xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/AArch64FastISel.cpp (revision 7a6dacac)

//===- AArch6464FastISel.cpp - AArch64 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 AArch64-specific support for the FastISel class. Some
// of the target-specific code is generated by tablegen in the file
// AArch64GenFastISel.inc, which is #included here.
//
//===----------------------------------------------------------------------===//

#include "AArch64.h"
#include "AArch64CallingConvention.h"
#include "AArch64MachineFunctionInfo.h"
#include "AArch64RegisterInfo.h"
#include "AArch64Subtarget.h"
#include "MCTargetDesc/AArch64AddressingModes.h"
#include "Utils/AArch64BaseInfo.h"
#include "llvm/ADT/APFloat.h"
#include "llvm/ADT/APInt.h"
#include "llvm/ADT/DenseMap.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/Analysis/BranchProbabilityInfo.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/ISDOpcodes.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineMemOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/MachineValueType.h"
#include "llvm/CodeGen/RuntimeLibcalls.h"
#include "llvm/CodeGen/ValueTypes.h"
#include "llvm/IR/Argument.h"
#include "llvm/IR/Attributes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/Constant.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/InstrTypes.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/IntrinsicsAArch64.h"
#include "llvm/IR/Operator.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/User.h"
#include "llvm/IR/Value.h"
#include "llvm/MC/MCInstrDesc.h"
#include "llvm/MC/MCRegisterInfo.h"
#include "llvm/MC/MCSymbol.h"
#include "llvm/Support/AtomicOrdering.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/CodeGen.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include <algorithm>
#include <cassert>
#include <cstdint>
#include <iterator>
#include <utility>

using namespace llvm;

namespace {

class AArch64FastISel final : public FastISel {
  class Address {
  public:
    using BaseKind = enum {
      RegBase,
      FrameIndexBase
    };

  private:
    BaseKind Kind = RegBase;
    AArch64_AM::ShiftExtendType ExtType = AArch64_AM::InvalidShiftExtend;
    union {
      unsigned Reg;
      int FI;
    } Base;
    unsigned OffsetReg = 0;
    unsigned Shift = 0;
    int64_t Offset = 0;
    const GlobalValue *GV = nullptr;

  public:
    Address() { Base.Reg = 0; }

    void setKind(BaseKind K) { Kind = K; }
    BaseKind getKind() const { return Kind; }
    void setExtendType(AArch64_AM::ShiftExtendType E) { ExtType = E; }
    AArch64_AM::ShiftExtendType getExtendType() const { return ExtType; }
    bool isRegBase() const { return Kind == RegBase; }
    bool isFIBase() const { return Kind == FrameIndexBase; }

    void setReg(unsigned Reg) {
      assert(isRegBase() && "Invalid base register access!");
      Base.Reg = Reg;
    }

    unsigned getReg() const {
      assert(isRegBase() && "Invalid base register access!");
      return Base.Reg;
    }

    void setOffsetReg(unsigned Reg) {
      OffsetReg = Reg;
    }

    unsigned getOffsetReg() const {
      return OffsetReg;
    }

    void setFI(unsigned FI) {
      assert(isFIBase() && "Invalid base frame index  access!");
      Base.FI = FI;
    }

    unsigned getFI() const {
      assert(isFIBase() && "Invalid base frame index access!");
      return Base.FI;
    }

    void setOffset(int64_t O) { Offset = O; }
    int64_t getOffset() { return Offset; }
    void setShift(unsigned S) { Shift = S; }
    unsigned getShift() { return Shift; }

    void setGlobalValue(const GlobalValue *G) { GV = G; }
    const GlobalValue *getGlobalValue() { return GV; }
  };

  /// Subtarget - Keep a pointer to the AArch64Subtarget around so that we can
  /// make the right decision when generating code for different targets.
  const AArch64Subtarget *Subtarget;
  LLVMContext *Context;

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

private:
  // Selection routines.
  bool selectAddSub(const Instruction *I);
  bool selectLogicalOp(const Instruction *I);
  bool selectLoad(const Instruction *I);
  bool selectStore(const Instruction *I);
  bool selectBranch(const Instruction *I);
  bool selectIndirectBr(const Instruction *I);
  bool selectCmp(const Instruction *I);
  bool selectSelect(const Instruction *I);
  bool selectFPExt(const Instruction *I);
  bool selectFPTrunc(const Instruction *I);
  bool selectFPToInt(const Instruction *I, bool Signed);
  bool selectIntToFP(const Instruction *I, bool Signed);
  bool selectRem(const Instruction *I, unsigned ISDOpcode);
  bool selectRet(const Instruction *I);
  bool selectTrunc(const Instruction *I);
  bool selectIntExt(const Instruction *I);
  bool selectMul(const Instruction *I);
  bool selectShift(const Instruction *I);
  bool selectBitCast(const Instruction *I);
  bool selectFRem(const Instruction *I);
  bool selectSDiv(const Instruction *I);
  bool selectGetElementPtr(const Instruction *I);
  bool selectAtomicCmpXchg(const AtomicCmpXchgInst *I);

  // Utility helper routines.
  bool isTypeLegal(Type *Ty, MVT &VT);
  bool isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed = false);
  bool isValueAvailable(const Value *V) const;
  bool computeAddress(const Value *Obj, Address &Addr, Type *Ty = nullptr);
  bool computeCallAddress(const Value *V, Address &Addr);
  bool simplifyAddress(Address &Addr, MVT VT);
  void addLoadStoreOperands(Address &Addr, const MachineInstrBuilder &MIB,
                            MachineMemOperand::Flags Flags,
                            unsigned ScaleFactor, MachineMemOperand *MMO);
  bool isMemCpySmall(uint64_t Len, MaybeAlign Alignment);
  bool tryEmitSmallMemCpy(Address Dest, Address Src, uint64_t Len,
                          MaybeAlign Alignment);
  bool foldXALUIntrinsic(AArch64CC::CondCode &CC, const Instruction *I,
                         const Value *Cond);
  bool optimizeIntExtLoad(const Instruction *I, MVT RetVT, MVT SrcVT);
  bool optimizeSelect(const SelectInst *SI);
  unsigned getRegForGEPIndex(const Value *Idx);

  // Emit helper routines.
  unsigned emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
                      const Value *RHS, bool SetFlags = false,
                      bool WantResult = true,  bool IsZExt = false);
  unsigned emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
                         unsigned RHSReg, bool SetFlags = false,
                         bool WantResult = true);
  unsigned emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
                         uint64_t Imm, bool SetFlags = false,
                         bool WantResult = true);
  unsigned emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
                         unsigned RHSReg, AArch64_AM::ShiftExtendType ShiftType,
                         uint64_t ShiftImm, bool SetFlags = false,
                         bool WantResult = true);
  unsigned emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
                         unsigned RHSReg, AArch64_AM::ShiftExtendType ExtType,
                         uint64_t ShiftImm, bool SetFlags = false,
                         bool WantResult = true);

  // Emit functions.
  bool emitCompareAndBranch(const BranchInst *BI);
  bool emitCmp(const Value *LHS, const Value *RHS, bool IsZExt);
  bool emitICmp(MVT RetVT, const Value *LHS, const Value *RHS, bool IsZExt);
  bool emitICmp_ri(MVT RetVT, unsigned LHSReg, uint64_t Imm);
  bool emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS);
  unsigned emitLoad(MVT VT, MVT ResultVT, Address Addr, bool WantZExt = true,
                    MachineMemOperand *MMO = nullptr);
  bool emitStore(MVT VT, unsigned SrcReg, Address Addr,
                 MachineMemOperand *MMO = nullptr);
  bool emitStoreRelease(MVT VT, unsigned SrcReg, unsigned AddrReg,
                        MachineMemOperand *MMO = nullptr);
  unsigned emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT, bool isZExt);
  unsigned emiti1Ext(unsigned SrcReg, MVT DestVT, bool isZExt);
  unsigned emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
                   bool SetFlags = false, bool WantResult = true,
                   bool IsZExt = false);
  unsigned emitAdd_ri_(MVT VT, unsigned Op0, int64_t Imm);
  unsigned emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
                   bool SetFlags = false, bool WantResult = true,
                   bool IsZExt = false);
  unsigned emitSubs_rr(MVT RetVT, unsigned LHSReg, unsigned RHSReg,
                       bool WantResult = true);
  unsigned emitSubs_rs(MVT RetVT, unsigned LHSReg, unsigned RHSReg,
                       AArch64_AM::ShiftExtendType ShiftType, uint64_t ShiftImm,
                       bool WantResult = true);
  unsigned emitLogicalOp(unsigned ISDOpc, MVT RetVT, const Value *LHS,
                         const Value *RHS);
  unsigned emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
                            uint64_t Imm);
  unsigned emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT, unsigned LHSReg,
                            unsigned RHSReg, uint64_t ShiftImm);
  unsigned emitAnd_ri(MVT RetVT, unsigned LHSReg, uint64_t Imm);
  unsigned emitMul_rr(MVT RetVT, unsigned Op0, unsigned Op1);
  unsigned emitSMULL_rr(MVT RetVT, unsigned Op0, unsigned Op1);
  unsigned emitUMULL_rr(MVT RetVT, unsigned Op0, unsigned Op1);
  unsigned emitLSL_rr(MVT RetVT, unsigned Op0Reg, unsigned Op1Reg);
  unsigned emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, uint64_t Imm,
                      bool IsZExt = true);
  unsigned emitLSR_rr(MVT RetVT, unsigned Op0Reg, unsigned Op1Reg);
  unsigned emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, uint64_t Imm,
                      bool IsZExt = true);
  unsigned emitASR_rr(MVT RetVT, unsigned Op0Reg, unsigned Op1Reg);
  unsigned emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0Reg, uint64_t Imm,
                      bool IsZExt = false);

  unsigned materializeInt(const ConstantInt *CI, MVT VT);
  unsigned materializeFP(const ConstantFP *CFP, MVT VT);
  unsigned materializeGV(const GlobalValue *GV);

  // Call handling routines.
private:
  CCAssignFn *CCAssignFnForCall(CallingConv::ID CC) const;
  bool processCallArgs(CallLoweringInfo &CLI, SmallVectorImpl<MVT> &ArgVTs,
                       unsigned &NumBytes);
  bool finishCall(CallLoweringInfo &CLI, unsigned NumBytes);

public:
  // Backend specific FastISel code.
  unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
  unsigned fastMaterializeConstant(const Constant *C) override;
  unsigned fastMaterializeFloatZero(const ConstantFP* CF) override;

  explicit AArch64FastISel(FunctionLoweringInfo &FuncInfo,
                           const TargetLibraryInfo *LibInfo)
      : FastISel(FuncInfo, LibInfo, /*SkipTargetIndependentISel=*/true) {
    Subtarget = &FuncInfo.MF->getSubtarget<AArch64Subtarget>();
    Context = &FuncInfo.Fn->getContext();
  }

  bool fastSelectInstruction(const Instruction *I) override;

#include "AArch64GenFastISel.inc"
};

} // end anonymous namespace

/// Check if the sign-/zero-extend will be a noop.
static bool isIntExtFree(const Instruction *I) {
  assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
         "Unexpected integer extend instruction.");
  assert(!I->getType()->isVectorTy() && I->getType()->isIntegerTy() &&
         "Unexpected value type.");
  bool IsZExt = isa<ZExtInst>(I);

  if (const auto *LI = dyn_cast<LoadInst>(I->getOperand(0)))
    if (LI->hasOneUse())
      return true;

  if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0)))
    if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr()))
      return true;

  return false;
}

/// Determine the implicit scale factor that is applied by a memory
/// operation for a given value type.
static unsigned getImplicitScaleFactor(MVT VT) {
  switch (VT.SimpleTy) {
  default:
    return 0;    // invalid
  case MVT::i1:  // fall-through
  case MVT::i8:
    return 1;
  case MVT::i16:
    return 2;
  case MVT::i32: // fall-through
  case MVT::f32:
    return 4;
  case MVT::i64: // fall-through
  case MVT::f64:
    return 8;
  }
}

CCAssignFn *AArch64FastISel::CCAssignFnForCall(CallingConv::ID CC) const {
  if (CC == CallingConv::GHC)
    return CC_AArch64_GHC;
  if (CC == CallingConv::CFGuard_Check)
    return CC_AArch64_Win64_CFGuard_Check;
  if (Subtarget->isTargetDarwin())
    return CC_AArch64_DarwinPCS;
  if (Subtarget->isTargetWindows())
    return CC_AArch64_Win64PCS;
  return CC_AArch64_AAPCS;
}

unsigned AArch64FastISel::fastMaterializeAlloca(const AllocaInst *AI) {
  assert(TLI.getValueType(DL, AI->getType(), true) == MVT::i64 &&
         "Alloca should always return a pointer.");

  // Don't handle dynamic allocas.
  if (!FuncInfo.StaticAllocaMap.count(AI))
    return 0;

  DenseMap<const AllocaInst *, int>::iterator SI =
      FuncInfo.StaticAllocaMap.find(AI);

  if (SI != FuncInfo.StaticAllocaMap.end()) {
    Register ResultReg = createResultReg(&AArch64::GPR64spRegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::ADDXri),
            ResultReg)
        .addFrameIndex(SI->second)
        .addImm(0)
        .addImm(0);
    return ResultReg;
  }

  return 0;
}

unsigned AArch64FastISel::materializeInt(const ConstantInt *CI, MVT VT) {
  if (VT > MVT::i64)
    return 0;

  if (!CI->isZero())
    return fastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());

  // Create a copy from the zero register to materialize a "0" value.
  const TargetRegisterClass *RC = (VT == MVT::i64) ? &AArch64::GPR64RegClass
                                                   : &AArch64::GPR32RegClass;
  unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
  Register ResultReg = createResultReg(RC);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(TargetOpcode::COPY),
          ResultReg).addReg(ZeroReg, getKillRegState(true));
  return ResultReg;
}

unsigned AArch64FastISel::materializeFP(const ConstantFP *CFP, MVT VT) {
  // Positive zero (+0.0) has to be materialized with a fmov from the zero
  // register, because the immediate version of fmov cannot encode zero.
  if (CFP->isNullValue())
    return fastMaterializeFloatZero(CFP);

  if (VT != MVT::f32 && VT != MVT::f64)
    return 0;

  const APFloat Val = CFP->getValueAPF();
  bool Is64Bit = (VT == MVT::f64);
  // This checks to see if we can use FMOV instructions to materialize
  // a constant, otherwise we have to materialize via the constant pool.
  int Imm =
      Is64Bit ? AArch64_AM::getFP64Imm(Val) : AArch64_AM::getFP32Imm(Val);
  if (Imm != -1) {
    unsigned Opc = Is64Bit ? AArch64::FMOVDi : AArch64::FMOVSi;
    return fastEmitInst_i(Opc, TLI.getRegClassFor(VT), Imm);
  }

  // For the large code model materialize the FP constant in code.
  if (TM.getCodeModel() == CodeModel::Large) {
    unsigned Opc1 = Is64Bit ? AArch64::MOVi64imm : AArch64::MOVi32imm;
    const TargetRegisterClass *RC = Is64Bit ?
        &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;

    Register TmpReg = createResultReg(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(Opc1), TmpReg)
        .addImm(CFP->getValueAPF().bitcastToAPInt().getZExtValue());

    Register ResultReg = createResultReg(TLI.getRegClassFor(VT));
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(TargetOpcode::COPY), ResultReg)
        .addReg(TmpReg, getKillRegState(true));

    return ResultReg;
  }

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

  unsigned CPI = MCP.getConstantPoolIndex(cast<Constant>(CFP), Alignment);
  Register ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::ADRP),
          ADRPReg).addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGE);

  unsigned Opc = Is64Bit ? AArch64::LDRDui : AArch64::LDRSui;
  Register ResultReg = createResultReg(TLI.getRegClassFor(VT));
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(Opc), ResultReg)
      .addReg(ADRPReg)
      .addConstantPoolIndex(CPI, 0, AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
  return ResultReg;
}

unsigned AArch64FastISel::materializeGV(const GlobalValue *GV) {
  // We can't handle thread-local variables quickly yet.
  if (GV->isThreadLocal())
    return 0;

  // MachO still uses GOT for large code-model accesses, but ELF requires
  // movz/movk sequences, which FastISel doesn't handle yet.
  if (!Subtarget->useSmallAddressing() && !Subtarget->isTargetMachO())
    return 0;

  unsigned OpFlags = Subtarget->ClassifyGlobalReference(GV, TM);

  EVT DestEVT = TLI.getValueType(DL, GV->getType(), true);
  if (!DestEVT.isSimple())
    return 0;

  Register ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
  unsigned ResultReg;

  if (OpFlags & AArch64II::MO_GOT) {
    // ADRP + LDRX
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::ADRP),
            ADRPReg)
        .addGlobalAddress(GV, 0, AArch64II::MO_PAGE | OpFlags);

    unsigned LdrOpc;
    if (Subtarget->isTargetILP32()) {
      ResultReg = createResultReg(&AArch64::GPR32RegClass);
      LdrOpc = AArch64::LDRWui;
    } else {
      ResultReg = createResultReg(&AArch64::GPR64RegClass);
      LdrOpc = AArch64::LDRXui;
    }
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(LdrOpc),
            ResultReg)
      .addReg(ADRPReg)
      .addGlobalAddress(GV, 0, AArch64II::MO_GOT | AArch64II::MO_PAGEOFF |
                        AArch64II::MO_NC | OpFlags);
    if (!Subtarget->isTargetILP32())
      return ResultReg;

    // LDRWui produces a 32-bit register, but pointers in-register are 64-bits
    // so we must extend the result on ILP32.
    Register Result64 = createResultReg(&AArch64::GPR64RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(TargetOpcode::SUBREG_TO_REG))
        .addDef(Result64)
        .addImm(0)
        .addReg(ResultReg, RegState::Kill)
        .addImm(AArch64::sub_32);
    return Result64;
  } else {
    // ADRP + ADDX
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::ADRP),
            ADRPReg)
        .addGlobalAddress(GV, 0, AArch64II::MO_PAGE | OpFlags);

    if (OpFlags & AArch64II::MO_TAGGED) {
      // MO_TAGGED on the page indicates a tagged address. Set the tag now.
      // We do so by creating a MOVK that sets bits 48-63 of the register to
      // (global address + 0x100000000 - PC) >> 48. This assumes that we're in
      // the small code model so we can assume a binary size of <= 4GB, which
      // makes the untagged PC relative offset positive. The binary must also be
      // loaded into address range [0, 2^48). Both of these properties need to
      // be ensured at runtime when using tagged addresses.
      //
      // TODO: There is duplicate logic in AArch64ExpandPseudoInsts.cpp that
      // also uses BuildMI for making an ADRP (+ MOVK) + ADD, but the operands
      // are not exactly 1:1 with FastISel so we cannot easily abstract this
      // out. At some point, it would be nice to find a way to not have this
      // duplciate code.
      unsigned DstReg = createResultReg(&AArch64::GPR64commonRegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::MOVKXi),
              DstReg)
          .addReg(ADRPReg)
          .addGlobalAddress(GV, /*Offset=*/0x100000000,
                            AArch64II::MO_PREL | AArch64II::MO_G3)
          .addImm(48);
      ADRPReg = DstReg;
    }

    ResultReg = createResultReg(&AArch64::GPR64spRegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::ADDXri),
            ResultReg)
        .addReg(ADRPReg)
        .addGlobalAddress(GV, 0,
                          AArch64II::MO_PAGEOFF | AArch64II::MO_NC | OpFlags)
        .addImm(0);
  }
  return ResultReg;
}

unsigned AArch64FastISel::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();
  // arm64_32 has 32-bit pointers held in 64-bit registers. Because of that,
  // 'null' pointers need to have a somewhat special treatment.
  if (isa<ConstantPointerNull>(C)) {
    assert(VT == MVT::i64 && "Expected 64-bit pointers");
    return materializeInt(ConstantInt::get(Type::getInt64Ty(*Context), 0), VT);
  }

  if (const auto *CI = dyn_cast<ConstantInt>(C))
    return materializeInt(CI, VT);
  else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
    return materializeFP(CFP, VT);
  else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
    return materializeGV(GV);

  return 0;
}

unsigned AArch64FastISel::fastMaterializeFloatZero(const ConstantFP* CFP) {
  assert(CFP->isNullValue() &&
         "Floating-point constant is not a positive zero.");
  MVT VT;
  if (!isTypeLegal(CFP->getType(), VT))
    return 0;

  if (VT != MVT::f32 && VT != MVT::f64)
    return 0;

  bool Is64Bit = (VT == MVT::f64);
  unsigned ZReg = Is64Bit ? AArch64::XZR : AArch64::WZR;
  unsigned Opc = Is64Bit ? AArch64::FMOVXDr : AArch64::FMOVWSr;
  return fastEmitInst_r(Opc, TLI.getRegClassFor(VT), ZReg);
}

/// Check if the multiply is by a power-of-2 constant.
static bool isMulPowOf2(const Value *I) {
  if (const auto *MI = dyn_cast<MulOperator>(I)) {
    if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(0)))
      if (C->getValue().isPowerOf2())
        return true;
    if (const auto *C = dyn_cast<ConstantInt>(MI->getOperand(1)))
      if (C->getValue().isPowerOf2())
        return true;
  }
  return false;
}

// Computes the address to get to an object.
bool AArch64FastISel::computeAddress(const Value *Obj, Address &Addr, Type *Ty)
{
  const User *U = nullptr;
  unsigned Opcode = Instruction::UserOp1;
  if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
    // Don't walk into other basic blocks unless the object is an alloca from
    // another block, otherwise it may not have a virtual register assigned.
    if (FuncInfo.StaticAllocaMap.count(static_cast<const AllocaInst *>(Obj)) ||
        FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
      Opcode = I->getOpcode();
      U = I;
    }
  } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
    Opcode = C->getOpcode();
    U = C;
  }

  if (auto *Ty = dyn_cast<PointerType>(Obj->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 through bitcasts.
    return computeAddress(U->getOperand(0), Addr, Ty);

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

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

  case Instruction::GetElementPtr: {
    Address SavedAddr = Addr;
    uint64_t TmpOffset = Addr.getOffset();

    // Iterate through the GEP folding the constants into offsets where
    // we can.
    for (gep_type_iterator GTI = gep_type_begin(U), E = gep_type_end(U);
         GTI != E; ++GTI) {
      const Value *Op = GTI.getOperand();
      if (StructType *STy = GTI.getStructTypeOrNull()) {
        const StructLayout *SL = DL.getStructLayout(STy);
        unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
        TmpOffset += SL->getElementOffset(Idx);
      } else {
        uint64_t S = GTI.getSequentialElementStride(DL);
        while (true) {
          if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
            // Constant-offset addressing.
            TmpOffset += 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));
            TmpOffset += CI->getSExtValue() * S;
            // Iterate on the other operand.
            Op = cast<AddOperator>(Op)->getOperand(0);
            continue;
          }
          // Unsupported
          goto unsupported_gep;
        }
      }
    }

    // Try to grab the base operand now.
    Addr.setOffset(TmpOffset);
    if (computeAddress(U->getOperand(0), Addr, Ty))
      return true;

    // We failed, restore everything and try the other options.
    Addr = SavedAddr;

  unsupported_gep:
    break;
  }
  case Instruction::Alloca: {
    const AllocaInst *AI = cast<AllocaInst>(Obj);
    DenseMap<const AllocaInst *, int>::iterator SI =
        FuncInfo.StaticAllocaMap.find(AI);
    if (SI != FuncInfo.StaticAllocaMap.end()) {
      Addr.setKind(Address::FrameIndexBase);
      Addr.setFI(SI->second);
      return true;
    }
    break;
  }
  case Instruction::Add: {
    // Adds of constants are common and easy enough.
    const Value *LHS = U->getOperand(0);
    const Value *RHS = U->getOperand(1);

    if (isa<ConstantInt>(LHS))
      std::swap(LHS, RHS);

    if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
      Addr.setOffset(Addr.getOffset() + CI->getSExtValue());
      return computeAddress(LHS, Addr, Ty);
    }

    Address Backup = Addr;
    if (computeAddress(LHS, Addr, Ty) && computeAddress(RHS, Addr, Ty))
      return true;
    Addr = Backup;

    break;
  }
  case Instruction::Sub: {
    // Subs of constants are common and easy enough.
    const Value *LHS = U->getOperand(0);
    const Value *RHS = U->getOperand(1);

    if (const ConstantInt *CI = dyn_cast<ConstantInt>(RHS)) {
      Addr.setOffset(Addr.getOffset() - CI->getSExtValue());
      return computeAddress(LHS, Addr, Ty);
    }
    break;
  }
  case Instruction::Shl: {
    if (Addr.getOffsetReg())
      break;

    const auto *CI = dyn_cast<ConstantInt>(U->getOperand(1));
    if (!CI)
      break;

    unsigned Val = CI->getZExtValue();
    if (Val < 1 || Val > 3)
      break;

    uint64_t NumBytes = 0;
    if (Ty && Ty->isSized()) {
      uint64_t NumBits = DL.getTypeSizeInBits(Ty);
      NumBytes = NumBits / 8;
      if (!isPowerOf2_64(NumBits))
        NumBytes = 0;
    }

    if (NumBytes != (1ULL << Val))
      break;

    Addr.setShift(Val);
    Addr.setExtendType(AArch64_AM::LSL);

    const Value *Src = U->getOperand(0);
    if (const auto *I = dyn_cast<Instruction>(Src)) {
      if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
        // Fold the zext or sext when it won't become a noop.
        if (const auto *ZE = dyn_cast<ZExtInst>(I)) {
          if (!isIntExtFree(ZE) &&
              ZE->getOperand(0)->getType()->isIntegerTy(32)) {
            Addr.setExtendType(AArch64_AM::UXTW);
            Src = ZE->getOperand(0);
          }
        } else if (const auto *SE = dyn_cast<SExtInst>(I)) {
          if (!isIntExtFree(SE) &&
              SE->getOperand(0)->getType()->isIntegerTy(32)) {
            Addr.setExtendType(AArch64_AM::SXTW);
            Src = SE->getOperand(0);
          }
        }
      }
    }

    if (const auto *AI = dyn_cast<BinaryOperator>(Src))
      if (AI->getOpcode() == Instruction::And) {
        const Value *LHS = AI->getOperand(0);
        const Value *RHS = AI->getOperand(1);

        if (const auto *C = dyn_cast<ConstantInt>(LHS))
          if (C->getValue() == 0xffffffff)
            std::swap(LHS, RHS);

        if (const auto *C = dyn_cast<ConstantInt>(RHS))
          if (C->getValue() == 0xffffffff) {
            Addr.setExtendType(AArch64_AM::UXTW);
            Register Reg = getRegForValue(LHS);
            if (!Reg)
              return false;
            Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, AArch64::sub_32);
            Addr.setOffsetReg(Reg);
            return true;
          }
      }

    Register Reg = getRegForValue(Src);
    if (!Reg)
      return false;
    Addr.setOffsetReg(Reg);
    return true;
  }
  case Instruction::Mul: {
    if (Addr.getOffsetReg())
      break;

    if (!isMulPowOf2(U))
      break;

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

    // Canonicalize power-of-2 value to the RHS.
    if (const auto *C = dyn_cast<ConstantInt>(LHS))
      if (C->getValue().isPowerOf2())
        std::swap(LHS, RHS);

    assert(isa<ConstantInt>(RHS) && "Expected an ConstantInt.");
    const auto *C = cast<ConstantInt>(RHS);
    unsigned Val = C->getValue().logBase2();
    if (Val < 1 || Val > 3)
      break;

    uint64_t NumBytes = 0;
    if (Ty && Ty->isSized()) {
      uint64_t NumBits = DL.getTypeSizeInBits(Ty);
      NumBytes = NumBits / 8;
      if (!isPowerOf2_64(NumBits))
        NumBytes = 0;
    }

    if (NumBytes != (1ULL << Val))
      break;

    Addr.setShift(Val);
    Addr.setExtendType(AArch64_AM::LSL);

    const Value *Src = LHS;
    if (const auto *I = dyn_cast<Instruction>(Src)) {
      if (FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB) {
        // Fold the zext or sext when it won't become a noop.
        if (const auto *ZE = dyn_cast<ZExtInst>(I)) {
          if (!isIntExtFree(ZE) &&
              ZE->getOperand(0)->getType()->isIntegerTy(32)) {
            Addr.setExtendType(AArch64_AM::UXTW);
            Src = ZE->getOperand(0);
          }
        } else if (const auto *SE = dyn_cast<SExtInst>(I)) {
          if (!isIntExtFree(SE) &&
              SE->getOperand(0)->getType()->isIntegerTy(32)) {
            Addr.setExtendType(AArch64_AM::SXTW);
            Src = SE->getOperand(0);
          }
        }
      }
    }

    Register Reg = getRegForValue(Src);
    if (!Reg)
      return false;
    Addr.setOffsetReg(Reg);
    return true;
  }
  case Instruction::And: {
    if (Addr.getOffsetReg())
      break;

    if (!Ty || DL.getTypeSizeInBits(Ty) != 8)
      break;

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

    if (const auto *C = dyn_cast<ConstantInt>(LHS))
      if (C->getValue() == 0xffffffff)
        std::swap(LHS, RHS);

    if (const auto *C = dyn_cast<ConstantInt>(RHS))
      if (C->getValue() == 0xffffffff) {
        Addr.setShift(0);
        Addr.setExtendType(AArch64_AM::LSL);
        Addr.setExtendType(AArch64_AM::UXTW);

        Register Reg = getRegForValue(LHS);
        if (!Reg)
          return false;
        Reg = fastEmitInst_extractsubreg(MVT::i32, Reg, AArch64::sub_32);
        Addr.setOffsetReg(Reg);
        return true;
      }
    break;
  }
  case Instruction::SExt:
  case Instruction::ZExt: {
    if (!Addr.getReg() || Addr.getOffsetReg())
      break;

    const Value *Src = nullptr;
    // Fold the zext or sext when it won't become a noop.
    if (const auto *ZE = dyn_cast<ZExtInst>(U)) {
      if (!isIntExtFree(ZE) && ZE->getOperand(0)->getType()->isIntegerTy(32)) {
        Addr.setExtendType(AArch64_AM::UXTW);
        Src = ZE->getOperand(0);
      }
    } else if (const auto *SE = dyn_cast<SExtInst>(U)) {
      if (!isIntExtFree(SE) && SE->getOperand(0)->getType()->isIntegerTy(32)) {
        Addr.setExtendType(AArch64_AM::SXTW);
        Src = SE->getOperand(0);
      }
    }

    if (!Src)
      break;

    Addr.setShift(0);
    Register Reg = getRegForValue(Src);
    if (!Reg)
      return false;
    Addr.setOffsetReg(Reg);
    return true;
  }
  } // end switch

  if (Addr.isRegBase() && !Addr.getReg()) {
    Register Reg = getRegForValue(Obj);
    if (!Reg)
      return false;
    Addr.setReg(Reg);
    return true;
  }

  if (!Addr.getOffsetReg()) {
    Register Reg = getRegForValue(Obj);
    if (!Reg)
      return false;
    Addr.setOffsetReg(Reg);
    return true;
  }

  return false;
}

bool AArch64FastISel::computeCallAddress(const Value *V, Address &Addr) {
  const User *U = nullptr;
  unsigned Opcode = Instruction::UserOp1;
  bool InMBB = true;

  if (const auto *I = dyn_cast<Instruction>(V)) {
    Opcode = I->getOpcode();
    U = I;
    InMBB = I->getParent() == FuncInfo.MBB->getBasicBlock();
  } else if (const auto *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 computeCallAddress(U->getOperand(0), Addr);
    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 computeCallAddress(U->getOperand(0), Addr);
    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 computeCallAddress(U->getOperand(0), Addr);
    break;
  }

  if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
    Addr.setGlobalValue(GV);
    return true;
  }

  // If all else fails, try to materialize the value in a register.
  if (!Addr.getGlobalValue()) {
    Addr.setReg(getRegForValue(V));
    return Addr.getReg() != 0;
  }

  return false;
}

bool AArch64FastISel::isTypeLegal(Type *Ty, MVT &VT) {
  EVT evt = TLI.getValueType(DL, Ty, true);

  if (Subtarget->isTargetILP32() && Ty->isPointerTy())
    return false;

  // Only handle simple types.
  if (evt == MVT::Other || !evt.isSimple())
    return false;
  VT = evt.getSimpleVT();

  // This is a legal type, but it's not something we handle in fast-isel.
  if (VT == MVT::f128)
    return false;

  // Handle all other legal types, i.e. a register that will directly hold this
  // value.
  return TLI.isTypeLegal(VT);
}

/// Determine if the value type is supported by FastISel.
///
/// FastISel for AArch64 can handle more value types than are legal. This adds
/// simple value type such as i1, i8, and i16.
bool AArch64FastISel::isTypeSupported(Type *Ty, MVT &VT, bool IsVectorAllowed) {
  if (Ty->isVectorTy() && !IsVectorAllowed)
    return false;

  if (isTypeLegal(Ty, VT))
    return true;

  // If this is a type than can be sign or zero-extended to a basic operation
  // go ahead and accept it now.
  if (VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16)
    return true;

  return false;
}

bool AArch64FastISel::isValueAvailable(const Value *V) const {
  if (!isa<Instruction>(V))
    return true;

  const auto *I = cast<Instruction>(V);
  return FuncInfo.MBBMap[I->getParent()] == FuncInfo.MBB;
}

bool AArch64FastISel::simplifyAddress(Address &Addr, MVT VT) {
  if (Subtarget->isTargetILP32())
    return false;

  unsigned ScaleFactor = getImplicitScaleFactor(VT);
  if (!ScaleFactor)
    return false;

  bool ImmediateOffsetNeedsLowering = false;
  bool RegisterOffsetNeedsLowering = false;
  int64_t Offset = Addr.getOffset();
  if (((Offset < 0) || (Offset & (ScaleFactor - 1))) && !isInt<9>(Offset))
    ImmediateOffsetNeedsLowering = true;
  else if (Offset > 0 && !(Offset & (ScaleFactor - 1)) &&
           !isUInt<12>(Offset / ScaleFactor))
    ImmediateOffsetNeedsLowering = true;

  // Cannot encode an offset register and an immediate offset in the same
  // instruction. Fold the immediate offset into the load/store instruction and
  // emit an additional add to take care of the offset register.
  if (!ImmediateOffsetNeedsLowering && Addr.getOffset() && Addr.getOffsetReg())
    RegisterOffsetNeedsLowering = true;

  // Cannot encode zero register as base.
  if (Addr.isRegBase() && Addr.getOffsetReg() && !Addr.getReg())
    RegisterOffsetNeedsLowering = true;

  // If this is a stack pointer and the offset needs to be simplified then put
  // the alloca address into a register, set the base type back to register and
  // continue. This should almost never happen.
  if ((ImmediateOffsetNeedsLowering || Addr.getOffsetReg()) && Addr.isFIBase())
  {
    Register ResultReg = createResultReg(&AArch64::GPR64spRegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::ADDXri),
            ResultReg)
      .addFrameIndex(Addr.getFI())
      .addImm(0)
      .addImm(0);
    Addr.setKind(Address::RegBase);
    Addr.setReg(ResultReg);
  }

  if (RegisterOffsetNeedsLowering) {
    unsigned ResultReg = 0;
    if (Addr.getReg()) {
      if (Addr.getExtendType() == AArch64_AM::SXTW ||
          Addr.getExtendType() == AArch64_AM::UXTW   )
        ResultReg = emitAddSub_rx(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
                                  Addr.getOffsetReg(), Addr.getExtendType(),
                                  Addr.getShift());
      else
        ResultReg = emitAddSub_rs(/*UseAdd=*/true, MVT::i64, Addr.getReg(),
                                  Addr.getOffsetReg(), AArch64_AM::LSL,
                                  Addr.getShift());
    } else {
      if (Addr.getExtendType() == AArch64_AM::UXTW)
        ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
                               Addr.getShift(), /*IsZExt=*/true);
      else if (Addr.getExtendType() == AArch64_AM::SXTW)
        ResultReg = emitLSL_ri(MVT::i64, MVT::i32, Addr.getOffsetReg(),
                               Addr.getShift(), /*IsZExt=*/false);
      else
        ResultReg = emitLSL_ri(MVT::i64, MVT::i64, Addr.getOffsetReg(),
                               Addr.getShift());
    }
    if (!ResultReg)
      return false;

    Addr.setReg(ResultReg);
    Addr.setOffsetReg(0);
    Addr.setShift(0);
    Addr.setExtendType(AArch64_AM::InvalidShiftExtend);
  }

  // Since the offset is too large for the load/store instruction get the
  // reg+offset into a register.
  if (ImmediateOffsetNeedsLowering) {
    unsigned ResultReg;
    if (Addr.getReg())
      // Try to fold the immediate into the add instruction.
      ResultReg = emitAdd_ri_(MVT::i64, Addr.getReg(), Offset);
    else
      ResultReg = fastEmit_i(MVT::i64, MVT::i64, ISD::Constant, Offset);

    if (!ResultReg)
      return false;
    Addr.setReg(ResultReg);
    Addr.setOffset(0);
  }
  return true;
}

void AArch64FastISel::addLoadStoreOperands(Address &Addr,
                                           const MachineInstrBuilder &MIB,
                                           MachineMemOperand::Flags Flags,
                                           unsigned ScaleFactor,
                                           MachineMemOperand *MMO) {
  int64_t Offset = Addr.getOffset() / ScaleFactor;
  // Frame base works a bit differently. Handle it separately.
  if (Addr.isFIBase()) {
    int FI = Addr.getFI();
    // FIXME: We shouldn't be using getObjectSize/getObjectAlignment.  The size
    // and alignment should be based on the VT.
    MMO = FuncInfo.MF->getMachineMemOperand(
        MachinePointerInfo::getFixedStack(*FuncInfo.MF, FI, Offset), Flags,
        MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
    // Now add the rest of the operands.
    MIB.addFrameIndex(FI).addImm(Offset);
  } else {
    assert(Addr.isRegBase() && "Unexpected address kind.");
    const MCInstrDesc &II = MIB->getDesc();
    unsigned Idx = (Flags & MachineMemOperand::MOStore) ? 1 : 0;
    Addr.setReg(
      constrainOperandRegClass(II, Addr.getReg(), II.getNumDefs()+Idx));
    Addr.setOffsetReg(
      constrainOperandRegClass(II, Addr.getOffsetReg(), II.getNumDefs()+Idx+1));
    if (Addr.getOffsetReg()) {
      assert(Addr.getOffset() == 0 && "Unexpected offset");
      bool IsSigned = Addr.getExtendType() == AArch64_AM::SXTW ||
                      Addr.getExtendType() == AArch64_AM::SXTX;
      MIB.addReg(Addr.getReg());
      MIB.addReg(Addr.getOffsetReg());
      MIB.addImm(IsSigned);
      MIB.addImm(Addr.getShift() != 0);
    } else
      MIB.addReg(Addr.getReg()).addImm(Offset);
  }

  if (MMO)
    MIB.addMemOperand(MMO);
}

unsigned AArch64FastISel::emitAddSub(bool UseAdd, MVT RetVT, const Value *LHS,
                                     const Value *RHS, bool SetFlags,
                                     bool WantResult,  bool IsZExt) {
  AArch64_AM::ShiftExtendType ExtendType = AArch64_AM::InvalidShiftExtend;
  bool NeedExtend = false;
  switch (RetVT.SimpleTy) {
  default:
    return 0;
  case MVT::i1:
    NeedExtend = true;
    break;
  case MVT::i8:
    NeedExtend = true;
    ExtendType = IsZExt ? AArch64_AM::UXTB : AArch64_AM::SXTB;
    break;
  case MVT::i16:
    NeedExtend = true;
    ExtendType = IsZExt ? AArch64_AM::UXTH : AArch64_AM::SXTH;
    break;
  case MVT::i32:  // fall-through
  case MVT::i64:
    break;
  }
  MVT SrcVT = RetVT;
  RetVT.SimpleTy = std::max(RetVT.SimpleTy, MVT::i32);

  // Canonicalize immediates to the RHS first.
  if (UseAdd && isa<Constant>(LHS) && !isa<Constant>(RHS))
    std::swap(LHS, RHS);

  // Canonicalize mul by power of 2 to the RHS.
  if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
    if (isMulPowOf2(LHS))
      std::swap(LHS, RHS);

  // Canonicalize shift immediate to the RHS.
  if (UseAdd && LHS->hasOneUse() && isValueAvailable(LHS))
    if (const auto *SI = dyn_cast<BinaryOperator>(LHS))
      if (isa<ConstantInt>(SI->getOperand(1)))
        if (SI->getOpcode() == Instruction::Shl  ||
            SI->getOpcode() == Instruction::LShr ||
            SI->getOpcode() == Instruction::AShr   )
          std::swap(LHS, RHS);

  Register LHSReg = getRegForValue(LHS);
  if (!LHSReg)
    return 0;

  if (NeedExtend)
    LHSReg = emitIntExt(SrcVT, LHSReg, RetVT, IsZExt);

  unsigned ResultReg = 0;
  if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
    uint64_t Imm = IsZExt ? C->getZExtValue() : C->getSExtValue();
    if (C->isNegative())
      ResultReg = emitAddSub_ri(!UseAdd, RetVT, LHSReg, -Imm, SetFlags,
                                WantResult);
    else
      ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, Imm, SetFlags,
                                WantResult);
  } else if (const auto *C = dyn_cast<Constant>(RHS))
    if (C->isNullValue())
      ResultReg = emitAddSub_ri(UseAdd, RetVT, LHSReg, 0, SetFlags, WantResult);

  if (ResultReg)
    return ResultReg;

  // Only extend the RHS within the instruction if there is a valid extend type.
  if (ExtendType != AArch64_AM::InvalidShiftExtend && RHS->hasOneUse() &&
      isValueAvailable(RHS)) {
    Register RHSReg = getRegForValue(RHS);
    if (!RHSReg)
      return 0;
    return emitAddSub_rx(UseAdd, RetVT, LHSReg, RHSReg, ExtendType, 0,
                         SetFlags, WantResult);
  }

  // Check if the mul can be folded into the instruction.
  if (RHS->hasOneUse() && isValueAvailable(RHS)) {
    if (isMulPowOf2(RHS)) {
      const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
      const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);

      if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
        if (C->getValue().isPowerOf2())
          std::swap(MulLHS, MulRHS);

      assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
      uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();
      Register RHSReg = getRegForValue(MulLHS);
      if (!RHSReg)
        return 0;
      ResultReg = emitAddSub_rs(UseAdd, RetVT, LHSReg, RHSReg, AArch64_AM::LSL,
                                ShiftVal, SetFlags, WantResult);
      if (ResultReg)
        return ResultReg;
    }
  }

  // Check if the shift can be folded into the instruction.
  if (RHS->hasOneUse() && isValueAvailable(RHS)) {
    if (const auto *SI = dyn_cast<BinaryOperator>(RHS)) {
      if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
        AArch64_AM::ShiftExtendType ShiftType = AArch64_AM::InvalidShiftExtend;
        switch (SI->getOpcode()) {
        default: break;
        case Instruction::Shl:  ShiftType = AArch64_AM::LSL; break;
        case Instruction::LShr: ShiftType = AArch64_AM::LSR; break;
        case Instruction::AShr: ShiftType = AArch64_AM::ASR; break;
        }
        uint64_t ShiftVal = C->getZExtValue();
        if (ShiftType != AArch64_AM::InvalidShiftExtend) {
          Register RHSReg = getRegForValue(SI->getOperand(0));
          if (!RHSReg)
            return 0;
          ResultReg = emitAddSub_rs(UseAdd, RetVT, LHSReg, RHSReg, ShiftType,
                                    ShiftVal, SetFlags, WantResult);
          if (ResultReg)
            return ResultReg;
        }
      }
    }
  }

  Register RHSReg = getRegForValue(RHS);
  if (!RHSReg)
    return 0;

  if (NeedExtend)
    RHSReg = emitIntExt(SrcVT, RHSReg, RetVT, IsZExt);

  return emitAddSub_rr(UseAdd, RetVT, LHSReg, RHSReg, SetFlags, WantResult);
}

unsigned AArch64FastISel::emitAddSub_rr(bool UseAdd, MVT RetVT, unsigned LHSReg,
                                        unsigned RHSReg, bool SetFlags,
                                        bool WantResult) {
  assert(LHSReg && RHSReg && "Invalid register number.");

  if (LHSReg == AArch64::SP || LHSReg == AArch64::WSP ||
      RHSReg == AArch64::SP || RHSReg == AArch64::WSP)
    return 0;

  if (RetVT != MVT::i32 && RetVT != MVT::i64)
    return 0;

  static const unsigned OpcTable[2][2][2] = {
    { { AArch64::SUBWrr,  AArch64::SUBXrr  },
      { AArch64::ADDWrr,  AArch64::ADDXrr  }  },
    { { AArch64::SUBSWrr, AArch64::SUBSXrr },
      { AArch64::ADDSWrr, AArch64::ADDSXrr }  }
  };
  bool Is64Bit = RetVT == MVT::i64;
  unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
  const TargetRegisterClass *RC =
      Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  unsigned ResultReg;
  if (WantResult)
    ResultReg = createResultReg(RC);
  else
    ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;

  const MCInstrDesc &II = TII.get(Opc);
  LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
  RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II, ResultReg)
      .addReg(LHSReg)
      .addReg(RHSReg);
  return ResultReg;
}

unsigned AArch64FastISel::emitAddSub_ri(bool UseAdd, MVT RetVT, unsigned LHSReg,
                                        uint64_t Imm, bool SetFlags,
                                        bool WantResult) {
  assert(LHSReg && "Invalid register number.");

  if (RetVT != MVT::i32 && RetVT != MVT::i64)
    return 0;

  unsigned ShiftImm;
  if (isUInt<12>(Imm))
    ShiftImm = 0;
  else if ((Imm & 0xfff000) == Imm) {
    ShiftImm = 12;
    Imm >>= 12;
  } else
    return 0;

  static const unsigned OpcTable[2][2][2] = {
    { { AArch64::SUBWri,  AArch64::SUBXri  },
      { AArch64::ADDWri,  AArch64::ADDXri  }  },
    { { AArch64::SUBSWri, AArch64::SUBSXri },
      { AArch64::ADDSWri, AArch64::ADDSXri }  }
  };
  bool Is64Bit = RetVT == MVT::i64;
  unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
  const TargetRegisterClass *RC;
  if (SetFlags)
    RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  else
    RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
  unsigned ResultReg;
  if (WantResult)
    ResultReg = createResultReg(RC);
  else
    ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;

  const MCInstrDesc &II = TII.get(Opc);
  LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II, ResultReg)
      .addReg(LHSReg)
      .addImm(Imm)
      .addImm(getShifterImm(AArch64_AM::LSL, ShiftImm));
  return ResultReg;
}

unsigned AArch64FastISel::emitAddSub_rs(bool UseAdd, MVT RetVT, unsigned LHSReg,
                                        unsigned RHSReg,
                                        AArch64_AM::ShiftExtendType ShiftType,
                                        uint64_t ShiftImm, bool SetFlags,
                                        bool WantResult) {
  assert(LHSReg && RHSReg && "Invalid register number.");
  assert(LHSReg != AArch64::SP && LHSReg != AArch64::WSP &&
         RHSReg != AArch64::SP && RHSReg != AArch64::WSP);

  if (RetVT != MVT::i32 && RetVT != MVT::i64)
    return 0;

  // Don't deal with undefined shifts.
  if (ShiftImm >= RetVT.getSizeInBits())
    return 0;

  static const unsigned OpcTable[2][2][2] = {
    { { AArch64::SUBWrs,  AArch64::SUBXrs  },
      { AArch64::ADDWrs,  AArch64::ADDXrs  }  },
    { { AArch64::SUBSWrs, AArch64::SUBSXrs },
      { AArch64::ADDSWrs, AArch64::ADDSXrs }  }
  };
  bool Is64Bit = RetVT == MVT::i64;
  unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
  const TargetRegisterClass *RC =
      Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  unsigned ResultReg;
  if (WantResult)
    ResultReg = createResultReg(RC);
  else
    ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;

  const MCInstrDesc &II = TII.get(Opc);
  LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
  RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II, ResultReg)
      .addReg(LHSReg)
      .addReg(RHSReg)
      .addImm(getShifterImm(ShiftType, ShiftImm));
  return ResultReg;
}

unsigned AArch64FastISel::emitAddSub_rx(bool UseAdd, MVT RetVT, unsigned LHSReg,
                                        unsigned RHSReg,
                                        AArch64_AM::ShiftExtendType ExtType,
                                        uint64_t ShiftImm, bool SetFlags,
                                        bool WantResult) {
  assert(LHSReg && RHSReg && "Invalid register number.");
  assert(LHSReg != AArch64::XZR && LHSReg != AArch64::WZR &&
         RHSReg != AArch64::XZR && RHSReg != AArch64::WZR);

  if (RetVT != MVT::i32 && RetVT != MVT::i64)
    return 0;

  if (ShiftImm >= 4)
    return 0;

  static const unsigned OpcTable[2][2][2] = {
    { { AArch64::SUBWrx,  AArch64::SUBXrx  },
      { AArch64::ADDWrx,  AArch64::ADDXrx  }  },
    { { AArch64::SUBSWrx, AArch64::SUBSXrx },
      { AArch64::ADDSWrx, AArch64::ADDSXrx }  }
  };
  bool Is64Bit = RetVT == MVT::i64;
  unsigned Opc = OpcTable[SetFlags][UseAdd][Is64Bit];
  const TargetRegisterClass *RC = nullptr;
  if (SetFlags)
    RC = Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  else
    RC = Is64Bit ? &AArch64::GPR64spRegClass : &AArch64::GPR32spRegClass;
  unsigned ResultReg;
  if (WantResult)
    ResultReg = createResultReg(RC);
  else
    ResultReg = Is64Bit ? AArch64::XZR : AArch64::WZR;

  const MCInstrDesc &II = TII.get(Opc);
  LHSReg = constrainOperandRegClass(II, LHSReg, II.getNumDefs());
  RHSReg = constrainOperandRegClass(II, RHSReg, II.getNumDefs() + 1);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II, ResultReg)
      .addReg(LHSReg)
      .addReg(RHSReg)
      .addImm(getArithExtendImm(ExtType, ShiftImm));
  return ResultReg;
}

bool AArch64FastISel::emitCmp(const Value *LHS, const Value *RHS, bool IsZExt) {
  Type *Ty = LHS->getType();
  EVT EVT = TLI.getValueType(DL, Ty, true);
  if (!EVT.isSimple())
    return false;
  MVT VT = EVT.getSimpleVT();

  switch (VT.SimpleTy) {
  default:
    return false;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
  case MVT::i64:
    return emitICmp(VT, LHS, RHS, IsZExt);
  case MVT::f32:
  case MVT::f64:
    return emitFCmp(VT, LHS, RHS);
  }
}

bool AArch64FastISel::emitICmp(MVT RetVT, const Value *LHS, const Value *RHS,
                               bool IsZExt) {
  return emitSub(RetVT, LHS, RHS, /*SetFlags=*/true, /*WantResult=*/false,
                 IsZExt) != 0;
}

bool AArch64FastISel::emitICmp_ri(MVT RetVT, unsigned LHSReg, uint64_t Imm) {
  return emitAddSub_ri(/*UseAdd=*/false, RetVT, LHSReg, Imm,
                       /*SetFlags=*/true, /*WantResult=*/false) != 0;
}

bool AArch64FastISel::emitFCmp(MVT RetVT, const Value *LHS, const Value *RHS) {
  if (RetVT != MVT::f32 && RetVT != MVT::f64)
    return false;

  // Check to see if the 2nd operand is a constant that we can encode directly
  // in the compare.
  bool UseImm = false;
  if (const auto *CFP = dyn_cast<ConstantFP>(RHS))
    if (CFP->isZero() && !CFP->isNegative())
      UseImm = true;

  Register LHSReg = getRegForValue(LHS);
  if (!LHSReg)
    return false;

  if (UseImm) {
    unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDri : AArch64::FCMPSri;
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(Opc))
        .addReg(LHSReg);
    return true;
  }

  Register RHSReg = getRegForValue(RHS);
  if (!RHSReg)
    return false;

  unsigned Opc = (RetVT == MVT::f64) ? AArch64::FCMPDrr : AArch64::FCMPSrr;
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(Opc))
      .addReg(LHSReg)
      .addReg(RHSReg);
  return true;
}

unsigned AArch64FastISel::emitAdd(MVT RetVT, const Value *LHS, const Value *RHS,
                                  bool SetFlags, bool WantResult, bool IsZExt) {
  return emitAddSub(/*UseAdd=*/true, RetVT, LHS, RHS, SetFlags, WantResult,
                    IsZExt);
}

/// This method is a wrapper to simplify add emission.
///
/// First try to emit an add with an immediate operand using emitAddSub_ri. If
/// that fails, then try to materialize the immediate into a register and use
/// emitAddSub_rr instead.
unsigned AArch64FastISel::emitAdd_ri_(MVT VT, unsigned Op0, int64_t Imm) {
  unsigned ResultReg;
  if (Imm < 0)
    ResultReg = emitAddSub_ri(false, VT, Op0, -Imm);
  else
    ResultReg = emitAddSub_ri(true, VT, Op0, Imm);

  if (ResultReg)
    return ResultReg;

  unsigned CReg = fastEmit_i(VT, VT, ISD::Constant, Imm);
  if (!CReg)
    return 0;

  ResultReg = emitAddSub_rr(true, VT, Op0, CReg);
  return ResultReg;
}

unsigned AArch64FastISel::emitSub(MVT RetVT, const Value *LHS, const Value *RHS,
                                  bool SetFlags, bool WantResult, bool IsZExt) {
  return emitAddSub(/*UseAdd=*/false, RetVT, LHS, RHS, SetFlags, WantResult,
                    IsZExt);
}

unsigned AArch64FastISel::emitSubs_rr(MVT RetVT, unsigned LHSReg,
                                      unsigned RHSReg, bool WantResult) {
  return emitAddSub_rr(/*UseAdd=*/false, RetVT, LHSReg, RHSReg,
                       /*SetFlags=*/true, WantResult);
}

unsigned AArch64FastISel::emitSubs_rs(MVT RetVT, unsigned LHSReg,
                                      unsigned RHSReg,
                                      AArch64_AM::ShiftExtendType ShiftType,
                                      uint64_t ShiftImm, bool WantResult) {
  return emitAddSub_rs(/*UseAdd=*/false, RetVT, LHSReg, RHSReg, ShiftType,
                       ShiftImm, /*SetFlags=*/true, WantResult);
}

unsigned AArch64FastISel::emitLogicalOp(unsigned ISDOpc, MVT RetVT,
                                        const Value *LHS, const Value *RHS) {
  // Canonicalize immediates to the RHS first.
  if (isa<ConstantInt>(LHS) && !isa<ConstantInt>(RHS))
    std::swap(LHS, RHS);

  // Canonicalize mul by power-of-2 to the RHS.
  if (LHS->hasOneUse() && isValueAvailable(LHS))
    if (isMulPowOf2(LHS))
      std::swap(LHS, RHS);

  // Canonicalize shift immediate to the RHS.
  if (LHS->hasOneUse() && isValueAvailable(LHS))
    if (const auto *SI = dyn_cast<ShlOperator>(LHS))
      if (isa<ConstantInt>(SI->getOperand(1)))
        std::swap(LHS, RHS);

  Register LHSReg = getRegForValue(LHS);
  if (!LHSReg)
    return 0;

  unsigned ResultReg = 0;
  if (const auto *C = dyn_cast<ConstantInt>(RHS)) {
    uint64_t Imm = C->getZExtValue();
    ResultReg = emitLogicalOp_ri(ISDOpc, RetVT, LHSReg, Imm);
  }
  if (ResultReg)
    return ResultReg;

  // Check if the mul can be folded into the instruction.
  if (RHS->hasOneUse() && isValueAvailable(RHS)) {
    if (isMulPowOf2(RHS)) {
      const Value *MulLHS = cast<MulOperator>(RHS)->getOperand(0);
      const Value *MulRHS = cast<MulOperator>(RHS)->getOperand(1);

      if (const auto *C = dyn_cast<ConstantInt>(MulLHS))
        if (C->getValue().isPowerOf2())
          std::swap(MulLHS, MulRHS);

      assert(isa<ConstantInt>(MulRHS) && "Expected a ConstantInt.");
      uint64_t ShiftVal = cast<ConstantInt>(MulRHS)->getValue().logBase2();

      Register RHSReg = getRegForValue(MulLHS);
      if (!RHSReg)
        return 0;
      ResultReg = emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, RHSReg, ShiftVal);
      if (ResultReg)
        return ResultReg;
    }
  }

  // Check if the shift can be folded into the instruction.
  if (RHS->hasOneUse() && isValueAvailable(RHS)) {
    if (const auto *SI = dyn_cast<ShlOperator>(RHS))
      if (const auto *C = dyn_cast<ConstantInt>(SI->getOperand(1))) {
        uint64_t ShiftVal = C->getZExtValue();
        Register RHSReg = getRegForValue(SI->getOperand(0));
        if (!RHSReg)
          return 0;
        ResultReg = emitLogicalOp_rs(ISDOpc, RetVT, LHSReg, RHSReg, ShiftVal);
        if (ResultReg)
          return ResultReg;
      }
  }

  Register RHSReg = getRegForValue(RHS);
  if (!RHSReg)
    return 0;

  MVT VT = std::max(MVT::i32, RetVT.SimpleTy);
  ResultReg = fastEmit_rr(VT, VT, ISDOpc, LHSReg, RHSReg);
  if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
    uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
    ResultReg = emitAnd_ri(MVT::i32, ResultReg, Mask);
  }
  return ResultReg;
}

unsigned AArch64FastISel::emitLogicalOp_ri(unsigned ISDOpc, MVT RetVT,
                                           unsigned LHSReg, uint64_t Imm) {
  static_assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR),
                "ISD nodes are not consecutive!");
  static const unsigned OpcTable[3][2] = {
    { AArch64::ANDWri, AArch64::ANDXri },
    { AArch64::ORRWri, AArch64::ORRXri },
    { AArch64::EORWri, AArch64::EORXri }
  };
  const TargetRegisterClass *RC;
  unsigned Opc;
  unsigned RegSize;
  switch (RetVT.SimpleTy) {
  default:
    return 0;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
  case MVT::i32: {
    unsigned Idx = ISDOpc - ISD::AND;
    Opc = OpcTable[Idx][0];
    RC = &AArch64::GPR32spRegClass;
    RegSize = 32;
    break;
  }
  case MVT::i64:
    Opc = OpcTable[ISDOpc - ISD::AND][1];
    RC = &AArch64::GPR64spRegClass;
    RegSize = 64;
    break;
  }

  if (!AArch64_AM::isLogicalImmediate(Imm, RegSize))
    return 0;

  Register ResultReg =
      fastEmitInst_ri(Opc, RC, LHSReg,
                      AArch64_AM::encodeLogicalImmediate(Imm, RegSize));
  if (RetVT >= MVT::i8 && RetVT <= MVT::i16 && ISDOpc != ISD::AND) {
    uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
    ResultReg = emitAnd_ri(MVT::i32, ResultReg, Mask);
  }
  return ResultReg;
}

unsigned AArch64FastISel::emitLogicalOp_rs(unsigned ISDOpc, MVT RetVT,
                                           unsigned LHSReg, unsigned RHSReg,
                                           uint64_t ShiftImm) {
  static_assert((ISD::AND + 1 == ISD::OR) && (ISD::AND + 2 == ISD::XOR),
                "ISD nodes are not consecutive!");
  static const unsigned OpcTable[3][2] = {
    { AArch64::ANDWrs, AArch64::ANDXrs },
    { AArch64::ORRWrs, AArch64::ORRXrs },
    { AArch64::EORWrs, AArch64::EORXrs }
  };

  // Don't deal with undefined shifts.
  if (ShiftImm >= RetVT.getSizeInBits())
    return 0;

  const TargetRegisterClass *RC;
  unsigned Opc;
  switch (RetVT.SimpleTy) {
  default:
    return 0;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
    Opc = OpcTable[ISDOpc - ISD::AND][0];
    RC = &AArch64::GPR32RegClass;
    break;
  case MVT::i64:
    Opc = OpcTable[ISDOpc - ISD::AND][1];
    RC = &AArch64::GPR64RegClass;
    break;
  }
  Register ResultReg =
      fastEmitInst_rri(Opc, RC, LHSReg, RHSReg,
                       AArch64_AM::getShifterImm(AArch64_AM::LSL, ShiftImm));
  if (RetVT >= MVT::i8 && RetVT <= MVT::i16) {
    uint64_t Mask = (RetVT == MVT::i8) ? 0xff : 0xffff;
    ResultReg = emitAnd_ri(MVT::i32, ResultReg, Mask);
  }
  return ResultReg;
}

unsigned AArch64FastISel::emitAnd_ri(MVT RetVT, unsigned LHSReg,
                                     uint64_t Imm) {
  return emitLogicalOp_ri(ISD::AND, RetVT, LHSReg, Imm);
}

unsigned AArch64FastISel::emitLoad(MVT VT, MVT RetVT, Address Addr,
                                   bool WantZExt, MachineMemOperand *MMO) {
  if (!TLI.allowsMisalignedMemoryAccesses(VT))
    return 0;

  // Simplify this down to something we can handle.
  if (!simplifyAddress(Addr, VT))
    return 0;

  unsigned ScaleFactor = getImplicitScaleFactor(VT);
  if (!ScaleFactor)
    llvm_unreachable("Unexpected value type.");

  // Negative offsets require unscaled, 9-bit, signed immediate offsets.
  // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
  bool UseScaled = true;
  if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
    UseScaled = false;
    ScaleFactor = 1;
  }

  static const unsigned GPOpcTable[2][8][4] = {
    // Sign-extend.
    { { AArch64::LDURSBWi,  AArch64::LDURSHWi,  AArch64::LDURWi,
        AArch64::LDURXi  },
      { AArch64::LDURSBXi,  AArch64::LDURSHXi,  AArch64::LDURSWi,
        AArch64::LDURXi  },
      { AArch64::LDRSBWui,  AArch64::LDRSHWui,  AArch64::LDRWui,
        AArch64::LDRXui  },
      { AArch64::LDRSBXui,  AArch64::LDRSHXui,  AArch64::LDRSWui,
        AArch64::LDRXui  },
      { AArch64::LDRSBWroX, AArch64::LDRSHWroX, AArch64::LDRWroX,
        AArch64::LDRXroX },
      { AArch64::LDRSBXroX, AArch64::LDRSHXroX, AArch64::LDRSWroX,
        AArch64::LDRXroX },
      { AArch64::LDRSBWroW, AArch64::LDRSHWroW, AArch64::LDRWroW,
        AArch64::LDRXroW },
      { AArch64::LDRSBXroW, AArch64::LDRSHXroW, AArch64::LDRSWroW,
        AArch64::LDRXroW }
    },
    // Zero-extend.
    { { AArch64::LDURBBi,   AArch64::LDURHHi,   AArch64::LDURWi,
        AArch64::LDURXi  },
      { AArch64::LDURBBi,   AArch64::LDURHHi,   AArch64::LDURWi,
        AArch64::LDURXi  },
      { AArch64::LDRBBui,   AArch64::LDRHHui,   AArch64::LDRWui,
        AArch64::LDRXui  },
      { AArch64::LDRBBui,   AArch64::LDRHHui,   AArch64::LDRWui,
        AArch64::LDRXui  },
      { AArch64::LDRBBroX,  AArch64::LDRHHroX,  AArch64::LDRWroX,
        AArch64::LDRXroX },
      { AArch64::LDRBBroX,  AArch64::LDRHHroX,  AArch64::LDRWroX,
        AArch64::LDRXroX },
      { AArch64::LDRBBroW,  AArch64::LDRHHroW,  AArch64::LDRWroW,
        AArch64::LDRXroW },
      { AArch64::LDRBBroW,  AArch64::LDRHHroW,  AArch64::LDRWroW,
        AArch64::LDRXroW }
    }
  };

  static const unsigned FPOpcTable[4][2] = {
    { AArch64::LDURSi,  AArch64::LDURDi  },
    { AArch64::LDRSui,  AArch64::LDRDui  },
    { AArch64::LDRSroX, AArch64::LDRDroX },
    { AArch64::LDRSroW, AArch64::LDRDroW }
  };

  unsigned Opc;
  const TargetRegisterClass *RC;
  bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
                      Addr.getOffsetReg();
  unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
  if (Addr.getExtendType() == AArch64_AM::UXTW ||
      Addr.getExtendType() == AArch64_AM::SXTW)
    Idx++;

  bool IsRet64Bit = RetVT == MVT::i64;
  switch (VT.SimpleTy) {
  default:
    llvm_unreachable("Unexpected value type.");
  case MVT::i1: // Intentional fall-through.
  case MVT::i8:
    Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][0];
    RC = (IsRet64Bit && !WantZExt) ?
             &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
    break;
  case MVT::i16:
    Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][1];
    RC = (IsRet64Bit && !WantZExt) ?
             &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
    break;
  case MVT::i32:
    Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][2];
    RC = (IsRet64Bit && !WantZExt) ?
             &AArch64::GPR64RegClass: &AArch64::GPR32RegClass;
    break;
  case MVT::i64:
    Opc = GPOpcTable[WantZExt][2 * Idx + IsRet64Bit][3];
    RC = &AArch64::GPR64RegClass;
    break;
  case MVT::f32:
    Opc = FPOpcTable[Idx][0];
    RC = &AArch64::FPR32RegClass;
    break;
  case MVT::f64:
    Opc = FPOpcTable[Idx][1];
    RC = &AArch64::FPR64RegClass;
    break;
  }

  // Create the base instruction, then add the operands.
  Register ResultReg = createResultReg(RC);
  MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
                                    TII.get(Opc), ResultReg);
  addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOLoad, ScaleFactor, MMO);

  // Loading an i1 requires special handling.
  if (VT == MVT::i1) {
    unsigned ANDReg = emitAnd_ri(MVT::i32, ResultReg, 1);
    assert(ANDReg && "Unexpected AND instruction emission failure.");
    ResultReg = ANDReg;
  }

  // For zero-extending loads to 64bit we emit a 32bit load and then convert
  // the 32bit reg to a 64bit reg.
  if (WantZExt && RetVT == MVT::i64 && VT <= MVT::i32) {
    Register Reg64 = createResultReg(&AArch64::GPR64RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(AArch64::SUBREG_TO_REG), Reg64)
        .addImm(0)
        .addReg(ResultReg, getKillRegState(true))
        .addImm(AArch64::sub_32);
    ResultReg = Reg64;
  }
  return ResultReg;
}

bool AArch64FastISel::selectAddSub(const Instruction *I) {
  MVT VT;
  if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
    return false;

  if (VT.isVector())
    return selectOperator(I, I->getOpcode());

  unsigned ResultReg;
  switch (I->getOpcode()) {
  default:
    llvm_unreachable("Unexpected instruction.");
  case Instruction::Add:
    ResultReg = emitAdd(VT, I->getOperand(0), I->getOperand(1));
    break;
  case Instruction::Sub:
    ResultReg = emitSub(VT, I->getOperand(0), I->getOperand(1));
    break;
  }
  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectLogicalOp(const Instruction *I) {
  MVT VT;
  if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
    return false;

  if (VT.isVector())
    return selectOperator(I, I->getOpcode());

  unsigned ResultReg;
  switch (I->getOpcode()) {
  default:
    llvm_unreachable("Unexpected instruction.");
  case Instruction::And:
    ResultReg = emitLogicalOp(ISD::AND, VT, I->getOperand(0), I->getOperand(1));
    break;
  case Instruction::Or:
    ResultReg = emitLogicalOp(ISD::OR, VT, I->getOperand(0), I->getOperand(1));
    break;
  case Instruction::Xor:
    ResultReg = emitLogicalOp(ISD::XOR, VT, I->getOperand(0), I->getOperand(1));
    break;
  }
  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectLoad(const Instruction *I) {
  MVT VT;
  // Verify we have a legal type before going any further.  Currently, we handle
  // simple types that will directly fit in a register (i32/f32/i64/f64) or
  // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
  if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true) ||
      cast<LoadInst>(I)->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;
    }
  }

  // See if we can handle this address.
  Address Addr;
  if (!computeAddress(I->getOperand(0), Addr, I->getType()))
    return false;

  // Fold the following sign-/zero-extend into the load instruction.
  bool WantZExt = true;
  MVT RetVT = VT;
  const Value *IntExtVal = nullptr;
  if (I->hasOneUse()) {
    if (const auto *ZE = dyn_cast<ZExtInst>(I->use_begin()->getUser())) {
      if (isTypeSupported(ZE->getType(), RetVT))
        IntExtVal = ZE;
      else
        RetVT = VT;
    } else if (const auto *SE = dyn_cast<SExtInst>(I->use_begin()->getUser())) {
      if (isTypeSupported(SE->getType(), RetVT))
        IntExtVal = SE;
      else
        RetVT = VT;
      WantZExt = false;
    }
  }

  unsigned ResultReg =
      emitLoad(VT, RetVT, Addr, WantZExt, createMachineMemOperandFor(I));
  if (!ResultReg)
    return false;

  // There are a few different cases we have to handle, because the load or the
  // sign-/zero-extend might not be selected by FastISel if we fall-back to
  // SelectionDAG. There is also an ordering issue when both instructions are in
  // different basic blocks.
  // 1.) The load instruction is selected by FastISel, but the integer extend
  //     not. This usually happens when the integer extend is in a different
  //     basic block and SelectionDAG took over for that basic block.
  // 2.) The load instruction is selected before the integer extend. This only
  //     happens when the integer extend is in a different basic block.
  // 3.) The load instruction is selected by SelectionDAG and the integer extend
  //     by FastISel. This happens if there are instructions between the load
  //     and the integer extend that couldn't be selected by FastISel.
  if (IntExtVal) {
    // The integer extend hasn't been emitted yet. FastISel or SelectionDAG
    // could select it. Emit a copy to subreg if necessary. FastISel will remove
    // it when it selects the integer extend.
    Register Reg = lookUpRegForValue(IntExtVal);
    auto *MI = MRI.getUniqueVRegDef(Reg);
    if (!MI) {
      if (RetVT == MVT::i64 && VT <= MVT::i32) {
        if (WantZExt) {
          // Delete the last emitted instruction from emitLoad (SUBREG_TO_REG).
          MachineBasicBlock::iterator I(std::prev(FuncInfo.InsertPt));
          ResultReg = std::prev(I)->getOperand(0).getReg();
          removeDeadCode(I, std::next(I));
        } else
          ResultReg = fastEmitInst_extractsubreg(MVT::i32, ResultReg,
                                                 AArch64::sub_32);
      }
      updateValueMap(I, ResultReg);
      return true;
    }

    // The integer extend has already been emitted - delete all the instructions
    // that have been emitted by the integer extend lowering code and use the
    // result from the load instruction directly.
    while (MI) {
      Reg = 0;
      for (auto &Opnd : MI->uses()) {
        if (Opnd.isReg()) {
          Reg = Opnd.getReg();
          break;
        }
      }
      MachineBasicBlock::iterator I(MI);
      removeDeadCode(I, std::next(I));
      MI = nullptr;
      if (Reg)
        MI = MRI.getUniqueVRegDef(Reg);
    }
    updateValueMap(IntExtVal, ResultReg);
    return true;
  }

  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::emitStoreRelease(MVT VT, unsigned SrcReg,
                                       unsigned AddrReg,
                                       MachineMemOperand *MMO) {
  unsigned Opc;
  switch (VT.SimpleTy) {
  default: return false;
  case MVT::i8:  Opc = AArch64::STLRB; break;
  case MVT::i16: Opc = AArch64::STLRH; break;
  case MVT::i32: Opc = AArch64::STLRW; break;
  case MVT::i64: Opc = AArch64::STLRX; break;
  }

  const MCInstrDesc &II = TII.get(Opc);
  SrcReg = constrainOperandRegClass(II, SrcReg, 0);
  AddrReg = constrainOperandRegClass(II, AddrReg, 1);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II)
      .addReg(SrcReg)
      .addReg(AddrReg)
      .addMemOperand(MMO);
  return true;
}

bool AArch64FastISel::emitStore(MVT VT, unsigned SrcReg, Address Addr,
                                MachineMemOperand *MMO) {
  if (!TLI.allowsMisalignedMemoryAccesses(VT))
    return false;

  // Simplify this down to something we can handle.
  if (!simplifyAddress(Addr, VT))
    return false;

  unsigned ScaleFactor = getImplicitScaleFactor(VT);
  if (!ScaleFactor)
    llvm_unreachable("Unexpected value type.");

  // Negative offsets require unscaled, 9-bit, signed immediate offsets.
  // Otherwise, we try using scaled, 12-bit, unsigned immediate offsets.
  bool UseScaled = true;
  if ((Addr.getOffset() < 0) || (Addr.getOffset() & (ScaleFactor - 1))) {
    UseScaled = false;
    ScaleFactor = 1;
  }

  static const unsigned OpcTable[4][6] = {
    { AArch64::STURBBi,  AArch64::STURHHi,  AArch64::STURWi,  AArch64::STURXi,
      AArch64::STURSi,   AArch64::STURDi },
    { AArch64::STRBBui,  AArch64::STRHHui,  AArch64::STRWui,  AArch64::STRXui,
      AArch64::STRSui,   AArch64::STRDui },
    { AArch64::STRBBroX, AArch64::STRHHroX, AArch64::STRWroX, AArch64::STRXroX,
      AArch64::STRSroX,  AArch64::STRDroX },
    { AArch64::STRBBroW, AArch64::STRHHroW, AArch64::STRWroW, AArch64::STRXroW,
      AArch64::STRSroW,  AArch64::STRDroW }
  };

  unsigned Opc;
  bool VTIsi1 = false;
  bool UseRegOffset = Addr.isRegBase() && !Addr.getOffset() && Addr.getReg() &&
                      Addr.getOffsetReg();
  unsigned Idx = UseRegOffset ? 2 : UseScaled ? 1 : 0;
  if (Addr.getExtendType() == AArch64_AM::UXTW ||
      Addr.getExtendType() == AArch64_AM::SXTW)
    Idx++;

  switch (VT.SimpleTy) {
  default: llvm_unreachable("Unexpected value type.");
  case MVT::i1:  VTIsi1 = true; [[fallthrough]];
  case MVT::i8:  Opc = OpcTable[Idx][0]; break;
  case MVT::i16: Opc = OpcTable[Idx][1]; break;
  case MVT::i32: Opc = OpcTable[Idx][2]; break;
  case MVT::i64: Opc = OpcTable[Idx][3]; break;
  case MVT::f32: Opc = OpcTable[Idx][4]; break;
  case MVT::f64: Opc = OpcTable[Idx][5]; break;
  }

  // Storing an i1 requires special handling.
  if (VTIsi1 && SrcReg != AArch64::WZR) {
    unsigned ANDReg = emitAnd_ri(MVT::i32, SrcReg, 1);
    assert(ANDReg && "Unexpected AND instruction emission failure.");
    SrcReg = ANDReg;
  }
  // Create the base instruction, then add the operands.
  const MCInstrDesc &II = TII.get(Opc);
  SrcReg = constrainOperandRegClass(II, SrcReg, II.getNumDefs());
  MachineInstrBuilder MIB =
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II).addReg(SrcReg);
  addLoadStoreOperands(Addr, MIB, MachineMemOperand::MOStore, ScaleFactor, MMO);

  return true;
}

bool AArch64FastISel::selectStore(const Instruction *I) {
  MVT VT;
  const Value *Op0 = I->getOperand(0);
  // Verify we have a legal type before going any further.  Currently, we handle
  // simple types that will directly fit in a register (i32/f32/i64/f64) or
  // those that can be sign or zero-extended to a basic operation (i1/i8/i16).
  if (!isTypeSupported(Op0->getType(), VT, /*IsVectorAllowed=*/true))
    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;
    }
  }

  // Get the value to be stored into a register. Use the zero register directly
  // when possible to avoid an unnecessary copy and a wasted register.
  unsigned SrcReg = 0;
  if (const auto *CI = dyn_cast<ConstantInt>(Op0)) {
    if (CI->isZero())
      SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
  } else if (const auto *CF = dyn_cast<ConstantFP>(Op0)) {
    if (CF->isZero() && !CF->isNegative()) {
      VT = MVT::getIntegerVT(VT.getSizeInBits());
      SrcReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
    }
  }

  if (!SrcReg)
    SrcReg = getRegForValue(Op0);

  if (!SrcReg)
    return false;

  auto *SI = cast<StoreInst>(I);

  // Try to emit a STLR for seq_cst/release.
  if (SI->isAtomic()) {
    AtomicOrdering Ord = SI->getOrdering();
    // The non-atomic instructions are sufficient for relaxed stores.
    if (isReleaseOrStronger(Ord)) {
      // The STLR addressing mode only supports a base reg; pass that directly.
      Register AddrReg = getRegForValue(PtrV);
      return emitStoreRelease(VT, SrcReg, AddrReg,
                              createMachineMemOperandFor(I));
    }
  }

  // See if we can handle this address.
  Address Addr;
  if (!computeAddress(PtrV, Addr, Op0->getType()))
    return false;

  if (!emitStore(VT, SrcReg, Addr, createMachineMemOperandFor(I)))
    return false;
  return true;
}

static AArch64CC::CondCode getCompareCC(CmpInst::Predicate Pred) {
  switch (Pred) {
  case CmpInst::FCMP_ONE:
  case CmpInst::FCMP_UEQ:
  default:
    // AL is our "false" for now. The other two need more compares.
    return AArch64CC::AL;
  case CmpInst::ICMP_EQ:
  case CmpInst::FCMP_OEQ:
    return AArch64CC::EQ;
  case CmpInst::ICMP_SGT:
  case CmpInst::FCMP_OGT:
    return AArch64CC::GT;
  case CmpInst::ICMP_SGE:
  case CmpInst::FCMP_OGE:
    return AArch64CC::GE;
  case CmpInst::ICMP_UGT:
  case CmpInst::FCMP_UGT:
    return AArch64CC::HI;
  case CmpInst::FCMP_OLT:
    return AArch64CC::MI;
  case CmpInst::ICMP_ULE:
  case CmpInst::FCMP_OLE:
    return AArch64CC::LS;
  case CmpInst::FCMP_ORD:
    return AArch64CC::VC;
  case CmpInst::FCMP_UNO:
    return AArch64CC::VS;
  case CmpInst::FCMP_UGE:
    return AArch64CC::PL;
  case CmpInst::ICMP_SLT:
  case CmpInst::FCMP_ULT:
    return AArch64CC::LT;
  case CmpInst::ICMP_SLE:
  case CmpInst::FCMP_ULE:
    return AArch64CC::LE;
  case CmpInst::FCMP_UNE:
  case CmpInst::ICMP_NE:
    return AArch64CC::NE;
  case CmpInst::ICMP_UGE:
    return AArch64CC::HS;
  case CmpInst::ICMP_ULT:
    return AArch64CC::LO;
  }
}

/// Try to emit a combined compare-and-branch instruction.
bool AArch64FastISel::emitCompareAndBranch(const BranchInst *BI) {
  // Speculation tracking/SLH assumes that optimized TB(N)Z/CB(N)Z instructions
  // will not be produced, as they are conditional branch instructions that do
  // not set flags.
  if (FuncInfo.MF->getFunction().hasFnAttribute(
          Attribute::SpeculativeLoadHardening))
    return false;

  assert(isa<CmpInst>(BI->getCondition()) && "Expected cmp instruction");
  const CmpInst *CI = cast<CmpInst>(BI->getCondition());
  CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);

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

  MVT VT;
  if (!isTypeSupported(LHS->getType(), VT))
    return false;

  unsigned BW = VT.getSizeInBits();
  if (BW > 64)
    return false;

  MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
  MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];

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

  int TestBit = -1;
  bool IsCmpNE;
  switch (Predicate) {
  default:
    return false;
  case CmpInst::ICMP_EQ:
  case CmpInst::ICMP_NE:
    if (isa<Constant>(LHS) && cast<Constant>(LHS)->isNullValue())
      std::swap(LHS, RHS);

    if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
      return false;

    if (const auto *AI = dyn_cast<BinaryOperator>(LHS))
      if (AI->getOpcode() == Instruction::And && isValueAvailable(AI)) {
        const Value *AndLHS = AI->getOperand(0);
        const Value *AndRHS = AI->getOperand(1);

        if (const auto *C = dyn_cast<ConstantInt>(AndLHS))
          if (C->getValue().isPowerOf2())
            std::swap(AndLHS, AndRHS);

        if (const auto *C = dyn_cast<ConstantInt>(AndRHS))
          if (C->getValue().isPowerOf2()) {
            TestBit = C->getValue().logBase2();
            LHS = AndLHS;
          }
      }

    if (VT == MVT::i1)
      TestBit = 0;

    IsCmpNE = Predicate == CmpInst::ICMP_NE;
    break;
  case CmpInst::ICMP_SLT:
  case CmpInst::ICMP_SGE:
    if (!isa<Constant>(RHS) || !cast<Constant>(RHS)->isNullValue())
      return false;

    TestBit = BW - 1;
    IsCmpNE = Predicate == CmpInst::ICMP_SLT;
    break;
  case CmpInst::ICMP_SGT:
  case CmpInst::ICMP_SLE:
    if (!isa<ConstantInt>(RHS))
      return false;

    if (cast<ConstantInt>(RHS)->getValue() != APInt(BW, -1, true))
      return false;

    TestBit = BW - 1;
    IsCmpNE = Predicate == CmpInst::ICMP_SLE;
    break;
  } // end switch

  static const unsigned OpcTable[2][2][2] = {
    { {AArch64::CBZW,  AArch64::CBZX },
      {AArch64::CBNZW, AArch64::CBNZX} },
    { {AArch64::TBZW,  AArch64::TBZX },
      {AArch64::TBNZW, AArch64::TBNZX} }
  };

  bool IsBitTest = TestBit != -1;
  bool Is64Bit = BW == 64;
  if (TestBit < 32 && TestBit >= 0)
    Is64Bit = false;

  unsigned Opc = OpcTable[IsBitTest][IsCmpNE][Is64Bit];
  const MCInstrDesc &II = TII.get(Opc);

  Register SrcReg = getRegForValue(LHS);
  if (!SrcReg)
    return false;

  if (BW == 64 && !Is64Bit)
    SrcReg = fastEmitInst_extractsubreg(MVT::i32, SrcReg, AArch64::sub_32);

  if ((BW < 32) && !IsBitTest)
    SrcReg = emitIntExt(VT, SrcReg, MVT::i32, /*isZExt=*/true);

  // Emit the combined compare and branch instruction.
  SrcReg = constrainOperandRegClass(II, SrcReg,  II.getNumDefs());
  MachineInstrBuilder MIB =
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(Opc))
          .addReg(SrcReg);
  if (IsBitTest)
    MIB.addImm(TestBit);
  MIB.addMBB(TBB);

  finishCondBranch(BI->getParent(), TBB, FBB);
  return true;
}

bool AArch64FastISel::selectBranch(const Instruction *I) {
  const BranchInst *BI = cast<BranchInst>(I);
  if (BI->isUnconditional()) {
    MachineBasicBlock *MSucc = FuncInfo.MBBMap[BI->getSuccessor(0)];
    fastEmitBranch(MSucc, BI->getDebugLoc());
    return true;
  }

  MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
  MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];

  if (const CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
    if (CI->hasOneUse() && isValueAvailable(CI)) {
      // Try to optimize or fold the cmp.
      CmpInst::Predicate Predicate = optimizeCmpPredicate(CI);
      switch (Predicate) {
      default:
        break;
      case CmpInst::FCMP_FALSE:
        fastEmitBranch(FBB, MIMD.getDL());
        return true;
      case CmpInst::FCMP_TRUE:
        fastEmitBranch(TBB, MIMD.getDL());
        return true;
      }

      // Try to emit a combined compare-and-branch first.
      if (emitCompareAndBranch(BI))
        return true;

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

      // Emit the cmp.
      if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
        return false;

      // FCMP_UEQ and FCMP_ONE cannot be checked with a single branch
      // instruction.
      AArch64CC::CondCode CC = getCompareCC(Predicate);
      AArch64CC::CondCode ExtraCC = AArch64CC::AL;
      switch (Predicate) {
      default:
        break;
      case CmpInst::FCMP_UEQ:
        ExtraCC = AArch64CC::EQ;
        CC = AArch64CC::VS;
        break;
      case CmpInst::FCMP_ONE:
        ExtraCC = AArch64CC::MI;
        CC = AArch64CC::GT;
        break;
      }
      assert((CC != AArch64CC::AL) && "Unexpected condition code.");

      // Emit the extra branch for FCMP_UEQ and FCMP_ONE.
      if (ExtraCC != AArch64CC::AL) {
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::Bcc))
            .addImm(ExtraCC)
            .addMBB(TBB);
      }

      // Emit the branch.
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::Bcc))
          .addImm(CC)
          .addMBB(TBB);

      finishCondBranch(BI->getParent(), TBB, FBB);
      return true;
    }
  } else if (const auto *CI = dyn_cast<ConstantInt>(BI->getCondition())) {
    uint64_t Imm = CI->getZExtValue();
    MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::B))
        .addMBB(Target);

    // Obtain the branch probability and add the target to the successor list.
    if (FuncInfo.BPI) {
      auto BranchProbability = FuncInfo.BPI->getEdgeProbability(
          BI->getParent(), Target->getBasicBlock());
      FuncInfo.MBB->addSuccessor(Target, BranchProbability);
    } else
      FuncInfo.MBB->addSuccessorWithoutProb(Target);
    return true;
  } else {
    AArch64CC::CondCode CC = AArch64CC::NE;
    if (foldXALUIntrinsic(CC, I, BI->getCondition())) {
      // Fake request the condition, otherwise the intrinsic might be completely
      // optimized away.
      Register CondReg = getRegForValue(BI->getCondition());
      if (!CondReg)
        return false;

      // Emit the branch.
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::Bcc))
        .addImm(CC)
        .addMBB(TBB);

      finishCondBranch(BI->getParent(), TBB, FBB);
      return true;
    }
  }

  Register CondReg = getRegForValue(BI->getCondition());
  if (CondReg == 0)
    return false;

  // i1 conditions come as i32 values, test the lowest bit with tb(n)z.
  unsigned Opcode = AArch64::TBNZW;
  if (FuncInfo.MBB->isLayoutSuccessor(TBB)) {
    std::swap(TBB, FBB);
    Opcode = AArch64::TBZW;
  }

  const MCInstrDesc &II = TII.get(Opcode);
  Register ConstrainedCondReg
    = constrainOperandRegClass(II, CondReg, II.getNumDefs());
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II)
      .addReg(ConstrainedCondReg)
      .addImm(0)
      .addMBB(TBB);

  finishCondBranch(BI->getParent(), TBB, FBB);
  return true;
}

bool AArch64FastISel::selectIndirectBr(const Instruction *I) {
  const IndirectBrInst *BI = cast<IndirectBrInst>(I);
  Register AddrReg = getRegForValue(BI->getOperand(0));
  if (AddrReg == 0)
    return false;

  // Emit the indirect branch.
  const MCInstrDesc &II = TII.get(AArch64::BR);
  AddrReg = constrainOperandRegClass(II, AddrReg,  II.getNumDefs());
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II).addReg(AddrReg);

  // Make sure the CFG is up-to-date.
  for (const auto *Succ : BI->successors())
    FuncInfo.MBB->addSuccessor(FuncInfo.MBBMap[Succ]);

  return true;
}

bool AArch64FastISel::selectCmp(const Instruction *I) {
  const CmpInst *CI = cast<CmpInst>(I);

  // Vectors of i1 are weird: bail out.
  if (CI->getType()->isVectorTy())
    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(&AArch64::GPR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(TargetOpcode::COPY), ResultReg)
        .addReg(AArch64::WZR, getKillRegState(true));
    break;
  case CmpInst::FCMP_TRUE:
    ResultReg = fastEmit_i(MVT::i32, MVT::i32, ISD::Constant, 1);
    break;
  }

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

  // Emit the cmp.
  if (!emitCmp(CI->getOperand(0), CI->getOperand(1), CI->isUnsigned()))
    return false;

  ResultReg = createResultReg(&AArch64::GPR32RegClass);

  // FCMP_UEQ and FCMP_ONE cannot be checked with a single instruction. These
  // condition codes are inverted, because they are used by CSINC.
  static unsigned CondCodeTable[2][2] = {
    { AArch64CC::NE, AArch64CC::VC },
    { AArch64CC::PL, AArch64CC::LE }
  };
  unsigned *CondCodes = nullptr;
  switch (Predicate) {
  default:
    break;
  case CmpInst::FCMP_UEQ:
    CondCodes = &CondCodeTable[0][0];
    break;
  case CmpInst::FCMP_ONE:
    CondCodes = &CondCodeTable[1][0];
    break;
  }

  if (CondCodes) {
    Register TmpReg1 = createResultReg(&AArch64::GPR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::CSINCWr),
            TmpReg1)
        .addReg(AArch64::WZR, getKillRegState(true))
        .addReg(AArch64::WZR, getKillRegState(true))
        .addImm(CondCodes[0]);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::CSINCWr),
            ResultReg)
        .addReg(TmpReg1, getKillRegState(true))
        .addReg(AArch64::WZR, getKillRegState(true))
        .addImm(CondCodes[1]);

    updateValueMap(I, ResultReg);
    return true;
  }

  // Now set a register based on the comparison.
  AArch64CC::CondCode CC = getCompareCC(Predicate);
  assert((CC != AArch64CC::AL) && "Unexpected condition code.");
  AArch64CC::CondCode invertedCC = getInvertedCondCode(CC);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::CSINCWr),
          ResultReg)
      .addReg(AArch64::WZR, getKillRegState(true))
      .addReg(AArch64::WZR, getKillRegState(true))
      .addImm(invertedCC);

  updateValueMap(I, ResultReg);
  return true;
}

/// Optimize selects of i1 if one of the operands has a 'true' or 'false'
/// value.
bool AArch64FastISel::optimizeSelect(const SelectInst *SI) {
  if (!SI->getType()->isIntegerTy(1))
    return false;

  const Value *Src1Val, *Src2Val;
  unsigned Opc = 0;
  bool NeedExtraOp = false;
  if (auto *CI = dyn_cast<ConstantInt>(SI->getTrueValue())) {
    if (CI->isOne()) {
      Src1Val = SI->getCondition();
      Src2Val = SI->getFalseValue();
      Opc = AArch64::ORRWrr;
    } else {
      assert(CI->isZero());
      Src1Val = SI->getFalseValue();
      Src2Val = SI->getCondition();
      Opc = AArch64::BICWrr;
    }
  } else if (auto *CI = dyn_cast<ConstantInt>(SI->getFalseValue())) {
    if (CI->isOne()) {
      Src1Val = SI->getCondition();
      Src2Val = SI->getTrueValue();
      Opc = AArch64::ORRWrr;
      NeedExtraOp = true;
    } else {
      assert(CI->isZero());
      Src1Val = SI->getCondition();
      Src2Val = SI->getTrueValue();
      Opc = AArch64::ANDWrr;
    }
  }

  if (!Opc)
    return false;

  Register Src1Reg = getRegForValue(Src1Val);
  if (!Src1Reg)
    return false;

  Register Src2Reg = getRegForValue(Src2Val);
  if (!Src2Reg)
    return false;

  if (NeedExtraOp)
    Src1Reg = emitLogicalOp_ri(ISD::XOR, MVT::i32, Src1Reg, 1);

  Register ResultReg = fastEmitInst_rr(Opc, &AArch64::GPR32RegClass, Src1Reg,
                                       Src2Reg);
  updateValueMap(SI, ResultReg);
  return true;
}

bool AArch64FastISel::selectSelect(const Instruction *I) {
  assert(isa<SelectInst>(I) && "Expected a select instruction.");
  MVT VT;
  if (!isTypeSupported(I->getType(), VT))
    return false;

  unsigned Opc;
  const TargetRegisterClass *RC;
  switch (VT.SimpleTy) {
  default:
    return false;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
    Opc = AArch64::CSELWr;
    RC = &AArch64::GPR32RegClass;
    break;
  case MVT::i64:
    Opc = AArch64::CSELXr;
    RC = &AArch64::GPR64RegClass;
    break;
  case MVT::f32:
    Opc = AArch64::FCSELSrrr;
    RC = &AArch64::FPR32RegClass;
    break;
  case MVT::f64:
    Opc = AArch64::FCSELDrrr;
    RC = &AArch64::FPR64RegClass;
    break;
  }

  const SelectInst *SI = cast<SelectInst>(I);
  const Value *Cond = SI->getCondition();
  AArch64CC::CondCode CC = AArch64CC::NE;
  AArch64CC::CondCode ExtraCC = AArch64CC::AL;

  if (optimizeSelect(SI))
    return true;

  // Try to pickup the flags, so we don't have to emit another compare.
  if (foldXALUIntrinsic(CC, I, Cond)) {
    // Fake request the condition to force emission of the XALU intrinsic.
    Register CondReg = getRegForValue(Cond);
    if (!CondReg)
      return false;
  } else if (isa<CmpInst>(Cond) && cast<CmpInst>(Cond)->hasOneUse() &&
             isValueAvailable(Cond)) {
    const auto *Cmp = cast<CmpInst>(Cond);
    // Try to optimize or fold the cmp.
    CmpInst::Predicate Predicate = optimizeCmpPredicate(Cmp);
    const Value *FoldSelect = nullptr;
    switch (Predicate) {
    default:
      break;
    case CmpInst::FCMP_FALSE:
      FoldSelect = SI->getFalseValue();
      break;
    case CmpInst::FCMP_TRUE:
      FoldSelect = SI->getTrueValue();
      break;
    }

    if (FoldSelect) {
      Register SrcReg = getRegForValue(FoldSelect);
      if (!SrcReg)
        return false;

      updateValueMap(I, SrcReg);
      return true;
    }

    // Emit the cmp.
    if (!emitCmp(Cmp->getOperand(0), Cmp->getOperand(1), Cmp->isUnsigned()))
      return false;

    // FCMP_UEQ and FCMP_ONE cannot be checked with a single select instruction.
    CC = getCompareCC(Predicate);
    switch (Predicate) {
    default:
      break;
    case CmpInst::FCMP_UEQ:
      ExtraCC = AArch64CC::EQ;
      CC = AArch64CC::VS;
      break;
    case CmpInst::FCMP_ONE:
      ExtraCC = AArch64CC::MI;
      CC = AArch64CC::GT;
      break;
    }
    assert((CC != AArch64CC::AL) && "Unexpected condition code.");
  } else {
    Register CondReg = getRegForValue(Cond);
    if (!CondReg)
      return false;

    const MCInstrDesc &II = TII.get(AArch64::ANDSWri);
    CondReg = constrainOperandRegClass(II, CondReg, 1);

    // Emit a TST instruction (ANDS wzr, reg, #imm).
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II,
            AArch64::WZR)
        .addReg(CondReg)
        .addImm(AArch64_AM::encodeLogicalImmediate(1, 32));
  }

  Register Src1Reg = getRegForValue(SI->getTrueValue());
  Register Src2Reg = getRegForValue(SI->getFalseValue());

  if (!Src1Reg || !Src2Reg)
    return false;

  if (ExtraCC != AArch64CC::AL)
    Src2Reg = fastEmitInst_rri(Opc, RC, Src1Reg, Src2Reg, ExtraCC);

  Register ResultReg = fastEmitInst_rri(Opc, RC, Src1Reg, Src2Reg, CC);
  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectFPExt(const Instruction *I) {
  Value *V = I->getOperand(0);
  if (!I->getType()->isDoubleTy() || !V->getType()->isFloatTy())
    return false;

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

  Register ResultReg = createResultReg(&AArch64::FPR64RegClass);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::FCVTDSr),
          ResultReg).addReg(Op);
  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectFPTrunc(const Instruction *I) {
  Value *V = I->getOperand(0);
  if (!I->getType()->isFloatTy() || !V->getType()->isDoubleTy())
    return false;

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

  Register ResultReg = createResultReg(&AArch64::FPR32RegClass);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::FCVTSDr),
          ResultReg).addReg(Op);
  updateValueMap(I, ResultReg);
  return true;
}

// FPToUI and FPToSI
bool AArch64FastISel::selectFPToInt(const Instruction *I, bool Signed) {
  MVT DestVT;
  if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
    return false;

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

  EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType(), true);
  if (SrcVT == MVT::f128 || SrcVT == MVT::f16)
    return false;

  unsigned Opc;
  if (SrcVT == MVT::f64) {
    if (Signed)
      Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWDr : AArch64::FCVTZSUXDr;
    else
      Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWDr : AArch64::FCVTZUUXDr;
  } else {
    if (Signed)
      Opc = (DestVT == MVT::i32) ? AArch64::FCVTZSUWSr : AArch64::FCVTZSUXSr;
    else
      Opc = (DestVT == MVT::i32) ? AArch64::FCVTZUUWSr : AArch64::FCVTZUUXSr;
  }
  Register ResultReg = createResultReg(
      DestVT == MVT::i32 ? &AArch64::GPR32RegClass : &AArch64::GPR64RegClass);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(Opc), ResultReg)
      .addReg(SrcReg);
  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectIntToFP(const Instruction *I, bool Signed) {
  MVT DestVT;
  if (!isTypeLegal(I->getType(), DestVT) || DestVT.isVector())
    return false;
  // Let regular ISEL handle FP16
  if (DestVT == MVT::f16)
    return false;

  assert((DestVT == MVT::f32 || DestVT == MVT::f64) &&
         "Unexpected value type.");

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

  EVT SrcVT = TLI.getValueType(DL, I->getOperand(0)->getType(), true);

  // Handle sign-extension.
  if (SrcVT == MVT::i16 || SrcVT == MVT::i8 || SrcVT == MVT::i1) {
    SrcReg =
        emitIntExt(SrcVT.getSimpleVT(), SrcReg, MVT::i32, /*isZExt*/ !Signed);
    if (!SrcReg)
      return false;
  }

  unsigned Opc;
  if (SrcVT == MVT::i64) {
    if (Signed)
      Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUXSri : AArch64::SCVTFUXDri;
    else
      Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUXSri : AArch64::UCVTFUXDri;
  } else {
    if (Signed)
      Opc = (DestVT == MVT::f32) ? AArch64::SCVTFUWSri : AArch64::SCVTFUWDri;
    else
      Opc = (DestVT == MVT::f32) ? AArch64::UCVTFUWSri : AArch64::UCVTFUWDri;
  }

  Register ResultReg = fastEmitInst_r(Opc, TLI.getRegClassFor(DestVT), SrcReg);
  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::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 && CC != CallingConv::Swift)
    return false;

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

  // Only handle simple cases of up to 8 GPR and FPR each.
  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::SwiftAsync) ||
        Arg.hasAttribute(Attribute::SwiftError) ||
        Arg.hasAttribute(Attribute::Nest))
      return false;

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

    EVT ArgVT = TLI.getValueType(DL, ArgTy);
    if (!ArgVT.isSimple())
      return false;

    MVT VT = ArgVT.getSimpleVT().SimpleTy;
    if (VT.isFloatingPoint() && !Subtarget->hasFPARMv8())
      return false;

    if (VT.isVector() &&
        (!Subtarget->hasNEON() || !Subtarget->isLittleEndian()))
      return false;

    if (VT >= MVT::i1 && VT <= MVT::i64)
      ++GPRCnt;
    else if ((VT >= MVT::f16 && VT <= MVT::f64) || VT.is64BitVector() ||
             VT.is128BitVector())
      ++FPRCnt;
    else
      return false;

    if (GPRCnt > 8 || FPRCnt > 8)
      return false;
  }

  static const MCPhysReg Registers[6][8] = {
    { AArch64::W0, AArch64::W1, AArch64::W2, AArch64::W3, AArch64::W4,
      AArch64::W5, AArch64::W6, AArch64::W7 },
    { AArch64::X0, AArch64::X1, AArch64::X2, AArch64::X3, AArch64::X4,
      AArch64::X5, AArch64::X6, AArch64::X7 },
    { AArch64::H0, AArch64::H1, AArch64::H2, AArch64::H3, AArch64::H4,
      AArch64::H5, AArch64::H6, AArch64::H7 },
    { AArch64::S0, AArch64::S1, AArch64::S2, AArch64::S3, AArch64::S4,
      AArch64::S5, AArch64::S6, AArch64::S7 },
    { AArch64::D0, AArch64::D1, AArch64::D2, AArch64::D3, AArch64::D4,
      AArch64::D5, AArch64::D6, AArch64::D7 },
    { AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3, AArch64::Q4,
      AArch64::Q5, AArch64::Q6, AArch64::Q7 }
  };

  unsigned GPRIdx = 0;
  unsigned FPRIdx = 0;
  for (auto const &Arg : F->args()) {
    MVT VT = TLI.getSimpleValueType(DL, Arg.getType());
    unsigned SrcReg;
    const TargetRegisterClass *RC;
    if (VT >= MVT::i1 && VT <= MVT::i32) {
      SrcReg = Registers[0][GPRIdx++];
      RC = &AArch64::GPR32RegClass;
      VT = MVT::i32;
    } else if (VT == MVT::i64) {
      SrcReg = Registers[1][GPRIdx++];
      RC = &AArch64::GPR64RegClass;
    } else if (VT == MVT::f16) {
      SrcReg = Registers[2][FPRIdx++];
      RC = &AArch64::FPR16RegClass;
    } else if (VT ==  MVT::f32) {
      SrcReg = Registers[3][FPRIdx++];
      RC = &AArch64::FPR32RegClass;
    } else if ((VT == MVT::f64) || VT.is64BitVector()) {
      SrcReg = Registers[4][FPRIdx++];
      RC = &AArch64::FPR64RegClass;
    } else if (VT.is128BitVector()) {
      SrcReg = Registers[5][FPRIdx++];
      RC = &AArch64::FPR128RegClass;
    } else
      llvm_unreachable("Unexpected value type.");

    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, MIMD,
            TII.get(TargetOpcode::COPY), ResultReg)
        .addReg(DstReg, getKillRegState(true));
    updateValueMap(&Arg, ResultReg);
  }
  return true;
}

bool AArch64FastISel::processCallArgs(CallLoweringInfo &CLI,
                                      SmallVectorImpl<MVT> &OutVTs,
                                      unsigned &NumBytes) {
  CallingConv::ID CC = CLI.CallConv;
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CC, false, *FuncInfo.MF, ArgLocs, *Context);
  CCInfo.AnalyzeCallOperands(OutVTs, CLI.OutFlags, CCAssignFnForCall(CC));

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

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

  // Process the args.
  for (CCValAssign &VA : ArgLocs) {
    const Value *ArgVal = CLI.OutVals[VA.getValNo()];
    MVT ArgVT = OutVTs[VA.getValNo()];

    Register ArgReg = getRegForValue(ArgVal);
    if (!ArgReg)
      return false;

    // Handle arg promotion: SExt, ZExt, AExt.
    switch (VA.getLocInfo()) {
    case CCValAssign::Full:
      break;
    case CCValAssign::SExt: {
      MVT DestVT = VA.getLocVT();
      MVT SrcVT = ArgVT;
      ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/false);
      if (!ArgReg)
        return false;
      break;
    }
    case CCValAssign::AExt:
    // Intentional fall-through.
    case CCValAssign::ZExt: {
      MVT DestVT = VA.getLocVT();
      MVT SrcVT = ArgVT;
      ArgReg = emitIntExt(SrcVT, ArgReg, DestVT, /*isZExt=*/true);
      if (!ArgReg)
        return false;
      break;
    }
    default:
      llvm_unreachable("Unknown arg promotion!");
    }

    // Now copy/store arg to correct locations.
    if (VA.isRegLoc() && !VA.needsCustom()) {
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
              TII.get(TargetOpcode::COPY), VA.getLocReg()).addReg(ArgReg);
      CLI.OutRegs.push_back(VA.getLocReg());
    } else if (VA.needsCustom()) {
      // FIXME: Handle custom args.
      return false;
    } else {
      assert(VA.isMemLoc() && "Assuming store on stack.");

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

      // Need to store on the stack.
      unsigned ArgSize = (ArgVT.getSizeInBits() + 7) / 8;

      unsigned BEAlign = 0;
      if (ArgSize < 8 && !Subtarget->isLittleEndian())
        BEAlign = 8 - ArgSize;

      Address Addr;
      Addr.setKind(Address::RegBase);
      Addr.setReg(AArch64::SP);
      Addr.setOffset(VA.getLocMemOffset() + BEAlign);

      Align Alignment = DL.getABITypeAlign(ArgVal->getType());
      MachineMemOperand *MMO = FuncInfo.MF->getMachineMemOperand(
          MachinePointerInfo::getStack(*FuncInfo.MF, Addr.getOffset()),
          MachineMemOperand::MOStore, ArgVT.getStoreSize(), Alignment);

      if (!emitStore(ArgVT, ArgReg, Addr, MMO))
        return false;
    }
  }
  return true;
}

bool AArch64FastISel::finishCall(CallLoweringInfo &CLI, unsigned NumBytes) {
  CallingConv::ID CC = CLI.CallConv;

  // Issue CALLSEQ_END
  unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AdjStackUp))
    .addImm(NumBytes).addImm(0);

  // Now the return values.
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CC, false, *FuncInfo.MF, RVLocs, *Context);
  CCInfo.AnalyzeCallResult(CLI.Ins, CCAssignFnForCall(CC));

  Register ResultReg = FuncInfo.CreateRegs(CLI.RetTy);
  for (unsigned i = 0; i != RVLocs.size(); ++i) {
    CCValAssign &VA = RVLocs[i];
    MVT CopyVT = VA.getValVT();
    unsigned CopyReg = ResultReg + i;

    // TODO: Handle big-endian results
    if (CopyVT.isVector() && !Subtarget->isLittleEndian())
      return false;

    // Copy result out of their specified physreg.
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(TargetOpcode::COPY),
            CopyReg)
        .addReg(VA.getLocReg());
    CLI.InRegs.push_back(VA.getLocReg());
  }

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

  return true;
}

bool AArch64FastISel::fastLowerCall(CallLoweringInfo &CLI) {
  CallingConv::ID CC  = CLI.CallConv;
  bool IsTailCall     = CLI.IsTailCall;
  bool IsVarArg       = CLI.IsVarArg;
  const Value *Callee = CLI.Callee;
  MCSymbol *Symbol = CLI.Symbol;

  if (!Callee && !Symbol)
    return false;

  // Allow SelectionDAG isel to handle calls to functions like setjmp that need
  // a bti instruction following the call.
  if (CLI.CB && CLI.CB->hasFnAttr(Attribute::ReturnsTwice) &&
      !Subtarget->noBTIAtReturnTwice() &&
      MF->getInfo<AArch64FunctionInfo>()->branchTargetEnforcement())
    return false;

  // Allow SelectionDAG isel to handle indirect calls with KCFI checks.
  if (CLI.CB && CLI.CB->isIndirectCall() &&
      CLI.CB->getOperandBundle(LLVMContext::OB_kcfi))
    return false;

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

  // FIXME: we could and should support this, but for now correctness at -O0 is
  // more important.
  if (Subtarget->isTargetILP32())
    return false;

  CodeModel::Model CM = TM.getCodeModel();
  // Only support the small-addressing and large code models.
  if (CM != CodeModel::Large && !Subtarget->useSmallAddressing())
    return false;

  // FIXME: Add large code model support for ELF.
  if (CM == CodeModel::Large && !Subtarget->isTargetMachO())
    return false;

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

  if (Subtarget->isWindowsArm64EC())
    return false;

  for (auto Flag : CLI.OutFlags)
    if (Flag.isInReg() || Flag.isSRet() || Flag.isNest() || Flag.isByVal() ||
        Flag.isSwiftSelf() || Flag.isSwiftAsync() || Flag.isSwiftError())
      return false;

  // Set up the argument vectors.
  SmallVector<MVT, 16> OutVTs;
  OutVTs.reserve(CLI.OutVals.size());

  for (auto *Val : CLI.OutVals) {
    MVT VT;
    if (!isTypeLegal(Val->getType(), VT) &&
        !(VT == MVT::i1 || VT == MVT::i8 || VT == MVT::i16))
      return false;

    // We don't handle vector parameters yet.
    if (VT.isVector() || VT.getSizeInBits() > 64)
      return false;

    OutVTs.push_back(VT);
  }

  Address Addr;
  if (Callee && !computeCallAddress(Callee, Addr))
    return false;

  // The weak function target may be zero; in that case we must use indirect
  // addressing via a stub on windows as it may be out of range for a
  // PC-relative jump.
  if (Subtarget->isTargetWindows() && Addr.getGlobalValue() &&
      Addr.getGlobalValue()->hasExternalWeakLinkage())
    return false;

  // Handle the arguments now that we've gotten them.
  unsigned NumBytes;
  if (!processCallArgs(CLI, OutVTs, NumBytes))
    return false;

  const AArch64RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
  if (RegInfo->isAnyArgRegReserved(*MF))
    RegInfo->emitReservedArgRegCallError(*MF);

  // Issue the call.
  MachineInstrBuilder MIB;
  if (Subtarget->useSmallAddressing()) {
    const MCInstrDesc &II =
        TII.get(Addr.getReg() ? getBLRCallOpcode(*MF) : (unsigned)AArch64::BL);
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II);
    if (Symbol)
      MIB.addSym(Symbol, 0);
    else if (Addr.getGlobalValue())
      MIB.addGlobalAddress(Addr.getGlobalValue(), 0, 0);
    else if (Addr.getReg()) {
      Register Reg = constrainOperandRegClass(II, Addr.getReg(), 0);
      MIB.addReg(Reg);
    } else
      return false;
  } else {
    unsigned CallReg = 0;
    if (Symbol) {
      Register ADRPReg = createResultReg(&AArch64::GPR64commonRegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::ADRP),
              ADRPReg)
          .addSym(Symbol, AArch64II::MO_GOT | AArch64II::MO_PAGE);

      CallReg = createResultReg(&AArch64::GPR64RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
              TII.get(AArch64::LDRXui), CallReg)
          .addReg(ADRPReg)
          .addSym(Symbol,
                  AArch64II::MO_GOT | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
    } else if (Addr.getGlobalValue())
      CallReg = materializeGV(Addr.getGlobalValue());
    else if (Addr.getReg())
      CallReg = Addr.getReg();

    if (!CallReg)
      return false;

    const MCInstrDesc &II = TII.get(getBLRCallOpcode(*MF));
    CallReg = constrainOperandRegClass(II, CallReg, 0);
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II).addReg(CallReg);
  }

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

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

  CLI.Call = MIB;

  // Finish off the call including any return values.
  return finishCall(CLI, NumBytes);
}

bool AArch64FastISel::isMemCpySmall(uint64_t Len, MaybeAlign Alignment) {
  if (Alignment)
    return Len / Alignment->value() <= 4;
  else
    return Len < 32;
}

bool AArch64FastISel::tryEmitSmallMemCpy(Address Dest, Address Src,
                                         uint64_t Len, MaybeAlign Alignment) {
  // Make sure we don't bloat code by inlining very large memcpy's.
  if (!isMemCpySmall(Len, Alignment))
    return false;

  int64_t UnscaledOffset = 0;
  Address OrigDest = Dest;
  Address OrigSrc = Src;

  while (Len) {
    MVT VT;
    if (!Alignment || *Alignment >= 8) {
      if (Len >= 8)
        VT = MVT::i64;
      else if (Len >= 4)
        VT = MVT::i32;
      else if (Len >= 2)
        VT = MVT::i16;
      else {
        VT = MVT::i8;
      }
    } else {
      assert(Alignment && "Alignment is set in this branch");
      // Bound based on alignment.
      if (Len >= 4 && *Alignment == 4)
        VT = MVT::i32;
      else if (Len >= 2 && *Alignment == 2)
        VT = MVT::i16;
      else {
        VT = MVT::i8;
      }
    }

    unsigned ResultReg = emitLoad(VT, VT, Src);
    if (!ResultReg)
      return false;

    if (!emitStore(VT, ResultReg, Dest))
      return false;

    int64_t Size = VT.getSizeInBits() / 8;
    Len -= Size;
    UnscaledOffset += Size;

    // We need to recompute the unscaled offset for each iteration.
    Dest.setOffset(OrigDest.getOffset() + UnscaledOffset);
    Src.setOffset(OrigSrc.getOffset() + UnscaledOffset);
  }

  return true;
}

/// 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 AArch64FastISel::foldXALUIntrinsic(AArch64CC::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;

  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);

  // Simplify multiplies.
  Intrinsic::ID IID = II->getIntrinsicID();
  switch (IID) {
  default:
    break;
  case Intrinsic::smul_with_overflow:
    if (const auto *C = dyn_cast<ConstantInt>(RHS))
      if (C->getValue() == 2)
        IID = Intrinsic::sadd_with_overflow;
    break;
  case Intrinsic::umul_with_overflow:
    if (const auto *C = dyn_cast<ConstantInt>(RHS))
      if (C->getValue() == 2)
        IID = Intrinsic::uadd_with_overflow;
    break;
  }

  AArch64CC::CondCode TmpCC;
  switch (IID) {
  default:
    return false;
  case Intrinsic::sadd_with_overflow:
  case Intrinsic::ssub_with_overflow:
    TmpCC = AArch64CC::VS;
    break;
  case Intrinsic::uadd_with_overflow:
    TmpCC = AArch64CC::HS;
    break;
  case Intrinsic::usub_with_overflow:
    TmpCC = AArch64CC::LO;
    break;
  case Intrinsic::smul_with_overflow:
  case Intrinsic::umul_with_overflow:
    TmpCC = AArch64CC::NE;
    break;
  }

  // Check if both instructions are in the same basic block.
  if (!isValueAvailable(II))
    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 AArch64FastISel::fastLowerIntrinsicCall(const IntrinsicInst *II) {
  // FIXME: Handle more intrinsics.
  switch (II->getIntrinsicID()) {
  default: return false;
  case Intrinsic::frameaddress: {
    MachineFrameInfo &MFI = FuncInfo.MF->getFrameInfo();
    MFI.setFrameAddressIsTaken(true);

    const AArch64RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
    Register FramePtr = RegInfo->getFrameRegister(*(FuncInfo.MF));
    Register SrcReg = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(TargetOpcode::COPY), SrcReg).addReg(FramePtr);
    // Recursively load frame address
    // ldr x0, [fp]
    // ldr x0, [x0]
    // ldr x0, [x0]
    // ...
    unsigned DestReg;
    unsigned Depth = cast<ConstantInt>(II->getOperand(0))->getZExtValue();
    while (Depth--) {
      DestReg = fastEmitInst_ri(AArch64::LDRXui, &AArch64::GPR64RegClass,
                                SrcReg, 0);
      assert(DestReg && "Unexpected LDR instruction emission failure.");
      SrcReg = DestReg;
    }

    updateValueMap(II, SrcReg);
    return true;
  }
  case Intrinsic::sponentry: {
    MachineFrameInfo &MFI = FuncInfo.MF->getFrameInfo();

    // SP = FP + Fixed Object + 16
    int FI = MFI.CreateFixedObject(4, 0, false);
    Register ResultReg = createResultReg(&AArch64::GPR64spRegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(AArch64::ADDXri), ResultReg)
            .addFrameIndex(FI)
            .addImm(0)
            .addImm(0);

    updateValueMap(II, ResultReg);
    return true;
  }
  case Intrinsic::memcpy:
  case Intrinsic::memmove: {
    const auto *MTI = cast<MemTransferInst>(II);
    // Don't handle volatile.
    if (MTI->isVolatile())
      return false;

    // Disable inlining for memmove before calls to ComputeAddress.  Otherwise,
    // we would emit dead code because we don't currently handle memmoves.
    bool IsMemCpy = (II->getIntrinsicID() == Intrinsic::memcpy);
    if (isa<ConstantInt>(MTI->getLength()) && IsMemCpy) {
      // Small memcpy's are common enough that we want to do them without a call
      // if possible.
      uint64_t Len = cast<ConstantInt>(MTI->getLength())->getZExtValue();
      MaybeAlign Alignment;
      if (MTI->getDestAlign() || MTI->getSourceAlign())
        Alignment = std::min(MTI->getDestAlign().valueOrOne(),
                             MTI->getSourceAlign().valueOrOne());
      if (isMemCpySmall(Len, Alignment)) {
        Address Dest, Src;
        if (!computeAddress(MTI->getRawDest(), Dest) ||
            !computeAddress(MTI->getRawSource(), Src))
          return false;
        if (tryEmitSmallMemCpy(Dest, Src, Len, Alignment))
          return true;
      }
    }

    if (!MTI->getLength()->getType()->isIntegerTy(64))
      return false;

    if (MTI->getSourceAddressSpace() > 255 || MTI->getDestAddressSpace() > 255)
      // Fast instruction selection doesn't support the special
      // address spaces.
      return false;

    const char *IntrMemName = isa<MemCpyInst>(II) ? "memcpy" : "memmove";
    return lowerCallTo(II, IntrMemName, II->arg_size() - 1);
  }
  case Intrinsic::memset: {
    const MemSetInst *MSI = cast<MemSetInst>(II);
    // Don't handle volatile.
    if (MSI->isVolatile())
      return false;

    if (!MSI->getLength()->getType()->isIntegerTy(64))
      return false;

    if (MSI->getDestAddressSpace() > 255)
      // Fast instruction selection doesn't support the special
      // address spaces.
      return false;

    return lowerCallTo(II, "memset", II->arg_size() - 1);
  }
  case Intrinsic::sin:
  case Intrinsic::cos:
  case Intrinsic::pow: {
    MVT RetVT;
    if (!isTypeLegal(II->getType(), RetVT))
      return false;

    if (RetVT != MVT::f32 && RetVT != MVT::f64)
      return false;

    static const RTLIB::Libcall LibCallTable[3][2] = {
      { RTLIB::SIN_F32, RTLIB::SIN_F64 },
      { RTLIB::COS_F32, RTLIB::COS_F64 },
      { RTLIB::POW_F32, RTLIB::POW_F64 }
    };
    RTLIB::Libcall LC;
    bool Is64Bit = RetVT == MVT::f64;
    switch (II->getIntrinsicID()) {
    default:
      llvm_unreachable("Unexpected intrinsic.");
    case Intrinsic::sin:
      LC = LibCallTable[0][Is64Bit];
      break;
    case Intrinsic::cos:
      LC = LibCallTable[1][Is64Bit];
      break;
    case Intrinsic::pow:
      LC = LibCallTable[2][Is64Bit];
      break;
    }

    ArgListTy Args;
    Args.reserve(II->arg_size());

    // Populate the argument list.
    for (auto &Arg : II->args()) {
      ArgListEntry Entry;
      Entry.Val = Arg;
      Entry.Ty = Arg->getType();
      Args.push_back(Entry);
    }

    CallLoweringInfo CLI;
    MCContext &Ctx = MF->getContext();
    CLI.setCallee(DL, Ctx, TLI.getLibcallCallingConv(LC), II->getType(),
                  TLI.getLibcallName(LC), std::move(Args));
    if (!lowerCallTo(CLI))
      return false;
    updateValueMap(II, CLI.ResultReg);
    return true;
  }
  case Intrinsic::fabs: {
    MVT VT;
    if (!isTypeLegal(II->getType(), VT))
      return false;

    unsigned Opc;
    switch (VT.SimpleTy) {
    default:
      return false;
    case MVT::f32:
      Opc = AArch64::FABSSr;
      break;
    case MVT::f64:
      Opc = AArch64::FABSDr;
      break;
    }
    Register SrcReg = getRegForValue(II->getOperand(0));
    if (!SrcReg)
      return false;
    Register ResultReg = createResultReg(TLI.getRegClassFor(VT));
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(Opc), ResultReg)
      .addReg(SrcReg);
    updateValueMap(II, ResultReg);
    return true;
  }
  case Intrinsic::trap:
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::BRK))
        .addImm(1);
    return true;
  case Intrinsic::debugtrap:
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::BRK))
        .addImm(0xF000);
    return true;

  case Intrinsic::sqrt: {
    Type *RetTy = II->getCalledFunction()->getReturnType();

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

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

    unsigned ResultReg = fastEmit_r(VT, VT, ISD::FSQRT, Op0Reg);
    if (!ResultReg)
      return false;

    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.
    const Function *Callee = II->getCalledFunction();
    auto *Ty = cast<StructType>(Callee->getReturnType());
    Type *RetTy = Ty->getTypeAtIndex(0U);

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

    if (VT != MVT::i32 && 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);

    // Simplify multiplies.
    Intrinsic::ID IID = II->getIntrinsicID();
    switch (IID) {
    default:
      break;
    case Intrinsic::smul_with_overflow:
      if (const auto *C = dyn_cast<ConstantInt>(RHS))
        if (C->getValue() == 2) {
          IID = Intrinsic::sadd_with_overflow;
          RHS = LHS;
        }
      break;
    case Intrinsic::umul_with_overflow:
      if (const auto *C = dyn_cast<ConstantInt>(RHS))
        if (C->getValue() == 2) {
          IID = Intrinsic::uadd_with_overflow;
          RHS = LHS;
        }
      break;
    }

    unsigned ResultReg1 = 0, ResultReg2 = 0, MulReg = 0;
    AArch64CC::CondCode CC = AArch64CC::Invalid;
    switch (IID) {
    default: llvm_unreachable("Unexpected intrinsic!");
    case Intrinsic::sadd_with_overflow:
      ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
      CC = AArch64CC::VS;
      break;
    case Intrinsic::uadd_with_overflow:
      ResultReg1 = emitAdd(VT, LHS, RHS, /*SetFlags=*/true);
      CC = AArch64CC::HS;
      break;
    case Intrinsic::ssub_with_overflow:
      ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
      CC = AArch64CC::VS;
      break;
    case Intrinsic::usub_with_overflow:
      ResultReg1 = emitSub(VT, LHS, RHS, /*SetFlags=*/true);
      CC = AArch64CC::LO;
      break;
    case Intrinsic::smul_with_overflow: {
      CC = AArch64CC::NE;
      Register LHSReg = getRegForValue(LHS);
      if (!LHSReg)
        return false;

      Register RHSReg = getRegForValue(RHS);
      if (!RHSReg)
        return false;

      if (VT == MVT::i32) {
        MulReg = emitSMULL_rr(MVT::i64, LHSReg, RHSReg);
        Register MulSubReg =
            fastEmitInst_extractsubreg(VT, MulReg, AArch64::sub_32);
        // cmp xreg, wreg, sxtw
        emitAddSub_rx(/*UseAdd=*/false, MVT::i64, MulReg, MulSubReg,
                      AArch64_AM::SXTW, /*ShiftImm=*/0, /*SetFlags=*/true,
                      /*WantResult=*/false);
        MulReg = MulSubReg;
      } else {
        assert(VT == MVT::i64 && "Unexpected value type.");
        // LHSReg and RHSReg cannot be killed by this Mul, since they are
        // reused in the next instruction.
        MulReg = emitMul_rr(VT, LHSReg, RHSReg);
        unsigned SMULHReg = fastEmit_rr(VT, VT, ISD::MULHS, LHSReg, RHSReg);
        emitSubs_rs(VT, SMULHReg, MulReg, AArch64_AM::ASR, 63,
                    /*WantResult=*/false);
      }
      break;
    }
    case Intrinsic::umul_with_overflow: {
      CC = AArch64CC::NE;
      Register LHSReg = getRegForValue(LHS);
      if (!LHSReg)
        return false;

      Register RHSReg = getRegForValue(RHS);
      if (!RHSReg)
        return false;

      if (VT == MVT::i32) {
        MulReg = emitUMULL_rr(MVT::i64, LHSReg, RHSReg);
        // tst xreg, #0xffffffff00000000
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
                TII.get(AArch64::ANDSXri), AArch64::XZR)
            .addReg(MulReg)
            .addImm(AArch64_AM::encodeLogicalImmediate(0xFFFFFFFF00000000, 64));
        MulReg = fastEmitInst_extractsubreg(VT, MulReg, AArch64::sub_32);
      } else {
        assert(VT == MVT::i64 && "Unexpected value type.");
        // LHSReg and RHSReg cannot be killed by this Mul, since they are
        // reused in the next instruction.
        MulReg = emitMul_rr(VT, LHSReg, RHSReg);
        unsigned UMULHReg = fastEmit_rr(VT, VT, ISD::MULHU, LHSReg, RHSReg);
        emitSubs_rr(VT, AArch64::XZR, UMULHReg, /*WantResult=*/false);
      }
      break;
    }
    }

    if (MulReg) {
      ResultReg1 = createResultReg(TLI.getRegClassFor(VT));
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
              TII.get(TargetOpcode::COPY), ResultReg1).addReg(MulReg);
    }

    if (!ResultReg1)
      return false;

    ResultReg2 = fastEmitInst_rri(AArch64::CSINCWr, &AArch64::GPR32RegClass,
                                  AArch64::WZR, AArch64::WZR,
                                  getInvertedCondCode(CC));
    (void)ResultReg2;
    assert((ResultReg1 + 1) == ResultReg2 &&
           "Nonconsecutive result registers.");
    updateValueMap(II, ResultReg1, 2);
    return true;
  }
  case Intrinsic::aarch64_crc32b:
  case Intrinsic::aarch64_crc32h:
  case Intrinsic::aarch64_crc32w:
  case Intrinsic::aarch64_crc32x:
  case Intrinsic::aarch64_crc32cb:
  case Intrinsic::aarch64_crc32ch:
  case Intrinsic::aarch64_crc32cw:
  case Intrinsic::aarch64_crc32cx: {
    if (!Subtarget->hasCRC())
      return false;

    unsigned Opc;
    switch (II->getIntrinsicID()) {
    default:
      llvm_unreachable("Unexpected intrinsic!");
    case Intrinsic::aarch64_crc32b:
      Opc = AArch64::CRC32Brr;
      break;
    case Intrinsic::aarch64_crc32h:
      Opc = AArch64::CRC32Hrr;
      break;
    case Intrinsic::aarch64_crc32w:
      Opc = AArch64::CRC32Wrr;
      break;
    case Intrinsic::aarch64_crc32x:
      Opc = AArch64::CRC32Xrr;
      break;
    case Intrinsic::aarch64_crc32cb:
      Opc = AArch64::CRC32CBrr;
      break;
    case Intrinsic::aarch64_crc32ch:
      Opc = AArch64::CRC32CHrr;
      break;
    case Intrinsic::aarch64_crc32cw:
      Opc = AArch64::CRC32CWrr;
      break;
    case Intrinsic::aarch64_crc32cx:
      Opc = AArch64::CRC32CXrr;
      break;
    }

    Register LHSReg = getRegForValue(II->getArgOperand(0));
    Register RHSReg = getRegForValue(II->getArgOperand(1));
    if (!LHSReg || !RHSReg)
      return false;

    Register ResultReg =
        fastEmitInst_rr(Opc, &AArch64::GPR32RegClass, LHSReg, RHSReg);
    updateValueMap(II, ResultReg);
    return true;
  }
  }
  return false;
}

bool AArch64FastISel::selectRet(const Instruction *I) {
  const ReturnInst *Ret = cast<ReturnInst>(I);
  const Function &F = *I->getParent()->getParent();

  if (!FuncInfo.CanLowerReturn)
    return false;

  if (F.isVarArg())
    return false;

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

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

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

  if (Ret->getNumOperands() > 0) {
    CallingConv::ID CC = F.getCallingConv();
    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_AArch64_AAPCS);

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

    CCValAssign &VA = ValLocs[0];
    const Value *RV = Ret->getOperand(0);

    // Don't bother handling odd stuff for now.
    if ((VA.getLocInfo() != CCValAssign::Full) &&
        (VA.getLocInfo() != CCValAssign::BCvt))
      return false;

    // Only handle register returns for now.
    if (!VA.isRegLoc())
      return false;

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

    unsigned SrcReg = Reg + VA.getValNo();
    Register DestReg = VA.getLocReg();
    // Avoid a cross-class copy. This is very unlikely.
    if (!MRI.getRegClass(SrcReg)->contains(DestReg))
      return false;

    EVT RVEVT = TLI.getValueType(DL, RV->getType());
    if (!RVEVT.isSimple())
      return false;

    // Vectors (of > 1 lane) in big endian need tricky handling.
    if (RVEVT.isVector() && RVEVT.getVectorElementCount().isVector() &&
        !Subtarget->isLittleEndian())
      return false;

    MVT RVVT = RVEVT.getSimpleVT();
    if (RVVT == MVT::f128)
      return false;

    MVT DestVT = VA.getValVT();
    // Special handling for extended integers.
    if (RVVT != DestVT) {
      if (RVVT != MVT::i1 && RVVT != MVT::i8 && RVVT != MVT::i16)
        return false;

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

      bool IsZExt = Outs[0].Flags.isZExt();
      SrcReg = emitIntExt(RVVT, SrcReg, DestVT, IsZExt);
      if (SrcReg == 0)
        return false;
    }

    // "Callee" (i.e. value producer) zero extends pointers at function
    // boundary.
    if (Subtarget->isTargetILP32() && RV->getType()->isPointerTy())
      SrcReg = emitAnd_ri(MVT::i64, SrcReg, 0xffffffff);

    // Make the copy.
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(TargetOpcode::COPY), DestReg).addReg(SrcReg);

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

  MachineInstrBuilder MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
                                    TII.get(AArch64::RET_ReallyLR));
  for (unsigned RetReg : RetRegs)
    MIB.addReg(RetReg, RegState::Implicit);
  return true;
}

bool AArch64FastISel::selectTrunc(const Instruction *I) {
  Type *DestTy = I->getType();
  Value *Op = I->getOperand(0);
  Type *SrcTy = Op->getType();

  EVT SrcEVT = TLI.getValueType(DL, SrcTy, true);
  EVT DestEVT = TLI.getValueType(DL, DestTy, true);
  if (!SrcEVT.isSimple())
    return false;
  if (!DestEVT.isSimple())
    return false;

  MVT SrcVT = SrcEVT.getSimpleVT();
  MVT DestVT = DestEVT.getSimpleVT();

  if (SrcVT != MVT::i64 && SrcVT != MVT::i32 && SrcVT != MVT::i16 &&
      SrcVT != MVT::i8)
    return false;
  if (DestVT != MVT::i32 && DestVT != MVT::i16 && DestVT != MVT::i8 &&
      DestVT != MVT::i1)
    return false;

  Register SrcReg = getRegForValue(Op);
  if (!SrcReg)
    return false;

  // If we're truncating from i64 to a smaller non-legal type then generate an
  // AND. Otherwise, we know the high bits are undefined and a truncate only
  // generate a COPY. We cannot mark the source register also as result
  // register, because this can incorrectly transfer the kill flag onto the
  // source register.
  unsigned ResultReg;
  if (SrcVT == MVT::i64) {
    uint64_t Mask = 0;
    switch (DestVT.SimpleTy) {
    default:
      // Trunc i64 to i32 is handled by the target-independent fast-isel.
      return false;
    case MVT::i1:
      Mask = 0x1;
      break;
    case MVT::i8:
      Mask = 0xff;
      break;
    case MVT::i16:
      Mask = 0xffff;
      break;
    }
    // Issue an extract_subreg to get the lower 32-bits.
    Register Reg32 = fastEmitInst_extractsubreg(MVT::i32, SrcReg,
                                                AArch64::sub_32);
    // Create the AND instruction which performs the actual truncation.
    ResultReg = emitAnd_ri(MVT::i32, Reg32, Mask);
    assert(ResultReg && "Unexpected AND instruction emission failure.");
  } else {
    ResultReg = createResultReg(&AArch64::GPR32RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(TargetOpcode::COPY), ResultReg)
        .addReg(SrcReg);
  }

  updateValueMap(I, ResultReg);
  return true;
}

unsigned AArch64FastISel::emiti1Ext(unsigned SrcReg, MVT DestVT, bool IsZExt) {
  assert((DestVT == MVT::i8 || DestVT == MVT::i16 || DestVT == MVT::i32 ||
          DestVT == MVT::i64) &&
         "Unexpected value type.");
  // Handle i8 and i16 as i32.
  if (DestVT == MVT::i8 || DestVT == MVT::i16)
    DestVT = MVT::i32;

  if (IsZExt) {
    unsigned ResultReg = emitAnd_ri(MVT::i32, SrcReg, 1);
    assert(ResultReg && "Unexpected AND instruction emission failure.");
    if (DestVT == MVT::i64) {
      // We're ZExt i1 to i64.  The ANDWri Wd, Ws, #1 implicitly clears the
      // upper 32 bits.  Emit a SUBREG_TO_REG to extend from Wd to Xd.
      Register Reg64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
              TII.get(AArch64::SUBREG_TO_REG), Reg64)
          .addImm(0)
          .addReg(ResultReg)
          .addImm(AArch64::sub_32);
      ResultReg = Reg64;
    }
    return ResultReg;
  } else {
    if (DestVT == MVT::i64) {
      // FIXME: We're SExt i1 to i64.
      return 0;
    }
    return fastEmitInst_rii(AArch64::SBFMWri, &AArch64::GPR32RegClass, SrcReg,
                            0, 0);
  }
}

unsigned AArch64FastISel::emitMul_rr(MVT RetVT, unsigned Op0, unsigned Op1) {
  unsigned Opc, ZReg;
  switch (RetVT.SimpleTy) {
  default: return 0;
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
    RetVT = MVT::i32;
    Opc = AArch64::MADDWrrr; ZReg = AArch64::WZR; break;
  case MVT::i64:
    Opc = AArch64::MADDXrrr; ZReg = AArch64::XZR; break;
  }

  const TargetRegisterClass *RC =
      (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  return fastEmitInst_rrr(Opc, RC, Op0, Op1, ZReg);
}

unsigned AArch64FastISel::emitSMULL_rr(MVT RetVT, unsigned Op0, unsigned Op1) {
  if (RetVT != MVT::i64)
    return 0;

  return fastEmitInst_rrr(AArch64::SMADDLrrr, &AArch64::GPR64RegClass,
                          Op0, Op1, AArch64::XZR);
}

unsigned AArch64FastISel::emitUMULL_rr(MVT RetVT, unsigned Op0, unsigned Op1) {
  if (RetVT != MVT::i64)
    return 0;

  return fastEmitInst_rrr(AArch64::UMADDLrrr, &AArch64::GPR64RegClass,
                          Op0, Op1, AArch64::XZR);
}

unsigned AArch64FastISel::emitLSL_rr(MVT RetVT, unsigned Op0Reg,
                                     unsigned Op1Reg) {
  unsigned Opc = 0;
  bool NeedTrunc = false;
  uint64_t Mask = 0;
  switch (RetVT.SimpleTy) {
  default: return 0;
  case MVT::i8:  Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xff;   break;
  case MVT::i16: Opc = AArch64::LSLVWr; NeedTrunc = true; Mask = 0xffff; break;
  case MVT::i32: Opc = AArch64::LSLVWr;                                  break;
  case MVT::i64: Opc = AArch64::LSLVXr;                                  break;
  }

  const TargetRegisterClass *RC =
      (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  if (NeedTrunc)
    Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Mask);

  Register ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op1Reg);
  if (NeedTrunc)
    ResultReg = emitAnd_ri(MVT::i32, ResultReg, Mask);
  return ResultReg;
}

unsigned AArch64FastISel::emitLSL_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
                                     uint64_t Shift, bool IsZExt) {
  assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
         "Unexpected source/return type pair.");
  assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
          SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
         "Unexpected source value type.");
  assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
          RetVT == MVT::i64) && "Unexpected return value type.");

  bool Is64Bit = (RetVT == MVT::i64);
  unsigned RegSize = Is64Bit ? 64 : 32;
  unsigned DstBits = RetVT.getSizeInBits();
  unsigned SrcBits = SrcVT.getSizeInBits();
  const TargetRegisterClass *RC =
      Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;

  // Just emit a copy for "zero" shifts.
  if (Shift == 0) {
    if (RetVT == SrcVT) {
      Register ResultReg = createResultReg(RC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
              TII.get(TargetOpcode::COPY), ResultReg)
          .addReg(Op0);
      return ResultReg;
    } else
      return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
  }

  // Don't deal with undefined shifts.
  if (Shift >= DstBits)
    return 0;

  // For immediate shifts we can fold the zero-/sign-extension into the shift.
  // {S|U}BFM Wd, Wn, #r, #s
  // Wd<32+s-r,32-r> = Wn<s:0> when r > s

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = shl i16 %1, 4
  // Wd<32+7-28,32-28> = Wn<7:0> <- clamp s to 7
  // 0b1111_1111_1111_1111__1111_1010_1010_0000 sext
  // 0b0000_0000_0000_0000__0000_0101_0101_0000 sext | zext
  // 0b0000_0000_0000_0000__0000_1010_1010_0000 zext

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = shl i16 %1, 8
  // Wd<32+7-24,32-24> = Wn<7:0>
  // 0b1111_1111_1111_1111__1010_1010_0000_0000 sext
  // 0b0000_0000_0000_0000__0101_0101_0000_0000 sext | zext
  // 0b0000_0000_0000_0000__1010_1010_0000_0000 zext

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = shl i16 %1, 12
  // Wd<32+3-20,32-20> = Wn<3:0>
  // 0b1111_1111_1111_1111__1010_0000_0000_0000 sext
  // 0b0000_0000_0000_0000__0101_0000_0000_0000 sext | zext
  // 0b0000_0000_0000_0000__1010_0000_0000_0000 zext

  unsigned ImmR = RegSize - Shift;
  // Limit the width to the length of the source type.
  unsigned ImmS = std::min<unsigned>(SrcBits - 1, DstBits - 1 - Shift);
  static const unsigned OpcTable[2][2] = {
    {AArch64::SBFMWri, AArch64::SBFMXri},
    {AArch64::UBFMWri, AArch64::UBFMXri}
  };
  unsigned Opc = OpcTable[IsZExt][Is64Bit];
  if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
    Register TmpReg = MRI.createVirtualRegister(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(AArch64::SUBREG_TO_REG), TmpReg)
        .addImm(0)
        .addReg(Op0)
        .addImm(AArch64::sub_32);
    Op0 = TmpReg;
  }
  return fastEmitInst_rii(Opc, RC, Op0, ImmR, ImmS);
}

unsigned AArch64FastISel::emitLSR_rr(MVT RetVT, unsigned Op0Reg,
                                     unsigned Op1Reg) {
  unsigned Opc = 0;
  bool NeedTrunc = false;
  uint64_t Mask = 0;
  switch (RetVT.SimpleTy) {
  default: return 0;
  case MVT::i8:  Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xff;   break;
  case MVT::i16: Opc = AArch64::LSRVWr; NeedTrunc = true; Mask = 0xffff; break;
  case MVT::i32: Opc = AArch64::LSRVWr; break;
  case MVT::i64: Opc = AArch64::LSRVXr; break;
  }

  const TargetRegisterClass *RC =
      (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  if (NeedTrunc) {
    Op0Reg = emitAnd_ri(MVT::i32, Op0Reg, Mask);
    Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Mask);
  }
  Register ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op1Reg);
  if (NeedTrunc)
    ResultReg = emitAnd_ri(MVT::i32, ResultReg, Mask);
  return ResultReg;
}

unsigned AArch64FastISel::emitLSR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
                                     uint64_t Shift, bool IsZExt) {
  assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
         "Unexpected source/return type pair.");
  assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
          SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
         "Unexpected source value type.");
  assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
          RetVT == MVT::i64) && "Unexpected return value type.");

  bool Is64Bit = (RetVT == MVT::i64);
  unsigned RegSize = Is64Bit ? 64 : 32;
  unsigned DstBits = RetVT.getSizeInBits();
  unsigned SrcBits = SrcVT.getSizeInBits();
  const TargetRegisterClass *RC =
      Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;

  // Just emit a copy for "zero" shifts.
  if (Shift == 0) {
    if (RetVT == SrcVT) {
      Register ResultReg = createResultReg(RC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
              TII.get(TargetOpcode::COPY), ResultReg)
      .addReg(Op0);
      return ResultReg;
    } else
      return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
  }

  // Don't deal with undefined shifts.
  if (Shift >= DstBits)
    return 0;

  // For immediate shifts we can fold the zero-/sign-extension into the shift.
  // {S|U}BFM Wd, Wn, #r, #s
  // Wd<s-r:0> = Wn<s:r> when r <= s

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = lshr i16 %1, 4
  // Wd<7-4:0> = Wn<7:4>
  // 0b0000_0000_0000_0000__0000_1111_1111_1010 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
  // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = lshr i16 %1, 8
  // Wd<7-7,0> = Wn<7:7>
  // 0b0000_0000_0000_0000__0000_0000_1111_1111 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = lshr i16 %1, 12
  // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
  // 0b0000_0000_0000_0000__0000_0000_0000_1111 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext

  if (Shift >= SrcBits && IsZExt)
    return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);

  // It is not possible to fold a sign-extend into the LShr instruction. In this
  // case emit a sign-extend.
  if (!IsZExt) {
    Op0 = emitIntExt(SrcVT, Op0, RetVT, IsZExt);
    if (!Op0)
      return 0;
    SrcVT = RetVT;
    SrcBits = SrcVT.getSizeInBits();
    IsZExt = true;
  }

  unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
  unsigned ImmS = SrcBits - 1;
  static const unsigned OpcTable[2][2] = {
    {AArch64::SBFMWri, AArch64::SBFMXri},
    {AArch64::UBFMWri, AArch64::UBFMXri}
  };
  unsigned Opc = OpcTable[IsZExt][Is64Bit];
  if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
    Register TmpReg = MRI.createVirtualRegister(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(AArch64::SUBREG_TO_REG), TmpReg)
        .addImm(0)
        .addReg(Op0)
        .addImm(AArch64::sub_32);
    Op0 = TmpReg;
  }
  return fastEmitInst_rii(Opc, RC, Op0, ImmR, ImmS);
}

unsigned AArch64FastISel::emitASR_rr(MVT RetVT, unsigned Op0Reg,
                                     unsigned Op1Reg) {
  unsigned Opc = 0;
  bool NeedTrunc = false;
  uint64_t Mask = 0;
  switch (RetVT.SimpleTy) {
  default: return 0;
  case MVT::i8:  Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xff;   break;
  case MVT::i16: Opc = AArch64::ASRVWr; NeedTrunc = true; Mask = 0xffff; break;
  case MVT::i32: Opc = AArch64::ASRVWr;                                  break;
  case MVT::i64: Opc = AArch64::ASRVXr;                                  break;
  }

  const TargetRegisterClass *RC =
      (RetVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  if (NeedTrunc) {
    Op0Reg = emitIntExt(RetVT, Op0Reg, MVT::i32, /*isZExt=*/false);
    Op1Reg = emitAnd_ri(MVT::i32, Op1Reg, Mask);
  }
  Register ResultReg = fastEmitInst_rr(Opc, RC, Op0Reg, Op1Reg);
  if (NeedTrunc)
    ResultReg = emitAnd_ri(MVT::i32, ResultReg, Mask);
  return ResultReg;
}

unsigned AArch64FastISel::emitASR_ri(MVT RetVT, MVT SrcVT, unsigned Op0,
                                     uint64_t Shift, bool IsZExt) {
  assert(RetVT.SimpleTy >= SrcVT.SimpleTy &&
         "Unexpected source/return type pair.");
  assert((SrcVT == MVT::i1 || SrcVT == MVT::i8 || SrcVT == MVT::i16 ||
          SrcVT == MVT::i32 || SrcVT == MVT::i64) &&
         "Unexpected source value type.");
  assert((RetVT == MVT::i8 || RetVT == MVT::i16 || RetVT == MVT::i32 ||
          RetVT == MVT::i64) && "Unexpected return value type.");

  bool Is64Bit = (RetVT == MVT::i64);
  unsigned RegSize = Is64Bit ? 64 : 32;
  unsigned DstBits = RetVT.getSizeInBits();
  unsigned SrcBits = SrcVT.getSizeInBits();
  const TargetRegisterClass *RC =
      Is64Bit ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;

  // Just emit a copy for "zero" shifts.
  if (Shift == 0) {
    if (RetVT == SrcVT) {
      Register ResultReg = createResultReg(RC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
              TII.get(TargetOpcode::COPY), ResultReg)
      .addReg(Op0);
      return ResultReg;
    } else
      return emitIntExt(SrcVT, Op0, RetVT, IsZExt);
  }

  // Don't deal with undefined shifts.
  if (Shift >= DstBits)
    return 0;

  // For immediate shifts we can fold the zero-/sign-extension into the shift.
  // {S|U}BFM Wd, Wn, #r, #s
  // Wd<s-r:0> = Wn<s:r> when r <= s

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = ashr i16 %1, 4
  // Wd<7-4:0> = Wn<7:4>
  // 0b1111_1111_1111_1111__1111_1111_1111_1010 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0101 sext | zext
  // 0b0000_0000_0000_0000__0000_0000_0000_1010 zext

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = ashr i16 %1, 8
  // Wd<7-7,0> = Wn<7:7>
  // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext

  // %1 = {s|z}ext i8 {0b1010_1010|0b0101_0101} to i16
  // %2 = ashr i16 %1, 12
  // Wd<7-7,0> = Wn<7:7> <- clamp r to 7
  // 0b1111_1111_1111_1111__1111_1111_1111_1111 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 sext
  // 0b0000_0000_0000_0000__0000_0000_0000_0000 zext

  if (Shift >= SrcBits && IsZExt)
    return materializeInt(ConstantInt::get(*Context, APInt(RegSize, 0)), RetVT);

  unsigned ImmR = std::min<unsigned>(SrcBits - 1, Shift);
  unsigned ImmS = SrcBits - 1;
  static const unsigned OpcTable[2][2] = {
    {AArch64::SBFMWri, AArch64::SBFMXri},
    {AArch64::UBFMWri, AArch64::UBFMXri}
  };
  unsigned Opc = OpcTable[IsZExt][Is64Bit];
  if (SrcVT.SimpleTy <= MVT::i32 && RetVT == MVT::i64) {
    Register TmpReg = MRI.createVirtualRegister(RC);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(AArch64::SUBREG_TO_REG), TmpReg)
        .addImm(0)
        .addReg(Op0)
        .addImm(AArch64::sub_32);
    Op0 = TmpReg;
  }
  return fastEmitInst_rii(Opc, RC, Op0, ImmR, ImmS);
}

unsigned AArch64FastISel::emitIntExt(MVT SrcVT, unsigned SrcReg, MVT DestVT,
                                     bool IsZExt) {
  assert(DestVT != MVT::i1 && "ZeroExt/SignExt an i1?");

  // FastISel does not have plumbing to deal with extensions where the SrcVT or
  // DestVT are odd things, so test to make sure that they are both types we can
  // handle (i1/i8/i16/i32 for SrcVT and i8/i16/i32/i64 for DestVT), otherwise
  // bail out to SelectionDAG.
  if (((DestVT != MVT::i8) && (DestVT != MVT::i16) &&
       (DestVT != MVT::i32) && (DestVT != MVT::i64)) ||
      ((SrcVT !=  MVT::i1) && (SrcVT !=  MVT::i8) &&
       (SrcVT !=  MVT::i16) && (SrcVT !=  MVT::i32)))
    return 0;

  unsigned Opc;
  unsigned Imm = 0;

  switch (SrcVT.SimpleTy) {
  default:
    return 0;
  case MVT::i1:
    return emiti1Ext(SrcReg, DestVT, IsZExt);
  case MVT::i8:
    if (DestVT == MVT::i64)
      Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
    else
      Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
    Imm = 7;
    break;
  case MVT::i16:
    if (DestVT == MVT::i64)
      Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
    else
      Opc = IsZExt ? AArch64::UBFMWri : AArch64::SBFMWri;
    Imm = 15;
    break;
  case MVT::i32:
    assert(DestVT == MVT::i64 && "IntExt i32 to i32?!?");
    Opc = IsZExt ? AArch64::UBFMXri : AArch64::SBFMXri;
    Imm = 31;
    break;
  }

  // Handle i8 and i16 as i32.
  if (DestVT == MVT::i8 || DestVT == MVT::i16)
    DestVT = MVT::i32;
  else if (DestVT == MVT::i64) {
    Register Src64 = MRI.createVirtualRegister(&AArch64::GPR64RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(AArch64::SUBREG_TO_REG), Src64)
        .addImm(0)
        .addReg(SrcReg)
        .addImm(AArch64::sub_32);
    SrcReg = Src64;
  }

  const TargetRegisterClass *RC =
      (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  return fastEmitInst_rii(Opc, RC, SrcReg, 0, Imm);
}

static bool isZExtLoad(const MachineInstr *LI) {
  switch (LI->getOpcode()) {
  default:
    return false;
  case AArch64::LDURBBi:
  case AArch64::LDURHHi:
  case AArch64::LDURWi:
  case AArch64::LDRBBui:
  case AArch64::LDRHHui:
  case AArch64::LDRWui:
  case AArch64::LDRBBroX:
  case AArch64::LDRHHroX:
  case AArch64::LDRWroX:
  case AArch64::LDRBBroW:
  case AArch64::LDRHHroW:
  case AArch64::LDRWroW:
    return true;
  }
}

static bool isSExtLoad(const MachineInstr *LI) {
  switch (LI->getOpcode()) {
  default:
    return false;
  case AArch64::LDURSBWi:
  case AArch64::LDURSHWi:
  case AArch64::LDURSBXi:
  case AArch64::LDURSHXi:
  case AArch64::LDURSWi:
  case AArch64::LDRSBWui:
  case AArch64::LDRSHWui:
  case AArch64::LDRSBXui:
  case AArch64::LDRSHXui:
  case AArch64::LDRSWui:
  case AArch64::LDRSBWroX:
  case AArch64::LDRSHWroX:
  case AArch64::LDRSBXroX:
  case AArch64::LDRSHXroX:
  case AArch64::LDRSWroX:
  case AArch64::LDRSBWroW:
  case AArch64::LDRSHWroW:
  case AArch64::LDRSBXroW:
  case AArch64::LDRSHXroW:
  case AArch64::LDRSWroW:
    return true;
  }
}

bool AArch64FastISel::optimizeIntExtLoad(const Instruction *I, MVT RetVT,
                                         MVT SrcVT) {
  const auto *LI = dyn_cast<LoadInst>(I->getOperand(0));
  if (!LI || !LI->hasOneUse())
    return false;

  // Check if the load instruction has already been selected.
  Register Reg = lookUpRegForValue(LI);
  if (!Reg)
    return false;

  MachineInstr *MI = MRI.getUniqueVRegDef(Reg);
  if (!MI)
    return false;

  // Check if the correct load instruction has been emitted - SelectionDAG might
  // have emitted a zero-extending load, but we need a sign-extending load.
  bool IsZExt = isa<ZExtInst>(I);
  const auto *LoadMI = MI;
  if (LoadMI->getOpcode() == TargetOpcode::COPY &&
      LoadMI->getOperand(1).getSubReg() == AArch64::sub_32) {
    Register LoadReg = MI->getOperand(1).getReg();
    LoadMI = MRI.getUniqueVRegDef(LoadReg);
    assert(LoadMI && "Expected valid instruction");
  }
  if (!(IsZExt && isZExtLoad(LoadMI)) && !(!IsZExt && isSExtLoad(LoadMI)))
    return false;

  // Nothing to be done.
  if (RetVT != MVT::i64 || SrcVT > MVT::i32) {
    updateValueMap(I, Reg);
    return true;
  }

  if (IsZExt) {
    Register Reg64 = createResultReg(&AArch64::GPR64RegClass);
    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
            TII.get(AArch64::SUBREG_TO_REG), Reg64)
        .addImm(0)
        .addReg(Reg, getKillRegState(true))
        .addImm(AArch64::sub_32);
    Reg = Reg64;
  } else {
    assert((MI->getOpcode() == TargetOpcode::COPY &&
            MI->getOperand(1).getSubReg() == AArch64::sub_32) &&
           "Expected copy instruction");
    Reg = MI->getOperand(1).getReg();
    MachineBasicBlock::iterator I(MI);
    removeDeadCode(I, std::next(I));
  }
  updateValueMap(I, Reg);
  return true;
}

bool AArch64FastISel::selectIntExt(const Instruction *I) {
  assert((isa<ZExtInst>(I) || isa<SExtInst>(I)) &&
         "Unexpected integer extend instruction.");
  MVT RetVT;
  MVT SrcVT;
  if (!isTypeSupported(I->getType(), RetVT))
    return false;

  if (!isTypeSupported(I->getOperand(0)->getType(), SrcVT))
    return false;

  // Try to optimize already sign-/zero-extended values from load instructions.
  if (optimizeIntExtLoad(I, RetVT, SrcVT))
    return true;

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

  // Try to optimize already sign-/zero-extended values from function arguments.
  bool IsZExt = isa<ZExtInst>(I);
  if (const auto *Arg = dyn_cast<Argument>(I->getOperand(0))) {
    if ((IsZExt && Arg->hasZExtAttr()) || (!IsZExt && Arg->hasSExtAttr())) {
      if (RetVT == MVT::i64 && SrcVT != MVT::i64) {
        Register ResultReg = createResultReg(&AArch64::GPR64RegClass);
        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD,
                TII.get(AArch64::SUBREG_TO_REG), ResultReg)
            .addImm(0)
            .addReg(SrcReg)
            .addImm(AArch64::sub_32);
        SrcReg = ResultReg;
      }

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

  unsigned ResultReg = emitIntExt(SrcVT, SrcReg, RetVT, IsZExt);
  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectRem(const Instruction *I, unsigned ISDOpcode) {
  EVT DestEVT = TLI.getValueType(DL, I->getType(), true);
  if (!DestEVT.isSimple())
    return false;

  MVT DestVT = DestEVT.getSimpleVT();
  if (DestVT != MVT::i64 && DestVT != MVT::i32)
    return false;

  unsigned DivOpc;
  bool Is64bit = (DestVT == MVT::i64);
  switch (ISDOpcode) {
  default:
    return false;
  case ISD::SREM:
    DivOpc = Is64bit ? AArch64::SDIVXr : AArch64::SDIVWr;
    break;
  case ISD::UREM:
    DivOpc = Is64bit ? AArch64::UDIVXr : AArch64::UDIVWr;
    break;
  }
  unsigned MSubOpc = Is64bit ? AArch64::MSUBXrrr : AArch64::MSUBWrrr;
  Register Src0Reg = getRegForValue(I->getOperand(0));
  if (!Src0Reg)
    return false;

  Register Src1Reg = getRegForValue(I->getOperand(1));
  if (!Src1Reg)
    return false;

  const TargetRegisterClass *RC =
      (DestVT == MVT::i64) ? &AArch64::GPR64RegClass : &AArch64::GPR32RegClass;
  Register QuotReg = fastEmitInst_rr(DivOpc, RC, Src0Reg, Src1Reg);
  assert(QuotReg && "Unexpected DIV instruction emission failure.");
  // The remainder is computed as numerator - (quotient * denominator) using the
  // MSUB instruction.
  Register ResultReg = fastEmitInst_rrr(MSubOpc, RC, QuotReg, Src1Reg, Src0Reg);
  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectMul(const Instruction *I) {
  MVT VT;
  if (!isTypeSupported(I->getType(), VT, /*IsVectorAllowed=*/true))
    return false;

  if (VT.isVector())
    return selectBinaryOp(I, ISD::MUL);

  const Value *Src0 = I->getOperand(0);
  const Value *Src1 = I->getOperand(1);
  if (const auto *C = dyn_cast<ConstantInt>(Src0))
    if (C->getValue().isPowerOf2())
      std::swap(Src0, Src1);

  // Try to simplify to a shift instruction.
  if (const auto *C = dyn_cast<ConstantInt>(Src1))
    if (C->getValue().isPowerOf2()) {
      uint64_t ShiftVal = C->getValue().logBase2();
      MVT SrcVT = VT;
      bool IsZExt = true;
      if (const auto *ZExt = dyn_cast<ZExtInst>(Src0)) {
        if (!isIntExtFree(ZExt)) {
          MVT VT;
          if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), VT)) {
            SrcVT = VT;
            IsZExt = true;
            Src0 = ZExt->getOperand(0);
          }
        }
      } else if (const auto *SExt = dyn_cast<SExtInst>(Src0)) {
        if (!isIntExtFree(SExt)) {
          MVT VT;
          if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), VT)) {
            SrcVT = VT;
            IsZExt = false;
            Src0 = SExt->getOperand(0);
          }
        }
      }

      Register Src0Reg = getRegForValue(Src0);
      if (!Src0Reg)
        return false;

      unsigned ResultReg =
          emitLSL_ri(VT, SrcVT, Src0Reg, ShiftVal, IsZExt);

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

  Register Src0Reg = getRegForValue(I->getOperand(0));
  if (!Src0Reg)
    return false;

  Register Src1Reg = getRegForValue(I->getOperand(1));
  if (!Src1Reg)
    return false;

  unsigned ResultReg = emitMul_rr(VT, Src0Reg, Src1Reg);

  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectShift(const Instruction *I) {
  MVT RetVT;
  if (!isTypeSupported(I->getType(), RetVT, /*IsVectorAllowed=*/true))
    return false;

  if (RetVT.isVector())
    return selectOperator(I, I->getOpcode());

  if (const auto *C = dyn_cast<ConstantInt>(I->getOperand(1))) {
    unsigned ResultReg = 0;
    uint64_t ShiftVal = C->getZExtValue();
    MVT SrcVT = RetVT;
    bool IsZExt = I->getOpcode() != Instruction::AShr;
    const Value *Op0 = I->getOperand(0);
    if (const auto *ZExt = dyn_cast<ZExtInst>(Op0)) {
      if (!isIntExtFree(ZExt)) {
        MVT TmpVT;
        if (isValueAvailable(ZExt) && isTypeSupported(ZExt->getSrcTy(), TmpVT)) {
          SrcVT = TmpVT;
          IsZExt = true;
          Op0 = ZExt->getOperand(0);
        }
      }
    } else if (const auto *SExt = dyn_cast<SExtInst>(Op0)) {
      if (!isIntExtFree(SExt)) {
        MVT TmpVT;
        if (isValueAvailable(SExt) && isTypeSupported(SExt->getSrcTy(), TmpVT)) {
          SrcVT = TmpVT;
          IsZExt = false;
          Op0 = SExt->getOperand(0);
        }
      }
    }

    Register Op0Reg = getRegForValue(Op0);
    if (!Op0Reg)
      return false;

    switch (I->getOpcode()) {
    default: llvm_unreachable("Unexpected instruction.");
    case Instruction::Shl:
      ResultReg = emitLSL_ri(RetVT, SrcVT, Op0Reg, ShiftVal, IsZExt);
      break;
    case Instruction::AShr:
      ResultReg = emitASR_ri(RetVT, SrcVT, Op0Reg, ShiftVal, IsZExt);
      break;
    case Instruction::LShr:
      ResultReg = emitLSR_ri(RetVT, SrcVT, Op0Reg, ShiftVal, IsZExt);
      break;
    }
    if (!ResultReg)
      return false;

    updateValueMap(I, ResultReg);
    return true;
  }

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

  Register Op1Reg = getRegForValue(I->getOperand(1));
  if (!Op1Reg)
    return false;

  unsigned ResultReg = 0;
  switch (I->getOpcode()) {
  default: llvm_unreachable("Unexpected instruction.");
  case Instruction::Shl:
    ResultReg = emitLSL_rr(RetVT, Op0Reg, Op1Reg);
    break;
  case Instruction::AShr:
    ResultReg = emitASR_rr(RetVT, Op0Reg, Op1Reg);
    break;
  case Instruction::LShr:
    ResultReg = emitLSR_rr(RetVT, Op0Reg, Op1Reg);
    break;
  }

  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectBitCast(const Instruction *I) {
  MVT RetVT, SrcVT;

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

  unsigned Opc;
  if (RetVT == MVT::f32 && SrcVT == MVT::i32)
    Opc = AArch64::FMOVWSr;
  else if (RetVT == MVT::f64 && SrcVT == MVT::i64)
    Opc = AArch64::FMOVXDr;
  else if (RetVT == MVT::i32 && SrcVT == MVT::f32)
    Opc = AArch64::FMOVSWr;
  else if (RetVT == MVT::i64 && SrcVT == MVT::f64)
    Opc = AArch64::FMOVDXr;
  else
    return false;

  const TargetRegisterClass *RC = nullptr;
  switch (RetVT.SimpleTy) {
  default: llvm_unreachable("Unexpected value type.");
  case MVT::i32: RC = &AArch64::GPR32RegClass; break;
  case MVT::i64: RC = &AArch64::GPR64RegClass; break;
  case MVT::f32: RC = &AArch64::FPR32RegClass; break;
  case MVT::f64: RC = &AArch64::FPR64RegClass; break;
  }
  Register Op0Reg = getRegForValue(I->getOperand(0));
  if (!Op0Reg)
    return false;

  Register ResultReg = fastEmitInst_r(Opc, RC, Op0Reg);
  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

bool AArch64FastISel::selectFRem(const Instruction *I) {
  MVT RetVT;
  if (!isTypeLegal(I->getType(), RetVT))
    return false;

  RTLIB::Libcall LC;
  switch (RetVT.SimpleTy) {
  default:
    return false;
  case MVT::f32:
    LC = RTLIB::REM_F32;
    break;
  case MVT::f64:
    LC = RTLIB::REM_F64;
    break;
  }

  ArgListTy Args;
  Args.reserve(I->getNumOperands());

  // Populate the argument list.
  for (auto &Arg : I->operands()) {
    ArgListEntry Entry;
    Entry.Val = Arg;
    Entry.Ty = Arg->getType();
    Args.push_back(Entry);
  }

  CallLoweringInfo CLI;
  MCContext &Ctx = MF->getContext();
  CLI.setCallee(DL, Ctx, TLI.getLibcallCallingConv(LC), I->getType(),
                TLI.getLibcallName(LC), std::move(Args));
  if (!lowerCallTo(CLI))
    return false;
  updateValueMap(I, CLI.ResultReg);
  return true;
}

bool AArch64FastISel::selectSDiv(const Instruction *I) {
  MVT VT;
  if (!isTypeLegal(I->getType(), VT))
    return false;

  if (!isa<ConstantInt>(I->getOperand(1)))
    return selectBinaryOp(I, ISD::SDIV);

  const APInt &C = cast<ConstantInt>(I->getOperand(1))->getValue();
  if ((VT != MVT::i32 && VT != MVT::i64) || !C ||
      !(C.isPowerOf2() || C.isNegatedPowerOf2()))
    return selectBinaryOp(I, ISD::SDIV);

  unsigned Lg2 = C.countr_zero();
  Register Src0Reg = getRegForValue(I->getOperand(0));
  if (!Src0Reg)
    return false;

  if (cast<BinaryOperator>(I)->isExact()) {
    unsigned ResultReg = emitASR_ri(VT, VT, Src0Reg, Lg2);
    if (!ResultReg)
      return false;
    updateValueMap(I, ResultReg);
    return true;
  }

  int64_t Pow2MinusOne = (1ULL << Lg2) - 1;
  unsigned AddReg = emitAdd_ri_(VT, Src0Reg, Pow2MinusOne);
  if (!AddReg)
    return false;

  // (Src0 < 0) ? Pow2 - 1 : 0;
  if (!emitICmp_ri(VT, Src0Reg, 0))
    return false;

  unsigned SelectOpc;
  const TargetRegisterClass *RC;
  if (VT == MVT::i64) {
    SelectOpc = AArch64::CSELXr;
    RC = &AArch64::GPR64RegClass;
  } else {
    SelectOpc = AArch64::CSELWr;
    RC = &AArch64::GPR32RegClass;
  }
  Register SelectReg = fastEmitInst_rri(SelectOpc, RC, AddReg, Src0Reg,
                                        AArch64CC::LT);
  if (!SelectReg)
    return false;

  // Divide by Pow2 --> ashr. If we're dividing by a negative value we must also
  // negate the result.
  unsigned ZeroReg = (VT == MVT::i64) ? AArch64::XZR : AArch64::WZR;
  unsigned ResultReg;
  if (C.isNegative())
    ResultReg = emitAddSub_rs(/*UseAdd=*/false, VT, ZeroReg, SelectReg,
                              AArch64_AM::ASR, Lg2);
  else
    ResultReg = emitASR_ri(VT, VT, SelectReg, Lg2);

  if (!ResultReg)
    return false;

  updateValueMap(I, ResultReg);
  return true;
}

/// This is mostly a copy of the existing FastISel getRegForGEPIndex code. We
/// have to duplicate it for AArch64, because otherwise we would fail during the
/// sign-extend emission.
unsigned AArch64FastISel::getRegForGEPIndex(const Value *Idx) {
  Register IdxN = getRegForValue(Idx);
  if (IdxN == 0)
    // Unhandled operand. Halt "fast" selection and bail.
    return 0;

  // If the index is smaller or larger than intptr_t, truncate or extend it.
  MVT PtrVT = TLI.getPointerTy(DL);
  EVT IdxVT = EVT::getEVT(Idx->getType(), /*HandleUnknown=*/false);
  if (IdxVT.bitsLT(PtrVT)) {
    IdxN = emitIntExt(IdxVT.getSimpleVT(), IdxN, PtrVT, /*isZExt=*/false);
  } else if (IdxVT.bitsGT(PtrVT))
    llvm_unreachable("AArch64 FastISel doesn't support types larger than i64");
  return IdxN;
}

/// This is mostly a copy of the existing FastISel GEP code, but we have to
/// duplicate it for AArch64, because otherwise we would bail out even for
/// simple cases. This is because the standard fastEmit functions don't cover
/// MUL at all and ADD is lowered very inefficientily.
bool AArch64FastISel::selectGetElementPtr(const Instruction *I) {
  if (Subtarget->isTargetILP32())
    return false;

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

  // Keep a running tab of the total offset to coalesce multiple N = N + Offset
  // into a single N = N + TotalOffset.
  uint64_t TotalOffs = 0;
  MVT VT = TLI.getPointerTy(DL);
  for (gep_type_iterator GTI = gep_type_begin(I), E = gep_type_end(I);
       GTI != E; ++GTI) {
    const Value *Idx = GTI.getOperand();
    if (auto *StTy = GTI.getStructTypeOrNull()) {
      unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
      // N = N + Offset
      if (Field)
        TotalOffs += DL.getStructLayout(StTy)->getElementOffset(Field);
    } else {
      // If this is a constant subscript, handle it quickly.
      if (const auto *CI = dyn_cast<ConstantInt>(Idx)) {
        if (CI->isZero())
          continue;
        // N = N + Offset
        TotalOffs += GTI.getSequentialElementStride(DL) *
                     cast<ConstantInt>(CI)->getSExtValue();
        continue;
      }
      if (TotalOffs) {
        N = emitAdd_ri_(VT, N, TotalOffs);
        if (!N)
          return false;
        TotalOffs = 0;
      }

      // N = N + Idx * ElementSize;
      uint64_t ElementSize = GTI.getSequentialElementStride(DL);
      unsigned IdxN = getRegForGEPIndex(Idx);
      if (!IdxN)
        return false;

      if (ElementSize != 1) {
        unsigned C = fastEmit_i(VT, VT, ISD::Constant, ElementSize);
        if (!C)
          return false;
        IdxN = emitMul_rr(VT, IdxN, C);
        if (!IdxN)
          return false;
      }
      N = fastEmit_rr(VT, VT, ISD::ADD, N, IdxN);
      if (!N)
        return false;
    }
  }
  if (TotalOffs) {
    N = emitAdd_ri_(VT, N, TotalOffs);
    if (!N)
      return false;
  }
  updateValueMap(I, N);
  return true;
}

bool AArch64FastISel::selectAtomicCmpXchg(const AtomicCmpXchgInst *I) {
  assert(TM.getOptLevel() == CodeGenOptLevel::None &&
         "cmpxchg survived AtomicExpand at optlevel > -O0");

  auto *RetPairTy = cast<StructType>(I->getType());
  Type *RetTy = RetPairTy->getTypeAtIndex(0U);
  assert(RetPairTy->getTypeAtIndex(1U)->isIntegerTy(1) &&
         "cmpxchg has a non-i1 status result");

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

  const TargetRegisterClass *ResRC;
  unsigned Opc, CmpOpc;
  // This only supports i32/i64, because i8/i16 aren't legal, and the generic
  // extractvalue selection doesn't support that.
  if (VT == MVT::i32) {
    Opc = AArch64::CMP_SWAP_32;
    CmpOpc = AArch64::SUBSWrs;
    ResRC = &AArch64::GPR32RegClass;
  } else if (VT == MVT::i64) {
    Opc = AArch64::CMP_SWAP_64;
    CmpOpc = AArch64::SUBSXrs;
    ResRC = &AArch64::GPR64RegClass;
  } else {
    return false;
  }

  const MCInstrDesc &II = TII.get(Opc);

  const Register AddrReg = constrainOperandRegClass(
      II, getRegForValue(I->getPointerOperand()), II.getNumDefs());
  const Register DesiredReg = constrainOperandRegClass(
      II, getRegForValue(I->getCompareOperand()), II.getNumDefs() + 1);
  const Register NewReg = constrainOperandRegClass(
      II, getRegForValue(I->getNewValOperand()), II.getNumDefs() + 2);

  const Register ResultReg1 = createResultReg(ResRC);
  const Register ResultReg2 = createResultReg(&AArch64::GPR32RegClass);
  const Register ScratchReg = createResultReg(&AArch64::GPR32RegClass);

  // FIXME: MachineMemOperand doesn't support cmpxchg yet.
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, II)
      .addDef(ResultReg1)
      .addDef(ScratchReg)
      .addUse(AddrReg)
      .addUse(DesiredReg)
      .addUse(NewReg);

  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(CmpOpc))
      .addDef(VT == MVT::i32 ? AArch64::WZR : AArch64::XZR)
      .addUse(ResultReg1)
      .addUse(DesiredReg)
      .addImm(0);

  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, MIMD, TII.get(AArch64::CSINCWr))
      .addDef(ResultReg2)
      .addUse(AArch64::WZR)
      .addUse(AArch64::WZR)
      .addImm(AArch64CC::NE);

  assert((ResultReg1 + 1) == ResultReg2 && "Nonconsecutive result registers.");
  updateValueMap(I, ResultReg1, 2);
  return true;
}

bool AArch64FastISel::fastSelectInstruction(const Instruction *I) {
  if (TLI.fallBackToDAGISel(*I))
    return false;
  switch (I->getOpcode()) {
  default:
    break;
  case Instruction::Add:
  case Instruction::Sub:
    return selectAddSub(I);
  case Instruction::Mul:
    return selectMul(I);
  case Instruction::SDiv:
    return selectSDiv(I);
  case Instruction::SRem:
    if (!selectBinaryOp(I, ISD::SREM))
      return selectRem(I, ISD::SREM);
    return true;
  case Instruction::URem:
    if (!selectBinaryOp(I, ISD::UREM))
      return selectRem(I, ISD::UREM);
    return true;
  case Instruction::Shl:
  case Instruction::LShr:
  case Instruction::AShr:
    return selectShift(I);
  case Instruction::And:
  case Instruction::Or:
  case Instruction::Xor:
    return selectLogicalOp(I);
  case Instruction::Br:
    return selectBranch(I);
  case Instruction::IndirectBr:
    return selectIndirectBr(I);
  case Instruction::BitCast:
    if (!FastISel::selectBitCast(I))
      return selectBitCast(I);
    return true;
  case Instruction::FPToSI:
    if (!selectCast(I, ISD::FP_TO_SINT))
      return selectFPToInt(I, /*Signed=*/true);
    return true;
  case Instruction::FPToUI:
    return selectFPToInt(I, /*Signed=*/false);
  case Instruction::ZExt:
  case Instruction::SExt:
    return selectIntExt(I);
  case Instruction::Trunc:
    if (!selectCast(I, ISD::TRUNCATE))
      return selectTrunc(I);
    return true;
  case Instruction::FPExt:
    return selectFPExt(I);
  case Instruction::FPTrunc:
    return selectFPTrunc(I);
  case Instruction::SIToFP:
    if (!selectCast(I, ISD::SINT_TO_FP))
      return selectIntToFP(I, /*Signed=*/true);
    return true;
  case Instruction::UIToFP:
    return selectIntToFP(I, /*Signed=*/false);
  case Instruction::Load:
    return selectLoad(I);
  case Instruction::Store:
    return selectStore(I);
  case Instruction::FCmp:
  case Instruction::ICmp:
    return selectCmp(I);
  case Instruction::Select:
    return selectSelect(I);
  case Instruction::Ret:
    return selectRet(I);
  case Instruction::FRem:
    return selectFRem(I);
  case Instruction::GetElementPtr:
    return selectGetElementPtr(I);
  case Instruction::AtomicCmpXchg:
    return selectAtomicCmpXchg(cast<AtomicCmpXchgInst>(I));
  }

  // fall-back to target-independent instruction selection.
  return selectOperator(I, I->getOpcode());
}

FastISel *AArch64::createFastISel(FunctionLoweringInfo &FuncInfo,
                                        const TargetLibraryInfo *LibInfo) {

  SMEAttrs CallerAttrs(*FuncInfo.Fn);
  if (CallerAttrs.hasZAState() || CallerAttrs.hasStreamingInterfaceOrBody() ||
      CallerAttrs.hasStreamingCompatibleInterface())
    return nullptr;
  return new AArch64FastISel(FuncInfo, LibInfo);
}