case X86Subtarget::isDarwin:
if (TM.getSubtarget<X86Subtarget>().is64Bit())
return new X8664_MachoTargetObjectFile();
- return new TargetLoweringObjectFileMachO();
+ return new X8632_MachoTargetObjectFile();
case X86Subtarget::isELF:
return new TargetLoweringObjectFileELF();
case X86Subtarget::isMingw:
if (Subtarget->is64Bit())
setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
+ setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
if (Subtarget->is64Bit()) {
setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
+ setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
}
// 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
}
- // Use the default ISD::DBG_STOPPOINT.
- setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
// FIXME - use subtarget debug flags
if (!Subtarget->isTargetDarwin() &&
!Subtarget->isTargetELF() &&
!Subtarget->isTargetCygMing()) {
- setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
}
setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
+ setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
+ for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
+ InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
+ setTruncStoreAction((MVT::SimpleValueType)VT,
+ (MVT::SimpleValueType)InnerVT, Expand);
+ setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
+ setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
+ setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
}
// FIXME: In order to prevent SSE instructions being expanded to MMX ones
setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
- setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
- setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
setTargetDAGCombine(ISD::SRL);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::MEMBARRIER);
+ setTargetDAGCombine(ISD::ZERO_EXTEND);
if (Subtarget->is64Bit())
setTargetDAGCombine(ISD::MUL);
computeRegisterProperties();
+ // Divide and reminder operations have no vector equivalent and can
+ // trap. Do a custom widening for these operations in which we never
+ // generate more divides/remainder than the original vector width.
+ for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
+ VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
+ if (!isTypeLegal((MVT::SimpleValueType)VT)) {
+ setOperationAction(ISD::SDIV, (MVT::SimpleValueType) VT, Custom);
+ setOperationAction(ISD::UDIV, (MVT::SimpleValueType) VT, Custom);
+ setOperationAction(ISD::SREM, (MVT::SimpleValueType) VT, Custom);
+ setOperationAction(ISD::UREM, (MVT::SimpleValueType) VT, Custom);
+ }
+ }
+
// 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
#include "X86GenCallingConv.inc"
+bool
+X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
+ const SmallVectorImpl<EVT> &OutTys,
+ const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
+ SelectionDAG &DAG) {
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
+ RVLocs, *DAG.getContext());
+ return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
+}
+
SDValue
X86TargetLowering::LowerReturn(SDValue Chain,
CallingConv::ID CallConv, bool isVarArg,
Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
Flag = Chain.getValue(1);
+
+ // RAX now acts like a return value.
+ MF.getRegInfo().addLiveOut(X86::RAX);
}
RetOps[0] = Chain; // Update chain.
// In case of tail call optimization mark all arguments mutable. Since they
// could be overwritten by lowering of arguments in case of a tail call.
int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
- VA.getLocMemOffset(), isImmutable);
+ VA.getLocMemOffset(), isImmutable, false);
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
if (Flags.isByVal())
return FIN;
// the start of the first vararg value... for expansion of llvm.va_start.
if (isVarArg) {
if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
- VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
+ VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
}
if (Is64Bit) {
unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
VarArgsGPOffset = NumIntRegs * 8;
VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
- TotalNumXMMRegs * 16, 16);
+ TotalNumXMMRegs * 16, 16,
+ false);
// Store the integer parameter registers.
SmallVector<SDValue, 8> MemOps;
// Calculate the new stack slot for the return address.
int SlotSize = Is64Bit ? 8 : 4;
int NewReturnAddrFI =
- MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
+ MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize,
+ true, false);
EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
// Create frame index.
int32_t Offset = VA.getLocMemOffset()+FPDiff;
uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
- FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
+ FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
FIN = DAG.getFrameIndex(FI, getPointerTy());
if (Flags.isByVal()) {
FPDiff, dl);
}
- // 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)) {
+ bool WasGlobalOrExternal = false;
+ if (getTargetMachine().getCodeModel() == CodeModel::Large) {
+ assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
+ // In the 64-bit large code model, we have to make all calls
+ // through a register, since the call instruction's 32-bit
+ // pc-relative offset may not be large enough to hold the whole
+ // address.
+ } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
+ WasGlobalOrExternal = true;
+ // 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.
+
// We should use extra load for direct calls to dllimported functions in
// non-JIT mode.
GlobalValue *GV = G->getGlobal();
G->getOffset(), OpFlags);
}
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
+ WasGlobalOrExternal = true;
unsigned char OpFlags = 0;
// On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
OpFlags);
- } else if (isTailCall) {
+ }
+
+ if (isTailCall && !WasGlobalOrExternal) {
unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
Chain = DAG.getCopyToReg(Chain, dl,
if (ReturnAddrIndex == 0) {
// Set up a frame object for the return address.
uint64_t SlotSize = TD->getPointerSize();
- ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
+ ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
+ true, false);
FuncInfo->setRAIndex(ReturnAddrIndex);
}
case ISD::SETNE: return X86::COND_NE;
case ISD::SETUO: return X86::COND_P;
case ISD::SETO: return X86::COND_NP;
+ case ISD::SETOEQ:
+ case ISD::SETUNE: return X86::COND_INVALID;
}
}
}
}
+/// isFPImmLegal - Returns true if the target can instruction select the
+/// specified FP immediate natively. If false, the legalizer will
+/// materialize the FP immediate as a load from a constant pool.
+bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
+ for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
+ if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
+ return true;
+ }
+ return false;
+}
+
/// isUndefOrInRange - Return true if Val is undef or if its value falls within
/// the specified range (L, H].
static bool isUndefOrInRange(int Val, int Low, int Hi) {
return ::isPSHUFLWMask(M, N->getValueType(0));
}
+/// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
+/// is suitable for input to PALIGNR.
+static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool hasSSSE3) {
+ int i, e = VT.getVectorNumElements();
+
+ // Do not handle v2i64 / v2f64 shuffles with palignr.
+ if (e < 4 || !hasSSSE3)
+ return false;
+
+ for (i = 0; i != e; ++i)
+ if (Mask[i] >= 0)
+ break;
+
+ // All undef, not a palignr.
+ if (i == e)
+ return false;
+
+ // Determine if it's ok to perform a palignr with only the LHS, since we
+ // don't have access to the actual shuffle elements to see if RHS is undef.
+ bool Unary = Mask[i] < (int)e;
+ bool NeedsUnary = false;
+
+ int s = Mask[i] - i;
+
+ // Check the rest of the elements to see if they are consecutive.
+ for (++i; i != e; ++i) {
+ int m = Mask[i];
+ if (m < 0)
+ continue;
+
+ Unary = Unary && (m < (int)e);
+ NeedsUnary = NeedsUnary || (m < s);
+
+ if (NeedsUnary && !Unary)
+ return false;
+ if (Unary && m != ((s+i) & (e-1)))
+ return false;
+ if (!Unary && m != (s+i))
+ return false;
+ }
+ return true;
+}
+
+bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
+ SmallVector<int, 8> M;
+ N->getMask(M);
+ return ::isPALIGNRMask(M, N->getValueType(0), true);
+}
+
/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to SHUFP*.
static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
isUndefOrEqual(N->getMaskElt(3), 3);
}
+/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
+/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
+/// <2, 3, 2, 3>
+bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
+ unsigned NumElems = N->getValueType(0).getVectorNumElements();
+
+ if (NumElems != 4)
+ return false;
+
+ return isUndefOrEqual(N->getMaskElt(0), 2) &&
+ isUndefOrEqual(N->getMaskElt(1), 3) &&
+ isUndefOrEqual(N->getMaskElt(2), 2) &&
+ isUndefOrEqual(N->getMaskElt(3), 3);
+}
+
/// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
return true;
}
-/// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
-/// and MOVLHPS.
-bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
+/// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a shuffle of elements that is suitable for input to MOVLHPS.
+bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
unsigned NumElems = N->getValueType(0).getVectorNumElements();
if (NumElems != 2 && NumElems != 4)
return true;
}
-/// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
-/// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
-/// <2, 3, 2, 3>
-bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
- unsigned NumElems = N->getValueType(0).getVectorNumElements();
-
- if (NumElems != 4)
- return false;
-
- return isUndefOrEqual(N->getMaskElt(0), 2) &&
- isUndefOrEqual(N->getMaskElt(1), 3) &&
- isUndefOrEqual(N->getMaskElt(2), 2) &&
- isUndefOrEqual(N->getMaskElt(3), 3);
-}
-
/// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to UNPCKL.
static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
}
/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
-/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
-/// instructions.
+/// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
int NumOperands = SVOp->getValueType(0).getVectorNumElements();
}
/// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
-/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
-/// instructions.
+/// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
unsigned Mask = 0;
}
/// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
-/// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
-/// instructions.
+/// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
unsigned Mask = 0;
return Mask;
}
+/// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
+/// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
+unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
+ EVT VVT = N->getValueType(0);
+ unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
+ int Val = 0;
+
+ unsigned i, e;
+ for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
+ Val = SVOp->getMaskElt(i);
+ if (Val >= 0)
+ break;
+ }
+ return (Val - i) * EltSize;
+}
+
/// isZeroNode - Returns true if Elt is a constant zero or a floating point
/// constant +0.0.
bool X86::isZeroNode(SDValue Elt) {
DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
}
+SDValue
+X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
+ SelectionDAG &DAG) {
+
+ // Check if the scalar load can be widened into a vector load. And if
+ // the address is "base + cst" see if the cst can be "absorbed" into
+ // the shuffle mask.
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
+ SDValue Ptr = LD->getBasePtr();
+ if (!ISD::isNormalLoad(LD) || LD->isVolatile())
+ return SDValue();
+ EVT PVT = LD->getValueType(0);
+ if (PVT != MVT::i32 && PVT != MVT::f32)
+ return SDValue();
+
+ int FI = -1;
+ int64_t Offset = 0;
+ if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
+ FI = FINode->getIndex();
+ Offset = 0;
+ } else if (Ptr.getOpcode() == ISD::ADD &&
+ isa<ConstantSDNode>(Ptr.getOperand(1)) &&
+ isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
+ FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
+ Offset = Ptr.getConstantOperandVal(1);
+ Ptr = Ptr.getOperand(0);
+ } else {
+ return SDValue();
+ }
+
+ SDValue Chain = LD->getChain();
+ // Make sure the stack object alignment is at least 16.
+ MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
+ if (DAG.InferPtrAlignment(Ptr) < 16) {
+ if (MFI->isFixedObjectIndex(FI)) {
+ // Can't change the alignment. Reference stack + offset explicitly
+ // if stack pointer is at least 16-byte aligned.
+ unsigned StackAlign = Subtarget->getStackAlignment();
+ if (StackAlign < 16)
+ return SDValue();
+ Offset = MFI->getObjectOffset(FI) + Offset;
+ SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
+ getPointerTy());
+ Ptr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
+ DAG.getConstant(Offset & ~15, getPointerTy()));
+ Offset %= 16;
+ } else {
+ MFI->setObjectAlignment(FI, 16);
+ }
+ }
+
+ // (Offset % 16) must be multiple of 4. Then address is then
+ // Ptr + (Offset & ~15).
+ if (Offset < 0)
+ return SDValue();
+ if ((Offset % 16) & 3)
+ return SDValue();
+ int64_t StartOffset = Offset & ~15;
+ if (StartOffset)
+ Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
+ Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
+
+ int EltNo = (Offset - StartOffset) >> 2;
+ int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
+ EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
+ SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
+ // Canonicalize it to a v4i32 shuffle.
+ V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ DAG.getVectorShuffle(MVT::v4i32, dl, V1,
+ DAG.getUNDEF(MVT::v4i32), &Mask[0]));
+ }
+
+ return SDValue();
+}
+
SDValue
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
DebugLoc dl = Op.getDebugLoc();
}
// Splat is obviously ok. Let legalizer expand it to a shuffle.
- if (Values.size() == 1)
+ if (Values.size() == 1) {
+ if (EVTBits == 32) {
+ // Instead of a shuffle like this:
+ // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
+ // Check if it's possible to issue this instead.
+ // shuffle (vload ptr)), undef, <1, 1, 1, 1>
+ unsigned Idx = CountTrailingZeros_32(NonZeros);
+ SDValue Item = Op.getOperand(Idx);
+ if (Op.getNode()->isOnlyUserOf(Item.getNode()))
+ return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
+ }
return SDValue();
+ }
// A vector full of immediates; various special cases are already
// handled, so this is best done with a single constant-pool load.
unsigned ShAmt = 0;
SDValue ShVal;
bool isShift = getSubtarget()->hasSSE2() &&
- isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
+ isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
if (isShift && ShVal.hasOneUse()) {
// If the shifted value has multiple uses, it may be cheaper to use
// v_set0 + movlhps or movhlps, etc.
- EVT EVT = VT.getVectorElementType();
- ShAmt *= EVT.getSizeInBits();
+ EVT EltVT = VT.getVectorElementType();
+ ShAmt *= EltVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
X86::isMOVSLDUPMask(SVOp) ||
X86::isMOVHLPSMask(SVOp) ||
- X86::isMOVHPMask(SVOp) ||
+ X86::isMOVLHPSMask(SVOp) ||
X86::isMOVLPMask(SVOp)))
return Op;
if (isShift) {
// No better options. Use a vshl / vsrl.
- EVT EVT = VT.getVectorElementType();
- ShAmt *= EVT.getSizeInBits();
+ EVT EltVT = VT.getVectorElementType();
+ ShAmt *= EltVT.getSizeInBits();
return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
}
MVT::v4i32, Vec),
Op.getOperand(1)));
// Transform it so it match pextrw which produces a 32-bit result.
- EVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
- SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
+ EVT EltVT = MVT::i32;
+ SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
Op.getOperand(0), Op.getOperand(1));
- SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
+ SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
DAG.getValueType(VT));
return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
} else if (VT.getSizeInBits() == 32) {
SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
EVT VT = Op.getValueType();
- EVT EVT = VT.getVectorElementType();
+ EVT EltVT = VT.getVectorElementType();
DebugLoc dl = Op.getDebugLoc();
SDValue N0 = Op.getOperand(0);
SDValue N1 = Op.getOperand(1);
SDValue N2 = Op.getOperand(2);
- if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
+ if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
isa<ConstantSDNode>(N2)) {
- unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
- : X86ISD::PINSRW;
+ unsigned Opc = (EltVT.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)
if (N2.getValueType() != MVT::i32)
N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
return DAG.getNode(Opc, dl, VT, N0, N1, N2);
- } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
+ } else if (EltVT == 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
// Create this as a scalar to vector..
N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
- } else if (EVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
+ } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
// PINSR* works with constant index.
return Op;
}
SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
EVT VT = Op.getValueType();
- EVT EVT = VT.getVectorElementType();
+ EVT EltVT = VT.getVectorElementType();
if (Subtarget->hasSSE41())
return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
- if (EVT == MVT::i8)
+ if (EltVT == MVT::i8)
return SDValue();
DebugLoc dl = Op.getDebugLoc();
SDValue N1 = Op.getOperand(1);
SDValue N2 = Op.getOperand(2);
- if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
+ if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
// Transform it so it match pinsrw which expects a 16-bit value in a GR32
// as its second argument.
if (N1.getValueType() != MVT::i32)
return Result;
}
+SDValue
+X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
+ // Create the TargetBlockAddressAddress node.
+ unsigned char OpFlags =
+ Subtarget->ClassifyBlockAddressReference();
+ CodeModel::Model M = getTargetMachine().getCodeModel();
+ BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
+ /*isTarget=*/true, OpFlags);
+
+ if (Subtarget->isPICStyleRIPRel() &&
+ (M == CodeModel::Small || M == CodeModel::Kernel))
+ Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
+ else
+ Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
+
+ // With PIC, the address is actually $g + Offset.
+ if (isGlobalRelativeToPICBase(OpFlags)) {
+ Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
+ DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
+ Result);
+ }
+
+ return Result;
+}
+
SDValue
X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
int64_t Offset,
GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
unsigned char OperandFlags) {
+ MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
DebugLoc dl = GA->getDebugLoc();
SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
SDValue Ops[] = { Chain, TGA };
Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
}
+
+ // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
+ MFI->setHasCalls(true);
+
SDValue Flag = Chain.getValue(1);
return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
}
DebugLoc dl = Op.getDebugLoc();
unsigned Size = SrcVT.getSizeInBits()/8;
MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
StackSlot,
// 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);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
Tys = DAG.getVTList(MVT::Other);
SmallVector<SDValue, 8> Ops;
// stack slot.
MachineFunction &MF = DAG.getMachineFunction();
unsigned MemSize = DstTy.getSizeInBits()/8;
- int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
unsigned Opc;
};
Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
Chain = Value.getValue(1);
- SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
+ SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
}
bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
+ if (X86CC == X86::COND_INVALID)
+ return SDValue();
SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
+
+ // Use sbb x, x to materialize carry bit into a GPR.
+ if (X86CC == X86::COND_B)
+ return DAG.getNode(ISD::AND, dl, MVT::i8,
+ DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), Cond),
+ DAG.getConstant(1, MVT::i8));
+
return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
DAG.getConstant(X86CC, MVT::i8), Cond);
}
DebugLoc dl = Op.getDebugLoc();
SDValue CC;
- if (Cond.getOpcode() == ISD::SETCC)
- Cond = LowerSETCC(Cond, DAG);
+ if (Cond.getOpcode() == ISD::SETCC) {
+ SDValue NewCond = LowerSETCC(Cond, DAG);
+ if (NewCond.getNode())
+ Cond = NewCond;
+ }
+
+ // Look pass (and (setcc_carry (cmp ...)), 1).
+ if (Cond.getOpcode() == ISD::AND &&
+ Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
+ if (C && C->getAPIntValue() == 1)
+ Cond = Cond.getOperand(0);
+ }
// 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) {
+ if (Cond.getOpcode() == X86ISD::SETCC ||
+ Cond.getOpcode() == X86ISD::SETCC_CARRY) {
CC = Cond.getOperand(0);
SDValue Cmp = Cond.getOperand(1);
DebugLoc dl = Op.getDebugLoc();
SDValue CC;
- if (Cond.getOpcode() == ISD::SETCC)
- Cond = LowerSETCC(Cond, DAG);
+ if (Cond.getOpcode() == ISD::SETCC) {
+ SDValue NewCond = LowerSETCC(Cond, DAG);
+ if (NewCond.getNode())
+ Cond = NewCond;
+ }
#if 0
// FIXME: LowerXALUO doesn't handle these!!
else if (Cond.getOpcode() == X86ISD::ADD ||
Cond = LowerXALUO(Cond, DAG);
#endif
+ // Look pass (and (setcc_carry (cmp ...)), 1).
+ if (Cond.getOpcode() == ISD::AND &&
+ Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
+ if (C && C->getAPIntValue() == 1)
+ Cond = Cond.getOperand(0);
+ }
+
// 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) {
+ if (Cond.getOpcode() == X86ISD::SETCC ||
+ Cond.getOpcode() == X86ISD::SETCC_CARRY) {
CC = Cond.getOperand(0);
SDValue Cmp = Cond.getOperand(1);
LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
false, false, false, false,
0, CallingConv::C, false, /*isReturnValueUsed=*/false,
- DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
+ DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
+ DAG.GetOrdering(Chain.getNode()));
return CallResult.second;
}
SDValue LHS = Op.getOperand(1);
SDValue RHS = Op.getOperand(2);
unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
+ assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
DAG.getConstant(X86CC, MVT::i8), Cond);
DebugLoc dl = Op.getDebugLoc();
// Save FP Control Word to stack slot
- int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
+ int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
+ case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
case ISD::SHL_PARTS:
case ISD::SRA_PARTS:
case ISD::SRL_PARTS: return LowerShift(Op, DAG);
Node->getOperand(2), DAG.getIntPtrConstant(0));
SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
Node->getOperand(2), DAG.getIntPtrConstant(1));
- // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
- // have a MemOperand. Pass the info through as a normal operand.
- SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
- SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
+ SDValue Ops[] = { Chain, In1, In2L, In2H };
SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
- SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
+ SDValue Result =
+ DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
+ cast<MemSDNode>(Node)->getMemOperand());
SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
Results.push_back(Result.getValue(2));
Results.push_back(edx.getValue(1));
return;
}
+ case ISD::SDIV:
+ case ISD::UDIV:
+ case ISD::SREM:
+ case ISD::UREM: {
+ EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
+ Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
+ return;
+ }
case ISD::ATOMIC_CMP_SWAP: {
EVT T = N->getValueType(0);
assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
case X86ISD::COMI: return "X86ISD::COMI";
case X86ISD::UCOMI: return "X86ISD::UCOMI";
case X86ISD::SETCC: return "X86ISD::SETCC";
+ case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
case X86ISD::CMOV: return "X86ISD::CMOV";
case X86ISD::BRCOND: return "X86ISD::BRCOND";
case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
if (VT.getSizeInBits() == 64)
return false;
- // FIXME: pshufb, blends, palignr, shifts.
+ // FIXME: pshufb, blends, shifts.
return (VT.getVectorNumElements() == 2 ||
ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
isMOVLMask(M, VT) ||
isPSHUFDMask(M, VT) ||
isPSHUFHWMask(M, VT) ||
isPSHUFLWMask(M, VT) ||
+ isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
isUNPCKLMask(M, VT) ||
isUNPCKHMask(M, VT) ||
isUNPCKL_v_undef_Mask(M, VT) ||
(*MIB).addOperand(*argOpers[i]);
MIB.addReg(t2);
assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
- (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
+ (*MIB).setMemRefs(bInstr->memoperands_begin(),
+ bInstr->memoperands_end());
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
MIB.addReg(EAXreg);
(*MIB).addOperand(*argOpers[i]);
assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
- (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
+ (*MIB).setMemRefs(bInstr->memoperands_begin(),
+ bInstr->memoperands_end());
MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
MIB.addReg(X86::EAX);
(*MIB).addOperand(*argOpers[i]);
MIB.addReg(t3);
assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
- (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
+ (*MIB).setMemRefs(mInstr->memoperands_begin(),
+ mInstr->memoperands_end());
MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
MIB.addReg(X86::EAX);
// all of this code can be replaced with that in the .td file.
MachineBasicBlock *
X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
- unsigned numArgs, bool memArg) const {
+ unsigned numArgs, bool memArg) const {
MachineFunction *F = BB->getParent();
DebugLoc dl = MI->getDebugLoc();
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
unsigned Opc;
-
- if (memArg) {
- Opc = numArgs == 3 ?
- X86::PCMPISTRM128rm :
- X86::PCMPESTRM128rm;
- } else {
- Opc = numArgs == 3 ?
- X86::PCMPISTRM128rr :
- X86::PCMPESTRM128rr;
- }
+ if (memArg)
+ Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
+ else
+ Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
// In the XMM save block, save all the XMM argument registers.
for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
+ MachineMemOperand *MMO =
+ F->getMachineMemOperand(
+ PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
+ MachineMemOperand::MOStore, Offset,
+ /*Size=*/16, /*Align=*/16);
BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
.addFrameIndex(RegSaveFrameIndex)
.addImm(/*Scale=*/1)
.addImm(/*Disp=*/Offset)
.addReg(/*Segment=*/0)
.addReg(MI->getOperand(i).getReg())
- .addMemOperand(MachineMemOperand(
- PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
- MachineMemOperand::MOStore, Offset,
- /*Size=*/16, /*Align=*/16));
+ .addMemOperand(MMO);
}
F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
MachineBasicBlock *
X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
- MachineBasicBlock *BB) const {
+ MachineBasicBlock *BB,
+ DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
-
+
// 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
const BasicBlock *LLVM_BB = BB->getBasicBlock();
MachineFunction::iterator It = BB;
++It;
-
+
// thisMBB:
// ...
// TrueVal = ...
BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
F->insert(It, copy0MBB);
F->insert(It, sinkMBB);
- // Update machine-CFG edges by transferring all successors of the current
+ // Update machine-CFG edges by first adding all successors of the current
// block to the new block which will contain the Phi node for the select.
- sinkMBB->transferSuccessors(BB);
-
+ // Also inform sdisel of the edge changes.
+ for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
+ E = BB->succ_end(); I != E; ++I) {
+ EM->insert(std::make_pair(*I, sinkMBB));
+ sinkMBB->addSuccessor(*I);
+ }
+ // Next, remove all successors of the current block, and add the true
+ // and fallthrough blocks as its successors.
+ while (!BB->succ_empty())
+ BB->removeSuccessor(BB->succ_begin());
// Add the true and fallthrough blocks as its successors.
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 ]
// ...
MachineBasicBlock *
X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
- MachineBasicBlock *BB) const {
+ MachineBasicBlock *BB,
+ DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
switch (MI->getOpcode()) {
default: assert(false && "Unexpected instr type to insert");
case X86::CMOV_GR8:
case X86::CMOV_V4F32:
case X86::CMOV_V2F64:
case X86::CMOV_V2I64:
- return EmitLoweredSelect(MI, BB);
+ return EmitLoweredSelect(MI, BB, EM);
case X86::FP32_TO_INT16_IN_MEM:
case X86::FP32_TO_INT32_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);
+ int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
// Load the old value of the high byte of the control word...
return TargetLowering::isGAPlusOffset(N, GA, Offset);
}
-static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
- const TargetLowering &TLI) {
- GlobalValue *GV;
- int64_t Offset = 0;
- if (TLI.isGAPlusOffset(Base, GV, Offset))
- return (GV->getAlignment() >= N && (Offset % N) == 0);
- // DAG combine handles the stack object case.
- return false;
-}
-
static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
- EVT EVT, LoadSDNode *&LDBase,
+ EVT EltVT, LoadSDNode *&LDBase,
unsigned &LastLoadedElt,
SelectionDAG &DAG, MachineFrameInfo *MFI,
const TargetLowering &TLI) {
continue;
LoadSDNode *LD = cast<LoadSDNode>(Elt);
- if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
+ if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
return false;
LastLoadedElt = i;
}
const TargetLowering &TLI) {
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
- EVT EVT = VT.getVectorElementType();
+ EVT EltVT = VT.getVectorElementType();
ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
unsigned NumElems = VT.getVectorNumElements();
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
LoadSDNode *LD = NULL;
unsigned LastLoadedElt;
- if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
+ if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
MFI, TLI))
return SDValue();
if (LastLoadedElt == NumElems - 1) {
- if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
+ if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
LD->getSrcValue(), LD->getSrcValueOffset(),
LD->isVolatile());
SDValue LHS = N->getOperand(1);
SDValue RHS = N->getOperand(2);
- // If we have SSE[12] support, try to form min/max nodes.
+ // If we have SSE[12] support, try to form min/max nodes. SSE min/max
+ // instructions have the peculiarity that if either operand is a NaN,
+ // they chose what we call the RHS operand (and as such are not symmetric).
+ // It happens that this matches the semantics of the common C idiom
+ // x<y?x:y and related forms, so we can recognize these cases.
if (Subtarget->hasSSE2() &&
(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
Cond.getOpcode() == ISD::SETCC) {
ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
unsigned Opcode = 0;
+ // Check for x CC y ? x : y.
if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
switch (CC) {
default: break;
- case ISD::SETOLE: // (X <= Y) ? X : Y -> min
+ case ISD::SETULT:
+ // This can be a min if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(RHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(LHS))
+ break;
+ }
+ Opcode = X86ISD::FMIN;
+ break;
+ case ISD::SETOLE:
+ // This can be a min if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(LHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(RHS))
+ break;
+ }
+ Opcode = X86ISD::FMIN;
+ break;
case ISD::SETULE:
- case ISD::SETLE:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
+ // This can be a min, but if either operand is a NaN we need it to
+ // preserve the original LHS.
+ std::swap(LHS, RHS);
+ case ISD::SETOLT:
case ISD::SETLT:
+ case ISD::SETLE:
Opcode = X86ISD::FMIN;
break;
- case ISD::SETOGT: // (X > Y) ? X : Y -> max
+ case ISD::SETOGE:
+ // This can be a max if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(LHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(RHS))
+ break;
+ }
+ Opcode = X86ISD::FMAX;
+ break;
case ISD::SETUGT:
+ // This can be a max if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(RHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(LHS))
+ break;
+ }
+ Opcode = X86ISD::FMAX;
+ break;
+ case ISD::SETUGE:
+ // This can be a max, but if either operand is a NaN we need it to
+ // preserve the original LHS.
+ std::swap(LHS, RHS);
+ case ISD::SETOGT:
case ISD::SETGT:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
case ISD::SETGE:
Opcode = X86ISD::FMAX;
break;
}
+ // Check for x CC y ? y : x -- a min/max with reversed arms.
} else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
switch (CC) {
default: break;
- case ISD::SETOGT:
- // This can use a min only if the LHS isn't NaN.
- if (DAG.isKnownNeverNaN(LHS))
- Opcode = X86ISD::FMIN;
- else if (DAG.isKnownNeverNaN(RHS)) {
- Opcode = X86ISD::FMIN;
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
+ case ISD::SETOGE:
+ // This can be a min if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(RHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(LHS))
+ break;
}
+ Opcode = X86ISD::FMIN;
break;
-
- case ISD::SETUGT: // (X > Y) ? Y : X -> min
+ case ISD::SETUGT:
+ // This can be a min if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(LHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(RHS))
+ break;
+ }
+ Opcode = X86ISD::FMIN;
+ break;
+ case ISD::SETUGE:
+ // This can be a min, but if either operand is a NaN we need it to
+ // preserve the original LHS.
+ std::swap(LHS, RHS);
+ case ISD::SETOGT:
case ISD::SETGT:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
case ISD::SETGE:
Opcode = X86ISD::FMIN;
break;
- case ISD::SETULE:
- // This can use a max only if the LHS isn't NaN.
- if (DAG.isKnownNeverNaN(LHS))
- Opcode = X86ISD::FMAX;
- else if (DAG.isKnownNeverNaN(RHS)) {
- Opcode = X86ISD::FMAX;
- // Put the potential NaN in the RHS so that SSE will preserve it.
- std::swap(LHS, RHS);
+ case ISD::SETULT:
+ // This can be a max if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(LHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(RHS))
+ break;
}
+ Opcode = X86ISD::FMAX;
break;
-
- case ISD::SETOLE: // (X <= Y) ? Y : X -> max
- case ISD::SETLE:
- if (!UnsafeFPMath) break;
- // FALL THROUGH.
- case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
+ case ISD::SETOLE:
+ // This can be a max if we can prove that at least one of the operands
+ // is not a nan.
+ if (!FiniteOnlyFPMath()) {
+ if (DAG.isKnownNeverNaN(RHS)) {
+ // Put the potential NaN in the RHS so that SSE will preserve it.
+ std::swap(LHS, RHS);
+ } else if (!DAG.isKnownNeverNaN(LHS))
+ break;
+ }
+ Opcode = X86ISD::FMAX;
+ break;
+ case ISD::SETULE:
+ // This can be a max, but if either operand is a NaN we need it to
+ // preserve the original LHS.
+ std::swap(LHS, RHS);
+ case ISD::SETOLT:
case ISD::SETLT:
+ case ISD::SETLE:
Opcode = X86ISD::FMAX;
break;
}
return SDValue();
}
+static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
+ EVT VT = N0.getValueType();
+
+ // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
+ // since the result of setcc_c is all zero's or all ones.
+ if (N1C && N0.getOpcode() == ISD::AND &&
+ N0.getOperand(1).getOpcode() == ISD::Constant) {
+ SDValue N00 = N0.getOperand(0);
+ if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
+ ((N00.getOpcode() == ISD::ANY_EXTEND ||
+ N00.getOpcode() == ISD::ZERO_EXTEND) &&
+ N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
+ APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
+ APInt ShAmt = N1C->getAPIntValue();
+ Mask = Mask.shl(ShAmt);
+ if (Mask != 0)
+ return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
+ N00, DAG.getConstant(Mask, VT));
+ }
+ }
+
+ return SDValue();
+}
/// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
/// when possible.
static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
+ if (!VT.isVector() && VT.isInteger() &&
+ N->getOpcode() == ISD::SHL)
+ return PerformSHLCombine(N, DAG);
+
// On X86 with SSE2 support, we can transform this to a vector shift if
// all elements are shifted by the same amount. We can't do this in legalize
// because the a constant vector is typically transformed to a constant pool
if (!Subtarget->hasSSE2())
return SDValue();
- EVT VT = N->getValueType(0);
if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
return SDValue();
}
}
+static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
+ // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
+ // (and (i32 x86isd::setcc_carry), 1)
+ // This eliminates the zext. This transformation is necessary because
+ // ISD::SETCC is always legalized to i8.
+ DebugLoc dl = N->getDebugLoc();
+ SDValue N0 = N->getOperand(0);
+ EVT VT = N->getValueType(0);
+ if (N0.getOpcode() == ISD::AND &&
+ N0.hasOneUse() &&
+ N0.getOperand(0).hasOneUse()) {
+ SDValue N00 = N0.getOperand(0);
+ if (N00.getOpcode() != X86ISD::SETCC_CARRY)
+ return SDValue();
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+ if (!C || C->getZExtValue() != 1)
+ return SDValue();
+ return DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
+ N00.getOperand(0), N00.getOperand(1)),
+ DAG.getConstant(1, VT));
+ }
+
+ return SDValue();
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
+ case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
}
return SDValue();
switch (Constraint[0]) {
default: break;
case 'r': // GENERAL_REGS
- case 'R': // LEGACY_REGS
case 'l': // INDEX_REGS
if (VT == MVT::i8)
return std::make_pair(0U, X86::GR8RegisterClass);
if (VT == MVT::i32 || !Subtarget->is64Bit())
return std::make_pair(0U, X86::GR32RegisterClass);
return std::make_pair(0U, X86::GR64RegisterClass);
+ case 'R': // LEGACY_REGS
+ if (VT == MVT::i8)
+ return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
+ if (VT == MVT::i16)
+ return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
+ if (VT == MVT::i32 || !Subtarget->is64Bit())
+ return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
+ return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
case 'f': // FP Stack registers.
// If SSE is enabled for this VT, use f80 to ensure the isel moves the
// value to the correct fpstack register class.
(Constraint[4] >= '0' && Constraint[4] <= '7') &&
Constraint[5] == ')' &&
Constraint[6] == '}') {
-
+
Res.first = X86::ST0+Constraint[4]-'0';
Res.second = X86::RFP80RegisterClass;
return Res;
}
-
+
// GCC allows "st(0)" to be called just plain "st".
- if (StringsEqualNoCase("{st}", Constraint)) {
+ if (StringRef("{st}").equals_lower(Constraint)) {
Res.first = X86::ST0;
Res.second = X86::RFP80RegisterClass;
return Res;
}
// flags -> EFLAGS
- if (StringsEqualNoCase("{flags}", Constraint)) {
+ if (StringRef("{flags}").equals_lower(Constraint)) {
Res.first = X86::EFLAGS;
Res.second = X86::CCRRegisterClass;
return Res;
}
-
+
// 'A' means EAX + EDX.
if (Constraint == "A") {
Res.first = X86::EAX;