else
setSchedulingPreference(Sched::RegPressure);
const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
+ TM.getSubtarget<X86Subtarget>().getRegisterInfo();
setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
// Bypass expensive divides on Atom when compiling with O2
// If we don't have F16C support, then lower half float conversions
// into library calls.
if (TM.Options.UseSoftFloat || !Subtarget->hasF16C()) {
- setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
- setOperationAction(ISD::FP32_TO_FP16, MVT::i16, Expand);
+ setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
+ setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
}
+ // There's never any support for operations beyond MVT::f32.
+ setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
+ setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
+ setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
+ setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
+
+ setLoadExtAction(ISD::EXTLOAD, MVT::f16, Expand);
+ setTruncStoreAction(MVT::f32, MVT::f16, Expand);
+ setTruncStoreAction(MVT::f64, MVT::f16, Expand);
+ setTruncStoreAction(MVT::f80, MVT::f16, Expand);
+
if (Subtarget->hasPOPCNT()) {
setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
} else {
setOperationAction(ISD::TRUNCATE, VT, Expand);
setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
- setOperationAction(ISD::ZERO_EXTEND_VECTOR_INREG, VT, Expand);
setOperationAction(ISD::ANY_EXTEND, VT, Expand);
setOperationAction(ISD::VSELECT, VT, Expand);
setOperationAction(ISD::SELECT_CC, VT, Expand);
(MVT::SimpleValueType)InnerVT, Expand);
setLoadExtAction(ISD::SEXTLOAD, VT, Expand);
setLoadExtAction(ISD::ZEXTLOAD, VT, Expand);
- setLoadExtAction(ISD::EXTLOAD, VT, Expand);
+
+ // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like types,
+ // we have to deal with them whether we ask for Expansion or not. Setting
+ // Expand causes its own optimisation problems though, so leave them legal.
+ if (VT.getVectorElementType() == MVT::i1)
+ setLoadExtAction(ISD::EXTLOAD, VT, Expand);
}
// FIXME: In order to prevent SSE instructions being expanded to MMX ones
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
}
+ // We support custom legalizing of sext and anyext loads for specific
+ // memory vector types which we can load as a scalar (or sequence of
+ // scalars) and extend in-register to a legal 128-bit vector type. For sext
+ // loads these must work with a single scalar load.
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i8, Custom);
+ if (Subtarget->is64Bit()) {
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v4i16, Custom);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v8i8, Custom);
+ }
+ setLoadExtAction(ISD::EXTLOAD, MVT::v2i8, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v2i16, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v2i32, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v4i8, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v4i16, Custom);
+ setLoadExtAction(ISD::EXTLOAD, MVT::v8i8, Custom);
+
setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
// some vselects for now.
setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
+ // SSE41 brings specific instructions for doing vector sign extend even in
+ // cases where we don't have SRA.
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i8, Custom);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i16, Custom);
+ setLoadExtAction(ISD::SEXTLOAD, MVT::v2i32, Custom);
+
// i8 and i16 vectors are custom , because the source register and source
// source memory operand types are not the same width. f32 vectors are
// custom since the immediate controlling the insert encodes additional
}
}// has AVX-512
+ if (!TM.Options.UseSoftFloat && Subtarget->hasBWI()) {
+ addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
+ addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
+ }
+
// SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
// of this type with custom code.
for (int VT = MVT::FIRST_VECTOR_VALUETYPE;
setPrefFunctionAlignment(4); // 2^4 bytes.
}
+// This has so far only been implemented for 64-bit MachO.
+bool X86TargetLowering::useLoadStackGuardNode() const {
+ return Subtarget->getTargetTriple().getObjectFormat() == Triple::MachO &&
+ Subtarget->is64Bit();
+}
+
TargetLoweringBase::LegalizeTypeAction
X86TargetLowering::getPreferredVectorAction(EVT VT) const {
if (ExperimentalVectorWideningLegalization &&
}
bool
-X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
- unsigned,
- bool *Fast) const {
+X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
+ unsigned,
+ unsigned,
+ bool *Fast) const {
if (Fast)
*Fast = Subtarget->isUnalignedMemAccessFast();
return true;
// Returns in ST0/ST1 are handled specially: these are pushed as operands to
// the RET instruction and handled by the FP Stackifier.
- if (VA.getLocReg() == X86::ST0 ||
- VA.getLocReg() == X86::ST1) {
+ if (VA.getLocReg() == X86::FP0 ||
+ VA.getLocReg() == X86::FP1) {
// If this is a copy from an xmm register to ST(0), use an FPExtend to
// change the value to the FP stack register class.
if (isScalarFPTypeInSSEReg(VA.getValVT()))
report_fatal_error("SSE register return with SSE disabled");
}
- SDValue Val;
-
- // If this is a call to a function that returns an fp value on the floating
- // point stack, we must guarantee the value is popped from the stack, so
- // a CopyFromReg is not good enough - the copy instruction may be eliminated
- // if the return value is not used. We use the FpPOP_RETVAL instruction
- // instead.
- if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
- // If we prefer to use the value in xmm registers, copy it out as f80 and
- // use a truncate to move it from fp stack reg to xmm reg.
- if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
- SDValue Ops[] = { Chain, InFlag };
- Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
- MVT::Other, MVT::Glue, Ops), 1);
- Val = Chain.getValue(0);
-
- // Round the f80 to the right size, which also moves it to the appropriate
- // xmm register.
- if (CopyVT != VA.getValVT())
- Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
- // This truncation won't change the value.
- DAG.getIntPtrConstant(1));
- } else {
- Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
- CopyVT, InFlag).getValue(1);
- Val = Chain.getValue(0);
- }
+ // If we prefer to use the value in xmm registers, copy it out as f80 and
+ // use a truncate to move it from fp stack reg to xmm reg.
+ if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
+ isScalarFPTypeInSSEReg(VA.getValVT()))
+ CopyVT = MVT::f80;
+
+ Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
+ CopyVT, InFlag).getValue(1);
+ SDValue Val = Chain.getValue(0);
+
+ if (CopyVT != VA.getValVT())
+ Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
+ // This truncation won't change the value.
+ DAG.getIntPtrConstant(1));
+
InFlag = Chain.getValue(2);
InVals.push_back(Val);
}
RC = &X86::VK8RegClass;
else if (RegVT == MVT::v16i1)
RC = &X86::VK16RegClass;
+ else if (RegVT == MVT::v32i1)
+ RC = &X86::VK32RegClass;
+ else if (RegVT == MVT::v64i1)
+ RC = &X86::VK64RegClass;
else
llvm_unreachable("Unknown argument type!");
TotalNumXMMRegs = 0;
if (IsWin64) {
- const TargetFrameLowering &TFI = *MF.getTarget().getFrameLowering();
+ const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
// Get to the caller-allocated home save location. Add 8 to account
// for the return address.
int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
// arguments passed in memory when using inalloca.
if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
NumBytesToPush = 0;
- assert(ArgLocs.back().getLocMemOffset() == 0 &&
- "an inalloca argument must be the only memory argument");
+ if (!ArgLocs.back().isMemLoc())
+ report_fatal_error("cannot use inalloca attribute on a register "
+ "parameter");
+ if (ArgLocs.back().getLocMemOffset() != 0)
+ report_fatal_error("any parameter with the inalloca attribute must be "
+ "the only memory argument");
}
if (!IsSibcall)
// Walk the register/memloc assignments, inserting copies/loads. In the case
// of tail call optimization arguments are handle later.
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
// Skip inalloca arguments, they have already been written.
ISD::ArgFlagsTy Flags = Outs[i].Flags;
RegsToPass[i].second.getValueType()));
// Add a register mask operand representing the call-preserved registers.
- const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
const uint32_t *Mask = TRI->getCallPreservedMask(CallConv);
assert(Mask && "Missing call preserved mask for calling convention");
Ops.push_back(DAG.getRegisterMask(Mask));
// If a tail called function callee has more arguments than the caller the
// caller needs to make sure that there is room to move the RETADDR to. This is
// achieved by reserving an area the size of the argument delta right after the
-// original REtADDR, but before the saved framepointer or the spilled registers
+// original RETADDR, but before the saved framepointer or the spilled registers
// e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
// stack layout:
// arg1
SelectionDAG& DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
const TargetMachine &TM = MF.getTarget();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
- const TargetFrameLowering &TFI = *TM.getFrameLowering();
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ TM.getSubtargetImpl()->getRegisterInfo());
+ const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
uint64_t AlignMask = StackAlignment - 1;
int64_t Offset = StackSize;
// Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
// emit a special epilogue.
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
if (RegInfo->needsStackRealignment(MF))
return false;
CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
CCValAssign &VA = RVLocs[i];
- if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
+ if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
return false;
}
}
MachineFrameInfo *MFI = MF.getFrameInfo();
const MachineRegisterInfo *MRI = &MF.getRegInfo();
const X86InstrInfo *TII =
- static_cast<const X86InstrInfo *>(DAG.getTarget().getInstrInfo());
+ static_cast<const X86InstrInfo *>(DAG.getSubtarget().getInstrInfo());
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign &VA = ArgLocs[i];
SDValue Arg = OutVals[i];
static bool isTargetShuffle(unsigned Opcode) {
switch(Opcode) {
default: return false;
+ case X86ISD::PSHUFB:
case X86ISD::PSHUFD:
case X86ISD::PSHUFHW:
case X86ISD::PSHUFLW:
SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
int ReturnAddrIndex = FuncInfo->getRAIndex();
return false;
}
-/// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
-/// their permute mask.
-static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
- SelectionDAG &DAG) {
- MVT VT = SVOp->getSimpleValueType(0);
- unsigned NumElems = VT.getVectorNumElements();
- SmallVector<int, 8> MaskVec;
-
- for (unsigned i = 0; i != NumElems; ++i) {
- int Idx = SVOp->getMaskElt(i);
- if (Idx >= 0) {
- if (Idx < (int)NumElems)
- Idx += NumElems;
- else
- Idx -= NumElems;
- }
- MaskVec.push_back(Idx);
- }
- return DAG.getVectorShuffle(VT, SDLoc(SVOp), SVOp->getOperand(1),
- SVOp->getOperand(0), &MaskVec[0]);
-}
-
/// ShouldXformToMOVHLPS - Return true if the node should be transformed to
/// match movhlps. The lower half elements should come from upper half of
/// V1 (and in order), and the upper half elements should come from the upper
}
/// getTargetShuffleMask - Calculates the shuffle mask corresponding to the
-/// target specific opcode. Returns true if the Mask could be calculated.
-/// Sets IsUnary to true if only uses one source.
+/// target specific opcode. Returns true if the Mask could be calculated. Sets
+/// IsUnary to true if only uses one source. Note that this will set IsUnary for
+/// shuffles which use a single input multiple times, and in those cases it will
+/// adjust the mask to only have indices within that single input.
static bool getTargetShuffleMask(SDNode *N, MVT VT,
SmallVectorImpl<int> &Mask, bool &IsUnary) {
unsigned NumElems = VT.getVectorNumElements();
SDValue ImmN;
IsUnary = false;
+ bool IsFakeUnary = false;
switch(N->getOpcode()) {
case X86ISD::SHUFP:
ImmN = N->getOperand(N->getNumOperands()-1);
DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::UNPCKH:
DecodeUNPCKHMask(VT, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::UNPCKL:
DecodeUNPCKLMask(VT, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::MOVHLPS:
DecodeMOVHLPSMask(NumElems, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::MOVLHPS:
DecodeMOVLHPSMask(NumElems, Mask);
+ IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
break;
case X86ISD::PALIGNR:
ImmN = N->getOperand(N->getNumOperands()-1);
DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
IsUnary = true;
break;
+ case X86ISD::PSHUFB: {
+ IsUnary = true;
+ SDValue MaskNode = N->getOperand(1);
+ while (MaskNode->getOpcode() == ISD::BITCAST)
+ MaskNode = MaskNode->getOperand(0);
+
+ if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
+ // If we have a build-vector, then things are easy.
+ EVT VT = MaskNode.getValueType();
+ assert(VT.isVector() &&
+ "Can't produce a non-vector with a build_vector!");
+ if (!VT.isInteger())
+ return false;
+
+ int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
+
+ SmallVector<uint64_t, 32> RawMask;
+ for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
+ auto *CN = dyn_cast<ConstantSDNode>(MaskNode->getOperand(i));
+ if (!CN)
+ return false;
+ APInt MaskElement = CN->getAPIntValue();
+
+ // We now have to decode the element which could be any integer size and
+ // extract each byte of it.
+ for (int j = 0; j < NumBytesPerElement; ++j) {
+ // Note that this is x86 and so always little endian: the low byte is
+ // the first byte of the mask.
+ RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
+ MaskElement = MaskElement.lshr(8);
+ }
+ }
+ DecodePSHUFBMask(RawMask, Mask);
+ break;
+ }
+
+ auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
+ if (!MaskLoad)
+ return false;
+
+ SDValue Ptr = MaskLoad->getBasePtr();
+ if (Ptr->getOpcode() == X86ISD::Wrapper)
+ Ptr = Ptr->getOperand(0);
+
+ auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
+ if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
+ return false;
+
+ if (auto *C = dyn_cast<ConstantDataSequential>(MaskCP->getConstVal())) {
+ // FIXME: Support AVX-512 here.
+ if (!C->getType()->isVectorTy() ||
+ (C->getNumElements() != 16 && C->getNumElements() != 32))
+ return false;
+
+ assert(C->getType()->isVectorTy() && "Expected a vector constant.");
+ DecodePSHUFBMask(C, Mask);
+ break;
+ }
+
+ return false;
+ }
case X86ISD::VPERMI:
ImmN = N->getOperand(N->getNumOperands()-1);
DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
default: llvm_unreachable("unknown target shuffle node");
}
+ // If we have a fake unary shuffle, the shuffle mask is spread across two
+ // inputs that are actually the same node. Re-map the mask to always point
+ // into the first input.
+ if (IsFakeUnary)
+ for (int &M : Mask)
+ if (M >= (int)Mask.size())
+ M -= Mask.size();
+
return true;
}
if (GoodInputs.size() == 2) {
// If the low inputs are spread across two dwords, pack them into
// a single dword.
- MoveMask[Mask[GoodInputs[0]] % 2 + MoveOffset] =
- Mask[GoodInputs[0]] - MaskOffset;
- MoveMask[Mask[GoodInputs[1]] % 2 + MoveOffset] =
- Mask[GoodInputs[1]] - MaskOffset;
- Mask[GoodInputs[0]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
- Mask[GoodInputs[1]] = Mask[GoodInputs[0]] % 2 + MoveOffset + MaskOffset;
+ MoveMask[MoveOffset] = Mask[GoodInputs[0]] - MaskOffset;
+ MoveMask[MoveOffset + 1] = Mask[GoodInputs[1]] - MaskOffset;
+ Mask[GoodInputs[0]] = MoveOffset + MaskOffset;
+ Mask[GoodInputs[1]] = MoveOffset + 1 + MaskOffset;
} else {
- // Otherwise pin the low inputs.
+ // Otherwise pin the good inputs.
for (int GoodInput : GoodInputs)
MoveMask[Mask[GoodInput] - MaskOffset] = Mask[GoodInput] - MaskOffset;
}
- int MoveMaskIdx =
- std::find(std::begin(MoveMask) + MoveOffset, std::end(MoveMask), -1) -
- std::begin(MoveMask);
- assert(MoveMaskIdx >= MoveOffset && "Established above");
-
if (BadInputs.size() == 2) {
+ // If we have two bad inputs then there may be either one or two good
+ // inputs fixed in place. Find a fixed input, and then find the *other*
+ // two adjacent indices by using modular arithmetic.
+ int GoodMaskIdx =
+ std::find_if(std::begin(MoveMask) + MoveOffset, std::end(MoveMask),
+ [](int M) { return M >= 0; }) -
+ std::begin(MoveMask);
+ int MoveMaskIdx =
+ (((GoodMaskIdx - MoveOffset) & ~1) + 2 % 4) + MoveOffset;
assert(MoveMask[MoveMaskIdx] == -1 && "Expected empty slot");
assert(MoveMask[MoveMaskIdx + 1] == -1 && "Expected empty slot");
- MoveMask[MoveMaskIdx + Mask[BadInputs[0]] % 2] =
- Mask[BadInputs[0]] - MaskOffset;
- MoveMask[MoveMaskIdx + Mask[BadInputs[1]] % 2] =
- Mask[BadInputs[1]] - MaskOffset;
- Mask[BadInputs[0]] = MoveMaskIdx + Mask[BadInputs[0]] % 2 + MaskOffset;
- Mask[BadInputs[1]] = MoveMaskIdx + Mask[BadInputs[1]] % 2 + MaskOffset;
+ MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
+ MoveMask[MoveMaskIdx + 1] = Mask[BadInputs[1]] - MaskOffset;
+ Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
+ Mask[BadInputs[1]] = MoveMaskIdx + 1 + MaskOffset;
} else {
assert(BadInputs.size() == 1 && "All sizes handled");
+ int MoveMaskIdx =
+ std::find(std::begin(MoveMask) + MoveOffset, std::end(MoveMask), -1) -
+ std::begin(MoveMask);
MoveMask[MoveMaskIdx] = Mask[BadInputs[0]] - MaskOffset;
Mask[BadInputs[0]] = MoveMaskIdx + MaskOffset;
}
DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, LoV, HiV));
}
+/// \brief Check whether a compaction lowering can be done by dropping even
+/// elements and compute how many times even elements must be dropped.
+///
+/// This handles shuffles which take every Nth element where N is a power of
+/// two. Example shuffle masks:
+///
+/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
+/// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
+/// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
+/// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
+/// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
+/// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
+///
+/// Any of these lanes can of course be undef.
+///
+/// This routine only supports N <= 3.
+/// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
+/// for larger N.
+///
+/// \returns N above, or the number of times even elements must be dropped if
+/// there is such a number. Otherwise returns zero.
+static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
+ // Figure out whether we're looping over two inputs or just one.
+ bool IsSingleInput = isSingleInputShuffleMask(Mask);
+
+ // The modulus for the shuffle vector entries is based on whether this is
+ // a single input or not.
+ int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
+ assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
+ "We should only be called with masks with a power-of-2 size!");
+
+ uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
+
+ // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
+ // and 2^3 simultaneously. This is because we may have ambiguity with
+ // partially undef inputs.
+ bool ViableForN[3] = {true, true, true};
+
+ for (int i = 0, e = Mask.size(); i < e; ++i) {
+ // Ignore undef lanes, we'll optimistically collapse them to the pattern we
+ // want.
+ if (Mask[i] == -1)
+ continue;
+
+ bool IsAnyViable = false;
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j]) {
+ uint64_t N = j + 1;
+
+ // The shuffle mask must be equal to (i * 2^N) % M.
+ if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
+ IsAnyViable = true;
+ else
+ ViableForN[j] = false;
+ }
+ // Early exit if we exhaust the possible powers of two.
+ if (!IsAnyViable)
+ break;
+ }
+
+ for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
+ if (ViableForN[j])
+ return j + 1;
+
+ // Return 0 as there is no viable power of two.
+ return 0;
+}
+
/// \brief Generic lowering of v16i8 shuffles.
///
/// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
if (isSingleInputShuffleMask(Mask)) {
// Check whether we can widen this to an i16 shuffle by duplicating bytes.
// Notably, this handles splat and partial-splat shuffles more efficiently.
+ // However, it only makes sense if the pre-duplication shuffle simplifies
+ // things significantly. Currently, this means we need to be able to
+ // express the pre-duplication shuffle as an i16 shuffle.
//
// FIXME: We should check for other patterns which can be widened into an
// i16 shuffle as well.
}
return true;
};
- if (canWidenViaDuplication(Mask)) {
+ auto tryToWidenViaDuplication = [&]() -> SDValue {
+ if (!canWidenViaDuplication(Mask))
+ return SDValue();
SmallVector<int, 4> LoInputs;
std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
[](int M) { return M >= 0 && M < 8; });
ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
- int ByteMask[16];
+ int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
SmallDenseMap<int, int, 8> LaneMap;
- for (int i = 0; i < 16; ++i)
- ByteMask[i] = -1;
for (int I : InPlaceInputs) {
- ByteMask[I] = I;
+ PreDupI16Shuffle[I/2] = I/2;
LaneMap[I] = I;
}
- int FreeByteIdx = 0;
- int TargetOffset = TargetLo ? 0 : 8;
- for (int I : MovingInputs) {
- // Walk the free index into the byte mask until we find an unoccupied
- // spot. We bound this to 8 steps to catch bugs, the pigeonhole
- // principle indicates that there *must* be a spot as we can only have
- // 8 duplicated inputs. We have to walk the index using modular
- // arithmetic to wrap around as necessary.
- // FIXME: We could do a much better job of picking an inexpensive slot
- // so this doesn't go through the worst case for the byte shuffle.
- for (int j = 0; j < 8 && ByteMask[FreeByteIdx + TargetOffset] != -1;
- ++j, FreeByteIdx = (FreeByteIdx + 1) % 8)
- ;
- assert(ByteMask[FreeByteIdx + TargetOffset] == -1 &&
- "Failed to find a free byte!");
- ByteMask[FreeByteIdx + TargetOffset] = I;
- LaneMap[I] = FreeByteIdx + TargetOffset;
+ int j = TargetLo ? 0 : 4, je = j + 4;
+ for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
+ // Check if j is already a shuffle of this input. This happens when
+ // there are two adjacent bytes after we move the low one.
+ if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
+ // If we haven't yet mapped the input, search for a slot into which
+ // we can map it.
+ while (j < je && PreDupI16Shuffle[j] != -1)
+ ++j;
+
+ if (j == je)
+ // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
+ return SDValue();
+
+ // Map this input with the i16 shuffle.
+ PreDupI16Shuffle[j] = MovingInputs[i] / 2;
+ }
+
+ // Update the lane map based on the mapping we ended up with.
+ LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
}
- V1 = DAG.getVectorShuffle(MVT::v16i8, DL, V1, DAG.getUNDEF(MVT::v16i8),
- ByteMask);
- for (int &M : Mask)
- if (M != -1)
- M = LaneMap[M];
+ V1 = DAG.getNode(
+ ISD::BITCAST, DL, MVT::v16i8,
+ DAG.getVectorShuffle(MVT::v8i16, DL,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
+ DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
// Unpack the bytes to form the i16s that will be shuffled into place.
V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
MVT::v16i8, V1, V1);
- int I16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
+ int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
for (int i = 0; i < 16; i += 2) {
if (Mask[i] != -1)
- I16Shuffle[i / 2] = Mask[i] - (TargetLo ? 0 : 8);
- assert(I16Shuffle[i / 2] < 8 && "Invalid v8 shuffle mask!");
+ PostDupI16Shuffle[i / 2] = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
+ assert(PostDupI16Shuffle[i / 2] < 8 && "Invalid v8 shuffle mask!");
}
- return DAG.getVectorShuffle(MVT::v8i16, DL,
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
- DAG.getUNDEF(MVT::v8i16), I16Shuffle);
- }
+ return DAG.getNode(
+ ISD::BITCAST, DL, MVT::v16i8,
+ DAG.getVectorShuffle(MVT::v8i16, DL,
+ DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1),
+ DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
+ };
+ if (SDValue V = tryToWidenViaDuplication())
+ return V;
}
// Check whether an interleaving lowering is likely to be more efficient.
return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, Evens, Odds);
}
- SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
- SDValue LoV1 =
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, Zero));
- SDValue HiV1 =
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, Zero));
- SDValue LoV2 =
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V2, Zero));
- SDValue HiV2 =
- DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
- DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V2, Zero));
+
+ // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
+ // with PSHUFB. It is important to do this before we attempt to generate any
+ // blends but after all of the single-input lowerings. If the single input
+ // lowerings can find an instruction sequence that is faster than a PSHUFB, we
+ // want to preserve that and we can DAG combine any longer sequences into
+ // a PSHUFB in the end. But once we start blending from multiple inputs,
+ // the complexity of DAG combining bad patterns back into PSHUFB is too high,
+ // and there are *very* few patterns that would actually be faster than the
+ // PSHUFB approach because of its ability to zero lanes.
+ //
+ // FIXME: The only exceptions to the above are blends which are exact
+ // interleavings with direct instructions supporting them. We currently don't
+ // handle those well here.
+ if (Subtarget->hasSSSE3()) {
+ SDValue V1Mask[16];
+ SDValue V2Mask[16];
+ for (int i = 0; i < 16; ++i)
+ if (Mask[i] == -1) {
+ V1Mask[i] = V2Mask[i] = DAG.getConstant(0x80, MVT::i8);
+ } else {
+ V1Mask[i] = DAG.getConstant(Mask[i] < 16 ? Mask[i] : 0x80, MVT::i8);
+ V2Mask[i] =
+ DAG.getConstant(Mask[i] < 16 ? 0x80 : Mask[i] - 16, MVT::i8);
+ }
+ V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
+ if (isSingleInputShuffleMask(Mask))
+ return V1; // Single inputs are easy.
+
+ // Otherwise, blend the two.
+ V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, V2,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
+ return DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
+ }
+
+ // Check whether a compaction lowering can be done. This handles shuffles
+ // which take every Nth element for some even N. See the helper function for
+ // details.
+ //
+ // We special case these as they can be particularly efficiently handled with
+ // the PACKUSB instruction on x86 and they show up in common patterns of
+ // rearranging bytes to truncate wide elements.
+ if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
+ // NumEvenDrops is the power of two stride of the elements. Another way of
+ // thinking about it is that we need to drop the even elements this many
+ // times to get the original input.
+ bool IsSingleInput = isSingleInputShuffleMask(Mask);
+
+ // First we need to zero all the dropped bytes.
+ assert(NumEvenDrops <= 3 &&
+ "No support for dropping even elements more than 3 times.");
+ // We use the mask type to pick which bytes are preserved based on how many
+ // elements are dropped.
+ MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
+ SDValue ByteClearMask =
+ DAG.getNode(ISD::BITCAST, DL, MVT::v16i8,
+ DAG.getConstant(0xFF, MaskVTs[NumEvenDrops - 1]));
+ V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
+ if (!IsSingleInput)
+ V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
+
+ // Now pack things back together.
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V1);
+ V2 = IsSingleInput ? V1 : DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V2);
+ SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
+ for (int i = 1; i < NumEvenDrops; ++i) {
+ Result = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, Result);
+ Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
+ }
+
+ return Result;
+ }
int V1LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
int V1HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
buildBlendMasks(LoMask, V1LoBlendMask, V2LoBlendMask);
buildBlendMasks(HiMask, V1HiBlendMask, V2HiBlendMask);
- SDValue V1Lo = DAG.getVectorShuffle(MVT::v8i16, DL, LoV1, HiV1, V1LoBlendMask);
- SDValue V2Lo = DAG.getVectorShuffle(MVT::v8i16, DL, LoV2, HiV2, V2LoBlendMask);
- SDValue V1Hi = DAG.getVectorShuffle(MVT::v8i16, DL, LoV1, HiV1, V1HiBlendMask);
- SDValue V2Hi = DAG.getVectorShuffle(MVT::v8i16, DL, LoV2, HiV2, V2HiBlendMask);
+ SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
+
+ auto buildLoAndHiV8s = [&](SDValue V, MutableArrayRef<int> LoBlendMask,
+ MutableArrayRef<int> HiBlendMask) {
+ SDValue V1, V2;
+ // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
+ // them out and avoid using UNPCK{L,H} to extract the elements of V as
+ // i16s.
+ if (std::none_of(LoBlendMask.begin(), LoBlendMask.end(),
+ [](int M) { return M >= 0 && M % 2 == 1; }) &&
+ std::none_of(HiBlendMask.begin(), HiBlendMask.end(),
+ [](int M) { return M >= 0 && M % 2 == 1; })) {
+ // Use a mask to drop the high bytes.
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
+ V1 = DAG.getNode(ISD::AND, DL, MVT::v8i16, V1,
+ DAG.getConstant(0x00FF, MVT::v8i16));
+
+ // This will be a single vector shuffle instead of a blend so nuke V2.
+ V2 = DAG.getUNDEF(MVT::v8i16);
+
+ // Squash the masks to point directly into V1.
+ for (int &M : LoBlendMask)
+ if (M >= 0)
+ M /= 2;
+ for (int &M : HiBlendMask)
+ if (M >= 0)
+ M /= 2;
+ } else {
+ // Otherwise just unpack the low half of V into V1 and the high half into
+ // V2 so that we can blend them as i16s.
+ V1 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+ DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
+ V2 = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16,
+ DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
+ }
+
+ SDValue BlendedLo = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, LoBlendMask);
+ SDValue BlendedHi = DAG.getVectorShuffle(MVT::v8i16, DL, V1, V2, HiBlendMask);
+ return std::make_pair(BlendedLo, BlendedHi);
+ };
+ SDValue V1Lo, V1Hi, V2Lo, V2Hi;
+ std::tie(V1Lo, V1Hi) = buildLoAndHiV8s(V1, V1LoBlendMask, V1HiBlendMask);
+ std::tie(V2Lo, V2Hi) = buildLoAndHiV8s(V2, V2LoBlendMask, V2HiBlendMask);
SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Lo, V2Lo, LoMask);
SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, V1Hi, V2Hi, HiMask);
// but in some cases the first operand may be transformed to UNDEF.
// In this case we should just commute the node.
if (V1IsUndef)
- return CommuteVectorShuffle(SVOp, DAG);
+ return DAG.getCommutedVectorShuffle(*SVOp);
// Check for non-undef masks pointing at an undef vector and make the masks
// undef as well. This makes it easier to match the shuffle based solely on
// V2. This allows us to match the shuffle pattern strictly on how many
// elements come from V1 without handling the symmetric cases.
if (NumV2Elements > NumV1Elements)
- return CommuteVectorShuffle(SVOp, DAG);
+ return DAG.getCommutedVectorShuffle(*SVOp);
// When the number of V1 and V2 elements are the same, try to minimize the
// number of uses of V2 in the low half of the vector.
else if (M >= 0)
++LowV1Elements;
if (LowV2Elements > LowV1Elements)
- return CommuteVectorShuffle(SVOp, DAG);
+ return DAG.getCommutedVectorShuffle(*SVOp);
}
// For each vector width, delegate to a specialized lowering routine.
To = V2;
DestIndex = std::find_if(Mask.begin(), Mask.end(), FromV1Predicate) -
Mask.begin();
+
+ // If we have 1 element from each vector, we have to check if we're
+ // changing V1's element's place. If so, we're done. Otherwise, we
+ // should assume we're changing V2's element's place and behave
+ // accordingly.
+ int FromV2 = std::count_if(Mask.begin(), Mask.end(), FromV2Predicate);
+ assert(DestIndex <= INT32_MAX && "truncated destination index");
+ if (FromV1 == FromV2 &&
+ static_cast<int>(DestIndex) == Mask[DestIndex] % 4) {
+ From = V2;
+ To = V1;
+ DestIndex =
+ std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
+ }
} else {
assert(std::count_if(Mask.begin(), Mask.end(), FromV2Predicate) == 1 &&
"More than one element from V1 and from V2, or no elements from one "
std::find_if(Mask.begin(), Mask.end(), FromV2Predicate) - Mask.begin();
}
+ // Get an index into the source vector in the range [0,4) (the mask is
+ // in the range [0,8) because it can address V1 and V2)
unsigned SrcIndex = Mask[DestIndex] % 4;
if (MayFoldLoad(From)) {
// Trivial case, when From comes from a load and is only used by the
// but in some cases the first operand may be transformed to UNDEF.
// In this case we should just commute the node.
if (V1IsUndef)
- return CommuteVectorShuffle(SVOp, DAG);
+ return DAG.getCommutedVectorShuffle(*SVOp);
// Vector shuffle lowering takes 3 steps:
//
if (ShouldXformToMOVHLPS(M, VT) ||
ShouldXformToMOVLP(V1.getNode(), V2.getNode(), M, VT))
- return CommuteVectorShuffle(SVOp, DAG);
+ return DAG.getCommutedVectorShuffle(*SVOp);
if (isShift) {
// No better options. Use a vshldq / vsrldq.
// Normalize the node to match x86 shuffle ops if needed
if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true)))
- return CommuteVectorShuffle(SVOp, DAG);
+ return DAG.getCommutedVectorShuffle(*SVOp);
// The checks below are all present in isShuffleMaskLegal, but they are
// inlined here right now to enable us to directly emit target specific
if (Subtarget->is64Bit())
IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, getPointerTy(), Chain,
IDX, MachinePointerInfo(), MVT::i32,
- false, false, 0);
+ false, false, false, 0);
else
IDX = DAG.getLoad(getPointerTy(), dl, Chain, IDX, MachinePointerInfo(),
false, false, false, 0);
// FIXME: Avoid the extend by constructing the right constant pool?
SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
FudgePtr, MachinePointerInfo::getConstantPool(),
- MVT::f32, false, false, 4);
+ MVT::f32, false, false, false, 4);
// Extend everything to 80 bits to force it to be done on x87.
SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
}
+// Lower vector extended loads using a shuffle. If SSSE3 is not available we
+// may emit an illegal shuffle but the expansion is still better than scalar
+// code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
+// we'll emit a shuffle and a arithmetic shift.
+// TODO: It is possible to support ZExt by zeroing the undef values during
+// the shuffle phase or after the shuffle.
+static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ MVT RegVT = Op.getSimpleValueType();
+ assert(RegVT.isVector() && "We only custom lower vector sext loads.");
+ assert(RegVT.isInteger() &&
+ "We only custom lower integer vector sext loads.");
+
+ // Nothing useful we can do without SSE2 shuffles.
+ assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
+
+ LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
+ SDLoc dl(Ld);
+ EVT MemVT = Ld->getMemoryVT();
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ unsigned RegSz = RegVT.getSizeInBits();
+
+ ISD::LoadExtType Ext = Ld->getExtensionType();
+
+ assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
+ && "Only anyext and sext are currently implemented.");
+ assert(MemVT != RegVT && "Cannot extend to the same type");
+ assert(MemVT.isVector() && "Must load a vector from memory");
+
+ unsigned NumElems = RegVT.getVectorNumElements();
+ unsigned MemSz = MemVT.getSizeInBits();
+ assert(RegSz > MemSz && "Register size must be greater than the mem size");
+
+ if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
+ // The only way in which we have a legal 256-bit vector result but not the
+ // integer 256-bit operations needed to directly lower a sextload is if we
+ // have AVX1 but not AVX2. In that case, we can always emit a sextload to
+ // a 128-bit vector and a normal sign_extend to 256-bits that should get
+ // correctly legalized. We do this late to allow the canonical form of
+ // sextload to persist throughout the rest of the DAG combiner -- it wants
+ // to fold together any extensions it can, and so will fuse a sign_extend
+ // of an sextload into an sextload targeting a wider value.
+ SDValue Load;
+ if (MemSz == 128) {
+ // Just switch this to a normal load.
+ assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
+ "it must be a legal 128-bit vector "
+ "type!");
+ Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
+ Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
+ Ld->isInvariant(), Ld->getAlignment());
+ } else {
+ assert(MemSz < 128 &&
+ "Can't extend a type wider than 128 bits to a 256 bit vector!");
+ // Do an sext load to a 128-bit vector type. We want to use the same
+ // number of elements, but elements half as wide. This will end up being
+ // recursively lowered by this routine, but will succeed as we definitely
+ // have all the necessary features if we're using AVX1.
+ EVT HalfEltVT =
+ EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
+ EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
+ Load =
+ DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
+ Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
+ Ld->isNonTemporal(), Ld->isInvariant(),
+ Ld->getAlignment());
+ }
+
+ // Replace chain users with the new chain.
+ assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
+
+ // Finally, do a normal sign-extend to the desired register.
+ return DAG.getSExtOrTrunc(Load, dl, RegVT);
+ }
+
+ // All sizes must be a power of two.
+ assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
+ "Non-power-of-two elements are not custom lowered!");
+
+ // Attempt to load the original value using scalar loads.
+ // Find the largest scalar type that divides the total loaded size.
+ MVT SclrLoadTy = MVT::i8;
+ for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
+ tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
+ MVT Tp = (MVT::SimpleValueType)tp;
+ if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
+ SclrLoadTy = Tp;
+ }
+ }
+
+ // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
+ if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
+ (64 <= MemSz))
+ SclrLoadTy = MVT::f64;
+
+ // Calculate the number of scalar loads that we need to perform
+ // in order to load our vector from memory.
+ unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
+
+ assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
+ "Can only lower sext loads with a single scalar load!");
+
+ unsigned loadRegZize = RegSz;
+ if (Ext == ISD::SEXTLOAD && RegSz == 256)
+ loadRegZize /= 2;
+
+ // Represent our vector as a sequence of elements which are the
+ // largest scalar that we can load.
+ EVT LoadUnitVecVT = EVT::getVectorVT(
+ *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
+
+ // Represent the data using the same element type that is stored in
+ // memory. In practice, we ''widen'' MemVT.
+ EVT WideVecVT =
+ EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
+ loadRegZize / MemVT.getScalarType().getSizeInBits());
+
+ assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
+ "Invalid vector type");
+
+ // We can't shuffle using an illegal type.
+ assert(TLI.isTypeLegal(WideVecVT) &&
+ "We only lower types that form legal widened vector types");
+
+ SmallVector<SDValue, 8> Chains;
+ SDValue Ptr = Ld->getBasePtr();
+ SDValue Increment =
+ DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, TLI.getPointerTy());
+ SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
+
+ for (unsigned i = 0; i < NumLoads; ++i) {
+ // Perform a single load.
+ SDValue ScalarLoad =
+ DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
+ Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
+ Ld->getAlignment());
+ Chains.push_back(ScalarLoad.getValue(1));
+ // Create the first element type using SCALAR_TO_VECTOR in order to avoid
+ // another round of DAGCombining.
+ if (i == 0)
+ Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
+ else
+ Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
+ ScalarLoad, DAG.getIntPtrConstant(i));
+
+ Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
+ }
+
+ SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
+
+ // Bitcast the loaded value to a vector of the original element type, in
+ // the size of the target vector type.
+ SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
+ unsigned SizeRatio = RegSz / MemSz;
+
+ if (Ext == ISD::SEXTLOAD) {
+ // If we have SSE4.1 we can directly emit a VSEXT node.
+ if (Subtarget->hasSSE41()) {
+ SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
+ return Sext;
+ }
+
+ // Otherwise we'll shuffle the small elements in the high bits of the
+ // larger type and perform an arithmetic shift. If the shift is not legal
+ // it's better to scalarize.
+ assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
+ "We can't implement an sext load without a arithmetic right shift!");
+
+ // Redistribute the loaded elements into the different locations.
+ SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
+
+ SDValue Shuff = DAG.getVectorShuffle(
+ WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
+
+ Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
+
+ // Build the arithmetic shift.
+ unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
+ MemVT.getVectorElementType().getSizeInBits();
+ Shuff =
+ DAG.getNode(ISD::SRA, dl, RegVT, Shuff, DAG.getConstant(Amt, RegVT));
+
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
+ return Shuff;
+ }
+
+ // Redistribute the loaded elements into the different locations.
+ SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i * SizeRatio] = i;
+
+ SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
+ DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
+
+ // Bitcast to the requested type.
+ Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
+ DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
+ return Shuff;
+}
+
// isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
// ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
// from the AND / OR.
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
Chain = SP.getValue(1);
unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
- const TargetFrameLowering &TFI = *DAG.getTarget().getFrameLowering();
+ const TargetFrameLowering &TFI = *DAG.getSubtarget().getFrameLowering();
unsigned StackAlign = TFI.getStackAlignment();
Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
if (Align > StackAlign)
Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
unsigned SPReg = RegInfo->getStackRegister();
SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
Chain = SP.getValue(1);
if (Depth > 0) {
SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), PtrVT);
return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
DAG.getNode(ISD::ADD, dl, PtrVT,
EVT VT = Op.getValueType();
SDLoc dl(Op); // FIXME probably not meaningful
unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
(FrameReg == X86::EBP && VT == MVT::i32)) &&
SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
SelectionDAG &DAG) const {
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize());
}
SDLoc dl (Op);
EVT PtrVT = getPointerTy();
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(DAG.getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ DAG.getSubtarget().getRegisterInfo());
unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
(FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
SDLoc dl (Op);
const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
- const TargetRegisterInfo* TRI = DAG.getTarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI = DAG.getSubtarget().getRegisterInfo();
if (Subtarget->is64Bit()) {
SDValue OutChains[6];
MachineFunction &MF = DAG.getMachineFunction();
const TargetMachine &TM = MF.getTarget();
- const TargetFrameLowering &TFI = *TM.getFrameLowering();
+ const TargetFrameLowering &TFI = *TM.getSubtargetImpl()->getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
MVT VT = Op.getSimpleValueType();
SDLoc DL(Op);
assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
(VT == MVT::v8i32 && Subtarget->hasInt256()));
- // Get the high parts.
- const int Mask[] = {1, 2, 3, 4, 5, 6, 7, 8};
- SDValue Hi0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
- SDValue Hi1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
+ // PMULxD operations multiply each even value (starting at 0) of LHS with
+ // the related value of RHS and produce a widen result.
+ // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
+ // => <2 x i64> <ae|cg>
+ //
+ // In other word, to have all the results, we need to perform two PMULxD:
+ // 1. one with the even values.
+ // 2. one with the odd values.
+ // To achieve #2, with need to place the odd values at an even position.
+ //
+ // Place the odd value at an even position (basically, shift all values 1
+ // step to the left):
+ const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
+ // <a|b|c|d> => <b|undef|d|undef>
+ SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
+ // <e|f|g|h> => <f|undef|h|undef>
+ SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
// Emit two multiplies, one for the lower 2 ints and one for the higher 2
// ints.
bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
unsigned Opcode =
(!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
+ // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
+ // => <2 x i64> <ae|cg>
SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
+ // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
+ // => <2 x i64> <bf|dh>
SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getNode(Opcode, dl, MulVT, Hi0, Hi1));
+ DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
// Shuffle it back into the right order.
- const int HighMask[] = {1, 5, 3, 7, 9, 13, 11, 15};
- SDValue Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
- const int LowMask[] = {0, 4, 2, 6, 8, 12, 10, 14};
- SDValue Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
+ SDValue Highs, Lows;
+ if (VT == MVT::v8i32) {
+ const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
+ Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
+ const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
+ Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
+ } else {
+ const int HighMask[] = {1, 5, 3, 7};
+ Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
+ const int LowMask[] = {1, 4, 2, 6};
+ Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
+ }
// If we have a signed multiply but no PMULDQ fix up the high parts of a
// unsigned multiply.
Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
}
- return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getValueType(), Highs, Lows);
+ // The first result of MUL_LOHI is actually the low value, followed by the
+ // high value.
+ SDValue Ops[] = {Lows, Highs};
+ return DAG.getMergeValues(Ops, dl);
}
static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &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::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
case ISD::FABS: return LowerFABS(Op, DAG);
case ISD::FNEG: return LowerFNEG(Op, DAG);
case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
return (SVT.getVectorNumElements() == 2 ||
ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
isMOVLMask(M, SVT) ||
+ isMOVHLPSMask(M, SVT) ||
isSHUFPMask(M, SVT) ||
isPSHUFDMask(M, SVT) ||
isPSHUFHWMask(M, SVT, Subtarget->hasInt256()) ||
MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
// Machine Information
- const TargetInstrInfo *TII = MBB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
XMMSaveMBB->addSuccessor(EndMBB);
// Now add the instructions.
- const TargetInstrInfo *TII = MBB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MBB->getParent()->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
unsigned CountReg = MI->getOperand(0).getReg();
MachineBasicBlock *
X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
MachineBasicBlock *BB) const {
- const TargetInstrInfo *TII = BB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
// To "insert" a SELECT_CC instruction, we actually have to insert the
// If the EFLAGS register isn't dead in the terminator, then claim that it's
// live into the sink and copy blocks.
- const TargetRegisterInfo* TRI = BB->getParent()->getTarget().getRegisterInfo();
+ const TargetRegisterInfo *TRI =
+ BB->getParent()->getSubtarget().getRegisterInfo();
if (!MI->killsRegister(X86::EFLAGS) &&
!checkAndUpdateEFLAGSKill(MI, BB, TRI)) {
copy0MBB->addLiveIn(X86::EFLAGS);
X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
bool Is64Bit) const {
MachineFunction *MF = BB->getParent();
- const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
const BasicBlock *LLVM_BB = BB->getBasicBlock();
BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
// Calls into a routine in libgcc to allocate more space from the heap.
- const uint32_t *RegMask =
- MF->getTarget().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
+ const uint32_t *RegMask = MF->getTarget()
+ .getSubtargetImpl()
+ ->getRegisterInfo()
+ ->getCallPreservedMask(CallingConv::C);
if (Is64Bit) {
BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
.addReg(sizeVReg);
MachineBasicBlock *
X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
MachineBasicBlock *BB) const {
- const TargetInstrInfo *TII = BB->getParent()->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = BB->getParent()->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
assert(!Subtarget->isTargetMacho());
// or EAX and doing an indirect call. The return value will then
// be in the normal return register.
MachineFunction *F = BB->getParent();
- const X86InstrInfo *TII
- = static_cast<const X86InstrInfo*>(F->getTarget().getInstrInfo());
+ const X86InstrInfo *TII =
+ static_cast<const X86InstrInfo *>(F->getSubtarget().getInstrInfo());
DebugLoc DL = MI->getDebugLoc();
assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
// Get a register mask for the lowered call.
// FIXME: The 32-bit calls have non-standard calling conventions. Use a
// proper register mask.
- const uint32_t *RegMask =
- F->getTarget().getRegisterInfo()->getCallPreservedMask(CallingConv::C);
+ const uint32_t *RegMask = F->getTarget()
+ .getSubtargetImpl()
+ ->getRegisterInfo()
+ ->getCallPreservedMask(CallingConv::C);
if (Subtarget->is64Bit()) {
MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
TII->get(X86::MOV64rm), X86::RDI)
MachineBasicBlock *MBB) const {
DebugLoc DL = MI->getDebugLoc();
MachineFunction *MF = MBB->getParent();
- const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
MachineRegisterInfo &MRI = MF->getRegInfo();
const BasicBlock *BB = MBB->getBasicBlock();
MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
.addMBB(restoreMBB);
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(MF->getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ MF->getSubtarget().getRegisterInfo());
MIB.addRegMask(RegInfo->getNoPreservedMask());
thisMBB->addSuccessor(mainMBB);
thisMBB->addSuccessor(restoreMBB);
MachineBasicBlock *MBB) const {
DebugLoc DL = MI->getDebugLoc();
MachineFunction *MF = MBB->getParent();
- const TargetInstrInfo *TII = MF->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
MachineRegisterInfo &MRI = MF->getRegInfo();
// Memory Reference
(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.
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(MF->getTarget().getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ MF->getSubtarget().getRegisterInfo());
unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
unsigned SP = RegInfo->getStackRegister();
default: llvm_unreachable("Unrecognized FMA variant.");
}
- const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
+ const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
MachineInstrBuilder MIB =
BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
.addOperand(MI->getOperand(0))
case X86::FP80_TO_INT32_IN_MEM:
case X86::FP80_TO_INT64_IN_MEM: {
MachineFunction *F = BB->getParent();
- const TargetInstrInfo *TII = F->getTarget().getInstrInfo();
+ const TargetInstrInfo *TII = F->getSubtarget().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
// Change the floating point control register to use "round towards zero"
case X86::VPCMPESTRM128MEM:
assert(Subtarget->hasSSE42() &&
"Target must have SSE4.2 or AVX features enabled");
- return EmitPCMPSTRM(MI, BB, BB->getParent()->getTarget().getInstrInfo());
+ return EmitPCMPSTRM(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
// String/text processing lowering.
case X86::PCMPISTRIREG:
case X86::VPCMPESTRIMEM:
assert(Subtarget->hasSSE42() &&
"Target must have SSE4.2 or AVX features enabled");
- return EmitPCMPSTRI(MI, BB, BB->getParent()->getTarget().getInstrInfo());
+ return EmitPCMPSTRI(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
// Thread synchronization.
case X86::MONITOR:
- return EmitMonitor(MI, BB, BB->getParent()->getTarget().getInstrInfo(), Subtarget);
+ return EmitMonitor(MI, BB, BB->getParent()->getSubtarget().getInstrInfo(),
+ Subtarget);
// xbegin
case X86::XBEGIN:
- return EmitXBegin(MI, BB, BB->getParent()->getTarget().getInstrInfo());
+ return EmitXBegin(MI, BB, BB->getParent()->getSubtarget().getInstrInfo());
case X86::VASTART_SAVE_XMM_REGS:
return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
return SDValue();
}
+/// \brief Combine an arbitrary chain of shuffles into a single instruction if
+/// possible.
+///
+/// This is the leaf of the recursive combinine below. When we have found some
+/// chain of single-use x86 shuffle instructions and accumulated the combined
+/// shuffle mask represented by them, this will try to pattern match that mask
+/// into either a single instruction if there is a special purpose instruction
+/// for this operation, or into a PSHUFB instruction which is a fully general
+/// instruction but should only be used to replace chains over a certain depth.
+static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
+ int Depth, bool HasPSHUFB, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
+
+ // Find the operand that enters the chain. Note that multiple uses are OK
+ // here, we're not going to remove the operand we find.
+ SDValue Input = Op.getOperand(0);
+ while (Input.getOpcode() == ISD::BITCAST)
+ Input = Input.getOperand(0);
+
+ MVT VT = Input.getSimpleValueType();
+ MVT RootVT = Root.getSimpleValueType();
+ SDLoc DL(Root);
+
+ // Just remove no-op shuffle masks.
+ if (Mask.size() == 1) {
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Input),
+ /*AddTo*/ true);
+ return true;
+ }
+
+ // Use the float domain if the operand type is a floating point type.
+ bool FloatDomain = VT.isFloatingPoint();
+
+ // If we don't have access to VEX encodings, the generic PSHUF instructions
+ // are preferable to some of the specialized forms despite requiring one more
+ // byte to encode because they can implicitly copy.
+ //
+ // IF we *do* have VEX encodings, than we can use shorter, more specific
+ // shuffle instructions freely as they can copy due to the extra register
+ // operand.
+ if (Subtarget->hasAVX()) {
+ // We have both floating point and integer variants of shuffles that dup
+ // either the low or high half of the vector.
+ if (Mask.equals(0, 0) || Mask.equals(1, 1)) {
+ bool Lo = Mask.equals(0, 0);
+ unsigned Shuffle = FloatDomain ? (Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS)
+ : (Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH);
+ if (Depth == 1 && Root->getOpcode() == Shuffle)
+ return false; // Nothing to do!
+ MVT ShuffleVT = FloatDomain ? MVT::v4f32 : MVT::v2i64;
+ Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
+ DCI.AddToWorklist(Op.getNode());
+ Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
+ DCI.AddToWorklist(Op.getNode());
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+ /*AddTo*/ true);
+ return true;
+ }
+
+ // FIXME: We should match UNPCKLPS and UNPCKHPS here.
+
+ // For the integer domain we have specialized instructions for duplicating
+ // any element size from the low or high half.
+ if (!FloatDomain &&
+ (Mask.equals(0, 0, 1, 1) || Mask.equals(2, 2, 3, 3) ||
+ Mask.equals(0, 0, 1, 1, 2, 2, 3, 3) ||
+ Mask.equals(4, 4, 5, 5, 6, 6, 7, 7) ||
+ Mask.equals(0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7) ||
+ Mask.equals(8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15,
+ 15))) {
+ bool Lo = Mask[0] == 0;
+ unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
+ if (Depth == 1 && Root->getOpcode() == Shuffle)
+ return false; // Nothing to do!
+ MVT ShuffleVT;
+ switch (Mask.size()) {
+ case 4: ShuffleVT = MVT::v4i32; break;
+ case 8: ShuffleVT = MVT::v8i16; break;
+ case 16: ShuffleVT = MVT::v16i8; break;
+ };
+ Op = DAG.getNode(ISD::BITCAST, DL, ShuffleVT, Input);
+ DCI.AddToWorklist(Op.getNode());
+ Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
+ DCI.AddToWorklist(Op.getNode());
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+ /*AddTo*/ true);
+ return true;
+ }
+ }
+
+ // Don't try to re-form single instruction chains under any circumstances now
+ // that we've done encoding canonicalization for them.
+ if (Depth < 2)
+ return false;
+
+ // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
+ // can replace them with a single PSHUFB instruction profitably. Intel's
+ // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
+ // in practice PSHUFB tends to be *very* fast so we're more aggressive.
+ if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
+ SmallVector<SDValue, 16> PSHUFBMask;
+ assert(Mask.size() <= 16 && "Can't shuffle elements smaller than bytes!");
+ int Ratio = 16 / Mask.size();
+ for (unsigned i = 0; i < 16; ++i) {
+ int M = Ratio * Mask[i / Ratio] + i % Ratio;
+ PSHUFBMask.push_back(DAG.getConstant(M, MVT::i8));
+ }
+ Op = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Input);
+ DCI.AddToWorklist(Op.getNode());
+ SDValue PSHUFBMaskOp =
+ DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, PSHUFBMask);
+ DCI.AddToWorklist(PSHUFBMaskOp.getNode());
+ Op = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, Op, PSHUFBMaskOp);
+ DCI.AddToWorklist(Op.getNode());
+ DCI.CombineTo(Root.getNode(), DAG.getNode(ISD::BITCAST, DL, RootVT, Op),
+ /*AddTo*/ true);
+ return true;
+ }
+
+ // Failed to find any combines.
+ return false;
+}
+
+/// \brief Fully generic combining of x86 shuffle instructions.
+///
+/// This should be the last combine run over the x86 shuffle instructions. Once
+/// they have been fully optimized, this will recursively consdier all chains
+/// of single-use shuffle instructions, build a generic model of the cumulative
+/// shuffle operation, and check for simpler instructions which implement this
+/// operation. We use this primarily for two purposes:
+///
+/// 1) Collapse generic shuffles to specialized single instructions when
+/// equivalent. In most cases, this is just an encoding size win, but
+/// sometimes we will collapse multiple generic shuffles into a single
+/// special-purpose shuffle.
+/// 2) Look for sequences of shuffle instructions with 3 or more total
+/// instructions, and replace them with the slightly more expensive SSSE3
+/// PSHUFB instruction if available. We do this as the last combining step
+/// to ensure we avoid using PSHUFB if we can implement the shuffle with
+/// a suitable short sequence of other instructions. The PHUFB will either
+/// use a register or have to read from memory and so is slightly (but only
+/// slightly) more expensive than the other shuffle instructions.
+///
+/// Because this is inherently a quadratic operation (for each shuffle in
+/// a chain, we recurse up the chain), the depth is limited to 8 instructions.
+/// This should never be an issue in practice as the shuffle lowering doesn't
+/// produce sequences of more than 8 instructions.
+///
+/// FIXME: We will currently miss some cases where the redundant shuffling
+/// would simplify under the threshold for PSHUFB formation because of
+/// combine-ordering. To fix this, we should do the redundant instruction
+/// combining in this recursive walk.
+static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
+ ArrayRef<int> IncomingMask, int Depth,
+ bool HasPSHUFB, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ // Bound the depth of our recursive combine because this is ultimately
+ // quadratic in nature.
+ if (Depth > 8)
+ return false;
+
+ // Directly rip through bitcasts to find the underlying operand.
+ while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
+ Op = Op.getOperand(0);
+
+ MVT VT = Op.getSimpleValueType();
+ if (!VT.isVector())
+ return false; // Bail if we hit a non-vector.
+ // FIXME: This routine should be taught about 256-bit shuffles, or a 256-bit
+ // version should be added.
+ if (VT.getSizeInBits() != 128)
+ return false;
+
+ assert(Root.getSimpleValueType().isVector() &&
+ "Shuffles operate on vector types!");
+ assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
+ "Can only combine shuffles of the same vector register size.");
+
+ if (!isTargetShuffle(Op.getOpcode()))
+ return false;
+ SmallVector<int, 16> OpMask;
+ bool IsUnary;
+ bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
+ // We only can combine unary shuffles which we can decode the mask for.
+ if (!HaveMask || !IsUnary)
+ return false;
+
+ assert(VT.getVectorNumElements() == OpMask.size() &&
+ "Different mask size from vector size!");
+
+ SmallVector<int, 16> Mask;
+ Mask.reserve(std::max(OpMask.size(), IncomingMask.size()));
+
+ // Merge this shuffle operation's mask into our accumulated mask. This is
+ // a bit tricky as the shuffle may have a different size from the root.
+ if (OpMask.size() == IncomingMask.size()) {
+ for (int M : IncomingMask)
+ Mask.push_back(OpMask[M]);
+ } else if (OpMask.size() < IncomingMask.size()) {
+ assert(IncomingMask.size() % OpMask.size() == 0 &&
+ "The smaller number of elements must divide the larger.");
+ int Ratio = IncomingMask.size() / OpMask.size();
+ for (int M : IncomingMask)
+ Mask.push_back(Ratio * OpMask[M / Ratio] + M % Ratio);
+ } else {
+ assert(OpMask.size() > IncomingMask.size() && "All other cases handled!");
+ assert(OpMask.size() % IncomingMask.size() == 0 &&
+ "The smaller number of elements must divide the larger.");
+ int Ratio = OpMask.size() / IncomingMask.size();
+ for (int i = 0, e = OpMask.size(); i < e; ++i)
+ Mask.push_back(OpMask[Ratio * IncomingMask[i / Ratio] + i % Ratio]);
+ }
+
+ // See if we can recurse into the operand to combine more things.
+ switch (Op.getOpcode()) {
+ case X86ISD::PSHUFB:
+ HasPSHUFB = true;
+ case X86ISD::PSHUFD:
+ case X86ISD::PSHUFHW:
+ case X86ISD::PSHUFLW:
+ if (Op.getOperand(0).hasOneUse() &&
+ combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
+ HasPSHUFB, DAG, DCI, Subtarget))
+ return true;
+ break;
+
+ case X86ISD::UNPCKL:
+ case X86ISD::UNPCKH:
+ assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
+ // We can't check for single use, we have to check that this shuffle is the only user.
+ if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
+ combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
+ HasPSHUFB, DAG, DCI, Subtarget))
+ return true;
+ break;
+ }
+
+ // Minor canonicalization of the accumulated shuffle mask to make it easier
+ // to match below. All this does is detect masks with squential pairs of
+ // elements, and shrink them to the half-width mask. It does this in a loop
+ // so it will reduce the size of the mask to the minimal width mask which
+ // performs an equivalent shuffle.
+ while (Mask.size() > 1) {
+ SmallVector<int, 16> NewMask;
+ for (int i = 0, e = Mask.size()/2; i < e; ++i) {
+ if (Mask[2*i] % 2 != 0 || Mask[2*i] != Mask[2*i + 1] + 1) {
+ NewMask.clear();
+ break;
+ }
+ NewMask.push_back(Mask[2*i] / 2);
+ }
+ if (NewMask.empty())
+ break;
+ Mask.swap(NewMask);
+ }
+
+ return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
+ Subtarget);
+}
+
/// \brief Get the PSHUF-style mask from PSHUF node.
///
/// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
return false;
continue;
+
+ case X86ISD::UNPCKL:
+ case X86ISD::UNPCKH:
+ // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
+ // shuffle into a preceding word shuffle.
+ if (V.getValueType() != MVT::v16i8 && V.getValueType() != MVT::v8i16)
+ return false;
+
+ // Search for a half-shuffle which we can combine with.
+ unsigned CombineOp =
+ V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
+ if (V.getOperand(0) != V.getOperand(1) ||
+ !V->isOnlyUserOf(V.getOperand(0).getNode()))
+ return false;
+ V = V.getOperand(0);
+ do {
+ switch (V.getOpcode()) {
+ default:
+ return false; // Nothing to combine.
+
+ case X86ISD::PSHUFLW:
+ case X86ISD::PSHUFHW:
+ if (V.getOpcode() == CombineOp)
+ break;
+
+ // Fallthrough!
+ case ISD::BITCAST:
+ V = V.getOperand(0);
+ continue;
+ }
+ break;
+ } while (V.hasOneUse());
+ break;
}
// Break out of the loop if we break out of the switch.
break;
SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
for (int &M : Mask)
M = VMask[M];
- V = DAG.getNode(X86ISD::PSHUFD, DL, V.getValueType(), V.getOperand(0),
+ V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
getV4X86ShuffleImm8ForMask(Mask, DAG));
// It is possible that one of the combinable shuffles was completely absorbed
return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, V);
}
+ // Look for shuffle patterns which can be implemented as a single unpack.
+ // FIXME: This doesn't handle the location of the PSHUFD generically, and
+ // only works when we have a PSHUFD followed by two half-shuffles.
+ if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
+ (V.getOpcode() == X86ISD::PSHUFLW ||
+ V.getOpcode() == X86ISD::PSHUFHW) &&
+ V.getOpcode() != N.getOpcode() &&
+ V.hasOneUse()) {
+ SDValue D = V.getOperand(0);
+ while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
+ D = D.getOperand(0);
+ if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
+ SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
+ SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
+ int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
+ int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
+ int WordMask[8];
+ for (int i = 0; i < 4; ++i) {
+ WordMask[i + NOffset] = Mask[i] + NOffset;
+ WordMask[i + VOffset] = VMask[i] + VOffset;
+ }
+ // Map the word mask through the DWord mask.
+ int MappedMask[8];
+ for (int i = 0; i < 8; ++i)
+ MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
+ const int UnpackLoMask[] = {0, 0, 1, 1, 2, 2, 3, 3};
+ const int UnpackHiMask[] = {4, 4, 5, 5, 6, 6, 7, 7};
+ if (std::equal(std::begin(MappedMask), std::end(MappedMask),
+ std::begin(UnpackLoMask)) ||
+ std::equal(std::begin(MappedMask), std::end(MappedMask),
+ std::begin(UnpackHiMask))) {
+ // We can replace all three shuffles with an unpack.
+ V = DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, D.getOperand(0));
+ DCI.AddToWorklist(V.getNode());
+ return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
+ : X86ISD::UNPCKH,
+ DL, MVT::v8i16, V, V);
+ }
+ }
+ }
+
break;
case X86ISD::PSHUFD:
SDValue N1 = N->getOperand(1);
EVT VT = N->getValueType(0);
- // Canonicalize shuffles that perform 'addsub' on packed float vectors
- // according to the rule:
- // (shuffle (FADD A, B), (FSUB A, B), Mask) ->
- // (shuffle (FSUB A, -B), (FADD A, -B), Mask)
- //
- // Where 'Mask' is:
- // <0,5,2,7> -- for v4f32 and v4f64 shuffles;
- // <0,3> -- for v2f64 shuffles;
- // <0,9,2,11,4,13,6,15> -- for v8f32 shuffles.
- //
- // This helps pattern-matching more SSE3/AVX ADDSUB instructions
- // during ISel stage.
- if (N->getOpcode() == ISD::VECTOR_SHUFFLE &&
- ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
- (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
- N0->getOpcode() == ISD::FADD && N1->getOpcode() == ISD::FSUB &&
- // Operands to the FADD and FSUB must be the same.
- ((N0->getOperand(0) == N1->getOperand(0) &&
- N0->getOperand(1) == N1->getOperand(1)) ||
- // FADD is commutable. See if by commuting the operands of the FADD
- // we would still be able to match the operands of the FSUB dag node.
- (N0->getOperand(1) == N1->getOperand(0) &&
- N0->getOperand(0) == N1->getOperand(1))) &&
- N0->getOperand(0)->getOpcode() != ISD::UNDEF &&
- N0->getOperand(1)->getOpcode() != ISD::UNDEF) {
-
- ShuffleVectorSDNode *SV = cast<ShuffleVectorSDNode>(N);
- unsigned NumElts = VT.getVectorNumElements();
- ArrayRef<int> Mask = SV->getMask();
- bool CanFold = true;
-
- for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i)
- CanFold = Mask[i] == (int)((i & 1) ? i + NumElts : i);
-
- if (CanFold) {
- SDValue Op0 = N1->getOperand(0);
- SDValue Op1 = DAG.getNode(ISD::FNEG, dl, VT, N1->getOperand(1));
- SDValue Sub = DAG.getNode(ISD::FSUB, dl, VT, Op0, Op1);
- SDValue Add = DAG.getNode(ISD::FADD, dl, VT, Op0, Op1);
- return DAG.getVectorShuffle(VT, dl, Sub, Add, Mask);
- }
- }
-
// Don't create instructions with illegal types after legalize types has run.
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
if (Shuffle.getNode())
return Shuffle;
+
+ // Try recursively combining arbitrary sequences of x86 shuffle
+ // instructions into higher-order shuffles. We do this after combining
+ // specific PSHUF instruction sequences into their minimal form so that we
+ // can evaluate how many specialized shuffle instructions are involved in
+ // a particular chain.
+ SmallVector<int, 1> NonceMask; // Just a placeholder.
+ NonceMask.push_back(0);
+ if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
+ /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
+ DCI, Subtarget))
+ return SDValue(); // This routine will use CombineTo to replace N.
}
return SDValue();
EVT MemVT = Ld->getMemoryVT();
SDLoc dl(Ld);
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- unsigned RegSz = RegVT.getSizeInBits();
// On Sandybridge unaligned 256bit loads are inefficient.
ISD::LoadExtType Ext = Ld->getExtensionType();
return DCI.CombineTo(N, NewVec, TF, true);
}
- // If this is a vector EXT Load then attempt to optimize it using a
- // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
- // expansion is still better than scalar code.
- // We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise we'll
- // emit a shuffle and a arithmetic shift.
- // 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() && Subtarget->hasSSE2() &&
- (Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)) {
- assert(MemVT != RegVT && "Cannot extend to the same type");
- assert(MemVT.isVector() && "Must load a vector from memory");
-
- unsigned NumElems = RegVT.getVectorNumElements();
- unsigned MemSz = MemVT.getSizeInBits();
- assert(RegSz > MemSz && "Register size must be greater than the mem size");
-
- if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256())
- return SDValue();
-
- // All sizes must be a power of two.
- if (!isPowerOf2_32(RegSz * MemSz * NumElems))
- return SDValue();
-
- // Attempt to load the original value using scalar loads.
- // Find the largest scalar type that divides the total loaded size.
- MVT SclrLoadTy = MVT::i8;
- for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
- tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
- MVT Tp = (MVT::SimpleValueType)tp;
- if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
- SclrLoadTy = Tp;
- }
- }
-
- // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
- if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
- (64 <= MemSz))
- SclrLoadTy = MVT::f64;
-
- // Calculate the number of scalar loads that we need to perform
- // in order to load our vector from memory.
- unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
- if (Ext == ISD::SEXTLOAD && NumLoads > 1)
- return SDValue();
-
- unsigned loadRegZize = RegSz;
- if (Ext == ISD::SEXTLOAD && RegSz == 256)
- loadRegZize /= 2;
-
- // Represent our vector as a sequence of elements which are the
- // largest scalar that we can load.
- EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
- loadRegZize/SclrLoadTy.getSizeInBits());
-
- // Represent the data using the same element type that is stored in
- // memory. In practice, we ''widen'' MemVT.
- EVT WideVecVT =
- EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
- loadRegZize/MemVT.getScalarType().getSizeInBits());
-
- assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
- "Invalid vector type");
-
- // We can't shuffle using an illegal type.
- if (!TLI.isTypeLegal(WideVecVT))
- return SDValue();
-
- SmallVector<SDValue, 8> Chains;
- SDValue Ptr = Ld->getBasePtr();
- SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits()/8,
- TLI.getPointerTy());
- SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
-
- for (unsigned i = 0; i < NumLoads; ++i) {
- // Perform a single load.
- SDValue ScalarLoad = DAG.getLoad(SclrLoadTy, dl, Ld->getChain(),
- Ptr, Ld->getPointerInfo(),
- Ld->isVolatile(), Ld->isNonTemporal(),
- Ld->isInvariant(), Ld->getAlignment());
- Chains.push_back(ScalarLoad.getValue(1));
- // Create the first element type using SCALAR_TO_VECTOR in order to avoid
- // another round of DAGCombining.
- if (i == 0)
- Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
- else
- Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
- ScalarLoad, DAG.getIntPtrConstant(i));
-
- Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
- }
-
- SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
-
- // Bitcast the loaded value to a vector of the original element type, in
- // the size of the target vector type.
- SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
- unsigned SizeRatio = RegSz/MemSz;
-
- if (Ext == ISD::SEXTLOAD) {
- // If we have SSE4.1 we can directly emit a VSEXT node.
- if (Subtarget->hasSSE41()) {
- SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
- return DCI.CombineTo(N, Sext, TF, true);
- }
-
- // Otherwise we'll shuffle the small elements in the high bits of the
- // larger type and perform an arithmetic shift. If the shift is not legal
- // it's better to scalarize.
- if (!TLI.isOperationLegalOrCustom(ISD::SRA, RegVT))
- return SDValue();
-
- // Redistribute the loaded elements into the different locations.
- SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
- for (unsigned i = 0; i != NumElems; ++i)
- ShuffleVec[i*SizeRatio + SizeRatio-1] = i;
-
- SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
- DAG.getUNDEF(WideVecVT),
- &ShuffleVec[0]);
-
- Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
-
- // Build the arithmetic shift.
- unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
- MemVT.getVectorElementType().getSizeInBits();
- Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
- DAG.getConstant(Amt, RegVT));
-
- return DCI.CombineTo(N, Shuff, TF, true);
- }
-
- // Redistribute the loaded elements into the different locations.
- SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
- for (unsigned i = 0; i != NumElems; ++i)
- ShuffleVec[i*SizeRatio] = i;
-
- SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
- DAG.getUNDEF(WideVecVT),
- &ShuffleVec[0]);
-
- // Bitcast to the requested type.
- Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
- // Replace the original load with the new sequence
- // and return the new chain.
- return DCI.CombineTo(N, Shuff, TF, true);
- }
-
return SDValue();
}
return SDValue();
}
+static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
+ SelectionDAG &DAG) {
+ // Take advantage of vector comparisons producing 0 or -1 in each lane to
+ // optimize away operation when it's from a constant.
+ //
+ // The general transformation is:
+ // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
+ // AND(VECTOR_CMP(x,y), constant2)
+ // constant2 = UNARYOP(constant)
+
+ // Early exit if this isn't a vector operation, the operand of the
+ // unary operation isn't a bitwise AND, or if the sizes of the operations
+ // aren't the same.
+ EVT VT = N->getValueType(0);
+ if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
+ N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
+ VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
+ return SDValue();
+
+ // Now check that the other operand of the AND is a constant. We could
+ // make the transformation for non-constant splats as well, but it's unclear
+ // that would be a benefit as it would not eliminate any operations, just
+ // perform one more step in scalar code before moving to the vector unit.
+ if (BuildVectorSDNode *BV =
+ dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
+ // Bail out if the vector isn't a constant.
+ if (!BV->isConstant())
+ return SDValue();
+
+ // Everything checks out. Build up the new and improved node.
+ SDLoc DL(N);
+ EVT IntVT = BV->getValueType(0);
+ // Create a new constant of the appropriate type for the transformed
+ // DAG.
+ SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
+ // The AND node needs bitcasts to/from an integer vector type around it.
+ SDValue MaskConst = DAG.getNode(ISD::BITCAST, DL, IntVT, SourceConst);
+ SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
+ N->getOperand(0)->getOperand(0), MaskConst);
+ SDValue Res = DAG.getNode(ISD::BITCAST, DL, VT, NewAnd);
+ return Res;
+ }
+
+ return SDValue();
+}
+
static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
const X86TargetLowering *XTLI) {
+ // First try to optimize away the conversion entirely when it's
+ // conditionally from a constant. Vectors only.
+ SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG);
+ if (Res != SDValue())
+ return Res;
+
+ // Now move on to more general possibilities.
SDValue Op0 = N->getOperand(0);
EVT InVT = Op0->getValueType(0);
case X86ISD::UNPCKL:
case X86ISD::MOVHLPS:
case X86ISD::MOVLHPS:
+ case X86ISD::PSHUFB:
case X86ISD::PSHUFD:
case X86ISD::PSHUFHW:
case X86ISD::PSHUFLW:
Constraint[5] == ')' &&
Constraint[6] == '}') {
- Res.first = X86::ST0+Constraint[4]-'0';
+ Res.first = X86::FP0+Constraint[4]-'0';
Res.second = &X86::RFP80RegClass;
return Res;
}
// GCC allows "st(0)" to be called just plain "st".
if (StringRef("{st}").equals_lower(Constraint)) {
- Res.first = X86::ST0;
+ Res.first = X86::FP0;
Res.second = &X86::RFP80RegClass;
return Res;
}