Target/X86/X86FastISel.cpp

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

#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
#include "X86RegisterInfo.h"
#include "X86Subtarget.h"
#include "X86TargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/DerivedTypes.h"
#include "llvm/GlobalVariable.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;

namespace {

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

  /// StackPtr - Register used as the stack pointer.
  ///
  unsigned StackPtr;

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

public:
  explicit X86FastISel(MachineFunction &mf,
                       MachineModuleInfo *mmi,
                       DwarfWriter *dw,
                       DenseMap<const Value *, unsigned> &vm,
                       DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
                       DenseMap<const AllocaInst *, int> &am
#ifndef NDEBUG
                       , SmallSet<Instruction*, 8> &cil
#endif
                       )
    : FastISel(mf, mmi, dw, vm, bm, am
#ifndef NDEBUG
               , cil
#endif
               ) {
    Subtarget = &TM.getSubtarget<X86Subtarget>();
    StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
    X86ScalarSSEf64 = Subtarget->hasSSE2();
    X86ScalarSSEf32 = Subtarget->hasSSE1();
  }

  virtual bool TargetSelectInstruction(Instruction *I);

#include "X86GenFastISel.inc"

private:
  bool X86FastEmitCompare(Value *LHS, Value *RHS, EVT VT);

  bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR);

  bool X86FastEmitStore(EVT VT, Value *Val,
                        const X86AddressMode &AM);
  bool X86FastEmitStore(EVT VT, unsigned Val,
                        const X86AddressMode &AM);

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

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

  bool X86SelectLoad(Instruction *I);

  bool X86SelectStore(Instruction *I);

  bool X86SelectCmp(Instruction *I);

  bool X86SelectZExt(Instruction *I);

  bool X86SelectBranch(Instruction *I);

  bool X86SelectShift(Instruction *I);

  bool X86SelectSelect(Instruction *I);

  bool X86SelectTrunc(Instruction *I);

  bool X86SelectFPExt(Instruction *I);
  bool X86SelectFPTrunc(Instruction *I);

  bool X86SelectExtractValue(Instruction *I);

  bool X86VisitIntrinsicCall(IntrinsicInst &I);
  bool X86SelectCall(Instruction *I);

  CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool isTailCall = false);

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

  unsigned TargetMaterializeConstant(Constant *C);

  unsigned TargetMaterializeAlloca(AllocaInst *C);

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

  bool isTypeLegal(const Type *Ty, EVT &VT, bool AllowI1 = false);
};

} // end anonymous namespace.

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

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

#include "X86GenCallingConv.inc"

/// CCAssignFnForCall - Selects the correct CCAssignFn for a given calling
/// convention.
CCAssignFn *X86FastISel::CCAssignFnForCall(CallingConv::ID CC,
                                           bool isTaillCall) {
  if (Subtarget->is64Bit()) {
    if (Subtarget->isTargetWin64())
      return CC_X86_Win64_C;
    else
      return CC_X86_64_C;
  }

  if (CC == CallingConv::X86_FastCall)
    return CC_X86_32_FastCall;
  else if (CC == CallingConv::Fast)
    return CC_X86_32_FastCC;
  else
    return CC_X86_32_C;
}

/// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
/// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
/// Return true and the result register by reference if it is possible.
bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
                                  unsigned &ResultReg) {
  // Get opcode and regclass of the output for the given load instruction.
  unsigned Opc = 0;
  const TargetRegisterClass *RC = NULL;
  switch (VT.getSimpleVT().SimpleTy) {
  default: return false;
  case MVT::i1:
  case MVT::i8:
    Opc = X86::MOV8rm;
    RC  = X86::GR8RegisterClass;
    break;
  case MVT::i16:
    Opc = X86::MOV16rm;
    RC  = X86::GR16RegisterClass;
    break;
  case MVT::i32:
    Opc = X86::MOV32rm;
    RC  = X86::GR32RegisterClass;
    break;
  case MVT::i64:
    // Must be in x86-64 mode.
    Opc = X86::MOV64rm;
    RC  = X86::GR64RegisterClass;
    break;
  case MVT::f32:
    if (Subtarget->hasSSE1()) {
      Opc = X86::MOVSSrm;
      RC  = X86::FR32RegisterClass;
    } else {
      Opc = X86::LD_Fp32m;
      RC  = X86::RFP32RegisterClass;
    }
    break;
  case MVT::f64:
    if (Subtarget->hasSSE2()) {
      Opc = X86::MOVSDrm;
      RC  = X86::FR64RegisterClass;
    } else {
      Opc = X86::LD_Fp64m;
      RC  = X86::RFP64RegisterClass;
    }
    break;
  case MVT::f80:
    // No f80 support yet.
    return false;
  }

  ResultReg = createResultReg(RC);
  addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
  return true;
}

/// X86FastEmitStore - Emit a machine instruction to store a value Val of
/// type VT. The address is either pre-computed, consisted of a base ptr, Ptr
/// and a displacement offset, or a GlobalAddress,
/// i.e. V. Return true if it is possible.
bool
X86FastISel::X86FastEmitStore(EVT VT, unsigned Val,
                              const X86AddressMode &AM) {
  // Get opcode and regclass of the output for the given store instruction.
  unsigned Opc = 0;
  switch (VT.getSimpleVT().SimpleTy) {
  case MVT::f80: // No f80 support yet.
  default: return false;
  case MVT::i1: {
    // Mask out all but lowest bit.
    unsigned AndResult = createResultReg(X86::GR8RegisterClass);
    BuildMI(MBB, DL,
            TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
    Val = AndResult;
  }
  // FALLTHROUGH, handling i1 as i8.
  case MVT::i8:  Opc = X86::MOV8mr;  break;
  case MVT::i16: Opc = X86::MOV16mr; break;
  case MVT::i32: Opc = X86::MOV32mr; break;
  case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
  case MVT::f32:
    Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
    break;
  case MVT::f64:
    Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
    break;
  }

  addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM).addReg(Val);
  return true;
}

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

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

    if (Opc) {
      addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM)
                             .addImm(Signed ? CI->getSExtValue() :
                                              CI->getZExtValue());
      return true;
    }
  }

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

  return X86FastEmitStore(VT, ValReg, AM);
}

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

  if (RR != 0) {
    ResultReg = RR;
    return true;
  } else
    return false;
}

/// X86SelectAddress - Attempt to fill in an address from the given value.
///
bool X86FastISel::X86SelectAddress(Value *V, X86AddressMode &AM) {
  User *U = NULL;
  unsigned Opcode = Instruction::UserOp1;
  if (Instruction *I = dyn_cast<Instruction>(V)) {
    Opcode = I->getOpcode();
    U = I;
  } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
    Opcode = C->getOpcode();
    U = C;
  }

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

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

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

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

  case Instruction::Add: {
    // Adds of constants are common and easy enough.
    if (ConstantInt *CI = dyn_cast<ConstantInt>(U->getOperand(1))) {
      uint64_t Disp = (int32_t)AM.Disp + (uint64_t)CI->getSExtValue();
      // They have to fit in the 32-bit signed displacement field though.
      if (isInt32(Disp)) {
        AM.Disp = (uint32_t)Disp;
        return X86SelectAddress(U->getOperand(0), AM);
      }
    }
    break;
  }

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

    // Pattern-match simple GEPs.
    uint64_t Disp = (int32_t)AM.Disp;
    unsigned IndexReg = AM.IndexReg;
    unsigned Scale = AM.Scale;
    gep_type_iterator GTI = gep_type_begin(U);
    // Iterate through the indices, folding what we can. Constants can be
    // folded, and one dynamic index can be handled, if the scale is supported.
    for (User::op_iterator i = U->op_begin() + 1, e = U->op_end();
         i != e; ++i, ++GTI) {
      Value *Op = *i;
      if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
        const StructLayout *SL = TD.getStructLayout(STy);
        unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
        Disp += SL->getElementOffset(Idx);
      } else {
        uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
        if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
          // Constant-offset addressing.
          Disp += CI->getSExtValue() * S;
        } else if (IndexReg == 0 &&
                   (!AM.GV || !Subtarget->isPICStyleRIPRel()) &&
                   (S == 1 || S == 2 || S == 4 || S == 8)) {
          // Scaled-index addressing.
          Scale = S;
          IndexReg = getRegForGEPIndex(Op);
          if (IndexReg == 0)
            return false;
        } else
          // Unsupported.
          goto unsupported_gep;
      }
    }
    // Check for displacement overflow.
    if (!isInt32(Disp))
      break;
    // Ok, the GEP indices were covered by constant-offset and scaled-index
    // addressing. Update the address state and move on to examining the base.
    AM.IndexReg = IndexReg;
    AM.Scale = Scale;
    AM.Disp = (uint32_t)Disp;
    if (X86SelectAddress(U->getOperand(0), AM))
      return true;

    // If we couldn't merge the sub value into this addr mode, revert back to
    // our address and just match the value instead of completely failing.
    AM = SavedAM;
    break;
  unsupported_gep:
    // Ok, the GEP indices weren't all covered.
    break;
  }
  }

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

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

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

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

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

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

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

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

      if (TLI.getPointerTy() == MVT::i64) {
        Opc = X86::MOV64rm;
        RC  = X86::GR64RegisterClass;

        if (Subtarget->isPICStyleRIPRel())
          StubAM.Base.Reg = X86::RIP;
      } else {
        Opc = X86::MOV32rm;
        RC  = X86::GR32RegisterClass;
      }

      LoadReg = createResultReg(RC);
      addFullAddress(BuildMI(MBB, DL, TII.get(Opc), LoadReg), StubAM);

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

    // Now construct the final address. Note that the Disp, Scale,
    // and Index values may already be set here.
    AM.Base.Reg = LoadReg;
    AM.GV = 0;
    return true;
  }

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

  return false;
}

/// X86SelectCallAddress - Attempt to fill in an address from the given value.
///
bool X86FastISel::X86SelectCallAddress(Value *V, X86AddressMode &AM) {
  User *U = NULL;
  unsigned Opcode = Instruction::UserOp1;
  if (Instruction *I = dyn_cast<Instruction>(V)) {
    Opcode = I->getOpcode();
    U = I;
  } else if (ConstantExpr *C = dyn_cast<ConstantExpr>(V)) {
    Opcode = C->getOpcode();
    U = C;
  }

  switch (Opcode) {
  default: break;
  case Instruction::BitCast:
    // Look past bitcasts.
    return X86SelectCallAddress(U->getOperand(0), AM);

  case Instruction::IntToPtr:
    // Look past no-op inttoptrs.
    if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
      return X86SelectCallAddress(U->getOperand(0), AM);
    break;

  case Instruction::PtrToInt:
    // Look past no-op ptrtoints.
    if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
      return X86SelectCallAddress(U->getOperand(0), AM);
    break;
  }

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

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

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

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

    // No ABI requires an extra load for anything other than DLLImport, which
    // we rejected above. Return a direct reference to the global.
    if (Subtarget->isPICStyleRIPRel()) {
      // Use rip-relative addressing if we can.  Above we verified that the
      // base and index registers are unused.
      assert(AM.Base.Reg == 0 && AM.IndexReg == 0);
      AM.Base.Reg = X86::RIP;
    } else if (Subtarget->isPICStyleStubPIC()) {
      AM.GVOpFlags = X86II::MO_PIC_BASE_OFFSET;
    } else if (Subtarget->isPICStyleGOT()) {
      AM.GVOpFlags = X86II::MO_GOTOFF;
    }

    return true;
  }

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

  return false;
}


/// X86SelectStore - Select and emit code to implement store instructions.
bool X86FastISel::X86SelectStore(Instruction* I) {
  EVT VT;
  if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
    return false;

  X86AddressMode AM;
  if (!X86SelectAddress(I->getOperand(1), AM))
    return false;

  return X86FastEmitStore(VT, I->getOperand(0), AM);
}

/// X86SelectLoad - Select and emit code to implement load instructions.
///
bool X86FastISel::X86SelectLoad(Instruction *I)  {
  EVT VT;
  if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
    return false;

  X86AddressMode AM;
  if (!X86SelectAddress(I->getOperand(0), AM))
    return false;

  unsigned ResultReg = 0;
  if (X86FastEmitLoad(VT, AM, ResultReg)) {
    UpdateValueMap(I, ResultReg);
    return true;
  }
  return false;
}

static unsigned X86ChooseCmpOpcode(EVT VT) {
  switch (VT.getSimpleVT().SimpleTy) {
  default:       return 0;
  case MVT::i8:  return X86::CMP8rr;
  case MVT::i16: return X86::CMP16rr;
  case MVT::i32: return X86::CMP32rr;
  case MVT::i64: return X86::CMP64rr;
  case MVT::f32: return X86::UCOMISSrr;
  case MVT::f64: return X86::UCOMISDrr;
  }
}

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

bool X86FastISel::X86FastEmitCompare(Value *Op0, Value *Op1, EVT VT) {
  unsigned Op0Reg = getRegForValue(Op0);
  if (Op0Reg == 0) return false;

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

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

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

  unsigned Op1Reg = getRegForValue(Op1);
  if (Op1Reg == 0) return false;
  BuildMI(MBB, DL, TII.get(CompareOpc)).addReg(Op0Reg).addReg(Op1Reg);

  return true;
}

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

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

  unsigned ResultReg = createResultReg(&X86::GR8RegClass);
  unsigned SetCCOpc;
  bool SwapArgs;  // false -> compare Op0, Op1.  true -> compare Op1, Op0.
  switch (CI->getPredicate()) {
  case CmpInst::FCMP_OEQ: {
    if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
      return false;

    unsigned EReg = createResultReg(&X86::GR8RegClass);
    unsigned NPReg = createResultReg(&X86::GR8RegClass);
    BuildMI(MBB, DL, TII.get(X86::SETEr), EReg);
    BuildMI(MBB, DL, TII.get(X86::SETNPr), NPReg);
    BuildMI(MBB, DL,
            TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
    UpdateValueMap(I, ResultReg);
    return true;
  }
  case CmpInst::FCMP_UNE: {
    if (!X86FastEmitCompare(CI->getOperand(0), CI->getOperand(1), VT))
      return false;

    unsigned NEReg = createResultReg(&X86::GR8RegClass);
    unsigned PReg = createResultReg(&X86::GR8RegClass);
    BuildMI(MBB, DL, TII.get(X86::SETNEr), NEReg);
    BuildMI(MBB, DL, TII.get(X86::SETPr), PReg);
    BuildMI(MBB, DL, TII.get(X86::OR8rr), ResultReg).addReg(PReg).addReg(NEReg);
    UpdateValueMap(I, ResultReg);
    return true;
  }
  case CmpInst::FCMP_OGT: SwapArgs = false; SetCCOpc = X86::SETAr;  break;
  case CmpInst::FCMP_OGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
  case CmpInst::FCMP_OLT: SwapArgs = true;  SetCCOpc = X86::SETAr;  break;
  case CmpInst::FCMP_OLE: SwapArgs = true;  SetCCOpc = X86::SETAEr; break;
  case CmpInst::FCMP_ONE: SwapArgs = false; SetCCOpc = X86::SETNEr; break;
  case CmpInst::FCMP_ORD: SwapArgs = false; SetCCOpc = X86::SETNPr; break;
  case CmpInst::FCMP_UNO: SwapArgs = false; SetCCOpc = X86::SETPr;  break;
  case CmpInst::FCMP_UEQ: SwapArgs = false; SetCCOpc = X86::SETEr;  break;
  case CmpInst::FCMP_UGT: SwapArgs = true;  SetCCOpc = X86::SETBr;  break;
  case CmpInst::FCMP_UGE: SwapArgs = true;  SetCCOpc = X86::SETBEr; break;
  case CmpInst::FCMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr;  break;
  case CmpInst::FCMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;

  case CmpInst::ICMP_EQ:  SwapArgs = false; SetCCOpc = X86::SETEr;  break;
  case CmpInst::ICMP_NE:  SwapArgs = false; SetCCOpc = X86::SETNEr; break;
  case CmpInst::ICMP_UGT: SwapArgs = false; SetCCOpc = X86::SETAr;  break;
  case CmpInst::ICMP_UGE: SwapArgs = false; SetCCOpc = X86::SETAEr; break;
  case CmpInst::ICMP_ULT: SwapArgs = false; SetCCOpc = X86::SETBr;  break;
  case CmpInst::ICMP_ULE: SwapArgs = false; SetCCOpc = X86::SETBEr; break;
  case CmpInst::ICMP_SGT: SwapArgs = false; SetCCOpc = X86::SETGr;  break;
  case CmpInst::ICMP_SGE: SwapArgs = false; SetCCOpc = X86::SETGEr; break;
  case CmpInst::ICMP_SLT: SwapArgs = false; SetCCOpc = X86::SETLr;  break;
  case CmpInst::ICMP_SLE: SwapArgs = false; SetCCOpc = X86::SETLEr; break;
  default:
    return false;
  }

  Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
  if (SwapArgs)
    std::swap(Op0, Op1);

  // Emit a compare of Op0/Op1.
  if (!X86FastEmitCompare(Op0, Op1, VT))
    return false;

  BuildMI(MBB, DL, TII.get(SetCCOpc), ResultReg);
  UpdateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectZExt(Instruction *I) {
  // Handle zero-extension from i1 to i8, which is common.
  if (I->getType()->isIntegerTy(8) &&
      I->getOperand(0)->getType()->isIntegerTy(1)) {
    unsigned ResultReg = getRegForValue(I->getOperand(0));
    if (ResultReg == 0) return false;
    // Set the high bits to zero.
    ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg);
    if (ResultReg == 0) return false;
    UpdateValueMap(I, ResultReg);
    return true;
  }

  return false;
}


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

  // Fold the common case of a conditional branch with a comparison.
  if (CmpInst *CI = dyn_cast<CmpInst>(BI->getCondition())) {
    if (CI->hasOneUse()) {
      EVT VT = TLI.getValueType(CI->getOperand(0)->getType());

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

      bool SwapArgs;  // false -> compare Op0, Op1.  true -> compare Op1, Op0.
      unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"

      switch (Predicate) {
      case CmpInst::FCMP_OEQ:
        std::swap(TrueMBB, FalseMBB);
        Predicate = CmpInst::FCMP_UNE;
        // FALL THROUGH
      case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
      case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA_4;  break;
      case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
      case CmpInst::FCMP_OLT: SwapArgs = true;  BranchOpc = X86::JA_4;  break;
      case CmpInst::FCMP_OLE: SwapArgs = true;  BranchOpc = X86::JAE_4; break;
      case CmpInst::FCMP_ONE: SwapArgs = false; BranchOpc = X86::JNE_4; break;
      case CmpInst::FCMP_ORD: SwapArgs = false; BranchOpc = X86::JNP_4; break;
      case CmpInst::FCMP_UNO: SwapArgs = false; BranchOpc = X86::JP_4;  break;
      case CmpInst::FCMP_UEQ: SwapArgs = false; BranchOpc = X86::JE_4;  break;
      case CmpInst::FCMP_UGT: SwapArgs = true;  BranchOpc = X86::JB_4;  break;
      case CmpInst::FCMP_UGE: SwapArgs = true;  BranchOpc = X86::JBE_4; break;
      case CmpInst::FCMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4;  break;
      case CmpInst::FCMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;

      case CmpInst::ICMP_EQ:  SwapArgs = false; BranchOpc = X86::JE_4;  break;
      case CmpInst::ICMP_NE:  SwapArgs = false; BranchOpc = X86::JNE_4; break;
      case CmpInst::ICMP_UGT: SwapArgs = false; BranchOpc = X86::JA_4;  break;
      case CmpInst::ICMP_UGE: SwapArgs = false; BranchOpc = X86::JAE_4; break;
      case CmpInst::ICMP_ULT: SwapArgs = false; BranchOpc = X86::JB_4;  break;
      case CmpInst::ICMP_ULE: SwapArgs = false; BranchOpc = X86::JBE_4; break;
      case CmpInst::ICMP_SGT: SwapArgs = false; BranchOpc = X86::JG_4;  break;
      case CmpInst::ICMP_SGE: SwapArgs = false; BranchOpc = X86::JGE_4; break;
      case CmpInst::ICMP_SLT: SwapArgs = false; BranchOpc = X86::JL_4;  break;
      case CmpInst::ICMP_SLE: SwapArgs = false; BranchOpc = X86::JLE_4; break;
      default:
        return false;
      }

      Value *Op0 = CI->getOperand(0), *Op1 = CI->getOperand(1);
      if (SwapArgs)
        std::swap(Op0, Op1);

      // Emit a compare of the LHS and RHS, setting the flags.
      if (!X86FastEmitCompare(Op0, Op1, VT))
        return false;

      BuildMI(MBB, DL, TII.get(BranchOpc)).addMBB(TrueMBB);

      if (Predicate == CmpInst::FCMP_UNE) {
        // X86 requires a second branch to handle UNE (and OEQ,
        // which is mapped to UNE above).
        BuildMI(MBB, DL, TII.get(X86::JP_4)).addMBB(TrueMBB);
      }

      FastEmitBranch(FalseMBB);
      MBB->addSuccessor(TrueMBB);
      return true;
    }
  } else if (ExtractValueInst *EI =
             dyn_cast<ExtractValueInst>(BI->getCondition())) {
    // Check to see if the branch instruction is from an "arithmetic with
    // overflow" intrinsic. The main way these intrinsics are used is:
    //
    //   %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
    //   %sum = extractvalue { i32, i1 } %t, 0
    //   %obit = extractvalue { i32, i1 } %t, 1
    //   br i1 %obit, label %overflow, label %normal
    //
    // The %sum and %obit are converted in an ADD and a SETO/SETB before
    // reaching the branch. Therefore, we search backwards through the MBB
    // looking for the SETO/SETB instruction. If an instruction modifies the
    // EFLAGS register before we reach the SETO/SETB instruction, then we can't
    // convert the branch into a JO/JB instruction.
    if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
      if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
          CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
        const MachineInstr *SetMI = 0;
        unsigned Reg = lookUpRegForValue(EI);

        for (MachineBasicBlock::const_reverse_iterator
               RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
          const MachineInstr &MI = *RI;

          if (MI.modifiesRegister(Reg)) {
            unsigned Src, Dst, SrcSR, DstSR;

            if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
              Reg = Src;
              continue;
            }

            SetMI = &MI;
            break;
          }

          const TargetInstrDesc &TID = MI.getDesc();
          if (TID.hasUnmodeledSideEffects() ||
              TID.hasImplicitDefOfPhysReg(X86::EFLAGS))
            break;
        }

        if (SetMI) {
          unsigned OpCode = SetMI->getOpcode();

          if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
            BuildMI(MBB, DL, TII.get(OpCode == X86::SETOr ?
                                        X86::JO_4 : X86::JB_4))
              .addMBB(TrueMBB);
            FastEmitBranch(FalseMBB);
            MBB->addSuccessor(TrueMBB);
            return true;
          }
        }
      }
    }
  }

  // Otherwise do a clumsy setcc and re-test it.
  unsigned OpReg = getRegForValue(BI->getCondition());
  if (OpReg == 0) return false;

  BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(OpReg).addReg(OpReg);
  BuildMI(MBB, DL, TII.get(X86::JNE_4)).addMBB(TrueMBB);
  FastEmitBranch(FalseMBB);
  MBB->addSuccessor(TrueMBB);
  return true;
}

bool X86FastISel::X86SelectShift(Instruction *I) {
  unsigned CReg = 0, OpReg = 0, OpImm = 0;
  const TargetRegisterClass *RC = NULL;
  if (I->getType()->isIntegerTy(8)) {
    CReg = X86::CL;
    RC = &X86::GR8RegClass;
    switch (I->getOpcode()) {
    case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
    case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
    case Instruction::Shl:  OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
    default: return false;
    }
  } else if (I->getType()->isIntegerTy(16)) {
    CReg = X86::CX;
    RC = &X86::GR16RegClass;
    switch (I->getOpcode()) {
    case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
    case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
    case Instruction::Shl:  OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
    default: return false;
    }
  } else if (I->getType()->isIntegerTy(32)) {
    CReg = X86::ECX;
    RC = &X86::GR32RegClass;
    switch (I->getOpcode()) {
    case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
    case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
    case Instruction::Shl:  OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
    default: return false;
    }
  } else if (I->getType()->isIntegerTy(64)) {
    CReg = X86::RCX;
    RC = &X86::GR64RegClass;
    switch (I->getOpcode()) {
    case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
    case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
    case Instruction::Shl:  OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
    default: return false;
    }
  } else {
    return false;
  }

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

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

  // Fold immediate in shl(x,3).
  if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
    unsigned ResultReg = createResultReg(RC);
    BuildMI(MBB, DL, TII.get(OpImm),
            ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
    UpdateValueMap(I, ResultReg);
    return true;
  }

  unsigned Op1Reg = getRegForValue(I->getOperand(1));
  if (Op1Reg == 0) return false;
  TII.copyRegToReg(*MBB, MBB->end(), CReg, Op1Reg, RC, RC);

  // The shift instruction uses X86::CL. If we defined a super-register
  // of X86::CL, emit an EXTRACT_SUBREG to precisely describe what
  // we're doing here.
  if (CReg != X86::CL)
    BuildMI(MBB, DL, TII.get(TargetOpcode::EXTRACT_SUBREG), X86::CL)
      .addReg(CReg).addImm(X86::SUBREG_8BIT);

  unsigned ResultReg = createResultReg(RC);
  BuildMI(MBB, DL, TII.get(OpReg), ResultReg).addReg(Op0Reg);
  UpdateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectSelect(Instruction *I) {
  EVT VT = TLI.getValueType(I->getType(), /*HandleUnknown=*/true);
  if (VT == MVT::Other || !isTypeLegal(I->getType(), VT))
    return false;

  unsigned Opc = 0;
  const TargetRegisterClass *RC = NULL;
  if (VT.getSimpleVT() == MVT::i16) {
    Opc = X86::CMOVE16rr;
    RC = &X86::GR16RegClass;
  } else if (VT.getSimpleVT() == MVT::i32) {
    Opc = X86::CMOVE32rr;
    RC = &X86::GR32RegClass;
  } else if (VT.getSimpleVT() == MVT::i64) {
    Opc = X86::CMOVE64rr;
    RC = &X86::GR64RegClass;
  } else {
    return false;
  }

  unsigned Op0Reg = getRegForValue(I->getOperand(0));
  if (Op0Reg == 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;

  BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(Op0Reg).addReg(Op0Reg);
  unsigned ResultReg = createResultReg(RC);
  BuildMI(MBB, DL, TII.get(Opc), ResultReg).addReg(Op1Reg).addReg(Op2Reg);
  UpdateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectFPExt(Instruction *I) {
  // fpext from float to double.
  if (Subtarget->hasSSE2() &&
      I->getType()->isDoubleTy()) {
    Value *V = I->getOperand(0);
    if (V->getType()->isFloatTy()) {
      unsigned OpReg = getRegForValue(V);
      if (OpReg == 0) return false;
      unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
      BuildMI(MBB, DL, TII.get(X86::CVTSS2SDrr), ResultReg).addReg(OpReg);
      UpdateValueMap(I, ResultReg);
      return true;
    }
  }

  return false;
}

bool X86FastISel::X86SelectFPTrunc(Instruction *I) {
  if (Subtarget->hasSSE2()) {
    if (I->getType()->isFloatTy()) {
      Value *V = I->getOperand(0);
      if (V->getType()->isDoubleTy()) {
        unsigned OpReg = getRegForValue(V);
        if (OpReg == 0) return false;
        unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
        BuildMI(MBB, DL, TII.get(X86::CVTSD2SSrr), ResultReg).addReg(OpReg);
        UpdateValueMap(I, ResultReg);
        return true;
      }
    }
  }

  return false;
}

bool X86FastISel::X86SelectTrunc(Instruction *I) {
  if (Subtarget->is64Bit())
    // All other cases should be handled by the tblgen generated code.
    return false;
  EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
  EVT DstVT = TLI.getValueType(I->getType());

  // This code only handles truncation to byte right now.
  if (DstVT != MVT::i8 && DstVT != MVT::i1)
    // All other cases should be handled by the tblgen generated code.
    return false;
  if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
    // All other cases should be handled by the tblgen generated code.
    return false;

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

  // First issue a copy to GR16_ABCD or GR32_ABCD.
  unsigned CopyOpc = (SrcVT == MVT::i16) ? X86::MOV16rr : X86::MOV32rr;
  const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
    ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
  unsigned CopyReg = createResultReg(CopyRC);
  BuildMI(MBB, DL, TII.get(CopyOpc), CopyReg).addReg(InputReg);

  // Then issue an extract_subreg.
  unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
                                                  CopyReg, X86::SUBREG_8BIT);
  if (!ResultReg)
    return false;

  UpdateValueMap(I, ResultReg);
  return true;
}

bool X86FastISel::X86SelectExtractValue(Instruction *I) {
  ExtractValueInst *EI = cast<ExtractValueInst>(I);
  Value *Agg = EI->getAggregateOperand();

  if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
    switch (CI->getIntrinsicID()) {
    default: break;
    case Intrinsic::sadd_with_overflow:
    case Intrinsic::uadd_with_overflow:
      // Cheat a little. We know that the registers for "add" and "seto" are
      // allocated sequentially. However, we only keep track of the register
      // for "add" in the value map. Use extractvalue's index to get the
      // correct register for "seto".
      UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
      return true;
    }
  }

  return false;
}

bool X86FastISel::X86VisitIntrinsicCall(IntrinsicInst &I) {
  // FIXME: Handle more intrinsics.
  switch (I.getIntrinsicID()) {
  default: return false;
  case Intrinsic::dbg_declare: {
    DbgDeclareInst *DI = cast<DbgDeclareInst>(&I);
    X86AddressMode AM;
    assert(DI->getAddress() && "Null address should be checked earlier!");
    if (!X86SelectAddress(DI->getAddress(), AM))
      return false;
    const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
    // FIXME may need to add RegState::Debug to any registers produced,
    // although ESP/EBP should be the only ones at the moment.
    addFullAddress(BuildMI(MBB, DL, II), AM).addImm(0).
                                        addMetadata(DI->getVariable());
    return true;
  }
  case Intrinsic::trap: {
    BuildMI(MBB, DL, TII.get(X86::TRAP));
    return true;
  }
  case Intrinsic::sadd_with_overflow:
  case Intrinsic::uadd_with_overflow: {
    // Replace "add with overflow" intrinsics with an "add" instruction followed
    // by a seto/setc instruction. Later on, when the "extractvalue"
    // instructions are encountered, we use the fact that two registers were
    // created sequentially to get the correct registers for the "sum" and the
    // "overflow bit".
    const Function *Callee = I.getCalledFunction();
    const Type *RetTy =
      cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));

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

    Value *Op1 = I.getOperand(1);
    Value *Op2 = I.getOperand(2);
    unsigned Reg1 = getRegForValue(Op1);
    unsigned Reg2 = getRegForValue(Op2);

    if (Reg1 == 0 || Reg2 == 0)
      // FIXME: Handle values *not* in registers.
      return false;

    unsigned OpC = 0;
    if (VT == MVT::i32)
      OpC = X86::ADD32rr;
    else if (VT == MVT::i64)
      OpC = X86::ADD64rr;
    else
      return false;

    unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
    BuildMI(MBB, DL, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
    unsigned DestReg1 = UpdateValueMap(&I, ResultReg);

    // If the add with overflow is an intra-block value then we just want to
    // create temporaries for it like normal.  If it is a cross-block value then
    // UpdateValueMap will return the cross-block register used.  Since we
    // *really* want the value to be live in the register pair known by
    // UpdateValueMap, we have to use DestReg1+1 as the destination register in
    // the cross block case.  In the non-cross-block case, we should just make
    // another register for the value.
    if (DestReg1 != ResultReg)
      ResultReg = DestReg1+1;
    else
      ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));

    unsigned Opc = X86::SETBr;
    if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
      Opc = X86::SETOr;
    BuildMI(MBB, DL, TII.get(Opc), ResultReg);
    return true;
  }
  }
}

bool X86FastISel::X86SelectCall(Instruction *I) {
  CallInst *CI = cast<CallInst>(I);
  Value *Callee = I->getOperand(0);

  // Can't handle inline asm yet.
  if (isa<InlineAsm>(Callee))
    return false;

  // Handle intrinsic calls.
  if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
    return X86VisitIntrinsicCall(*II);

  // Handle only C and fastcc calling conventions for now.
  CallSite CS(CI);
  CallingConv::ID CC = CS.getCallingConv();
  if (CC != CallingConv::C &&
      CC != CallingConv::Fast &&
      CC != CallingConv::X86_FastCall)
    return false;

  // fastcc with -tailcallopt is intended to provide a guaranteed
  // tail call optimization. Fastisel doesn't know how to do that.
  if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
    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 = CS.getType();
  EVT RetVT;
  if (RetTy->isVoidTy())
    RetVT = MVT::isVoid;
  else if (!isTypeLegal(RetTy, RetVT, true))
    return false;

  // Materialize callee address in a register. FIXME: GV address can be
  // handled with a CALLpcrel32 instead.
  X86AddressMode CalleeAM;
  if (!X86SelectCallAddress(Callee, CalleeAM))
    return false;
  unsigned CalleeOp = 0;
  GlobalValue *GV = 0;
  if (CalleeAM.GV != 0) {
    GV = CalleeAM.GV;
  } else if (CalleeAM.Base.Reg != 0) {
    CalleeOp = CalleeAM.Base.Reg;
  } else
    return false;

  // Allow calls which produce i1 results.
  bool AndToI1 = false;
  if (RetVT == MVT::i1) {
    RetVT = MVT::i8;
    AndToI1 = true;
  }

  // Deal with call operands first.
  SmallVector<Value*, 8> ArgVals;
  SmallVector<unsigned, 8> Args;
  SmallVector<EVT, 8> ArgVTs;
  SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
  Args.reserve(CS.arg_size());
  ArgVals.reserve(CS.arg_size());
  ArgVTs.reserve(CS.arg_size());
  ArgFlags.reserve(CS.arg_size());
  for (CallSite::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(Arg);
    ArgVals.push_back(*i);
    ArgVTs.push_back(ArgVT);
    ArgFlags.push_back(Flags);
  }

  // Analyze operands of the call, assigning locations to each operand.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
  CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CCAssignFnForCall(CC));

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

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

  // Process argument: walk the register/memloc assignments, inserting
  // copies / loads.
  SmallVector<unsigned, 4> RegArgs;
  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
    CCValAssign &VA = ArgLocs[i];
    unsigned Arg = Args[VA.getValNo()];
    EVT ArgVT = ArgVTs[VA.getValNo()];

    // Promote the value if needed.
    switch (VA.getLocInfo()) {
    default: llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full: break;
    case CCValAssign::SExt: {
      bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
                                       Arg, ArgVT, Arg);
      assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
      Emitted = true;
      ArgVT = VA.getLocVT();
      break;
    }
    case CCValAssign::ZExt: {
      bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
                                       Arg, ArgVT, Arg);
      assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
      Emitted = true;
      ArgVT = VA.getLocVT();
      break;
    }
    case CCValAssign::AExt: {
      bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
                                       Arg, ArgVT, Arg);
      if (!Emitted)
        Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
                                    Arg, ArgVT, Arg);
      if (!Emitted)
        Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
                                    Arg, ArgVT, Arg);

      assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
      ArgVT = VA.getLocVT();
      break;
    }
    case CCValAssign::BCvt: {
      unsigned BC = FastEmit_r(ArgVT.getSimpleVT(), VA.getLocVT().getSimpleVT(),
                               ISD::BIT_CONVERT, Arg);
      assert(BC != 0 && "Failed to emit a bitcast!");
      Arg = BC;
      ArgVT = VA.getLocVT();
      break;
    }
    }

    if (VA.isRegLoc()) {
      TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
      bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
                                      Arg, RC, RC);
      assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
      Emitted = true;
      RegArgs.push_back(VA.getLocReg());
    } else {
      unsigned LocMemOffset = VA.getLocMemOffset();
      X86AddressMode AM;
      AM.Base.Reg = StackPtr;
      AM.Disp = LocMemOffset;
      Value *ArgVal = ArgVals[VA.getValNo()];

      // If this is a really simple value, emit this with the Value* version of
      // X86FastEmitStore.  If it isn't simple, we don't want to do this, as it
      // can cause us to reevaluate the argument.
      if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
        X86FastEmitStore(ArgVT, ArgVal, AM);
      else
        X86FastEmitStore(ArgVT, Arg, AM);
    }
  }

  // ELF / PIC requires GOT in the EBX register before function calls via PLT
  // GOT pointer.
  if (Subtarget->isPICStyleGOT()) {
    TargetRegisterClass *RC = X86::GR32RegisterClass;
    unsigned Base = getInstrInfo()->getGlobalBaseReg(&MF);
    bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC);
    assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
    Emitted = true;
  }

  // Issue the call.
  MachineInstrBuilder MIB;
  if (CalleeOp) {
    // Register-indirect call.
    unsigned CallOpc = Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r;
    MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addReg(CalleeOp);

  } else {
    // Direct call.
    assert(GV && "Not a direct call");
    unsigned CallOpc =
      Subtarget->is64Bit() ? X86::CALL64pcrel32 : X86::CALLpcrel32;

    // See if we need any target-specific flags on the GV operand.
    unsigned char OpFlags = 0;

    // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
    // external symbols most go through the PLT in PIC mode.  If the symbol
    // has hidden or protected visibility, or if it is static or local, then
    // we don't need to use the PLT - we can directly call it.
    if (Subtarget->isTargetELF() &&
        TM.getRelocationModel() == Reloc::PIC_ &&
        GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
      OpFlags = X86II::MO_PLT;
    } else if (Subtarget->isPICStyleStubAny() &&
               (GV->isDeclaration() || GV->isWeakForLinker()) &&
               Subtarget->getDarwinVers() < 9) {
      // PC-relative references to external symbols should go through $stub,
      // unless we're building with the leopard linker or later, which
      // automatically synthesizes these stubs.
      OpFlags = X86II::MO_DARWIN_STUB;
    }


    MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addGlobalAddress(GV, 0, OpFlags);
  }

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

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

  // Issue CALLSEQ_END
  unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
  BuildMI(MBB, DL, TII.get(AdjStackUp)).addImm(NumBytes).addImm(0);

  // Now handle call return value (if any).
  if (RetVT.getSimpleVT().SimpleTy != MVT::isVoid) {
    SmallVector<CCValAssign, 16> RVLocs;
    CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
    CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);

    // Copy all of the result registers out of their specified physreg.
    assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
    EVT CopyVT = RVLocs[0].getValVT();
    TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
    TargetRegisterClass *SrcRC = DstRC;

    // If this is a call to a function that returns an fp value on the x87 fp
    // stack, but where we prefer to use the value in xmm registers, copy it
    // out as F80 and use a truncate to move it from fp stack reg to xmm reg.
    if ((RVLocs[0].getLocReg() == X86::ST0 ||
         RVLocs[0].getLocReg() == X86::ST1) &&
        isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
      CopyVT = MVT::f80;
      SrcRC = X86::RSTRegisterClass;
      DstRC = X86::RFP80RegisterClass;
    }

    unsigned ResultReg = createResultReg(DstRC);
    bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
                                    RVLocs[0].getLocReg(), DstRC, SrcRC);
    assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
    Emitted = true;
    if (CopyVT != RVLocs[0].getValVT()) {
      // Round the F80 the right size, which also moves to the appropriate xmm
      // register. This is accomplished by storing the F80 value in memory and
      // then loading it back. Ewww...
      EVT ResVT = RVLocs[0].getValVT();
      unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
      unsigned MemSize = ResVT.getSizeInBits()/8;
      int FI = MFI.CreateStackObject(MemSize, MemSize, false);
      addFrameReference(BuildMI(MBB, DL, TII.get(Opc)), FI).addReg(ResultReg);
      DstRC = ResVT == MVT::f32
        ? X86::FR32RegisterClass : X86::FR64RegisterClass;
      Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
      ResultReg = createResultReg(DstRC);
      addFrameReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), FI);
    }

    if (AndToI1) {
      // Mask out all but lowest bit for some call which produces an i1.
      unsigned AndResult = createResultReg(X86::GR8RegisterClass);
      BuildMI(MBB, DL,
              TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
      ResultReg = AndResult;
    }

    UpdateValueMap(I, ResultReg);
  }

  return true;
}


bool
X86FastISel::TargetSelectInstruction(Instruction *I)  {
  switch (I->getOpcode()) {
  default: break;
  case Instruction::Load:
    return X86SelectLoad(I);
  case Instruction::Store:
    return X86SelectStore(I);
  case Instruction::ICmp:
  case Instruction::FCmp:
    return X86SelectCmp(I);
  case Instruction::ZExt:
    return X86SelectZExt(I);
  case Instruction::Br:
    return X86SelectBranch(I);
  case Instruction::Call:
    return X86SelectCall(I);
  case Instruction::LShr:
  case Instruction::AShr:
  case Instruction::Shl:
    return X86SelectShift(I);
  case Instruction::Select:
    return X86SelectSelect(I);
  case Instruction::Trunc:
    return X86SelectTrunc(I);
  case Instruction::FPExt:
    return X86SelectFPExt(I);
  case Instruction::FPTrunc:
    return X86SelectFPTrunc(I);
  case Instruction::ExtractValue:
    return X86SelectExtractValue(I);
  case Instruction::IntToPtr: // Deliberate fall-through.
  case Instruction::PtrToInt: {
    EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
    EVT DstVT = TLI.getValueType(I->getType());
    if (DstVT.bitsGT(SrcVT))
      return X86SelectZExt(I);
    if (DstVT.bitsLT(SrcVT))
      return X86SelectTrunc(I);
    unsigned Reg = getRegForValue(I->getOperand(0));
    if (Reg == 0) return false;
    UpdateValueMap(I, Reg);
    return true;
  }
  }

  return false;
}

unsigned X86FastISel::TargetMaterializeConstant(Constant *C) {
  EVT VT;
  if (!isTypeLegal(C->getType(), VT))
    return false;

  // Get opcode and regclass of the output for the given load instruction.
  unsigned Opc = 0;
  const TargetRegisterClass *RC = NULL;
  switch (VT.getSimpleVT().SimpleTy) {
  default: return false;
  case MVT::i8:
    Opc = X86::MOV8rm;
    RC  = X86::GR8RegisterClass;
    break;
  case MVT::i16:
    Opc = X86::MOV16rm;
    RC  = X86::GR16RegisterClass;
    break;
  case MVT::i32:
    Opc = X86::MOV32rm;
    RC  = X86::GR32RegisterClass;
    break;
  case MVT::i64:
    // Must be in x86-64 mode.
    Opc = X86::MOV64rm;
    RC  = X86::GR64RegisterClass;
    break;
  case MVT::f32:
    if (Subtarget->hasSSE1()) {
      Opc = X86::MOVSSrm;
      RC  = X86::FR32RegisterClass;
    } else {
      Opc = X86::LD_Fp32m;
      RC  = X86::RFP32RegisterClass;
    }
    break;
  case MVT::f64:
    if (Subtarget->hasSSE2()) {
      Opc = X86::MOVSDrm;
      RC  = X86::FR64RegisterClass;
    } else {
      Opc = X86::LD_Fp64m;
      RC  = X86::RFP64RegisterClass;
    }
    break;
  case MVT::f80:
    // No f80 support yet.
    return false;
  }

  // Materialize addresses with LEA instructions.
  if (isa<GlobalValue>(C)) {
    X86AddressMode AM;
    if (X86SelectAddress(C, AM)) {
      if (TLI.getPointerTy() == MVT::i32)
        Opc = X86::LEA32r;
      else
        Opc = X86::LEA64r;
      unsigned ResultReg = createResultReg(RC);
      addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
      return ResultReg;
    }
    return 0;
  }

  // MachineConstantPool wants an explicit alignment.
  unsigned Align = TD.getPrefTypeAlignment(C->getType());
  if (Align == 0) {
    // Alignment of vector types.  FIXME!
    Align = TD.getTypeAllocSize(C->getType());
  }

  // x86-32 PIC requires a PIC base register for constant pools.
  unsigned PICBase = 0;
  unsigned char OpFlag = 0;
  if (Subtarget->isPICStyleStubPIC()) { // Not dynamic-no-pic
    OpFlag = X86II::MO_PIC_BASE_OFFSET;
    PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
  } else if (Subtarget->isPICStyleGOT()) {
    OpFlag = X86II::MO_GOTOFF;
    PICBase = getInstrInfo()->getGlobalBaseReg(&MF);
  } else if (Subtarget->isPICStyleRIPRel() &&
             TM.getCodeModel() == CodeModel::Small) {
    PICBase = X86::RIP;
  }

  // Create the load from the constant pool.
  unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
  unsigned ResultReg = createResultReg(RC);
  addConstantPoolReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg),
                           MCPOffset, PICBase, OpFlag);

  return ResultReg;
}

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

  X86AddressMode AM;
  if (!X86SelectAddress(C, AM))
    return 0;
  unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
  TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
  unsigned ResultReg = createResultReg(RC);
  addLeaAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
  return ResultReg;
}

namespace llvm {
  llvm::FastISel *X86::createFastISel(MachineFunction &mf,
                        MachineModuleInfo *mmi,
                        DwarfWriter *dw,
                        DenseMap<const Value *, unsigned> &vm,
                        DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
                        DenseMap<const AllocaInst *, int> &am
#ifndef NDEBUG
                        , SmallSet<Instruction*, 8> &cil
#endif
                        ) {
    return new X86FastISel(mf, mmi, dw, vm, bm, am
#ifndef NDEBUG
                           , cil
#endif
                           );
  }
}