? DAG.getConstant(0, DL, WideIntVT) : DAG.getUNDEF(WideIntVT);
    V = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideIntVT, Tmp, V,
                    DAG.getVectorIdxConstant(0, DL));
    SDValue Res, Chain;
    if (IsStrict) {
      Res = DAG.getNode(ISD::STRICT_UINT_TO_FP, DL, {WideVT, MVT::Other},
                        {Op->getOperand(0), V});
      Chain = Res.getValue(1);
    } else {
      Res = DAG.getNode(ISD::UINT_TO_FP, DL, WideVT, V);
    }
 
    Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, Res,
                      DAG.getVectorIdxConstant(0, DL));
 
    if (IsStrict)
      return DAG.getMergeValues({Res, Chain}, DL);
    return Res;
  }
 
  if (Subtarget.hasAVX() && VecIntVT == MVT::v4i32 &&
      Op->getSimpleValueType(0) == MVT::v4f64) {
    SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i64, V);
    Constant *Bias = ConstantFP::get(
        *DAG.getContext(),
        APFloat(APFloat::IEEEdouble(), APInt(64, 0x4330000000000000ULL)));
    auto PtrVT = DAG.getTargetLoweringInfo().getPointerTy(DAG.getDataLayout());
    SDValue CPIdx = DAG.getConstantPool(Bias, PtrVT, Align(8));
    SDVTList Tys = DAG.getVTList(MVT::v4f64, MVT::Other);
    SDValue Ops[] = {DAG.getEntryNode(), CPIdx};
    SDValue VBias = DAG.getMemIntrinsicNode(
        X86ISD::VBROADCAST_LOAD, DL, Tys, Ops, MVT::f64,
        MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), Align(8),
        MachineMemOperand::MOLoad);
 
    SDValue Or = DAG.getNode(ISD::OR, DL, MVT::v4i64, ZExtIn,
                             DAG.getBitcast(MVT::v4i64, VBias));
    Or = DAG.getBitcast(MVT::v4f64, Or);
 
    if (IsStrict)
      return DAG.getNode(ISD::STRICT_FSUB, DL, {MVT::v4f64, MVT::Other},
                         {Op.getOperand(0), Or, VBias});
    return DAG.getNode(ISD::FSUB, DL, MVT::v4f64, Or, VBias);
  }
 
  // The algorithm is the following:
  // #ifdef __SSE4_1__
  //     uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
  //     uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
  //                                 (uint4) 0x53000000, 0xaa);
  // #else
  //     uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
  //     uint4 hi = (v >> 16) | (uint4) 0x53000000;
  // #endif
  //     float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
  //     return (float4) lo + fhi;
 
  bool Is128 = VecIntVT == MVT::v4i32;
  MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
  // If we convert to something else than the supported type, e.g., to v4f64,
  // abort early.
  if (VecFloatVT != Op->getSimpleValueType(0))
    return SDValue();
 
  // In the #idef/#else code, we have in common:
  // - The vector of constants:
  // -- 0x4b000000
  // -- 0x53000000
  // - A shift:
  // -- v >> 16
 
  // Create the splat vector for 0x4b000000.
  SDValue VecCstLow = DAG.getConstant(0x4b000000, DL, VecIntVT);
  // Create the splat vector for 0x53000000.
  SDValue VecCstHigh = DAG.getConstant(0x53000000, DL, VecIntVT);
 
  // Create the right shift.
  SDValue VecCstShift = DAG.getConstant(16, DL, VecIntVT);
  SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
 
  SDValue Low, High;
  if (Subtarget.hasSSE41()) {
    MVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
    //     uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
    SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
    SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
    // Low will be bitcasted right away, so do not bother bitcasting back to its
    // original type.
    Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
                      VecCstLowBitcast, DAG.getTargetConstant(0xaa, DL, MVT::i8));
    //     uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
    //                                 (uint4) 0x53000000, 0xaa);
    SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
    SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
    // High will be bitcasted right away, so do not bother bitcasting back to
    // its original type.
    High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
                       VecCstHighBitcast, DAG.getTargetConstant(0xaa, DL, MVT::i8));
  } else {
    SDValue VecCstMask = DAG.getConstant(0xffff, DL, VecIntVT);
    //     uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
    SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
    Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
 
    //     uint4 hi = (v >> 16) | (uint4) 0x53000000;
    High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
  }
 
  // Create the vector constant for (0x1.0p39f + 0x1.0p23f).
  SDValue VecCstFSub = DAG.getConstantFP(
      APFloat(APFloat::IEEEsingle(), APInt(32, 0x53000080)), DL, VecFloatVT);
 
  //     float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
  // NOTE: By using fsub of a positive constant instead of fadd of a negative
  // constant, we avoid reassociation in MachineCombiner when reassoc is
  // enabled. See PR24512.
  SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
  // TODO: Are there any fast-math-flags to propagate here?
  //     (float4) lo;
  SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
  //     return (float4) lo + fhi;
  if (IsStrict) {
    SDValue FHigh = DAG.getNode(ISD::STRICT_FSUB, DL, {VecFloatVT, MVT::Other},
                                {Op.getOperand(0), HighBitcast, VecCstFSub});
    return DAG.getNode(ISD::STRICT_FADD, DL, {VecFloatVT, MVT::Other},
                       {FHigh.getValue(1), LowBitcast, FHigh});
  }
 
  SDValue FHigh =
      DAG.getNode(ISD::FSUB, DL, VecFloatVT, HighBitcast, VecCstFSub);
  return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
}
 
static SDValue lowerUINT_TO_FP_vec(SDValue Op, const SDLoc &dl, SelectionDAG &DAG,
                                   const X86Subtarget &Subtarget) {
  unsigned OpNo = Op.getNode()->isStrictFPOpcode() ? 1 : 0;
  SDValue N0 = Op.getOperand(OpNo);
  MVT SrcVT = N0.getSimpleValueType();
 
  switch (SrcVT.SimpleTy) {
  default:
    llvm_unreachable("Custom UINT_TO_FP is not supported!");
  case MVT::v2i32:
    return lowerUINT_TO_FP_v2i32(Op, dl, DAG, Subtarget);
  case MVT::v4i32:
  case MVT::v8i32:
    return lowerUINT_TO_FP_vXi32(Op, dl, DAG, Subtarget);
  case MVT::v2i64:
  case MVT::v4i64:
    return lowerINT_TO_FP_vXi64(Op, dl, DAG, Subtarget);
  }
}
 
SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
                                           SelectionDAG &DAG) const {
  bool IsStrict = Op->isStrictFPOpcode();
  unsigned OpNo = IsStrict ? 1 : 0;
  SDValue Src = Op.getOperand(OpNo);
  SDLoc dl(Op);
  auto PtrVT = getPointerTy(DAG.getDataLayout());
  MVT SrcVT = Src.getSimpleValueType();
  MVT DstVT = Op->getSimpleValueType(0);
  SDValue Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode();
 
  // Bail out when we don't have native conversion instructions.
  if (DstVT == MVT::f128)
    return SDValue();
 
  if (isBF16orSoftF16(DstVT, Subtarget))
    return promoteXINT_TO_FP(Op, dl, DAG);
  else if (isLegalConversion(SrcVT, DstVT, false, Subtarget))
    return Op;
 
  if (SDValue V = lowerFPToIntToFP(Op, dl, DAG, Subtarget))
    return V;
 
  if (DstVT.isVector())
    return lowerUINT_TO_FP_vec(Op, dl, DAG, Subtarget);
 
  if (Subtarget.isTargetWin64() && SrcVT == MVT::i128)
    return LowerWin64_INT128_TO_FP(Op, DAG);
 
  if (SDValue Extract = vectorizeExtractedCast(Op, dl, DAG, Subtarget))
    return Extract;
 
  if (Subtarget.hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
      (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget.is64Bit()))) {
    // Conversions from unsigned i32 to f32/f64 are legal,
    // using VCVTUSI2SS/SD.  Same for i64 in 64-bit mode.
    return Op;
  }
 
  // Promote i32 to i64 and use a signed conversion on 64-bit targets.
  if (SrcVT == MVT::i32 && Subtarget.is64Bit()) {
    Src = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Src);
    if (IsStrict)
      return DAG.getNode(ISD::STRICT_SINT_TO_FP, dl, {DstVT, MVT::Other},
                         {Chain, Src});
    return DAG.getNode(ISD::SINT_TO_FP, dl, DstVT, Src);
  }
 
  if (SDValue V = LowerI64IntToFP_AVX512DQ(Op, dl, DAG, Subtarget))
    return V;
  if (SDValue V = LowerI64IntToFP16(Op, dl, DAG, Subtarget))
    return V;
 
  // The transform for i64->f64 isn't correct for 0 when rounding to negative
  // infinity. It produces -0.0, so disable under strictfp.
  if (SrcVT == MVT::i64 && DstVT == MVT::f64 && Subtarget.hasSSE2() &&
      !IsStrict)
    return LowerUINT_TO_FP_i64(Op, dl, DAG, Subtarget);
  // The transform for i32->f64/f32 isn't correct for 0 when rounding to
  // negative infinity. So disable under strictfp. Using FILD instead.
  if (SrcVT == MVT::i32 && Subtarget.hasSSE2() && DstVT != MVT::f80 &&
      !IsStrict)
    return LowerUINT_TO_FP_i32(Op, dl, DAG, Subtarget);
  if (Subtarget.is64Bit() && SrcVT == MVT::i64 &&
      (DstVT == MVT::f32 || DstVT == MVT::f64))
    return SDValue();
 
  // Make a 64-bit buffer, and use it to build an FILD.
  SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64, 8);
  int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
  Align SlotAlign(8);
  MachinePointerInfo MPI =
      MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI);
  if (SrcVT == MVT::i32) {
    SDValue OffsetSlot =
        DAG.getMemBasePlusOffset(StackSlot, TypeSize::getFixed(4), dl);
    SDValue Store1 = DAG.getStore(Chain, dl, Src, StackSlot, MPI, SlotAlign);
    SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
                                  OffsetSlot, MPI.getWithOffset(4), SlotAlign);
    std::pair<SDValue, SDValue> Tmp =
        BuildFILD(DstVT, MVT::i64, dl, Store2, StackSlot, MPI, SlotAlign, DAG);
    if (IsStrict)
      return DAG.getMergeValues({Tmp.first, Tmp.second}, dl);
 
    return Tmp.first;
  }
 
  assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
  SDValue ValueToStore = Src;
  if (isScalarFPTypeInSSEReg(Op.getValueType()) && !Subtarget.is64Bit()) {
    // Bitcasting to f64 here allows us to do a single 64-bit store from
    // an SSE register, avoiding the store forwarding penalty that would come
    // with two 32-bit stores.
    ValueToStore = DAG.getBitcast(MVT::f64, ValueToStore);
  }
  SDValue Store =
      DAG.getStore(Chain, dl, ValueToStore, StackSlot, MPI, SlotAlign);
  // For i64 source, we need to add the appropriate power of 2 if the input
  // was negative. We must be careful to do the computation in x87 extended
  // precision, not in SSE.
  SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
  SDValue Ops[] = {Store, StackSlot};
  SDValue Fild =
      DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, MVT::i64, MPI,
                              SlotAlign, MachineMemOperand::MOLoad);
  Chain = Fild.getValue(1);
 
  // Check whether the sign bit is set.
  SDValue SignSet = DAG.getSetCC(
      dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
      Op.getOperand(OpNo), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
 
  // Build a 64 bit pair (FF, 0) in the constant pool, with FF in the hi bits.
  APInt FF(64, 0x5F80000000000000ULL);
  SDValue FudgePtr =
      DAG.getConstantPool(ConstantInt::get(*DAG.getContext(), FF), PtrVT);
  Align CPAlignment = cast<ConstantPoolSDNode>(FudgePtr)->getAlign();
 
  // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
  SDValue Zero = DAG.getIntPtrConstant(0, dl);
  SDValue Four = DAG.getIntPtrConstant(4, dl);
  SDValue Offset = DAG.getSelect(dl, Zero.getValueType(), SignSet, Four, Zero);
  FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
 
  // Load the value out, extending it from f32 to f80.
  SDValue Fudge = DAG.getExtLoad(
      ISD::EXTLOAD, dl, MVT::f80, Chain, FudgePtr,
      MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
      CPAlignment);
  Chain = Fudge.getValue(1);
  // Extend everything to 80 bits to force it to be done on x87.
  // TODO: Are there any fast-math-flags to propagate here?
  if (IsStrict) {
    unsigned Opc = ISD::STRICT_FADD;
    // Windows needs the precision control changed to 80bits around this add.
    if (Subtarget.isOSWindows() && DstVT == MVT::f32)
      Opc = X86ISD::STRICT_FP80_ADD;
 
    SDValue Add =
        DAG.getNode(Opc, dl, {MVT::f80, MVT::Other}, {Chain, Fild, Fudge});
    // STRICT_FP_ROUND can't handle equal types.
    if (DstVT == MVT::f80)
      return Add;
    return DAG.getNode(ISD::STRICT_FP_ROUND, dl, {DstVT, MVT::Other},
                       {Add.getValue(1), Add,
                        DAG.getIntPtrConstant(0, dl, /*isTarget=*/true)});
  }
  unsigned Opc = ISD::FADD;
  // Windows needs the precision control changed to 80bits around this add.
  if (Subtarget.isOSWindows() && DstVT == MVT::f32)
    Opc = X86ISD::FP80_ADD;
 
  SDValue Add = DAG.getNode(Opc, dl, MVT::f80, Fild, Fudge);
  return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
                     DAG.getIntPtrConstant(0, dl, /*isTarget=*/true));
}
 
// If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
// is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
// just return an SDValue().
// Otherwise it is assumed to be a conversion from one of f32, f64 or f80
// to i16, i32 or i64, and we lower it to a legal sequence and return the
// result.
SDValue X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
                                           bool IsSigned,
                                           SDValue &Chain) const {
  bool IsStrict = Op->isStrictFPOpcode();
  SDLoc DL(Op);
 
  EVT DstTy = Op.getValueType();
  SDValue Value = Op.getOperand(IsStrict ? 1 : 0);
  EVT TheVT = Value.getValueType();
  auto PtrVT = getPointerTy(DAG.getDataLayout());
 
  if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
    // f16 must be promoted before using the lowering in this routine.
    // fp128 does not use this lowering.
    return SDValue();
  }
 
  // If using FIST to compute an unsigned i64, we'll need some fixup
  // to handle values above the maximum signed i64.  A FIST is always
  // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
  bool UnsignedFixup = !IsSigned && DstTy == MVT::i64;
 
  // FIXME: This does not generate an invalid exception if the input does not
  // fit in i32. PR44019
  if (!IsSigned && DstTy != MVT::i64) {
    // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
    // The low 32 bits of the fist result will have the correct uint32 result.
    assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
    DstTy = MVT::i64;
  }
 
  assert(DstTy.getSimpleVT() <= MVT::i64 &&
         DstTy.getSimpleVT() >= MVT::i16 &&
         "Unknown FP_TO_INT to lower!");
 
  // We lower FP->int64 into FISTP64 followed by a load from a temporary
  // stack slot.
  MachineFunction &MF = DAG.getMachineFunction();
  unsigned MemSize = DstTy.getStoreSize();
  int SSFI =
      MF.getFrameInfo().CreateStackObject(MemSize, Align(MemSize), false);
  SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
 
  Chain = IsStrict ? Op.getOperand(0) : DAG.getEntryNode();
 
  SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
 
  if (UnsignedFixup) {
    //
    // Conversion to unsigned i64 is implemented with a select,
    // depending on whether the source value fits in the range
    // of a signed i64.  Let Thresh be the FP equivalent of
    // 0x8000000000000000ULL.
    //
    //  Adjust = (Value >= Thresh) ? 0x80000000 : 0;
    //  FltOfs = (Value >= Thresh) ? 0x80000000 : 0;
    //  FistSrc = (Value - FltOfs);
    //  Fist-to-mem64 FistSrc
    //  Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
    //  to XOR'ing the high 32 bits with Adjust.
    //
    // Being a power of 2, Thresh is exactly representable in all FP formats.
    // For X87 we'd like to use the smallest FP type for this constant, but
    // for DAG type consistency we have to match the FP operand type.
 
    APFloat Thresh(APFloat::IEEEsingle(), APInt(32, 0x5f000000));
    [[maybe_unused]] APFloat::opStatus Status = APFloat::opOK;
    bool LosesInfo = false;
    if (TheVT == MVT::f64)
      // The rounding mode is irrelevant as the conversion should be exact.
      Status = Thresh.convert(APFloat::IEEEdouble(), APFloat::rmNearestTiesToEven,
                              &LosesInfo);
    else if (TheVT == MVT::f80)
      Status = Thresh.convert(APFloat::x87DoubleExtended(),
                              APFloat::rmNearestTiesToEven, &LosesInfo);
 
    assert(Status == APFloat::opOK && !LosesInfo &&
           "FP conversion should have been exact");
 
    SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
 
    EVT ResVT = getSetCCResultType(DAG.getDataLayout(),
                                   *DAG.getContext(), TheVT);
    SDValue Cmp;
    if (IsStrict) {
      Cmp = DAG.getSetCC(DL, ResVT, Value, ThreshVal, ISD::SETGE, Chain,
                         /*IsSignaling*/ true);
      Chain = Cmp.getValue(1);
    } else {
      Cmp = DAG.getSetCC(DL, ResVT, Value, ThreshVal, ISD::SETGE);
    }
 
    // Our preferred lowering of
    //
    // (Value >= Thresh) ? 0x8000000000000000ULL : 0
    //
    // is
    //
    // (Value >= Thresh) << 63
    //
    // but since we can get here after LegalOperations, DAGCombine might do the
    // wrong thing if we create a select. So, directly create the preferred
    // version.
    SDValue Zext = DAG.getNode(ISD::ZERO_EXTEND,