X86ISelLowering.cpp

//===-- X86ISelLowering.h - X86 DAG Lowering Interface ----------*- C++ -*-===//
//
//                     The LLVM Compiler Infrastructure
//
// This file was developed by Chris Lattner and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the interfaces that X86 uses to lower LLVM code into a
// selection DAG.
//
//===----------------------------------------------------------------------===//

#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
#include "X86TargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
#include "llvm/ADT/VectorExtras.h"
#include "llvm/Analysis/ScalarEvolutionExpressions.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SSARegMap.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;

// FIXME: temporary.
#include "llvm/Support/CommandLine.h"
static cl::opt<bool> EnableFastCC("enable-x86-fastcc", cl::Hidden,
                                  cl::desc("Enable fastcc on X86"));

X86TargetLowering::X86TargetLowering(TargetMachine &TM)
  : TargetLowering(TM) {
  Subtarget = &TM.getSubtarget<X86Subtarget>();
  X86ScalarSSE = Subtarget->hasSSE2();

  // Set up the TargetLowering object.

  // X86 is weird, it always uses i8 for shift amounts and setcc results.
  setShiftAmountType(MVT::i8);
  setSetCCResultType(MVT::i8);
  setSetCCResultContents(ZeroOrOneSetCCResult);
  setSchedulingPreference(SchedulingForRegPressure);
  setShiftAmountFlavor(Mask);   // shl X, 32 == shl X, 0
  setStackPointerRegisterToSaveRestore(X86::ESP);

  // Add legal addressing mode scale values.
  addLegalAddressScale(8);
  addLegalAddressScale(4);
  addLegalAddressScale(2);
  // Enter the ones which require both scale + index last. These are more
  // expensive.
  addLegalAddressScale(9);
  addLegalAddressScale(5);
  addLegalAddressScale(3);
  
  // Set up the register classes.
  addRegisterClass(MVT::i8, X86::R8RegisterClass);
  addRegisterClass(MVT::i16, X86::R16RegisterClass);
  addRegisterClass(MVT::i32, X86::R32RegisterClass);

  // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
  // operation.
  setOperationAction(ISD::UINT_TO_FP       , MVT::i1   , Promote);
  setOperationAction(ISD::UINT_TO_FP       , MVT::i8   , Promote);
  setOperationAction(ISD::UINT_TO_FP       , MVT::i16  , Promote);

  if (X86ScalarSSE)
    // No SSE i64 SINT_TO_FP, so expand i32 UINT_TO_FP instead.
    setOperationAction(ISD::UINT_TO_FP     , MVT::i32  , Expand);
  else
    setOperationAction(ISD::UINT_TO_FP     , MVT::i32  , Promote);

  // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
  // this operation.
  setOperationAction(ISD::SINT_TO_FP       , MVT::i1   , Promote);
  setOperationAction(ISD::SINT_TO_FP       , MVT::i8   , Promote);
  // SSE has no i16 to fp conversion, only i32
  if (X86ScalarSSE)
    setOperationAction(ISD::SINT_TO_FP     , MVT::i16  , Promote);
  else {
    setOperationAction(ISD::SINT_TO_FP     , MVT::i16  , Custom);
    setOperationAction(ISD::SINT_TO_FP     , MVT::i32  , Custom);
  }

  // We can handle SINT_TO_FP and FP_TO_SINT from/to i64 even though i64
  // isn't legal.
  setOperationAction(ISD::SINT_TO_FP       , MVT::i64  , Custom);
  setOperationAction(ISD::FP_TO_SINT       , MVT::i64  , Custom);

  // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
  // this operation.
  setOperationAction(ISD::FP_TO_SINT       , MVT::i1   , Promote);
  setOperationAction(ISD::FP_TO_SINT       , MVT::i8   , Promote);

  if (X86ScalarSSE) {
    setOperationAction(ISD::FP_TO_SINT     , MVT::i16  , Promote);
  } else {
    setOperationAction(ISD::FP_TO_SINT     , MVT::i16  , Custom);
    setOperationAction(ISD::FP_TO_SINT     , MVT::i32  , Custom);
  }

  // Handle FP_TO_UINT by promoting the destination to a larger signed
  // conversion.
  setOperationAction(ISD::FP_TO_UINT       , MVT::i1   , Promote);
  setOperationAction(ISD::FP_TO_UINT       , MVT::i8   , Promote);
  setOperationAction(ISD::FP_TO_UINT       , MVT::i16  , Promote);

  if (X86ScalarSSE && !Subtarget->hasSSE3())
    // Expand FP_TO_UINT into a select.
    // FIXME: We would like to use a Custom expander here eventually to do
    // the optimal thing for SSE vs. the default expansion in the legalizer.
    setOperationAction(ISD::FP_TO_UINT     , MVT::i32  , Expand);
  else
    // With SSE3 we can use fisttpll to convert to a signed i64.
    setOperationAction(ISD::FP_TO_UINT     , MVT::i32  , Promote);

  setOperationAction(ISD::BIT_CONVERT      , MVT::f32  , Expand);
  setOperationAction(ISD::BIT_CONVERT      , MVT::i32  , Expand);

  setOperationAction(ISD::BRCOND           , MVT::Other, Custom);
  setOperationAction(ISD::BR_CC            , MVT::Other, Expand);
  setOperationAction(ISD::SELECT_CC        , MVT::Other, Expand);
  setOperationAction(ISD::MEMMOVE          , MVT::Other, Expand);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16  , Expand);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8   , Expand);
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1   , Expand);
  setOperationAction(ISD::FP_ROUND_INREG   , MVT::f32  , Expand);
  setOperationAction(ISD::SEXTLOAD         , MVT::i1   , Expand);
  setOperationAction(ISD::FREM             , MVT::f64  , Expand);
  setOperationAction(ISD::CTPOP            , MVT::i8   , Expand);
  setOperationAction(ISD::CTTZ             , MVT::i8   , Expand);
  setOperationAction(ISD::CTLZ             , MVT::i8   , Expand);
  setOperationAction(ISD::CTPOP            , MVT::i16  , Expand);
  setOperationAction(ISD::CTTZ             , MVT::i16  , Expand);
  setOperationAction(ISD::CTLZ             , MVT::i16  , Expand);
  setOperationAction(ISD::CTPOP            , MVT::i32  , Expand);
  setOperationAction(ISD::CTTZ             , MVT::i32  , Expand);
  setOperationAction(ISD::CTLZ             , MVT::i32  , Expand);
  setOperationAction(ISD::READCYCLECOUNTER , MVT::i64  , Custom);
  setOperationAction(ISD::BSWAP            , MVT::i16  , Expand);

  // These should be promoted to a larger select which is supported.
  setOperationAction(ISD::SELECT           , MVT::i1   , Promote);
  setOperationAction(ISD::SELECT           , MVT::i8   , Promote);

  // X86 wants to expand cmov itself.
  setOperationAction(ISD::SELECT          , MVT::i16  , Custom);
  setOperationAction(ISD::SELECT          , MVT::i32  , Custom);
  setOperationAction(ISD::SELECT          , MVT::f32  , Custom);
  setOperationAction(ISD::SELECT          , MVT::f64  , Custom);
  setOperationAction(ISD::SETCC           , MVT::i8   , Custom);
  setOperationAction(ISD::SETCC           , MVT::i16  , Custom);
  setOperationAction(ISD::SETCC           , MVT::i32  , Custom);
  setOperationAction(ISD::SETCC           , MVT::f32  , Custom);
  setOperationAction(ISD::SETCC           , MVT::f64  , Custom);
  // X86 ret instruction may pop stack.
  setOperationAction(ISD::RET             , MVT::Other, Custom);
  // Darwin ABI issue.
  setOperationAction(ISD::ConstantPool    , MVT::i32  , Custom);
  setOperationAction(ISD::GlobalAddress   , MVT::i32  , Custom);
  setOperationAction(ISD::ExternalSymbol  , MVT::i32  , Custom);
  // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
  setOperationAction(ISD::SHL_PARTS       , MVT::i32  , Custom);
  setOperationAction(ISD::SRA_PARTS       , MVT::i32  , Custom);
  setOperationAction(ISD::SRL_PARTS       , MVT::i32  , Custom);
  // X86 wants to expand memset / memcpy itself.
  setOperationAction(ISD::MEMSET          , MVT::Other, Custom);
  setOperationAction(ISD::MEMCPY          , MVT::Other, Custom);

  // We don't have line number support yet.
  setOperationAction(ISD::LOCATION, MVT::Other, Expand);
  setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
  // FIXME - use subtarget debug flags
  if (!TM.getSubtarget<X86Subtarget>().isTargetDarwin())
    setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);

  // VASTART needs to be custom lowered to use the VarArgsFrameIndex
  setOperationAction(ISD::VASTART           , MVT::Other, Custom);
  
  // Use the default implementation.
  setOperationAction(ISD::VAARG             , MVT::Other, Expand);
  setOperationAction(ISD::VACOPY            , MVT::Other, Expand);
  setOperationAction(ISD::VAEND             , MVT::Other, Expand);
  setOperationAction(ISD::STACKSAVE,          MVT::Other, Expand); 
  setOperationAction(ISD::STACKRESTORE,       MVT::Other, Expand);
  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32  , Expand);

  setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
  setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);

  if (X86ScalarSSE) {
    // Set up the FP register classes.
    addRegisterClass(MVT::f32, X86::FR32RegisterClass);
    addRegisterClass(MVT::f64, X86::FR64RegisterClass);

    // SSE has no load+extend ops
    setOperationAction(ISD::EXTLOAD,  MVT::f32, Expand);
    setOperationAction(ISD::ZEXTLOAD, MVT::f32, Expand);

    // Use ANDPD to simulate FABS.
    setOperationAction(ISD::FABS , MVT::f64, Custom);
    setOperationAction(ISD::FABS , MVT::f32, Custom);

    // Use XORP to simulate FNEG.
    setOperationAction(ISD::FNEG , MVT::f64, Custom);
    setOperationAction(ISD::FNEG , MVT::f32, Custom);

    // We don't support sin/cos/fmod
    setOperationAction(ISD::FSIN , MVT::f64, Expand);
    setOperationAction(ISD::FCOS , MVT::f64, Expand);
    setOperationAction(ISD::FREM , MVT::f64, Expand);
    setOperationAction(ISD::FSIN , MVT::f32, Expand);
    setOperationAction(ISD::FCOS , MVT::f32, Expand);
    setOperationAction(ISD::FREM , MVT::f32, Expand);

    // Expand FP immediates into loads from the stack, except for the special
    // cases we handle.
    setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
    setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
    addLegalFPImmediate(+0.0); // xorps / xorpd
  } else {
    // Set up the FP register classes.
    addRegisterClass(MVT::f64, X86::RFPRegisterClass);
    
    setOperationAction(ISD::UNDEF, MVT::f64, Expand);
    
    if (!UnsafeFPMath) {
      setOperationAction(ISD::FSIN           , MVT::f64  , Expand);
      setOperationAction(ISD::FCOS           , MVT::f64  , Expand);
    }

    setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
    addLegalFPImmediate(+0.0); // FLD0
    addLegalFPImmediate(+1.0); // FLD1
    addLegalFPImmediate(-0.0); // FLD0/FCHS
    addLegalFPImmediate(-1.0); // FLD1/FCHS
  }

  // First set operation action for all vector types to expand. Then we
  // will selectively turn on ones that can be effectively codegen'd.
  for (unsigned VT = (unsigned)MVT::Vector + 1;
       VT != (unsigned)MVT::LAST_VALUETYPE; VT++) {
    setOperationAction(ISD::ADD , (MVT::ValueType)VT, Expand);
    setOperationAction(ISD::SUB , (MVT::ValueType)VT, Expand);
    setOperationAction(ISD::MUL , (MVT::ValueType)VT, Expand);
    setOperationAction(ISD::LOAD, (MVT::ValueType)VT, Expand);
  }

  if (TM.getSubtarget<X86Subtarget>().hasMMX()) {
    addRegisterClass(MVT::v8i8,  X86::VR64RegisterClass);
    addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
    addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);

    // FIXME: add MMX packed arithmetics
    setOperationAction(ISD::ConstantVec, MVT::v8i8,  Expand);
    setOperationAction(ISD::ConstantVec, MVT::v4i16, Expand);
    setOperationAction(ISD::ConstantVec, MVT::v2i32, Expand);
  }

  if (TM.getSubtarget<X86Subtarget>().hasSSE1()) {
    addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);

    setOperationAction(ISD::ADD        , MVT::v4f32, Legal);
    setOperationAction(ISD::SUB        , MVT::v4f32, Legal);
    setOperationAction(ISD::MUL        , MVT::v4f32, Legal);
    setOperationAction(ISD::LOAD       , MVT::v4f32, Legal);
    setOperationAction(ISD::ConstantVec, MVT::v4f32, Expand);
  }

  if (TM.getSubtarget<X86Subtarget>().hasSSE2()) {
    addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
    addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
    addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
    addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
    addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);


    setOperationAction(ISD::ADD        , MVT::v2f64, Legal);
    setOperationAction(ISD::SUB        , MVT::v2f64, Legal);
    setOperationAction(ISD::MUL        , MVT::v2f64, Legal);
    setOperationAction(ISD::LOAD       , MVT::v2f64, Legal);
    setOperationAction(ISD::ConstantVec, MVT::v2f64, Expand);
    setOperationAction(ISD::ConstantVec, MVT::v16i8, Expand);
    setOperationAction(ISD::ConstantVec, MVT::v8i16, Expand);
    setOperationAction(ISD::ConstantVec, MVT::v4i32, Expand);
    setOperationAction(ISD::ConstantVec, MVT::v2i64, Expand);
  }

  computeRegisterProperties();

  // FIXME: These should be based on subtarget info. Plus, the values should
  // be smaller when we are in optimizing for size mode.
  maxStoresPerMemset = 16; // For %llvm.memset -> sequence of stores
  maxStoresPerMemcpy = 16; // For %llvm.memcpy -> sequence of stores
  maxStoresPerMemmove = 16; // For %llvm.memmove -> sequence of stores
  allowUnalignedMemoryAccesses = true; // x86 supports it!
}

std::vector<SDOperand>
X86TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
  if (F.getCallingConv() == CallingConv::Fast && EnableFastCC)
    return LowerFastCCArguments(F, DAG);
  return LowerCCCArguments(F, DAG);
}

std::pair<SDOperand, SDOperand>
X86TargetLowering::LowerCallTo(SDOperand Chain, const Type *RetTy,
                               bool isVarArg, unsigned CallingConv,
                               bool isTailCall,
                               SDOperand Callee, ArgListTy &Args,
                               SelectionDAG &DAG) {
  assert((!isVarArg || CallingConv == CallingConv::C) &&
         "Only C takes varargs!");

  // If the callee is a GlobalAddress node (quite common, every direct call is)
  // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee))
    Callee = DAG.getTargetGlobalAddress(G->getGlobal(), getPointerTy());
  else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee))
    Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy());

  if (CallingConv == CallingConv::Fast && EnableFastCC)
    return LowerFastCCCallTo(Chain, RetTy, isTailCall, Callee, Args, DAG);
  return  LowerCCCCallTo(Chain, RetTy, isVarArg, isTailCall, Callee, Args, DAG);
}

//===----------------------------------------------------------------------===//
//                    C Calling Convention implementation
//===----------------------------------------------------------------------===//

std::vector<SDOperand>
X86TargetLowering::LowerCCCArguments(Function &F, SelectionDAG &DAG) {
  std::vector<SDOperand> ArgValues;

  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();

  // Add DAG nodes to load the arguments...  On entry to a function on the X86,
  // the stack frame looks like this:
  //
  // [ESP] -- return address
  // [ESP + 4] -- first argument (leftmost lexically)
  // [ESP + 8] -- second argument, if first argument is four bytes in size
  //    ...
  //
  unsigned ArgOffset = 0;   // Frame mechanisms handle retaddr slot
  for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
    MVT::ValueType ObjectVT = getValueType(I->getType());
    unsigned ArgIncrement = 4;
    unsigned ObjSize;
    switch (ObjectVT) {
    default: assert(0 && "Unhandled argument type!");
    case MVT::i1:
    case MVT::i8:  ObjSize = 1;                break;
    case MVT::i16: ObjSize = 2;                break;
    case MVT::i32: ObjSize = 4;                break;
    case MVT::i64: ObjSize = ArgIncrement = 8; break;
    case MVT::f32: ObjSize = 4;                break;
    case MVT::f64: ObjSize = ArgIncrement = 8; break;
    }
    // Create the frame index object for this incoming parameter...
    int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);

    // Create the SelectionDAG nodes corresponding to a load from this parameter
    SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);

    // Don't codegen dead arguments.  FIXME: remove this check when we can nuke
    // dead loads.
    SDOperand ArgValue;
    if (!I->use_empty())
      ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
                             DAG.getSrcValue(NULL));
    else {
      if (MVT::isInteger(ObjectVT))
        ArgValue = DAG.getConstant(0, ObjectVT);
      else
        ArgValue = DAG.getConstantFP(0, ObjectVT);
    }
    ArgValues.push_back(ArgValue);

    ArgOffset += ArgIncrement;   // Move on to the next argument...
  }

  // If the function takes variable number of arguments, make a frame index for
  // the start of the first vararg value... for expansion of llvm.va_start.
  if (F.isVarArg())
    VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
  ReturnAddrIndex = 0;     // No return address slot generated yet.
  BytesToPopOnReturn = 0;  // Callee pops nothing.
  BytesCallerReserves = ArgOffset;

  // Finally, inform the code generator which regs we return values in.
  switch (getValueType(F.getReturnType())) {
  default: assert(0 && "Unknown type!");
  case MVT::isVoid: break;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
    MF.addLiveOut(X86::EAX);
    break;
  case MVT::i64:
    MF.addLiveOut(X86::EAX);
    MF.addLiveOut(X86::EDX);
    break;
  case MVT::f32:
  case MVT::f64:
    MF.addLiveOut(X86::ST0);
    break;
  }
  return ArgValues;
}

std::pair<SDOperand, SDOperand>
X86TargetLowering::LowerCCCCallTo(SDOperand Chain, const Type *RetTy,
                                  bool isVarArg, bool isTailCall,
                                  SDOperand Callee, ArgListTy &Args,
                                  SelectionDAG &DAG) {
  // Count how many bytes are to be pushed on the stack.
  unsigned NumBytes = 0;

  if (Args.empty()) {
    // Save zero bytes.
    Chain = DAG.getCALLSEQ_START(Chain, DAG.getConstant(0, getPointerTy()));
  } else {
    for (unsigned i = 0, e = Args.size(); i != e; ++i)
      switch (getValueType(Args[i].second)) {
      default: assert(0 && "Unknown value type!");
      case MVT::i1:
      case MVT::i8:
      case MVT::i16:
      case MVT::i32:
      case MVT::f32:
        NumBytes += 4;
        break;
      case MVT::i64:
      case MVT::f64:
        NumBytes += 8;
        break;
      }

    Chain = DAG.getCALLSEQ_START(Chain,
                                 DAG.getConstant(NumBytes, getPointerTy()));

    // Arguments go on the stack in reverse order, as specified by the ABI.
    unsigned ArgOffset = 0;
    SDOperand StackPtr = DAG.getRegister(X86::ESP, MVT::i32);
    std::vector<SDOperand> Stores;

    for (unsigned i = 0, e = Args.size(); i != e; ++i) {
      SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
      PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);

      switch (getValueType(Args[i].second)) {
      default: assert(0 && "Unexpected ValueType for argument!");
      case MVT::i1:
      case MVT::i8:
      case MVT::i16:
        // Promote the integer to 32 bits.  If the input type is signed use a
        // sign extend, otherwise use a zero extend.
        if (Args[i].second->isSigned())
          Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
        else
          Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);

        // FALL THROUGH
      case MVT::i32:
      case MVT::f32:
        Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
                                     Args[i].first, PtrOff,
                                     DAG.getSrcValue(NULL)));
        ArgOffset += 4;
        break;
      case MVT::i64:
      case MVT::f64:
        Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
                                     Args[i].first, PtrOff,
                                     DAG.getSrcValue(NULL)));
        ArgOffset += 8;
        break;
      }
    }
    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);
  }

  std::vector<MVT::ValueType> RetVals;
  MVT::ValueType RetTyVT = getValueType(RetTy);
  RetVals.push_back(MVT::Other);

  // The result values produced have to be legal.  Promote the result.
  switch (RetTyVT) {
  case MVT::isVoid: break;
  default:
    RetVals.push_back(RetTyVT);
    break;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
    RetVals.push_back(MVT::i32);
    break;
  case MVT::f32:
    if (X86ScalarSSE)
      RetVals.push_back(MVT::f32);
    else
      RetVals.push_back(MVT::f64);
    break;
  case MVT::i64:
    RetVals.push_back(MVT::i32);
    RetVals.push_back(MVT::i32);
    break;
  }

  std::vector<MVT::ValueType> NodeTys;
  NodeTys.push_back(MVT::Other);   // Returns a chain
  NodeTys.push_back(MVT::Flag);    // Returns a flag for retval copy to use.
  std::vector<SDOperand> Ops;
  Ops.push_back(Chain);
  Ops.push_back(Callee);

  // FIXME: Do not generate X86ISD::TAILCALL for now.
  Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops);
  SDOperand InFlag = Chain.getValue(1);

  NodeTys.clear();
  NodeTys.push_back(MVT::Other);   // Returns a chain
  NodeTys.push_back(MVT::Flag);    // Returns a flag for retval copy to use.
  Ops.clear();
  Ops.push_back(Chain);
  Ops.push_back(DAG.getConstant(NumBytes, getPointerTy()));
  Ops.push_back(DAG.getConstant(0, getPointerTy()));
  Ops.push_back(InFlag);
  Chain = DAG.getNode(ISD::CALLSEQ_END, NodeTys, Ops);
  InFlag = Chain.getValue(1);
  
  SDOperand RetVal;
  if (RetTyVT != MVT::isVoid) {
    switch (RetTyVT) {
    default: assert(0 && "Unknown value type to return!");
    case MVT::i1:
    case MVT::i8:
      RetVal = DAG.getCopyFromReg(Chain, X86::AL, MVT::i8, InFlag);
      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);
}

//===----------------------------------------------------------------------===//
//                    Fast Calling Convention implementation
//===----------------------------------------------------------------------===//
//
// The X86 'fast' calling convention passes up to two integer arguments in
// registers (an appropriate portion of EAX/EDX), passes arguments in C order,
// and requires that the callee pop its arguments off the stack (allowing proper
// tail calls), and has the same return value conventions as C calling convs.
//
// This calling convention always arranges for the callee pop value to be 8n+4
// bytes, which is needed for tail recursion elimination and stack alignment
// reasons.
//
// Note that this can be enhanced in the future to pass fp vals in registers
// (when we have a global fp allocator) and do other tricks.
//

/// AddLiveIn - This helper function adds the specified physical register to the
/// MachineFunction as a live in value.  It also creates a corresponding virtual
/// register for it.
static unsigned AddLiveIn(MachineFunction &MF, unsigned PReg,
                          TargetRegisterClass *RC) {
  assert(RC->contains(PReg) && "Not the correct regclass!");
  unsigned VReg = MF.getSSARegMap()->createVirtualRegister(RC);
  MF.addLiveIn(PReg, VReg);
  return VReg;
}

// FASTCC_NUM_INT_ARGS_INREGS - This is the max number of integer arguments
// to pass in registers.  0 is none, 1 is is "use EAX", 2 is "use EAX and
// EDX".  Anything more is illegal.
//
// FIXME: The linscan register allocator currently has problem with
// coallescing.  At the time of this writing, whenever it decides to coallesce
// a physreg with a virtreg, this increases the size of the physreg's live
// range, and the live range cannot ever be reduced.  This causes problems if
// too many physregs are coalleced with virtregs, which can cause the register
// allocator to wedge itself.
//
// This code triggers this problem more often if we pass args in registers,
// so disable it until this is fixed.
//
// NOTE: this isn't marked const, so that GCC doesn't emit annoying warnings
// about code being dead.
//
static unsigned FASTCC_NUM_INT_ARGS_INREGS = 0;


std::vector<SDOperand>
X86TargetLowering::LowerFastCCArguments(Function &F, SelectionDAG &DAG) {
  std::vector<SDOperand> ArgValues;

  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();

  // Add DAG nodes to load the arguments...  On entry to a function the stack
  // frame looks like this:
  //
  // [ESP] -- return address
  // [ESP + 4] -- first nonreg argument (leftmost lexically)
  // [ESP + 8] -- second nonreg argument, if first argument is 4 bytes in size
  //    ...
  unsigned ArgOffset = 0;   // Frame mechanisms handle retaddr slot

  // Keep track of the number of integer regs passed so far.  This can be either
  // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
  // used).
  unsigned NumIntRegs = 0;
  
  for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
    MVT::ValueType ObjectVT = getValueType(I->getType());
    unsigned ArgIncrement = 4;
    unsigned ObjSize = 0;
    SDOperand ArgValue;

    switch (ObjectVT) {
    default: assert(0 && "Unhandled argument type!");
    case MVT::i1:
    case MVT::i8:
      if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
        if (!I->use_empty()) {
          unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DL : X86::AL,
                                    X86::R8RegisterClass);
          ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i8);
          DAG.setRoot(ArgValue.getValue(1));
          if (ObjectVT == MVT::i1)
            // FIXME: Should insert a assertzext here.
            ArgValue = DAG.getNode(ISD::TRUNCATE, MVT::i1, ArgValue);
        }
        ++NumIntRegs;
        break;
      }

      ObjSize = 1;
      break;
    case MVT::i16:
      if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
        if (!I->use_empty()) {
          unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::DX : X86::AX,
                                    X86::R16RegisterClass);
          ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i16);
          DAG.setRoot(ArgValue.getValue(1));
        }
        ++NumIntRegs;
        break;
      }
      ObjSize = 2;
      break;
    case MVT::i32:
      if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
        if (!I->use_empty()) {
          unsigned VReg = AddLiveIn(MF, NumIntRegs ? X86::EDX : X86::EAX,
                                    X86::R32RegisterClass);
          ArgValue = DAG.getCopyFromReg(DAG.getRoot(), VReg, MVT::i32);
          DAG.setRoot(ArgValue.getValue(1));
        }
        ++NumIntRegs;
        break;
      }
      ObjSize = 4;
      break;
    case MVT::i64:
      if (NumIntRegs+2 <= FASTCC_NUM_INT_ARGS_INREGS) {
        if (!I->use_empty()) {
          unsigned BotReg = AddLiveIn(MF, X86::EAX, X86::R32RegisterClass);
          unsigned TopReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);

          SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
          SDOperand Hi  = DAG.getCopyFromReg(Low.getValue(1), TopReg, MVT::i32);
          DAG.setRoot(Hi.getValue(1));

          ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
        }
        NumIntRegs += 2;
        break;
      } else if (NumIntRegs+1 <= FASTCC_NUM_INT_ARGS_INREGS) {
        if (!I->use_empty()) {
          unsigned BotReg = AddLiveIn(MF, X86::EDX, X86::R32RegisterClass);
          SDOperand Low = DAG.getCopyFromReg(DAG.getRoot(), BotReg, MVT::i32);
          DAG.setRoot(Low.getValue(1));

          // Load the high part from memory.
          // Create the frame index object for this incoming parameter...
          int FI = MFI->CreateFixedObject(4, ArgOffset);
          SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
          SDOperand Hi = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
                                     DAG.getSrcValue(NULL));
          ArgValue = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, Low, Hi);
        }
        ArgOffset += 4;
        NumIntRegs = FASTCC_NUM_INT_ARGS_INREGS;
        break;
      }
      ObjSize = ArgIncrement = 8;
      break;
    case MVT::f32: ObjSize = 4;                break;
    case MVT::f64: ObjSize = ArgIncrement = 8; break;
    }

    // Don't codegen dead arguments.  FIXME: remove this check when we can nuke
    // dead loads.
    if (ObjSize && !I->use_empty()) {
      // Create the frame index object for this incoming parameter...
      int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);

      // Create the SelectionDAG nodes corresponding to a load from this
      // parameter.
      SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);

      ArgValue = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
                             DAG.getSrcValue(NULL));
    } else if (ArgValue.Val == 0) {
      if (MVT::isInteger(ObjectVT))
        ArgValue = DAG.getConstant(0, ObjectVT);
      else
        ArgValue = DAG.getConstantFP(0, ObjectVT);
    }
    ArgValues.push_back(ArgValue);

    if (ObjSize)
      ArgOffset += ArgIncrement;   // Move on to the next argument.
  }

  // Make sure the instruction takes 8n+4 bytes to make sure the start of the
  // arguments and the arguments after the retaddr has been pushed are aligned.
  if ((ArgOffset & 7) == 0)
    ArgOffset += 4;

  VarArgsFrameIndex = 0xAAAAAAA;   // fastcc functions can't have varargs.
  ReturnAddrIndex = 0;             // No return address slot generated yet.
  BytesToPopOnReturn = ArgOffset;  // Callee pops all stack arguments.
  BytesCallerReserves = 0;

  // Finally, inform the code generator which regs we return values in.
  switch (getValueType(F.getReturnType())) {
  default: assert(0 && "Unknown type!");
  case MVT::isVoid: break;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
  case MVT::i32:
    MF.addLiveOut(X86::EAX);
    break;
  case MVT::i64:
    MF.addLiveOut(X86::EAX);
    MF.addLiveOut(X86::EDX);
    break;
  case MVT::f32:
  case MVT::f64:
    MF.addLiveOut(X86::ST0);
    break;
  }
  return ArgValues;
}

std::pair<SDOperand, SDOperand>
X86TargetLowering::LowerFastCCCallTo(SDOperand Chain, const Type *RetTy,
                                     bool isTailCall, SDOperand Callee,
                                     ArgListTy &Args, SelectionDAG &DAG) {
  // Count how many bytes are to be pushed on the stack.
  unsigned NumBytes = 0;

  // Keep track of the number of integer regs passed so far.  This can be either
  // 0 (neither EAX or EDX used), 1 (EAX is used) or 2 (EAX and EDX are both
  // used).
  unsigned NumIntRegs = 0;

  for (unsigned i = 0, e = Args.size(); i != e; ++i)
    switch (getValueType(Args[i].second)) {
    default: assert(0 && "Unknown value type!");
    case MVT::i1:
    case MVT::i8:
    case MVT::i16:
    case MVT::i32:
      if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
        ++NumIntRegs;
        break;
      }
      // fall through
    case MVT::f32:
      NumBytes += 4;
      break;
    case MVT::i64:
      if (NumIntRegs+2 <= FASTCC_NUM_INT_ARGS_INREGS) {
        NumIntRegs += 2;
        break;
      } else if (NumIntRegs+1 <= FASTCC_NUM_INT_ARGS_INREGS) {
        NumIntRegs = FASTCC_NUM_INT_ARGS_INREGS;
        NumBytes += 4;
        break;
      }

      // fall through
    case MVT::f64:
      NumBytes += 8;
      break;
    }

  // Make sure the instruction takes 8n+4 bytes to make sure the start of the
  // arguments and the arguments after the retaddr has been pushed are aligned.
  if ((NumBytes & 7) == 0)
    NumBytes += 4;

  Chain = DAG.getCALLSEQ_START(Chain,DAG.getConstant(NumBytes, getPointerTy()));

  // Arguments go on the stack in reverse order, as specified by the ABI.
  unsigned ArgOffset = 0;
  SDOperand StackPtr = DAG.getRegister(X86::ESP, MVT::i32);
  NumIntRegs = 0;
  std::vector<SDOperand> Stores;
  std::vector<SDOperand> RegValuesToPass;
  for (unsigned i = 0, e = Args.size(); i != e; ++i) {
    switch (getValueType(Args[i].second)) {
    default: assert(0 && "Unexpected ValueType for argument!");
    case MVT::i1:
      Args[i].first = DAG.getNode(ISD::ANY_EXTEND, MVT::i8, Args[i].first);
      // Fall through.
    case MVT::i8:
    case MVT::i16:
    case MVT::i32:
      if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
        RegValuesToPass.push_back(Args[i].first);
        ++NumIntRegs;
        break;
      }
      // Fall through
    case MVT::f32: {
      SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
      PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
      Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
                                   Args[i].first, PtrOff,
                                   DAG.getSrcValue(NULL)));
      ArgOffset += 4;
      break;
    }
    case MVT::i64:
       // Can pass (at least) part of it in regs?
      if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
        SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
                                   Args[i].first, DAG.getConstant(1, MVT::i32));
        SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
                                   Args[i].first, DAG.getConstant(0, MVT::i32));
        RegValuesToPass.push_back(Lo);
        ++NumIntRegs;
        
        // Pass both parts in regs?
        if (NumIntRegs < FASTCC_NUM_INT_ARGS_INREGS) {
          RegValuesToPass.push_back(Hi);
          ++NumIntRegs;
        } else {
          // Pass the high part in memory.
          SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
          PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
          Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
                                       Hi, PtrOff, DAG.getSrcValue(NULL)));
          ArgOffset += 4;
        }
        break;
      }
      // Fall through
    case MVT::f64:
      SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
      PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
      Stores.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
                                   Args[i].first, PtrOff,
                                   DAG.getSrcValue(NULL)));
      ArgOffset += 8;
      break;
    }
  }
  if (!Stores.empty())
    Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, Stores);

  // Make sure the instruction takes 8n+4 bytes to make sure the start of the
  // arguments and the arguments after the retaddr has been pushed are aligned.
  if ((ArgOffset & 7) == 0)
    ArgOffset += 4;

  std::vector<MVT::ValueType> RetVals;
  MVT::ValueType RetTyVT = getValueType(RetTy);

  RetVals.push_back(MVT::Other);

  // The result values produced have to be legal.  Promote the result.
  switch (RetTyVT) {
  case MVT::isVoid: break;
  default:
    RetVals.push_back(RetTyVT);
    break;
  case MVT::i1:
  case MVT::i8:
  case MVT::i16:
    RetVals.push_back(MVT::i32);
    break;
  case MVT::f32:
    if (X86ScalarSSE)
      RetVals.push_back(MVT::f32);
    else
      RetVals.push_back(MVT::f64);
    break;
  case MVT::i64:
    RetVals.push_back(MVT::i32);
    RetVals.push_back(MVT::i32);
    break;
  }

  // Build a sequence of copy-to-reg nodes chained together with token chain
  // and flag operands which copy the outgoing args into registers.
  SDOperand InFlag;
  for (unsigned i = 0, e = RegValuesToPass.size(); i != e; ++i) {
    unsigned CCReg;
    SDOperand RegToPass = RegValuesToPass[i];
    switch (RegToPass.getValueType()) {
    default: assert(0 && "Bad thing to pass in regs");
    case MVT::i8:
      CCReg = (i == 0) ? X86::AL  : X86::DL;
      break;
    case MVT::i16:
      CCReg = (i == 0) ? X86::AX  : X86::DX;
      break;
    case MVT::i32:
      CCReg = (i == 0) ? X86::EAX : X86::EDX;
      break;
    }

    Chain = DAG.getCopyToReg(Chain, CCReg, RegToPass, InFlag);
    InFlag = Chain.getValue(1);