X86ISelLowering.cpp

        X86::isUNPCKHMask(PermMask.Val))
      return Op;
  }

  // Try PSHUF* first, then SHUFP*.
  // MMX doesn't have PSHUFD but it does have PSHUFW. While it's theoretically
  // possible to shuffle a v2i32 using PSHUFW, that's not yet implemented.
  if (isMMX && NumElems == 4 && X86::isPSHUFDMask(PermMask.Val)) {
    if (V2.getOpcode() != ISD::UNDEF)
      return DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1,
                         DAG.getNode(ISD::UNDEF, VT), PermMask);
    return Op;
  }

  if (!isMMX) {
    if (Subtarget->hasSSE2() &&
        (X86::isPSHUFDMask(PermMask.Val) ||
         X86::isPSHUFHWMask(PermMask.Val) ||
         X86::isPSHUFLWMask(PermMask.Val))) {
      MVT RVT = VT;
      if (VT == MVT::v4f32) {
        RVT = MVT::v4i32;
        Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT,
                         DAG.getNode(ISD::BIT_CONVERT, RVT, V1),
                         DAG.getNode(ISD::UNDEF, RVT), PermMask);
      } else if (V2.getOpcode() != ISD::UNDEF)
        Op = DAG.getNode(ISD::VECTOR_SHUFFLE, RVT, V1,
                         DAG.getNode(ISD::UNDEF, RVT), PermMask);
      if (RVT != VT)
        Op = DAG.getNode(ISD::BIT_CONVERT, VT, Op);
      return Op;
    }

    // Binary or unary shufps.
    if (X86::isSHUFPMask(PermMask.Val) ||
        (V2.getOpcode() == ISD::UNDEF && X86::isPSHUFDMask(PermMask.Val)))
      return Op;
  }

  // Handle v8i16 specifically since SSE can do byte extraction and insertion.
  if (VT == MVT::v8i16) {
    SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(V1, V2, PermMask, DAG, *this);
    if (NewOp.Val)
      return NewOp;
  }

  // Handle all 4 wide cases with a number of shuffles except for MMX.
  if (NumElems == 4 && !isMMX)
    return LowerVECTOR_SHUFFLE_4wide(V1, V2, PermMask, VT, DAG);

  return SDValue();
}

SDValue
X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
                                                SelectionDAG &DAG) {
  MVT VT = Op.getValueType();
  if (VT.getSizeInBits() == 8) {
    SDValue Extract = DAG.getNode(X86ISD::PEXTRB, MVT::i32,
                                    Op.getOperand(0), Op.getOperand(1));
    SDValue Assert  = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
                                    DAG.getValueType(VT));
    return DAG.getNode(ISD::TRUNCATE, VT, Assert);
  } else if (VT.getSizeInBits() == 16) {
    SDValue Extract = DAG.getNode(X86ISD::PEXTRW, MVT::i32,
                                    Op.getOperand(0), Op.getOperand(1));
    SDValue Assert  = DAG.getNode(ISD::AssertZext, MVT::i32, Extract,
                                    DAG.getValueType(VT));
    return DAG.getNode(ISD::TRUNCATE, VT, Assert);
  } else if (VT == MVT::f32) {
    // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
    // the result back to FR32 register. It's only worth matching if the
    // result has a single use which is a store or a bitcast to i32.
    if (!Op.hasOneUse())
      return SDValue();
    SDNode *User = *Op.Val->use_begin();
    if (User->getOpcode() != ISD::STORE &&
        (User->getOpcode() != ISD::BIT_CONVERT ||
         User->getValueType(0) != MVT::i32))
      return SDValue();
    SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
                    DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Op.getOperand(0)),
                                    Op.getOperand(1));
    return DAG.getNode(ISD::BIT_CONVERT, MVT::f32, Extract);
  }
  return SDValue();
}


SDValue
X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
  if (!isa<ConstantSDNode>(Op.getOperand(1)))
    return SDValue();

  if (Subtarget->hasSSE41()) {
    SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
    if (Res.Val)
      return Res;
  }

  MVT VT = Op.getValueType();
  // TODO: handle v16i8.
  if (VT.getSizeInBits() == 16) {
    SDValue Vec = Op.getOperand(0);
    unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
    if (Idx == 0)
      return DAG.getNode(ISD::TRUNCATE, MVT::i16,
                         DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::i32,
                                 DAG.getNode(ISD::BIT_CONVERT, MVT::v4i32, Vec),
                                     Op.getOperand(1)));
    // Transform it so it match pextrw which produces a 32-bit result.
    MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
    SDValue Extract = DAG.getNode(X86ISD::PEXTRW, EVT,
                                    Op.getOperand(0), Op.getOperand(1));
    SDValue Assert  = DAG.getNode(ISD::AssertZext, EVT, Extract,
                                    DAG.getValueType(VT));
    return DAG.getNode(ISD::TRUNCATE, VT, Assert);
  } else if (VT.getSizeInBits() == 32) {
    unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
    if (Idx == 0)
      return Op;
    // SHUFPS the element to the lowest double word, then movss.
    MVT MaskVT = MVT::getIntVectorWithNumElements(4);
    SmallVector<SDValue, 8> IdxVec;
    IdxVec.
      push_back(DAG.getConstant(Idx, MaskVT.getVectorElementType()));
    IdxVec.
      push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
    IdxVec.
      push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
    IdxVec.
      push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
    SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
                                 &IdxVec[0], IdxVec.size());
    SDValue Vec = Op.getOperand(0);
    Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
                      Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
                       DAG.getIntPtrConstant(0));
  } else if (VT.getSizeInBits() == 64) {
    // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
    // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
    //        to match extract_elt for f64.
    unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getValue();
    if (Idx == 0)
      return Op;

    // UNPCKHPD the element to the lowest double word, then movsd.
    // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
    // to a f64mem, the whole operation is folded into a single MOVHPDmr.
    MVT MaskVT = MVT::getIntVectorWithNumElements(2);
    SmallVector<SDValue, 8> IdxVec;
    IdxVec.push_back(DAG.getConstant(1, MaskVT.getVectorElementType()));
    IdxVec.
      push_back(DAG.getNode(ISD::UNDEF, MaskVT.getVectorElementType()));
    SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, MaskVT,
                                 &IdxVec[0], IdxVec.size());
    SDValue Vec = Op.getOperand(0);
    Vec = DAG.getNode(ISD::VECTOR_SHUFFLE, Vec.getValueType(),
                      Vec, DAG.getNode(ISD::UNDEF, Vec.getValueType()), Mask);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, VT, Vec,
                       DAG.getIntPtrConstant(0));
  }

  return SDValue();
}

SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
  MVT VT = Op.getValueType();
  MVT EVT = VT.getVectorElementType();

  SDValue N0 = Op.getOperand(0);
  SDValue N1 = Op.getOperand(1);
  SDValue N2 = Op.getOperand(2);

  if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
      isa<ConstantSDNode>(N2)) {
    unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
                                                  : X86ISD::PINSRW;
    // Transform it so it match pinsr{b,w} which expects a GR32 as its second
    // argument.
    if (N1.getValueType() != MVT::i32)
      N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
    if (N2.getValueType() != MVT::i32)
      N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
    return DAG.getNode(Opc, VT, N0, N1, N2);
  } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
    // Bits [7:6] of the constant are the source select.  This will always be
    //  zero here.  The DAG Combiner may combine an extract_elt index into these
    //  bits.  For example (insert (extract, 3), 2) could be matched by putting
    //  the '3' into bits [7:6] of X86ISD::INSERTPS.
    // Bits [5:4] of the constant are the destination select.  This is the 
    //  value of the incoming immediate.
    // Bits [3:0] of the constant are the zero mask.  The DAG Combiner may 
    //   combine either bitwise AND or insert of float 0.0 to set these bits.
    N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue() << 4);
    return DAG.getNode(X86ISD::INSERTPS, VT, N0, N1, N2);
  }
  return SDValue();
}

SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
  MVT VT = Op.getValueType();
  MVT EVT = VT.getVectorElementType();

  if (Subtarget->hasSSE41())
    return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);

  if (EVT == MVT::i8)
    return SDValue();

  SDValue N0 = Op.getOperand(0);
  SDValue N1 = Op.getOperand(1);
  SDValue N2 = Op.getOperand(2);

  if (EVT.getSizeInBits() == 16) {
    // Transform it so it match pinsrw which expects a 16-bit value in a GR32
    // as its second argument.
    if (N1.getValueType() != MVT::i32)
      N1 = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, N1);
    if (N2.getValueType() != MVT::i32)
      N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getValue());
    return DAG.getNode(X86ISD::PINSRW, VT, N0, N1, N2);
  }
  return SDValue();
}

SDValue
X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
  if (Op.getValueType() == MVT::v2f32)
    return DAG.getNode(ISD::BIT_CONVERT, MVT::v2f32,
                       DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2i32,
                                   DAG.getNode(ISD::BIT_CONVERT, MVT::i32,
                                               Op.getOperand(0))));

  SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, Op.getOperand(0));
  MVT VT = MVT::v2i32;
  switch (Op.getValueType().getSimpleVT()) {
  default: break;
  case MVT::v16i8:
  case MVT::v8i16:
    VT = MVT::v4i32;
    break;
  }
  return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(),
                     DAG.getNode(ISD::SCALAR_TO_VECTOR, VT, AnyExt));
}

// ConstantPool, JumpTable, 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.
SDValue
X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
  SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(),
                                               getPointerTy(),
                                               CP->getAlignment());
  Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
  // With PIC, the address is actually $g + Offset.
  if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
      !Subtarget->isPICStyleRIPRel()) {
    Result = DAG.getNode(ISD::ADD, getPointerTy(),
                         DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
                         Result);
  }

  return Result;
}

SDValue
X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
  GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
  SDValue Result = DAG.getTargetGlobalAddress(GV, getPointerTy());
  Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
  // With PIC, the address is actually $g + Offset.
  if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
      !Subtarget->isPICStyleRIPRel()) {
    Result = DAG.getNode(ISD::ADD, getPointerTy(),
                         DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
                         Result);
  }
  
  // For Darwin & Mingw32, 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. Platform check is inside GVRequiresExtraLoad() call
  // The same applies for external symbols during PIC codegen
  if (Subtarget->GVRequiresExtraLoad(GV, getTargetMachine(), false))
    Result = DAG.getLoad(getPointerTy(), DAG.getEntryNode(), Result,
                         PseudoSourceValue::getGOT(), 0);

  return Result;
}

// Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
static SDValue
LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
                                const MVT PtrVT) {
  SDValue InFlag;
  SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), X86::EBX,
                                     DAG.getNode(X86ISD::GlobalBaseReg,
                                                 PtrVT), InFlag);
  InFlag = Chain.getValue(1);

  // emit leal symbol@TLSGD(,%ebx,1), %eax
  SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
  SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
                                             GA->getValueType(0),
                                             GA->getOffset());
  SDValue Ops[] = { Chain,  TGA, InFlag };
  SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 3);
  InFlag = Result.getValue(2);
  Chain = Result.getValue(1);

  // call ___tls_get_addr. This function receives its argument in
  // the register EAX.
  Chain = DAG.getCopyToReg(Chain, X86::EAX, Result, InFlag);
  InFlag = Chain.getValue(1);

  NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
  SDValue Ops1[] = { Chain,
                      DAG.getTargetExternalSymbol("___tls_get_addr",
                                                  PtrVT),
                      DAG.getRegister(X86::EAX, PtrVT),
                      DAG.getRegister(X86::EBX, PtrVT),
                      InFlag };
  Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 5);
  InFlag = Chain.getValue(1);

  return DAG.getCopyFromReg(Chain, X86::EAX, PtrVT, InFlag);
}

// Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
static SDValue
LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
                                const MVT PtrVT) {
  SDValue InFlag, Chain;

  // emit leaq symbol@TLSGD(%rip), %rdi
  SDVTList NodeTys = DAG.getVTList(PtrVT, MVT::Other, MVT::Flag);
  SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
                                             GA->getValueType(0),
                                             GA->getOffset());
  SDValue Ops[]  = { DAG.getEntryNode(), TGA};
  SDValue Result = DAG.getNode(X86ISD::TLSADDR, NodeTys, Ops, 2);
  Chain  = Result.getValue(1);
  InFlag = Result.getValue(2);

  // call __tls_get_addr. This function receives its argument in
  // the register RDI.
  Chain = DAG.getCopyToReg(Chain, X86::RDI, Result, InFlag);
  InFlag = Chain.getValue(1);

  NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
  SDValue Ops1[] = { Chain,
                      DAG.getTargetExternalSymbol("__tls_get_addr",
                                                  PtrVT),
                      DAG.getRegister(X86::RDI, PtrVT),
                      InFlag };
  Chain = DAG.getNode(X86ISD::CALL, NodeTys, Ops1, 4);
  InFlag = Chain.getValue(1);

  return DAG.getCopyFromReg(Chain, X86::RAX, PtrVT, InFlag);
}

// Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
// "local exec" model.
static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
                                     const MVT PtrVT) {
  // Get the Thread Pointer
  SDValue ThreadPointer = DAG.getNode(X86ISD::THREAD_POINTER, PtrVT);
  // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
  // exec)
  SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
                                             GA->getValueType(0),
                                             GA->getOffset());
  SDValue Offset = DAG.getNode(X86ISD::Wrapper, PtrVT, TGA);

  if (GA->getGlobal()->isDeclaration()) // initial exec TLS model
    Offset = DAG.getLoad(PtrVT, DAG.getEntryNode(), Offset,
                         PseudoSourceValue::getGOT(), 0);

  // The address of the thread local variable is the add of the thread
  // pointer with the offset of the variable.
  return DAG.getNode(ISD::ADD, PtrVT, ThreadPointer, Offset);
}

SDValue
X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
  // TODO: implement the "local dynamic" model
  // TODO: implement the "initial exec"model for pic executables
  assert(Subtarget->isTargetELF() &&
         "TLS not implemented for non-ELF targets");
  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
  // If the relocation model is PIC, use the "General Dynamic" TLS Model,
  // otherwise use the "Local Exec"TLS Model
  if (Subtarget->is64Bit()) {
    return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
  } else {
    if (getTargetMachine().getRelocationModel() == Reloc::PIC_)
      return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
    else
      return LowerToTLSExecModel(GA, DAG, getPointerTy());
  }
}

SDValue
X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
  const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
  SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy());
  Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
  // With PIC, the address is actually $g + Offset.
  if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
      !Subtarget->isPICStyleRIPRel()) {
    Result = DAG.getNode(ISD::ADD, getPointerTy(),
                         DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
                         Result);
  }

  return Result;
}

SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
  JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
  SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy());
  Result = DAG.getNode(X86ISD::Wrapper, getPointerTy(), Result);
  // With PIC, the address is actually $g + Offset.
  if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
      !Subtarget->isPICStyleRIPRel()) {
    Result = DAG.getNode(ISD::ADD, getPointerTy(),
                         DAG.getNode(X86ISD::GlobalBaseReg, getPointerTy()),
                         Result);
  }

  return Result;
}

/// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
/// take a 2 x i32 value to shift plus a shift amount. 
SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
  MVT VT = Op.getValueType();
  unsigned VTBits = VT.getSizeInBits();
  bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
  SDValue ShOpLo = Op.getOperand(0);
  SDValue ShOpHi = Op.getOperand(1);
  SDValue ShAmt  = Op.getOperand(2);
  SDValue Tmp1 = isSRA ?
    DAG.getNode(ISD::SRA, VT, ShOpHi, DAG.getConstant(VTBits - 1, MVT::i8)) :
    DAG.getConstant(0, VT);

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

  SDValue AndNode = DAG.getNode(ISD::AND, MVT::i8, ShAmt,
                                  DAG.getConstant(VTBits, MVT::i8));
  SDValue Cond = DAG.getNode(X86ISD::CMP, VT,
                               AndNode, DAG.getConstant(0, MVT::i8));

  SDValue Hi, Lo;
  SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
  SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
  SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };

  if (Op.getOpcode() == ISD::SHL_PARTS) {
    Hi = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
    Lo = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
  } else {
    Lo = DAG.getNode(X86ISD::CMOV, VT, Ops0, 4);
    Hi = DAG.getNode(X86ISD::CMOV, VT, Ops1, 4);
  }

  SDValue Ops[2] = { Lo, Hi };
  return DAG.getMergeValues(Ops, 2);
}

SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
  MVT SrcVT = Op.getOperand(0).getValueType();
  assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
         "Unknown SINT_TO_FP to lower!");
  
  // These are really Legal; caller falls through into that case.
  if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
    return SDValue();
  if (SrcVT == MVT::i64 && Op.getValueType() != MVT::f80 && 
      Subtarget->is64Bit())
    return SDValue();
  
  unsigned Size = SrcVT.getSizeInBits()/8;
  MachineFunction &MF = DAG.getMachineFunction();
  int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
  SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
  SDValue Chain = DAG.getStore(DAG.getEntryNode(), Op.getOperand(0),
                                 StackSlot,
                                 PseudoSourceValue::getFixedStack(SSFI), 0);

  // Build the FILD
  SDVTList Tys;
  bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
  if (useSSE)
    Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
  else
    Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
  SmallVector<SDValue, 8> Ops;
  Ops.push_back(Chain);
  Ops.push_back(StackSlot);
  Ops.push_back(DAG.getValueType(SrcVT));
  SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD,
                                 Tys, &Ops[0], Ops.size());

  if (useSSE) {
    Chain = Result.getValue(1);
    SDValue 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);
    SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
    Tys = DAG.getVTList(MVT::Other);
    SmallVector<SDValue, 8> 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[0], Ops.size());
    Result = DAG.getLoad(Op.getValueType(), Chain, StackSlot,
                         PseudoSourceValue::getFixedStack(SSFI), 0);
  }

  return Result;
}

std::pair<SDValue,SDValue> X86TargetLowering::
FP_TO_SINTHelper(SDValue Op, SelectionDAG &DAG) {
  assert(Op.getValueType().getSimpleVT() <= MVT::i64 &&
         Op.getValueType().getSimpleVT() >= MVT::i16 &&
         "Unknown FP_TO_SINT to lower!");

  // These are really Legal.
  if (Op.getValueType() == MVT::i32 && 
      isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
    return std::make_pair(SDValue(), SDValue());
  if (Subtarget->is64Bit() &&
      Op.getValueType() == MVT::i64 &&
      Op.getOperand(0).getValueType() != MVT::f80)
    return std::make_pair(SDValue(), SDValue());

  // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
  // stack slot.
  MachineFunction &MF = DAG.getMachineFunction();
  unsigned MemSize = Op.getValueType().getSizeInBits()/8;
  int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
  SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
  unsigned Opc;
  switch (Op.getValueType().getSimpleVT()) {
  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;
  }

  SDValue Chain = DAG.getEntryNode();
  SDValue Value = Op.getOperand(0);
  if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
    assert(Op.getValueType() == MVT::i64 && "Invalid FP_TO_SINT to lower!");
    Chain = DAG.getStore(Chain, Value, StackSlot,
                         PseudoSourceValue::getFixedStack(SSFI), 0);
    SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
    SDValue Ops[] = {
      Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
    };
    Value = DAG.getNode(X86ISD::FLD, Tys, Ops, 3);
    Chain = Value.getValue(1);
    SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
    StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
  }

  // Build the FP_TO_INT*_IN_MEM
  SDValue Ops[] = { Chain, Value, StackSlot };
  SDValue FIST = DAG.getNode(Opc, MVT::Other, Ops, 3);

  return std::make_pair(FIST, StackSlot);
}

SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
  std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(Op, DAG);
  SDValue FIST = Vals.first, StackSlot = Vals.second;
  if (FIST.Val == 0) return SDValue();
  
  // Load the result.
  return DAG.getLoad(Op.getValueType(), FIST, StackSlot, NULL, 0);
}

SDNode *X86TargetLowering::ExpandFP_TO_SINT(SDNode *N, SelectionDAG &DAG) {
  std::pair<SDValue,SDValue> Vals = FP_TO_SINTHelper(SDValue(N, 0), DAG);
  SDValue FIST = Vals.first, StackSlot = Vals.second;
  if (FIST.Val == 0) return 0;

  MVT VT = N->getValueType(0);

  // Return a load from the stack slot.
  SDValue Res = DAG.getLoad(VT, FIST, StackSlot, NULL, 0);

  // Use MERGE_VALUES to drop the chain result value and get a node with one
  // result.  This requires turning off getMergeValues simplification, since
  // otherwise it will give us Res back.
  return DAG.getMergeValues(&Res, 1, false).Val;
}

SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
  MVT VT = Op.getValueType();
  MVT EltVT = VT;
  if (VT.isVector())
    EltVT = VT.getVectorElementType();
  std::vector<Constant*> CV;
  if (EltVT == MVT::f64) {
    Constant *C = ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63))));
    CV.push_back(C);
    CV.push_back(C);
  } else {
    Constant *C = ConstantFP::get(APFloat(APInt(32, ~(1U << 31))));
    CV.push_back(C);
    CV.push_back(C);
    CV.push_back(C);
    CV.push_back(C);
  }
  Constant *C = ConstantVector::get(CV);
  SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
  SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
                               PseudoSourceValue::getConstantPool(), 0,
                               false, 16);
  return DAG.getNode(X86ISD::FAND, VT, Op.getOperand(0), Mask);
}

SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
  MVT VT = Op.getValueType();
  MVT EltVT = VT;
  unsigned EltNum = 1;
  if (VT.isVector()) {
    EltVT = VT.getVectorElementType();
    EltNum = VT.getVectorNumElements();
  }
  std::vector<Constant*> CV;
  if (EltVT == MVT::f64) {
    Constant *C = ConstantFP::get(APFloat(APInt(64, 1ULL << 63)));
    CV.push_back(C);
    CV.push_back(C);
  } else {
    Constant *C = ConstantFP::get(APFloat(APInt(32, 1U << 31)));
    CV.push_back(C);
    CV.push_back(C);
    CV.push_back(C);
    CV.push_back(C);
  }
  Constant *C = ConstantVector::get(CV);
  SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
  SDValue Mask = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
                               PseudoSourceValue::getConstantPool(), 0,
                               false, 16);
  if (VT.isVector()) {
    return DAG.getNode(ISD::BIT_CONVERT, VT,
                       DAG.getNode(ISD::XOR, MVT::v2i64,
                    DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Op.getOperand(0)),
                    DAG.getNode(ISD::BIT_CONVERT, MVT::v2i64, Mask)));
  } else {
    return DAG.getNode(X86ISD::FXOR, VT, Op.getOperand(0), Mask);
  }
}

SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  MVT VT = Op.getValueType();
  MVT SrcVT = Op1.getValueType();

  // If second operand is smaller, extend it first.
  if (SrcVT.bitsLT(VT)) {
    Op1 = DAG.getNode(ISD::FP_EXTEND, VT, Op1);
    SrcVT = VT;
  }
  // And if it is bigger, shrink it first.
  if (SrcVT.bitsGT(VT)) {
    Op1 = DAG.getNode(ISD::FP_ROUND, VT, Op1, DAG.getIntPtrConstant(1));
    SrcVT = VT;
  }

  // At this point the operands and the result should have the same
  // type, and that won't be f80 since that is not custom lowered.

  // First get the sign bit of second operand.
  std::vector<Constant*> CV;
  if (SrcVT == MVT::f64) {
    CV.push_back(ConstantFP::get(APFloat(APInt(64, 1ULL << 63))));
    CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
  } else {
    CV.push_back(ConstantFP::get(APFloat(APInt(32, 1U << 31))));
    CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
    CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
    CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
  }
  Constant *C = ConstantVector::get(CV);
  SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
  SDValue Mask1 = DAG.getLoad(SrcVT, DAG.getEntryNode(), CPIdx,
                                PseudoSourceValue::getConstantPool(), 0,
                                false, 16);
  SDValue SignBit = DAG.getNode(X86ISD::FAND, SrcVT, Op1, Mask1);

  // Shift sign bit right or left if the two operands have different types.
  if (SrcVT.bitsGT(VT)) {
    // Op0 is MVT::f32, Op1 is MVT::f64.
    SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, MVT::v2f64, SignBit);
    SignBit = DAG.getNode(X86ISD::FSRL, MVT::v2f64, SignBit,
                          DAG.getConstant(32, MVT::i32));
    SignBit = DAG.getNode(ISD::BIT_CONVERT, MVT::v4f32, SignBit);
    SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, MVT::f32, SignBit,
                          DAG.getIntPtrConstant(0));
  }

  // Clear first operand sign bit.
  CV.clear();
  if (VT == MVT::f64) {
    CV.push_back(ConstantFP::get(APFloat(APInt(64, ~(1ULL << 63)))));
    CV.push_back(ConstantFP::get(APFloat(APInt(64, 0))));
  } else {
    CV.push_back(ConstantFP::get(APFloat(APInt(32, ~(1U << 31)))));
    CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
    CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
    CV.push_back(ConstantFP::get(APFloat(APInt(32, 0))));
  }
  C = ConstantVector::get(CV);
  CPIdx = DAG.getConstantPool(C, getPointerTy(), 4);
  SDValue Mask2 = DAG.getLoad(VT, DAG.getEntryNode(), CPIdx,
                                PseudoSourceValue::getConstantPool(), 0,
                                false, 16);
  SDValue Val = DAG.getNode(X86ISD::FAND, VT, Op0, Mask2);

  // Or the value with the sign bit.
  return DAG.getNode(X86ISD::FOR, VT, Val, SignBit);
}

SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
  assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
  SDValue Cond;
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  SDValue CC = Op.getOperand(2);
  ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
  bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
  unsigned X86CC;

  if (translateX86CC(cast<CondCodeSDNode>(CC)->get(), isFP, X86CC,
                     Op0, Op1, DAG)) {
    Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
    return DAG.getNode(X86ISD::SETCC, MVT::i8,
                       DAG.getConstant(X86CC, MVT::i8), Cond);
  }

  assert(isFP && "Illegal integer SetCC!");

  Cond = DAG.getNode(X86ISD::CMP, MVT::i32, Op0, Op1);
  switch (SetCCOpcode) {
  default: assert(false && "Illegal floating point SetCC!");
  case ISD::SETOEQ: {  // !PF & ZF
    SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
                                 DAG.getConstant(X86::COND_NP, MVT::i8), Cond);
    SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
                                 DAG.getConstant(X86::COND_E, MVT::i8), Cond);
    return DAG.getNode(ISD::AND, MVT::i8, Tmp1, Tmp2);
  }
  case ISD::SETUNE: {  // PF | !ZF
    SDValue Tmp1 = DAG.getNode(X86ISD::SETCC, MVT::i8,
                                 DAG.getConstant(X86::COND_P, MVT::i8), Cond);
    SDValue Tmp2 = DAG.getNode(X86ISD::SETCC, MVT::i8,
                                 DAG.getConstant(X86::COND_NE, MVT::i8), Cond);
    return DAG.getNode(ISD::OR, MVT::i8, Tmp1, Tmp2);
  }
  }
}

SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
  SDValue Cond;
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  SDValue CC = Op.getOperand(2);
  MVT VT = Op.getValueType();
  ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
  bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();

  if (isFP) {
    unsigned SSECC = 8;
    MVT VT0 = Op0.getValueType();
    assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
    unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
    bool Swap = false;

    switch (SetCCOpcode) {
    default: break;
    case ISD::SETOEQ:
    case ISD::SETEQ:  SSECC = 0; break;
    case ISD::SETOGT: 
    case ISD::SETGT: Swap = true; // Fallthrough
    case ISD::SETLT:
    case ISD::SETOLT: SSECC = 1; break;
    case ISD::SETOGE:
    case ISD::SETGE: Swap = true; // Fallthrough
    case ISD::SETLE:
    case ISD::SETOLE: SSECC = 2; break;
    case ISD::SETUO:  SSECC = 3; break;
    case ISD::SETUNE:
    case ISD::SETNE:  SSECC = 4; break;
    case ISD::SETULE: Swap = true;
    case ISD::SETUGE: SSECC = 5; break;
    case ISD::SETULT: Swap = true;
    case ISD::SETUGT: SSECC = 6; break;
    case ISD::SETO:   SSECC = 7; break;
    }
    if (Swap)
      std::swap(Op0, Op1);

    // In the two special cases we can't handle, emit two comparisons.
    if (SSECC == 8) {
      if (SetCCOpcode == ISD::SETUEQ) {
        SDValue UNORD, EQ;
        UNORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
        EQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
        return DAG.getNode(ISD::OR, VT, UNORD, EQ);
      }
      else if (SetCCOpcode == ISD::SETONE) {
        SDValue ORD, NEQ;
        ORD = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
        NEQ = DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
        return DAG.getNode(ISD::AND, VT, ORD, NEQ);
      }
      assert(0 && "Illegal FP comparison");
    }
    // Handle all other FP comparisons here.
    return DAG.getNode(Opc, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
  }
  
  // We are handling one of the integer comparisons here.  Since SSE only has
  // GT and EQ comparisons for integer, swapping operands and multiple
  // operations may be required for some comparisons.
  unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
  bool Swap = false, Invert = false, FlipSigns = false;
  
  switch (VT.getSimpleVT()) {
  default: break;
  case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
  case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
  case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
  case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
  }
  
  switch (SetCCOpcode) {
  default: break;
  case ISD::SETNE:  Invert = true;
  case ISD::SETEQ:  Opc = EQOpc; break;
  case ISD::SETLT:  Swap = true;
  case ISD::SETGT:  Opc = GTOpc; break;
  case ISD::SETGE:  Swap = true;
  case ISD::SETLE:  Opc = GTOpc; Invert = true; break;
  case ISD::SETULT: Swap = true;
  case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
  case ISD::SETUGE: Swap = true;
  case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
  }
  if (Swap)
    std::swap(Op0, Op1);
  
  // Since SSE has no unsigned integer comparisons, we need to flip  the sign
  // bits of the inputs before performing those operations.
  if (FlipSigns) {
    MVT EltVT = VT.getVectorElementType();
    SDValue SignBit = DAG.getConstant(EltVT.getIntegerVTSignBit(), EltVT);
    std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
    SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, VT, &SignBits[0],
                                    SignBits.size());
    Op0 = DAG.getNode(ISD::XOR, VT, Op0, SignVec);
    Op1 = DAG.getNode(ISD::XOR, VT, Op1, SignVec);
  }
  
  SDValue Result = DAG.getNode(Opc, VT, Op0, Op1);

  // If the logical-not of the result is required, perform that now.
  if (Invert) {
    MVT EltVT = VT.getVectorElementType();
    SDValue NegOne = DAG.getConstant(EltVT.getIntegerVTBitMask(), EltVT);
    std::vector<SDValue> NegOnes(VT.getVectorNumElements(), NegOne);
    SDValue NegOneV = DAG.getNode(ISD::BUILD_VECTOR, VT, &NegOnes[0],
                                    NegOnes.size());
    Result = DAG.getNode(ISD::XOR, VT, Result, NegOneV);
  }
  return Result;
}

SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
  bool addTest = true;
  SDValue Cond  = Op.getOperand(0);
  SDValue CC;

  if (Cond.getOpcode() == ISD::SETCC)
    Cond = LowerSETCC(Cond, DAG);

  // If condition flag is set by a X86ISD::CMP, then use it as the condition
  // setting operand in place of the X86ISD::SETCC.
  if (Cond.getOpcode() == X86ISD::SETCC) {
    CC = Cond.getOperand(0);

    SDValue Cmp = Cond.getOperand(1);
    unsigned Opc = Cmp.getOpcode();
    MVT VT = Op.getValueType();
    
    bool IllegalFPCMov = false;
    if (VT.isFloatingPoint() && !VT.isVector() &&
        !isScalarFPTypeInSSEReg(VT))  // FPStack?
      IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSignExtended());
    
    if ((Opc == X86ISD::CMP ||
         Opc == X86ISD::COMI ||
         Opc == X86ISD::UCOMI) && !IllegalFPCMov) {
      Cond = Cmp;
      addTest = false;
    }
  }

  if (addTest) {
    CC = DAG.getConstant(X86::COND_NE, MVT::i8);
    Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
  }

  const MVT *VTs = DAG.getNodeValueTypes(Op.getValueType(),
                                                    MVT::Flag);
  SmallVector<SDValue, 4> 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, VTs, 2, &Ops[0], Ops.size());
}

SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
  bool addTest = true;
  SDValue Chain = Op.getOperand(0);
  SDValue Cond  = Op.getOperand(1);
  SDValue Dest  = Op.getOperand(2);
  SDValue CC;

  if (Cond.getOpcode() == ISD::SETCC)
    Cond = LowerSETCC(Cond, DAG);

  // If condition flag is set by a X86ISD::CMP, then use it as the condition
  // setting operand in place of the X86ISD::SETCC.
  if (Cond.getOpcode() == X86ISD::SETCC) {
    CC = Cond.getOperand(0);

    SDValue Cmp = Cond.getOperand(1);
    unsigned Opc = Cmp.getOpcode();
    if (Opc == X86ISD::CMP ||
        Opc == X86ISD::COMI ||
        Opc == X86ISD::UCOMI) {
      Cond = Cmp;
      addTest = false;
    }
  }

  if (addTest) {
    CC = DAG.getConstant(X86::COND_NE, MVT::i8);
    Cond= DAG.getNode(X86ISD::CMP, MVT::i32, Cond, DAG.getConstant(0, MVT::i8));
  }
  return DAG.getNode(X86ISD::BRCOND, Op.getValueType(),
                     Chain, Op.getOperand(2), CC, Cond);
}


// Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
// Calls to _alloca is needed to probe the stack when allocating more than 4k
// bytes in one go. Touching the stack at 4K increments is necessary to ensure
// that the guard pages used by the OS virtual memory manager are allocated in
// correct sequence.
SDValue
X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
                                           SelectionDAG &DAG) {
  assert(Subtarget->isTargetCygMing() &&
         "This should be used only on Cygwin/Mingw targets");