setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
if (Subtarget->is64Bit()) {
- setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
+ if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
+ // f32/f64 are legal, f80 is custom.
+ setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
+ else
+ setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
} else if (!Subtarget->useSoftFloat()) {
// We have an algorithm for SSE2->double, and we turn this into a
// 64-bit FILD followed by conditional FADD for other targets.
setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
// We have an algorithm for SSE2, and we turn this into a 64-bit
- // FILD for other targets.
+ // FILD or VCVTUSI2SS/SD for other targets.
setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
}
setOperationAction(ISD::SETCC , MVT::i64 , Custom);
}
setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
- setOperationAction(ISD::CATCHRET , MVT::Other, Custom);
// NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
// SjLj exception handling but a light-weight setjmp/longjmp replacement to
// support continuation, user-level threading, and etc.. As a result, no
setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
+ setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
+ setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
+ setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
+ // ISD::CTTZ v2i64 - scalarization is faster.
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
+ // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
+
// Custom lower build_vector, vector_shuffle, and extract_vector_elt.
for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
MVT VT = (MVT::SimpleValueType)i;
setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
+ setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
+ setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
+ setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
+ setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
+
if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
setOperationAction(ISD::FMA, MVT::v8f32, Legal);
setOperationAction(ISD::FMA, MVT::v4f64, Legal);
if (Subtarget->hasInt256())
setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
-
// Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
MVT VT = (MVT::SimpleValueType)i;
setOperationAction(ISD::FMA, MVT::v8f64, Legal);
setOperationAction(ISD::FMA, MVT::v16f32, Legal);
- // FIXME: [US]INT_TO_FP are not legal for f80.
- setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
- setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
- if (Subtarget->is64Bit()) {
- setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
- setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
- }
setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
if (Subtarget->hasDQI()) {
setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
+
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
}
if (Subtarget->hasVLX() && Subtarget->hasCDI()) {
setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
+
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
}
if (Subtarget->hasDQI()) {
setOperationAction(ISD::MUL, MVT::v2i64, Legal);
setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
+ setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
setOperationAction(ISD::AND, MVT::v8i32, Legal);
setOperationAction(ISD::OR, MVT::v8i32, Legal);
setOperationAction(ISD::UMULO, VT, Custom);
}
-
if (!Subtarget->is64Bit()) {
// These libcalls are not available in 32-bit.
setLibcallName(RTLIB::SHL_I128, nullptr);
#include "X86GenCallingConv.inc"
-bool
-X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
- MachineFunction &MF, bool isVarArg,
- const SmallVectorImpl<ISD::OutputArg> &Outs,
- LLVMContext &Context) const {
+bool X86TargetLowering::CanLowerReturn(
+ CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
+ const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
SmallVector<CCValAssign, 16> RVLocs;
CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
return CCInfo.CheckReturn(Outs, RetCC_X86);
return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
}
-SDValue
-X86TargetLowering::LowerFormalArguments(SDValue Chain,
- CallingConv::ID CallConv,
- bool isVarArg,
- const SmallVectorImpl<ISD::InputArg> &Ins,
- SDLoc dl,
- SelectionDAG &DAG,
- SmallVectorImpl<SDValue> &InVals)
- const {
+SDValue X86TargetLowering::LowerFormalArguments(
+ SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
+ const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
+ SmallVectorImpl<SDValue> &InVals) const {
MachineFunction &MF = DAG.getMachineFunction();
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
const Function *WinEHParent = nullptr;
if (MMI.hasWinEHFuncInfo(Fn))
WinEHParent = MMI.getWinEHParent(Fn);
- bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
// Figure out if XMM registers are in use.
if (!MemOps.empty())
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
- } else if (IsWin64 && IsWinEHOutlined) {
- // Get to the caller-allocated home save location. Add 8 to account
- // for the return address.
- int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
- FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
- /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
-
- MMI.getWinEHFuncInfo(Fn)
- .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
- FuncInfo->getRegSaveFrameIndex();
-
- // Store the second integer parameter (rdx) into rsp+16 relative to the
- // stack pointer at the entry of the function.
- SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
- getPointerTy(DAG.getDataLayout()));
- unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
- SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
- Chain = DAG.getStore(
- Val.getValue(1), dl, Val, RSFIN,
- MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
- FuncInfo->getRegSaveFrameIndex()),
- /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
}
if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
/// Check whether the call is eligible for tail call optimization. Targets
/// that want to do tail call optimization should implement this function.
-bool
-X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
- CallingConv::ID CalleeCC,
- bool isVarArg,
- bool isCalleeStructRet,
- bool isCallerStructRet,
- Type *RetTy,
- const SmallVectorImpl<ISD::OutputArg> &Outs,
- const SmallVectorImpl<SDValue> &OutVals,
- const SmallVectorImpl<ISD::InputArg> &Ins,
- SelectionDAG &DAG) const {
+bool X86TargetLowering::IsEligibleForTailCallOptimization(
+ SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
+ bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
+ const SmallVectorImpl<ISD::OutputArg> &Outs,
+ const SmallVectorImpl<SDValue> &OutVals,
+ const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
return false;
/// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
/// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
/// Then bitcast to their original type, ensuring they get CSE'd.
-static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
- SDLoc dl) {
+static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG, SDLoc dl) {
assert(VT.isVector() && "Expected a vector type");
SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
SDValue Vec;
- if (VT.is256BitVector()) {
- if (HasInt256) { // AVX2
+ if (VT.is512BitVector()) {
+ SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
+ Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
+ } else if (VT.is256BitVector()) {
+ if (Subtarget->hasInt256()) { // AVX2
SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
} else { // AVX
return Op;
if (!VT.is512BitVector())
- return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
+ return getOnesVector(VT, Subtarget, DAG, dl);
}
BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
return LowerAVXCONCAT_VECTORS(Op, DAG);
}
-
//===----------------------------------------------------------------------===//
// Vector shuffle lowering
//
return Zeroable;
}
+// X86 has dedicated unpack instructions that can handle specific blend
+// operations: UNPCKH and UNPCKL.
+static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
+ SDValue V1, SDValue V2,
+ SelectionDAG &DAG) {
+ int NumElts = VT.getVectorNumElements();
+ bool Unpckl = true;
+ bool Unpckh = true;
+ bool UnpcklSwapped = true;
+ bool UnpckhSwapped = true;
+ int NumEltsInLane = 128 / VT.getScalarSizeInBits();
+
+ for (int i = 0; i < NumElts; ++i) {
+ unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
+
+ int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
+ int HiPos = LoPos + NumEltsInLane / 2;
+ int LoPosSwapped = (LoPos + NumElts) % (NumElts * 2);
+ int HiPosSwapped = (HiPos + NumElts) % (NumElts * 2);
+
+ if (Mask[i] == -1)
+ continue;
+ if (Mask[i] != LoPos)
+ Unpckl = false;
+ if (Mask[i] != HiPos)
+ Unpckh = false;
+ if (Mask[i] != LoPosSwapped)
+ UnpcklSwapped = false;
+ if (Mask[i] != HiPosSwapped)
+ UnpckhSwapped = false;
+ if (!Unpckl && !Unpckh && !UnpcklSwapped && !UnpckhSwapped)
+ return SDValue();
+ }
+ if (Unpckl)
+ return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
+ if (Unpckh)
+ return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
+ if (UnpcklSwapped)
+ return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
+ if (UnpckhSwapped)
+ return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
+
+ llvm_unreachable("Unexpected result of UNPCK mask analysis");
+ return SDValue();
+}
+
/// \brief Try to emit a bitmask instruction for a shuffle.
///
/// This handles cases where we can model a blend exactly as a bitmask due to
///
/// Given a specific number of elements, element bit width, and extension
/// stride, produce either a zero or any extension based on the available
-/// features of the subtarget.
+/// features of the subtarget. The extended elements are consecutive and
+/// begin and can start from an offseted element index in the input; to
+/// avoid excess shuffling the offset must either being in the bottom lane
+/// or at the start of a higher lane. All extended elements must be from
+/// the same lane.
static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
- SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
+ SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
assert(Scale > 1 && "Need a scale to extend.");
- int NumElements = VT.getVectorNumElements();
int EltBits = VT.getScalarSizeInBits();
+ int NumElements = VT.getVectorNumElements();
+ int NumEltsPerLane = 128 / EltBits;
+ int OffsetLane = Offset / NumEltsPerLane;
assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
"Only 8, 16, and 32 bit elements can be extended.");
assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
+ assert(0 <= Offset && "Extension offset must be positive.");
+ assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
+ "Extension offset must be in the first lane or start an upper lane.");
+
+ // Check that an index is in same lane as the base offset.
+ auto SafeOffset = [&](int Idx) {
+ return OffsetLane == (Idx / NumEltsPerLane);
+ };
+
+ // Shift along an input so that the offset base moves to the first element.
+ auto ShuffleOffset = [&](SDValue V) {
+ if (!Offset)
+ return V;
+
+ SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
+ for (int i = 0; i * Scale < NumElements; ++i) {
+ int SrcIdx = i + Offset;
+ ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
+ }
+ return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
+ };
// Found a valid zext mask! Try various lowering strategies based on the
// input type and available ISA extensions.
if (Subtarget->hasSSE41()) {
+ // Not worth offseting 128-bit vectors if scale == 2, a pattern using
+ // PUNPCK will catch this in a later shuffle match.
+ if (Offset && Scale == 2 && VT.getSizeInBits() == 128)
+ return SDValue();
MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
NumElements / Scale);
- return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
+ InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
+ return DAG.getBitcast(VT, InputV);
}
+ assert(VT.getSizeInBits() == 128 && "Only 128-bit vectors can be extended.");
+
// For any extends we can cheat for larger element sizes and use shuffle
// instructions that can fold with a load and/or copy.
if (AnyExt && EltBits == 32) {
- int PSHUFDMask[4] = {0, -1, 1, -1};
+ int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
+ -1};
return DAG.getBitcast(
VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
DAG.getBitcast(MVT::v4i32, InputV),
getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
}
if (AnyExt && EltBits == 16 && Scale > 2) {
- int PSHUFDMask[4] = {0, -1, 0, -1};
+ int PSHUFDMask[4] = {Offset / 2, -1,
+ SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
DAG.getBitcast(MVT::v4i32, InputV),
getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
- int PSHUFHWMask[4] = {1, -1, -1, -1};
+ int PSHUFWMask[4] = {1, -1, -1, -1};
+ unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
return DAG.getBitcast(
- VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
+ VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
DAG.getBitcast(MVT::v8i16, InputV),
- getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
+ getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
}
// The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
+ int LoIdx = Offset * EltBits;
SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
DAG.getConstant(EltBits, DL, MVT::i8),
- DAG.getConstant(0, DL, MVT::i8)));
- if (isUndefInRange(Mask, NumElements/2, NumElements/2))
+ DAG.getConstant(LoIdx, DL, MVT::i8)));
+
+ if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
+ !SafeOffset(Offset + 1))
return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
- SDValue Hi =
- DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
- DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
- DAG.getConstant(EltBits, DL, MVT::i8),
- DAG.getConstant(EltBits, DL, MVT::i8)));
+ int HiIdx = (Offset + 1) * EltBits;
+ SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
+ DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
+ DAG.getConstant(EltBits, DL, MVT::i8),
+ DAG.getConstant(HiIdx, DL, MVT::i8)));
return DAG.getNode(ISD::BITCAST, DL, VT,
DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
}
if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
assert(NumElements == 16 && "Unexpected byte vector width!");
SDValue PSHUFBMask[16];
- for (int i = 0; i < 16; ++i)
- PSHUFBMask[i] =
- DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
+ for (int i = 0; i < 16; ++i) {
+ int Idx = Offset + (i / Scale);
+ PSHUFBMask[i] = DAG.getConstant(
+ (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
+ }
InputV = DAG.getBitcast(MVT::v16i8, InputV);
return DAG.getBitcast(VT,
DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
MVT::v16i8, PSHUFBMask)));
}
+ // If we are extending from an (odd)offset, shuffle them by 1 element.
+ if (Offset & 1) {
+ SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
+ for (int i = 1; i < NumElements; ++i)
+ ShMask[i - 1] = i;
+ InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
+ Offset--;
+ }
+
// Otherwise emit a sequence of unpacks.
do {
+ unsigned UnpackLoHi = X86ISD::UNPCKL;
+ if (Offset >= (NumElements / 2)) {
+ UnpackLoHi = X86ISD::UNPCKH;
+ Offset -= (NumElements / 2);
+ }
+
MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
: getZeroVector(InputVT, Subtarget, DAG, DL);
InputV = DAG.getBitcast(InputVT, InputV);
- InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
+ InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
Scale /= 2;
EltBits *= 2;
NumElements /= 2;
SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
int Bits = VT.getSizeInBits();
+ int NumLanes = Bits / 128;
int NumElements = VT.getVectorNumElements();
+ int NumEltsPerLane = NumElements / NumLanes;
assert(VT.getScalarSizeInBits() <= 32 &&
"Exceeds 32-bit integer zero extension limit");
assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
auto Lower = [&](int Scale) -> SDValue {
SDValue InputV;
bool AnyExt = true;
+ int Offset = 0;
+ int Matches = 0;
for (int i = 0; i < NumElements; ++i) {
- if (Mask[i] == -1)
+ int M = Mask[i];
+ if (M == -1)
continue; // Valid anywhere but doesn't tell us anything.
if (i % Scale != 0) {
// Each of the extended elements need to be zeroable.
// Each of the base elements needs to be consecutive indices into the
// same input vector.
- SDValue V = Mask[i] < NumElements ? V1 : V2;
- if (!InputV)
+ SDValue V = M < NumElements ? V1 : V2;
+ M = M % NumElements;
+ if (!InputV) {
InputV = V;
- else if (InputV != V)
+ Offset = M - (i / Scale);
+ } else if (InputV != V)
return SDValue(); // Flip-flopping inputs.
- if (Mask[i] % NumElements != i / Scale)
+ // Offset must start in the lowest 128-bit lane or at the start of an
+ // upper lane.
+ // FIXME: Is it ever worth allowing a negative base offset?
+ if (!((0 <= Offset && Offset < NumEltsPerLane) ||
+ (Offset % NumEltsPerLane) == 0))
+ return SDValue();
+
+ // If we are offsetting, all referenced entries must come from the same
+ // lane.
+ if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
+ return SDValue();
+
+ if ((M % NumElements) != (Offset + (i / Scale)))
return SDValue(); // Non-consecutive strided elements.
+ Matches++;
}
// If we fail to find an input, we have a zero-shuffle which should always
if (!InputV)
return SDValue();
+ // If we are offsetting, don't extend if we only match a single input, we
+ // can always do better by using a basic PSHUF or PUNPCK.
+ if (Offset != 0 && Matches < 2)
+ return SDValue();
+
return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
- DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
+ DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
};
// The widest scale possible for extending is to a 64-bit integer.
/// because for floating point vectors we have a generalized SHUFPS lowering
/// strategy that handles everything that doesn't *exactly* match an unpack,
/// making this clever lowering unnecessary.
-static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
- SDValue V2, ArrayRef<int> Mask,
- SelectionDAG &DAG) {
+static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
+ SDValue V1, SDValue V2,
+ ArrayRef<int> Mask,
+ SelectionDAG &DAG) {
assert(!VT.isFloatingPoint() &&
"This routine only supports integer vectors.");
assert(!isSingleInputShuffleMask(Mask) &&
Mask, DAG);
// Try to lower by permuting the inputs into an unpack instruction.
- if (SDValue Unpack =
- lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
+ if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
+ V2, Mask, DAG))
return Unpack;
// We implement this with SHUFPS because it can blend from two vectors.
lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
return BitBlend;
- if (SDValue Unpack =
- lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
+ if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
+ V2, Mask, DAG))
return Unpack;
// If we can't directly blend but can use PSHUFB, that will be better as it
// FIXME: It might be worth trying to detect if the unpack-feeding
// shuffles will both be pshufb, in which case we shouldn't bother with
// this.
- if (SDValue Unpack =
- lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
+ if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
+ DL, MVT::v16i8, V1, V2, Mask, DAG))
return Unpack;
}
DL, VT, V1, V2, Mask, Subtarget, DAG))
return Insertion;
- // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
- // check for those subtargets here and avoid much of the subtarget querying in
- // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
- // ability to manipulate a 256-bit vector with integer types. Since we'll use
- // floating point types there eventually, just immediately cast everything to
- // a float and operate entirely in that domain.
+ // There is a really nice hard cut-over between AVX1 and AVX2 that means we
+ // can check for those subtargets here and avoid much of the subtarget
+ // querying in the per-vector-type lowering routines. With AVX1 we have
+ // essentially *zero* ability to manipulate a 256-bit vector with integer
+ // types. Since we'll use floating point types there eventually, just
+ // immediately cast everything to a float and operate entirely in that domain.
if (VT.isInteger() && !Subtarget->hasAVX2()) {
int ElementBits = VT.getScalarSizeInBits();
if (ElementBits < 32)
return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
}
-// X86 has dedicated unpack instructions that can handle specific blend
-// operations: UNPCKH and UNPCKL.
-static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT,
- ArrayRef<int> Mask, SDValue V1,
- SDValue V2, SelectionDAG &DAG) {
- int NumElts = VT.getVectorNumElements();
- bool Unpckl = true;
- bool Unpckh = true;
- bool UnpcklSwapped = true;
- bool UnpckhSwapped = true;
- int NumEltsInLane = 128 / VT.getScalarSizeInBits();
-
- for (int i = 0; i < NumElts ; ++i) {
- unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
-
- int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
- int HiPos = LoPos + NumEltsInLane / 2;
- int LoPosSwapped = (LoPos + NumElts) % (NumElts * 2);
- int HiPosSwapped = (HiPos + NumElts) % (NumElts * 2);
-
- if (Mask[i] == -1)
- continue;
- if (Mask[i] != LoPos)
- Unpckl = false;
- if (Mask[i] != HiPos)
- Unpckh = false;
- if (Mask[i] != LoPosSwapped)
- UnpcklSwapped = false;
- if (Mask[i] != HiPosSwapped)
- UnpckhSwapped = false;
- if (!Unpckl && !Unpckh && !UnpcklSwapped && !UnpckhSwapped)
- return SDValue();
- }
- if (Unpckl)
- return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
- if (Unpckh)
- return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
- if (UnpcklSwapped)
- return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
- if (UnpckhSwapped)
- return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
-
- llvm_unreachable("Unexpected result of UNPCK mask analysis");
- return SDValue();
-}
-
/// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
const X86Subtarget *Subtarget,
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
- SDValue UnpckNode =
- lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG);
- if (UnpckNode)
- return UnpckNode;
+ if (SDValue Unpck =
+ lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
+ return Unpck;
return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
}
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
- SDValue UnpckNode =
- lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG);
- if (UnpckNode)
- return UnpckNode;
+ if (SDValue Unpck =
+ lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
+ return Unpck;
return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
}
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
- SDValue UnpckNode =
- lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG);
- if (UnpckNode)
- return UnpckNode;
+ if (SDValue Unpck =
+ lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
+ return Unpck;
return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
}
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
- SDValue UnpckNode =
- lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG);
- if (UnpckNode)
- return UnpckNode;
+ if (SDValue Unpck =
+ lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
+ return Unpck;
return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
}
return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
}
+// Lower vXi1 vector shuffles.
+// There is no a dedicated instruction on AVX-512 that shuffles the masks.
+// The only way to shuffle bits is to sign-extend the mask vector to SIMD
+// vector, shuffle and then truncate it back.
+static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
+ MVT VT, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDLoc DL(Op);
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ ArrayRef<int> Mask = SVOp->getMask();
+ assert(Subtarget->hasAVX512() &&
+ "Cannot lower 512-bit vectors w/o basic ISA!");
+ EVT ExtVT;
+ switch (VT.SimpleTy) {
+ default:
+ assert(false && "Expected a vector of i1 elements");
+ break;
+ case MVT::v2i1:
+ ExtVT = MVT::v2i64;
+ break;
+ case MVT::v4i1:
+ ExtVT = MVT::v4i32;
+ break;
+ case MVT::v8i1:
+ ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
+ break;
+ case MVT::v16i1:
+ ExtVT = MVT::v16i32;
+ break;
+ case MVT::v32i1:
+ ExtVT = MVT::v32i16;
+ break;
+ case MVT::v64i1:
+ ExtVT = MVT::v64i8;
+ break;
+ }
+
+ if (ISD::isBuildVectorAllZeros(V1.getNode()))
+ V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
+ else if (ISD::isBuildVectorAllOnes(V1.getNode()))
+ V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
+ else
+ V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
+
+ if (V2.isUndef())
+ V2 = DAG.getUNDEF(ExtVT);
+ else if (ISD::isBuildVectorAllZeros(V2.getNode()))
+ V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
+ else if (ISD::isBuildVectorAllOnes(V2.getNode()))
+ V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
+ else
+ V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
+ return DAG.getNode(ISD::TRUNCATE, DL, VT,
+ DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
+}
/// \brief Top-level lowering for x86 vector shuffles.
///
/// This handles decomposition, canonicalization, and lowering of all x86
MVT VT = Op.getSimpleValueType();
int NumElements = VT.getVectorNumElements();
SDLoc dl(Op);
+ bool Is1BitVector = (VT.getScalarType() == MVT::i1);
- assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
+ assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
+ "Can't lower MMX shuffles");
bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
// elements wider than 64 bits, but it might be interesting to form i128
// integers to handle flipping the low and high halves of AVX 256-bit vectors.
SmallVector<int, 16> WidenedMask;
- if (VT.getScalarSizeInBits() < 64 &&
+ if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
canWidenShuffleElements(Mask, WidenedMask)) {
MVT NewEltVT = VT.isFloatingPoint()
? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
if (VT.getSizeInBits() == 512)
return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
+ if (Is1BitVector)
+ return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
llvm_unreachable("Unimplemented!");
}
// If needed, look through a bitcast to get to the load.
if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
SubVec2 = SubVec2.getOperand(0);
-
+
if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
bool Fast;
unsigned Alignment = FirstLd->getAlignment();
MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
false, false, false, 16);
SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
+ // TODO: Are there any fast-math-flags to propagate here?
SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
SDValue Result;
DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
// Subtract the bias.
+ // TODO: Are there any fast-math-flags to propagate here?
SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
// Handle final rounding.
// float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
+ // TODO: Are there any fast-math-flags to propagate here?
SDValue FHigh =
DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
// return (float4) lo + fhi;
MVT SrcVT = N0.getSimpleValueType();
MVT DstVT = Op.getSimpleValueType();
+
+ if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
+ (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
+ // Conversions from unsigned i32 to f32/f64 are legal,
+ // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
+ return Op;
+ }
+
if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
return LowerUINT_TO_FP_i64(Op, DAG);
if (SrcVT == MVT::i32 && X86ScalarSSEf64)
MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
false, false, false, 4);
// Extend everything to 80 bits to force it to be done on x87.
+ // TODO: Are there any fast-math-flags to propagate here?
SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
DAG.getIntPtrConstant(0, dl));
case X86ISD::CMPM:
case X86ISD::CMPMU:
return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
+ case X86ISD::VFPCLASS:
+ return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
case X86ISD::VTRUNC:
case X86ISD::VTRUNCS:
case X86ISD::VTRUNCUS:
SDValue PreservedSrc,
const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
- if (isAllOnes(Mask))
- return Op;
+ if (isAllOnes(Mask))
+ return Op;
- EVT VT = Op.getValueType();
- SDLoc dl(Op);
- // The mask should be of type MVT::i1
- SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
+ EVT VT = Op.getValueType();
+ SDLoc dl(Op);
+ // The mask should be of type MVT::i1
+ SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
- if (PreservedSrc.getOpcode() == ISD::UNDEF)
- PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
- return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
+ if (Op.getOpcode() == X86ISD::FSETCC)
+ return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
+
+ if (PreservedSrc.getOpcode() == ISD::UNDEF)
+ PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
+ return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
}
static int getSEHRegistrationNodeSize(const Function *Fn) {
return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
Mask, PassThru, Subtarget, DAG);
}
+ case INTR_TYPE_SCALAR_MASK: {
+ SDValue Src1 = Op.getOperand(1);
+ SDValue Src2 = Op.getOperand(2);
+ SDValue passThru = Op.getOperand(3);
+ SDValue Mask = Op.getOperand(4);
+ return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
+ Mask, passThru, Subtarget, DAG);
+ }
case INTR_TYPE_SCALAR_MASK_RM: {
SDValue Src1 = Op.getOperand(1);
SDValue Src2 = Op.getOperand(2);
Mask, PassThru, Subtarget, DAG);
}
}
- return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
- Src1,Src2),
+ // TODO: Intrinsics should have fast-math-flags to propagate.
+ return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
Mask, PassThru, Subtarget, DAG);
}
case INTR_TYPE_2OP_MASK_RM: {
SDValue Src2 = Op.getOperand(2);
SDValue PassThru = Op.getOperand(3);
SDValue Mask = Op.getOperand(4);
- // We specify 2 possible modes for intrinsics, with/without rounding modes.
+ // We specify 2 possible modes for intrinsics, with/without rounding
+ // modes.
// First, we check if the intrinsic have rounding mode (6 operands),
// if not, we set rounding mode to "current".
SDValue Rnd;
SDValue Imm = Op.getOperand(3);
SDValue PassThru = Op.getOperand(4);
SDValue Mask = Op.getOperand(5);
- // We specify 2 possible modes for intrinsics, with/without rounding modes.
+ // We specify 2 possible modes for intrinsics, with/without rounding
+ // modes.
// First, we check if the intrinsic have rounding mode (7 operands),
// if not, we set rounding mode to "current".
SDValue Rnd;
Mask, PassThru, Subtarget, DAG);
}
case INTR_TYPE_3OP_IMM8_MASK:
- case INTR_TYPE_3OP_MASK: {
+ case INTR_TYPE_3OP_MASK:
+ case INSERT_SUBVEC: {
SDValue Src1 = Op.getOperand(1);
SDValue Src2 = Op.getOperand(2);
SDValue Src3 = Op.getOperand(3);
if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
+ else if (IntrData->Type == INSERT_SUBVEC) {
+ // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
+ assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
+ unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
+ Imm *= Src2.getValueType().getVectorNumElements();
+ Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
+ }
+
// We specify 2 possible opcodes for intrinsics with rounding modes.
// First, we check if the intrinsic may have non-default rounding mode,
// (IntrData->Opc1 != 0), then we check the rounding mode operand.
Src1, Src2, Src3),
Mask, PassThru, Subtarget, DAG);
}
+ case FPCLASS: {
+ // FPclass intrinsics with mask
+ SDValue Src1 = Op.getOperand(1);
+ EVT VT = Src1.getValueType();
+ EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
+ VT.getVectorNumElements());
+ SDValue Imm = Op.getOperand(2);
+ SDValue Mask = Op.getOperand(3);
+ EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
+ Mask.getValueType().getSizeInBits());
+ SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
+ SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
+ DAG.getTargetConstant(0, dl, MaskVT),
+ Subtarget, DAG);
+ SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
+ DAG.getUNDEF(BitcastVT), FPclassMask,
+ DAG.getIntPtrConstant(0, dl));
+ return DAG.getBitcast(Op.getValueType(), Res);
+ }
case CMP_MASK:
case CMP_MASK_CC: {
// Comparison intrinsics with masks.
DAG.getIntPtrConstant(0, dl));
return DAG.getBitcast(Op.getValueType(), Res);
}
+ case CMP_MASK_SCALAR_CC: {
+ SDValue Src1 = Op.getOperand(1);
+ SDValue Src2 = Op.getOperand(2);
+ SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
+ SDValue Mask = Op.getOperand(4);
+
+ SDValue Cmp;
+ if (IntrData->Opc1 != 0) {
+ SDValue Rnd = Op.getOperand(5);
+ if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
+ X86::STATIC_ROUNDING::CUR_DIRECTION)
+ Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
+ }
+ //default rounding mode
+ if(!Cmp.getNode())
+ Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
+
+ SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
+ DAG.getTargetConstant(0, dl,
+ MVT::i1),
+ Subtarget, DAG);
+
+ return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
+ DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
+ DAG.getValueType(MVT::i1));
+ }
case COMI: { // Comparison intrinsics
ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
SDValue LHS = Op.getOperand(1);
DAG.getRegister(StoreAddrReg, PtrVT));
}
-SDValue X86TargetLowering::LowerCATCHRET(SDValue Op, SelectionDAG &DAG) const {
- SDValue Chain = Op.getOperand(0);
- SDValue Dest = Op.getOperand(1);
- SDLoc DL(Op);
-
- MVT PtrVT = getPointerTy(DAG.getDataLayout());
-
- MachineFunction &MF = DAG.getMachineFunction();
- if (isAsynchronousEHPersonality(
- classifyEHPersonality(MF.getFunction()->getPersonalityFn()))) {
- // For SEH, codegen catchret as a branch for now.
- // FIXME: Insert something to restore the frame.
- return DAG.getNode(ISD::BR, DL, MVT::Other, Chain, Dest);
- }
-
-
- unsigned ReturnReg = (PtrVT == MVT::i64 ? X86::RAX : X86::EAX);
-
- // Load the address of the destination block.
- MachineBasicBlock *DestMBB = cast<BasicBlockSDNode>(Dest)->getBasicBlock();
- SDValue BlockPtr = DAG.getMCSymbol(DestMBB->getSymbol(), PtrVT);
- unsigned WrapperKind =
- Subtarget->isPICStyleRIPRel() ? X86ISD::WrapperRIP : X86ISD::Wrapper;
- SDValue WrappedPtr = DAG.getNode(WrapperKind, DL, PtrVT, BlockPtr);
- Chain = DAG.getCopyToReg(Chain, DL, ReturnReg, WrappedPtr);
- return DAG.getNode(X86ISD::CATCHRET, DL, MVT::Other, Chain,
- DAG.getRegister(ReturnReg, PtrVT));
-}
-
SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
SelectionDAG &DAG) const {
SDLoc DL(Op);
static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
- unsigned NumBits = VT.getSizeInBits();
+ unsigned NumBits = VT.getScalarSizeInBits();
SDLoc dl(Op);
- Op = Op.getOperand(0);
+
+ if (VT.isVector()) {
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+
+ SDValue N0 = Op.getOperand(0);
+ SDValue Zero = DAG.getConstant(0, dl, VT);
+
+ // lsb(x) = (x & -x)
+ SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
+ DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
+
+ // cttz_undef(x) = (width - 1) - ctlz(lsb)
+ if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
+ TLI.isOperationLegal(ISD::CTLZ, VT)) {
+ SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
+ return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
+ DAG.getNode(ISD::CTLZ, dl, VT, LSB));
+ }
+
+ // cttz(x) = ctpop(lsb - 1)
+ SDValue One = DAG.getConstant(1, dl, VT);
+ return DAG.getNode(ISD::CTPOP, dl, VT,
+ DAG.getNode(ISD::SUB, dl, VT, LSB, One));
+ }
+
+ assert(Op.getOpcode() == ISD::CTTZ &&
+ "Only scalar CTTZ requires custom lowering");
// Issue a bsf (scan bits forward) which also sets EFLAGS.
SDVTList VTs = DAG.getVTList(VT, MVT::i32);
- Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
+ Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
// If src is zero (i.e. bsf sets ZF), returns NumBits.
SDValue Ops[] = {
// Note: this turns large loads into lock cmpxchg8b/16b.
// FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
-bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
+TargetLowering::AtomicExpansionKind
+X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
- return needsCmpXchgNb(PTy->getElementType());
+ return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
+ : AtomicExpansionKind::None;
}
-TargetLoweringBase::AtomicRMWExpansionKind
+TargetLowering::AtomicExpansionKind
X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
Type *MemType = AI->getType();
// If the operand is too big, we must see if cmpxchg8/16b is available
// and default to library calls otherwise.
if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
- return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
- : AtomicRMWExpansionKind::None;
+ return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
+ : AtomicExpansionKind::None;
}
AtomicRMWInst::BinOp Op = AI->getOperation();
case AtomicRMWInst::Add:
case AtomicRMWInst::Sub:
// It's better to use xadd, xsub or xchg for these in all cases.
- return AtomicRMWExpansionKind::None;
+ return AtomicExpansionKind::None;
case AtomicRMWInst::Or:
case AtomicRMWInst::And:
case AtomicRMWInst::Xor:
// If the atomicrmw's result isn't actually used, we can just add a "lock"
// prefix to a normal instruction for these operations.
- return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
- : AtomicRMWExpansionKind::None;
+ return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
+ : AtomicExpansionKind::None;
case AtomicRMWInst::Nand:
case AtomicRMWInst::Max:
case AtomicRMWInst::Min:
case AtomicRMWInst::UMin:
// These always require a non-trivial set of data operations on x86. We must
// use a cmpxchg loop.
- return AtomicRMWExpansionKind::CmpXChg;
+ return AtomicExpansionKind::CmpXChg;
}
}
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::CATCHRET: return LowerCATCHRET(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::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
case ISD::CTLZ: return LowerCTLZ(Op, DAG);
case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
- case ISD::CTTZ: return LowerCTTZ(Op, DAG);
+ case ISD::CTTZ:
+ case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
case ISD::UMUL_LOHI:
case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
DAG.getBitcast(MVT::v2i64, VBias));
Or = DAG.getBitcast(MVT::v2f64, Or);
+ // TODO: Are there any fast-math-flags to propagate here?
SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
return;
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::CATCHRET: return "X86ISD::CATCHRET";
case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
+ case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
}
return nullptr;
}
// 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::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;
+ 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
return V;
}
-/// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
+/// \brief Search for a combinable shuffle across a chain ending in pshuflw or
+/// pshufhw.
///
/// We walk up the chain, skipping shuffles of the other half and looking
/// through shuffles which switch halves trying to find a shuffle of the same
InputVector.getOpcode() == ISD::BITCAST &&
dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
uint64_t ExtractedElt =
- cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
+ cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
uint64_t InputValue =
- cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
+ cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
uint64_t Res = (InputValue >> ExtractedElt) & 1;
return DAG.getConstant(Res, dl, MVT::i1);
}
}
} // BEXTR
+ // If both input operands are being cast from floating point types,
+ // try to convert this into a floating point logic node to avoid
+ // unnecessary moves from SSE to integer registers.
+ // FIXME: Split this into a helper function, so it can also be used with
+ // or/xor combining.
+ if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
+ ((Subtarget->hasSSE1() && VT == MVT::i32) ||
+ (Subtarget->hasSSE2() && VT == MVT::i64))) {
+ SDValue N00 = N0.getOperand(0);
+ SDValue N10 = N1.getOperand(0);
+ EVT N00Type = N00.getValueType();
+ EVT N10Type = N10.getValueType();
+ if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
+ SDValue FLogic = DAG.getNode(X86ISD::FAND, DL, N00Type, N00, N10);
+ return DAG.getBitcast(VT, FLogic);
+ }
+ }
+
return SDValue();
}
ShuffleVec[i] = i * SizeRatio;
// Can't shuffle using an illegal type.
- assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
- && "WideVecVT should be legal");
+ assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
+ "WideVecVT should be legal");
WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
}
ISD::NON_EXTLOAD);
SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
-
}
/// PerformMSTORECombine - Resolve truncating stores
static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
ShuffleVec[i] = i * SizeRatio;
// Can't shuffle using an illegal type.
- assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
- && "WideVecVT should be legal");
+ assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
+ "WideVecVT should be legal");
SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
DAG.getUNDEF(WideVecVT),
}
/// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
-static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
// F[X]OR(0.0, x) -> x
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
if (C->getValueAPF().isPosZero())
return N->getOperand(0);
+
+ EVT VT = N->getValueType(0);
+ if (VT.is512BitVector() && !Subtarget->hasDQI()) {
+ SDLoc dl(N);
+ MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
+ MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
+
+ SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
+ SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
+ unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
+ SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
+ return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
+ }
return SDValue();
}
case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
case X86ISD::FXOR:
- case X86ISD::FOR: return PerformFORCombine(N, DAG);
+ case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
case X86ISD::FMIN:
case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
case X86ISD::FAND: return PerformFANDCombine(N, DAG);