lib/Target/ARM/ARMFastISel.cpp

//===-- ARMFastISel.cpp - ARM FastISel implementation ---------------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the ARM-specific support for the FastISel class. Some
// of the target-specific code is generated by tablegen in the file
// ARMGenFastISel.inc, which is #included here.
//
//===----------------------------------------------------------------------===//

#include "ARM.h"
#include "ARMBaseInstrInfo.h"
#include "ARMCallingConv.h"
#include "ARMRegisterInfo.h"
#include "ARMTargetMachine.h"
#include "ARMSubtarget.h"
#include "ARMConstantPoolValue.h"
#include "llvm/CallingConv.h"
#include "llvm/DerivedTypes.h"
#include "llvm/GlobalVariable.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Module.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;

static cl::opt<bool>
EnableARMFastISel("arm-fast-isel",
                    cl::desc("Turn on experimental ARM fast-isel support"),
                    cl::init(false), cl::Hidden);

namespace {

class ARMFastISel : public FastISel {

  /// Subtarget - Keep a pointer to the ARMSubtarget around so that we can
  /// make the right decision when generating code for different targets.
  const ARMSubtarget *Subtarget;
  const TargetMachine &TM;
  const TargetInstrInfo &TII;
  const TargetLowering &TLI;
  ARMFunctionInfo *AFI;

  // Convenience variables to avoid some queries.
  bool isThumb;
  LLVMContext *Context;

  public:
    explicit ARMFastISel(FunctionLoweringInfo &funcInfo)
    : FastISel(funcInfo),
      TM(funcInfo.MF->getTarget()),
      TII(*TM.getInstrInfo()),
      TLI(*TM.getTargetLowering()) {
      Subtarget = &TM.getSubtarget<ARMSubtarget>();
      AFI = funcInfo.MF->getInfo<ARMFunctionInfo>();
      isThumb = AFI->isThumbFunction();
      Context = &funcInfo.Fn->getContext();
    }

    // Code from FastISel.cpp.
    virtual unsigned FastEmitInst_(unsigned MachineInstOpcode,
                                   const TargetRegisterClass *RC);
    virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
                                    const TargetRegisterClass *RC,
                                    unsigned Op0, bool Op0IsKill);
    virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
                                     const TargetRegisterClass *RC,
                                     unsigned Op0, bool Op0IsKill,
                                     unsigned Op1, bool Op1IsKill);
    virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
                                     const TargetRegisterClass *RC,
                                     unsigned Op0, bool Op0IsKill,
                                     uint64_t Imm);
    virtual unsigned FastEmitInst_rf(unsigned MachineInstOpcode,
                                     const TargetRegisterClass *RC,
                                     unsigned Op0, bool Op0IsKill,
                                     const ConstantFP *FPImm);
    virtual unsigned FastEmitInst_i(unsigned MachineInstOpcode,
                                    const TargetRegisterClass *RC,
                                    uint64_t Imm);
    virtual unsigned FastEmitInst_rri(unsigned MachineInstOpcode,
                                      const TargetRegisterClass *RC,
                                      unsigned Op0, bool Op0IsKill,
                                      unsigned Op1, bool Op1IsKill,
                                      uint64_t Imm);
    virtual unsigned FastEmitInst_extractsubreg(MVT RetVT,
                                                unsigned Op0, bool Op0IsKill,
                                                uint32_t Idx);

    // Backend specific FastISel code.
    virtual bool TargetSelectInstruction(const Instruction *I);
    virtual unsigned TargetMaterializeConstant(const Constant *C);
    virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);

  #include "ARMGenFastISel.inc"

    // Instruction selection routines.
  private:
    virtual bool SelectLoad(const Instruction *I);
    virtual bool SelectStore(const Instruction *I);
    virtual bool SelectBranch(const Instruction *I);
    virtual bool SelectCmp(const Instruction *I);
    virtual bool SelectFPExt(const Instruction *I);
    virtual bool SelectFPTrunc(const Instruction *I);
    virtual bool SelectBinaryOp(const Instruction *I, unsigned ISDOpcode);
    virtual bool SelectSIToFP(const Instruction *I);
    virtual bool SelectFPToSI(const Instruction *I);
    virtual bool SelectSDiv(const Instruction *I);
    virtual bool SelectSRem(const Instruction *I);
    virtual bool SelectCall(const Instruction *I);
    virtual bool SelectSelect(const Instruction *I);

    // Utility routines.
  private:
    bool isTypeLegal(const Type *Ty, EVT &VT);
    bool isLoadTypeLegal(const Type *Ty, EVT &VT);
    bool ARMEmitLoad(EVT VT, unsigned &ResultReg, unsigned Reg, int Offset);
    bool ARMEmitStore(EVT VT, unsigned SrcReg, unsigned Reg, int Offset);
    bool ARMComputeRegOffset(const Value *Obj, unsigned &Reg, int &Offset);
    void ARMSimplifyRegOffset(unsigned &Reg, int &Offset, EVT VT);
    unsigned ARMMaterializeFP(const ConstantFP *CFP, EVT VT);
    unsigned ARMMaterializeInt(const Constant *C, EVT VT);
    unsigned ARMMaterializeGV(const GlobalValue *GV, EVT VT);
    unsigned ARMMoveToFPReg(EVT VT, unsigned SrcReg);
    unsigned ARMMoveToIntReg(EVT VT, unsigned SrcReg);

    // Call handling routines.
  private:
    CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool Return);
    bool ProcessCallArgs(SmallVectorImpl<Value*> &Args,
                         SmallVectorImpl<unsigned> &ArgRegs,
                         SmallVectorImpl<EVT> &ArgVTs,
                         SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
                         SmallVectorImpl<unsigned> &RegArgs,
                         CallingConv::ID CC,
                         unsigned &NumBytes);
    bool FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
                    const Instruction *I, CallingConv::ID CC,
                    unsigned &NumBytes);
    bool ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call);

    // OptionalDef handling routines.
  private:
    bool DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR);
    const MachineInstrBuilder &AddOptionalDefs(const MachineInstrBuilder &MIB);
};

} // end anonymous namespace

#include "ARMGenCallingConv.inc"

// DefinesOptionalPredicate - This is different from DefinesPredicate in that
// we don't care about implicit defs here, just places we'll need to add a
// default CCReg argument. Sets CPSR if we're setting CPSR instead of CCR.
bool ARMFastISel::DefinesOptionalPredicate(MachineInstr *MI, bool *CPSR) {
  const TargetInstrDesc &TID = MI->getDesc();
  if (!TID.hasOptionalDef())
    return false;

  // Look to see if our OptionalDef is defining CPSR or CCR.
  for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
    const MachineOperand &MO = MI->getOperand(i);
    if (!MO.isReg() || !MO.isDef()) continue;
    if (MO.getReg() == ARM::CPSR)
      *CPSR = true;
  }
  return true;
}

// If the machine is predicable go ahead and add the predicate operands, if
// it needs default CC operands add those.
const MachineInstrBuilder &
ARMFastISel::AddOptionalDefs(const MachineInstrBuilder &MIB) {
  MachineInstr *MI = &*MIB;

  // Do we use a predicate?
  if (TII.isPredicable(MI))
    AddDefaultPred(MIB);

  // Do we optionally set a predicate?  Preds is size > 0 iff the predicate
  // defines CPSR. All other OptionalDefines in ARM are the CCR register.
  bool CPSR = false;
  if (DefinesOptionalPredicate(MI, &CPSR)) {
    if (CPSR)
      AddDefaultT1CC(MIB);
    else
      AddDefaultCC(MIB);
  }
  return MIB;
}

unsigned ARMFastISel::FastEmitInst_(unsigned MachineInstOpcode,
                                    const TargetRegisterClass* RC) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg));
  return ResultReg;
}

unsigned ARMFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
                                     const TargetRegisterClass *RC,
                                     unsigned Op0, bool Op0IsKill) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
                   .addReg(Op0, Op0IsKill * RegState::Kill));
  else {
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
                   .addReg(Op0, Op0IsKill * RegState::Kill));
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                   TII.get(TargetOpcode::COPY), ResultReg)
                   .addReg(II.ImplicitDefs[0]));
  }
  return ResultReg;
}

unsigned ARMFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
                                      const TargetRegisterClass *RC,
                                      unsigned Op0, bool Op0IsKill,
                                      unsigned Op1, bool Op1IsKill) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addReg(Op1, Op1IsKill * RegState::Kill));
  else {
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addReg(Op1, Op1IsKill * RegState::Kill));
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                           TII.get(TargetOpcode::COPY), ResultReg)
                   .addReg(II.ImplicitDefs[0]));
  }
  return ResultReg;
}

unsigned ARMFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
                                      const TargetRegisterClass *RC,
                                      unsigned Op0, bool Op0IsKill,
                                      uint64_t Imm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addImm(Imm));
  else {
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addImm(Imm));
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                           TII.get(TargetOpcode::COPY), ResultReg)
                   .addReg(II.ImplicitDefs[0]));
  }
  return ResultReg;
}

unsigned ARMFastISel::FastEmitInst_rf(unsigned MachineInstOpcode,
                                      const TargetRegisterClass *RC,
                                      unsigned Op0, bool Op0IsKill,
                                      const ConstantFP *FPImm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addFPImm(FPImm));
  else {
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addFPImm(FPImm));
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                           TII.get(TargetOpcode::COPY), ResultReg)
                   .addReg(II.ImplicitDefs[0]));
  }
  return ResultReg;
}

unsigned ARMFastISel::FastEmitInst_rri(unsigned MachineInstOpcode,
                                       const TargetRegisterClass *RC,
                                       unsigned Op0, bool Op0IsKill,
                                       unsigned Op1, bool Op1IsKill,
                                       uint64_t Imm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addReg(Op1, Op1IsKill * RegState::Kill)
                   .addImm(Imm));
  else {
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
                   .addReg(Op0, Op0IsKill * RegState::Kill)
                   .addReg(Op1, Op1IsKill * RegState::Kill)
                   .addImm(Imm));
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                           TII.get(TargetOpcode::COPY), ResultReg)
                   .addReg(II.ImplicitDefs[0]));
  }
  return ResultReg;
}

unsigned ARMFastISel::FastEmitInst_i(unsigned MachineInstOpcode,
                                     const TargetRegisterClass *RC,
                                     uint64_t Imm) {
  unsigned ResultReg = createResultReg(RC);
  const TargetInstrDesc &II = TII.get(MachineInstOpcode);

  if (II.getNumDefs() >= 1)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II, ResultReg)
                   .addImm(Imm));
  else {
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II)
                   .addImm(Imm));
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                           TII.get(TargetOpcode::COPY), ResultReg)
                   .addReg(II.ImplicitDefs[0]));
  }
  return ResultReg;
}

unsigned ARMFastISel::FastEmitInst_extractsubreg(MVT RetVT,
                                                 unsigned Op0, bool Op0IsKill,
                                                 uint32_t Idx) {
  unsigned ResultReg = createResultReg(TLI.getRegClassFor(RetVT));
  assert(TargetRegisterInfo::isVirtualRegister(Op0) &&
         "Cannot yet extract from physregs");
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
                         DL, TII.get(TargetOpcode::COPY), ResultReg)
                 .addReg(Op0, getKillRegState(Op0IsKill), Idx));
  return ResultReg;
}

// TODO: Don't worry about 64-bit now, but when this is fixed remove the
// checks from the various callers.
unsigned ARMFastISel::ARMMoveToFPReg(EVT VT, unsigned SrcReg) {
  if (VT.getSimpleVT().SimpleTy == MVT::f64) return 0;

  unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                          TII.get(ARM::VMOVRS), MoveReg)
                  .addReg(SrcReg));
  return MoveReg;
}

unsigned ARMFastISel::ARMMoveToIntReg(EVT VT, unsigned SrcReg) {
  if (VT.getSimpleVT().SimpleTy == MVT::i64) return 0;

  unsigned MoveReg = createResultReg(TLI.getRegClassFor(VT));
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                          TII.get(ARM::VMOVSR), MoveReg)
                  .addReg(SrcReg));
  return MoveReg;
}

// For double width floating point we need to materialize two constants
// (the high and the low) into integer registers then use a move to get
// the combined constant into an FP reg.
unsigned ARMFastISel::ARMMaterializeFP(const ConstantFP *CFP, EVT VT) {
  const APFloat Val = CFP->getValueAPF();
  bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64;

  // This checks to see if we can use VFP3 instructions to materialize
  // a constant, otherwise we have to go through the constant pool.
  if (TLI.isFPImmLegal(Val, VT)) {
    unsigned Opc = is64bit ? ARM::FCONSTD : ARM::FCONSTS;
    unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
                            DestReg)
                    .addFPImm(CFP));
    return DestReg;
  }

  // Require VFP2 for loading fp constants.
  if (!Subtarget->hasVFP2()) return false;

  // MachineConstantPool wants an explicit alignment.
  unsigned Align = TD.getPrefTypeAlignment(CFP->getType());
  if (Align == 0) {
    // TODO: Figure out if this is correct.
    Align = TD.getTypeAllocSize(CFP->getType());
  }
  unsigned Idx = MCP.getConstantPoolIndex(cast<Constant>(CFP), Align);
  unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
  unsigned Opc = is64bit ? ARM::VLDRD : ARM::VLDRS;

  // The extra reg is for addrmode5.
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
                          DestReg)
                  .addConstantPoolIndex(Idx)
                  .addReg(0));
  return DestReg;
}

unsigned ARMFastISel::ARMMaterializeInt(const Constant *C, EVT VT) {

  // For now 32-bit only.
  if (VT.getSimpleVT().SimpleTy != MVT::i32) return false;

  // MachineConstantPool wants an explicit alignment.
  unsigned Align = TD.getPrefTypeAlignment(C->getType());
  if (Align == 0) {
    // TODO: Figure out if this is correct.
    Align = TD.getTypeAllocSize(C->getType());
  }
  unsigned Idx = MCP.getConstantPoolIndex(C, Align);
  unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));

  if (isThumb)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                            TII.get(ARM::t2LDRpci), DestReg)
                    .addConstantPoolIndex(Idx));
  else
    // The extra reg and immediate are for addrmode2.
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                            TII.get(ARM::LDRcp), DestReg)
                    .addConstantPoolIndex(Idx)
                    .addReg(0).addImm(0));

  return DestReg;
}

unsigned ARMFastISel::ARMMaterializeGV(const GlobalValue *GV, EVT VT) {
  // For now 32-bit only.
  if (VT.getSimpleVT().SimpleTy != MVT::i32) return 0;

  Reloc::Model RelocM = TM.getRelocationModel();

  // TODO: No external globals for now.
  if (Subtarget->GVIsIndirectSymbol(GV, RelocM)) return 0;

  // TODO: Need more magic for ARM PIC.
  if (!isThumb && (RelocM == Reloc::PIC_)) return 0;

  // MachineConstantPool wants an explicit alignment.
  unsigned Align = TD.getPrefTypeAlignment(GV->getType());
  if (Align == 0) {
    // TODO: Figure out if this is correct.
    Align = TD.getTypeAllocSize(GV->getType());
  }

  // Grab index.
  unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb() ? 4 : 8);
  unsigned Id = AFI->createConstPoolEntryUId();
  ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, Id,
                                                       ARMCP::CPValue, PCAdj);
  unsigned Idx = MCP.getConstantPoolIndex(CPV, Align);

  // Load value.
  MachineInstrBuilder MIB;
  unsigned DestReg = createResultReg(TLI.getRegClassFor(VT));
  if (isThumb) {
    unsigned Opc = (RelocM != Reloc::PIC_) ? ARM::t2LDRpci : ARM::t2LDRpci_pic;
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), DestReg)
          .addConstantPoolIndex(Idx);
    if (RelocM == Reloc::PIC_)
      MIB.addImm(Id);
  } else {
    // The extra reg and immediate are for addrmode2.
    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(ARM::LDRcp),
                  DestReg)
          .addConstantPoolIndex(Idx)
          .addReg(0).addImm(0);
  }
  AddOptionalDefs(MIB);
  return DestReg;
}

unsigned ARMFastISel::TargetMaterializeConstant(const Constant *C) {
  EVT VT = TLI.getValueType(C->getType(), true);

  // Only handle simple types.
  if (!VT.isSimple()) return 0;

  if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
    return ARMMaterializeFP(CFP, VT);
  else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
    return ARMMaterializeGV(GV, VT);
  else if (isa<ConstantInt>(C))
    return ARMMaterializeInt(C, VT);

  return 0;
}

unsigned ARMFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
  // Don't handle dynamic allocas.
  if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;

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

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

  // This will get lowered later into the correct offsets and registers
  // via rewriteXFrameIndex.
  if (SI != FuncInfo.StaticAllocaMap.end()) {
    TargetRegisterClass* RC = TLI.getRegClassFor(VT);
    unsigned ResultReg = createResultReg(RC);
    unsigned Opc = isThumb ? ARM::t2ADDri : ARM::ADDri;
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, *FuncInfo.InsertPt, DL,
                            TII.get(Opc), ResultReg)
                            .addFrameIndex(SI->second)
                            .addImm(0));
    return ResultReg;
  }

  return 0;
}

bool ARMFastISel::isTypeLegal(const Type *Ty, EVT &VT) {
  VT = TLI.getValueType(Ty, true);

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

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

bool ARMFastISel::isLoadTypeLegal(const Type *Ty, EVT &VT) {
  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::i8 || VT == MVT::i16)
    return true;

  return false;
}

// Computes the Reg+Offset to get to an object.
bool ARMFastISel::ARMComputeRegOffset(const Value *Obj, unsigned &Reg,
                                      int &Offset) {
  // Some boilerplate from the X86 FastISel.
  const User *U = NULL;
  unsigned Opcode = Instruction::UserOp1;
  if (const Instruction *I = dyn_cast<Instruction>(Obj)) {
    // Don't walk into other basic blocks; it's possible we haven't
    // visited them yet, so the instructions may not yet be assigned
    // virtual registers.
    if (FuncInfo.MBBMap[I->getParent()] != FuncInfo.MBB)
      return false;
    Opcode = I->getOpcode();
    U = I;
  } else if (const ConstantExpr *C = dyn_cast<ConstantExpr>(Obj)) {
    Opcode = C->getOpcode();
    U = C;
  }

  if (const PointerType *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 ARMComputeRegOffset(U->getOperand(0), Reg, Offset);
    }
    case Instruction::IntToPtr: {
      // Look past no-op inttoptrs.
      if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
        return ARMComputeRegOffset(U->getOperand(0), Reg, Offset);
      break;
    }
    case Instruction::PtrToInt: {
      // Look past no-op ptrtoints.
      if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
        return ARMComputeRegOffset(U->getOperand(0), Reg, Offset);
      break;
    }
    case Instruction::GetElementPtr: {
      int SavedOffset = Offset;
      unsigned SavedReg = Reg;
      int TmpOffset = Offset;

      // Iterate through the GEP folding the constants into offsets where
      // we can.
      gep_type_iterator GTI = gep_type_begin(U);
      for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
           i != e; ++i, ++GTI) {
        const Value *Op = *i;
        if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
          const StructLayout *SL = TD.getStructLayout(STy);
          unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
          TmpOffset += SL->getElementOffset(Idx);
        } else {
          uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
          SmallVector<const Value *, 4> Worklist;
          Worklist.push_back(Op);
          do {
            Op = Worklist.pop_back_val();
            if (const ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
              // Constant-offset addressing.
              TmpOffset += CI->getSExtValue() * S;
            } else if (0 && isa<AddOperator>(Op) &&
                       isa<ConstantInt>(cast<AddOperator>(Op)->getOperand(1))) {
              // An add with a constant operand. Fold the constant.
              ConstantInt *CI =
                cast<ConstantInt>(cast<AddOperator>(Op)->getOperand(1));
              TmpOffset += CI->getSExtValue() * S;
              // Add the other operand back to the work list.
              Worklist.push_back(cast<AddOperator>(Op)->getOperand(0));
            } else
              goto unsupported_gep;
          } while (!Worklist.empty());
        }
      }

      // Try to grab the base operand now.
      Offset = TmpOffset;
      if (ARMComputeRegOffset(U->getOperand(0), Reg, Offset)) return true;

      // We failed, restore everything and try the other options.
      Offset = SavedOffset;
      Reg = SavedReg;

      unsupported_gep:
      break;
    }
    case Instruction::Alloca: {
      // TODO: Fix this to do intermediate loads, etc.
      if (Offset != 0) return false;

      const AllocaInst *AI = cast<AllocaInst>(Obj);
      DenseMap<const AllocaInst*, int>::iterator SI =
        FuncInfo.StaticAllocaMap.find(AI);
      if (SI != FuncInfo.StaticAllocaMap.end()) {
        Reg = ARM::SP;
        Offset = SI->second;
        return true;
      }
      // Don't handle dynamic allocas.
      assert(!FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(Obj)) &&
             "Alloca should have been handled earlier!");
      return false;
    }
  }

  // Materialize the global variable's address into a reg which can
  // then be used later to load the variable.
  if (const GlobalValue *GV = dyn_cast<GlobalValue>(Obj)) {
    unsigned Tmp = ARMMaterializeGV(GV, TLI.getValueType(Obj->getType()));
    if (Tmp == 0) return false;

    Reg = Tmp;
    return true;
  }

  // Try to get this in a register if nothing else has worked.
  if (Reg == 0) Reg = getRegForValue(Obj);
  return Reg != 0;
}

void ARMFastISel::ARMSimplifyRegOffset(unsigned &Reg, int &Offset, EVT VT) {

  // Since the offset may be too large for the load instruction
  // get the reg+offset into a register.
  if (Reg != ARM::SP && Offset != 0) {
    ARMCC::CondCodes Pred = ARMCC::AL;
    unsigned PredReg = 0;

    TargetRegisterClass *RC = isThumb ? ARM::tGPRRegisterClass :
      ARM::GPRRegisterClass;
    unsigned BaseReg = createResultReg(RC);

    if (!isThumb)
      emitARMRegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                              BaseReg, Reg, Offset, Pred, PredReg,
                              static_cast<const ARMBaseInstrInfo&>(TII));
    else {
      assert(AFI->isThumb2Function());
      emitT2RegPlusImmediate(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                             BaseReg, Reg, Offset, Pred, PredReg,
                             static_cast<const ARMBaseInstrInfo&>(TII));
    }
    Offset = 0;
    Reg = BaseReg;
  }
}

bool ARMFastISel::ARMEmitLoad(EVT VT, unsigned &ResultReg,
                              unsigned Reg, int Offset) {

  assert(VT.isSimple() && "Non-simple types are invalid here!");
  unsigned Opc;
  TargetRegisterClass *RC;
  bool isFloat = false;
  switch (VT.getSimpleVT().SimpleTy) {
    default:
      // This is mostly going to be Neon/vector support.
      return false;
    case MVT::i16:
      Opc = isThumb ? ARM::t2LDRHi8 : ARM::LDRH;
      RC = ARM::GPRRegisterClass;
      VT = MVT::i32;
      break;
    case MVT::i8:
      Opc = isThumb ? ARM::t2LDRBi8 : ARM::LDRB;
      RC = ARM::GPRRegisterClass;
      VT = MVT::i32;
      break;
    case MVT::i32:
      Opc = isThumb ? ARM::t2LDRi8 : ARM::LDR;
      RC = ARM::GPRRegisterClass;
      break;
    case MVT::f32:
      Opc = ARM::VLDRS;
      RC = TLI.getRegClassFor(VT);
      isFloat = true;
      break;
    case MVT::f64:
      Opc = ARM::VLDRD;
      RC = TLI.getRegClassFor(VT);
      isFloat = true;
      break;
  }

  ResultReg = createResultReg(RC);

  // For now with the additions above the offset should be zero - thus we
  // can always fit into an i8.
  assert((Reg == ARM::SP || Offset == 0) &&
          "Offset not zero and not a stack load!");

  if (Reg == ARM::SP)
    TII.loadRegFromStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
                             ResultReg, Offset, RC,
                             TM.getRegisterInfo());
  // The thumb and floating point instructions both take 2 operands, ARM takes
  // another register.
  else if (isFloat || isThumb)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                            TII.get(Opc), ResultReg)
                    .addReg(Reg).addImm(Offset));
  else
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                            TII.get(Opc), ResultReg)
                    .addReg(Reg).addReg(0).addImm(Offset));
  return true;
}

bool ARMFastISel::SelectLoad(const Instruction *I) {
  // Verify we have a legal type before going any further.
  EVT VT;
  if (!isLoadTypeLegal(I->getType(), VT))
    return false;

  // Our register and offset with innocuous defaults.
  unsigned Reg = 0;
  int Offset = 0;

  // See if we can handle this as Reg + Offset
  if (!ARMComputeRegOffset(I->getOperand(0), Reg, Offset))
    return false;

  ARMSimplifyRegOffset(Reg, Offset, VT);

  unsigned ResultReg;
  if (!ARMEmitLoad(VT, ResultReg, Reg, Offset)) return false;

  UpdateValueMap(I, ResultReg);
  return true;
}

bool ARMFastISel::ARMEmitStore(EVT VT, unsigned SrcReg,
                               unsigned DstReg, int Offset) {
  unsigned StrOpc;
  bool isFloat = false;
  // VT is set here only for use in the alloca stores below - those are promoted
  // to reg size always.
  switch (VT.getSimpleVT().SimpleTy) {
    default: return false;
    case MVT::i1:
    case MVT::i8:
      VT = MVT::i32;
      StrOpc = isThumb ? ARM::t2STRBi8 : ARM::STRB;
      break;
    case MVT::i16:
      VT = MVT::i32;
      StrOpc = isThumb ? ARM::t2STRHi8 : ARM::STRH;
      break;
    case MVT::i32: StrOpc = isThumb ? ARM::t2STRi8 : ARM::STR; break;
    case MVT::f32:
      if (!Subtarget->hasVFP2()) return false;
      StrOpc = ARM::VSTRS;
      isFloat = true;
      break;
    case MVT::f64:
      if (!Subtarget->hasVFP2()) return false;
      StrOpc = ARM::VSTRD;
      isFloat = true;
      break;
  }

  if (DstReg == ARM::SP)
    TII.storeRegToStackSlot(*FuncInfo.MBB, *FuncInfo.InsertPt,
                            SrcReg, true /*isKill*/, Offset,
                            TLI.getRegClassFor(VT), TM.getRegisterInfo());
  // The thumb addressing mode has operands swapped from the arm addressing
  // mode, the floating point one only has two operands.
  if (isFloat || isThumb)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                            TII.get(StrOpc))
                    .addReg(SrcReg).addReg(DstReg).addImm(Offset));
  else
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                            TII.get(StrOpc))
                    .addReg(SrcReg).addReg(DstReg).addReg(0).addImm(Offset));

  return true;
}

bool ARMFastISel::SelectStore(const Instruction *I) {
  Value *Op0 = I->getOperand(0);
  unsigned SrcReg = 0;

  // Yay type legalization
  EVT VT;
  if (!isLoadTypeLegal(I->getOperand(0)->getType(), VT))
    return false;

  // Get the value to be stored into a register.
  SrcReg = getRegForValue(Op0);
  if (SrcReg == 0)
    return false;

  // Our register and offset with innocuous defaults.
  unsigned Reg = 0;
  int Offset = 0;

  // See if we can handle this as Reg + Offset
  if (!ARMComputeRegOffset(I->getOperand(1), Reg, Offset))
    return false;

  ARMSimplifyRegOffset(Reg, Offset, VT);

  if (!ARMEmitStore(VT, SrcReg, Reg, Offset)) return false;

  return true;
}

static ARMCC::CondCodes getComparePred(CmpInst::Predicate Pred) {
  switch (Pred) {
    // Needs two compares...
    case CmpInst::FCMP_ONE:
    case CmpInst::FCMP_UEQ:
    default:
      assert(false && "Unhandled CmpInst::Predicate!");
      return ARMCC::AL;
    case CmpInst::ICMP_EQ:
    case CmpInst::FCMP_OEQ:
      return ARMCC::EQ;
    case CmpInst::ICMP_SGT:
    case CmpInst::FCMP_OGT:
      return ARMCC::GT;
    case CmpInst::ICMP_SGE:
    case CmpInst::FCMP_OGE:
      return ARMCC::GE;
    case CmpInst::ICMP_UGT:
    case CmpInst::FCMP_UGT:
      return ARMCC::HI;
    case CmpInst::FCMP_OLT:
      return ARMCC::MI;
    case CmpInst::ICMP_ULE:
    case CmpInst::FCMP_OLE:
      return ARMCC::LS;
    case CmpInst::FCMP_ORD:
      return ARMCC::VC;
    case CmpInst::FCMP_UNO:
      return ARMCC::VS;
    case CmpInst::FCMP_UGE:
      return ARMCC::PL;
    case CmpInst::ICMP_SLT:
    case CmpInst::FCMP_ULT:
      return ARMCC::LT;
    case CmpInst::ICMP_SLE:
    case CmpInst::FCMP_ULE:
      return ARMCC::LE;
    case CmpInst::FCMP_UNE:
    case CmpInst::ICMP_NE:
      return ARMCC::NE;
    case CmpInst::ICMP_UGE:
      return ARMCC::HS;
    case CmpInst::ICMP_ULT:
      return ARMCC::LO;
  }
}

bool ARMFastISel::SelectBranch(const Instruction *I) {
  const BranchInst *BI = cast<BranchInst>(I);
  MachineBasicBlock *TBB = FuncInfo.MBBMap[BI->getSuccessor(0)];
  MachineBasicBlock *FBB = FuncInfo.MBBMap[BI->getSuccessor(1)];

  // Simple branch support.
  // TODO: Try to avoid the re-computation in some places.
  unsigned CondReg = getRegForValue(BI->getCondition());
  if (CondReg == 0) return false;

  // Re-set the flags just in case.
  unsigned CmpOpc = isThumb ? ARM::t2CMPri : ARM::CMPri;
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
                  .addReg(CondReg).addImm(1));

  unsigned BrOpc = isThumb ? ARM::t2Bcc : ARM::Bcc;
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BrOpc))
                  .addMBB(TBB).addImm(ARMCC::EQ).addReg(ARM::CPSR);
  FastEmitBranch(FBB, DL);
  FuncInfo.MBB->addSuccessor(TBB);
  return true;
}

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

  EVT VT;
  const Type *Ty = CI->getOperand(0)->getType();
  if (!isTypeLegal(Ty, VT))
    return false;

  bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
  if (isFloat && !Subtarget->hasVFP2())
    return false;

  unsigned CmpOpc;
  unsigned CondReg;
  switch (VT.getSimpleVT().SimpleTy) {
    default: return false;
    // TODO: Verify compares.
    case MVT::f32:
      CmpOpc = ARM::VCMPES;
      CondReg = ARM::FPSCR;
      break;
    case MVT::f64:
      CmpOpc = ARM::VCMPED;
      CondReg = ARM::FPSCR;
      break;
    case MVT::i32:
      CmpOpc = isThumb ? ARM::t2CMPrr : ARM::CMPrr;
      CondReg = ARM::CPSR;
      break;
  }

  // Get the compare predicate.
  ARMCC::CondCodes ARMPred = getComparePred(CI->getPredicate());

  // We may not handle every CC for now.
  if (ARMPred == ARMCC::AL) return false;

  unsigned Arg1 = getRegForValue(CI->getOperand(0));
  if (Arg1 == 0) return false;

  unsigned Arg2 = getRegForValue(CI->getOperand(1));
  if (Arg2 == 0) return false;

  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
                  .addReg(Arg1).addReg(Arg2));

  // For floating point we need to move the result to a comparison register
  // that we can then use for branches.
  if (isFloat)
    AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                            TII.get(ARM::FMSTAT)));

  // Now set a register based on the comparison. Explicitly set the predicates
  // here.
  unsigned MovCCOpc = isThumb ? ARM::t2MOVCCi : ARM::MOVCCi;
  TargetRegisterClass *RC = isThumb ? ARM::rGPRRegisterClass
                                    : ARM::GPRRegisterClass;
  unsigned DestReg = createResultReg(RC);
  Constant *Zero
    = ConstantInt::get(Type::getInt32Ty(*Context), 0);
  unsigned ZeroReg = TargetMaterializeConstant(Zero);
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), DestReg)
          .addReg(ZeroReg).addImm(1)
          .addImm(ARMPred).addReg(CondReg);

  UpdateValueMap(I, DestReg);
  return true;
}

bool ARMFastISel::SelectFPExt(const Instruction *I) {
  // Make sure we have VFP and that we're extending float to double.
  if (!Subtarget->hasVFP2()) return false;

  Value *V = I->getOperand(0);
  if (!I->getType()->isDoubleTy() ||
      !V->getType()->isFloatTy()) return false;

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

  unsigned Result = createResultReg(ARM::DPRRegisterClass);
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                          TII.get(ARM::VCVTDS), Result)
                  .addReg(Op));
  UpdateValueMap(I, Result);
  return true;
}

bool ARMFastISel::SelectFPTrunc(const Instruction *I) {
  // Make sure we have VFP and that we're truncating double to float.
  if (!Subtarget->hasVFP2()) return false;

  Value *V = I->getOperand(0);
  if (!(I->getType()->isFloatTy() &&
        V->getType()->isDoubleTy())) return false;

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

  unsigned Result = createResultReg(ARM::SPRRegisterClass);
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                          TII.get(ARM::VCVTSD), Result)
                  .addReg(Op));
  UpdateValueMap(I, Result);
  return true;
}

bool ARMFastISel::SelectSIToFP(const Instruction *I) {
  // Make sure we have VFP.
  if (!Subtarget->hasVFP2()) return false;

  EVT DstVT;
  const Type *Ty = I->getType();
  if (!isTypeLegal(Ty, DstVT))
    return false;

  unsigned Op = getRegForValue(I->getOperand(0));
  if (Op == 0) return false;

  // The conversion routine works on fp-reg to fp-reg and the operand above
  // was an integer, move it to the fp registers if possible.
  unsigned FP = ARMMoveToFPReg(MVT::f32, Op);
  if (FP == 0) return false;

  unsigned Opc;
  if (Ty->isFloatTy()) Opc = ARM::VSITOS;
  else if (Ty->isDoubleTy()) Opc = ARM::VSITOD;
  else return 0;

  unsigned ResultReg = createResultReg(TLI.getRegClassFor(DstVT));
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
                          ResultReg)
                  .addReg(FP));
  UpdateValueMap(I, ResultReg);
  return true;
}

bool ARMFastISel::SelectFPToSI(const Instruction *I) {
  // Make sure we have VFP.
  if (!Subtarget->hasVFP2()) return false;

  EVT DstVT;
  const Type *RetTy = I->getType();
  if (!isTypeLegal(RetTy, DstVT))
    return false;

  unsigned Op = getRegForValue(I->getOperand(0));
  if (Op == 0) return false;

  unsigned Opc;
  const Type *OpTy = I->getOperand(0)->getType();
  if (OpTy->isFloatTy()) Opc = ARM::VTOSIZS;
  else if (OpTy->isDoubleTy()) Opc = ARM::VTOSIZD;
  else return 0;

  // f64->s32 or f32->s32 both need an intermediate f32 reg.
  unsigned ResultReg = createResultReg(TLI.getRegClassFor(MVT::f32));
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc),
                          ResultReg)
                  .addReg(Op));

  // This result needs to be in an integer register, but the conversion only
  // takes place in fp-regs.
  unsigned IntReg = ARMMoveToIntReg(DstVT, ResultReg);
  if (IntReg == 0) return false;

  UpdateValueMap(I, IntReg);
  return true;
}

bool ARMFastISel::SelectSelect(const Instruction *I) {
  EVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
  if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
    return false;

  // Things need to be register sized for register moves.
  if (VT.getSimpleVT().SimpleTy != MVT::i32) return false;
  const TargetRegisterClass *RC = TLI.getRegClassFor(VT);

  unsigned CondReg = getRegForValue(I->getOperand(0));
  if (CondReg == 0) return false;
  unsigned Op1Reg = getRegForValue(I->getOperand(1));
  if (Op1Reg == 0) return false;
  unsigned Op2Reg = getRegForValue(I->getOperand(2));
  if (Op2Reg == 0) return false;

  unsigned CmpOpc = isThumb ? ARM::t2TSTri : ARM::TSTri;
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CmpOpc))
                  .addReg(CondReg).addImm(1));
  unsigned ResultReg = createResultReg(RC);
  unsigned MovCCOpc = isThumb ? ARM::t2MOVCCr : ARM::MOVCCr;
  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(MovCCOpc), ResultReg)
    .addReg(Op1Reg).addReg(Op2Reg)
    .addImm(ARMCC::EQ).addReg(ARM::CPSR);
  UpdateValueMap(I, ResultReg);
  return true;
}

bool ARMFastISel::SelectSDiv(const Instruction *I) {
  EVT VT;
  const Type *Ty = I->getType();
  if (!isTypeLegal(Ty, VT))
    return false;

  // If we have integer div support we should have selected this automagically.
  // In case we have a real miss go ahead and return false and we'll pick
  // it up later.
  if (Subtarget->hasDivide()) return false;

  // Otherwise emit a libcall.
  RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
  if (VT == MVT::i8)
    LC = RTLIB::SDIV_I8;
  else if (VT == MVT::i16)
    LC = RTLIB::SDIV_I16;
  else if (VT == MVT::i32)
    LC = RTLIB::SDIV_I32;
  else if (VT == MVT::i64)
    LC = RTLIB::SDIV_I64;
  else if (VT == MVT::i128)
    LC = RTLIB::SDIV_I128;
  assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SDIV!");

  return ARMEmitLibcall(I, LC);
}

bool ARMFastISel::SelectSRem(const Instruction *I) {
  EVT VT;
  const Type *Ty = I->getType();
  if (!isTypeLegal(Ty, VT))
    return false;

  RTLIB::Libcall LC = RTLIB::UNKNOWN_LIBCALL;
  if (VT == MVT::i8)
    LC = RTLIB::SREM_I8;
  else if (VT == MVT::i16)
    LC = RTLIB::SREM_I16;
  else if (VT == MVT::i32)
    LC = RTLIB::SREM_I32;
  else if (VT == MVT::i64)
    LC = RTLIB::SREM_I64;
  else if (VT == MVT::i128)
    LC = RTLIB::SREM_I128;
  assert(LC != RTLIB::UNKNOWN_LIBCALL && "Unsupported SREM!");

  return ARMEmitLibcall(I, LC);
}

bool ARMFastISel::SelectBinaryOp(const Instruction *I, unsigned ISDOpcode) {
  EVT VT  = TLI.getValueType(I->getType(), true);

  // We can get here in the case when we want to use NEON for our fp
  // operations, but can't figure out how to. Just use the vfp instructions
  // if we have them.
  // FIXME: It'd be nice to use NEON instructions.
  const Type *Ty = I->getType();
  bool isFloat = (Ty->isDoubleTy() || Ty->isFloatTy());
  if (isFloat && !Subtarget->hasVFP2())
    return false;

  unsigned Op1 = getRegForValue(I->getOperand(0));
  if (Op1 == 0) return false;

  unsigned Op2 = getRegForValue(I->getOperand(1));
  if (Op2 == 0) return false;

  unsigned Opc;
  bool is64bit = VT.getSimpleVT().SimpleTy == MVT::f64 ||
                 VT.getSimpleVT().SimpleTy == MVT::i64;
  switch (ISDOpcode) {
    default: return false;
    case ISD::FADD:
      Opc = is64bit ? ARM::VADDD : ARM::VADDS;
      break;
    case ISD::FSUB:
      Opc = is64bit ? ARM::VSUBD : ARM::VSUBS;
      break;
    case ISD::FMUL:
      Opc = is64bit ? ARM::VMULD : ARM::VMULS;
      break;
  }
  unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                          TII.get(Opc), ResultReg)
                  .addReg(Op1).addReg(Op2));
  UpdateValueMap(I, ResultReg);
  return true;
}

// Call Handling Code

// This is largely taken directly from CCAssignFnForNode - we don't support
// varargs in FastISel so that part has been removed.
// TODO: We may not support all of this.
CCAssignFn *ARMFastISel::CCAssignFnForCall(CallingConv::ID CC, bool Return) {
  switch (CC) {
  default:
    llvm_unreachable("Unsupported calling convention");
  case CallingConv::C:
  case CallingConv::Fast:
    // Use target triple & subtarget features to do actual dispatch.
    if (Subtarget->isAAPCS_ABI()) {
      if (Subtarget->hasVFP2() &&
          FloatABIType == FloatABI::Hard)
        return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
      else
        return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
    } else
        return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
  case CallingConv::ARM_AAPCS_VFP:
    return (Return ? RetCC_ARM_AAPCS_VFP: CC_ARM_AAPCS_VFP);
  case CallingConv::ARM_AAPCS:
    return (Return ? RetCC_ARM_AAPCS: CC_ARM_AAPCS);
  case CallingConv::ARM_APCS:
    return (Return ? RetCC_ARM_APCS: CC_ARM_APCS);
  }
}

bool ARMFastISel::ProcessCallArgs(SmallVectorImpl<Value*> &Args,
                                  SmallVectorImpl<unsigned> &ArgRegs,
                                  SmallVectorImpl<EVT> &ArgVTs,
                                  SmallVectorImpl<ISD::ArgFlagsTy> &ArgFlags,
                                  SmallVectorImpl<unsigned> &RegArgs,
                                  CallingConv::ID CC,
                                  unsigned &NumBytes) {
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CC, false, TM, ArgLocs, *Context);
  CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC, false));

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

  // Issue CALLSEQ_START
  unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                          TII.get(AdjStackDown))
                  .addImm(NumBytes));

  // Process the args.
  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
    CCValAssign &VA = ArgLocs[i];
    unsigned Arg = ArgRegs[VA.getValNo()];
    EVT ArgVT = ArgVTs[VA.getValNo()];

    // Handle arg promotion, etc.
    switch (VA.getLocInfo()) {
      case CCValAssign::Full: break;
      default:
      // TODO: Handle arg promotion.
      return false;
    }

    // Now copy/store arg to correct locations.
    // TODO: We need custom lowering for f64 args.
    if (VA.isRegLoc() && !VA.needsCustom()) {
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
              VA.getLocReg())
      .addReg(Arg);
      RegArgs.push_back(VA.getLocReg());
    } else {
      // Need to store
      return false;
    }
  }

  return true;
}

bool ARMFastISel::FinishCall(EVT RetVT, SmallVectorImpl<unsigned> &UsedRegs,
                             const Instruction *I, CallingConv::ID CC,
                             unsigned &NumBytes) {
  // Issue CALLSEQ_END
  unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
  AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                          TII.get(AdjStackUp))
                  .addImm(NumBytes).addImm(0));

  // Now the return value.
  if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) {
    SmallVector<CCValAssign, 16> RVLocs;
    CCState CCInfo(CC, false, TM, RVLocs, *Context);
    CCInfo.AnalyzeCallResult(RetVT, CCAssignFnForCall(CC, true));

    // Copy all of the result registers out of their specified physreg.
    if (RVLocs.size() == 2 && RetVT.getSimpleVT().SimpleTy == MVT::f64) {
      // For this move we copy into two registers and then move into the
      // double fp reg we want.
      // TODO: Are the copies necessary?
      TargetRegisterClass *CopyRC = TLI.getRegClassFor(MVT::i32);
      unsigned Copy1 = createResultReg(CopyRC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
              Copy1).addReg(RVLocs[0].getLocReg());
      UsedRegs.push_back(RVLocs[0].getLocReg());

      unsigned Copy2 = createResultReg(CopyRC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
              Copy2).addReg(RVLocs[1].getLocReg());
      UsedRegs.push_back(RVLocs[1].getLocReg());

      EVT DestVT = RVLocs[0].getValVT();
      TargetRegisterClass* DstRC = TLI.getRegClassFor(DestVT);
      unsigned ResultReg = createResultReg(DstRC);
      AddOptionalDefs(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
                              TII.get(ARM::VMOVDRR), ResultReg)
                      .addReg(Copy1).addReg(Copy2));

      // Finally update the result.
      UpdateValueMap(I, ResultReg);
    } else {
      assert(RVLocs.size() == 1 &&"Can't handle non-double multi-reg retvals!");
      EVT CopyVT = RVLocs[0].getValVT();
      TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);

      unsigned ResultReg = createResultReg(DstRC);
      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
              ResultReg).addReg(RVLocs[0].getLocReg());
      UsedRegs.push_back(RVLocs[0].getLocReg());

      // Finally update the result.
      UpdateValueMap(I, ResultReg);
    }
  }

  return true;
}

// A quick function that will emit a call for a named libcall in F with the
// vector of passed arguments for the Instruction in I. We can assume that we
// can emit a call for any libcall we can produce. This is an abridged version
// of the full call infrastructure since we won't need to worry about things
// like computed function pointers or strange arguments at call sites.
// TODO: Try to unify this and the normal call bits for ARM, then try to unify
// with X86.
bool ARMFastISel::ARMEmitLibcall(const Instruction *I, RTLIB::Libcall Call) {
  CallingConv::ID CC = TLI.getLibcallCallingConv(Call);

  // Handle *simple* calls for now.
  const Type *RetTy = I->getType();
  EVT RetVT;
  if (RetTy->isVoidTy())
    RetVT = MVT::isVoid;
  else if (!isTypeLegal(RetTy, RetVT))
    return false;

  // For now we're using BLX etc on the assumption that we have v5t ops.
  if (!Subtarget->hasV5TOps()) return false;

  // Set up the argument vectors.
  SmallVector<Value*, 8> Args;
  SmallVector<unsigned, 8> ArgRegs;
  SmallVector<EVT, 8> ArgVTs;
  SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
  Args.reserve(I->getNumOperands());
  ArgRegs.reserve(I->getNumOperands());
  ArgVTs.reserve(I->getNumOperands());
  ArgFlags.reserve(I->getNumOperands());
  for (unsigned i = 0; i < I->getNumOperands(); ++i) {
    Value *Op = I->getOperand(i);
    unsigned Arg = getRegForValue(Op);
    if (Arg == 0) return false;

    const Type *ArgTy = Op->getType();
    EVT ArgVT;
    if (!isTypeLegal(ArgTy, ArgVT)) return false;

    ISD::ArgFlagsTy Flags;
    unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
    Flags.setOrigAlign(OriginalAlignment);

    Args.push_back(Op);
    ArgRegs.push_back(Arg);
    ArgVTs.push_back(ArgVT);
    ArgFlags.push_back(Flags);
  }

  // Handle the arguments now that we've gotten them.
  SmallVector<unsigned, 4> RegArgs;
  unsigned NumBytes;
  if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
    return false;

  // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
  // TODO: Turn this into the table of arm call ops.
  MachineInstrBuilder MIB;
  unsigned CallOpc;
  if(isThumb)
    CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
  else
    CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
  MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
        .addExternalSymbol(TLI.getLibcallName(Call));

  // Add implicit physical register uses to the call.
  for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
    MIB.addReg(RegArgs[i]);

  // Finish off the call including any return values.
  SmallVector<unsigned, 4> UsedRegs;
  if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;

  // Set all unused physreg defs as dead.
  static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);

  return true;
}

bool ARMFastISel::SelectCall(const Instruction *I) {
  const CallInst *CI = cast<CallInst>(I);
  const Value *Callee = CI->getCalledValue();

  // Can't handle inline asm or worry about intrinsics yet.
  if (isa<InlineAsm>(Callee) || isa<IntrinsicInst>(CI)) return false;

  // Only handle global variable Callees that are direct calls.
  const GlobalValue *GV = dyn_cast<GlobalValue>(Callee);
  if (!GV || Subtarget->GVIsIndirectSymbol(GV, TM.getRelocationModel()))
    return false;

  // Check the calling convention.
  ImmutableCallSite CS(CI);
  CallingConv::ID CC = CS.getCallingConv();
  // TODO: Avoid some calling conventions?
  if (CC != CallingConv::C) {
    // errs() << "Can't handle calling convention: " << CC << "\n";
    return false;
  }

  // Let SDISel handle vararg functions.
  const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
  const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
  if (FTy->isVarArg())
    return false;

  // Handle *simple* calls for now.
  const Type *RetTy = I->getType();
  EVT RetVT;
  if (RetTy->isVoidTy())
    RetVT = MVT::isVoid;
  else if (!isTypeLegal(RetTy, RetVT))
    return false;

  // For now we're using BLX etc on the assumption that we have v5t ops.
  // TODO: Maybe?
  if (!Subtarget->hasV5TOps()) return false;

  // Set up the argument vectors.
  SmallVector<Value*, 8> Args;
  SmallVector<unsigned, 8> ArgRegs;
  SmallVector<EVT, 8> ArgVTs;
  SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
  Args.reserve(CS.arg_size());
  ArgRegs.reserve(CS.arg_size());
  ArgVTs.reserve(CS.arg_size());
  ArgFlags.reserve(CS.arg_size());
  for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
       i != e; ++i) {
    unsigned Arg = getRegForValue(*i);

    if (Arg == 0)
      return false;
    ISD::ArgFlagsTy Flags;
    unsigned AttrInd = i - CS.arg_begin() + 1;
    if (CS.paramHasAttr(AttrInd, Attribute::SExt))
      Flags.setSExt();
    if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
      Flags.setZExt();

         // FIXME: Only handle *easy* calls for now.
    if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
        CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
        CS.paramHasAttr(AttrInd, Attribute::Nest) ||
        CS.paramHasAttr(AttrInd, Attribute::ByVal))
      return false;

    const Type *ArgTy = (*i)->getType();
    EVT ArgVT;
    if (!isTypeLegal(ArgTy, ArgVT))
      return false;
    unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
    Flags.setOrigAlign(OriginalAlignment);

    Args.push_back(*i);
    ArgRegs.push_back(Arg);
    ArgVTs.push_back(ArgVT);
    ArgFlags.push_back(Flags);
  }

  // Handle the arguments now that we've gotten them.
  SmallVector<unsigned, 4> RegArgs;
  unsigned NumBytes;
  if (!ProcessCallArgs(Args, ArgRegs, ArgVTs, ArgFlags, RegArgs, CC, NumBytes))
    return false;

  // Issue the call, BLXr9 for darwin, BLX otherwise. This uses V5 ops.
  // TODO: Turn this into the table of arm call ops.
  MachineInstrBuilder MIB;
  unsigned CallOpc;
  if(isThumb)
    CallOpc = Subtarget->isTargetDarwin() ? ARM::tBLXi_r9 : ARM::tBLXi;
  else
    CallOpc = Subtarget->isTargetDarwin() ? ARM::BLr9 : ARM::BL;
  MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
              .addGlobalAddress(GV, 0, 0);

  // Add implicit physical register uses to the call.
  for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
    MIB.addReg(RegArgs[i]);

  // Finish off the call including any return values.
  SmallVector<unsigned, 4> UsedRegs;
  if (!FinishCall(RetVT, UsedRegs, I, CC, NumBytes)) return false;

  // Set all unused physreg defs as dead.
  static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);

  return true;

}

// TODO: SoftFP support.
bool ARMFastISel::TargetSelectInstruction(const Instruction *I) {
  // No Thumb-1 for now.
  if (isThumb && !AFI->isThumb2Function()) return false;

  switch (I->getOpcode()) {
    case Instruction::Load:
      return SelectLoad(I);
    case Instruction::Store:
      return SelectStore(I);
    case Instruction::Br:
      return SelectBranch(I);
    case Instruction::ICmp:
    case Instruction::FCmp:
      return SelectCmp(I);
    case Instruction::FPExt:
      return SelectFPExt(I);
    case Instruction::FPTrunc:
      return SelectFPTrunc(I);
    case Instruction::SIToFP:
      return SelectSIToFP(I);
    case Instruction::FPToSI:
      return SelectFPToSI(I);
    case Instruction::FAdd:
      return SelectBinaryOp(I, ISD::FADD);
    case Instruction::FSub:
      return SelectBinaryOp(I, ISD::FSUB);
    case Instruction::FMul:
      return SelectBinaryOp(I, ISD::FMUL);
    case Instruction::SDiv:
      return SelectSDiv(I);
    case Instruction::SRem:
      return SelectSRem(I);
    case Instruction::Call:
      return SelectCall(I);
    case Instruction::Select:
      return SelectSelect(I);
    default: break;
  }
  return false;
}

namespace llvm {
  llvm::FastISel *ARM::createFastISel(FunctionLoweringInfo &funcInfo) {
    // Completely untested on non-darwin.
    const TargetMachine &TM = funcInfo.MF->getTarget();
    const ARMSubtarget *Subtarget = &TM.getSubtarget<ARMSubtarget>();
    if (Subtarget->isTargetDarwin() && EnableARMFastISel)
      return new ARMFastISel(funcInfo);
    return 0;
  }
}