xref: /llvm-project-15.0.7/llvm/lib/Target/X86/X86FastISel.cpp (revision cfa1376a)

//===-- X86FastISel.cpp - X86 FastISel 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 X86-specific support for the FastISel class. Much
// of the target-specific code is generated by tablegen in the file
// X86GenFastISel.inc, which is #included here.
//
//===----------------------------------------------------------------------===//

#include "X86.h"
#include "X86CallingConv.h"
#include "X86InstrBuilder.h"
#include "X86InstrInfo.h"
#include "X86MachineFunctionInfo.h"
#include "X86RegisterInfo.h"
#include "X86Subtarget.h"
#include "X86TargetMachine.h"
#include "llvm/Analysis/BranchProbabilityInfo.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/DebugInfo.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/GlobalAlias.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/IntrinsicsX86.h"
#include "llvm/IR/Operator.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCSymbol.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;

namespace {

class X86FastISel final : public FastISel {
  /// Subtarget - Keep a pointer to the X86Subtarget around so that we can
  /// make the right decision when generating code for different targets.
  const X86Subtarget *Subtarget;

  /// X86ScalarSSEf32, X86ScalarSSEf64 - Select between SSE or x87
  /// floating point ops.
  /// When SSE is available, use it for f32 operations.
  /// When SSE2 is available, use it for f64 operations.
  bool X86ScalarSSEf64;
  bool X86ScalarSSEf32;

public:
  explicit X86FastISel(FunctionLoweringInfo &funcInfo,
                       const TargetLibraryInfo *libInfo)
      : FastISel(funcInfo, libInfo) {
    Subtarget = &funcInfo.MF->getSubtarget<X86Subtarget>();
    X86ScalarSSEf64 = Subtarget->hasSSE2();
    X86ScalarSSEf32 = Subtarget->hasSSE1();
  }

  bool fastSelectInstruction(const Instruction *I) override;

  /// The specified machine instr operand is a vreg, and that
  /// vreg is being provided by the specified load instruction.  If possible,
  /// try to fold the load as an operand to the instruction, returning true if
  /// possible.
  bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
                           const LoadInst *LI) override;

  bool fastLowerArguments() override;
  bool fastLowerCall(CallLoweringInfo &CLI) override;
  bool fastLowerIntrinsicCall(const IntrinsicInst *II) override;

#include "X86GenFastISel.inc"

private:
  bool X86FastEmitCompare(const Value *LHS, const Value *RHS, EVT VT,
                          const DebugLoc &DL);

  bool X86FastEmitLoad(MVT VT, X86AddressMode &AM, MachineMemOperand *MMO,
                       unsigned &ResultReg, unsigned Alignment = 1);

  bool X86FastEmitStore(EVT VT, const Value *Val, X86AddressMode &AM,
                        MachineMemOperand *MMO = nullptr, bool Aligned = false);
  bool X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
                        X86AddressMode &AM,
                        MachineMemOperand *MMO = nullptr, bool Aligned = false);

  bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
                         unsigned &ResultReg);

  bool X86SelectAddress(const Value *V, X86AddressMode &AM);
  bool X86SelectCallAddress(const Value *V, X86AddressMode &AM);

  bool X86SelectLoad(const Instruction *I);

  bool X86SelectStore(const Instruction *I);

  bool X86SelectRet(const Instruction *I);

  bool X86SelectCmp(const Instruction *I);

  bool X86SelectZExt(const Instruction *I);

  bool X86SelectSExt(const Instruction *I);

  bool X86SelectBranch(const Instruction *I);

  bool X86SelectShift(const Instruction *I);

  bool X86SelectDivRem(const Instruction *I);

  bool X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I);

  bool X86FastEmitSSESelect(MVT RetVT, const Instruction *I);

  bool X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I);

  bool X86SelectSelect(const Instruction *I);

  bool X86SelectTrunc(const Instruction *I);

  bool X86SelectFPExtOrFPTrunc(const Instruction *I, unsigned Opc,
                               const TargetRegisterClass *RC);

  bool X86SelectFPExt(const Instruction *I);
  bool X86SelectFPTrunc(const Instruction *I);
  bool X86SelectSIToFP(const Instruction *I);
  bool X86SelectUIToFP(const Instruction *I);
  bool X86SelectIntToFP(const Instruction *I, bool IsSigned);

  const X86InstrInfo *getInstrInfo() const {
    return Subtarget->getInstrInfo();
  }
  const X86TargetMachine *getTargetMachine() const {
    return static_cast<const X86TargetMachine *>(&TM);
  }

  bool handleConstantAddresses(const Value *V, X86AddressMode &AM);

  unsigned X86MaterializeInt(const ConstantInt *CI, MVT VT);
  unsigned X86MaterializeFP(const ConstantFP *CFP, MVT VT);
  unsigned X86MaterializeGV(const GlobalValue *GV, MVT VT);
  unsigned fastMaterializeConstant(const Constant *C) override;

  unsigned fastMaterializeAlloca(const AllocaInst *C) override;

  unsigned fastMaterializeFloatZero(const ConstantFP *CF) override;

  /// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
  /// computed in an SSE register, not on the X87 floating point stack.
  bool isScalarFPTypeInSSEReg(EVT VT) const {
    return (VT == MVT::f64 && X86ScalarSSEf64) || // f64 is when SSE2
      (VT == MVT::f32 && X86ScalarSSEf32);   // f32 is when SSE1
  }

  bool isTypeLegal(Type *Ty, MVT &VT, bool AllowI1 = false);

  bool IsMemcpySmall(uint64_t Len);

  bool TryEmitSmallMemcpy(X86AddressMode DestAM,
                          X86AddressMode SrcAM, uint64_t Len);

  bool foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I,
                            const Value *Cond);

  const MachineInstrBuilder &addFullAddress(const MachineInstrBuilder &MIB,
                                            X86AddressMode &AM);

  unsigned fastEmitInst_rrrr(unsigned MachineInstOpcode,
                             const TargetRegisterClass *RC, unsigned Op0,
                             bool Op0IsKill, unsigned Op1, bool Op1IsKill,
                             unsigned Op2, bool Op2IsKill, unsigned Op3,
                             bool Op3IsKill);
};

} // end anonymous namespace.

static std::pair<unsigned, bool>
getX86SSEConditionCode(CmpInst::Predicate Predicate) {
  unsigned CC;
  bool NeedSwap = false;

  // SSE Condition code mapping:
  //  0 - EQ
  //  1 - LT
  //  2 - LE
  //  3 - UNORD
  //  4 - NEQ
  //  5 - NLT
  //  6 - NLE
  //  7 - ORD
  switch (Predicate) {
  default: llvm_unreachable("Unexpected predicate");
  case CmpInst::FCMP_OEQ: CC = 0;          break;
  case CmpInst::FCMP_OGT: NeedSwap = true; LLVM_FALLTHROUGH;
  case CmpInst::FCMP_OLT: CC = 1;          break;
  case CmpInst::FCMP_OGE: NeedSwap = true; LLVM_FALLTHROUGH;
  case CmpInst::FCMP_OLE: CC = 2;          break;
  case CmpInst::FCMP_UNO: CC = 3;          break;
  case CmpInst::FCMP_UNE: CC = 4;          break;
  case CmpInst::FCMP_ULE: NeedSwap = true; LLVM_FALLTHROUGH;
  case CmpInst::FCMP_UGE: CC = 5;          break;
  case CmpInst::FCMP_ULT: NeedSwap = true; LLVM_FALLTHROUGH;
  case CmpInst::FCMP_UGT: CC = 6;          break;
  case CmpInst::FCMP_ORD: CC = 7;          break;
  case CmpInst::FCMP_UEQ: CC = 8;          break;
  case CmpInst::FCMP_ONE: CC = 12;         break;
  }

  return std::make_pair(CC, NeedSwap);
}

/// Adds a complex addressing mode to the given machine instr builder.
/// Note, this will constrain the index register.  If its not possible to
/// constrain the given index register, then a new one will be created.  The
/// IndexReg field of the addressing mode will be updated to match in this case.
const MachineInstrBuilder &
X86FastISel::addFullAddress(const MachineInstrBuilder &MIB,
                            X86AddressMode &AM) {
  // First constrain the index register.  It needs to be a GR64_NOSP.
  AM.IndexReg = constrainOperandRegClass(MIB->getDesc(), AM.IndexReg,
                                         MIB->getNumOperands() +
                                         X86::AddrIndexReg);
  return ::addFullAddress(MIB, AM);
}

/// Check if it is possible to fold the condition from the XALU intrinsic
/// into the user. The condition code will only be updated on success.
bool X86FastISel::foldX86XALUIntrinsic(X86::CondCode &CC, const Instruction *I,
                                       const Value *Cond) {
  if (!isa<ExtractValueInst>(Cond))
    return false;

  const auto *EV = cast<ExtractValueInst>(Cond);
  if (!isa<IntrinsicInst>(EV->getAggregateOperand()))
    return false;

  const auto *II = cast<IntrinsicInst>(EV->getAggregateOperand());
  MVT RetVT;
  const Function *Callee = II->getCalledFunction();
  Type *RetTy =
    cast<StructType>(Callee->getReturnType())->getTypeAtIndex(0U);
  if (!isTypeLegal(RetTy, RetVT))
    return false;

  if (RetVT != MVT::i32 && RetVT != MVT::i64)
    return false;

  X86::CondCode TmpCC;
  switch (II->getIntrinsicID()) {
  default: return false;
  case Intrinsic::sadd_with_overflow:
  case Intrinsic::ssub_with_overflow:
  case Intrinsic::smul_with_overflow:
  case Intrinsic::umul_with_overflow: TmpCC = X86::COND_O; break;
  case Intrinsic::uadd_with_overflow:
  case Intrinsic::usub_with_overflow: TmpCC = X86::COND_B; break;
  }

  // Check if both instructions are in the same basic block.
  if (II->getParent() != I->getParent())
    return false;

  // Make sure nothing is in the way
  BasicBlock::const_iterator Start(I);
  BasicBlock::const_iterator End(II);
  for (auto Itr = std::prev(Start); Itr != End; --Itr) {
    // We only expect extractvalue instructions between the intrinsic and the
    // instruction to be selected.
    if (!isa<ExtractValueInst>(Itr))
      return false;

    // Check that the extractvalue operand comes from the intrinsic.
    const auto *EVI = cast<ExtractValueInst>(Itr);
    if (EVI->getAggregateOperand() != II)
      return false;
  }

  CC = TmpCC;
  return true;
}

bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) {
  EVT evt = TLI.getValueType(DL, Ty, /*AllowUnknown=*/true);
  if (evt == MVT::Other || !evt.isSimple())
    // Unhandled type. Halt "fast" selection and bail.
    return false;

  VT = evt.getSimpleVT();
  // For now, require SSE/SSE2 for performing floating-point operations,
  // since x87 requires additional work.
  if (VT == MVT::f64 && !X86ScalarSSEf64)
    return false;
  if (VT == MVT::f32 && !X86ScalarSSEf32)
    return false;
  // Similarly, no f80 support yet.
  if (VT == MVT::f80)
    return false;
  // We only handle legal types. For example, on x86-32 the instruction
  // selector contains all of the 64-bit instructions from x86-64,
  // under the assumption that i64 won't be used if the target doesn't
  // support it.
  return (AllowI1 && VT == MVT::i1) || TLI.isTypeLegal(VT);
}

/// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
/// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
/// Return true and the result register by reference if it is possible.
bool X86FastISel::X86FastEmitLoad(MVT VT, X86AddressMode &AM,
                                  MachineMemOperand *MMO, unsigned &ResultReg,
                                  unsigned Alignment) {
  bool HasSSE41 = Subtarget->hasSSE41();
  bool HasAVX = Subtarget->hasAVX();
  bool HasAVX2 = Subtarget->hasAVX2();
  bool HasAVX512 = Subtarget->hasAVX512();
  bool HasVLX = Subtarget->hasVLX();
  bool IsNonTemporal = MMO && MMO->isNonTemporal();

  // Treat i1 loads the same as i8 loads. Masking will be done when storing.
  if (VT == MVT::i1)
    VT = MVT::i8;

  // Get opcode and regclass of the output for the given load instruction.
  unsigned Opc = 0;
  switch (VT.SimpleTy) {
  default: return false;
  case MVT::i8:
    Opc = X86::MOV8rm;
    break;
  case MVT::i16:
    Opc = X86::MOV16rm;
    break;
  case MVT::i32:
    Opc = X86::MOV32rm;
    break;
  case MVT::i64:
    // Must be in x86-64 mode.
    Opc = X86::MOV64rm;
    break;
  case MVT::f32:
    if (X86ScalarSSEf32)
      Opc = HasAVX512 ? X86::VMOVSSZrm_alt :
            HasAVX    ? X86::VMOVSSrm_alt :
                        X86::MOVSSrm_alt;
    else
      Opc = X86::LD_Fp32m;
    break;
  case MVT::f64:
    if (X86ScalarSSEf64)
      Opc = HasAVX512 ? X86::VMOVSDZrm_alt :
            HasAVX    ? X86::VMOVSDrm_alt :
                        X86::MOVSDrm_alt;
    else
      Opc = X86::LD_Fp64m;
    break;
  case MVT::f80:
    // No f80 support yet.
    return false;
  case MVT::v4f32:
    if (IsNonTemporal && Alignment >= 16 && HasSSE41)
      Opc = HasVLX ? X86::VMOVNTDQAZ128rm :
            HasAVX ? X86::VMOVNTDQArm : X86::MOVNTDQArm;
    else if (Alignment >= 16)
      Opc = HasVLX ? X86::VMOVAPSZ128rm :
            HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm;
    else
      Opc = HasVLX ? X86::VMOVUPSZ128rm :
            HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm;
    break;
  case MVT::v2f64:
    if (IsNonTemporal && Alignment >= 16 && HasSSE41)
      Opc = HasVLX ? X86::VMOVNTDQAZ128rm :
            HasAVX ? X86::VMOVNTDQArm : X86::MOVNTDQArm;
    else if (Alignment >= 16)
      Opc = HasVLX ? X86::VMOVAPDZ128rm :
            HasAVX ? X86::VMOVAPDrm : X86::MOVAPDrm;
    else
      Opc = HasVLX ? X86::VMOVUPDZ128rm :
            HasAVX ? X86::VMOVUPDrm : X86::MOVUPDrm;
    break;
  case MVT::v4i32:
  case MVT::v2i64:
  case MVT::v8i16:
  case MVT::v16i8:
    if (IsNonTemporal && Alignment >= 16 && HasSSE41)
      Opc = HasVLX ? X86::VMOVNTDQAZ128rm :
            HasAVX ? X86::VMOVNTDQArm : X86::MOVNTDQArm;
    else if (Alignment >= 16)
      Opc = HasVLX ? X86::VMOVDQA64Z128rm :
            HasAVX ? X86::VMOVDQArm : X86::MOVDQArm;
    else
      Opc = HasVLX ? X86::VMOVDQU64Z128rm :
            HasAVX ? X86::VMOVDQUrm : X86::MOVDQUrm;
    break;
  case MVT::v8f32:
    assert(HasAVX);
    if (IsNonTemporal && Alignment >= 32 && HasAVX2)
      Opc = HasVLX ? X86::VMOVNTDQAZ256rm : X86::VMOVNTDQAYrm;
    else if (IsNonTemporal && Alignment >= 16)
      return false; // Force split for X86::VMOVNTDQArm
    else if (Alignment >= 32)
      Opc = HasVLX ? X86::VMOVAPSZ256rm : X86::VMOVAPSYrm;
    else
      Opc = HasVLX ? X86::VMOVUPSZ256rm : X86::VMOVUPSYrm;
    break;
  case MVT::v4f64:
    assert(HasAVX);
    if (IsNonTemporal && Alignment >= 32 && HasAVX2)
      Opc = HasVLX ? X86::VMOVNTDQAZ256rm : X86::VMOVNTDQAYrm;
    else if (IsNonTemporal && Alignment >= 16)
      return false; // Force split for X86::VMOVNTDQArm
    else if (Alignment >= 32)
      Opc = HasVLX ? X86::VMOVAPDZ256rm : X86::VMOVAPDYrm;
    else
      Opc = HasVLX ? X86::VMOVUPDZ256rm : X86::VMOVUPDYrm;
    break;
  case MVT::v8i32:
  case MVT::v4i64:
  case MVT::v16i16:
  case MVT::v32i8:
    assert(HasAVX);
    if (IsNonTemporal && Alignment >= 32 && HasAVX2)
      Opc = HasVLX ? X86::VMOVNTDQAZ256rm : X86::VMOVNTDQAYrm;
    else if (IsNonTemporal && Alignment >= 16)
      return false; // Force split for X86::VMOVNTDQArm
    else if (Alignment >= 32)
      Opc = HasVLX ? X86::VMOVDQA64Z256rm : X86::VMOVDQAYrm;
    else
      Opc = HasVLX ? X86::VMOVDQU64Z256rm : X86::VMOVDQUYrm;
    break;
  case MVT::v16f32:
    assert(HasAVX512);
    if (IsNonTemporal && Alignment >= 64)
      Opc = X86::VMOVNTDQAZrm;
    else
      Opc = (Alignment >= 64) ? X86::VMOVAPSZrm : X86::VMOVUPSZrm;
    break;
  case MVT::v8f64:
    assert(HasAVX512);
    if (IsNonTemporal && Alignment >= 64)
      Opc = X86::VMOVNTDQAZrm;
    else
      Opc = (Alignment >= 64) ? X86::VMOVAPDZrm : X86::VMOVUPDZrm;
    break;
  case MVT::v8i64:
  case MVT::v16i32:
  case MVT::v32i16:
  case MVT::v64i8:
    assert(HasAVX512);
    // Note: There are a lot more choices based on type with AVX-512, but
    // there's really no advantage when the load isn't masked.
    if (IsNonTemporal && Alignment >= 64)
      Opc = X86::VMOVNTDQAZrm;
    else
      Opc = (Alignment >= 64) ? X86::VMOVDQA64Zrm : X86::VMOVDQU64Zrm;
    break;
  }

  const TargetRegisterClass *RC = TLI.getRegClassFor(VT);

  ResultReg = createResultReg(RC);
  MachineInstrBuilder MIB =
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg);
  addFullAddress(MIB, AM);
  if (MMO)
    MIB->addMemOperand(*FuncInfo.MF, MMO);
  return true;
}

/// X86FastEmitStore - Emit a machine instruction to store a value Val of
/// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
/// and a displacement offset, or a GlobalAddress,
/// i.e. V. Return true if it is possible.
bool X86FastISel::X86FastEmitStore(EVT VT, unsigned ValReg, bool ValIsKill,
                                   X86AddressMode &AM,
                                   MachineMemOperand *MMO, bool Aligned) {
  bool HasSSE1 = Subtarget->hasSSE1();
  bool HasSSE2 = Subtarget->hasSSE2();
  bool HasSSE4A = Subtarget->hasSSE4A();
  bool HasAVX = Subtarget->hasAVX();
  bool HasAVX512 = Subtarget->hasAVX512();
  bool HasVLX = Subtarget->hasVLX();
  bool IsNonTemporal = MMO && MMO->isNonTemporal();

  // Get opcode and regclass of the output for the given store instruction.
  unsigned Opc = 0;
  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::f80: // No f80 support yet.
  default: return false;
  case MVT::i1: {
    // Mask out all but lowest bit.
    Register AndResult = createResultReg(&X86::GR8RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(X86::AND8ri), AndResult)
      .addReg(ValReg, getKillRegState(ValIsKill)).addImm(1);
    ValReg = AndResult;
    LLVM_FALLTHROUGH; // handle i1 as i8.
  }
  case MVT::i8:  Opc = X86::MOV8mr;  break;
  case MVT::i16: Opc = X86::MOV16mr; break;
  case MVT::i32:
    Opc = (IsNonTemporal && HasSSE2) ? X86::MOVNTImr : X86::MOV32mr;
    break;
  case MVT::i64:
    // Must be in x86-64 mode.
    Opc = (IsNonTemporal && HasSSE2) ? X86::MOVNTI_64mr : X86::MOV64mr;
    break;
  case MVT::f32:
    if (X86ScalarSSEf32) {
      if (IsNonTemporal && HasSSE4A)
        Opc = X86::MOVNTSS;
      else
        Opc = HasAVX512 ? X86::VMOVSSZmr :
              HasAVX ? X86::VMOVSSmr : X86::MOVSSmr;
    } else
      Opc = X86::ST_Fp32m;
    break;
  case MVT::f64:
    if (X86ScalarSSEf32) {
      if (IsNonTemporal && HasSSE4A)
        Opc = X86::MOVNTSD;
      else
        Opc = HasAVX512 ? X86::VMOVSDZmr :
              HasAVX ? X86::VMOVSDmr : X86::MOVSDmr;
    } else
      Opc = X86::ST_Fp64m;
    break;
  case MVT::x86mmx:
    Opc = (IsNonTemporal && HasSSE1) ? X86::MMX_MOVNTQmr : X86::MMX_MOVQ64mr;
    break;
  case MVT::v4f32:
    if (Aligned) {
      if (IsNonTemporal)
        Opc = HasVLX ? X86::VMOVNTPSZ128mr :
              HasAVX ? X86::VMOVNTPSmr : X86::MOVNTPSmr;
      else
        Opc = HasVLX ? X86::VMOVAPSZ128mr :
              HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr;
    } else
      Opc = HasVLX ? X86::VMOVUPSZ128mr :
            HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr;
    break;
  case MVT::v2f64:
    if (Aligned) {
      if (IsNonTemporal)
        Opc = HasVLX ? X86::VMOVNTPDZ128mr :
              HasAVX ? X86::VMOVNTPDmr : X86::MOVNTPDmr;
      else
        Opc = HasVLX ? X86::VMOVAPDZ128mr :
              HasAVX ? X86::VMOVAPDmr : X86::MOVAPDmr;
    } else
      Opc = HasVLX ? X86::VMOVUPDZ128mr :
            HasAVX ? X86::VMOVUPDmr : X86::MOVUPDmr;
    break;
  case MVT::v4i32:
  case MVT::v2i64:
  case MVT::v8i16:
  case MVT::v16i8:
    if (Aligned) {
      if (IsNonTemporal)
        Opc = HasVLX ? X86::VMOVNTDQZ128mr :
              HasAVX ? X86::VMOVNTDQmr : X86::MOVNTDQmr;
      else
        Opc = HasVLX ? X86::VMOVDQA64Z128mr :
              HasAVX ? X86::VMOVDQAmr : X86::MOVDQAmr;
    } else
      Opc = HasVLX ? X86::VMOVDQU64Z128mr :
            HasAVX ? X86::VMOVDQUmr : X86::MOVDQUmr;
    break;
  case MVT::v8f32:
    assert(HasAVX);
    if (Aligned) {
      if (IsNonTemporal)
        Opc = HasVLX ? X86::VMOVNTPSZ256mr : X86::VMOVNTPSYmr;
      else
        Opc = HasVLX ? X86::VMOVAPSZ256mr : X86::VMOVAPSYmr;
    } else
      Opc = HasVLX ? X86::VMOVUPSZ256mr : X86::VMOVUPSYmr;
    break;
  case MVT::v4f64:
    assert(HasAVX);
    if (Aligned) {
      if (IsNonTemporal)
        Opc = HasVLX ? X86::VMOVNTPDZ256mr : X86::VMOVNTPDYmr;
      else
        Opc = HasVLX ? X86::VMOVAPDZ256mr : X86::VMOVAPDYmr;
    } else
      Opc = HasVLX ? X86::VMOVUPDZ256mr : X86::VMOVUPDYmr;
    break;
  case MVT::v8i32:
  case MVT::v4i64:
  case MVT::v16i16:
  case MVT::v32i8:
    assert(HasAVX);
    if (Aligned) {
      if (IsNonTemporal)
        Opc = HasVLX ? X86::VMOVNTDQZ256mr : X86::VMOVNTDQYmr;
      else
        Opc = HasVLX ? X86::VMOVDQA64Z256mr : X86::VMOVDQAYmr;
    } else
      Opc = HasVLX ? X86::VMOVDQU64Z256mr : X86::VMOVDQUYmr;
    break;
  case MVT::v16f32:
    assert(HasAVX512);
    if (Aligned)
      Opc = IsNonTemporal ? X86::VMOVNTPSZmr : X86::VMOVAPSZmr;
    else
      Opc = X86::VMOVUPSZmr;
    break;
  case MVT::v8f64:
    assert(HasAVX512);
    if (Aligned) {
      Opc = IsNonTemporal ? X86::VMOVNTPDZmr : X86::VMOVAPDZmr;
    } else
      Opc = X86::VMOVUPDZmr;
    break;
  case MVT::v8i64:
  case MVT::v16i32:
  case MVT::v32i16:
  case MVT::v64i8:
    assert(HasAVX512);
    // Note: There are a lot more choices based on type with AVX-512, but
    // there's really no advantage when the store isn't masked.
    if (Aligned)
      Opc = IsNonTemporal ? X86::VMOVNTDQZmr : X86::VMOVDQA64Zmr;
    else
      Opc = X86::VMOVDQU64Zmr;
    break;
  }

  const MCInstrDesc &Desc = TII.get(Opc);
  // Some of the instructions in the previous switch use FR128 instead
  // of FR32 for ValReg. Make sure the register we feed the instruction
  // matches its register class constraints.
  // Note: This is fine to do a copy from FR32 to FR128, this is the
  // same registers behind the scene and actually why it did not trigger
  // any bugs before.
  ValReg = constrainOperandRegClass(Desc, ValReg, Desc.getNumOperands() - 1);
  MachineInstrBuilder MIB =
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, Desc);
  addFullAddress(MIB, AM).addReg(ValReg, getKillRegState(ValIsKill));
  if (MMO)
    MIB->addMemOperand(*FuncInfo.MF, MMO);

  return true;
}

bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
                                   X86AddressMode &AM,
                                   MachineMemOperand *MMO, bool Aligned) {
  // Handle 'null' like i32/i64 0.
  if (isa<ConstantPointerNull>(Val))
    Val = Constant::getNullValue(DL.getIntPtrType(Val->getContext()));

  // If this is a store of a simple constant, fold the constant into the store.
  if (const ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
    unsigned Opc = 0;
    bool Signed = true;
    switch (VT.getSimpleVT().SimpleTy) {
    default: break;
    case MVT::i1:
      Signed = false;
      LLVM_FALLTHROUGH; // Handle as i8.
    case MVT::i8:  Opc = X86::MOV8mi;  break;
    case MVT::i16: Opc = X86::MOV16mi; break;
    case MVT::i32: Opc = X86::MOV32mi; break;
    case MVT::i64:
      // Must be a 32-bit sign extended value.
      if (isInt<32>(CI->getSExtValue()))
        Opc = X86::MOV64mi32;
      break;
    }

    if (Opc) {
      MachineInstrBuilder MIB =
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc));
      addFullAddress(MIB, AM).addImm(Signed ? (uint64_t) CI->getSExtValue()
                                            : CI->getZExtValue());
      if (MMO)
        MIB->addMemOperand(*FuncInfo.MF, MMO);
      return true;
    }
  }

  Register ValReg = getRegForValue(Val);
  if (ValReg == 0)
    return false;

  bool ValKill = hasTrivialKill(Val);
  return X86FastEmitStore(VT, ValReg, ValKill, AM, MMO, Aligned);
}

/// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
/// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
/// ISD::SIGN_EXTEND).
bool X86FastISel::X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT,
                                    unsigned Src, EVT SrcVT,
                                    unsigned &ResultReg) {
  unsigned RR = fastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
                           Src, /*TODO: Kill=*/false);
  if (RR == 0)
    return false;

  ResultReg = RR;
  return true;
}

bool X86FastISel::handleConstantAddresses(const Value *V, X86AddressMode &AM) {
  // Handle constant address.
  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
    // Can't handle alternate code models yet.
    if (TM.getCodeModel() != CodeModel::Small)
      return false;

    // Can't handle TLS yet.
    if (GV->isThreadLocal())
      return false;

    // Can't handle !absolute_symbol references yet.
    if (GV->isAbsoluteSymbolRef())
      return false;

    // RIP-relative addresses can't have additional register operands, so if
    // we've already folded stuff into the addressing mode, just force the
    // global value into its own register, which we can use as the basereg.
    if (!Subtarget->isPICStyleRIPRel() ||
        (AM.Base.Reg == 0 && AM.IndexReg == 0)) {
      // Okay, we've committed to selecting this global. Set up the address.
      AM.GV = GV;

      // Allow the subtarget to classify the global.
      unsigned char GVFlags = Subtarget->classifyGlobalReference(GV);

      // If this reference is relative to the pic base, set it now.
      if (isGlobalRelativeToPICBase(GVFlags)) {
        // FIXME: How do we know Base.Reg is free??
        AM.Base.Reg = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
      }

      // Unless the ABI requires an extra load, return a direct reference to
      // the global.
      if (!isGlobalStubReference(GVFlags)) {
        if (Subtarget->isPICStyleRIPRel()) {
          // Use rip-relative addressing if we can.  Above we verified that the
          // base and index registers are unused.
          assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
          AM.Base.Reg = X86::RIP;
        }
        AM.GVOpFlags = GVFlags;
        return true;
      }

      // Ok, we need to do a load from a stub.  If we've already loaded from
      // this stub, reuse the loaded pointer, otherwise emit the load now.
      DenseMap<const Value *, Register>::iterator I = LocalValueMap.find(V);
      Register LoadReg;
      if (I != LocalValueMap.end() && I->second) {
        LoadReg = I->second;
      } else {
        // Issue load from stub.
        unsigned Opc = 0;
        const TargetRegisterClass *RC = nullptr;
        X86AddressMode StubAM;
        StubAM.Base.Reg = AM.Base.Reg;
        StubAM.GV = GV;
        StubAM.GVOpFlags = GVFlags;

        // Prepare for inserting code in the local-value area.
        SavePoint SaveInsertPt = enterLocalValueArea();

        if (TLI.getPointerTy(DL) == MVT::i64) {
          Opc = X86::MOV64rm;
          RC  = &X86::GR64RegClass;

          if (Subtarget->isPICStyleRIPRel())
            StubAM.Base.Reg = X86::RIP;
        } else {
          Opc = X86::MOV32rm;
          RC  = &X86::GR32RegClass;
        }

        LoadReg = createResultReg(RC);
        MachineInstrBuilder LoadMI =
          BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), LoadReg);
        addFullAddress(LoadMI, StubAM);

        // Ok, back to normal mode.
        leaveLocalValueArea(SaveInsertPt);

        // Prevent loading GV stub multiple times in same MBB.
        LocalValueMap[V] = LoadReg;
      }

      // Now construct the final address. Note that the Disp, Scale,
      // and Index values may already be set here.
      AM.Base.Reg = LoadReg;
      AM.GV = nullptr;
      return true;
    }
  }

  // If all else fails, try to materialize the value in a register.
  if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
    if (AM.Base.Reg == 0) {
      AM.Base.Reg = getRegForValue(V);
      return AM.Base.Reg != 0;
    }
    if (AM.IndexReg == 0) {
      assert(AM.Scale == 1 && "Scale with no index!");
      AM.IndexReg = getRegForValue(V);
      return AM.IndexReg != 0;
    }
  }

  return false;
}

/// X86SelectAddress - Attempt to fill in an address from the given value.
///
bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
  SmallVector<const Value *, 32> GEPs;
redo_gep:
  const User *U = nullptr;
  unsigned Opcode = Instruction::UserOp1;
  if (const Instruction *I = dyn_cast<Instruction>(V)) {
    // Don't walk into other basic blocks; it's possible we haven't
    // visited them yet, so the instructions may not yet be assigned
    // virtual registers.
    if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(V)) ||
        FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
      Opcode = I->getOpcode();
      U = I;
    }
  } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
    Opcode = C->getOpcode();
    U = C;
  }

  if (PointerType *Ty = dyn_cast<PointerType>(V->getType()))
    if (Ty->getAddressSpace() > 255)
      // Fast instruction selection doesn't support the special
      // address spaces.
      return false;

  switch (Opcode) {
  default: break;
  case Instruction::BitCast:
    // Look past bitcasts.
    return X86SelectAddress(U->getOperand(0), AM);

  case Instruction::IntToPtr:
    // Look past no-op inttoptrs.
    if (TLI.getValueType(DL, U->getOperand(0)->getType()) ==
        TLI.getPointerTy(DL))
      return X86SelectAddress(U->getOperand(0), AM);
    break;

  case Instruction::PtrToInt:
    // Look past no-op ptrtoints.
    if (TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL))
      return X86SelectAddress(U->getOperand(0), AM);
    break;

  case Instruction::Alloca: {
    // Do static allocas.
    const AllocaInst *A = cast<AllocaInst>(V);
    DenseMap<const AllocaInst *, int>::iterator SI =
      FuncInfo.StaticAllocaMap.find(A);
    if (SI != FuncInfo.StaticAllocaMap.end()) {
      AM.BaseType = X86AddressMode::FrameIndexBase;
      AM.Base.FrameIndex = SI->second;
      return true;
    }
    break;
  }

  case Instruction::Add: {
    // Adds of constants are common and easy enough.
    if (const ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
      uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
      // They have to fit in the 32-bit signed displacement field though.
      if (isInt<32>(Disp)) {
        AM.Disp = (uint32_t)Disp;
        return X86SelectAddress(U->getOperand(0), AM);
      }
    }
    break;
  }

  case Instruction::GetElementPtr: {
    X86AddressMode SavedAM = AM;

    // Pattern-match simple GEPs.
    uint64_t Disp = (int32_t)AM.Disp;
    unsigned IndexReg = AM.IndexReg;
    unsigned Scale = AM.Scale;
    gep_type_iterator GTI = gep_type_begin(U);
    // Iterate through the indices, folding what we can. Constants can be
    // folded, and one dynamic index can be handled, if the scale is supported.
    for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
         i != e; ++i, ++GTI) {
      const Value *Op = *i;
      if (StructType *STy = GTI.getStructTypeOrNull()) {
        const StructLayout *SL = DL.getStructLayout(STy);
        Disp += SL->getElementOffset(cast<ConstantInt>(Op)->getZExtValue());
        continue;
      }

      // A array/variable index is always of the form i*S where S is the
      // constant scale size.  See if we can push the scale into immediates.
      uint64_t S = DL.getTypeAllocSize(GTI.getIndexedType());
      for (;;) {
        if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
          // Constant-offset addressing.
          Disp += CI->getSExtValue() * S;
          break;
        }
        if (canFoldAddIntoGEP(U, Op)) {
          // A compatible add with a constant operand. Fold the constant.
          ConstantInt *CI =
            cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
          Disp += CI->getSExtValue() * S;
          // Iterate on the other operand.
          Op = cast<AddOperator>(Op)->getOperand(0);
          continue;
        }
        if (IndexReg == 0 &&
            (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
            (S == 1 || S == 2 || S == 4 || S == 8)) {
          // Scaled-index addressing.
          Scale = S;
          IndexReg = getRegForGEPIndex(Op).first;
          if (IndexReg == 0)
            return false;
          break;
        }
        // Unsupported.
        goto unsupported_gep;
      }
    }

    // Check for displacement overflow.
    if (!isInt<32>(Disp))
      break;

    AM.IndexReg = IndexReg;
    AM.Scale = Scale;
    AM.Disp = (uint32_t)Disp;
    GEPs.push_back(V);

    if (const GetElementPtrInst *GEP =
          dyn_cast<GetElementPtrInst>(U->getOperand(0))) {
      // Ok, the GEP indices were covered by constant-offset and scaled-index
      // addressing. Update the address state and move on to examining the base.
      V = GEP;
      goto redo_gep;
    } else if (X86SelectAddress(U->getOperand(0), AM)) {
      return true;
    }

    // If we couldn't merge the gep value into this addr mode, revert back to
    // our address and just match the value instead of completely failing.
    AM = SavedAM;

    for (const Value *I : reverse(GEPs))
      if (handleConstantAddresses(I, AM))
        return true;

    return false;
  unsupported_gep:
    // Ok, the GEP indices weren't all covered.
    break;
  }
  }

  return handleConstantAddresses(V, AM);
}

/// X86SelectCallAddress - Attempt to fill in an address from the given value.
///
bool X86FastISel::X86SelectCallAddress(const Value *V, X86AddressMode &AM) {
  const User *U = nullptr;
  unsigned Opcode = Instruction::UserOp1;
  const Instruction *I = dyn_cast<Instruction>(V);
  // Record if the value is defined in the same basic block.
  //
  // This information is crucial to know whether or not folding an
  // operand is valid.
  // Indeed, FastISel generates or reuses a virtual register for all
  // operands of all instructions it selects. Obviously, the definition and
  // its uses must use the same virtual register otherwise the produced
  // code is incorrect.
  // Before instruction selection, FunctionLoweringInfo::set sets the virtual
  // registers for values that are alive across basic blocks. This ensures
  // that the values are consistently set between across basic block, even
  // if different instruction selection mechanisms are used (e.g., a mix of
  // SDISel and FastISel).
  // For values local to a basic block, the instruction selection process
  // generates these virtual registers with whatever method is appropriate
  // for its needs. In particular, FastISel and SDISel do not share the way
  // local virtual registers are set.
  // Therefore, this is impossible (or at least unsafe) to share values
  // between basic blocks unless they use the same instruction selection
  // method, which is not guarantee for X86.
  // Moreover, things like hasOneUse could not be used accurately, if we
  // allow to reference values across basic blocks whereas they are not
  // alive across basic blocks initially.
  bool InMBB = true;
  if (I) {
    Opcode = I->getOpcode();
    U = I;
    InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
  } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
    Opcode = C->getOpcode();
    U = C;
  }

  switch (Opcode) {
  default: break;
  case Instruction::BitCast:
    // Look past bitcasts if its operand is in the same BB.
    if (InMBB)
      return X86SelectCallAddress(U->getOperand(0), AM);
    break;

  case Instruction::IntToPtr:
    // Look past no-op inttoptrs if its operand is in the same BB.
    if (InMBB &&
        TLI.getValueType(DL, U->getOperand(0)->getType()) ==
            TLI.getPointerTy(DL))
      return X86SelectCallAddress(U->getOperand(0), AM);
    break;

  case Instruction::PtrToInt:
    // Look past no-op ptrtoints if its operand is in the same BB.
    if (InMBB && TLI.getValueType(DL, U->getType()) == TLI.getPointerTy(DL))
      return X86SelectCallAddress(U->getOperand(0), AM);
    break;
  }

  // Handle constant address.
  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
    // Can't handle alternate code models yet.
    if (TM.getCodeModel() != CodeModel::Small)
      return false;

    // RIP-relative addresses can't have additional register operands.
    if (Subtarget->isPICStyleRIPRel() &&
        (AM.Base.Reg != 0 || AM.IndexReg != 0))
      return false;

    // Can't handle TLS.
    if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
      if (GVar->isThreadLocal())
        return false;

    // Okay, we've committed to selecting this global. Set up the basic address.
    AM.GV = GV;

    // Return a direct reference to the global. Fastisel can handle calls to
    // functions that require loads, such as dllimport and nonlazybind
    // functions.
    if (Subtarget->isPICStyleRIPRel()) {
      // Use rip-relative addressing if we can.  Above we verified that the
      // base and index registers are unused.
      assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
      AM.Base.Reg = X86::RIP;
    } else {
      AM.GVOpFlags = Subtarget->classifyLocalReference(nullptr);
    }

    return true;
  }

  // If all else fails, try to materialize the value in a register.
  if (!AM.GV || !Subtarget->isPICStyleRIPRel()) {
    if (AM.Base.Reg == 0) {
      AM.Base.Reg = getRegForValue(V);
      return AM.Base.Reg != 0;
    }
    if (AM.IndexReg == 0) {
      assert(AM.Scale == 1 && "Scale with no index!");
      AM.IndexReg = getRegForValue(V);
      return AM.IndexReg != 0;
    }
  }

  return false;
}


/// X86SelectStore - Select and emit code to implement store instructions.
bool X86FastISel::X86SelectStore(const Instruction *I) {
  // Atomic stores need special handling.
  const StoreInst *S = cast<StoreInst>(I);

  if (S->isAtomic())
    return false;

  const Value *PtrV = I->getOperand(1);
  if (TLI.supportSwiftError()) {
    // Swifterror values can come from either a function parameter with
    // swifterror attribute or an alloca with swifterror attribute.
    if (const Argument *Arg = dyn_cast<Argument>(PtrV)) {
      if (Arg->hasSwiftErrorAttr())
        return false;
    }

    if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(PtrV)) {
      if (Alloca->isSwiftError())
        return false;
    }
  }

  const Value *Val = S->getValueOperand();
  const Value *Ptr = S->getPointerOperand();

  MVT VT;
  if (!isTypeLegal(Val->getType(), VT, /*AllowI1=*/true))
    return false;

  Align Alignment = S->getAlign();
  Align ABIAlignment = DL.getABITypeAlign(Val->getType());
  bool Aligned = Alignment >= ABIAlignment;

  X86AddressMode AM;
  if (!X86SelectAddress(Ptr, AM))
    return false;

  return X86FastEmitStore(VT, Val, AM, createMachineMemOperandFor(I), Aligned);
}

/// X86SelectRet - Select and emit code to implement ret instructions.
bool X86FastISel::X86SelectRet(const Instruction *I) {
  const ReturnInst *Ret = cast<ReturnInst>(I);
  const Function &F = *I->getParent()->getParent();
  const X86MachineFunctionInfo *X86MFInfo =
      FuncInfo.MF->getInfo<X86MachineFunctionInfo>();

  if (!FuncInfo.CanLowerReturn)
    return false;

  if (TLI.supportSwiftError() &&
      F.getAttributes().hasAttrSomewhere(Attribute::SwiftError))
    return false;

  if (TLI.supportSplitCSR(FuncInfo.MF))
    return false;

  CallingConv::ID CC = F.getCallingConv();
  if (CC != CallingConv::C &&
      CC != CallingConv::Fast &&
      CC != CallingConv::Tail &&
      CC != CallingConv::X86_FastCall &&
      CC != CallingConv::X86_StdCall &&
      CC != CallingConv::X86_ThisCall &&
      CC != CallingConv::X86_64_SysV &&
      CC != CallingConv::Win64)
    return false;

  // Don't handle popping bytes if they don't fit the ret's immediate.
  if (!isUInt<16>(X86MFInfo->getBytesToPopOnReturn()))
    return false;

  // fastcc with -tailcallopt is intended to provide a guaranteed
  // tail call optimization. Fastisel doesn't know how to do that.
  if ((CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) ||
      CC == CallingConv::Tail)
    return false;

  // Let SDISel handle vararg functions.
  if (F.isVarArg())
    return false;

  // Build a list of return value registers.
  SmallVector<unsigned, 4> RetRegs;

  if (Ret->getNumOperands() > 0) {
    SmallVector<ISD::OutputArg, 4> Outs;
    GetReturnInfo(CC, F.getReturnType(), F.getAttributes(), Outs, TLI, DL);

    // Analyze operands of the call, assigning locations to each operand.
    SmallVector<CCValAssign, 16> ValLocs;
    CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
    CCInfo.AnalyzeReturn(Outs, RetCC_X86);

    const Value *RV = Ret->getOperand(0);
    Register Reg = getRegForValue(RV);
    if (Reg == 0)
      return false;

    // Only handle a single return value for now.
    if (ValLocs.size() != 1)
      return false;

    CCValAssign &VA = ValLocs[0];

    // Don't bother handling odd stuff for now.
    if (VA.getLocInfo() != CCValAssign::Full)
      return false;
    // Only handle register returns for now.
    if (!VA.isRegLoc())
      return false;

    // The calling-convention tables for x87 returns don't tell
    // the whole story.
    if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
      return false;

    unsigned SrcReg = Reg + VA.getValNo();
    EVT SrcVT = TLI.getValueType(DL, RV->getType());
    EVT DstVT = VA.getValVT();
    // Special handling for extended integers.
    if (SrcVT != DstVT) {
      if (SrcVT != MVT::i1 && SrcVT != MVT::i8 && SrcVT != MVT::i16)
        return false;

      if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
        return false;

      assert(DstVT == MVT::i32 && "X86 should always ext to i32");

      if (SrcVT == MVT::i1) {
        if (Outs[0].Flags.isSExt())
          return false;
        // TODO
        SrcReg = fastEmitZExtFromI1(MVT::i8, SrcReg, /*Op0IsKill=*/false);
        SrcVT = MVT::i8;
      }
      unsigned Op = Outs[0].Flags.isZExt() ? ISD::ZERO_EXTEND :
                                             ISD::SIGN_EXTEND;
      // TODO
      SrcReg = fastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Op, SrcReg,
                          /*Op0IsKill=*/false);
    }

    // Make the copy.
    Register DstReg = VA.getLocReg();
    const TargetRegisterClass *SrcRC = MRI.getRegClass(SrcReg);
    // Avoid a cross-class copy. This is very unlikely.
    if (!SrcRC->contains(DstReg))
      return false;
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), DstReg).addReg(SrcReg);

    // Add register to return instruction.
    RetRegs.push_back(VA.getLocReg());
  }

  // Swift calling convention does not require we copy the sret argument
  // into %rax/%eax for the return, and SRetReturnReg is not set for Swift.

  // All x86 ABIs require that for returning structs by value we copy
  // the sret argument into %rax/%eax (depending on ABI) for the return.
  // We saved the argument into a virtual register in the entry block,
  // so now we copy the value out and into %rax/%eax.
  if (F.hasStructRetAttr() && CC != CallingConv::Swift) {
    Register Reg = X86MFInfo->getSRetReturnReg();
    assert(Reg &&
           "SRetReturnReg should have been set in LowerFormalArguments()!");
    unsigned RetReg = Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX;
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), RetReg).addReg(Reg);
    RetRegs.push_back(RetReg);
  }

  // Now emit the RET.
  MachineInstrBuilder MIB;
  if (X86MFInfo->getBytesToPopOnReturn()) {
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                  TII.get(Subtarget->is64Bit() ? X86::RETIQ : X86::RETIL))
              .addImm(X86MFInfo->getBytesToPopOnReturn());
  } else {
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                  TII.get(Subtarget->is64Bit() ? X86::RETQ : X86::RETL));
  }
  for (unsigned i = 0, e = RetRegs.size(); i != e; ++i)
    MIB.addReg(RetRegs[i], RegState::Implicit);
  return true;
}

/// X86SelectLoad - Select and emit code to implement load instructions.
///
bool X86FastISel::X86SelectLoad(const Instruction *I) {
  const LoadInst *LI = cast<LoadInst>(I);

  // Atomic loads need special handling.
  if (LI->isAtomic())
    return false;

  const Value *SV = I->getOperand(0);
  if (TLI.supportSwiftError()) {
    // Swifterror values can come from either a function parameter with
    // swifterror attribute or an alloca with swifterror attribute.
    if (const Argument *Arg = dyn_cast<Argument>(SV)) {
      if (Arg->hasSwiftErrorAttr())
        return false;
    }

    if (const AllocaInst *Alloca = dyn_cast<AllocaInst>(SV)) {
      if (Alloca->isSwiftError())
        return false;
    }
  }

  MVT VT;
  if (!isTypeLegal(LI->getType(), VT, /*AllowI1=*/true))
    return false;

  const Value *Ptr = LI->getPointerOperand();

  X86AddressMode AM;
  if (!X86SelectAddress(Ptr, AM))
    return false;

  unsigned ResultReg = 0;
  if (!X86FastEmitLoad(VT, AM, createMachineMemOperandFor(LI), ResultReg,
                       LI->getAlign().value()))
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
  bool HasAVX512 = Subtarget->hasAVX512();
  bool HasAVX = Subtarget->hasAVX();
  bool X86ScalarSSEf32 = Subtarget->hasSSE1();
  bool X86ScalarSSEf64 = Subtarget->hasSSE2();

  switch (VT.getSimpleVT().SimpleTy) {
  default:       return 0;
  case MVT::i8:  return X86::CMP8rr;
  case MVT::i16: return X86::CMP16rr;
  case MVT::i32: return X86::CMP32rr;
  case MVT::i64: return X86::CMP64rr;
  case MVT::f32:
    return X86ScalarSSEf32
               ? (HasAVX512 ? X86::VUCOMISSZrr
                            : HasAVX ? X86::VUCOMISSrr : X86::UCOMISSrr)
               : 0;
  case MVT::f64:
    return X86ScalarSSEf64
               ? (HasAVX512 ? X86::VUCOMISDZrr
                            : HasAVX ? X86::VUCOMISDrr : X86::UCOMISDrr)
               : 0;
  }
}

/// If we have a comparison with RHS as the RHS  of the comparison, return an
/// opcode that works for the compare (e.g. CMP32ri) otherwise return 0.
static unsigned X86ChooseCmpImmediateOpcode(EVT VT, const ConstantInt *RHSC) {
  int64_t Val = RHSC->getSExtValue();
  switch (VT.getSimpleVT().SimpleTy) {
  // Otherwise, we can't fold the immediate into this comparison.
  default:
    return 0;
  case MVT::i8:
    return X86::CMP8ri;
  case MVT::i16:
    if (isInt<8>(Val))
      return X86::CMP16ri8;
    return X86::CMP16ri;
  case MVT::i32:
    if (isInt<8>(Val))
      return X86::CMP32ri8;
    return X86::CMP32ri;
  case MVT::i64:
    if (isInt<8>(Val))
      return X86::CMP64ri8;
    // 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
    // field.
    if (isInt<32>(Val))
      return X86::CMP64ri32;
    return 0;
  }
}

bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1, EVT VT,
                                     const DebugLoc &CurDbgLoc) {
  Register Op0Reg = getRegForValue(Op0);
  if (Op0Reg == 0) return false;

  // Handle 'null' like i32/i64 0.
  if (isa<ConstantPointerNull>(Op1))
    Op1 = Constant::getNullValue(DL.getIntPtrType(Op0->getContext()));

  // We have two options: compare with register or immediate.  If the RHS of
  // the compare is an immediate that we can fold into this compare, use
  // CMPri, otherwise use CMPrr.
  if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
    if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, CurDbgLoc, TII.get(CompareImmOpc))
        .addReg(Op0Reg)
        .addImm(Op1C->getSExtValue());
      return true;
    }
  }

  unsigned CompareOpc = X86ChooseCmpOpcode(VT, Subtarget);
  if (CompareOpc == 0) return false;

  Register Op1Reg = getRegForValue(Op1);
  if (Op1Reg == 0) return false;
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, CurDbgLoc, TII.get(CompareOpc))
    .addReg(Op0Reg)
    .addReg(Op1Reg);

  return true;
}

bool X86FastISel::X86SelectCmp(const Instruction *I) {
  const CmpInst *CI = cast<CmpInst>(I);

  MVT VT;
  if (!isTypeLegal(I->getOperand(0)->getType(), VT))
    return false;

  // Try to optimize or fold the cmp.
  CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
  unsigned ResultReg = 0;
  switch (Predicate) {
  default: break;
  case CmpInst::FCMP_FALSE: {
    ResultReg = createResultReg(&X86::GR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV32r0),
            ResultReg);
    ResultReg = fastEmitInst_extractsubreg(MVT::i8, ResultReg,
                                           /*Op0IsKill=*/true, X86::sub_8bit);
    if (!ResultReg)
      return false;
    break;
  }
  case CmpInst::FCMP_TRUE: {
    ResultReg = createResultReg(&X86::GR8RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri),
            ResultReg).addImm(1);
    break;
  }
  }

  if (ResultReg) {
    updateValueMap(I, ResultReg);
    return true;
  }

  const Value *LHS = CI->getOperand(0);
  const Value *RHS = CI->getOperand(1);

  // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0.
  // We don't have to materialize a zero constant for this case and can just use
  // %x again on the RHS.
  if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) {
    const auto *RHSC = dyn_cast<ConstantFP>(RHS);
    if (RHSC && RHSC->isNullValue())
      RHS = LHS;
  }

  // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction.
  static const uint16_t SETFOpcTable[2][3] = {
    { X86::COND_E,  X86::COND_NP, X86::AND8rr },
    { X86::COND_NE, X86::COND_P,  X86::OR8rr  }
  };
  const uint16_t *SETFOpc = nullptr;
  switch (Predicate) {
  default: break;
  case CmpInst::FCMP_OEQ: SETFOpc = &SETFOpcTable[0][0]; break;
  case CmpInst::FCMP_UNE: SETFOpc = &SETFOpcTable[1][0]; break;
  }

  ResultReg = createResultReg(&X86::GR8RegClass);
  if (SETFOpc) {
    if (!X86FastEmitCompare(LHS, RHS, VT, I->getDebugLoc()))
      return false;

    Register FlagReg1 = createResultReg(&X86::GR8RegClass);
    Register FlagReg2 = createResultReg(&X86::GR8RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETCCr),
            FlagReg1).addImm(SETFOpc[0]);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETCCr),
            FlagReg2).addImm(SETFOpc[1]);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(SETFOpc[2]),
            ResultReg).addReg(FlagReg1).addReg(FlagReg2);
    updateValueMap(I, ResultReg);
    return true;
  }

  X86::CondCode CC;
  bool SwapArgs;
  std::tie(CC, SwapArgs) = X86::getX86ConditionCode(Predicate);
  assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");

  if (SwapArgs)
    std::swap(LHS, RHS);

  // Emit a compare of LHS/RHS.
  if (!X86FastEmitCompare(LHS, RHS, VT, I->getDebugLoc()))
    return false;

  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETCCr),
          ResultReg).addImm(CC);
  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectZExt(const Instruction *I) {
  EVT DstVT = TLI.getValueType(DL, I->getType());
  if (!TLI.isTypeLegal(DstVT))
    return false;

  Register ResultReg = getRegForValue(I->getOperand(0));
  if (ResultReg == 0)
    return false;

  // Handle zero-extension from i1 to i8, which is common.
  MVT SrcVT = TLI.getSimpleValueType(DL, I->getOperand(0)->getType());
  if (SrcVT == MVT::i1) {
    // Set the high bits to zero.
    ResultReg = fastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
    SrcVT = MVT::i8;

    if (ResultReg == 0)
      return false;
  }

  if (DstVT == MVT::i64) {
    // Handle extension to 64-bits via sub-register shenanigans.
    unsigned MovInst;

    switch (SrcVT.SimpleTy) {
    case MVT::i8:  MovInst = X86::MOVZX32rr8;  break;
    case MVT::i16: MovInst = X86::MOVZX32rr16; break;
    case MVT::i32: MovInst = X86::MOV32rr;     break;
    default: llvm_unreachable("Unexpected zext to i64 source type");
    }

    Register Result32 = createResultReg(&X86::GR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(MovInst), Result32)
      .addReg(ResultReg);

    ResultReg = createResultReg(&X86::GR64RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::SUBREG_TO_REG),
            ResultReg)
      .addImm(0).addReg(Result32).addImm(X86::sub_32bit);
  } else if (DstVT == MVT::i16) {
    // i8->i16 doesn't exist in the autogenerated isel table. Need to zero
    // extend to 32-bits and then extract down to 16-bits.
    Register Result32 = createResultReg(&X86::GR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOVZX32rr8),
            Result32).addReg(ResultReg);

    ResultReg = fastEmitInst_extractsubreg(MVT::i16, Result32,
                                           /*Op0IsKill=*/true, X86::sub_16bit);
  } else if (DstVT != MVT::i8) {
    ResultReg = fastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::ZERO_EXTEND,
                           ResultReg, /*Op0IsKill=*/true);
    if (ResultReg == 0)
      return false;
  }

  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectSExt(const Instruction *I) {
  EVT DstVT = TLI.getValueType(DL, I->getType());
  if (!TLI.isTypeLegal(DstVT))
    return false;

  Register ResultReg = getRegForValue(I->getOperand(0));
  if (ResultReg == 0)
    return false;

  // Handle sign-extension from i1 to i8.
  MVT SrcVT = TLI.getSimpleValueType(DL, I->getOperand(0)->getType());
  if (SrcVT == MVT::i1) {
    // Set the high bits to zero.
    Register ZExtReg = fastEmitZExtFromI1(MVT::i8, ResultReg,
                                          /*TODO: Kill=*/false);
    if (ZExtReg == 0)
      return false;

    // Negate the result to make an 8-bit sign extended value.
    ResultReg = createResultReg(&X86::GR8RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::NEG8r),
            ResultReg).addReg(ZExtReg);

    SrcVT = MVT::i8;
  }

  if (DstVT == MVT::i16) {
    // i8->i16 doesn't exist in the autogenerated isel table. Need to sign
    // extend to 32-bits and then extract down to 16-bits.
    Register Result32 = createResultReg(&X86::GR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOVSX32rr8),
            Result32).addReg(ResultReg);

    ResultReg = fastEmitInst_extractsubreg(MVT::i16, Result32,
                                           /*Op0IsKill=*/true, X86::sub_16bit);
  } else if (DstVT != MVT::i8) {
    ResultReg = fastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::SIGN_EXTEND,
                           ResultReg, /*Op0IsKill=*/true);
    if (ResultReg == 0)
      return false;
  }

  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectBranch(const Instruction *I) {
  // Unconditional branches are selected by tablegen-generated code.
  // Handle a conditional branch.
  const BranchInst *BI = cast<BranchInst>(I);
  MachineBasicBlock *TrueMBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
  MachineBasicBlock *FalseMBB = FuncInfo.MBBMap[BI->getSuccessor(1)];

  // Fold the common case of a conditional branch with a comparison
  // in the same block (values defined on other blocks may not have
  // initialized registers).
  X86::CondCode CC;
  if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
    if (CI->hasOneUse() && CI->getParent() == I->getParent()) {
      EVT VT = TLI.getValueType(DL, CI->getOperand(0)->getType());

      // Try to optimize or fold the cmp.
      CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
      switch (Predicate) {
      default: break;
      case CmpInst::FCMP_FALSE: fastEmitBranch(FalseMBB, DbgLoc); return true;
      case CmpInst::FCMP_TRUE:  fastEmitBranch(TrueMBB, DbgLoc); return true;
      }

      const Value *CmpLHS = CI->getOperand(0);
      const Value *CmpRHS = CI->getOperand(1);

      // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x,
      // 0.0.
      // We don't have to materialize a zero constant for this case and can just
      // use %x again on the RHS.
      if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) {
        const auto *CmpRHSC = dyn_cast<ConstantFP>(CmpRHS);
        if (CmpRHSC && CmpRHSC->isNullValue())
          CmpRHS = CmpLHS;
      }

      // Try to take advantage of fallthrough opportunities.
      if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
        std::swap(TrueMBB, FalseMBB);
        Predicate = CmpInst::getInversePredicate(Predicate);
      }

      // FCMP_OEQ and FCMP_UNE cannot be expressed with a single flag/condition
      // code check. Instead two branch instructions are required to check all
      // the flags. First we change the predicate to a supported condition code,
      // which will be the first branch. Later one we will emit the second
      // branch.
      bool NeedExtraBranch = false;
      switch (Predicate) {
      default: break;
      case CmpInst::FCMP_OEQ:
        std::swap(TrueMBB, FalseMBB);
        LLVM_FALLTHROUGH;
      case CmpInst::FCMP_UNE:
        NeedExtraBranch = true;
        Predicate = CmpInst::FCMP_ONE;
        break;
      }

      bool SwapArgs;
      std::tie(CC, SwapArgs) = X86::getX86ConditionCode(Predicate);
      assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");

      if (SwapArgs)
        std::swap(CmpLHS, CmpRHS);

      // Emit a compare of the LHS and RHS, setting the flags.
      if (!X86FastEmitCompare(CmpLHS, CmpRHS, VT, CI->getDebugLoc()))
        return false;

      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JCC_1))
        .addMBB(TrueMBB).addImm(CC);

      // X86 requires a second branch to handle UNE (and OEQ, which is mapped
      // to UNE above).
      if (NeedExtraBranch) {
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JCC_1))
          .addMBB(TrueMBB).addImm(X86::COND_P);
      }

      finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
      return true;
    }
  } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
    // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which
    // typically happen for _Bool and C++ bools.
    MVT SourceVT;
    if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
        isTypeLegal(TI->getOperand(0)->getType(), SourceVT)) {
      unsigned TestOpc = 0;
      switch (SourceVT.SimpleTy) {
      default: break;
      case MVT::i8:  TestOpc = X86::TEST8ri; break;
      case MVT::i16: TestOpc = X86::TEST16ri; break;
      case MVT::i32: TestOpc = X86::TEST32ri; break;
      case MVT::i64: TestOpc = X86::TEST64ri32; break;
      }
      if (TestOpc) {
        Register OpReg = getRegForValue(TI->getOperand(0));
        if (OpReg == 0) return false;

        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TestOpc))
          .addReg(OpReg).addImm(1);

        unsigned JmpCond = X86::COND_NE;
        if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
          std::swap(TrueMBB, FalseMBB);
          JmpCond = X86::COND_E;
        }

        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JCC_1))
          .addMBB(TrueMBB).addImm(JmpCond);

        finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
        return true;
      }
    }
  } else if (foldX86XALUIntrinsic(CC, BI, BI->getCondition())) {
    // Fake request the condition, otherwise the intrinsic might be completely
    // optimized away.
    Register TmpReg = getRegForValue(BI->getCondition());
    if (TmpReg == 0)
      return false;

    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JCC_1))
      .addMBB(TrueMBB).addImm(CC);
    finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
    return true;
  }

  // Otherwise do a clumsy setcc and re-test it.
  // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used
  // in an explicit cast, so make sure to handle that correctly.
  Register OpReg = getRegForValue(BI->getCondition());
  if (OpReg == 0) return false;

  // In case OpReg is a K register, COPY to a GPR
  if (MRI.getRegClass(OpReg) == &X86::VK1RegClass) {
    unsigned KOpReg = OpReg;
    OpReg = createResultReg(&X86::GR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), OpReg)
        .addReg(KOpReg);
    OpReg = fastEmitInst_extractsubreg(MVT::i8, OpReg, /*Op0IsKill=*/true,
                                       X86::sub_8bit);
  }
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
      .addReg(OpReg)
      .addImm(1);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::JCC_1))
    .addMBB(TrueMBB).addImm(X86::COND_NE);
  finishCondBranch(BI->getParent(), TrueMBB, FalseMBB);
  return true;
}

bool X86FastISel::X86SelectShift(const Instruction *I) {
  unsigned CReg = 0, OpReg = 0;
  const TargetRegisterClass *RC = nullptr;
  if (I->getType()->isIntegerTy(8)) {
    CReg = X86::CL;
    RC = &X86::GR8RegClass;
    switch (I->getOpcode()) {
    case Instruction::LShr: OpReg = X86::SHR8rCL; break;
    case Instruction::AShr: OpReg = X86::SAR8rCL; break;
    case Instruction::Shl:  OpReg = X86::SHL8rCL; break;
    default: return false;
    }
  } else if (I->getType()->isIntegerTy(16)) {
    CReg = X86::CX;
    RC = &X86::GR16RegClass;
    switch (I->getOpcode()) {
    default: llvm_unreachable("Unexpected shift opcode");
    case Instruction::LShr: OpReg = X86::SHR16rCL; break;
    case Instruction::AShr: OpReg = X86::SAR16rCL; break;
    case Instruction::Shl:  OpReg = X86::SHL16rCL; break;
    }
  } else if (I->getType()->isIntegerTy(32)) {
    CReg = X86::ECX;
    RC = &X86::GR32RegClass;
    switch (I->getOpcode()) {
    default: llvm_unreachable("Unexpected shift opcode");
    case Instruction::LShr: OpReg = X86::SHR32rCL; break;
    case Instruction::AShr: OpReg = X86::SAR32rCL; break;
    case Instruction::Shl:  OpReg = X86::SHL32rCL; break;
    }
  } else if (I->getType()->isIntegerTy(64)) {
    CReg = X86::RCX;
    RC = &X86::GR64RegClass;
    switch (I->getOpcode()) {
    default: llvm_unreachable("Unexpected shift opcode");
    case Instruction::LShr: OpReg = X86::SHR64rCL; break;
    case Instruction::AShr: OpReg = X86::SAR64rCL; break;
    case Instruction::Shl:  OpReg = X86::SHL64rCL; break;
    }
  } else {
    return false;
  }

  MVT VT;
  if (!isTypeLegal(I->getType(), VT))
    return false;

  Register Op0Reg = getRegForValue(I->getOperand(0));
  if (Op0Reg == 0) return false;

  Register Op1Reg = getRegForValue(I->getOperand(1));
  if (Op1Reg == 0) return false;
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpcode::COPY),
          CReg).addReg(Op1Reg);

  // The shift instruction uses X86::CL. If we defined a super-register
  // of X86::CL, emit a subreg KILL to precisely describe what we're doing here.
  if (CReg != X86::CL)
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::KILL), X86::CL)
      .addReg(CReg, RegState::Kill);

  Register ResultReg = createResultReg(RC);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(OpReg), ResultReg)
    .addReg(Op0Reg);
  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectDivRem(const Instruction *I) {
  const static unsigned NumTypes = 4; // i8, i16, i32, i64
  const static unsigned NumOps   = 4; // SDiv, SRem, UDiv, URem
  const static bool S = true;  // IsSigned
  const static bool U = false; // !IsSigned
  const static unsigned Copy = TargetOpcode::COPY;
  // For the X86 DIV/IDIV instruction, in most cases the dividend
  // (numerator) must be in a specific register pair highreg:lowreg,
  // producing the quotient in lowreg and the remainder in highreg.
  // For most data types, to set up the instruction, the dividend is
  // copied into lowreg, and lowreg is sign-extended or zero-extended
  // into highreg.  The exception is i8, where the dividend is defined
  // as a single register rather than a register pair, and we
  // therefore directly sign-extend or zero-extend the dividend into
  // lowreg, instead of copying, and ignore the highreg.
  const static struct DivRemEntry {
    // The following portion depends only on the data type.
    const TargetRegisterClass *RC;
    unsigned LowInReg;  // low part of the register pair
    unsigned HighInReg; // high part of the register pair
    // The following portion depends on both the data type and the operation.
    struct DivRemResult {
    unsigned OpDivRem;        // The specific DIV/IDIV opcode to use.
    unsigned OpSignExtend;    // Opcode for sign-extending lowreg into
                              // highreg, or copying a zero into highreg.
    unsigned OpCopy;          // Opcode for copying dividend into lowreg, or
                              // zero/sign-extending into lowreg for i8.
    unsigned DivRemResultReg; // Register containing the desired result.
    bool IsOpSigned;          // Whether to use signed or unsigned form.
    } ResultTable[NumOps];
  } OpTable[NumTypes] = {
    { &X86::GR8RegClass,  X86::AX,  0, {
        { X86::IDIV8r,  0,            X86::MOVSX16rr8, X86::AL,  S }, // SDiv
        { X86::IDIV8r,  0,            X86::MOVSX16rr8, X86::AH,  S }, // SRem
        { X86::DIV8r,   0,            X86::MOVZX16rr8, X86::AL,  U }, // UDiv
        { X86::DIV8r,   0,            X86::MOVZX16rr8, X86::AH,  U }, // URem
      }
    }, // i8
    { &X86::GR16RegClass, X86::AX,  X86::DX, {
        { X86::IDIV16r, X86::CWD,     Copy,            X86::AX,  S }, // SDiv
        { X86::IDIV16r, X86::CWD,     Copy,            X86::DX,  S }, // SRem
        { X86::DIV16r,  X86::MOV32r0, Copy,            X86::AX,  U }, // UDiv
        { X86::DIV16r,  X86::MOV32r0, Copy,            X86::DX,  U }, // URem
      }
    }, // i16
    { &X86::GR32RegClass, X86::EAX, X86::EDX, {
        { X86::IDIV32r, X86::CDQ,     Copy,            X86::EAX, S }, // SDiv
        { X86::IDIV32r, X86::CDQ,     Copy,            X86::EDX, S }, // SRem
        { X86::DIV32r,  X86::MOV32r0, Copy,            X86::EAX, U }, // UDiv
        { X86::DIV32r,  X86::MOV32r0, Copy,            X86::EDX, U }, // URem
      }
    }, // i32
    { &X86::GR64RegClass, X86::RAX, X86::RDX, {
        { X86::IDIV64r, X86::CQO,     Copy,            X86::RAX, S }, // SDiv
        { X86::IDIV64r, X86::CQO,     Copy,            X86::RDX, S }, // SRem
        { X86::DIV64r,  X86::MOV32r0, Copy,            X86::RAX, U }, // UDiv
        { X86::DIV64r,  X86::MOV32r0, Copy,            X86::RDX, U }, // URem
      }
    }, // i64
  };

  MVT VT;
  if (!isTypeLegal(I->getType(), VT))
    return false;

  unsigned TypeIndex, OpIndex;
  switch (VT.SimpleTy) {
  default: return false;
  case MVT::i8:  TypeIndex = 0; break;
  case MVT::i16: TypeIndex = 1; break;
  case MVT::i32: TypeIndex = 2; break;
  case MVT::i64: TypeIndex = 3;
    if (!Subtarget->is64Bit())
      return false;
    break;
  }

  switch (I->getOpcode()) {
  default: llvm_unreachable("Unexpected div/rem opcode");
  case Instruction::SDiv: OpIndex = 0; break;
  case Instruction::SRem: OpIndex = 1; break;
  case Instruction::UDiv: OpIndex = 2; break;
  case Instruction::URem: OpIndex = 3; break;
  }

  const DivRemEntry &TypeEntry = OpTable[TypeIndex];
  const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex];
  Register Op0Reg = getRegForValue(I->getOperand(0));
  if (Op0Reg == 0)
    return false;
  Register Op1Reg = getRegForValue(I->getOperand(1));
  if (Op1Reg == 0)
    return false;

  // Move op0 into low-order input register.
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
          TII.get(OpEntry.OpCopy), TypeEntry.LowInReg).addReg(Op0Reg);
  // Zero-extend or sign-extend into high-order input register.
  if (OpEntry.OpSignExtend) {
    if (OpEntry.IsOpSigned)
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(OpEntry.OpSignExtend));
    else {
      Register Zero32 = createResultReg(&X86::GR32RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(X86::MOV32r0), Zero32);

      // Copy the zero into the appropriate sub/super/identical physical
      // register. Unfortunately the operations needed are not uniform enough
      // to fit neatly into the table above.
      if (VT == MVT::i16) {
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                TII.get(Copy), TypeEntry.HighInReg)
          .addReg(Zero32, 0, X86::sub_16bit);
      } else if (VT == MVT::i32) {
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                TII.get(Copy), TypeEntry.HighInReg)
            .addReg(Zero32);
      } else if (VT == MVT::i64) {
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg)
            .addImm(0).addReg(Zero32).addImm(X86::sub_32bit);
      }
    }
  }
  // Generate the DIV/IDIV instruction.
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
          TII.get(OpEntry.OpDivRem)).addReg(Op1Reg);
  // For i8 remainder, we can't reference ah directly, as we'll end
  // up with bogus copies like %r9b = COPY %ah. Reference ax
  // instead to prevent ah references in a rex instruction.
  //
  // The current assumption of the fast register allocator is that isel
  // won't generate explicit references to the GR8_NOREX registers. If
  // the allocator and/or the backend get enhanced to be more robust in
  // that regard, this can be, and should be, removed.
  unsigned ResultReg = 0;
  if ((I->getOpcode() == Instruction::SRem ||
       I->getOpcode() == Instruction::URem) &&
      OpEntry.DivRemResultReg == X86::AH && Subtarget->is64Bit()) {
    Register SourceSuperReg = createResultReg(&X86::GR16RegClass);
    Register ResultSuperReg = createResultReg(&X86::GR16RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(Copy), SourceSuperReg).addReg(X86::AX);

    // Shift AX right by 8 bits instead of using AH.
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SHR16ri),
            ResultSuperReg).addReg(SourceSuperReg).addImm(8);

    // Now reference the 8-bit subreg of the result.
    ResultReg = fastEmitInst_extractsubreg(MVT::i8, ResultSuperReg,
                                           /*Op0IsKill=*/true, X86::sub_8bit);
  }
  // Copy the result out of the physreg if we haven't already.
  if (!ResultReg) {
    ResultReg = createResultReg(TypeEntry.RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Copy), ResultReg)
        .addReg(OpEntry.DivRemResultReg);
  }
  updateValueMap(I, ResultReg);

  return true;
}

/// Emit a conditional move instruction (if the are supported) to lower
/// the select.
bool X86FastISel::X86FastEmitCMoveSelect(MVT RetVT, const Instruction *I) {
  // Check if the subtarget supports these instructions.
  if (!Subtarget->hasCMov())
    return false;

  // FIXME: Add support for i8.
  if (RetVT < MVT::i16 || RetVT > MVT::i64)
    return false;

  const Value *Cond = I->getOperand(0);
  const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
  bool NeedTest = true;
  X86::CondCode CC = X86::COND_NE;

  // Optimize conditions coming from a compare if both instructions are in the
  // same basic block (values defined in other basic blocks may not have
  // initialized registers).
  const auto *CI = dyn_cast<CmpInst>(Cond);
  if (CI && (CI->getParent() == I->getParent())) {
    CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);

    // FCMP_OEQ and FCMP_UNE cannot be checked with a single instruction.
    static const uint16_t SETFOpcTable[2][3] = {
      { X86::COND_NP, X86::COND_E,  X86::TEST8rr },
      { X86::COND_P,  X86::COND_NE, X86::OR8rr   }
    };
    const uint16_t *SETFOpc = nullptr;
    switch (Predicate) {
    default: break;
    case CmpInst::FCMP_OEQ:
      SETFOpc = &SETFOpcTable[0][0];
      Predicate = CmpInst::ICMP_NE;
      break;
    case CmpInst::FCMP_UNE:
      SETFOpc = &SETFOpcTable[1][0];
      Predicate = CmpInst::ICMP_NE;
      break;
    }

    bool NeedSwap;
    std::tie(CC, NeedSwap) = X86::getX86ConditionCode(Predicate);
    assert(CC <= X86::LAST_VALID_COND && "Unexpected condition code.");

    const Value *CmpLHS = CI->getOperand(0);
    const Value *CmpRHS = CI->getOperand(1);
    if (NeedSwap)
      std::swap(CmpLHS, CmpRHS);

    EVT CmpVT = TLI.getValueType(DL, CmpLHS->getType());
    // Emit a compare of the LHS and RHS, setting the flags.
    if (!X86FastEmitCompare(CmpLHS, CmpRHS, CmpVT, CI->getDebugLoc()))
      return false;

    if (SETFOpc) {
      Register FlagReg1 = createResultReg(&X86::GR8RegClass);
      Register FlagReg2 = createResultReg(&X86::GR8RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETCCr),
              FlagReg1).addImm(SETFOpc[0]);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETCCr),
              FlagReg2).addImm(SETFOpc[1]);
      auto const &II = TII.get(SETFOpc[2]);
      if (II.getNumDefs()) {
        Register TmpReg = createResultReg(&X86::GR8RegClass);
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, TmpReg)
          .addReg(FlagReg2).addReg(FlagReg1);
      } else {
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
          .addReg(FlagReg2).addReg(FlagReg1);
      }
    }
    NeedTest = false;
  } else if (foldX86XALUIntrinsic(CC, I, Cond)) {
    // Fake request the condition, otherwise the intrinsic might be completely
    // optimized away.
    Register TmpReg = getRegForValue(Cond);
    if (TmpReg == 0)
      return false;

    NeedTest = false;
  }

  if (NeedTest) {
    // Selects operate on i1, however, CondReg is 8 bits width and may contain
    // garbage. Indeed, only the less significant bit is supposed to be
    // accurate. If we read more than the lsb, we may see non-zero values
    // whereas lsb is zero. Therefore, we have to truncate Op0Reg to i1 for
    // the select. This is achieved by performing TEST against 1.
    Register CondReg = getRegForValue(Cond);
    if (CondReg == 0)
      return false;
    bool CondIsKill = hasTrivialKill(Cond);

    // In case OpReg is a K register, COPY to a GPR
    if (MRI.getRegClass(CondReg) == &X86::VK1RegClass) {
      unsigned KCondReg = CondReg;
      CondReg = createResultReg(&X86::GR32RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::COPY), CondReg)
          .addReg(KCondReg, getKillRegState(CondIsKill));
      CondReg = fastEmitInst_extractsubreg(MVT::i8, CondReg, /*Op0IsKill=*/true,
                                           X86::sub_8bit);
    }
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
        .addReg(CondReg, getKillRegState(CondIsKill))
        .addImm(1);
  }

  const Value *LHS = I->getOperand(1);
  const Value *RHS = I->getOperand(2);

  Register RHSReg = getRegForValue(RHS);
  bool RHSIsKill = hasTrivialKill(RHS);

  Register LHSReg = getRegForValue(LHS);
  bool LHSIsKill = hasTrivialKill(LHS);

  if (!LHSReg || !RHSReg)
    return false;

  const TargetRegisterInfo &TRI = *Subtarget->getRegisterInfo();
  unsigned Opc = X86::getCMovOpcode(TRI.getRegSizeInBits(*RC)/8);
  Register ResultReg = fastEmitInst_rri(Opc, RC, RHSReg, RHSIsKill,
                                        LHSReg, LHSIsKill, CC);
  updateValueMap(I, ResultReg);
  return true;
}

/// Emit SSE or AVX instructions to lower the select.
///
/// Try to use SSE1/SSE2 instructions to simulate a select without branches.
/// This lowers fp selects into a CMP/AND/ANDN/OR sequence when the necessary
/// SSE instructions are available. If AVX is available, try to use a VBLENDV.
bool X86FastISel::X86FastEmitSSESelect(MVT RetVT, const Instruction *I) {
  // Optimize conditions coming from a compare if both instructions are in the
  // same basic block (values defined in other basic blocks may not have
  // initialized registers).
  const auto *CI = dyn_cast<FCmpInst>(I->getOperand(0));
  if (!CI || (CI->getParent() != I->getParent()))
    return false;

  if (I->getType() != CI->getOperand(0)->getType() ||
      !((Subtarget->hasSSE1() && RetVT == MVT::f32) ||
        (Subtarget->hasSSE2() && RetVT == MVT::f64)))
    return false;

  const Value *CmpLHS = CI->getOperand(0);
  const Value *CmpRHS = CI->getOperand(1);
  CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);

  // The optimizer might have replaced fcmp oeq %x, %x with fcmp ord %x, 0.0.
  // We don't have to materialize a zero constant for this case and can just use
  // %x again on the RHS.
  if (Predicate == CmpInst::FCMP_ORD || Predicate == CmpInst::FCMP_UNO) {
    const auto *CmpRHSC = dyn_cast<ConstantFP>(CmpRHS);
    if (CmpRHSC && CmpRHSC->isNullValue())
      CmpRHS = CmpLHS;
  }

  unsigned CC;
  bool NeedSwap;
  std::tie(CC, NeedSwap) = getX86SSEConditionCode(Predicate);
  if (CC > 7 && !Subtarget->hasAVX())
    return false;

  if (NeedSwap)
    std::swap(CmpLHS, CmpRHS);

  const Value *LHS = I->getOperand(1);
  const Value *RHS = I->getOperand(2);

  Register LHSReg = getRegForValue(LHS);
  bool LHSIsKill = hasTrivialKill(LHS);

  Register RHSReg = getRegForValue(RHS);
  bool RHSIsKill = hasTrivialKill(RHS);

  Register CmpLHSReg = getRegForValue(CmpLHS);
  bool CmpLHSIsKill = hasTrivialKill(CmpLHS);

  Register CmpRHSReg = getRegForValue(CmpRHS);
  bool CmpRHSIsKill = hasTrivialKill(CmpRHS);

  if (!LHSReg || !RHSReg || !CmpLHSReg || !CmpRHSReg)
    return false;

  const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
  unsigned ResultReg;

  if (Subtarget->hasAVX512()) {
    // If we have AVX512 we can use a mask compare and masked movss/sd.
    const TargetRegisterClass *VR128X = &X86::VR128XRegClass;
    const TargetRegisterClass *VK1 = &X86::VK1RegClass;

    unsigned CmpOpcode =
      (RetVT == MVT::f32) ? X86::VCMPSSZrr : X86::VCMPSDZrr;
    Register CmpReg = fastEmitInst_rri(CmpOpcode, VK1, CmpLHSReg, CmpLHSIsKill,
                                       CmpRHSReg, CmpRHSIsKill, CC);

    // Need an IMPLICIT_DEF for the input that is used to generate the upper
    // bits of the result register since its not based on any of the inputs.
    Register ImplicitDefReg = createResultReg(VR128X);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);

    // Place RHSReg is the passthru of the masked movss/sd operation and put
    // LHS in the input. The mask input comes from the compare.
    unsigned MovOpcode =
      (RetVT == MVT::f32) ? X86::VMOVSSZrrk : X86::VMOVSDZrrk;
    unsigned MovReg = fastEmitInst_rrrr(MovOpcode, VR128X, RHSReg, RHSIsKill,
                                        CmpReg, true, ImplicitDefReg, true,
                                        LHSReg, LHSIsKill);

    ResultReg = createResultReg(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), ResultReg).addReg(MovReg);

  } else if (Subtarget->hasAVX()) {
    const TargetRegisterClass *VR128 = &X86::VR128RegClass;

    // If we have AVX, create 1 blendv instead of 3 logic instructions.
    // Blendv was introduced with SSE 4.1, but the 2 register form implicitly
    // uses XMM0 as the selection register. That may need just as many
    // instructions as the AND/ANDN/OR sequence due to register moves, so
    // don't bother.
    unsigned CmpOpcode =
      (RetVT == MVT::f32) ? X86::VCMPSSrr : X86::VCMPSDrr;
    unsigned BlendOpcode =
      (RetVT == MVT::f32) ? X86::VBLENDVPSrr : X86::VBLENDVPDrr;

    Register CmpReg = fastEmitInst_rri(CmpOpcode, RC, CmpLHSReg, CmpLHSIsKill,
                                       CmpRHSReg, CmpRHSIsKill, CC);
    Register VBlendReg = fastEmitInst_rrr(BlendOpcode, VR128, RHSReg, RHSIsKill,
                                          LHSReg, LHSIsKill, CmpReg, true);
    ResultReg = createResultReg(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), ResultReg).addReg(VBlendReg);
  } else {
    // Choose the SSE instruction sequence based on data type (float or double).
    static const uint16_t OpcTable[2][4] = {
      { X86::CMPSSrr,  X86::ANDPSrr,  X86::ANDNPSrr,  X86::ORPSrr  },
      { X86::CMPSDrr,  X86::ANDPDrr,  X86::ANDNPDrr,  X86::ORPDrr  }
    };

    const uint16_t *Opc = nullptr;
    switch (RetVT.SimpleTy) {
    default: return false;
    case MVT::f32: Opc = &OpcTable[0][0]; break;
    case MVT::f64: Opc = &OpcTable[1][0]; break;
    }

    const TargetRegisterClass *VR128 = &X86::VR128RegClass;
    Register CmpReg = fastEmitInst_rri(Opc[0], RC, CmpLHSReg, CmpLHSIsKill,
                                       CmpRHSReg, CmpRHSIsKill, CC);
    Register AndReg = fastEmitInst_rr(Opc[1], VR128, CmpReg,
                                      /*Op0IsKill=*/false, LHSReg, LHSIsKill);
    Register AndNReg = fastEmitInst_rr(Opc[2], VR128, CmpReg,
                                       /*Op0IsKill=*/true, RHSReg, RHSIsKill);
    Register OrReg = fastEmitInst_rr(Opc[3], VR128, AndNReg, /*Op0IsKill=*/true,
                                     AndReg, /*Op1IsKill=*/true);
    ResultReg = createResultReg(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), ResultReg).addReg(OrReg);
  }
  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86FastEmitPseudoSelect(MVT RetVT, const Instruction *I) {
  // These are pseudo CMOV instructions and will be later expanded into control-
  // flow.
  unsigned Opc;
  switch (RetVT.SimpleTy) {
  default: return false;
  case MVT::i8:  Opc = X86::CMOV_GR8;  break;
  case MVT::i16: Opc = X86::CMOV_GR16; break;
  case MVT::i32: Opc = X86::CMOV_GR32; break;
  case MVT::f32: Opc = Subtarget->hasAVX512() ? X86::CMOV_FR32X
                                              : X86::CMOV_FR32; break;
  case MVT::f64: Opc = Subtarget->hasAVX512() ? X86::CMOV_FR64X
                                              : X86::CMOV_FR64; break;
  }

  const Value *Cond = I->getOperand(0);
  X86::CondCode CC = X86::COND_NE;

  // Optimize conditions coming from a compare if both instructions are in the
  // same basic block (values defined in other basic blocks may not have
  // initialized registers).
  const auto *CI = dyn_cast<CmpInst>(Cond);
  if (CI && (CI->getParent() == I->getParent())) {
    bool NeedSwap;
    std::tie(CC, NeedSwap) = X86::getX86ConditionCode(CI->getPredicate());
    if (CC > X86::LAST_VALID_COND)
      return false;

    const Value *CmpLHS = CI->getOperand(0);
    const Value *CmpRHS = CI->getOperand(1);

    if (NeedSwap)
      std::swap(CmpLHS, CmpRHS);

    EVT CmpVT = TLI.getValueType(DL, CmpLHS->getType());
    if (!X86FastEmitCompare(CmpLHS, CmpRHS, CmpVT, CI->getDebugLoc()))
      return false;
  } else {
    Register CondReg = getRegForValue(Cond);
    if (CondReg == 0)
      return false;
    bool CondIsKill = hasTrivialKill(Cond);

    // In case OpReg is a K register, COPY to a GPR
    if (MRI.getRegClass(CondReg) == &X86::VK1RegClass) {
      unsigned KCondReg = CondReg;
      CondReg = createResultReg(&X86::GR32RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::COPY), CondReg)
          .addReg(KCondReg, getKillRegState(CondIsKill));
      CondReg = fastEmitInst_extractsubreg(MVT::i8, CondReg, /*Op0IsKill=*/true,
                                           X86::sub_8bit);
    }
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TEST8ri))
        .addReg(CondReg, getKillRegState(CondIsKill))
        .addImm(1);
  }

  const Value *LHS = I->getOperand(1);
  const Value *RHS = I->getOperand(2);

  Register LHSReg = getRegForValue(LHS);
  bool LHSIsKill = hasTrivialKill(LHS);

  Register RHSReg = getRegForValue(RHS);
  bool RHSIsKill = hasTrivialKill(RHS);

  if (!LHSReg || !RHSReg)
    return false;

  const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);

  Register ResultReg =
    fastEmitInst_rri(Opc, RC, RHSReg, RHSIsKill, LHSReg, LHSIsKill, CC);
  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectSelect(const Instruction *I) {
  MVT RetVT;
  if (!isTypeLegal(I->getType(), RetVT))
    return false;

  // Check if we can fold the select.
  if (const auto *CI = dyn_cast<CmpInst>(I->getOperand(0))) {
    CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
    const Value *Opnd = nullptr;
    switch (Predicate) {
    default:                              break;
    case CmpInst::FCMP_FALSE: Opnd = I->getOperand(2); break;
    case CmpInst::FCMP_TRUE:  Opnd = I->getOperand(1); break;
    }
    // No need for a select anymore - this is an unconditional move.
    if (Opnd) {
      Register OpReg = getRegForValue(Opnd);
      if (OpReg == 0)
        return false;
      bool OpIsKill = hasTrivialKill(Opnd);
      const TargetRegisterClass *RC = TLI.getRegClassFor(RetVT);
      Register ResultReg = createResultReg(RC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::COPY), ResultReg)
        .addReg(OpReg, getKillRegState(OpIsKill));
      updateValueMap(I, ResultReg);
      return true;
    }
  }

  // First try to use real conditional move instructions.
  if (X86FastEmitCMoveSelect(RetVT, I))
    return true;

  // Try to use a sequence of SSE instructions to simulate a conditional move.
  if (X86FastEmitSSESelect(RetVT, I))
    return true;

  // Fall-back to pseudo conditional move instructions, which will be later
  // converted to control-flow.
  if (X86FastEmitPseudoSelect(RetVT, I))
    return true;

  return false;
}

// Common code for X86SelectSIToFP and X86SelectUIToFP.
bool X86FastISel::X86SelectIntToFP(const Instruction *I, bool IsSigned) {
  // The target-independent selection algorithm in FastISel already knows how
  // to select a SINT_TO_FP if the target is SSE but not AVX.
  // Early exit if the subtarget doesn't have AVX.
  // Unsigned conversion requires avx512.
  bool HasAVX512 = Subtarget->hasAVX512();
  if (!Subtarget->hasAVX() || (!IsSigned && !HasAVX512))
    return false;

  // TODO: We could sign extend narrower types.
  MVT SrcVT = TLI.getSimpleValueType(DL, I->getOperand(0)->getType());
  if (SrcVT != MVT::i32 && SrcVT != MVT::i64)
    return false;

  // Select integer to float/double conversion.
  Register OpReg = getRegForValue(I->getOperand(0));
  if (OpReg == 0)
    return false;

  unsigned Opcode;

  static const uint16_t SCvtOpc[2][2][2] = {
    { { X86::VCVTSI2SSrr,  X86::VCVTSI642SSrr },
      { X86::VCVTSI2SDrr,  X86::VCVTSI642SDrr } },
    { { X86::VCVTSI2SSZrr, X86::VCVTSI642SSZrr },
      { X86::VCVTSI2SDZrr, X86::VCVTSI642SDZrr } },
  };
  static const uint16_t UCvtOpc[2][2] = {
    { X86::VCVTUSI2SSZrr, X86::VCVTUSI642SSZrr },
    { X86::VCVTUSI2SDZrr, X86::VCVTUSI642SDZrr },
  };
  bool Is64Bit = SrcVT == MVT::i64;

  if (I->getType()->isDoubleTy()) {
    // s/uitofp int -> double
    Opcode = IsSigned ? SCvtOpc[HasAVX512][1][Is64Bit] : UCvtOpc[1][Is64Bit];
  } else if (I->getType()->isFloatTy()) {
    // s/uitofp int -> float
    Opcode = IsSigned ? SCvtOpc[HasAVX512][0][Is64Bit] : UCvtOpc[0][Is64Bit];
  } else
    return false;

  MVT DstVT = TLI.getValueType(DL, I->getType()).getSimpleVT();
  const TargetRegisterClass *RC = TLI.getRegClassFor(DstVT);
  Register ImplicitDefReg = createResultReg(RC);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
          TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
  Register ResultReg =
      fastEmitInst_rr(Opcode, RC, ImplicitDefReg, true, OpReg, false);
  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectSIToFP(const Instruction *I) {
  return X86SelectIntToFP(I, /*IsSigned*/true);
}

bool X86FastISel::X86SelectUIToFP(const Instruction *I) {
  return X86SelectIntToFP(I, /*IsSigned*/false);
}

// Helper method used by X86SelectFPExt and X86SelectFPTrunc.
bool X86FastISel::X86SelectFPExtOrFPTrunc(const Instruction *I,
                                          unsigned TargetOpc,
                                          const TargetRegisterClass *RC) {
  assert((I->getOpcode() == Instruction::FPExt ||
          I->getOpcode() == Instruction::FPTrunc) &&
         "Instruction must be an FPExt or FPTrunc!");
  bool HasAVX = Subtarget->hasAVX();

  Register OpReg = getRegForValue(I->getOperand(0));
  if (OpReg == 0)
    return false;

  unsigned ImplicitDefReg;
  if (HasAVX) {
    ImplicitDefReg = createResultReg(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);

  }

  Register ResultReg = createResultReg(RC);
  MachineInstrBuilder MIB;
  MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(TargetOpc),
                ResultReg);

  if (HasAVX)
    MIB.addReg(ImplicitDefReg);

  MIB.addReg(OpReg);
  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectFPExt(const Instruction *I) {
  if (X86ScalarSSEf64 && I->getType()->isDoubleTy() &&
      I->getOperand(0)->getType()->isFloatTy()) {
    bool HasAVX512 = Subtarget->hasAVX512();
    // fpext from float to double.
    unsigned Opc =
        HasAVX512 ? X86::VCVTSS2SDZrr
                  : Subtarget->hasAVX() ? X86::VCVTSS2SDrr : X86::CVTSS2SDrr;
    return X86SelectFPExtOrFPTrunc(I, Opc, TLI.getRegClassFor(MVT::f64));
  }

  return false;
}

bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
  if (X86ScalarSSEf64 && I->getType()->isFloatTy() &&
      I->getOperand(0)->getType()->isDoubleTy()) {
    bool HasAVX512 = Subtarget->hasAVX512();
    // fptrunc from double to float.
    unsigned Opc =
        HasAVX512 ? X86::VCVTSD2SSZrr
                  : Subtarget->hasAVX() ? X86::VCVTSD2SSrr : X86::CVTSD2SSrr;
    return X86SelectFPExtOrFPTrunc(I, Opc, TLI.getRegClassFor(MVT::f32));
  }

  return false;
}

bool X86FastISel::X86SelectTrunc(const Instruction *I) {
  EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType());
  EVT DstVT = TLI.getValueType(DL, I->getType());

  // This code only handles truncation to byte.
  if (DstVT != MVT::i8 && DstVT != MVT::i1)
    return false;
  if (!TLI.isTypeLegal(SrcVT))
    return false;

  Register InputReg = getRegForValue(I->getOperand(0));
  if (!InputReg)
    // Unhandled operand.  Halt "fast" selection and bail.
    return false;

  if (SrcVT == MVT::i8) {
    // Truncate from i8 to i1; no code needed.
    updateValueMap(I, InputReg);
    return true;
  }

  // Issue an extract_subreg.
  Register ResultReg = fastEmitInst_extractsubreg(MVT::i8,
                                                  InputReg, false,
                                                  X86::sub_8bit);
  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::IsMemcpySmall(uint64_t Len) {
  return Len <= (Subtarget->is64Bit() ? 32 : 16);
}

bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM,
                                     X86AddressMode SrcAM, uint64_t Len) {

  // Make sure we don't bloat code by inlining very large memcpy's.
  if (!IsMemcpySmall(Len))
    return false;

  bool i64Legal = Subtarget->is64Bit();

  // We don't care about alignment here since we just emit integer accesses.
  while (Len) {
    MVT VT;
    if (Len >= 8 && i64Legal)
      VT = MVT::i64;
    else if (Len >= 4)
      VT = MVT::i32;
    else if (Len >= 2)
      VT = MVT::i16;
    else
      VT = MVT::i8;

    unsigned Reg;
    bool RV = X86FastEmitLoad(VT, SrcAM, nullptr, Reg);
    RV &= X86FastEmitStore(VT, Reg, /*ValIsKill=*/true, DestAM);
    assert(RV && "Failed to emit load or store??");

    unsigned Size = VT.getSizeInBits()/8;
    Len -= Size;
    DestAM.Disp += Size;
    SrcAM.Disp += Size;
  }

  return true;
}

bool X86FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
  // FIXME: Handle more intrinsics.
  switch (II->getIntrinsicID()) {
  default: return false;
  case Intrinsic::convert_from_fp16:
  case Intrinsic::convert_to_fp16: {
    if (Subtarget->useSoftFloat() || !Subtarget->hasF16C())
      return false;

    const Value *Op = II->getArgOperand(0);
    Register InputReg = getRegForValue(Op);
    if (InputReg == 0)
      return false;

    // F16C only allows converting from float to half and from half to float.
    bool IsFloatToHalf = II->getIntrinsicID() == Intrinsic::convert_to_fp16;
    if (IsFloatToHalf) {
      if (!Op->getType()->isFloatTy())
        return false;
    } else {
      if (!II->getType()->isFloatTy())
        return false;
    }

    unsigned ResultReg = 0;
    const TargetRegisterClass *RC = TLI.getRegClassFor(MVT::v8i16);
    if (IsFloatToHalf) {
      // 'InputReg' is implicitly promoted from register class FR32 to
      // register class VR128 by method 'constrainOperandRegClass' which is
      // directly called by 'fastEmitInst_ri'.
      // Instruction VCVTPS2PHrr takes an extra immediate operand which is
      // used to provide rounding control: use MXCSR.RC, encoded as 0b100.
      // It's consistent with the other FP instructions, which are usually
      // controlled by MXCSR.
      unsigned Opc = Subtarget->hasVLX() ? X86::VCVTPS2PHZ128rr
                                         : X86::VCVTPS2PHrr;
      InputReg = fastEmitInst_ri(Opc, RC, InputReg, false, 4);

      // Move the lower 32-bits of ResultReg to another register of class GR32.
      Opc = Subtarget->hasAVX512() ? X86::VMOVPDI2DIZrr
                                   : X86::VMOVPDI2DIrr;
      ResultReg = createResultReg(&X86::GR32RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
          .addReg(InputReg, RegState::Kill);

      // The result value is in the lower 16-bits of ResultReg.
      unsigned RegIdx = X86::sub_16bit;
      ResultReg = fastEmitInst_extractsubreg(MVT::i16, ResultReg, true, RegIdx);
    } else {
      assert(Op->getType()->isIntegerTy(16) && "Expected a 16-bit integer!");
      // Explicitly zero-extend the input to 32-bit.
      InputReg = fastEmit_r(MVT::i16, MVT::i32, ISD::ZERO_EXTEND, InputReg,
                            /*Op0IsKill=*/false);

      // The following SCALAR_TO_VECTOR will be expanded into a VMOVDI2PDIrr.
      InputReg = fastEmit_r(MVT::i32, MVT::v4i32, ISD::SCALAR_TO_VECTOR,
                            InputReg, /*Op0IsKill=*/true);

      unsigned Opc = Subtarget->hasVLX() ? X86::VCVTPH2PSZ128rr
                                         : X86::VCVTPH2PSrr;
      InputReg = fastEmitInst_r(Opc, RC, InputReg, /*Op0IsKill=*/true);

      // The result value is in the lower 32-bits of ResultReg.
      // Emit an explicit copy from register class VR128 to register class FR32.
      ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::COPY), ResultReg)
          .addReg(InputReg, RegState::Kill);
    }

    updateValueMap(II, ResultReg);
    return true;
  }
  case Intrinsic::frameaddress: {
    MachineFunction *MF = FuncInfo.MF;
    if (MF->getTarget().getMCAsmInfo()->usesWindowsCFI())
      return false;

    Type *RetTy = II->getCalledFunction()->getReturnType();

    MVT VT;
    if (!isTypeLegal(RetTy, VT))
      return false;

    unsigned Opc;
    const TargetRegisterClass *RC = nullptr;

    switch (VT.SimpleTy) {
    default: llvm_unreachable("Invalid result type for frameaddress.");
    case MVT::i32: Opc = X86::MOV32rm; RC = &X86::GR32RegClass; break;
    case MVT::i64: Opc = X86::MOV64rm; RC = &X86::GR64RegClass; break;
    }

    // This needs to be set before we call getPtrSizedFrameRegister, otherwise
    // we get the wrong frame register.
    MachineFrameInfo &MFI = MF->getFrameInfo();
    MFI.setFrameAddressIsTaken(true);

    const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
    unsigned FrameReg = RegInfo->getPtrSizedFrameRegister(*MF);
    assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
            (FrameReg == X86::EBP && VT == MVT::i32)) &&
           "Invalid Frame Register!");

    // Always make a copy of the frame register to a vreg first, so that we
    // never directly reference the frame register (the TwoAddressInstruction-
    // Pass doesn't like that).
    Register SrcReg = createResultReg(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), SrcReg).addReg(FrameReg);

    // Now recursively load from the frame address.
    // movq (%rbp), %rax
    // movq (%rax), %rax
    // movq (%rax), %rax
    // ...
    unsigned DestReg;
    unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
    while (Depth--) {
      DestReg = createResultReg(RC);
      addDirectMem(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                           TII.get(Opc), DestReg), SrcReg);
      SrcReg = DestReg;
    }

    updateValueMap(II, SrcReg);
    return true;
  }
  case Intrinsic::memcpy: {
    const MemCpyInst *MCI = cast<MemCpyInst>(II);
    // Don't handle volatile or variable length memcpys.
    if (MCI->isVolatile())
      return false;

    if (isa<ConstantInt>(MCI->getLength())) {
      // Small memcpy's are common enough that we want to do them
      // without a call if possible.
      uint64_t Len = cast<ConstantInt>(MCI->getLength())->getZExtValue();
      if (IsMemcpySmall(Len)) {
        X86AddressMode DestAM, SrcAM;
        if (!X86SelectAddress(MCI->getRawDest(), DestAM) ||
            !X86SelectAddress(MCI->getRawSource(), SrcAM))
          return false;
        TryEmitSmallMemcpy(DestAM, SrcAM, Len);
        return true;
      }
    }

    unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
    if (!MCI->getLength()->getType()->isIntegerTy(SizeWidth))
      return false;

    if (MCI->getSourceAddressSpace() > 255 || MCI->getDestAddressSpace() > 255)
      return false;

    return lowerCallTo(II, "memcpy", II->getNumArgOperands() - 1);
  }
  case Intrinsic::memset: {
    const MemSetInst *MSI = cast<MemSetInst>(II);

    if (MSI->isVolatile())
      return false;

    unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
    if (!MSI->getLength()->getType()->isIntegerTy(SizeWidth))
      return false;

    if (MSI->getDestAddressSpace() > 255)
      return false;

    return lowerCallTo(II, "memset", II->getNumArgOperands() - 1);
  }
  case Intrinsic::stackprotector: {
    // Emit code to store the stack guard onto the stack.
    EVT PtrTy = TLI.getPointerTy(DL);

    const Value *Op1 = II->getArgOperand(0); // The guard's value.
    const AllocaInst *Slot = cast<AllocaInst>(II->getArgOperand(1));

    MFI.setStackProtectorIndex(FuncInfo.StaticAllocaMap[Slot]);

    // Grab the frame index.
    X86AddressMode AM;
    if (!X86SelectAddress(Slot, AM)) return false;
    if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
    return true;
  }
  case Intrinsic::dbg_declare: {
    const DbgDeclareInst *DI = cast<DbgDeclareInst>(II);
    X86AddressMode AM;
    assert(DI->getAddress() && "Null address should be checked earlier!");
    if (!X86SelectAddress(DI->getAddress(), AM))
      return false;
    const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
    // FIXME may need to add RegState::Debug to any registers produced,
    // although ESP/EBP should be the only ones at the moment.
    assert(DI->getVariable()->isValidLocationForIntrinsic(DbgLoc) &&
           "Expected inlined-at fields to agree");
    addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II), AM)
        .addImm(0)
        .addMetadata(DI->getVariable())
        .addMetadata(DI->getExpression());
    return true;
  }
  case Intrinsic::trap: {
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::TRAP));
    return true;
  }
  case Intrinsic::sqrt: {
    if (!Subtarget->hasSSE1())
      return false;

    Type *RetTy = II->getCalledFunction()->getReturnType();

    MVT VT;
    if (!isTypeLegal(RetTy, VT))
      return false;

    // Unfortunately we can't use fastEmit_r, because the AVX version of FSQRT
    // is not generated by FastISel yet.
    // FIXME: Update this code once tablegen can handle it.
    static const uint16_t SqrtOpc[3][2] = {
      { X86::SQRTSSr,   X86::SQRTSDr },
      { X86::VSQRTSSr,  X86::VSQRTSDr },
      { X86::VSQRTSSZr, X86::VSQRTSDZr },
    };
    unsigned AVXLevel = Subtarget->hasAVX512() ? 2 :
                        Subtarget->hasAVX()    ? 1 :
                                                 0;
    unsigned Opc;
    switch (VT.SimpleTy) {
    default: return false;
    case MVT::f32: Opc = SqrtOpc[AVXLevel][0]; break;
    case MVT::f64: Opc = SqrtOpc[AVXLevel][1]; break;
    }

    const Value *SrcVal = II->getArgOperand(0);
    Register SrcReg = getRegForValue(SrcVal);

    if (SrcReg == 0)
      return false;

    const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
    unsigned ImplicitDefReg = 0;
    if (AVXLevel > 0) {
      ImplicitDefReg = createResultReg(RC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::IMPLICIT_DEF), ImplicitDefReg);
    }

    Register ResultReg = createResultReg(RC);
    MachineInstrBuilder MIB;
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc),
                  ResultReg);

    if (ImplicitDefReg)
      MIB.addReg(ImplicitDefReg);

    MIB.addReg(SrcReg);

    updateValueMap(II, ResultReg);
    return true;
  }
  case Intrinsic::sadd_with_overflow:
  case Intrinsic::uadd_with_overflow:
  case Intrinsic::ssub_with_overflow:
  case Intrinsic::usub_with_overflow:
  case Intrinsic::smul_with_overflow:
  case Intrinsic::umul_with_overflow: {
    // This implements the basic lowering of the xalu with overflow intrinsics
    // into add/sub/mul followed by either seto or setb.
    const Function *Callee = II->getCalledFunction();
    auto *Ty = cast<StructType>(Callee->getReturnType());
    Type *RetTy = Ty->getTypeAtIndex(0U);
    assert(Ty->getTypeAtIndex(1)->isIntegerTy() &&
           Ty->getTypeAtIndex(1)->getScalarSizeInBits() == 1 &&
           "Overflow value expected to be an i1");

    MVT VT;
    if (!isTypeLegal(RetTy, VT))
      return false;

    if (VT < MVT::i8 || VT > MVT::i64)
      return false;

    const Value *LHS = II->getArgOperand(0);
    const Value *RHS = II->getArgOperand(1);

    // Canonicalize immediate to the RHS.
    if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS) && II->isCommutative())
      std::swap(LHS, RHS);

    unsigned BaseOpc, CondCode;
    switch (II->getIntrinsicID()) {
    default: llvm_unreachable("Unexpected intrinsic!");
    case Intrinsic::sadd_with_overflow:
      BaseOpc = ISD::ADD; CondCode = X86::COND_O; break;
    case Intrinsic::uadd_with_overflow:
      BaseOpc = ISD::ADD; CondCode = X86::COND_B; break;
    case Intrinsic::ssub_with_overflow:
      BaseOpc = ISD::SUB; CondCode = X86::COND_O; break;
    case Intrinsic::usub_with_overflow:
      BaseOpc = ISD::SUB; CondCode = X86::COND_B; break;
    case Intrinsic::smul_with_overflow:
      BaseOpc = X86ISD::SMUL; CondCode = X86::COND_O; break;
    case Intrinsic::umul_with_overflow:
      BaseOpc = X86ISD::UMUL; CondCode = X86::COND_O; break;
    }

    Register LHSReg = getRegForValue(LHS);
    if (LHSReg == 0)
      return false;
    bool LHSIsKill = hasTrivialKill(LHS);

    unsigned ResultReg = 0;
    // Check if we have an immediate version.
    if (const auto *CI = dyn_cast<ConstantInt>(RHS)) {
      static const uint16_t Opc[2][4] = {
        { X86::INC8r, X86::INC16r, X86::INC32r, X86::INC64r },
        { X86::DEC8r, X86::DEC16r, X86::DEC32r, X86::DEC64r }
      };

      if (CI->isOne() && (BaseOpc == ISD::ADD || BaseOpc == ISD::SUB) &&
          CondCode == X86::COND_O) {
        // We can use INC/DEC.
        ResultReg = createResultReg(TLI.getRegClassFor(VT));
        bool IsDec = BaseOpc == ISD::SUB;
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                TII.get(Opc[IsDec][VT.SimpleTy-MVT::i8]), ResultReg)
          .addReg(LHSReg, getKillRegState(LHSIsKill));
      } else
        ResultReg = fastEmit_ri(VT, VT, BaseOpc, LHSReg, LHSIsKill,
                                CI->getZExtValue());
    }

    unsigned RHSReg;
    bool RHSIsKill;
    if (!ResultReg) {
      RHSReg = getRegForValue(RHS);
      if (RHSReg == 0)
        return false;
      RHSIsKill = hasTrivialKill(RHS);
      ResultReg = fastEmit_rr(VT, VT, BaseOpc, LHSReg, LHSIsKill, RHSReg,
                              RHSIsKill);
    }

    // FastISel doesn't have a pattern for all X86::MUL*r and X86::IMUL*r. Emit
    // it manually.
    if (BaseOpc == X86ISD::UMUL && !ResultReg) {
      static const uint16_t MULOpc[] =
        { X86::MUL8r, X86::MUL16r, X86::MUL32r, X86::MUL64r };
      static const MCPhysReg Reg[] = { X86::AL, X86::AX, X86::EAX, X86::RAX };
      // First copy the first operand into RAX, which is an implicit input to
      // the X86::MUL*r instruction.
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::COPY), Reg[VT.SimpleTy-MVT::i8])
        .addReg(LHSReg, getKillRegState(LHSIsKill));
      ResultReg = fastEmitInst_r(MULOpc[VT.SimpleTy-MVT::i8],
                                 TLI.getRegClassFor(VT), RHSReg, RHSIsKill);
    } else if (BaseOpc == X86ISD::SMUL && !ResultReg) {
      static const uint16_t MULOpc[] =
        { X86::IMUL8r, X86::IMUL16rr, X86::IMUL32rr, X86::IMUL64rr };
      if (VT == MVT::i8) {
        // Copy the first operand into AL, which is an implicit input to the
        // X86::IMUL8r instruction.
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
               TII.get(TargetOpcode::COPY), X86::AL)
          .addReg(LHSReg, getKillRegState(LHSIsKill));
        ResultReg = fastEmitInst_r(MULOpc[0], TLI.getRegClassFor(VT), RHSReg,
                                   RHSIsKill);
      } else
        ResultReg = fastEmitInst_rr(MULOpc[VT.SimpleTy-MVT::i8],
                                    TLI.getRegClassFor(VT), LHSReg, LHSIsKill,
                                    RHSReg, RHSIsKill);
    }

    if (!ResultReg)
      return false;

    // Assign to a GPR since the overflow return value is lowered to a SETcc.
    Register ResultReg2 = createResultReg(&X86::GR8RegClass);
    assert((ResultReg+1) == ResultReg2 && "Nonconsecutive result registers.");
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::SETCCr),
            ResultReg2).addImm(CondCode);

    updateValueMap(II, ResultReg, 2);
    return true;
  }
  case Intrinsic::x86_sse_cvttss2si:
  case Intrinsic::x86_sse_cvttss2si64:
  case Intrinsic::x86_sse2_cvttsd2si:
  case Intrinsic::x86_sse2_cvttsd2si64: {
    bool IsInputDouble;
    switch (II->getIntrinsicID()) {
    default: llvm_unreachable("Unexpected intrinsic.");
    case Intrinsic::x86_sse_cvttss2si:
    case Intrinsic::x86_sse_cvttss2si64:
      if (!Subtarget->hasSSE1())
        return false;
      IsInputDouble = false;
      break;
    case Intrinsic::x86_sse2_cvttsd2si:
    case Intrinsic::x86_sse2_cvttsd2si64:
      if (!Subtarget->hasSSE2())
        return false;
      IsInputDouble = true;
      break;
    }

    Type *RetTy = II->getCalledFunction()->getReturnType();
    MVT VT;
    if (!isTypeLegal(RetTy, VT))
      return false;

    static const uint16_t CvtOpc[3][2][2] = {
      { { X86::CVTTSS2SIrr,   X86::CVTTSS2SI64rr },
        { X86::CVTTSD2SIrr,   X86::CVTTSD2SI64rr } },
      { { X86::VCVTTSS2SIrr,  X86::VCVTTSS2SI64rr },
        { X86::VCVTTSD2SIrr,  X86::VCVTTSD2SI64rr } },
      { { X86::VCVTTSS2SIZrr, X86::VCVTTSS2SI64Zrr },
        { X86::VCVTTSD2SIZrr, X86::VCVTTSD2SI64Zrr } },
    };
    unsigned AVXLevel = Subtarget->hasAVX512() ? 2 :
                        Subtarget->hasAVX()    ? 1 :
                                                 0;
    unsigned Opc;
    switch (VT.SimpleTy) {
    default: llvm_unreachable("Unexpected result type.");
    case MVT::i32: Opc = CvtOpc[AVXLevel][IsInputDouble][0]; break;
    case MVT::i64: Opc = CvtOpc[AVXLevel][IsInputDouble][1]; break;
    }

    // Check if we can fold insertelement instructions into the convert.
    const Value *Op = II->getArgOperand(0);
    while (auto *IE = dyn_cast<InsertElementInst>(Op)) {
      const Value *Index = IE->getOperand(2);
      if (!isa<ConstantInt>(Index))
        break;
      unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();

      if (Idx == 0) {
        Op = IE->getOperand(1);
        break;
      }
      Op = IE->getOperand(0);
    }

    Register Reg = getRegForValue(Op);
    if (Reg == 0)
      return false;

    Register ResultReg = createResultReg(TLI.getRegClassFor(VT));
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
      .addReg(Reg);

    updateValueMap(II, ResultReg);
    return true;
  }
  }
}

bool X86FastISel::fastLowerArguments() {
  if (!FuncInfo.CanLowerReturn)
    return false;

  const Function *F = FuncInfo.Fn;
  if (F->isVarArg())
    return false;

  CallingConv::ID CC = F->getCallingConv();
  if (CC != CallingConv::C)
    return false;

  if (Subtarget->isCallingConvWin64(CC))
    return false;

  if (!Subtarget->is64Bit())
    return false;

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

  // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments.
  unsigned GPRCnt = 0;
  unsigned FPRCnt = 0;
  for (auto const &Arg : F->args()) {
    if (Arg.hasAttribute(Attribute::ByVal) ||
        Arg.hasAttribute(Attribute::InReg) ||
        Arg.hasAttribute(Attribute::StructRet) ||
        Arg.hasAttribute(Attribute::SwiftSelf) ||
        Arg.hasAttribute(Attribute::SwiftError) ||
        Arg.hasAttribute(Attribute::Nest))
      return false;

    Type *ArgTy = Arg.getType();
    if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
      return false;

    EVT ArgVT = TLI.getValueType(DL, ArgTy);
    if (!ArgVT.isSimple()) return false;
    switch (ArgVT.getSimpleVT().SimpleTy) {
    default: return false;
    case MVT::i32:
    case MVT::i64:
      ++GPRCnt;
      break;
    case MVT::f32:
    case MVT::f64:
      if (!Subtarget->hasSSE1())
        return false;
      ++FPRCnt;
      break;
    }

    if (GPRCnt > 6)
      return false;

    if (FPRCnt > 8)
      return false;
  }

  static const MCPhysReg GPR32ArgRegs[] = {
    X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
  };
  static const MCPhysReg GPR64ArgRegs[] = {
    X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9
  };
  static const MCPhysReg XMMArgRegs[] = {
    X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
    X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
  };

  unsigned GPRIdx = 0;
  unsigned FPRIdx = 0;
  for (auto const &Arg : F->args()) {
    MVT VT = TLI.getSimpleValueType(DL, Arg.getType());
    const TargetRegisterClass *RC = TLI.getRegClassFor(VT);
    unsigned SrcReg;
    switch (VT.SimpleTy) {
    default: llvm_unreachable("Unexpected value type.");
    case MVT::i32: SrcReg = GPR32ArgRegs[GPRIdx++]; break;
    case MVT::i64: SrcReg = GPR64ArgRegs[GPRIdx++]; break;
    case MVT::f32: LLVM_FALLTHROUGH;
    case MVT::f64: SrcReg = XMMArgRegs[FPRIdx++]; break;
    }
    Register DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
    // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
    // Without this, EmitLiveInCopies may eliminate the livein if its only
    // use is a bitcast (which isn't turned into an instruction).
    Register ResultReg = createResultReg(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), ResultReg)
      .addReg(DstReg, getKillRegState(true));
    updateValueMap(&Arg, ResultReg);
  }
  return true;
}

static unsigned computeBytesPoppedByCalleeForSRet(const X86Subtarget *Subtarget,
                                                  CallingConv::ID CC,
                                                  const CallBase *CB) {
  if (Subtarget->is64Bit())
    return 0;
  if (Subtarget->getTargetTriple().isOSMSVCRT())
    return 0;
  if (CC == CallingConv::Fast || CC == CallingConv::GHC ||
      CC == CallingConv::HiPE || CC == CallingConv::Tail)
    return 0;

  if (CB)
    if (CB->arg_empty() || !CB->paramHasAttr(0, Attribute::StructRet) ||
        CB->paramHasAttr(0, Attribute::InReg) || Subtarget->isTargetMCU())
      return 0;

  return 4;
}

bool X86FastISel::fastLowerCall(CallLoweringInfo &CLI) {
  auto &OutVals       = CLI.OutVals;
  auto &OutFlags      = CLI.OutFlags;
  auto &OutRegs       = CLI.OutRegs;
  auto &Ins           = CLI.Ins;
  auto &InRegs        = CLI.InRegs;
  CallingConv::ID CC  = CLI.CallConv;
  bool &IsTailCall    = CLI.IsTailCall;
  bool IsVarArg       = CLI.IsVarArg;
  const Value *Callee = CLI.Callee;
  MCSymbol *Symbol = CLI.Symbol;

  bool Is64Bit        = Subtarget->is64Bit();
  bool IsWin64        = Subtarget->isCallingConvWin64(CC);

  const CallInst *CI = dyn_cast_or_null<CallInst>(CLI.CB);
  const Function *CalledFn = CI ? CI->getCalledFunction() : nullptr;

  // Call / invoke instructions with NoCfCheck attribute require special
  // handling.
  const auto *II = dyn_cast_or_null<InvokeInst>(CLI.CB);
  if ((CI && CI->doesNoCfCheck()) || (II && II->doesNoCfCheck()))
    return false;

  // Functions with no_caller_saved_registers that need special handling.
  if ((CI && CI->hasFnAttr("no_caller_saved_registers")) ||
      (CalledFn && CalledFn->hasFnAttribute("no_caller_saved_registers")))
    return false;

  // Functions using thunks for indirect calls need to use SDISel.
  if (Subtarget->useIndirectThunkCalls())
    return false;

  // Handle only C, fastcc, and webkit_js calling conventions for now.
  switch (CC) {
  default: return false;
  case CallingConv::C:
  case CallingConv::Fast:
  case CallingConv::Tail:
  case CallingConv::WebKit_JS:
  case CallingConv::Swift:
  case CallingConv::X86_FastCall:
  case CallingConv::X86_StdCall:
  case CallingConv::X86_ThisCall:
  case CallingConv::Win64:
  case CallingConv::X86_64_SysV:
  case CallingConv::CFGuard_Check:
    break;
  }

  // Allow SelectionDAG isel to handle tail calls.
  if (IsTailCall)
    return false;

  // fastcc with -tailcallopt is intended to provide a guaranteed
  // tail call optimization. Fastisel doesn't know how to do that.
  if ((CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt) ||
      CC == CallingConv::Tail)
    return false;

  // Don't know how to handle Win64 varargs yet.  Nothing special needed for
  // x86-32. Special handling for x86-64 is implemented.
  if (IsVarArg && IsWin64)
    return false;

  // Don't know about inalloca yet.
  if (CLI.CB && CLI.CB->hasInAllocaArgument())
    return false;

  for (auto Flag : CLI.OutFlags)
    if (Flag.isSwiftError() || Flag.isPreallocated())
      return false;

  SmallVector<MVT, 16> OutVTs;
  SmallVector<unsigned, 16> ArgRegs;

  // If this is a constant i1/i8/i16 argument, promote to i32 to avoid an extra
  // instruction. This is safe because it is common to all FastISel supported
  // calling conventions on x86.
  for (int i = 0, e = OutVals.size(); i != e; ++i) {
    Value *&Val = OutVals[i];
    ISD::ArgFlagsTy Flags = OutFlags[i];
    if (auto *CI = dyn_cast<ConstantInt>(Val)) {
      if (CI->getBitWidth() < 32) {
        if (Flags.isSExt())
          Val = ConstantExpr::getSExt(CI, Type::getInt32Ty(CI->getContext()));
        else
          Val = ConstantExpr::getZExt(CI, Type::getInt32Ty(CI->getContext()));
      }
    }

    // Passing bools around ends up doing a trunc to i1 and passing it.
    // Codegen this as an argument + "and 1".
    MVT VT;
    auto *TI = dyn_cast<TruncInst>(Val);
    unsigned ResultReg;
    if (TI && TI->getType()->isIntegerTy(1) && CLI.CB &&
        (TI->getParent() == CLI.CB->getParent()) && TI->hasOneUse()) {
      Value *PrevVal = TI->getOperand(0);
      ResultReg = getRegForValue(PrevVal);

      if (!ResultReg)
        return false;

      if (!isTypeLegal(PrevVal->getType(), VT))
        return false;

      ResultReg =
        fastEmit_ri(VT, VT, ISD::AND, ResultReg, hasTrivialKill(PrevVal), 1);
    } else {
      if (!isTypeLegal(Val->getType(), VT) ||
          (VT.isVector() && VT.getVectorElementType() == MVT::i1))
        return false;
      ResultReg = getRegForValue(Val);
    }

    if (!ResultReg)
      return false;

    ArgRegs.push_back(ResultReg);
    OutVTs.push_back(VT);
  }

  // Analyze operands of the call, assigning locations to each operand.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, ArgLocs, CLI.RetTy->getContext());

  // Allocate shadow area for Win64
  if (IsWin64)
    CCInfo.AllocateStack(32, Align(8));

  CCInfo.AnalyzeCallOperands(OutVTs, OutFlags, CC_X86);

  // Get a count of how many bytes are to be pushed on the stack.
  unsigned NumBytes = CCInfo.getAlignedCallFrameSize();

  // Issue CALLSEQ_START
  unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
    .addImm(NumBytes).addImm(0).addImm(0);

  // Walk the register/memloc assignments, inserting copies/loads.
  const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
    CCValAssign const &VA = ArgLocs[i];
    const Value *ArgVal = OutVals[VA.getValNo()];
    MVT ArgVT = OutVTs[VA.getValNo()];

    if (ArgVT == MVT::x86mmx)
      return false;

    unsigned ArgReg = ArgRegs[VA.getValNo()];

    // Promote the value if needed.
    switch (VA.getLocInfo()) {
    case CCValAssign::Full: break;
    case CCValAssign::SExt: {
      assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
             "Unexpected extend");

      if (ArgVT == MVT::i1)
        return false;

      bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), ArgReg,
                                       ArgVT, ArgReg);
      assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
      ArgVT = VA.getLocVT();
      break;
    }
    case CCValAssign::ZExt: {
      assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
             "Unexpected extend");

      // Handle zero-extension from i1 to i8, which is common.
      if (ArgVT == MVT::i1) {
        // Set the high bits to zero.
        ArgReg = fastEmitZExtFromI1(MVT::i8, ArgReg, /*TODO: Kill=*/false);
        ArgVT = MVT::i8;

        if (ArgReg == 0)
          return false;
      }

      bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), ArgReg,
                                       ArgVT, ArgReg);
      assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
      ArgVT = VA.getLocVT();
      break;
    }
    case CCValAssign::AExt: {
      assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
             "Unexpected extend");
      bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(), ArgReg,
                                       ArgVT, ArgReg);
      if (!Emitted)
        Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(), ArgReg,
                                    ArgVT, ArgReg);
      if (!Emitted)
        Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(), ArgReg,
                                    ArgVT, ArgReg);

      assert(Emitted && "Failed to emit a aext!"); (void)Emitted;
      ArgVT = VA.getLocVT();
      break;
    }
    case CCValAssign::BCvt: {
      ArgReg = fastEmit_r(ArgVT, VA.getLocVT(), ISD::BITCAST, ArgReg,
                          /*TODO: Kill=*/false);
      assert(ArgReg && "Failed to emit a bitcast!");
      ArgVT = VA.getLocVT();
      break;
    }
    case CCValAssign::VExt:
      // VExt has not been implemented, so this should be impossible to reach
      // for now.  However, fallback to Selection DAG isel once implemented.
      return false;
    case CCValAssign::AExtUpper:
    case CCValAssign::SExtUpper:
    case CCValAssign::ZExtUpper:
    case CCValAssign::FPExt:
    case CCValAssign::Trunc:
      llvm_unreachable("Unexpected loc info!");
    case CCValAssign::Indirect:
      // FIXME: Indirect doesn't need extending, but fast-isel doesn't fully
      // support this.
      return false;
    }

    if (VA.isRegLoc()) {
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
      OutRegs.push_back(VA.getLocReg());
    } else {
      assert(VA.isMemLoc() && "Unknown value location!");

      // Don't emit stores for undef values.
      if (isa<UndefValue>(ArgVal))
        continue;

      unsigned LocMemOffset = VA.getLocMemOffset();
      X86AddressMode AM;
      AM.Base.Reg = RegInfo->getStackRegister();
      AM.Disp = LocMemOffset;
      ISD::ArgFlagsTy Flags = OutFlags[VA.getValNo()];
      Align Alignment = DL.getABITypeAlign(ArgVal->getType());
      MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
          MachinePointerInfo::getStack(*FuncInfo.MF, LocMemOffset),
          MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);
      if (Flags.isByVal()) {
        X86AddressMode SrcAM;
        SrcAM.Base.Reg = ArgReg;
        if (!TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize()))
          return false;
      } else if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal)) {
        // If this is a really simple value, emit this with the Value* version
        // of X86FastEmitStore.  If it isn't simple, we don't want to do this,
        // as it can cause us to reevaluate the argument.
        if (!X86FastEmitStore(ArgVT, ArgVal, AM, MMO))
          return false;
      } else {
        bool ValIsKill = hasTrivialKill(ArgVal);
        if (!X86FastEmitStore(ArgVT, ArgReg, ValIsKill, AM, MMO))
          return false;
      }
    }
  }

  // ELF / PIC requires GOT in the EBX register before function calls via PLT
  // GOT pointer.
  if (Subtarget->isPICStyleGOT()) {
    unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), X86::EBX).addReg(Base);
  }

  if (Is64Bit && IsVarArg && !IsWin64) {
    // From AMD64 ABI document:
    // For calls that may call functions that use varargs or stdargs
    // (prototype-less calls or calls to functions containing ellipsis (...) in
    // the declaration) %al is used as hidden argument to specify the number
    // of SSE registers used. The contents of %al do not need to match exactly
    // the number of registers, but must be an ubound on the number of SSE
    // registers used and is in the range 0 - 8 inclusive.

    // Count the number of XMM registers allocated.
    static const MCPhysReg XMMArgRegs[] = {
      X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
      X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
    };
    unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
    assert((Subtarget->hasSSE1() || !NumXMMRegs)
           && "SSE registers cannot be used when SSE is disabled");
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV8ri),
            X86::AL).addImm(NumXMMRegs);
  }

  // Materialize callee address in a register. FIXME: GV address can be
  // handled with a CALLpcrel32 instead.
  X86AddressMode CalleeAM;
  if (!X86SelectCallAddress(Callee, CalleeAM))
    return false;

  unsigned CalleeOp = 0;
  const GlobalValue *GV = nullptr;
  if (CalleeAM.GV != nullptr) {
    GV = CalleeAM.GV;
  } else if (CalleeAM.Base.Reg != 0) {
    CalleeOp = CalleeAM.Base.Reg;
  } else
    return false;

  // Issue the call.
  MachineInstrBuilder MIB;
  if (CalleeOp) {
    // Register-indirect call.
    unsigned CallOpc = Is64Bit ? X86::CALL64r : X86::CALL32r;
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc))
      .addReg(CalleeOp);
  } else {
    // Direct call.
    assert(GV && "Not a direct call");
    // See if we need any target-specific flags on the GV operand.
    unsigned char OpFlags = Subtarget->classifyGlobalFunctionReference(GV);

    // This will be a direct call, or an indirect call through memory for
    // NonLazyBind calls or dllimport calls.
    bool NeedLoad = OpFlags == X86II::MO_DLLIMPORT ||
                    OpFlags == X86II::MO_GOTPCREL ||
                    OpFlags == X86II::MO_COFFSTUB;
    unsigned CallOpc = NeedLoad
                           ? (Is64Bit ? X86::CALL64m : X86::CALL32m)
                           : (Is64Bit ? X86::CALL64pcrel32 : X86::CALLpcrel32);

    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(CallOpc));
    if (NeedLoad)
      MIB.addReg(Is64Bit ? X86::RIP : 0).addImm(1).addReg(0);
    if (Symbol)
      MIB.addSym(Symbol, OpFlags);
    else
      MIB.addGlobalAddress(GV, 0, OpFlags);
    if (NeedLoad)
      MIB.addReg(0);
  }

  // Add a register mask operand representing the call-preserved registers.
  // Proper defs for return values will be added by setPhysRegsDeadExcept().
  MIB.addRegMask(TRI.getCallPreservedMask(*FuncInfo.MF, CC));

  // Add an implicit use GOT pointer in EBX.
  if (Subtarget->isPICStyleGOT())
    MIB.addReg(X86::EBX, RegState::Implicit);

  if (Is64Bit && IsVarArg && !IsWin64)
    MIB.addReg(X86::AL, RegState::Implicit);

  // Add implicit physical register uses to the call.
  for (auto Reg : OutRegs)
    MIB.addReg(Reg, RegState::Implicit);

  // Issue CALLSEQ_END
  unsigned NumBytesForCalleeToPop =
      X86::isCalleePop(CC, Subtarget->is64Bit(), IsVarArg,
                       TM.Options.GuaranteedTailCallOpt)
          ? NumBytes // Callee pops everything.
          : computeBytesPoppedByCalleeForSRet(Subtarget, CC, CLI.CB);
  unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackUp))
    .addImm(NumBytes).addImm(NumBytesForCalleeToPop);

  // Now handle call return values.
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCRetInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs,
                    CLI.RetTy->getContext());
  CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86);

  // Copy all of the result registers out of their specified physreg.
  Register ResultReg = FuncInfo.CreateRegs(CLI.RetTy);
  for (unsigned i = 0; i != RVLocs.size(); ++i) {
    CCValAssign &VA = RVLocs[i];
    EVT CopyVT = VA.getValVT();
    unsigned CopyReg = ResultReg + i;
    Register SrcReg = VA.getLocReg();

    // If this is x86-64, and we disabled SSE, we can't return FP values
    if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
        ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
      report_fatal_error("SSE register return with SSE disabled");
    }

    // If we prefer to use the value in xmm registers, copy it out as f80 and
    // use a truncate to move it from fp stack reg to xmm reg.
    if ((SrcReg == X86::FP0 || SrcReg == X86::FP1) &&
        isScalarFPTypeInSSEReg(VA.getValVT())) {
      CopyVT = MVT::f80;
      CopyReg = createResultReg(&X86::RFP80RegClass);
    }

    // Copy out the result.
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), CopyReg).addReg(SrcReg);
    InRegs.push_back(VA.getLocReg());

    // Round the f80 to the right size, which also moves it to the appropriate
    // xmm register. This is accomplished by storing the f80 value in memory
    // and then loading it back.
    if (CopyVT != VA.getValVT()) {
      EVT ResVT = VA.getValVT();
      unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
      unsigned MemSize = ResVT.getSizeInBits()/8;
      int FI = MFI.CreateStackObject(MemSize, Align(MemSize), false);
      addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                                TII.get(Opc)), FI)
        .addReg(CopyReg);
      Opc = ResVT == MVT::f32 ? X86::MOVSSrm_alt : X86::MOVSDrm_alt;
      addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                                TII.get(Opc), ResultReg + i), FI);
    }
  }

  CLI.ResultReg = ResultReg;
  CLI.NumResultRegs = RVLocs.size();
  CLI.Call = MIB;

  return true;
}

bool
X86FastISel::fastSelectInstruction(const Instruction *I)  {
  switch (I->getOpcode()) {
  default: break;
  case Instruction::Load:
    return X86SelectLoad(I);
  case Instruction::Store:
    return X86SelectStore(I);
  case Instruction::Ret:
    return X86SelectRet(I);
  case Instruction::ICmp:
  case Instruction::FCmp:
    return X86SelectCmp(I);
  case Instruction::ZExt:
    return X86SelectZExt(I);
  case Instruction::SExt:
    return X86SelectSExt(I);
  case Instruction::Br:
    return X86SelectBranch(I);
  case Instruction::LShr:
  case Instruction::AShr:
  case Instruction::Shl:
    return X86SelectShift(I);
  case Instruction::SDiv:
  case Instruction::UDiv:
  case Instruction::SRem:
  case Instruction::URem:
    return X86SelectDivRem(I);
  case Instruction::Select:
    return X86SelectSelect(I);
  case Instruction::Trunc:
    return X86SelectTrunc(I);
  case Instruction::FPExt:
    return X86SelectFPExt(I);
  case Instruction::FPTrunc:
    return X86SelectFPTrunc(I);
  case Instruction::SIToFP:
    return X86SelectSIToFP(I);
  case Instruction::UIToFP:
    return X86SelectUIToFP(I);
  case Instruction::IntToPtr: // Deliberate fall-through.
  case Instruction::PtrToInt: {
    EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType());
    EVT DstVT = TLI.getValueType(DL, I->getType());
    if (DstVT.bitsGT(SrcVT))
      return X86SelectZExt(I);
    if (DstVT.bitsLT(SrcVT))
      return X86SelectTrunc(I);
    Register Reg = getRegForValue(I->getOperand(0));
    if (Reg == 0) return false;
    updateValueMap(I, Reg);
    return true;
  }
  case Instruction::BitCast: {
    // Select SSE2/AVX bitcasts between 128/256/512 bit vector types.
    if (!Subtarget->hasSSE2())
      return false;

    MVT SrcVT, DstVT;
    if (!isTypeLegal(I->getOperand(0)->getType(), SrcVT) ||
        !isTypeLegal(I->getType(), DstVT))
      return false;

    // Only allow vectors that use xmm/ymm/zmm.
    if (!SrcVT.isVector() || !DstVT.isVector() ||
        SrcVT.getVectorElementType() == MVT::i1 ||
        DstVT.getVectorElementType() == MVT::i1)
      return false;

    Register Reg = getRegForValue(I->getOperand(0));
    if (!Reg)
      return false;

    // Emit a reg-reg copy so we don't propagate cached known bits information
    // with the wrong VT if we fall out of fast isel after selecting this.
    const TargetRegisterClass *DstClass = TLI.getRegClassFor(DstVT);
    Register ResultReg = createResultReg(DstClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::COPY), ResultReg).addReg(Reg);

    updateValueMap(I, ResultReg);
    return true;
  }
  }

  return false;
}

unsigned X86FastISel::X86MaterializeInt(const ConstantInt *CI, MVT VT) {
  if (VT > MVT::i64)
    return 0;

  uint64_t Imm = CI->getZExtValue();
  if (Imm == 0) {
    Register SrcReg = fastEmitInst_(X86::MOV32r0, &X86::GR32RegClass);
    switch (VT.SimpleTy) {
    default: llvm_unreachable("Unexpected value type");
    case MVT::i1:
    case MVT::i8:
      return fastEmitInst_extractsubreg(MVT::i8, SrcReg, /*Op0IsKill=*/true,
                                        X86::sub_8bit);
    case MVT::i16:
      return fastEmitInst_extractsubreg(MVT::i16, SrcReg, /*Op0IsKill=*/true,
                                        X86::sub_16bit);
    case MVT::i32:
      return SrcReg;
    case MVT::i64: {
      Register ResultReg = createResultReg(&X86::GR64RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
              TII.get(TargetOpcode::SUBREG_TO_REG), ResultReg)
        .addImm(0).addReg(SrcReg).addImm(X86::sub_32bit);
      return ResultReg;
    }
    }
  }

  unsigned Opc = 0;
  switch (VT.SimpleTy) {
  default: llvm_unreachable("Unexpected value type");
  case MVT::i1:
    VT = MVT::i8;
    LLVM_FALLTHROUGH;
  case MVT::i8:  Opc = X86::MOV8ri;  break;
  case MVT::i16: Opc = X86::MOV16ri; break;
  case MVT::i32: Opc = X86::MOV32ri; break;
  case MVT::i64: {
    if (isUInt<32>(Imm))
      Opc = X86::MOV32ri64;
    else if (isInt<32>(Imm))
      Opc = X86::MOV64ri32;
    else
      Opc = X86::MOV64ri;
    break;
  }
  }
  return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
}

unsigned X86FastISel::X86MaterializeFP(const ConstantFP *CFP, MVT VT) {
  if (CFP->isNullValue())
    return fastMaterializeFloatZero(CFP);

  // Can't handle alternate code models yet.
  CodeModel::Model CM = TM.getCodeModel();
  if (CM != CodeModel::Small && CM != CodeModel::Large)
    return 0;

  // Get opcode and regclass of the output for the given load instruction.
  unsigned Opc = 0;
  bool HasAVX = Subtarget->hasAVX();
  bool HasAVX512 = Subtarget->hasAVX512();
  switch (VT.SimpleTy) {
  default: return 0;
  case MVT::f32:
    if (X86ScalarSSEf32)
      Opc = HasAVX512 ? X86::VMOVSSZrm_alt :
            HasAVX    ? X86::VMOVSSrm_alt :
                        X86::MOVSSrm_alt;
    else
      Opc = X86::LD_Fp32m;
    break;
  case MVT::f64:
    if (X86ScalarSSEf64)
      Opc = HasAVX512 ? X86::VMOVSDZrm_alt :
            HasAVX    ? X86::VMOVSDrm_alt :
                        X86::MOVSDrm_alt;
    else
      Opc = X86::LD_Fp64m;
    break;
  case MVT::f80:
    // No f80 support yet.
    return 0;
  }

  // MachineConstantPool wants an explicit alignment.
  Align Alignment = DL.getPrefTypeAlign(CFP->getType());

  // x86-32 PIC requires a PIC base register for constant pools.
  unsigned PICBase = 0;
  unsigned char OpFlag = Subtarget->classifyLocalReference(nullptr);
  if (OpFlag == X86II::MO_PIC_BASE_OFFSET)
    PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
  else if (OpFlag == X86II::MO_GOTOFF)
    PICBase = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
  else if (Subtarget->is64Bit() && TM.getCodeModel() == CodeModel::Small)
    PICBase = X86::RIP;

  // Create the load from the constant pool.
  unsigned CPI = MCP.getConstantPoolIndex(CFP, Alignment);
  Register ResultReg = createResultReg(TLI.getRegClassFor(VT.SimpleTy));

  // Large code model only applies to 64-bit mode.
  if (Subtarget->is64Bit() && CM == CodeModel::Large) {
    Register AddrReg = createResultReg(&X86::GR64RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV64ri),
            AddrReg)
      .addConstantPoolIndex(CPI, 0, OpFlag);
    MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                                      TII.get(Opc), ResultReg);
    addRegReg(MIB, AddrReg, false, PICBase, false);
    MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
        MachinePointerInfo::getConstantPool(*FuncInfo.MF),
        MachineMemOperand::MOLoad, DL.getPointerSize(), Alignment);
    MIB->addMemOperand(*FuncInfo.MF, MMO);
    return ResultReg;
  }

  addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                                   TII.get(Opc), ResultReg),
                           CPI, PICBase, OpFlag);
  return ResultReg;
}

unsigned X86FastISel::X86MaterializeGV(const GlobalValue *GV, MVT VT) {
  // Can't handle alternate code models yet.
  if (TM.getCodeModel() != CodeModel::Small)
    return 0;

  // Materialize addresses with LEA/MOV instructions.
  X86AddressMode AM;
  if (X86SelectAddress(GV, AM)) {
    // If the expression is just a basereg, then we're done, otherwise we need
    // to emit an LEA.
    if (AM.BaseType == X86AddressMode::RegBase &&
        AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr)
      return AM.Base.Reg;

    Register ResultReg = createResultReg(TLI.getRegClassFor(VT));
    if (TM.getRelocationModel() == Reloc::Static &&
        TLI.getPointerTy(DL) == MVT::i64) {
      // The displacement code could be more than 32 bits away so we need to use
      // an instruction with a 64 bit immediate
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV64ri),
              ResultReg)
        .addGlobalAddress(GV);
    } else {
      unsigned Opc =
          TLI.getPointerTy(DL) == MVT::i32
              ? (Subtarget->isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r)
              : X86::LEA64r;
      addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                             TII.get(Opc), ResultReg), AM);
    }
    return ResultReg;
  }
  return 0;
}

unsigned X86FastISel::fastMaterializeConstant(const Constant *C) {
  EVT CEVT = TLI.getValueType(DL, C->getType(), true);

  // Only handle simple types.
  if (!CEVT.isSimple())
    return 0;
  MVT VT = CEVT.getSimpleVT();

  if (const auto *CI = dyn_cast<ConstantInt>(C))
    return X86MaterializeInt(CI, VT);
  else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
    return X86MaterializeFP(CFP, VT);
  else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
    return X86MaterializeGV(GV, VT);

  return 0;
}

unsigned X86FastISel::fastMaterializeAlloca(const AllocaInst *C) {
  // Fail on dynamic allocas. At this point, getRegForValue has already
  // checked its CSE maps, so if we're here trying to handle a dynamic
  // alloca, we're not going to succeed. X86SelectAddress has a
  // check for dynamic allocas, because it's called directly from
  // various places, but targetMaterializeAlloca also needs a check
  // in order to avoid recursion between getRegForValue,
  // X86SelectAddrss, and targetMaterializeAlloca.
  if (!FuncInfo.StaticAllocaMap.count(C))
    return 0;
  assert(C->isStaticAlloca() && "dynamic alloca in the static alloca map?");

  X86AddressMode AM;
  if (!X86SelectAddress(C, AM))
    return 0;
  unsigned Opc =
      TLI.getPointerTy(DL) == MVT::i32
          ? (Subtarget->isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r)
          : X86::LEA64r;
  const TargetRegisterClass *RC = TLI.getRegClassFor(TLI.getPointerTy(DL));
  Register ResultReg = createResultReg(RC);
  addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
                         TII.get(Opc), ResultReg), AM);
  return ResultReg;
}

unsigned X86FastISel::fastMaterializeFloatZero(const ConstantFP *CF) {
  MVT VT;
  if (!isTypeLegal(CF->getType(), VT))
    return 0;

  // Get opcode and regclass for the given zero.
  bool HasAVX512 = Subtarget->hasAVX512();
  unsigned Opc = 0;
  switch (VT.SimpleTy) {
  default: return 0;
  case MVT::f32:
    if (X86ScalarSSEf32)
      Opc = HasAVX512 ? X86::AVX512_FsFLD0SS : X86::FsFLD0SS;
    else
      Opc = X86::LD_Fp032;
    break;
  case MVT::f64:
    if (X86ScalarSSEf64)
      Opc = HasAVX512 ? X86::AVX512_FsFLD0SD : X86::FsFLD0SD;
    else
      Opc = X86::LD_Fp064;
    break;
  case MVT::f80:
    // No f80 support yet.
    return 0;
  }

  Register ResultReg = createResultReg(TLI.getRegClassFor(VT));
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg);
  return ResultReg;
}


bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
                                      const LoadInst *LI) {
  const Value *Ptr = LI->getPointerOperand();
  X86AddressMode AM;
  if (!X86SelectAddress(Ptr, AM))
    return false;

  const X86InstrInfo &XII = (const X86InstrInfo &)TII;

  unsigned Size = DL.getTypeAllocSize(LI->getType());

  SmallVector<MachineOperand, 8> AddrOps;
  AM.getFullAddress(AddrOps);

  MachineInstr *Result = XII.foldMemoryOperandImpl(
      *FuncInfo.MF, *MI, OpNo, AddrOps, FuncInfo.InsertPt, Size, LI->getAlign(),
      /*AllowCommute=*/true);
  if (!Result)
    return false;

  // The index register could be in the wrong register class.  Unfortunately,
  // foldMemoryOperandImpl could have commuted the instruction so its not enough
  // to just look at OpNo + the offset to the index reg.  We actually need to
  // scan the instruction to find the index reg and see if its the correct reg
  // class.
  unsigned OperandNo = 0;
  for (MachineInstr::mop_iterator I = Result->operands_begin(),
       E = Result->operands_end(); I != E; ++I, ++OperandNo) {
    MachineOperand &MO = *I;
    if (!MO.isReg() || MO.isDef() || MO.getReg() != AM.IndexReg)
      continue;
    // Found the index reg, now try to rewrite it.
    Register IndexReg = constrainOperandRegClass(Result->getDesc(),
                                                 MO.getReg(), OperandNo);
    if (IndexReg == MO.getReg())
      continue;
    MO.setReg(IndexReg);
  }

  Result->addMemOperand(*FuncInfo.MF, createMachineMemOperandFor(LI));
  Result->cloneInstrSymbols(*FuncInfo.MF, *MI);
  MachineBasicBlock::iterator I(MI);
  removeDeadCode(I, std::next(I));
  return true;
}

unsigned X86FastISel::fastEmitInst_rrrr(unsigned MachineInstOpcode,
                                        const TargetRegisterClass *RC,
                                        unsigned Op0, bool Op0IsKill,
                                        unsigned Op1, bool Op1IsKill,
                                        unsigned Op2, bool Op2IsKill,
                                        unsigned Op3, bool Op3IsKill) {
  const MCInstrDesc &II = TII.get(MachineInstOpcode);

  Register ResultReg = createResultReg(RC);
  Op0 = constrainOperandRegClass(II, Op0, II.getNumDefs());
  Op1 = constrainOperandRegClass(II, Op1, II.getNumDefs() + 1);
  Op2 = constrainOperandRegClass(II, Op2, II.getNumDefs() + 2);
  Op3 = constrainOperandRegClass(II, Op3, II.getNumDefs() + 3);

  if (II.getNumDefs() >= 1)
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II, ResultReg)
        .addReg(Op0, getKillRegState(Op0IsKill))
        .addReg(Op1, getKillRegState(Op1IsKill))
        .addReg(Op2, getKillRegState(Op2IsKill))
        .addReg(Op3, getKillRegState(Op3IsKill));
  else {
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, II)
        .addReg(Op0, getKillRegState(Op0IsKill))
        .addReg(Op1, getKillRegState(Op1IsKill))
        .addReg(Op2, getKillRegState(Op2IsKill))
        .addReg(Op3, getKillRegState(Op3IsKill));
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
            TII.get(TargetOpcode::COPY), ResultReg).addReg(II.ImplicitDefs[0]);
  }
  return ResultReg;
}


namespace llvm {
  FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo,
                                const TargetLibraryInfo *libInfo) {
    return new X86FastISel(funcInfo, libInfo);
  }
}