Subtarget = &TM.getSubtarget<X86Subtarget>();
X86ScalarSSEf64 = Subtarget->hasSSE2();
X86ScalarSSEf32 = Subtarget->hasSSE1();
- X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
RegInfo = TM.getRegisterInfo();
TD = getDataLayout();
setSchedulingPreference(Sched::ILP);
else
setSchedulingPreference(Sched::RegPressure);
- setStackPointerRegisterToSaveRestore(X86StackPtr);
+ setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
// Bypass i32 with i8 on Atom when compiling with O2
if (Subtarget->hasSlowDivide() && TM.getOptLevel() >= CodeGenOpt::Default)
setOperationAction(ISD::SETCC , MVT::i64 , Custom);
}
setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
+ // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intened to support
+ // SjLj exception handling but a light-weight setjmp/longjmp replacement to
+ // support continuation, user-level threading, and etc.. As a result, no
+ // other SjLj exception interfaces are implemented and please don't build
+ // your own exception handling based on them.
+ // LLVM/Clang supports zero-cost DWARF exception handling.
+ setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
+ setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
// Darwin ABI issue.
setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
setOperationAction(ISD::TRAP, MVT::Other, Legal);
+ setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
// VASTART needs to be custom lowered to use the VarArgsFrameIndex
setOperationAction(ISD::VASTART , MVT::Other, Custom);
setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
+ setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
+ setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
+ // As there is no 64-bit GPR available, we need build a special custom
+ // sequence to convert from v2i32 to v2f32.
+ if (!Subtarget->is64Bit())
+ setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
+
+ setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
+ setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
+
setLoadExtAction(ISD::EXTLOAD, MVT::v2f32, Legal);
}
setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
setOperationAction(ISD::FABS, MVT::v4f64, Custom);
+ setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
+
+ setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
+
setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
+ setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
+ setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
+ setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
+
setLoadExtAction(ISD::EXTLOAD, MVT::v4f32, Legal);
setOperationAction(ISD::SRL, MVT::v16i16, Custom);
setTargetDAGCombine(ISD::ANY_EXTEND);
setTargetDAGCombine(ISD::SIGN_EXTEND);
setTargetDAGCombine(ISD::TRUNCATE);
- setTargetDAGCombine(ISD::UINT_TO_FP);
setTargetDAGCombine(ISD::SINT_TO_FP);
setTargetDAGCombine(ISD::SETCC);
- setTargetDAGCombine(ISD::FP_TO_SINT);
if (Subtarget->is64Bit())
setTargetDAGCombine(ISD::MUL);
setTargetDAGCombine(ISD::XOR);
// cases like PR2962. This should be removed when PR2962 is fixed.
const Function *F = MF.getFunction();
if (IsZeroVal &&
- !F->getFnAttributes().hasNoImplicitFloatAttr()) {
+ !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat)) {
if (Size >= 16 &&
(Subtarget->isUnalignedMemAccessFast() ||
((DstAlign == 0 || DstAlign >= 16) &&
unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
TotalNumIntRegs);
- bool NoImplicitFloatOps = Fn->getFnAttributes().hasNoImplicitFloatAttr();
+ bool NoImplicitFloatOps = Fn->getFnAttributes().
+ hasAttribute(Attributes::NoImplicitFloat);
assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
"SSE register cannot be used when SSE is disabled!");
assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
/// optimization is performed and it is required (FPDiff!=0).
static SDValue
EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
- SDValue Chain, SDValue RetAddrFrIdx,
- bool Is64Bit, int FPDiff, DebugLoc dl) {
+ SDValue Chain, SDValue RetAddrFrIdx, EVT PtrVT,
+ unsigned SlotSize, int FPDiff, DebugLoc dl) {
// Store the return address to the appropriate stack slot.
if (!FPDiff) return Chain;
// Calculate the new stack slot for the return address.
- int SlotSize = Is64Bit ? 8 : 4;
int NewReturnAddrFI =
MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
- EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
- SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
+ SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
MachinePointerInfo::getFixedStack(NewReturnAddrFI),
false, false, 0);
int FPDiff = 0;
if (isTailCall && !IsSibcall) {
// Lower arguments at fp - stackoffset + fpdiff.
- unsigned NumBytesCallerPushed =
- MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
+ X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
+ unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
+
FPDiff = NumBytesCallerPushed - NumBytes;
// Set the delta of movement of the returnaddr stackslot.
// But only set if delta is greater than previous delta.
- if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
- MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
+ if (FPDiff < X86Info->getTCReturnAddrDelta())
+ X86Info->setTCReturnAddrDelta(FPDiff);
}
if (!IsSibcall)
} else if (!IsSibcall && (!isTailCall || isByVal)) {
assert(VA.isMemLoc());
if (StackPtr.getNode() == 0)
- StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
+ StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
+ getPointerTy());
MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
dl, DAG, VA, Flags));
}
// Copy relative to framepointer.
SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
if (StackPtr.getNode() == 0)
- StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
+ StackPtr = DAG.getCopyFromReg(Chain, dl,
+ RegInfo->getStackRegister(),
getPointerTy());
Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
&MemOpChains2[0], MemOpChains2.size());
// Store the return address to the appropriate stack slot.
- Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
+ Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
+ getPointerTy(), RegInfo->getSlotSize(),
FPDiff, dl);
}
OpFlags = X86II::MO_DARWIN_STUB;
} else if (Subtarget->isPICStyleRIPRel() &&
isa<Function>(GV) &&
- cast<Function>(GV)->getFnAttributes().hasNonLazyBindAttr()) {
+ cast<Function>(GV)->getFnAttributes().
+ hasAttribute(Attributes::NonLazyBind)) {
// If the function is marked as non-lazy, generate an indirect call
// which loads from the GOT directly. This avoids runtime overhead
// at the cost of eager binding (and one extra byte of encoding).
unsigned StackAlignment = TFI.getStackAlignment();
uint64_t AlignMask = StackAlignment - 1;
int64_t Offset = StackSize;
- uint64_t SlotSize = TD->getPointerSize();
+ unsigned SlotSize = RegInfo->getSlotSize();
if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
// Number smaller than 12 so just add the difference.
Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
if (ReturnAddrIndex == 0) {
// Set up a frame object for the return address.
- uint64_t SlotSize = TD->getPointerSize();
+ unsigned SlotSize = RegInfo->getSlotSize();
ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
false);
FuncInfo->setRAIndex(ReturnAddrIndex);
MVT ShufVT = V.getValueType().getSimpleVT();
unsigned NumElems = ShufVT.getVectorNumElements();
SmallVector<int, 16> ShuffleMask;
- SDValue ImmN;
bool IsUnary;
if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
return SDValue();
}
-// LowerVectorFpExtend - Recognize the scalarized FP_EXTEND from v2f32 to v2f64
-// and convert it into X86ISD::VFPEXT due to the current ISD::FP_EXTEND has the
-// constraint of matching input/output vector elements.
SDValue
-X86TargetLowering::LowerVectorFpExtend(SDValue &Op, SelectionDAG &DAG) const {
- DebugLoc DL = Op.getDebugLoc();
- SDNode *N = Op.getNode();
+X86TargetLowering::buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) const {
EVT VT = Op.getValueType();
- unsigned NumElts = Op.getNumOperands();
- // Check supported types and sub-targets.
- //
- // Only v2f32 -> v2f64 needs special handling.
- if (VT != MVT::v2f64 || !Subtarget->hasSSE2())
+ // Skip if insert_vec_elt is not supported.
+ if (!isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
return SDValue();
- SDValue VecIn;
- EVT VecInVT;
- SmallVector<int, 8> Mask;
- EVT SrcVT = MVT::Other;
+ DebugLoc DL = Op.getDebugLoc();
+ unsigned NumElems = Op.getNumOperands();
- // Check the patterns could be translated into X86vfpext.
- for (unsigned i = 0; i < NumElts; ++i) {
- SDValue In = N->getOperand(i);
- unsigned Opcode = In.getOpcode();
+ SDValue VecIn1;
+ SDValue VecIn2;
+ SmallVector<unsigned, 4> InsertIndices;
+ SmallVector<int, 8> Mask(NumElems, -1);
- // Skip if the element is undefined.
- if (Opcode == ISD::UNDEF) {
- Mask.push_back(-1);
+ for (unsigned i = 0; i != NumElems; ++i) {
+ unsigned Opc = Op.getOperand(i).getOpcode();
+
+ if (Opc == ISD::UNDEF)
continue;
- }
- // Quit if one of the elements is not defined from 'fpext'.
- if (Opcode != ISD::FP_EXTEND)
- return SDValue();
+ if (Opc != ISD::EXTRACT_VECTOR_ELT) {
+ // Quit if more than 1 elements need inserting.
+ if (InsertIndices.size() > 1)
+ return SDValue();
+
+ InsertIndices.push_back(i);
+ continue;
+ }
- // Check how the source of 'fpext' is defined.
- SDValue L2In = In.getOperand(0);
- EVT L2InVT = L2In.getValueType();
+ SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
+ SDValue ExtIdx = Op.getOperand(i).getOperand(1);
- // Check the original type
- if (SrcVT == MVT::Other)
- SrcVT = L2InVT;
- else if (SrcVT != L2InVT) // Quit if non-homogenous typed.
+ // Quit if extracted from vector of different type.
+ if (ExtractedFromVec.getValueType() != VT)
return SDValue();
- // Check whether the value being 'fpext'ed is extracted from the same
- // source.
- Opcode = L2In.getOpcode();
-
- // Quit if it's not extracted with a constant index.
- if (Opcode != ISD::EXTRACT_VECTOR_ELT ||
- !isa<ConstantSDNode>(L2In.getOperand(1)))
+ // Quit if non-constant index.
+ if (!isa<ConstantSDNode>(ExtIdx))
return SDValue();
- SDValue ExtractedFromVec = L2In.getOperand(0);
+ if (VecIn1.getNode() == 0)
+ VecIn1 = ExtractedFromVec;
+ else if (VecIn1 != ExtractedFromVec) {
+ if (VecIn2.getNode() == 0)
+ VecIn2 = ExtractedFromVec;
+ else if (VecIn2 != ExtractedFromVec)
+ // Quit if more than 2 vectors to shuffle
+ return SDValue();
+ }
- if (VecIn.getNode() == 0) {
- VecIn = ExtractedFromVec;
- VecInVT = ExtractedFromVec.getValueType();
- } else if (VecIn != ExtractedFromVec) // Quit if built from more than 1 vec.
- return SDValue();
+ unsigned Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
- Mask.push_back(cast<ConstantSDNode>(L2In.getOperand(1))->getZExtValue());
+ if (ExtractedFromVec == VecIn1)
+ Mask[i] = Idx;
+ else if (ExtractedFromVec == VecIn2)
+ Mask[i] = Idx + NumElems;
}
- // Quit if all operands of BUILD_VECTOR are undefined.
- if (!VecIn.getNode())
+ if (VecIn1.getNode() == 0)
return SDValue();
- // Fill the remaining mask as undef.
- for (unsigned i = NumElts; i < VecInVT.getVectorNumElements(); ++i)
- Mask.push_back(-1);
+ VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
+ SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
+ for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
+ unsigned Idx = InsertIndices[i];
+ NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
+ DAG.getIntPtrConstant(Idx));
+ }
- return DAG.getNode(X86ISD::VFPEXT, DL, VT,
- DAG.getVectorShuffle(VecInVT, DL,
- VecIn, DAG.getUNDEF(VecInVT),
- &Mask[0]));
+ return NV;
}
SDValue
if (Broadcast.getNode())
return Broadcast;
- SDValue FpExt = LowerVectorFpExtend(Op, DAG);
- if (FpExt.getNode())
- return FpExt;
-
unsigned EVTBits = ExtVT.getSizeInBits();
unsigned NumZero = 0;
if (LD.getNode())
return LD;
+ // Check for a build vector from mostly shuffle plus few inserting.
+ SDValue Sh = buildFromShuffleMostly(Op, DAG);
+ if (Sh.getNode())
+ return Sh;
+
// For SSE 4.1, use insertps to put the high elements into the low element.
if (getSubtarget()->hasSSE41()) {
SDValue Result;
}
static bool MayFoldVectorLoad(SDValue V) {
- if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
+ while (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
V = V.getOperand(0);
+
if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
V = V.getOperand(0);
if (V.hasOneUse() && V.getOpcode() == ISD::BUILD_VECTOR &&
V.getNumOperands() == 2 && V.getOperand(1).getOpcode() == ISD::UNDEF)
// BUILD_VECTOR (load), undef
V = V.getOperand(0);
- if (MayFoldLoad(V))
- return true;
- return false;
+
+ return MayFoldLoad(V);
}
// FIXME: the version above should always be used. Since there's
getShuffleSHUFImmediate(SVOp), DAG);
}
+// Reduce a vector shuffle to zext.
+SDValue
+X86TargetLowering::lowerVectorIntExtend(SDValue Op, SelectionDAG &DAG) const {
+ // PMOVZX is only available from SSE41.
+ if (!Subtarget->hasSSE41())
+ return SDValue();
+
+ EVT VT = Op.getValueType();
+
+ // Only AVX2 support 256-bit vector integer extending.
+ if (!Subtarget->hasAVX2() && VT.is256BitVector())
+ return SDValue();
+
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ DebugLoc DL = Op.getDebugLoc();
+ SDValue V1 = Op.getOperand(0);
+ SDValue V2 = Op.getOperand(1);
+ unsigned NumElems = VT.getVectorNumElements();
+
+ // Extending is an unary operation and the element type of the source vector
+ // won't be equal to or larger than i64.
+ if (V2.getOpcode() != ISD::UNDEF || !VT.isInteger() ||
+ VT.getVectorElementType() == MVT::i64)
+ return SDValue();
+
+ // Find the expansion ratio, e.g. expanding from i8 to i32 has a ratio of 4.
+ unsigned Shift = 1; // Start from 2, i.e. 1 << 1.
+ while ((1U << Shift) < NumElems) {
+ if (SVOp->getMaskElt(1U << Shift) == 1)
+ break;
+ Shift += 1;
+ // The maximal ratio is 8, i.e. from i8 to i64.
+ if (Shift > 3)
+ return SDValue();
+ }
+
+ // Check the shuffle mask.
+ unsigned Mask = (1U << Shift) - 1;
+ for (unsigned i = 0; i != NumElems; ++i) {
+ int EltIdx = SVOp->getMaskElt(i);
+ if ((i & Mask) != 0 && EltIdx != -1)
+ return SDValue();
+ if ((i & Mask) == 0 && (unsigned)EltIdx != (i >> Shift))
+ return SDValue();
+ }
+
+ unsigned NBits = VT.getVectorElementType().getSizeInBits() << Shift;
+ EVT NeVT = EVT::getIntegerVT(*DAG.getContext(), NBits);
+ EVT NVT = EVT::getVectorVT(*DAG.getContext(), NeVT, NumElems >> Shift);
+
+ if (!isTypeLegal(NVT))
+ return SDValue();
+
+ // Simplify the operand as it's prepared to be fed into shuffle.
+ unsigned SignificantBits = NVT.getSizeInBits() >> Shift;
+ if (V1.getOpcode() == ISD::BITCAST &&
+ V1.getOperand(0).getOpcode() == ISD::SCALAR_TO_VECTOR &&
+ V1.getOperand(0).getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+ V1.getOperand(0)
+ .getOperand(0).getValueType().getSizeInBits() == SignificantBits) {
+ // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
+ SDValue V = V1.getOperand(0).getOperand(0).getOperand(0);
+ ConstantSDNode *CIdx =
+ dyn_cast<ConstantSDNode>(V1.getOperand(0).getOperand(0).getOperand(1));
+ // If it's foldable, i.e. normal load with single use, we will let code
+ // selection to fold it. Otherwise, we will short the conversion sequence.
+ if (CIdx && CIdx->getZExtValue() == 0 &&
+ (!ISD::isNormalLoad(V.getNode()) || !V.hasOneUse()))
+ V1 = DAG.getNode(ISD::BITCAST, DL, V1.getValueType(), V);
+ }
+
+ return DAG.getNode(ISD::BITCAST, DL, VT,
+ DAG.getNode(X86ISD::VZEXT, DL, NVT, V1));
+}
+
SDValue
X86TargetLowering::NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG) const {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
return PromoteSplat(SVOp, DAG);
}
+ // Check integer expanding shuffles.
+ SDValue NewOp = lowerVectorIntExtend(Op, DAG);
+ if (NewOp.getNode())
+ return NewOp;
+
// If the shuffle can be profitably rewritten as a narrower shuffle, then
// do it!
if (VT == MVT::v8i16 || VT == MVT::v16i8 ||
bool HasAVX = Subtarget->hasAVX();
bool HasAVX2 = Subtarget->hasAVX2();
MachineFunction &MF = DAG.getMachineFunction();
- bool OptForSize = MF.getFunction()->getFnAttributes().hasOptimizeForSizeAttr();
+ bool OptForSize = MF.getFunction()->getFnAttributes().
+ hasAttribute(Attributes::OptimizeForSize);
assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
return Sub;
}
+SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
+ SelectionDAG &DAG) const {
+ SDValue N0 = Op.getOperand(0);
+ EVT SVT = N0.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+
+ assert((SVT == MVT::v4i8 || SVT == MVT::v4i16 ||
+ SVT == MVT::v8i8 || SVT == MVT::v8i16) &&
+ "Custom UINT_TO_FP is not supported!");
+
+ EVT NVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, SVT.getVectorNumElements());
+ return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
+ DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
+}
+
SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
SelectionDAG &DAG) const {
SDValue N0 = Op.getOperand(0);
DebugLoc dl = Op.getDebugLoc();
+ if (Op.getValueType().isVector())
+ return lowerUINT_TO_FP_vec(Op, DAG);
+
// Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
// optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
// the optimization here.
}
}
+SDValue X86TargetLowering::lowerZERO_EXTEND(SDValue Op, SelectionDAG &DAG) const {
+ DebugLoc DL = Op.getDebugLoc();
+ EVT VT = Op.getValueType();
+ SDValue In = Op.getOperand(0);
+ EVT SVT = In.getValueType();
+
+ if (!VT.is256BitVector() || !SVT.is128BitVector() ||
+ VT.getVectorNumElements() != SVT.getVectorNumElements())
+ return SDValue();
+
+ assert(Subtarget->hasAVX() && "256-bit vector is observed without AVX!");
+
+ // AVX2 has better support of integer extending.
+ if (Subtarget->hasAVX2())
+ return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
+
+ SDValue Lo = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32, In);
+ static const int Mask[] = {4, 5, 6, 7, -1, -1, -1, -1};
+ SDValue Hi = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32,
+ DAG.getVectorShuffle(MVT::v8i16, DL, In, DAG.getUNDEF(MVT::v8i16), &Mask[0]));
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
+}
+
+SDValue X86TargetLowering::lowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
+ DebugLoc DL = Op.getDebugLoc();
+ EVT VT = Op.getValueType();
+ EVT SVT = Op.getOperand(0).getValueType();
+
+ if (!VT.is128BitVector() || !SVT.is256BitVector() ||
+ VT.getVectorNumElements() != SVT.getVectorNumElements())
+ return SDValue();
+
+ assert(Subtarget->hasAVX() && "256-bit vector is observed without AVX!");
+
+ unsigned NumElems = VT.getVectorNumElements();
+ EVT NVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
+ NumElems * 2);
+
+ SDValue In = Op.getOperand(0);
+ SmallVector<int, 16> MaskVec(NumElems * 2, -1);
+ // Prepare truncation shuffle mask
+ for (unsigned i = 0; i != NumElems; ++i)
+ MaskVec[i] = i * 2;
+ SDValue V = DAG.getVectorShuffle(NVT, DL,
+ DAG.getNode(ISD::BITCAST, DL, NVT, In),
+ DAG.getUNDEF(NVT), &MaskVec[0]);
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
+ DAG.getIntPtrConstant(0));
+}
+
SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
SelectionDAG &DAG) const {
- if (Op.getValueType().isVector())
+ if (Op.getValueType().isVector()) {
+ if (Op.getValueType() == MVT::v8i16)
+ return DAG.getNode(ISD::TRUNCATE, Op.getDebugLoc(), Op.getValueType(),
+ DAG.getNode(ISD::FP_TO_SINT, Op.getDebugLoc(),
+ MVT::v8i32, Op.getOperand(0)));
return SDValue();
+ }
std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
/*IsSigned=*/ true, /*IsReplace=*/ false);
return FIST;
}
+SDValue X86TargetLowering::lowerFP_EXTEND(SDValue Op,
+ SelectionDAG &DAG) const {
+ DebugLoc DL = Op.getDebugLoc();
+ EVT VT = Op.getValueType();
+ SDValue In = Op.getOperand(0);
+ EVT SVT = In.getValueType();
+
+ assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
+
+ return DAG.getNode(X86ISD::VFPEXT, DL, VT,
+ DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
+ In, DAG.getUNDEF(SVT)));
+}
+
SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) const {
LLVMContext *Context = DAG.getContext();
DebugLoc dl = Op.getDebugLoc();
}
}
+ // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
+ // widen the cmov and push the truncate through. This avoids introducing a new
+ // branch during isel and doesn't add any extensions.
+ if (Op.getValueType() == MVT::i8 &&
+ Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
+ SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
+ if (T1.getValueType() == T2.getValueType() &&
+ // Blacklist CopyFromReg to avoid partial register stalls.
+ T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
+ SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
+ SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
+ return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
+ }
+ }
+
// X86ISD::CMOV means set the result (which is operand 1) to the RHS if
// condition is true.
SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
Flag = Chain.getValue(1);
- Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
+ Chain = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
+ SPTy).getValue(1);
SDValue Ops1[2] = { Chain.getValue(0), Chain };
return DAG.getMergeValues(Ops1, 2, dl);
// Sanity Check: Make sure using fp_offset makes sense.
assert(!getTargetMachine().Options.UseSoftFloat &&
!(DAG.getMachineFunction()
- .getFunction()->getFnAttributes().hasNoImplicitFloatAttr()) &&
+ .getFunction()->getFnAttributes()
+ .hasAttribute(Attributes::NoImplicitFloat)) &&
Subtarget->hasSSE1());
}
unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
DebugLoc dl = Op.getDebugLoc();
+ EVT PtrVT = getPointerTy();
if (Depth > 0) {
SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
SDValue Offset =
- DAG.getConstant(TD->getPointerSize(),
- Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
- return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
- DAG.getNode(ISD::ADD, dl, getPointerTy(),
+ DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
+ return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
+ DAG.getNode(ISD::ADD, dl, PtrVT,
FrameAddr, Offset),
MachinePointerInfo(), false, false, false, 0);
}
// Just load the return address.
SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
- return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
+ return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
RetAddrFI, MachinePointerInfo(), false, false, false, 0);
}
SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
SelectionDAG &DAG) const {
- return DAG.getIntPtrConstant(2*TD->getPointerSize());
+ return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
}
SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
- DAG.getIntPtrConstant(TD->getPointerSize()));
+ DAG.getIntPtrConstant(RegInfo->getSlotSize()));
StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
false, false, 0);
Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
}
+SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
+ SelectionDAG &DAG) const {
+ DebugLoc DL = Op.getDebugLoc();
+ return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
+ DAG.getVTList(MVT::i32, MVT::Other),
+ Op.getOperand(0), Op.getOperand(1));
+}
+
+SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
+ SelectionDAG &DAG) const {
+ DebugLoc DL = Op.getDebugLoc();
+ return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
+ Op.getOperand(0), Op.getOperand(1));
+}
+
static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
return Op.getOperand(0);
}
for (FunctionType::param_iterator I = FTy->param_begin(),
E = FTy->param_end(); I != E; ++I, ++Idx)
- if (Attrs.getParamAttributes(Idx).hasInRegAttr())
+ if (Attrs.getParamAttributes(Idx).hasAttribute(Attributes::InReg))
// FIXME: should only count parameters that are lowered to integers.
InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
+ case ISD::TRUNCATE: return lowerTRUNCATE(Op, DAG);
+ case ISD::ZERO_EXTEND: return lowerZERO_EXTEND(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
+ case ISD::FP_EXTEND: return lowerFP_EXTEND(Op, DAG);
case ISD::FABS: return LowerFABS(Op, DAG);
case ISD::FNEG: return LowerFNEG(Op, DAG);
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
+ case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
+ case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
}
return;
}
+ case ISD::UINT_TO_FP: {
+ if (N->getOperand(0).getValueType() != MVT::v2i32 &&
+ N->getValueType(0) != MVT::v2f32)
+ return;
+ SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
+ N->getOperand(0));
+ SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
+ MVT::f64);
+ SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
+ SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, VBias));
+ Or = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or);
+ SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
+ Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
+ return;
+ }
+ case ISD::FP_ROUND: {
+ SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
+ Results.push_back(V);
+ return;
+ }
case ISD::READCYCLECOUNTER: {
SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue TheChain = N->getOperand(0);
case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
+ case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
+ case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
case X86ISD::VSEXT_MOVL: return "X86ISD::VSEXT_MOVL";
case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
+ case X86ISD::VZEXT: return "X86ISD::VZEXT";
+ case X86ISD::VSEXT: return "X86ISD::VSEXT";
case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
+ case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
case X86ISD::VSHL: return "X86ISD::VSHL";
return BB;
}
+MachineBasicBlock *
+X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
+ MachineBasicBlock *MBB) const {
+ DebugLoc DL = MI->getDebugLoc();
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+
+ MachineFunction *MF = MBB->getParent();
+ MachineRegisterInfo &MRI = MF->getRegInfo();
+
+ const BasicBlock *BB = MBB->getBasicBlock();
+ MachineFunction::iterator I = MBB;
+ ++I;
+
+ // Memory Reference
+ MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
+ MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
+
+ unsigned DstReg;
+ unsigned MemOpndSlot = 0;
+
+ unsigned CurOp = 0;
+
+ DstReg = MI->getOperand(CurOp++).getReg();
+ const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
+ assert(RC->hasType(MVT::i32) && "Invalid destination!");
+ unsigned mainDstReg = MRI.createVirtualRegister(RC);
+ unsigned restoreDstReg = MRI.createVirtualRegister(RC);
+
+ MemOpndSlot = CurOp;
+
+ MVT PVT = getPointerTy();
+ assert((PVT == MVT::i64 || PVT == MVT::i32) &&
+ "Invalid Pointer Size!");
+
+ // For v = setjmp(buf), we generate
+ //
+ // thisMBB:
+ // buf[LabelOffset] = restoreMBB
+ // SjLjSetup restoreMBB
+ //
+ // mainMBB:
+ // v_main = 0
+ //
+ // sinkMBB:
+ // v = phi(main, restore)
+ //
+ // restoreMBB:
+ // v_restore = 1
+
+ MachineBasicBlock *thisMBB = MBB;
+ MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
+ MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
+ MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
+ MF->insert(I, mainMBB);
+ MF->insert(I, sinkMBB);
+ MF->push_back(restoreMBB);
+
+ MachineInstrBuilder MIB;
+
+ // Transfer the remainder of BB and its successor edges to sinkMBB.
+ sinkMBB->splice(sinkMBB->begin(), MBB,
+ llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
+ sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
+
+ // thisMBB:
+ unsigned PtrStoreOpc = 0;
+ unsigned LabelReg = 0;
+ const int64_t LabelOffset = 1 * PVT.getStoreSize();
+ Reloc::Model RM = getTargetMachine().getRelocationModel();
+ bool UseImmLabel = (getTargetMachine().getCodeModel() == CodeModel::Small) &&
+ (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
+
+ // Prepare IP either in reg or imm.
+ if (!UseImmLabel) {
+ PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
+ const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
+ LabelReg = MRI.createVirtualRegister(PtrRC);
+ if (Subtarget->is64Bit()) {
+ MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
+ .addReg(X86::RIP)
+ .addImm(0)
+ .addReg(0)
+ .addMBB(restoreMBB)
+ .addReg(0);
+ } else {
+ const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
+ MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
+ .addReg(XII->getGlobalBaseReg(MF))
+ .addImm(0)
+ .addReg(0)
+ .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
+ .addReg(0);
+ }
+ } else
+ PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
+ // Store IP
+ MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+ if (i == X86::AddrDisp)
+ MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
+ else
+ MIB.addOperand(MI->getOperand(MemOpndSlot + i));
+ }
+ if (!UseImmLabel)
+ MIB.addReg(LabelReg);
+ else
+ MIB.addMBB(restoreMBB);
+ MIB.setMemRefs(MMOBegin, MMOEnd);
+ // Setup
+ MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
+ .addMBB(restoreMBB);
+ MIB.addRegMask(RegInfo->getNoPreservedMask());
+ thisMBB->addSuccessor(mainMBB);
+ thisMBB->addSuccessor(restoreMBB);
+
+ // mainMBB:
+ // EAX = 0
+ BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
+ mainMBB->addSuccessor(sinkMBB);
+
+ // sinkMBB:
+ BuildMI(*sinkMBB, sinkMBB->begin(), DL,
+ TII->get(X86::PHI), DstReg)
+ .addReg(mainDstReg).addMBB(mainMBB)
+ .addReg(restoreDstReg).addMBB(restoreMBB);
+
+ // restoreMBB:
+ BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
+ BuildMI(restoreMBB, DL, TII->get(X86::JMP_4)).addMBB(sinkMBB);
+ restoreMBB->addSuccessor(sinkMBB);
+
+ MI->eraseFromParent();
+ return sinkMBB;
+}
+
+MachineBasicBlock *
+X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
+ MachineBasicBlock *MBB) const {
+ DebugLoc DL = MI->getDebugLoc();
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+
+ MachineFunction *MF = MBB->getParent();
+ MachineRegisterInfo &MRI = MF->getRegInfo();
+
+ // Memory Reference
+ MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
+ MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
+
+ MVT PVT = getPointerTy();
+ assert((PVT == MVT::i64 || PVT == MVT::i32) &&
+ "Invalid Pointer Size!");
+
+ const TargetRegisterClass *RC =
+ (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
+ unsigned Tmp = MRI.createVirtualRegister(RC);
+ // Since FP is only updated here but NOT referenced, it's treated as GPR.
+ unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
+ unsigned SP = RegInfo->getStackRegister();
+
+ MachineInstrBuilder MIB;
+
+ const int64_t LabelOffset = 1 * PVT.getStoreSize();
+ const int64_t SPOffset = 2 * PVT.getStoreSize();
+
+ unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
+ unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
+
+ // Reload FP
+ MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
+ MIB.addOperand(MI->getOperand(i));
+ MIB.setMemRefs(MMOBegin, MMOEnd);
+ // Reload IP
+ MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+ if (i == X86::AddrDisp)
+ MIB.addDisp(MI->getOperand(i), LabelOffset);
+ else
+ MIB.addOperand(MI->getOperand(i));
+ }
+ MIB.setMemRefs(MMOBegin, MMOEnd);
+ // Reload SP
+ MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
+ for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
+ if (i == X86::AddrDisp)
+ MIB.addDisp(MI->getOperand(i), SPOffset);
+ else
+ MIB.addOperand(MI->getOperand(i));
+ }
+ MIB.setMemRefs(MMOBegin, MMOEnd);
+ // Jump
+ BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
+
+ MI->eraseFromParent();
+ return MBB;
+}
+
MachineBasicBlock *
X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
MachineBasicBlock *BB) const {
case X86::VAARG_64:
return EmitVAARG64WithCustomInserter(MI, BB);
+
+ case X86::EH_SjLj_SetJmp32:
+ case X86::EH_SjLj_SetJmp64:
+ return emitEHSjLjSetJmp(MI, BB);
+
+ case X86::EH_SjLj_LongJmp32:
+ case X86::EH_SjLj_LongJmp64:
+ return emitEHSjLjLongJmp(MI, BB);
}
}
return NewOp;
SDValue InputVector = N->getOperand(0);
+ // Detect whether we are trying to convert from mmx to i32 and the bitcast
+ // from mmx to v2i32 has a single usage.
+ if (InputVector.getNode()->getOpcode() == llvm::ISD::BITCAST &&
+ InputVector.getNode()->getOperand(0).getValueType() == MVT::x86mmx &&
+ InputVector.hasOneUse() && N->getValueType(0) == MVT::i32)
+ return DAG.getNode(X86ISD::MMX_MOVD2W, InputVector.getDebugLoc(),
+ N->getValueType(0),
+ InputVector.getNode()->getOperand(0));
// Only operate on vectors of 4 elements, where the alternative shuffling
// gets to be more expensive.
if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
CC = X86::GetOppositeBranchCondition(CC);
std::swap(TrueC, FalseC);
+ std::swap(TrueOp, FalseOp);
}
// Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
}
}
}
+
+ // Handle these cases:
+ // (select (x != c), e, c) -> select (x != c), e, x),
+ // (select (x == c), c, e) -> select (x == c), x, e)
+ // where the c is an integer constant, and the "select" is the combination
+ // of CMOV and CMP.
+ //
+ // The rationale for this change is that the conditional-move from a constant
+ // needs two instructions, however, conditional-move from a register needs
+ // only one instruction.
+ //
+ // CAVEAT: By replacing a constant with a symbolic value, it may obscure
+ // some instruction-combining opportunities. This opt needs to be
+ // postponed as late as possible.
+ //
+ if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
+ // the DCI.xxxx conditions are provided to postpone the optimization as
+ // late as possible.
+
+ ConstantSDNode *CmpAgainst = 0;
+ if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
+ (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
+ dyn_cast<ConstantSDNode>(Cond.getOperand(0)) == 0) {
+
+ if (CC == X86::COND_NE &&
+ CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
+ CC = X86::GetOppositeBranchCondition(CC);
+ std::swap(TrueOp, FalseOp);
+ }
+
+ if (CC == X86::COND_E &&
+ CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
+ SDValue Ops[] = { FalseOp, Cond.getOperand(0),
+ DAG.getConstant(CC, MVT::i8), Cond };
+ return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops,
+ array_lengthof(Ops));
+ }
+ }
+ }
+
return SDValue();
}
ISD::LoadExtType Ext = Ld->getExtensionType();
// If this is a vector EXT Load then attempt to optimize it using a
- // shuffle. We need SSE4 for the shuffles.
+ // shuffle. We need SSSE3 shuffles.
// TODO: It is possible to support ZExt by zeroing the undef values
// during the shuffle phase or after the shuffle.
if (RegVT.isVector() && RegVT.isInteger() &&
- Ext == ISD::EXTLOAD && Subtarget->hasSSE41()) {
+ Ext == ISD::EXTLOAD && Subtarget->hasSSSE3()) {
assert(MemVT != RegVT && "Cannot extend to the same type");
assert(MemVT.isVector() && "Must load a vector from memory");
return SDValue();
const Function *F = DAG.getMachineFunction().getFunction();
- bool NoImplicitFloatOps = F->getFnAttributes().hasNoImplicitFloatAttr();
+ bool NoImplicitFloatOps = F->getFnAttributes().
+ hasAttribute(Attributes::NoImplicitFloat);
bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
&& Subtarget->hasSSE2();
if ((VT.isVector() ||
return SDValue();
}
+// Helper function of PerformSETCCCombine. It is to materialize "setb reg"
+// as "sbb reg,reg", since it can be extended without zext and produces
+// an all-ones bit which is more useful than 0/1 in some cases.
+static SDValue MaterializeSETB(DebugLoc DL, SDValue EFLAGS, SelectionDAG &DAG) {
+ return DAG.getNode(ISD::AND, DL, MVT::i8,
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
+ DAG.getConstant(X86::COND_B, MVT::i8), EFLAGS),
+ DAG.getConstant(1, MVT::i8));
+}
+
// Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
SDValue EFLAGS = N->getOperand(1);
+ if (CC == X86::COND_A) {
+ // Try to convert COND_A into COND_B in an attempt to facilitate
+ // materializing "setb reg".
+ //
+ // Do not flip "e > c", where "c" is a constant, because Cmp instruction
+ // cannot take an immediate as its first operand.
+ //
+ if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
+ EFLAGS.getValueType().isInteger() &&
+ !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
+ SDValue NewSub = DAG.getNode(X86ISD::SUB, EFLAGS.getDebugLoc(),
+ EFLAGS.getNode()->getVTList(),
+ EFLAGS.getOperand(1), EFLAGS.getOperand(0));
+ SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
+ return MaterializeSETB(DL, NewEFLAGS, DAG);
+ }
+ }
+
// Materialize "setb reg" as "sbb reg,reg", since it can be extended without
// a zext and produces an all-ones bit which is more useful than 0/1 in some
// cases.
if (CC == X86::COND_B)
- return DAG.getNode(ISD::AND, DL, MVT::i8,
- DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
- DAG.getConstant(CC, MVT::i8), EFLAGS),
- DAG.getConstant(1, MVT::i8));
+ return MaterializeSETB(DL, EFLAGS, DAG);
SDValue Flags;
return SDValue();
}
-static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG) {
- SDValue Op0 = N->getOperand(0);
- EVT InVT = Op0->getValueType(0);
-
- // UINT_TO_FP(v4i8) -> SINT_TO_FP(ZEXT(v4i8 to v4i32))
- if (InVT == MVT::v8i8 || InVT == MVT::v4i8) {
- DebugLoc dl = N->getDebugLoc();
- MVT DstVT = InVT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
- SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
- // Notice that we use SINT_TO_FP because we know that the high bits
- // are zero and SINT_TO_FP is better supported by the hardware.
- return DAG.getNode(ISD::SINT_TO_FP, dl, N->getValueType(0), P);
- }
-
- return SDValue();
-}
-
static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
const X86TargetLowering *XTLI) {
SDValue Op0 = N->getOperand(0);
return SDValue();
}
-static SDValue PerformFP_TO_SINTCombine(SDNode *N, SelectionDAG &DAG) {
- EVT VT = N->getValueType(0);
-
- // v4i8 = FP_TO_SINT() -> v4i8 = TRUNCATE (V4i32 = FP_TO_SINT()
- if (VT == MVT::v8i8 || VT == MVT::v4i8) {
- DebugLoc dl = N->getDebugLoc();
- MVT DstVT = VT == MVT::v4i8 ? MVT::v4i32 : MVT::v8i32;
- SDValue I = DAG.getNode(ISD::FP_TO_SINT, dl, DstVT, N->getOperand(0));
- return DAG.getNode(ISD::TRUNCATE, dl, VT, I);
- }
-
- return SDValue();
-}
-
// Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
X86TargetLowering::DAGCombinerInfo &DCI) {
return OptimizeConditionalInDecrement(N, DAG);
}
+/// performVZEXTCombine - Performs build vector combines
+static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ // (vzext (bitcast (vzext (x)) -> (vzext x)
+ SDValue In = N->getOperand(0);
+ while (In.getOpcode() == ISD::BITCAST)
+ In = In.getOperand(0);
+
+ if (In.getOpcode() != X86ISD::VZEXT)
+ return SDValue();
+
+ return DAG.getNode(X86ISD::VZEXT, N->getDebugLoc(), N->getValueType(0), In.getOperand(0));
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
- case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG);
case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
- case ISD::FP_TO_SINT: return PerformFP_TO_SINTCombine(N, DAG);
case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
case X86ISD::FXOR:
case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
+ case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
case X86ISD::SHUFP: // Handle all target specific shuffles
case X86ISD::PALIGN:
case X86ISD::UNPCKH:
return Res;
}
+
+//===----------------------------------------------------------------------===//
+//
+// X86 cost model.
+//
+//===----------------------------------------------------------------------===//
+
+struct X86CostTblEntry {
+ int ISD;
+ MVT Type;
+ unsigned Cost;
+};
+
+unsigned
+X86VectorTargetTransformInfo::getArithmeticInstrCost(unsigned Opcode,
+ Type *Ty) const {
+ // Legalize the type.
+ std::pair<unsigned, MVT> LT =
+ getTypeLegalizationCost(Ty->getContext(), TLI->getValueType(Ty));
+
+ int ISD = InstructionOpcodeToISD(Opcode);
+ assert(ISD && "Invalid opcode");
+
+ const X86Subtarget &ST =
+ TLI->getTargetMachine().getSubtarget<X86Subtarget>();
+
+ static const X86CostTblEntry AVX1CostTable[] = {
+ // We don't have to scalarize unsupported ops. We can issue two half-sized
+ // operations and we only need to extract the upper YMM half.
+ // Two ops + 1 extract + 1 insert = 4.
+ { ISD::MUL, MVT::v8i32, 4 },
+ { ISD::SUB, MVT::v8i32, 4 },
+ { ISD::ADD, MVT::v8i32, 4 },
+ { ISD::MUL, MVT::v4i64, 4 },
+ { ISD::SUB, MVT::v4i64, 4 },
+ { ISD::ADD, MVT::v4i64, 4 },
+ };
+
+ // Look for AVX1 lowering tricks.
+ if (ST.hasAVX())
+ for (unsigned int i = 0, e = array_lengthof(AVX1CostTable); i < e; ++i) {
+ if (AVX1CostTable[i].ISD == ISD && AVX1CostTable[i].Type == LT.second)
+ return LT.first * AVX1CostTable[i].Cost;
+ }
+
+ // Fallback to the default implementation.
+ return VectorTargetTransformImpl::getArithmeticInstrCost(Opcode, Ty);
+}
+
+unsigned
+X86VectorTargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val,
+ unsigned Index) const {
+ assert(Val->isVectorTy() && "This must be a vector type");
+
+ if (Index != -1u) {
+ // Legalize the type.
+ std::pair<unsigned, MVT> LT =
+ getTypeLegalizationCost(Val->getContext(), TLI->getValueType(Val));
+
+ // This type is legalized to a scalar type.
+ if (!LT.second.isVector())
+ return 0;
+
+ // The type may be split. Normalize the index to the new type.
+ unsigned Width = LT.second.getVectorNumElements();
+ Index = Index % Width;
+
+ // Floating point scalars are already located in index #0.
+ if (Val->getScalarType()->isFloatingPointTy() && Index == 0)
+ return 0;
+ }
+
+ return VectorTargetTransformImpl::getVectorInstrCost(Opcode, Val, Index);
+}
+
projects / oota-llvm.git / blobdiff