X86ISelLowering.cpp

      Chain = RetVal.getValue(1);
      if (RetTyVT == MVT::i1) 
        RetVal = DAG.getNode(ISD::TRUNCATE, MVT::i1, RetVal);
      break;
    case MVT::i16:
      RetVal = DAG.getCopyFromReg(Chain, X86::AX, MVT::i16, InFlag);
      Chain = RetVal.getValue(1);
      break;
    case MVT::i32:
      RetVal = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
      Chain = RetVal.getValue(1);
      break;
    case MVT::i64: {
      SDOperand Lo = DAG.getCopyFromReg(Chain, X86::EAX, MVT::i32, InFlag);
      SDOperand Hi = DAG.getCopyFromReg(Lo.getValue(1), X86::EDX, MVT::i32, 
                                        Lo.getValue(2));
      RetVal = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Lo, Hi);
      Chain = Hi.getValue(1);
      break;
    }
    case MVT::f32:
    case MVT::f64: {
      std::vector<MVT::ValueType> Tys;
      Tys.push_back(MVT::f64);
      Tys.push_back(MVT::Other);
      Tys.push_back(MVT::Flag);
      std::vector<SDOperand> Ops;
      Ops.push_back(Chain);
      Ops.push_back(InFlag);
      RetVal = DAG.getNode(X86ISD::FP_GET_RESULT, Tys, Ops);
      Chain  = RetVal.getValue(1);
      InFlag = RetVal.getValue(2);
      if (X86ScalarSSE) {
        // FIXME: Currently the FST is flagged to the FP_GET_RESULT. This
        // shouldn't be necessary except that RFP cannot be live across
        // multiple blocks. When stackifier is fixed, they can be uncoupled.
        MachineFunction &MF = DAG.getMachineFunction();
        int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
        SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
        Tys.clear();
        Tys.push_back(MVT::Other);
        Ops.clear();
        Ops.push_back(Chain);
        Ops.push_back(RetVal);
        Ops.push_back(StackSlot);
        Ops.push_back(DAG.getValueType(RetTyVT));
        Ops.push_back(InFlag);
        Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
        RetVal = DAG.getLoad(RetTyVT, Chain, StackSlot,
                             DAG.getSrcValue(NULL));
        Chain = RetVal.getValue(1);
      }

      if (RetTyVT == MVT::f32 && !X86ScalarSSE)
        // FIXME: we would really like to remember that this FP_ROUND
        // operation is okay to eliminate if we allow excess FP precision.
        RetVal = DAG.getNode(ISD::FP_ROUND, MVT::f32, RetVal);
      break;
    }
    }
  }

  return std::make_pair(RetVal, Chain);
}

SDOperand X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
  if (ReturnAddrIndex == 0) {
    // Set up a frame object for the return address.
    MachineFunction &MF = DAG.getMachineFunction();
    ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(4, -4);
  }

  return DAG.getFrameIndex(ReturnAddrIndex, MVT::i32);
}



std::pair<SDOperand, SDOperand> X86TargetLowering::
LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
                        SelectionDAG &DAG) {
  SDOperand Result;
  if (Depth)        // Depths > 0 not supported yet!
    Result = DAG.getConstant(0, getPointerTy());
  else {
    SDOperand RetAddrFI = getReturnAddressFrameIndex(DAG);
    if (!isFrameAddress)
      // Just load the return address
      Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(), RetAddrFI,
                           DAG.getSrcValue(NULL));
    else
      Result = DAG.getNode(ISD::SUB, MVT::i32, RetAddrFI,
                           DAG.getConstant(4, MVT::i32));
  }
  return std::make_pair(Result, Chain);
}

/// getCondBrOpcodeForX86CC - Returns the X86 conditional branch opcode
/// which corresponds to the condition code.
static unsigned getCondBrOpcodeForX86CC(unsigned X86CC) {
  switch (X86CC) {
  default: assert(0 && "Unknown X86 conditional code!");
  case X86ISD::COND_A:  return X86::JA;
  case X86ISD::COND_AE: return X86::JAE;
  case X86ISD::COND_B:  return X86::JB;
  case X86ISD::COND_BE: return X86::JBE;
  case X86ISD::COND_E:  return X86::JE;
  case X86ISD::COND_G:  return X86::JG;
  case X86ISD::COND_GE: return X86::JGE;
  case X86ISD::COND_L:  return X86::JL;
  case X86ISD::COND_LE: return X86::JLE;
  case X86ISD::COND_NE: return X86::JNE;
  case X86ISD::COND_NO: return X86::JNO;
  case X86ISD::COND_NP: return X86::JNP;
  case X86ISD::COND_NS: return X86::JNS;
  case X86ISD::COND_O:  return X86::JO;
  case X86ISD::COND_P:  return X86::JP;
  case X86ISD::COND_S:  return X86::JS;
  }
}

/// translateX86CC - do a one to one translation of a ISD::CondCode to the X86
/// specific condition code. It returns a false if it cannot do a direct
/// translation. X86CC is the translated CondCode. Flip is set to true if the
/// the order of comparison operands should be flipped.
static bool translateX86CC(SDOperand CC, bool isFP, unsigned &X86CC,
                           bool &Flip) {
  ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
  Flip = false;
  X86CC = X86ISD::COND_INVALID;
  if (!isFP) {
    switch (SetCCOpcode) {
    default: break;
    case ISD::SETEQ:  X86CC = X86ISD::COND_E;  break;
    case ISD::SETGT:  X86CC = X86ISD::COND_G;  break;
    case ISD::SETGE:  X86CC = X86ISD::COND_GE; break;
    case ISD::SETLT:  X86CC = X86ISD::COND_L;  break;
    case ISD::SETLE:  X86CC = X86ISD::COND_LE; break;
    case ISD::SETNE:  X86CC = X86ISD::COND_NE; break;
    case ISD::SETULT: X86CC = X86ISD::COND_B;  break;
    case ISD::SETUGT: X86CC = X86ISD::COND_A;  break;
    case ISD::SETULE: X86CC = X86ISD::COND_BE; break;
    case ISD::SETUGE: X86CC = X86ISD::COND_AE; break;
    }
  } else {
    // On a floating point condition, the flags are set as follows:
    // ZF  PF  CF   op
    //  0 | 0 | 0 | X > Y
    //  0 | 0 | 1 | X < Y
    //  1 | 0 | 0 | X == Y
    //  1 | 1 | 1 | unordered
    switch (SetCCOpcode) {
    default: break;
    case ISD::SETUEQ:
    case ISD::SETEQ: X86CC = X86ISD::COND_E;  break;
    case ISD::SETOLE: Flip = true; // Fallthrough
    case ISD::SETOGT:
    case ISD::SETGT: X86CC = X86ISD::COND_A;  break;
    case ISD::SETOLT: Flip = true; // Fallthrough
    case ISD::SETOGE:
    case ISD::SETGE: X86CC = X86ISD::COND_AE; break;
    case ISD::SETUGE: Flip = true; // Fallthrough
    case ISD::SETULT:
    case ISD::SETLT: X86CC = X86ISD::COND_B;  break;
    case ISD::SETUGT: Flip = true; // Fallthrough
    case ISD::SETULE:
    case ISD::SETLE: X86CC = X86ISD::COND_BE; break;
    case ISD::SETONE:
    case ISD::SETNE: X86CC = X86ISD::COND_NE; break;
    case ISD::SETUO: X86CC = X86ISD::COND_P;  break;
    case ISD::SETO:  X86CC = X86ISD::COND_NP; break;
    }
  }

  return X86CC != X86ISD::COND_INVALID;
}

/// hasFPCMov - is there a floating point cmov for the specific X86 condition
/// code. Current x86 isa includes the following FP cmov instructions:
/// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
static bool hasFPCMov(unsigned X86CC) {
  switch (X86CC) {
  default:
    return false;
  case X86ISD::COND_B:
  case X86ISD::COND_BE:
  case X86ISD::COND_E:
  case X86ISD::COND_P:
  case X86ISD::COND_A:
  case X86ISD::COND_AE:
  case X86ISD::COND_NE:
  case X86ISD::COND_NP:
    return true;
  }
}

MachineBasicBlock *
X86TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
                                           MachineBasicBlock *BB) {
  switch (MI->getOpcode()) {
  default: assert(false && "Unexpected instr type to insert");
  case X86::CMOV_FR32:
  case X86::CMOV_FR64: {
    // To "insert" a SELECT_CC instruction, we actually have to insert the
    // diamond control-flow pattern.  The incoming instruction knows the
    // destination vreg to set, the condition code register to branch on, the
    // true/false values to select between, and a branch opcode to use.
    const BasicBlock *LLVM_BB = BB->getBasicBlock();
    ilist<MachineBasicBlock>::iterator It = BB;
    ++It;
  
    //  thisMBB:
    //  ...
    //   TrueVal = ...
    //   cmpTY ccX, r1, r2
    //   bCC copy1MBB
    //   fallthrough --> copy0MBB
    MachineBasicBlock *thisMBB = BB;
    MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
    MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
    unsigned Opc = getCondBrOpcodeForX86CC(MI->getOperand(3).getImmedValue());
    BuildMI(BB, Opc, 1).addMBB(sinkMBB);
    MachineFunction *F = BB->getParent();
    F->getBasicBlockList().insert(It, copy0MBB);
    F->getBasicBlockList().insert(It, sinkMBB);
    // Update machine-CFG edges
    BB->addSuccessor(copy0MBB);
    BB->addSuccessor(sinkMBB);
  
    //  copy0MBB:
    //   %FalseValue = ...
    //   # fallthrough to sinkMBB
    BB = copy0MBB;
  
    // Update machine-CFG edges
    BB->addSuccessor(sinkMBB);
  
    //  sinkMBB:
    //   %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
    //  ...
    BB = sinkMBB;
    BuildMI(BB, X86::PHI, 4, MI->getOperand(0).getReg())
      .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
      .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);

    delete MI;   // The pseudo instruction is gone now.
    return BB;
  }

  case X86::FP_TO_INT16_IN_MEM:
  case X86::FP_TO_INT32_IN_MEM:
  case X86::FP_TO_INT64_IN_MEM: {
    // Change the floating point control register to use "round towards zero"
    // mode when truncating to an integer value.
    MachineFunction *F = BB->getParent();
    int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
    addFrameReference(BuildMI(BB, X86::FNSTCW16m, 4), CWFrameIdx);

    // Load the old value of the high byte of the control word...
    unsigned OldCW =
      F->getSSARegMap()->createVirtualRegister(X86::R16RegisterClass);
    addFrameReference(BuildMI(BB, X86::MOV16rm, 4, OldCW), CWFrameIdx);

    // Set the high part to be round to zero...
    addFrameReference(BuildMI(BB, X86::MOV16mi, 5), CWFrameIdx).addImm(0xC7F);

    // Reload the modified control word now...
    addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);

    // Restore the memory image of control word to original value
    addFrameReference(BuildMI(BB, X86::MOV16mr, 5), CWFrameIdx).addReg(OldCW);

    // Get the X86 opcode to use.
    unsigned Opc;
    switch (MI->getOpcode()) {
    default: assert(0 && "illegal opcode!");
    case X86::FP_TO_INT16_IN_MEM: Opc = X86::FpIST16m; break;
    case X86::FP_TO_INT32_IN_MEM: Opc = X86::FpIST32m; break;
    case X86::FP_TO_INT64_IN_MEM: Opc = X86::FpIST64m; break;
    }

    X86AddressMode AM;
    MachineOperand &Op = MI->getOperand(0);
    if (Op.isRegister()) {
      AM.BaseType = X86AddressMode::RegBase;
      AM.Base.Reg = Op.getReg();
    } else {
      AM.BaseType = X86AddressMode::FrameIndexBase;
      AM.Base.FrameIndex = Op.getFrameIndex();
    }
    Op = MI->getOperand(1);
    if (Op.isImmediate())
      AM.Scale = Op.getImmedValue();
    Op = MI->getOperand(2);
    if (Op.isImmediate())
      AM.IndexReg = Op.getImmedValue();
    Op = MI->getOperand(3);
    if (Op.isGlobalAddress()) {
      AM.GV = Op.getGlobal();
    } else {
      AM.Disp = Op.getImmedValue();
    }
    addFullAddress(BuildMI(BB, Opc, 5), AM).addReg(MI->getOperand(4).getReg());

    // Reload the original control word now.
    addFrameReference(BuildMI(BB, X86::FLDCW16m, 4), CWFrameIdx);

    delete MI;   // The pseudo instruction is gone now.
    return BB;
  }
  }
}


//===----------------------------------------------------------------------===//
//                           X86 Custom Lowering Hooks
//===----------------------------------------------------------------------===//

/// LowerOperation - Provide custom lowering hooks for some operations.
///
SDOperand X86TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
  switch (Op.getOpcode()) {
  default: assert(0 && "Should not custom lower this!");
  case ISD::SHL_PARTS:
  case ISD::SRA_PARTS:
  case ISD::SRL_PARTS: {
    assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
           "Not an i64 shift!");
    bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
    SDOperand ShOpLo = Op.getOperand(0);
    SDOperand ShOpHi = Op.getOperand(1);
    SDOperand ShAmt  = Op.getOperand(2);
    SDOperand Tmp1 = isSRA ? DAG.getNode(ISD::SRA, MVT::i32, ShOpHi,
                                         DAG.getConstant(31, MVT::i8))
                           : DAG.getConstant(0, MVT::i32);

    SDOperand Tmp2, Tmp3;
    if (Op.getOpcode() == ISD::SHL_PARTS) {
      Tmp2 = DAG.getNode(X86ISD::SHLD, MVT::i32, ShOpHi, ShOpLo, ShAmt);
      Tmp3 = DAG.getNode(ISD::SHL, MVT::i32, ShOpLo, ShAmt);
    } else {
      Tmp2 = DAG.getNode(X86ISD::SHRD, MVT::i32, ShOpLo, ShOpHi, ShAmt);
      Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, MVT::i32, ShOpHi, ShAmt);
    }

    SDOperand InFlag = DAG.getNode(X86ISD::TEST, MVT::Flag,
                                   ShAmt, DAG.getConstant(32, MVT::i8));

    SDOperand Hi, Lo;
    SDOperand CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);

    std::vector<MVT::ValueType> Tys;
    Tys.push_back(MVT::i32);
    Tys.push_back(MVT::Flag);
    std::vector<SDOperand> Ops;
    if (Op.getOpcode() == ISD::SHL_PARTS) {
      Ops.push_back(Tmp2);
      Ops.push_back(Tmp3);
      Ops.push_back(CC);
      Ops.push_back(InFlag);
      Hi = DAG.getNode(X86ISD::CMOV, Tys, Ops);
      InFlag = Hi.getValue(1);

      Ops.clear();
      Ops.push_back(Tmp3);
      Ops.push_back(Tmp1);
      Ops.push_back(CC);
      Ops.push_back(InFlag);
      Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
    } else {
      Ops.push_back(Tmp2);
      Ops.push_back(Tmp3);
      Ops.push_back(CC);
      Ops.push_back(InFlag);
      Lo = DAG.getNode(X86ISD::CMOV, Tys, Ops);
      InFlag = Lo.getValue(1);

      Ops.clear();
      Ops.push_back(Tmp3);
      Ops.push_back(Tmp1);
      Ops.push_back(CC);
      Ops.push_back(InFlag);
      Hi = DAG.getNode(X86ISD::CMOV, Tys, Ops);
    }

    Tys.clear();
    Tys.push_back(MVT::i32);
    Tys.push_back(MVT::i32);
    Ops.clear();
    Ops.push_back(Lo);
    Ops.push_back(Hi);
    return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
  }
  case ISD::SINT_TO_FP: {
    assert(Op.getOperand(0).getValueType() <= MVT::i64 &&
           Op.getOperand(0).getValueType() >= MVT::i16 &&
           "Unknown SINT_TO_FP to lower!");

    SDOperand Result;
    MVT::ValueType SrcVT = Op.getOperand(0).getValueType();
    unsigned Size = MVT::getSizeInBits(SrcVT)/8;
    MachineFunction &MF = DAG.getMachineFunction();
    int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
    SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
    SDOperand Chain = DAG.getNode(ISD::STORE, MVT::Other,
                                  DAG.getEntryNode(), Op.getOperand(0),
                                  StackSlot, DAG.getSrcValue(NULL));

    // Build the FILD
    std::vector<MVT::ValueType> Tys;
    Tys.push_back(MVT::f64);
    Tys.push_back(MVT::Other);
    if (X86ScalarSSE) Tys.push_back(MVT::Flag);
    std::vector<SDOperand> Ops;
    Ops.push_back(Chain);
    Ops.push_back(StackSlot);
    Ops.push_back(DAG.getValueType(SrcVT));
    Result = DAG.getNode(X86ScalarSSE ? X86ISD::FILD_FLAG :X86ISD::FILD,
                         Tys, Ops);

    if (X86ScalarSSE) {
      Chain = Result.getValue(1);
      SDOperand InFlag = Result.getValue(2);

      // FIXME: Currently the FST is flagged to the FILD_FLAG. This
      // shouldn't be necessary except that RFP cannot be live across
      // multiple blocks. When stackifier is fixed, they can be uncoupled.
      MachineFunction &MF = DAG.getMachineFunction();
      int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
      SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
      std::vector<MVT::ValueType> Tys;
      Tys.push_back(MVT::Other);
      std::vector<SDOperand> Ops;
      Ops.push_back(Chain);
      Ops.push_back(Result);
      Ops.push_back(StackSlot);
      Ops.push_back(DAG.getValueType(Op.getValueType()));
      Ops.push_back(InFlag);
      Chain = DAG.getNode(X86ISD::FST, Tys, Ops);
      Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
                           DAG.getSrcValue(NULL));
    }

    return Result;
  }
  case ISD::FP_TO_SINT: {
    assert(Op.getValueType() <= MVT::i64 && Op.getValueType() >= MVT::i16 &&
           "Unknown FP_TO_SINT to lower!");
    // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
    // stack slot.
    MachineFunction &MF = DAG.getMachineFunction();
    unsigned MemSize = MVT::getSizeInBits(Op.getValueType())/8;
    int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
    SDOperand StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());

    unsigned Opc;
    switch (Op.getValueType()) {
    default: assert(0 && "Invalid FP_TO_SINT to lower!");
    case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
    case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
    case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
    }

    SDOperand Chain = DAG.getEntryNode();
    SDOperand Value = Op.getOperand(0);
    if (X86ScalarSSE) {
      assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
      Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value, StackSlot, 
                          DAG.getSrcValue(0));
      std::vector<MVT::ValueType> Tys;
      Tys.push_back(MVT::f64);
      Tys.push_back(MVT::Other);
      std::vector<SDOperand> Ops;
      Ops.push_back(Chain);
      Ops.push_back(StackSlot);
      Ops.push_back(DAG.getValueType(Op.getOperand(0).getValueType()));
      Value = DAG.getNode(X86ISD::FLD, Tys, Ops);
      Chain = Value.getValue(1);
      SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
      StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
    }

    // Build the FP_TO_INT*_IN_MEM
    std::vector<SDOperand> Ops;
    Ops.push_back(Chain);
    Ops.push_back(Value);
    Ops.push_back(StackSlot);
    SDOperand FIST = DAG.getNode(Opc, MVT::Other, Ops);

    // Load the result.
    return DAG.getLoad(Op.getValueType(), FIST, StackSlot,
                       DAG.getSrcValue(NULL));
  }
  case ISD::READCYCLECOUNTER: {
    std::vector<MVT::ValueType> Tys;
    Tys.push_back(MVT::Other);
    Tys.push_back(MVT::Flag);
    std::vector<SDOperand> Ops;
    Ops.push_back(Op.getOperand(0));
    SDOperand rd = DAG.getNode(X86ISD::RDTSC_DAG, Tys, Ops);
    Ops.clear();
    Ops.push_back(DAG.getCopyFromReg(rd, X86::EAX, MVT::i32, rd.getValue(1)));
    Ops.push_back(DAG.getCopyFromReg(Ops[0].getValue(1), X86::EDX, 
                                     MVT::i32, Ops[0].getValue(2)));
    Ops.push_back(Ops[1].getValue(1));
    Tys[0] = Tys[1] = MVT::i32;
    Tys.push_back(MVT::Other);
    return DAG.getNode(ISD::MERGE_VALUES, Tys, Ops);
  }
  case ISD::FABS: {
    MVT::ValueType VT = Op.getValueType();
    const Type *OpNTy =  MVT::getTypeForValueType(VT);
    std::vector<Constant*> CV;
    if (VT == MVT::f64) {
      CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(~(1ULL << 63))));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
    } else {
      CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(~(1U << 31))));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
    }
    Constant *CS = ConstantStruct::get(CV);
    SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
    SDOperand Mask 
      = DAG.getNode(X86ISD::LOAD_PACK,
                    VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
    return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
  }
  case ISD::FNEG: {
    MVT::ValueType VT = Op.getValueType();
    const Type *OpNTy =  MVT::getTypeForValueType(VT);
    std::vector<Constant*> CV;
    if (VT == MVT::f64) {
      CV.push_back(ConstantFP::get(OpNTy, BitsToDouble(1ULL << 63)));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
    } else {
      CV.push_back(ConstantFP::get(OpNTy, BitsToFloat(1U << 31)));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
      CV.push_back(ConstantFP::get(OpNTy, 0.0));
    }
    Constant *CS = ConstantStruct::get(CV);
    SDOperand CPIdx = DAG.getConstantPool(CS, getPointerTy(), 4);
    SDOperand Mask 
      = DAG.getNode(X86ISD::LOAD_PACK,
                    VT, DAG.getEntryNode(), CPIdx, DAG.getSrcValue(NULL));
    return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
  }
  case ISD::SETCC: {
    assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
    SDOperand Cond;
    SDOperand CC = Op.getOperand(2);
    ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
    bool isFP = MVT::isFloatingPoint(Op.getOperand(1).getValueType());
    bool Flip;
    unsigned X86CC;
    if (translateX86CC(CC, isFP, X86CC, Flip)) {
      if (Flip)
        Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
                           Op.getOperand(1), Op.getOperand(0));
      else
        Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
                           Op.getOperand(0), Op.getOperand(1));
      return DAG.getNode(X86ISD::SETCC, MVT::i8, 
                         DAG.getConstant(X86CC, MVT::i8), Cond);
    } else {
      assert(isFP && "Illegal integer SetCC!");

      Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
                         Op.getOperand(0), Op.getOperand(1));
      std::vector<MVT::ValueType> Tys;
      std::vector<SDOperand> Ops;
      switch (SetCCOpcode) {
      default: assert(false && "Illegal floating point SetCC!");
      case ISD::SETOEQ: {  // !PF & ZF
        Tys.push_back(MVT::i8);
        Tys.push_back(MVT::Flag);
        Ops.push_back(DAG.getConstant(X86ISD::COND_NP, MVT::i8));
        Ops.push_back(Cond);
        SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
        SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
                                     DAG.getConstant(X86ISD::COND_E, MVT::i8),
                                     Tmp1.getValue(1));
        return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
      }
      case ISD::SETUNE: {  // PF | !ZF
        Tys.push_back(MVT::i8);
        Tys.push_back(MVT::Flag);
        Ops.push_back(DAG.getConstant(X86ISD::COND_P, MVT::i8));
        Ops.push_back(Cond);
        SDOperand Tmp1 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
        SDOperand Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
                                     DAG.getConstant(X86ISD::COND_NE, MVT::i8),
                                     Tmp1.getValue(1));
        return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
      }
      }
    }
  }
  case ISD::SELECT: {
    MVT::ValueType VT = Op.getValueType();
    bool isFP      = MVT::isFloatingPoint(VT);
    bool isFPStack = isFP && !X86ScalarSSE;
    bool isFPSSE   = isFP && X86ScalarSSE;
    bool addTest   = false;
    SDOperand Op0 = Op.getOperand(0);
    SDOperand Cond, CC;
    if (Op0.getOpcode() == ISD::SETCC)
      Op0 = LowerOperation(Op0, DAG);

    if (Op0.getOpcode() == X86ISD::SETCC) {
      // If condition flag is set by a X86ISD::CMP, then make a copy of it
      // (since flag operand cannot be shared). If the X86ISD::SETCC does not
      // have another use it will be eliminated.
      // If the X86ISD::SETCC has more than one use, then it's probably better
      // to use a test instead of duplicating the X86ISD::CMP (for register
      // pressure reason).
      if (Op0.getOperand(1).getOpcode() == X86ISD::CMP) {
        if (!Op0.hasOneUse()) {
          std::vector<MVT::ValueType> Tys;
          for (unsigned i = 0; i < Op0.Val->getNumValues(); ++i)
            Tys.push_back(Op0.Val->getValueType(i));
          std::vector<SDOperand> Ops;
          for (unsigned i = 0; i < Op0.getNumOperands(); ++i)
            Ops.push_back(Op0.getOperand(i));
          Op0 = DAG.getNode(X86ISD::SETCC, Tys, Ops);
        }

        CC   = Op0.getOperand(0);
        Cond = Op0.getOperand(1);
        // Make a copy as flag result cannot be used by more than one.
        Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
                           Cond.getOperand(0), Cond.getOperand(1));
        addTest =
          isFPStack && !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
      } else
        addTest = true;
    } else
      addTest = true;

    if (addTest) {
      CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
      Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Op0, Op0);
    }

    std::vector<MVT::ValueType> Tys;
    Tys.push_back(Op.getValueType());
    Tys.push_back(MVT::Flag);
    std::vector<SDOperand> Ops;
    // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
    // condition is true.
    Ops.push_back(Op.getOperand(2));
    Ops.push_back(Op.getOperand(1));
    Ops.push_back(CC);
    Ops.push_back(Cond);
    return DAG.getNode(X86ISD::CMOV, Tys, Ops);
  }
  case ISD::BRCOND: {
    bool addTest = false;
    SDOperand Cond  = Op.getOperand(1);
    SDOperand Dest  = Op.getOperand(2);
    SDOperand CC;
    if (Cond.getOpcode() == ISD::SETCC)
      Cond = LowerOperation(Cond, DAG);

    if (Cond.getOpcode() == X86ISD::SETCC) {
      // If condition flag is set by a X86ISD::CMP, then make a copy of it
      // (since flag operand cannot be shared). If the X86ISD::SETCC does not
      // have another use it will be eliminated.
      // If the X86ISD::SETCC has more than one use, then it's probably better
      // to use a test instead of duplicating the X86ISD::CMP (for register
      // pressure reason).
      if (Cond.getOperand(1).getOpcode() == X86ISD::CMP) {
        if (!Cond.hasOneUse()) {
          std::vector<MVT::ValueType> Tys;
          for (unsigned i = 0; i < Cond.Val->getNumValues(); ++i)
            Tys.push_back(Cond.Val->getValueType(i));
          std::vector<SDOperand> Ops;
          for (unsigned i = 0; i < Cond.getNumOperands(); ++i)
            Ops.push_back(Cond.getOperand(i));
          Cond = DAG.getNode(X86ISD::SETCC, Tys, Ops);
        }

        CC   = Cond.getOperand(0);
        Cond = Cond.getOperand(1);
        // Make a copy as flag result cannot be used by more than one.
        Cond = DAG.getNode(X86ISD::CMP, MVT::Flag,
                           Cond.getOperand(0), Cond.getOperand(1));
      } else
        addTest = true;
    } else
      addTest = true;

    if (addTest) {
      CC = DAG.getConstant(X86ISD::COND_NE, MVT::i8);
      Cond = DAG.getNode(X86ISD::TEST, MVT::Flag, Cond, Cond);
    }
    return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
                       Op.getOperand(0), Op.getOperand(2), CC, Cond);
  }
  case ISD::MEMSET: {
    SDOperand InFlag(0, 0);
    SDOperand Chain = Op.getOperand(0);
    unsigned Align =
      (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
    if (Align == 0) Align = 1;

    ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
    // If not DWORD aligned, call memset if size is less than the threshold.
    // It knows how to align to the right boundary first.
    if ((Align & 3) != 0 ||
        !(I && I->getValue() >= Subtarget->getMinRepStrSizeThreshold())) {
      MVT::ValueType IntPtr = getPointerTy();
      const Type *IntPtrTy = getTargetData().getIntPtrType();
      std::vector<std::pair<SDOperand, const Type*> > Args;
      Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
      // Extend the ubyte argument to be an int value for the call.
      SDOperand Val = DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Op.getOperand(2));
      Args.push_back(std::make_pair(Val, IntPtrTy));
      Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
      std::pair<SDOperand,SDOperand> CallResult =
        LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
                    DAG.getExternalSymbol("memset", IntPtr), Args, DAG);
      return CallResult.second;
    }

    MVT::ValueType AVT;
    SDOperand Count;
    ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Op.getOperand(2));
    unsigned BytesLeft = 0;
    if (ValC) {
      unsigned ValReg;
      unsigned Val = ValC->getValue() & 255;

      // If the value is a constant, then we can potentially use larger sets.
      switch (Align & 3) {
      case 2:   // WORD aligned
        AVT = MVT::i16;
        Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
        BytesLeft = I->getValue() % 2;
        Val    = (Val << 8) | Val;
        ValReg = X86::AX;
        break;
      case 0:   // DWORD aligned
        AVT = MVT::i32;
        Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
        BytesLeft = I->getValue() % 4;
        Val = (Val << 8)  | Val;
        Val = (Val << 16) | Val;
        ValReg = X86::EAX;
        break;
      default:  // Byte aligned
        AVT = MVT::i8;
        Count = Op.getOperand(3);
        ValReg = X86::AL;
        break;
      }

      Chain  = DAG.getCopyToReg(Chain, ValReg, DAG.getConstant(Val, AVT),
                                InFlag);
      InFlag = Chain.getValue(1);
    } else {
      AVT = MVT::i8;
      Count  = Op.getOperand(3);
      Chain  = DAG.getCopyToReg(Chain, X86::AL, Op.getOperand(2), InFlag);
      InFlag = Chain.getValue(1);
    }

    Chain  = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
    InFlag = Chain.getValue(1);
    Chain  = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
    InFlag = Chain.getValue(1);
    Chain  = DAG.getNode(X86ISD::REP_STOS, MVT::Other, Chain,
                         DAG.getValueType(AVT), InFlag);

    if (BytesLeft) {
      // Issue stores for the last 1 - 3 bytes.
      SDOperand Value;
      unsigned Val = ValC->getValue() & 255;
      unsigned Offset = I->getValue() - BytesLeft;
      SDOperand DstAddr = Op.getOperand(1);
      MVT::ValueType AddrVT = DstAddr.getValueType();
      if (BytesLeft >= 2) {
        Value = DAG.getConstant((Val << 8) | Val, MVT::i16);
        Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
                            DAG.getNode(ISD::ADD, AddrVT, DstAddr,
                                        DAG.getConstant(Offset, AddrVT)),
                            DAG.getSrcValue(NULL));
        BytesLeft -= 2;
        Offset += 2;
      }

      if (BytesLeft == 1) {
        Value = DAG.getConstant(Val, MVT::i8);
        Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
                            DAG.getNode(ISD::ADD, AddrVT, DstAddr,
                                        DAG.getConstant(Offset, AddrVT)),
                            DAG.getSrcValue(NULL));
      }
    }

    return Chain;
  }
  case ISD::MEMCPY: {
    SDOperand Chain = Op.getOperand(0);
    unsigned Align =
      (unsigned)cast<ConstantSDNode>(Op.getOperand(4))->getValue();
    if (Align == 0) Align = 1;

    ConstantSDNode *I = dyn_cast<ConstantSDNode>(Op.getOperand(3));
    // If not DWORD aligned, call memcpy if size is less than the threshold.
    // It knows how to align to the right boundary first.
    if ((Align & 3) != 0 ||
        !(I && I->getValue() >= Subtarget->getMinRepStrSizeThreshold())) {
      MVT::ValueType IntPtr = getPointerTy();
      const Type *IntPtrTy = getTargetData().getIntPtrType();
      std::vector<std::pair<SDOperand, const Type*> > Args;
      Args.push_back(std::make_pair(Op.getOperand(1), IntPtrTy));
      Args.push_back(std::make_pair(Op.getOperand(2), IntPtrTy));
      Args.push_back(std::make_pair(Op.getOperand(3), IntPtrTy));
      std::pair<SDOperand,SDOperand> CallResult =
        LowerCallTo(Chain, Type::VoidTy, false, CallingConv::C, false,
                    DAG.getExternalSymbol("memcpy", IntPtr), Args, DAG);
      return CallResult.second;
    }

    MVT::ValueType AVT;
    SDOperand Count;
    unsigned BytesLeft = 0;
    switch (Align & 3) {
    case 2:   // WORD aligned
      AVT = MVT::i16;
      Count = DAG.getConstant(I->getValue() / 2, MVT::i32);
      BytesLeft = I->getValue() % 2;
      break;
    case 0:   // DWORD aligned
      AVT = MVT::i32;
      Count = DAG.getConstant(I->getValue() / 4, MVT::i32);
      BytesLeft = I->getValue() % 4;
      break;
    default:  // Byte aligned
      AVT = MVT::i8;
      Count = Op.getOperand(3);
      break;
    }

    SDOperand InFlag(0, 0);
    Chain  = DAG.getCopyToReg(Chain, X86::ECX, Count, InFlag);
    InFlag = Chain.getValue(1);
    Chain  = DAG.getCopyToReg(Chain, X86::EDI, Op.getOperand(1), InFlag);
    InFlag = Chain.getValue(1);
    Chain  = DAG.getCopyToReg(Chain, X86::ESI, Op.getOperand(2), InFlag);
    InFlag = Chain.getValue(1);
    Chain = DAG.getNode(X86ISD::REP_MOVS, MVT::Other, Chain,
                        DAG.getValueType(AVT), InFlag);

    if (BytesLeft) {
      // Issue loads and stores for the last 1 - 3 bytes.
      unsigned Offset = I->getValue() - BytesLeft;
      SDOperand DstAddr = Op.getOperand(1);
      MVT::ValueType DstVT = DstAddr.getValueType();
      SDOperand SrcAddr = Op.getOperand(2);
      MVT::ValueType SrcVT = SrcAddr.getValueType();
      SDOperand Value;
      if (BytesLeft >= 2) {
        Value = DAG.getLoad(MVT::i16, Chain,
                            DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
                                        DAG.getConstant(Offset, SrcVT)),
                            DAG.getSrcValue(NULL));
        Chain = Value.getValue(1);
        Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
                            DAG.getNode(ISD::ADD, DstVT, DstAddr,
                                        DAG.getConstant(Offset, DstVT)),
                            DAG.getSrcValue(NULL));
        BytesLeft -= 2;
        Offset += 2;
      }

      if (BytesLeft == 1) {
        Value = DAG.getLoad(MVT::i8, Chain,
                            DAG.getNode(ISD::ADD, SrcVT, SrcAddr,
                                        DAG.getConstant(Offset, SrcVT)),
                            DAG.getSrcValue(NULL));
        Chain = Value.getValue(1);
        Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
                            DAG.getNode(ISD::ADD, DstVT, DstAddr,
                                        DAG.getConstant(Offset, DstVT)),
                            DAG.getSrcValue(NULL));
      }
    }

    return Chain;
  }

  // ConstantPool, GlobalAddress, and ExternalSymbol are lowered as their
  // target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
  // one of the above mentioned nodes. It has to be wrapped because otherwise
  // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
  // be used to form addressing mode. These wrapped nodes will be selected
  // into MOV32ri.
  case ISD::ConstantPool: {
    ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
    SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
                         DAG.getTargetConstantPool(CP->get(), getPointerTy(),
                                                   CP->getAlignment()));
    if (getTargetMachine().getSubtarget<X86Subtarget>().isTargetDarwin()) {
      // With PIC, the address is actually $g + Offset.
      if (getTargetMachine().getRelocationModel() == Reloc::PIC)
        Result = DAG.getNode(ISD::ADD, getPointerTy(),
                DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);    
    }

    return Result;
  }
  case ISD::GlobalAddress: {
    GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
    SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
                         DAG.getTargetGlobalAddress(GV, getPointerTy()));
    if (getTargetMachine().
        getSubtarget<X86Subtarget>().isTargetDarwin()) {
      // With PIC, the address is actually $g + Offset.
      if (getTargetMachine().getRelocationModel() == Reloc::PIC)
        Result = DAG.getNode(ISD::ADD, getPointerTy(),
                    DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);

      // For Darwin, external and weak symbols are indirect, so we want to load
      // the value at address GV, not the value of GV itself.  This means that
      // the GlobalAddress must be in the base or index register of the address,
      // not the GV offset field.
      if (getTargetMachine().getRelocationModel() != Reloc::Static &&
          (GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() ||
           (GV->isExternal() && !GV->hasNotBeenReadFromBytecode())))
        Result = DAG.getLoad(MVT::i32, DAG.getEntryNode(),
                             Result, DAG.getSrcValue(NULL));
    }

    return Result;
  }
  case ISD::ExternalSymbol: {
    const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
    SDOperand Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(),
                         DAG.getTargetExternalSymbol(Sym, getPointerTy()));
    if (getTargetMachine().
        getSubtarget<X86Subtarget>().isTargetDarwin()) {
      // With PIC, the address is actually $g + Offset.
      if (getTargetMachine().getRelocationModel() == Reloc::PIC)
        Result = DAG.getNode(ISD::ADD, getPointerTy(),
                    DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()), Result);
    }

    return Result;
  }
  case ISD::VASTART: {
    // vastart just stores the address of the VarArgsFrameIndex slot into the
    // memory location argument.
    // FIXME: Replace MVT::i32 with PointerTy
    SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
    return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR, 
                       Op.getOperand(1), Op.getOperand(2));
  }
  case ISD::RET: {
    SDOperand Copy;
    
    switch(Op.getNumOperands()) {
    default:
      assert(0 && "Do not know how to return this many arguments!");
      abort();
    case 1: 
      return DAG.getNode(X86ISD::RET_FLAG, MVT::Other, Op.getOperand(0),
                         DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
    case 2: {
      MVT::ValueType ArgVT = Op.getOperand(1).getValueType();
      if (MVT::isInteger(ArgVT))
        Copy = DAG.getCopyToReg(Op.getOperand(0), X86::EAX, Op.getOperand(1),
                                SDOperand());
      else if (!X86ScalarSSE) {
        std::vector<MVT::ValueType> Tys;
        Tys.push_back(MVT::Other);
        Tys.push_back(MVT::Flag);
        std::vector<SDOperand> Ops;
        Ops.push_back(Op.getOperand(0));
        Ops.push_back(Op.getOperand(1));
        Copy = DAG.getNode(X86ISD::FP_SET_RESULT, Tys, Ops);
      } else {
        SDOperand MemLoc;
        SDOperand Chain = Op.getOperand(0);
        SDOperand Value = Op.getOperand(1);

        if (Value.getOpcode() == ISD::LOAD &&
            (Chain == Value.getValue(1) || Chain == Value.getOperand(0))) {
          Chain  = Value.getOperand(0);
          MemLoc = Value.getOperand(1);
        } else {
          // Spill the value to memory and reload it into top of stack.
          unsigned Size = MVT::getSizeInBits(ArgVT)/8;
          MachineFunction &MF = DAG.getMachineFunction();
          int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
          MemLoc = DAG.getFrameIndex(SSFI, getPointerTy());
          Chain = DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), 
                              Value, MemLoc, DAG.getSrcValue(0));