//===- X86InstructionSelector.cpp -----------------------------------------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
/// \file
/// This file implements the targeting of the InstructionSelector class for
/// X86.
/// \todo This should be generated by TableGen.
//===----------------------------------------------------------------------===//

#include "MCTargetDesc/X86BaseInfo.h"
#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86InstrInfo.h"
#include "X86RegisterBankInfo.h"
#include "X86RegisterInfo.h"
#include "X86Subtarget.h"
#include "X86TargetMachine.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelector.h"
#include "llvm/CodeGen/GlobalISel/InstructionSelectorImpl.h"
#include "llvm/CodeGen/GlobalISel/RegisterBank.h"
#include "llvm/CodeGen/GlobalISel/Utils.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/TargetOpcodes.h"
#include "llvm/CodeGen/TargetRegisterInfo.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/IntrinsicsX86.h"
#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/LowLevelTypeImpl.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include <cassert>
#include <cstdint>
#include <tuple>

#define DEBUG_TYPE "X86-isel"

using namespace llvm;

namespace {

#define GET_GLOBALISEL_PREDICATE_BITSET
#include "X86GenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATE_BITSET

class X86InstructionSelector : public InstructionSelector {
public:
  X86InstructionSelector(const X86TargetMachine &TM, const X86Subtarget &STI,
                         const X86RegisterBankInfo &RBI);

  bool select(MachineInstr &I) override;
  static const char *getName() { return DEBUG_TYPE; }

private:
  /// tblgen-erated 'select' implementation, used as the initial selector for
  /// the patterns that don't require complex C++.
  bool selectImpl(MachineInstr &I, CodeGenCoverage &CoverageInfo) const;

  // TODO: remove after supported by Tablegen-erated instruction selection.
  unsigned getLoadStoreOp(const LLT &Ty, const RegisterBank &RB, unsigned Opc,
                          Align Alignment) const;

  bool selectLoadStoreOp(MachineInstr &I, MachineRegisterInfo &MRI,
                         MachineFunction &MF) const;
  bool selectFrameIndexOrGep(MachineInstr &I, MachineRegisterInfo &MRI,
                             MachineFunction &MF) const;
  bool selectGlobalValue(MachineInstr &I, MachineRegisterInfo &MRI,
                         MachineFunction &MF) const;
  bool selectConstant(MachineInstr &I, MachineRegisterInfo &MRI,
                      MachineFunction &MF) const;
  bool selectTruncOrPtrToInt(MachineInstr &I, MachineRegisterInfo &MRI,
                             MachineFunction &MF) const;
  bool selectZext(MachineInstr &I, MachineRegisterInfo &MRI,
                  MachineFunction &MF) const;
  bool selectAnyext(MachineInstr &I, MachineRegisterInfo &MRI,
                    MachineFunction &MF) const;
  bool selectCmp(MachineInstr &I, MachineRegisterInfo &MRI,
                 MachineFunction &MF) const;
  bool selectFCmp(MachineInstr &I, MachineRegisterInfo &MRI,
                  MachineFunction &MF) const;
  bool selectUadde(MachineInstr &I, MachineRegisterInfo &MRI,
                   MachineFunction &MF) const;
  bool selectCopy(MachineInstr &I, MachineRegisterInfo &MRI) const;
  bool selectUnmergeValues(MachineInstr &I, MachineRegisterInfo &MRI,
                           MachineFunction &MF);
  bool selectMergeValues(MachineInstr &I, MachineRegisterInfo &MRI,
                         MachineFunction &MF);
  bool selectInsert(MachineInstr &I, MachineRegisterInfo &MRI,
                    MachineFunction &MF) const;
  bool selectExtract(MachineInstr &I, MachineRegisterInfo &MRI,
                     MachineFunction &MF) const;
  bool selectCondBranch(MachineInstr &I, MachineRegisterInfo &MRI,
                        MachineFunction &MF) const;
  bool selectTurnIntoCOPY(MachineInstr &I, MachineRegisterInfo &MRI,
                          const unsigned DstReg,
                          const TargetRegisterClass *DstRC,
                          const unsigned SrcReg,
                          const TargetRegisterClass *SrcRC) const;
  bool materializeFP(MachineInstr &I, MachineRegisterInfo &MRI,
                     MachineFunction &MF) const;
  bool selectImplicitDefOrPHI(MachineInstr &I, MachineRegisterInfo &MRI) const;
  bool selectDivRem(MachineInstr &I, MachineRegisterInfo &MRI,
                    MachineFunction &MF) const;
  bool selectIntrinsicWSideEffects(MachineInstr &I, MachineRegisterInfo &MRI,
                                   MachineFunction &MF) const;

  // emit insert subreg instruction and insert it before MachineInstr &I
  bool emitInsertSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I,
                        MachineRegisterInfo &MRI, MachineFunction &MF) const;
  // emit extract subreg instruction and insert it before MachineInstr &I
  bool emitExtractSubreg(unsigned DstReg, unsigned SrcReg, MachineInstr &I,
                         MachineRegisterInfo &MRI, MachineFunction &MF) const;

  const TargetRegisterClass *getRegClass(LLT Ty, const RegisterBank &RB) const;
  const TargetRegisterClass *getRegClass(LLT Ty, unsigned Reg,
                                         MachineRegisterInfo &MRI) const;

  const X86TargetMachine &TM;
  const X86Subtarget &STI;
  const X86InstrInfo &TII;
  const X86RegisterInfo &TRI;
  const X86RegisterBankInfo &RBI;

#define GET_GLOBALISEL_PREDICATES_DECL
#include "X86GenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_DECL

#define GET_GLOBALISEL_TEMPORARIES_DECL
#include "X86GenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_DECL
};

} // end anonymous namespace

#define GET_GLOBALISEL_IMPL
#include "X86GenGlobalISel.inc"
#undef GET_GLOBALISEL_IMPL

X86InstructionSelector::X86InstructionSelector(const X86TargetMachine &TM,
                                               const X86Subtarget &STI,
                                               const X86RegisterBankInfo &RBI)
    : InstructionSelector(), TM(TM), STI(STI), TII(*STI.getInstrInfo()),
      TRI(*STI.getRegisterInfo()), RBI(RBI),
#define GET_GLOBALISEL_PREDICATES_INIT
#include "X86GenGlobalISel.inc"
#undef GET_GLOBALISEL_PREDICATES_INIT
#define GET_GLOBALISEL_TEMPORARIES_INIT
#include "X86GenGlobalISel.inc"
#undef GET_GLOBALISEL_TEMPORARIES_INIT
{
}

// FIXME: This should be target-independent, inferred from the types declared
// for each class in the bank.
const TargetRegisterClass *
X86InstructionSelector::getRegClass(LLT Ty, const RegisterBank &RB) const {
  if (RB.getID() == X86::GPRRegBankID) {
    if (Ty.getSizeInBits() <= 8)
      return &X86::GR8RegClass;
    if (Ty.getSizeInBits() == 16)
      return &X86::GR16RegClass;
    if (Ty.getSizeInBits() == 32)
      return &X86::GR32RegClass;
    if (Ty.getSizeInBits() == 64)
      return &X86::GR64RegClass;
  }
  if (RB.getID() == X86::VECRRegBankID) {
    if (Ty.getSizeInBits() == 32)
      return STI.hasAVX512() ? &X86::FR32XRegClass : &X86::FR32RegClass;
    if (Ty.getSizeInBits() == 64)
      return STI.hasAVX512() ? &X86::FR64XRegClass : &X86::FR64RegClass;
    if (Ty.getSizeInBits() == 128)
      return STI.hasAVX512() ? &X86::VR128XRegClass : &X86::VR128RegClass;
    if (Ty.getSizeInBits() == 256)
      return STI.hasAVX512() ? &X86::VR256XRegClass : &X86::VR256RegClass;
    if (Ty.getSizeInBits() == 512)
      return &X86::VR512RegClass;
  }

  llvm_unreachable("Unknown RegBank!");
}

const TargetRegisterClass *
X86InstructionSelector::getRegClass(LLT Ty, unsigned Reg,
                                    MachineRegisterInfo &MRI) const {
  const RegisterBank &RegBank = *RBI.getRegBank(Reg, MRI, TRI);
  return getRegClass(Ty, RegBank);
}

static unsigned getSubRegIndex(const TargetRegisterClass *RC) {
  unsigned SubIdx = X86::NoSubRegister;
  if (RC == &X86::GR32RegClass) {
    SubIdx = X86::sub_32bit;
  } else if (RC == &X86::GR16RegClass) {
    SubIdx = X86::sub_16bit;
  } else if (RC == &X86::GR8RegClass) {
    SubIdx = X86::sub_8bit;
  }

  return SubIdx;
}

static const TargetRegisterClass *getRegClassFromGRPhysReg(Register Reg) {
  assert(Reg.isPhysical());
  if (X86::GR64RegClass.contains(Reg))
    return &X86::GR64RegClass;
  if (X86::GR32RegClass.contains(Reg))
    return &X86::GR32RegClass;
  if (X86::GR16RegClass.contains(Reg))
    return &X86::GR16RegClass;
  if (X86::GR8RegClass.contains(Reg))
    return &X86::GR8RegClass;

  llvm_unreachable("Unknown RegClass for PhysReg!");
}

// Set X86 Opcode and constrain DestReg.
bool X86InstructionSelector::selectCopy(MachineInstr &I,
                                        MachineRegisterInfo &MRI) const {
  Register DstReg = I.getOperand(0).getReg();
  const unsigned DstSize = RBI.getSizeInBits(DstReg, MRI, TRI);
  const RegisterBank &DstRegBank = *RBI.getRegBank(DstReg, MRI, TRI);

  Register SrcReg = I.getOperand(1).getReg();
  const unsigned SrcSize = RBI.getSizeInBits(SrcReg, MRI, TRI);
  const RegisterBank &SrcRegBank = *RBI.getRegBank(SrcReg, MRI, TRI);

  if (DstReg.isPhysical()) {
    assert(I.isCopy() && "Generic operators do not allow physical registers");

    if (DstSize > SrcSize && SrcRegBank.getID() == X86::GPRRegBankID &&
        DstRegBank.getID() == X86::GPRRegBankID) {

      const TargetRegisterClass *SrcRC =
          getRegClass(MRI.getType(SrcReg), SrcRegBank);
      const TargetRegisterClass *DstRC = getRegClassFromGRPhysReg(DstReg);

      if (SrcRC != DstRC) {
        // This case can be generated by ABI lowering, performe anyext
        Register ExtSrc = MRI.createVirtualRegister(DstRC);
        BuildMI(*I.getParent(), I, I.getDebugLoc(),
                TII.get(TargetOpcode::SUBREG_TO_REG))
            .addDef(ExtSrc)
            .addImm(0)
            .addReg(SrcReg)
            .addImm(getSubRegIndex(SrcRC));

        I.getOperand(1).setReg(ExtSrc);
      }
    }

    return true;
  }

  assert((!SrcReg.isPhysical() || I.isCopy()) &&
         "No phys reg on generic operators");
  assert((DstSize == SrcSize ||
          // Copies are a mean to setup initial types, the number of
          // bits may not exactly match.
          (SrcReg.isPhysical() &&
           DstSize <= RBI.getSizeInBits(SrcReg, MRI, TRI))) &&
         "Copy with different width?!");

  const TargetRegisterClass *DstRC =
      getRegClass(MRI.getType(DstReg), DstRegBank);

  if (SrcRegBank.getID() == X86::GPRRegBankID &&
      DstRegBank.getID() == X86::GPRRegBankID && SrcSize > DstSize &&
      SrcReg.isPhysical()) {
    // Change the physical register to performe truncate.

    const TargetRegisterClass *SrcRC = getRegClassFromGRPhysReg(SrcReg);

    if (DstRC != SrcRC) {
      I.getOperand(1).setSubReg(getSubRegIndex(DstRC));
      I.getOperand(1).substPhysReg(SrcReg, TRI);
    }
  }

  // No need to constrain SrcReg. It will get constrained when
  // we hit another of its use or its defs.
  // Copies do not have constraints.
  const TargetRegisterClass *OldRC = MRI.getRegClassOrNull(DstReg);
  if (!OldRC || !DstRC->hasSubClassEq(OldRC)) {
    if (!RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
      LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
                        << " operand\n");
      return false;
    }
  }
  I.setDesc(TII.get(X86::COPY));
  return true;
}

bool X86InstructionSelector::select(MachineInstr &I) {
  assert(I.getParent() && "Instruction should be in a basic block!");
  assert(I.getParent()->getParent() && "Instruction should be in a function!");

  MachineBasicBlock &MBB = *I.getParent();
  MachineFunction &MF = *MBB.getParent();
  MachineRegisterInfo &MRI = MF.getRegInfo();

  unsigned Opcode = I.getOpcode();
  if (!isPreISelGenericOpcode(Opcode)) {
    // Certain non-generic instructions also need some special handling.

    if (Opcode == TargetOpcode::LOAD_STACK_GUARD)
      return false;

    if (I.isCopy())
      return selectCopy(I, MRI);

    return true;
  }

  assert(I.getNumOperands() == I.getNumExplicitOperands() &&
         "Generic instruction has unexpected implicit operands\n");

  if (selectImpl(I, *CoverageInfo))
    return true;

  LLVM_DEBUG(dbgs() << " C++ instruction selection: "; I.print(dbgs()));

  // TODO: This should be implemented by tblgen.
  switch (I.getOpcode()) {
  default:
    return false;
  case TargetOpcode::G_STORE:
  case TargetOpcode::G_LOAD:
    return selectLoadStoreOp(I, MRI, MF);
  case TargetOpcode::G_PTR_ADD:
  case TargetOpcode::G_FRAME_INDEX:
    return selectFrameIndexOrGep(I, MRI, MF);
  case TargetOpcode::G_GLOBAL_VALUE:
    return selectGlobalValue(I, MRI, MF);
  case TargetOpcode::G_CONSTANT:
    return selectConstant(I, MRI, MF);
  case TargetOpcode::G_FCONSTANT:
    return materializeFP(I, MRI, MF);
  case TargetOpcode::G_PTRTOINT:
  case TargetOpcode::G_TRUNC:
    return selectTruncOrPtrToInt(I, MRI, MF);
  case TargetOpcode::G_INTTOPTR:
    return selectCopy(I, MRI);
  case TargetOpcode::G_ZEXT:
    return selectZext(I, MRI, MF);
  case TargetOpcode::G_ANYEXT:
    return selectAnyext(I, MRI, MF);
  case TargetOpcode::G_ICMP:
    return selectCmp(I, MRI, MF);
  case TargetOpcode::G_FCMP:
    return selectFCmp(I, MRI, MF);
  case TargetOpcode::G_UADDE:
    return selectUadde(I, MRI, MF);
  case TargetOpcode::G_UNMERGE_VALUES:
    return selectUnmergeValues(I, MRI, MF);
  case TargetOpcode::G_MERGE_VALUES:
  case TargetOpcode::G_CONCAT_VECTORS:
    return selectMergeValues(I, MRI, MF);
  case TargetOpcode::G_EXTRACT:
    return selectExtract(I, MRI, MF);
  case TargetOpcode::G_INSERT:
    return selectInsert(I, MRI, MF);
  case TargetOpcode::G_BRCOND:
    return selectCondBranch(I, MRI, MF);
  case TargetOpcode::G_IMPLICIT_DEF:
  case TargetOpcode::G_PHI:
    return selectImplicitDefOrPHI(I, MRI);
  case TargetOpcode::G_SDIV:
  case TargetOpcode::G_UDIV:
  case TargetOpcode::G_SREM:
  case TargetOpcode::G_UREM:
    return selectDivRem(I, MRI, MF);
  case TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS:
    return selectIntrinsicWSideEffects(I, MRI, MF);
  }

  return false;
}

unsigned X86InstructionSelector::getLoadStoreOp(const LLT &Ty,
                                                const RegisterBank &RB,
                                                unsigned Opc,
                                                Align Alignment) const {
  bool Isload = (Opc == TargetOpcode::G_LOAD);
  bool HasAVX = STI.hasAVX();
  bool HasAVX512 = STI.hasAVX512();
  bool HasVLX = STI.hasVLX();

  if (Ty == LLT::scalar(8)) {
    if (X86::GPRRegBankID == RB.getID())
      return Isload ? X86::MOV8rm : X86::MOV8mr;
  } else if (Ty == LLT::scalar(16)) {
    if (X86::GPRRegBankID == RB.getID())
      return Isload ? X86::MOV16rm : X86::MOV16mr;
  } else if (Ty == LLT::scalar(32) || Ty == LLT::pointer(0, 32)) {
    if (X86::GPRRegBankID == RB.getID())
      return Isload ? X86::MOV32rm : X86::MOV32mr;
    if (X86::VECRRegBankID == RB.getID())
      return Isload ? (HasAVX512 ? X86::VMOVSSZrm_alt :
                       HasAVX    ? X86::VMOVSSrm_alt :
                                   X86::MOVSSrm_alt)
                    : (HasAVX512 ? X86::VMOVSSZmr :
                       HasAVX    ? X86::VMOVSSmr :
                                   X86::MOVSSmr);
  } else if (Ty == LLT::scalar(64) || Ty == LLT::pointer(0, 64)) {
    if (X86::GPRRegBankID == RB.getID())
      return Isload ? X86::MOV64rm : X86::MOV64mr;
    if (X86::VECRRegBankID == RB.getID())
      return Isload ? (HasAVX512 ? X86::VMOVSDZrm_alt :
                       HasAVX    ? X86::VMOVSDrm_alt :
                                   X86::MOVSDrm_alt)
                    : (HasAVX512 ? X86::VMOVSDZmr :
                       HasAVX    ? X86::VMOVSDmr :
                                   X86::MOVSDmr);
  } else if (Ty.isVector() && Ty.getSizeInBits() == 128) {
    if (Alignment >= Align(16))
      return Isload ? (HasVLX ? X86::VMOVAPSZ128rm
                              : HasAVX512
                                    ? X86::VMOVAPSZ128rm_NOVLX
                                    : HasAVX ? X86::VMOVAPSrm : X86::MOVAPSrm)
                    : (HasVLX ? X86::VMOVAPSZ128mr
                              : HasAVX512
                                    ? X86::VMOVAPSZ128mr_NOVLX
                                    : HasAVX ? X86::VMOVAPSmr : X86::MOVAPSmr);
    else
      return Isload ? (HasVLX ? X86::VMOVUPSZ128rm
                              : HasAVX512
                                    ? X86::VMOVUPSZ128rm_NOVLX
                                    : HasAVX ? X86::VMOVUPSrm : X86::MOVUPSrm)
                    : (HasVLX ? X86::VMOVUPSZ128mr
                              : HasAVX512
                                    ? X86::VMOVUPSZ128mr_NOVLX
                                    : HasAVX ? X86::VMOVUPSmr : X86::MOVUPSmr);
  } else if (Ty.isVector() && Ty.getSizeInBits() == 256) {
    if (Alignment >= Align(32))
      return Isload ? (HasVLX ? X86::VMOVAPSZ256rm
                              : HasAVX512 ? X86::VMOVAPSZ256rm_NOVLX
                                          : X86::VMOVAPSYrm)
                    : (HasVLX ? X86::VMOVAPSZ256mr
                              : HasAVX512 ? X86::VMOVAPSZ256mr_NOVLX
                                          : X86::VMOVAPSYmr);
    else
      return Isload ? (HasVLX ? X86::VMOVUPSZ256rm
                              : HasAVX512 ? X86::VMOVUPSZ256rm_NOVLX
                                          : X86::VMOVUPSYrm)
                    : (HasVLX ? X86::VMOVUPSZ256mr
                              : HasAVX512 ? X86::VMOVUPSZ256mr_NOVLX
                                          : X86::VMOVUPSYmr);
  } else if (Ty.isVector() && Ty.getSizeInBits() == 512) {
    if (Alignment >= Align(64))
      return Isload ? X86::VMOVAPSZrm : X86::VMOVAPSZmr;
    else
      return Isload ? X86::VMOVUPSZrm : X86::VMOVUPSZmr;
  }
  return Opc;
}

// Fill in an address from the given instruction.
static void X86SelectAddress(const MachineInstr &I,
                             const MachineRegisterInfo &MRI,
                             X86AddressMode &AM) {
  assert(I.getOperand(0).isReg() && "unsupported opperand.");
  assert(MRI.getType(I.getOperand(0).getReg()).isPointer() &&
         "unsupported type.");

  if (I.getOpcode() == TargetOpcode::G_PTR_ADD) {
    if (auto COff = getConstantVRegSExtVal(I.getOperand(2).getReg(), MRI)) {
      int64_t Imm = *COff;
      if (isInt<32>(Imm)) { // Check for displacement overflow.
        AM.Disp = static_cast<int32_t>(Imm);
        AM.Base.Reg = I.getOperand(1).getReg();
        return;
      }
    }
  } else if (I.getOpcode() == TargetOpcode::G_FRAME_INDEX) {
    AM.Base.FrameIndex = I.getOperand(1).getIndex();
    AM.BaseType = X86AddressMode::FrameIndexBase;
    return;
  }

  // Default behavior.
  AM.Base.Reg = I.getOperand(0).getReg();
}

bool X86InstructionSelector::selectLoadStoreOp(MachineInstr &I,
                                               MachineRegisterInfo &MRI,
                                               MachineFunction &MF) const {
  unsigned Opc = I.getOpcode();

  assert((Opc == TargetOpcode::G_STORE || Opc == TargetOpcode::G_LOAD) &&
         "unexpected instruction");

  const Register DefReg = I.getOperand(0).getReg();
  LLT Ty = MRI.getType(DefReg);
  const RegisterBank &RB = *RBI.getRegBank(DefReg, MRI, TRI);

  assert(I.hasOneMemOperand());
  auto &MemOp = **I.memoperands_begin();
  if (MemOp.isAtomic()) {
    // Note: for unordered operations, we rely on the fact the appropriate MMO
    // is already on the instruction we're mutating, and thus we don't need to
    // make any changes.  So long as we select an opcode which is capable of
    // loading or storing the appropriate size atomically, the rest of the
    // backend is required to respect the MMO state. 
    if (!MemOp.isUnordered()) {
      LLVM_DEBUG(dbgs() << "Atomic ordering not supported yet\n");
      return false;
    }
    if (MemOp.getAlign() < Ty.getSizeInBits() / 8) {
      LLVM_DEBUG(dbgs() << "Unaligned atomics not supported yet\n");
      return false;
    }
  }

  unsigned NewOpc = getLoadStoreOp(Ty, RB, Opc, MemOp.getAlign());
  if (NewOpc == Opc)
    return false;

  X86AddressMode AM;
  X86SelectAddress(*MRI.getVRegDef(I.getOperand(1).getReg()), MRI, AM);

  I.setDesc(TII.get(NewOpc));
  MachineInstrBuilder MIB(MF, I);
  if (Opc == TargetOpcode::G_LOAD) {
    I.RemoveOperand(1);
    addFullAddress(MIB, AM);
  } else {
    // G_STORE (VAL, Addr), X86Store instruction (Addr, VAL)
    I.RemoveOperand(1);
    I.RemoveOperand(0);
    addFullAddress(MIB, AM).addUse(DefReg);
  }
  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}

static unsigned getLeaOP(LLT Ty, const X86Subtarget &STI) {
  if (Ty == LLT::pointer(0, 64))
    return X86::LEA64r;
  else if (Ty == LLT::pointer(0, 32))
    return STI.isTarget64BitILP32() ? X86::LEA64_32r : X86::LEA32r;
  else
    llvm_unreachable("Can't get LEA opcode. Unsupported type.");
}

bool X86InstructionSelector::selectFrameIndexOrGep(MachineInstr &I,
                                                   MachineRegisterInfo &MRI,
                                                   MachineFunction &MF) const {
  unsigned Opc = I.getOpcode();

  assert((Opc == TargetOpcode::G_FRAME_INDEX || Opc == TargetOpcode::G_PTR_ADD) &&
         "unexpected instruction");

  const Register DefReg = I.getOperand(0).getReg();
  LLT Ty = MRI.getType(DefReg);

  // Use LEA to calculate frame index and GEP
  unsigned NewOpc = getLeaOP(Ty, STI);
  I.setDesc(TII.get(NewOpc));
  MachineInstrBuilder MIB(MF, I);

  if (Opc == TargetOpcode::G_FRAME_INDEX) {
    addOffset(MIB, 0);
  } else {
    MachineOperand &InxOp = I.getOperand(2);
    I.addOperand(InxOp);        // set IndexReg
    InxOp.ChangeToImmediate(1); // set Scale
    MIB.addImm(0).addReg(0);
  }

  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}

bool X86InstructionSelector::selectGlobalValue(MachineInstr &I,
                                               MachineRegisterInfo &MRI,
                                               MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_GLOBAL_VALUE) &&
         "unexpected instruction");

  auto GV = I.getOperand(1).getGlobal();
  if (GV->isThreadLocal()) {
    return false; // TODO: we don't support TLS yet.
  }

  // Can't handle alternate code models yet.
  if (TM.getCodeModel() != CodeModel::Small)
    return false;

  X86AddressMode AM;
  AM.GV = GV;
  AM.GVOpFlags = STI.classifyGlobalReference(GV);

  // TODO: The ABI requires an extra load. not supported yet.
  if (isGlobalStubReference(AM.GVOpFlags))
    return false;

  // TODO: This reference is relative to the pic base. not supported yet.
  if (isGlobalRelativeToPICBase(AM.GVOpFlags))
    return false;

  if (STI.isPICStyleRIPRel()) {
    // Use rip-relative addressing.
    assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
    AM.Base.Reg = X86::RIP;
  }

  const Register DefReg = I.getOperand(0).getReg();
  LLT Ty = MRI.getType(DefReg);
  unsigned NewOpc = getLeaOP(Ty, STI);

  I.setDesc(TII.get(NewOpc));
  MachineInstrBuilder MIB(MF, I);

  I.RemoveOperand(1);
  addFullAddress(MIB, AM);

  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}

bool X86InstructionSelector::selectConstant(MachineInstr &I,
                                            MachineRegisterInfo &MRI,
                                            MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_CONSTANT) &&
         "unexpected instruction");

  const Register DefReg = I.getOperand(0).getReg();
  LLT Ty = MRI.getType(DefReg);

  if (RBI.getRegBank(DefReg, MRI, TRI)->getID() != X86::GPRRegBankID)
    return false;

  uint64_t Val = 0;
  if (I.getOperand(1).isCImm()) {
    Val = I.getOperand(1).getCImm()->getZExtValue();
    I.getOperand(1).ChangeToImmediate(Val);
  } else if (I.getOperand(1).isImm()) {
    Val = I.getOperand(1).getImm();
  } else
    llvm_unreachable("Unsupported operand type.");

  unsigned NewOpc;
  switch (Ty.getSizeInBits()) {
  case 8:
    NewOpc = X86::MOV8ri;
    break;
  case 16:
    NewOpc = X86::MOV16ri;
    break;
  case 32:
    NewOpc = X86::MOV32ri;
    break;
  case 64:
    // TODO: in case isUInt<32>(Val), X86::MOV32ri can be used
    if (isInt<32>(Val))
      NewOpc = X86::MOV64ri32;
    else
      NewOpc = X86::MOV64ri;
    break;
  default:
    llvm_unreachable("Can't select G_CONSTANT, unsupported type.");
  }

  I.setDesc(TII.get(NewOpc));
  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}

// Helper function for selectTruncOrPtrToInt and selectAnyext.
// Returns true if DstRC lives on a floating register class and
// SrcRC lives on a 128-bit vector class.
static bool canTurnIntoCOPY(const TargetRegisterClass *DstRC,
                            const TargetRegisterClass *SrcRC) {
  return (DstRC == &X86::FR32RegClass || DstRC == &X86::FR32XRegClass ||
          DstRC == &X86::FR64RegClass || DstRC == &X86::FR64XRegClass) &&
         (SrcRC == &X86::VR128RegClass || SrcRC == &X86::VR128XRegClass);
}

bool X86InstructionSelector::selectTurnIntoCOPY(
    MachineInstr &I, MachineRegisterInfo &MRI, const unsigned DstReg,
    const TargetRegisterClass *DstRC, const unsigned SrcReg,
    const TargetRegisterClass *SrcRC) const {

  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
                      << " operand\n");
    return false;
  }
  I.setDesc(TII.get(X86::COPY));
  return true;
}

bool X86InstructionSelector::selectTruncOrPtrToInt(MachineInstr &I,
                                                   MachineRegisterInfo &MRI,
                                                   MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_TRUNC ||
          I.getOpcode() == TargetOpcode::G_PTRTOINT) &&
         "unexpected instruction");

  const Register DstReg = I.getOperand(0).getReg();
  const Register SrcReg = I.getOperand(1).getReg();

  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);

  const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI);
  const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI);

  if (DstRB.getID() != SrcRB.getID()) {
    LLVM_DEBUG(dbgs() << TII.getName(I.getOpcode())
                      << " input/output on different banks\n");
    return false;
  }

  const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB);
  const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB);

  if (!DstRC || !SrcRC)
    return false;

  // If that's truncation of the value that lives on the vector class and goes
  // into the floating class, just replace it with copy, as we are able to
  // select it as a regular move.
  if (canTurnIntoCOPY(DstRC, SrcRC))
    return selectTurnIntoCOPY(I, MRI, DstReg, DstRC, SrcReg, SrcRC);

  if (DstRB.getID() != X86::GPRRegBankID)
    return false;

  unsigned SubIdx;
  if (DstRC == SrcRC) {
    // Nothing to be done
    SubIdx = X86::NoSubRegister;
  } else if (DstRC == &X86::GR32RegClass) {
    SubIdx = X86::sub_32bit;
  } else if (DstRC == &X86::GR16RegClass) {
    SubIdx = X86::sub_16bit;
  } else if (DstRC == &X86::GR8RegClass) {
    SubIdx = X86::sub_8bit;
  } else {
    return false;
  }

  SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx);

  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
                      << "\n");
    return false;
  }

  I.getOperand(1).setSubReg(SubIdx);

  I.setDesc(TII.get(X86::COPY));
  return true;
}

bool X86InstructionSelector::selectZext(MachineInstr &I,
                                        MachineRegisterInfo &MRI,
                                        MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_ZEXT) && "unexpected instruction");

  const Register DstReg = I.getOperand(0).getReg();
  const Register SrcReg = I.getOperand(1).getReg();

  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);

  assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(16)) &&
         "8=>16 Zext is handled by tablegen");
  assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(32)) &&
         "8=>32 Zext is handled by tablegen");
  assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(32)) &&
         "16=>32 Zext is handled by tablegen");
  assert(!(SrcTy == LLT::scalar(8) && DstTy == LLT::scalar(64)) &&
         "8=>64 Zext is handled by tablegen");
  assert(!(SrcTy == LLT::scalar(16) && DstTy == LLT::scalar(64)) &&
         "16=>64 Zext is handled by tablegen");
  assert(!(SrcTy == LLT::scalar(32) && DstTy == LLT::scalar(64)) &&
         "32=>64 Zext is handled by tablegen");

  if (SrcTy != LLT::scalar(1))
    return false;

  unsigned AndOpc;
  if (DstTy == LLT::scalar(8))
    AndOpc = X86::AND8ri;
  else if (DstTy == LLT::scalar(16))
    AndOpc = X86::AND16ri8;
  else if (DstTy == LLT::scalar(32))
    AndOpc = X86::AND32ri8;
  else if (DstTy == LLT::scalar(64))
    AndOpc = X86::AND64ri8;
  else
    return false;

  Register DefReg = SrcReg;
  if (DstTy != LLT::scalar(8)) {
    Register ImpDefReg =
        MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI));
    BuildMI(*I.getParent(), I, I.getDebugLoc(),
            TII.get(TargetOpcode::IMPLICIT_DEF), ImpDefReg);

    DefReg = MRI.createVirtualRegister(getRegClass(DstTy, DstReg, MRI));
    BuildMI(*I.getParent(), I, I.getDebugLoc(),
            TII.get(TargetOpcode::INSERT_SUBREG), DefReg)
        .addReg(ImpDefReg)
        .addReg(SrcReg)
        .addImm(X86::sub_8bit);
  }

  MachineInstr &AndInst =
      *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(AndOpc), DstReg)
           .addReg(DefReg)
           .addImm(1);

  constrainSelectedInstRegOperands(AndInst, TII, TRI, RBI);

  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectAnyext(MachineInstr &I,
                                          MachineRegisterInfo &MRI,
                                          MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_ANYEXT) && "unexpected instruction");

  const Register DstReg = I.getOperand(0).getReg();
  const Register SrcReg = I.getOperand(1).getReg();

  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);

  const RegisterBank &DstRB = *RBI.getRegBank(DstReg, MRI, TRI);
  const RegisterBank &SrcRB = *RBI.getRegBank(SrcReg, MRI, TRI);

  assert(DstRB.getID() == SrcRB.getID() &&
         "G_ANYEXT input/output on different banks\n");

  assert(DstTy.getSizeInBits() > SrcTy.getSizeInBits() &&
         "G_ANYEXT incorrect operand size");

  const TargetRegisterClass *DstRC = getRegClass(DstTy, DstRB);
  const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcRB);

  // If that's ANY_EXT of the value that lives on the floating class and goes
  // into the vector class, just replace it with copy, as we are able to select
  // it as a regular move.
  if (canTurnIntoCOPY(SrcRC, DstRC))
    return selectTurnIntoCOPY(I, MRI, SrcReg, SrcRC, DstReg, DstRC);

  if (DstRB.getID() != X86::GPRRegBankID)
    return false;

  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
                      << " operand\n");
    return false;
  }

  if (SrcRC == DstRC) {
    I.setDesc(TII.get(X86::COPY));
    return true;
  }

  BuildMI(*I.getParent(), I, I.getDebugLoc(),
          TII.get(TargetOpcode::SUBREG_TO_REG))
      .addDef(DstReg)
      .addImm(0)
      .addReg(SrcReg)
      .addImm(getSubRegIndex(SrcRC));

  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectCmp(MachineInstr &I,
                                       MachineRegisterInfo &MRI,
                                       MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_ICMP) && "unexpected instruction");

  X86::CondCode CC;
  bool SwapArgs;
  std::tie(CC, SwapArgs) = X86::getX86ConditionCode(
      (CmpInst::Predicate)I.getOperand(1).getPredicate());

  Register LHS = I.getOperand(2).getReg();
  Register RHS = I.getOperand(3).getReg();

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

  unsigned OpCmp;
  LLT Ty = MRI.getType(LHS);

  switch (Ty.getSizeInBits()) {
  default:
    return false;
  case 8:
    OpCmp = X86::CMP8rr;
    break;
  case 16:
    OpCmp = X86::CMP16rr;
    break;
  case 32:
    OpCmp = X86::CMP32rr;
    break;
  case 64:
    OpCmp = X86::CMP64rr;
    break;
  }

  MachineInstr &CmpInst =
      *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
           .addReg(LHS)
           .addReg(RHS);

  MachineInstr &SetInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
                                   TII.get(X86::SETCCr), I.getOperand(0).getReg()).addImm(CC);

  constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
  constrainSelectedInstRegOperands(SetInst, TII, TRI, RBI);

  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectFCmp(MachineInstr &I,
                                        MachineRegisterInfo &MRI,
                                        MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_FCMP) && "unexpected instruction");

  Register LhsReg = I.getOperand(2).getReg();
  Register RhsReg = I.getOperand(3).getReg();
  CmpInst::Predicate Predicate =
      (CmpInst::Predicate)I.getOperand(1).getPredicate();

  // 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;
  }

  // Compute the opcode for the CMP instruction.
  unsigned OpCmp;
  LLT Ty = MRI.getType(LhsReg);
  switch (Ty.getSizeInBits()) {
  default:
    return false;
  case 32:
    OpCmp = X86::UCOMISSrr;
    break;
  case 64:
    OpCmp = X86::UCOMISDrr;
    break;
  }

  Register ResultReg = I.getOperand(0).getReg();
  RBI.constrainGenericRegister(
      ResultReg,
      *getRegClass(LLT::scalar(8), *RBI.getRegBank(ResultReg, MRI, TRI)), MRI);
  if (SETFOpc) {
    MachineInstr &CmpInst =
        *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
             .addReg(LhsReg)
             .addReg(RhsReg);

    Register FlagReg1 = MRI.createVirtualRegister(&X86::GR8RegClass);
    Register FlagReg2 = MRI.createVirtualRegister(&X86::GR8RegClass);
    MachineInstr &Set1 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
                                  TII.get(X86::SETCCr), FlagReg1).addImm(SETFOpc[0]);
    MachineInstr &Set2 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
                                  TII.get(X86::SETCCr), FlagReg2).addImm(SETFOpc[1]);
    MachineInstr &Set3 = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
                                  TII.get(SETFOpc[2]), ResultReg)
                              .addReg(FlagReg1)
                              .addReg(FlagReg2);
    constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
    constrainSelectedInstRegOperands(Set1, TII, TRI, RBI);
    constrainSelectedInstRegOperands(Set2, TII, TRI, RBI);
    constrainSelectedInstRegOperands(Set3, TII, TRI, RBI);

    I.eraseFromParent();
    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(LhsReg, RhsReg);

  // Emit a compare of LHS/RHS.
  MachineInstr &CmpInst =
      *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpCmp))
           .addReg(LhsReg)
           .addReg(RhsReg);

  MachineInstr &Set =
      *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SETCCr), ResultReg).addImm(CC);
  constrainSelectedInstRegOperands(CmpInst, TII, TRI, RBI);
  constrainSelectedInstRegOperands(Set, TII, TRI, RBI);
  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectUadde(MachineInstr &I,
                                         MachineRegisterInfo &MRI,
                                         MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_UADDE) && "unexpected instruction");

  const Register DstReg = I.getOperand(0).getReg();
  const Register CarryOutReg = I.getOperand(1).getReg();
  const Register Op0Reg = I.getOperand(2).getReg();
  const Register Op1Reg = I.getOperand(3).getReg();
  Register CarryInReg = I.getOperand(4).getReg();

  const LLT DstTy = MRI.getType(DstReg);

  if (DstTy != LLT::scalar(32))
    return false;

  // find CarryIn def instruction.
  MachineInstr *Def = MRI.getVRegDef(CarryInReg);
  while (Def->getOpcode() == TargetOpcode::G_TRUNC) {
    CarryInReg = Def->getOperand(1).getReg();
    Def = MRI.getVRegDef(CarryInReg);
  }

  unsigned Opcode;
  if (Def->getOpcode() == TargetOpcode::G_UADDE) {
    // carry set by prev ADD.

    BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), X86::EFLAGS)
        .addReg(CarryInReg);

    if (!RBI.constrainGenericRegister(CarryInReg, X86::GR32RegClass, MRI))
      return false;

    Opcode = X86::ADC32rr;
  } else if (auto val = getConstantVRegVal(CarryInReg, MRI)) {
    // carry is constant, support only 0.
    if (*val != 0)
      return false;

    Opcode = X86::ADD32rr;
  } else
    return false;

  MachineInstr &AddInst =
      *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Opcode), DstReg)
           .addReg(Op0Reg)
           .addReg(Op1Reg);

  BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), CarryOutReg)
      .addReg(X86::EFLAGS);

  if (!constrainSelectedInstRegOperands(AddInst, TII, TRI, RBI) ||
      !RBI.constrainGenericRegister(CarryOutReg, X86::GR32RegClass, MRI))
    return false;

  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectExtract(MachineInstr &I,
                                           MachineRegisterInfo &MRI,
                                           MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_EXTRACT) &&
         "unexpected instruction");

  const Register DstReg = I.getOperand(0).getReg();
  const Register SrcReg = I.getOperand(1).getReg();
  int64_t Index = I.getOperand(2).getImm();

  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);

  // Meanwile handle vector type only.
  if (!DstTy.isVector())
    return false;

  if (Index % DstTy.getSizeInBits() != 0)
    return false; // Not extract subvector.

  if (Index == 0) {
    // Replace by extract subreg copy.
    if (!emitExtractSubreg(DstReg, SrcReg, I, MRI, MF))
      return false;

    I.eraseFromParent();
    return true;
  }

  bool HasAVX = STI.hasAVX();
  bool HasAVX512 = STI.hasAVX512();
  bool HasVLX = STI.hasVLX();

  if (SrcTy.getSizeInBits() == 256 && DstTy.getSizeInBits() == 128) {
    if (HasVLX)
      I.setDesc(TII.get(X86::VEXTRACTF32x4Z256rr));
    else if (HasAVX)
      I.setDesc(TII.get(X86::VEXTRACTF128rr));
    else
      return false;
  } else if (SrcTy.getSizeInBits() == 512 && HasAVX512) {
    if (DstTy.getSizeInBits() == 128)
      I.setDesc(TII.get(X86::VEXTRACTF32x4Zrr));
    else if (DstTy.getSizeInBits() == 256)
      I.setDesc(TII.get(X86::VEXTRACTF64x4Zrr));
    else
      return false;
  } else
    return false;

  // Convert to X86 VEXTRACT immediate.
  Index = Index / DstTy.getSizeInBits();
  I.getOperand(2).setImm(Index);

  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}

bool X86InstructionSelector::emitExtractSubreg(unsigned DstReg, unsigned SrcReg,
                                               MachineInstr &I,
                                               MachineRegisterInfo &MRI,
                                               MachineFunction &MF) const {
  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);
  unsigned SubIdx = X86::NoSubRegister;

  if (!DstTy.isVector() || !SrcTy.isVector())
    return false;

  assert(SrcTy.getSizeInBits() > DstTy.getSizeInBits() &&
         "Incorrect Src/Dst register size");

  if (DstTy.getSizeInBits() == 128)
    SubIdx = X86::sub_xmm;
  else if (DstTy.getSizeInBits() == 256)
    SubIdx = X86::sub_ymm;
  else
    return false;

  const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI);
  const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI);

  SrcRC = TRI.getSubClassWithSubReg(SrcRC, SubIdx);

  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain EXTRACT_SUBREG\n");
    return false;
  }

  BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY), DstReg)
      .addReg(SrcReg, 0, SubIdx);

  return true;
}

bool X86InstructionSelector::emitInsertSubreg(unsigned DstReg, unsigned SrcReg,
                                              MachineInstr &I,
                                              MachineRegisterInfo &MRI,
                                              MachineFunction &MF) const {
  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg);
  unsigned SubIdx = X86::NoSubRegister;

  // TODO: support scalar types
  if (!DstTy.isVector() || !SrcTy.isVector())
    return false;

  assert(SrcTy.getSizeInBits() < DstTy.getSizeInBits() &&
         "Incorrect Src/Dst register size");

  if (SrcTy.getSizeInBits() == 128)
    SubIdx = X86::sub_xmm;
  else if (SrcTy.getSizeInBits() == 256)
    SubIdx = X86::sub_ymm;
  else
    return false;

  const TargetRegisterClass *SrcRC = getRegClass(SrcTy, SrcReg, MRI);
  const TargetRegisterClass *DstRC = getRegClass(DstTy, DstReg, MRI);

  if (!RBI.constrainGenericRegister(SrcReg, *SrcRC, MRI) ||
      !RBI.constrainGenericRegister(DstReg, *DstRC, MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain INSERT_SUBREG\n");
    return false;
  }

  BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::COPY))
      .addReg(DstReg, RegState::DefineNoRead, SubIdx)
      .addReg(SrcReg);

  return true;
}

bool X86InstructionSelector::selectInsert(MachineInstr &I,
                                          MachineRegisterInfo &MRI,
                                          MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_INSERT) && "unexpected instruction");

  const Register DstReg = I.getOperand(0).getReg();
  const Register SrcReg = I.getOperand(1).getReg();
  const Register InsertReg = I.getOperand(2).getReg();
  int64_t Index = I.getOperand(3).getImm();

  const LLT DstTy = MRI.getType(DstReg);
  const LLT InsertRegTy = MRI.getType(InsertReg);

  // Meanwile handle vector type only.
  if (!DstTy.isVector())
    return false;

  if (Index % InsertRegTy.getSizeInBits() != 0)
    return false; // Not insert subvector.

  if (Index == 0 && MRI.getVRegDef(SrcReg)->isImplicitDef()) {
    // Replace by subreg copy.
    if (!emitInsertSubreg(DstReg, InsertReg, I, MRI, MF))
      return false;

    I.eraseFromParent();
    return true;
  }

  bool HasAVX = STI.hasAVX();
  bool HasAVX512 = STI.hasAVX512();
  bool HasVLX = STI.hasVLX();

  if (DstTy.getSizeInBits() == 256 && InsertRegTy.getSizeInBits() == 128) {
    if (HasVLX)
      I.setDesc(TII.get(X86::VINSERTF32x4Z256rr));
    else if (HasAVX)
      I.setDesc(TII.get(X86::VINSERTF128rr));
    else
      return false;
  } else if (DstTy.getSizeInBits() == 512 && HasAVX512) {
    if (InsertRegTy.getSizeInBits() == 128)
      I.setDesc(TII.get(X86::VINSERTF32x4Zrr));
    else if (InsertRegTy.getSizeInBits() == 256)
      I.setDesc(TII.get(X86::VINSERTF64x4Zrr));
    else
      return false;
  } else
    return false;

  // Convert to X86 VINSERT immediate.
  Index = Index / InsertRegTy.getSizeInBits();

  I.getOperand(3).setImm(Index);

  return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
}

bool X86InstructionSelector::selectUnmergeValues(
    MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) {
  assert((I.getOpcode() == TargetOpcode::G_UNMERGE_VALUES) &&
         "unexpected instruction");

  // Split to extracts.
  unsigned NumDefs = I.getNumOperands() - 1;
  Register SrcReg = I.getOperand(NumDefs).getReg();
  unsigned DefSize = MRI.getType(I.getOperand(0).getReg()).getSizeInBits();

  for (unsigned Idx = 0; Idx < NumDefs; ++Idx) {
    MachineInstr &ExtrInst =
        *BuildMI(*I.getParent(), I, I.getDebugLoc(),
                 TII.get(TargetOpcode::G_EXTRACT), I.getOperand(Idx).getReg())
             .addReg(SrcReg)
             .addImm(Idx * DefSize);

    if (!select(ExtrInst))
      return false;
  }

  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectMergeValues(
    MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) {
  assert((I.getOpcode() == TargetOpcode::G_MERGE_VALUES ||
          I.getOpcode() == TargetOpcode::G_CONCAT_VECTORS) &&
         "unexpected instruction");

  // Split to inserts.
  Register DstReg = I.getOperand(0).getReg();
  Register SrcReg0 = I.getOperand(1).getReg();

  const LLT DstTy = MRI.getType(DstReg);
  const LLT SrcTy = MRI.getType(SrcReg0);
  unsigned SrcSize = SrcTy.getSizeInBits();

  const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI);

  // For the first src use insertSubReg.
  Register DefReg = MRI.createGenericVirtualRegister(DstTy);
  MRI.setRegBank(DefReg, RegBank);
  if (!emitInsertSubreg(DefReg, I.getOperand(1).getReg(), I, MRI, MF))
    return false;

  for (unsigned Idx = 2; Idx < I.getNumOperands(); ++Idx) {
    Register Tmp = MRI.createGenericVirtualRegister(DstTy);
    MRI.setRegBank(Tmp, RegBank);

    MachineInstr &InsertInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
                                        TII.get(TargetOpcode::G_INSERT), Tmp)
                                    .addReg(DefReg)
                                    .addReg(I.getOperand(Idx).getReg())
                                    .addImm((Idx - 1) * SrcSize);

    DefReg = Tmp;

    if (!select(InsertInst))
      return false;
  }

  MachineInstr &CopyInst = *BuildMI(*I.getParent(), I, I.getDebugLoc(),
                                    TII.get(TargetOpcode::COPY), DstReg)
                                .addReg(DefReg);

  if (!select(CopyInst))
    return false;

  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectCondBranch(MachineInstr &I,
                                              MachineRegisterInfo &MRI,
                                              MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_BRCOND) && "unexpected instruction");

  const Register CondReg = I.getOperand(0).getReg();
  MachineBasicBlock *DestMBB = I.getOperand(1).getMBB();

  MachineInstr &TestInst =
      *BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TEST8ri))
           .addReg(CondReg)
           .addImm(1);
  BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::JCC_1))
      .addMBB(DestMBB).addImm(X86::COND_NE);

  constrainSelectedInstRegOperands(TestInst, TII, TRI, RBI);

  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::materializeFP(MachineInstr &I,
                                           MachineRegisterInfo &MRI,
                                           MachineFunction &MF) const {
  assert((I.getOpcode() == TargetOpcode::G_FCONSTANT) &&
         "unexpected instruction");

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

  const Register DstReg = I.getOperand(0).getReg();
  const LLT DstTy = MRI.getType(DstReg);
  const RegisterBank &RegBank = *RBI.getRegBank(DstReg, MRI, TRI);
  Align Alignment = Align(DstTy.getSizeInBytes());
  const DebugLoc &DbgLoc = I.getDebugLoc();

  unsigned Opc =
      getLoadStoreOp(DstTy, RegBank, TargetOpcode::G_LOAD, Alignment);

  // Create the load from the constant pool.
  const ConstantFP *CFP = I.getOperand(1).getFPImm();
  unsigned CPI = MF.getConstantPool()->getConstantPoolIndex(CFP, Alignment);
  MachineInstr *LoadInst = nullptr;
  unsigned char OpFlag = STI.classifyLocalReference(nullptr);

  if (CM == CodeModel::Large && STI.is64Bit()) {
    // Under X86-64 non-small code model, GV (and friends) are 64-bits, so
    // they cannot be folded into immediate fields.

    Register AddrReg = MRI.createVirtualRegister(&X86::GR64RegClass);
    BuildMI(*I.getParent(), I, DbgLoc, TII.get(X86::MOV64ri), AddrReg)
        .addConstantPoolIndex(CPI, 0, OpFlag);

    MachineMemOperand *MMO = MF.getMachineMemOperand(
        MachinePointerInfo::getConstantPool(MF), MachineMemOperand::MOLoad,
        MF.getDataLayout().getPointerSize(), Alignment);

    LoadInst =
        addDirectMem(BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg),
                     AddrReg)
            .addMemOperand(MMO);

  } else if (CM == CodeModel::Small || !STI.is64Bit()) {
    // Handle the case when globals fit in our immediate field.
    // This is true for X86-32 always and X86-64 when in -mcmodel=small mode.

    // x86-32 PIC requires a PIC base register for constant pools.
    unsigned PICBase = 0;
    if (OpFlag == X86II::MO_PIC_BASE_OFFSET || OpFlag == X86II::MO_GOTOFF) {
      // PICBase can be allocated by TII.getGlobalBaseReg(&MF).
      // In DAGISEL the code that initialize it generated by the CGBR pass.
      return false; // TODO support the mode.
    } else if (STI.is64Bit() && TM.getCodeModel() == CodeModel::Small)
      PICBase = X86::RIP;

    LoadInst = addConstantPoolReference(
        BuildMI(*I.getParent(), I, DbgLoc, TII.get(Opc), DstReg), CPI, PICBase,
        OpFlag);
  } else
    return false;

  constrainSelectedInstRegOperands(*LoadInst, TII, TRI, RBI);
  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectImplicitDefOrPHI(
    MachineInstr &I, MachineRegisterInfo &MRI) const {
  assert((I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF ||
          I.getOpcode() == TargetOpcode::G_PHI) &&
         "unexpected instruction");

  Register DstReg = I.getOperand(0).getReg();

  if (!MRI.getRegClassOrNull(DstReg)) {
    const LLT DstTy = MRI.getType(DstReg);
    const TargetRegisterClass *RC = getRegClass(DstTy, DstReg, MRI);

    if (!RBI.constrainGenericRegister(DstReg, *RC, MRI)) {
      LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
                        << " operand\n");
      return false;
    }
  }

  if (I.getOpcode() == TargetOpcode::G_IMPLICIT_DEF)
    I.setDesc(TII.get(X86::IMPLICIT_DEF));
  else
    I.setDesc(TII.get(X86::PHI));

  return true;
}

bool X86InstructionSelector::selectDivRem(MachineInstr &I,
                                          MachineRegisterInfo &MRI,
                                          MachineFunction &MF) const {
  // The implementation of this function is taken from X86FastISel.
  assert((I.getOpcode() == TargetOpcode::G_SDIV ||
          I.getOpcode() == TargetOpcode::G_SREM ||
          I.getOpcode() == TargetOpcode::G_UDIV ||
          I.getOpcode() == TargetOpcode::G_UREM) &&
         "unexpected instruction");

  const Register DstReg = I.getOperand(0).getReg();
  const Register Op1Reg = I.getOperand(1).getReg();
  const Register Op2Reg = I.getOperand(2).getReg();

  const LLT RegTy = MRI.getType(DstReg);
  assert(RegTy == MRI.getType(Op1Reg) && RegTy == MRI.getType(Op2Reg) &&
         "Arguments and return value types must match");

  const RegisterBank *RegRB = RBI.getRegBank(DstReg, MRI, TRI);
  if (!RegRB || RegRB->getID() != X86::GPRRegBankID)
    return false;

  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 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 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 the dividend
  // into lowreg, instead of copying, and ignore the highreg.
  const static struct DivRemEntry {
    // The following portion depends only on the data type.
    unsigned SizeInBits;
    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] = {
      {8,
       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
      {16,
       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
      {32,
       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
      {64,
       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
  };

  auto OpEntryIt = llvm::find_if(OpTable, [RegTy](const DivRemEntry &El) {
    return El.SizeInBits == RegTy.getSizeInBits();
  });
  if (OpEntryIt == std::end(OpTable))
    return false;

  unsigned OpIndex;
  switch (I.getOpcode()) {
  default:
    llvm_unreachable("Unexpected div/rem opcode");
  case TargetOpcode::G_SDIV:
    OpIndex = 0;
    break;
  case TargetOpcode::G_SREM:
    OpIndex = 1;
    break;
  case TargetOpcode::G_UDIV:
    OpIndex = 2;
    break;
  case TargetOpcode::G_UREM:
    OpIndex = 3;
    break;
  }

  const DivRemEntry &TypeEntry = *OpEntryIt;
  const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex];

  const TargetRegisterClass *RegRC = getRegClass(RegTy, *RegRB);
  if (!RBI.constrainGenericRegister(Op1Reg, *RegRC, MRI) ||
      !RBI.constrainGenericRegister(Op2Reg, *RegRC, MRI) ||
      !RBI.constrainGenericRegister(DstReg, *RegRC, MRI)) {
    LLVM_DEBUG(dbgs() << "Failed to constrain " << TII.getName(I.getOpcode())
                      << " operand\n");
    return false;
  }

  // Move op1 into low-order input register.
  BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpCopy),
          TypeEntry.LowInReg)
      .addReg(Op1Reg);
  // Zero-extend or sign-extend into high-order input register.
  if (OpEntry.OpSignExtend) {
    if (OpEntry.IsOpSigned)
      BuildMI(*I.getParent(), I, I.getDebugLoc(),
              TII.get(OpEntry.OpSignExtend));
    else {
      Register Zero32 = MRI.createVirtualRegister(&X86::GR32RegClass);
      BuildMI(*I.getParent(), I, I.getDebugLoc(), 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 (RegTy.getSizeInBits() == 16) {
        BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy),
                TypeEntry.HighInReg)
            .addReg(Zero32, 0, X86::sub_16bit);
      } else if (RegTy.getSizeInBits() == 32) {
        BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy),
                TypeEntry.HighInReg)
            .addReg(Zero32);
      } else if (RegTy.getSizeInBits() == 64) {
        BuildMI(*I.getParent(), I, I.getDebugLoc(),
                TII.get(TargetOpcode::SUBREG_TO_REG), TypeEntry.HighInReg)
            .addImm(0)
            .addReg(Zero32)
            .addImm(X86::sub_32bit);
      }
    }
  }
  // Generate the DIV/IDIV instruction.
  BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(OpEntry.OpDivRem))
      .addReg(Op2Reg);
  // 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.
  if ((I.getOpcode() == Instruction::SRem ||
       I.getOpcode() == Instruction::URem) &&
      OpEntry.DivRemResultReg == X86::AH && STI.is64Bit()) {
    Register SourceSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass);
    Register ResultSuperReg = MRI.createVirtualRegister(&X86::GR16RegClass);
    BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(Copy), SourceSuperReg)
        .addReg(X86::AX);

    // Shift AX right by 8 bits instead of using AH.
    BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::SHR16ri),
            ResultSuperReg)
        .addReg(SourceSuperReg)
        .addImm(8);

    // Now reference the 8-bit subreg of the result.
    BuildMI(*I.getParent(), I, I.getDebugLoc(),
            TII.get(TargetOpcode::SUBREG_TO_REG))
        .addDef(DstReg)
        .addImm(0)
        .addReg(ResultSuperReg)
        .addImm(X86::sub_8bit);
  } else {
    BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(TargetOpcode::COPY),
            DstReg)
        .addReg(OpEntry.DivRemResultReg);
  }
  I.eraseFromParent();
  return true;
}

bool X86InstructionSelector::selectIntrinsicWSideEffects(
    MachineInstr &I, MachineRegisterInfo &MRI, MachineFunction &MF) const {

  assert(I.getOpcode() == TargetOpcode::G_INTRINSIC_W_SIDE_EFFECTS &&
         "unexpected instruction");

  if (I.getOperand(0).getIntrinsicID() != Intrinsic::trap)
    return false;

  BuildMI(*I.getParent(), I, I.getDebugLoc(), TII.get(X86::TRAP));

  I.eraseFromParent();
  return true;
}

InstructionSelector *
llvm::createX86InstructionSelector(const X86TargetMachine &TM,
                                   X86Subtarget &Subtarget,
                                   X86RegisterBankInfo &RBI) {
  return new X86InstructionSelector(TM, Subtarget, RBI);
}