setOperationAction(ISD::FREM , MVT::f80 , Expand);
setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Expand);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i16 , Expand);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i32 , Expand);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i64 , Expand);
if (Subtarget->hasBMI()) {
setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
} else {
setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
}
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i8 , Expand);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i16 , Expand);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i32 , Expand);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF , MVT::i64 , Expand);
if (Subtarget->hasLZCNT()) {
setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
} else {
setOperationAction(ISD::FCOS , MVT::f80 , Expand);
}
+ setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
+ setOperationAction(ISD::FCEIL, MVT::f80, Expand);
+ setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
+ setOperationAction(ISD::FRINT, MVT::f80, Expand);
+ setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
setOperationAction(ISD::FMA, MVT::f80, Expand);
}
setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::CTTZ_ZERO_UNDEF, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::CTLZ_ZERO_UNDEF, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
- setPrefLoopAlignment(16);
+ setPrefLoopAlignment(4); // 2^4 bytes.
benefitFromCodePlacementOpt = true;
- setPrefFunctionAlignment(4);
+ setPrefFunctionAlignment(4); // 2^4 bytes.
}
case X86ISD::MOVDDUP:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
- case X86ISD::UNPCKLP:
- case X86ISD::PUNPCKL:
- case X86ISD::UNPCKHP:
- case X86ISD::PUNPCKH:
+ case X86ISD::UNPCKL:
+ case X86ISD::UNPCKH:
case X86ISD::VPERMILP:
case X86ISD::VPERM2X128:
return true;
case X86ISD::MOVLPD:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
- case X86ISD::UNPCKLP:
- case X86ISD::PUNPCKL:
- case X86ISD::UNPCKHP:
- case X86ISD::PUNPCKH:
+ case X86ISD::UNPCKL:
+ case X86ISD::UNPCKH:
return DAG.getNode(Opc, dl, VT, V1, V2);
}
return SDValue();
static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool hasSSSE3OrAVX) {
int i, e = VT.getVectorNumElements();
- if (VT.getSizeInBits() != 128 && VT.getSizeInBits() != 64)
+ if (VT.getSizeInBits() != 128)
return false;
// Do not handle v2i64 / v2f64 shuffles with palignr.
/// specifies a shuffle of elements that is suitable for input to 256-bit
/// VSHUFPSY.
static bool isVSHUFPYMask(const SmallVectorImpl<int> &Mask, EVT VT,
- bool HasAVX) {
- int NumElems = VT.getVectorNumElements();
-
- if (!HasAVX || VT.getSizeInBits() != 256)
- return false;
-
- if (NumElems != 4 && NumElems != 8)
- return false;
-
- // VSHUFPSY divides the resulting vector into 4 chunks.
- // The sources are also splitted into 4 chunks, and each destination
- // chunk must come from a different source chunk.
- //
- // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
- // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
- //
- // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
- // Y3..Y0, Y3..Y0, X3..X0, X3..X0
- //
- // VSHUFPDY divides the resulting vector into 4 chunks.
- // The sources are also splitted into 4 chunks, and each destination
- // chunk must come from a different source chunk.
- //
- // SRC1 => X3 X2 X1 X0
- // SRC2 => Y3 Y2 Y1 Y0
- //
- // DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
- //
- int QuarterSize = NumElems/4;
- int HalfSize = QuarterSize*2;
- for (int i = 0; i < QuarterSize; ++i)
- if (!isUndefOrInRange(Mask[i], 0, HalfSize))
- return false;
- for (int i = QuarterSize; i < QuarterSize*2; ++i)
- if (!isUndefOrInRange(Mask[i], NumElems, NumElems+HalfSize))
- return false;
-
- // For VSHUFPSY, the mask of the second half must be the same as the first
- // but with the appropriate offsets. This works in the same way as
- // VPERMILPS works with masks.
- for (int i = QuarterSize*2; i < QuarterSize*3; ++i) {
- if (!isUndefOrInRange(Mask[i], HalfSize, NumElems))
- return false;
- if (NumElems == 4)
- continue;
- // VSHUFPSY handling
- int FstHalfIdx = i-HalfSize;
- if (Mask[FstHalfIdx] < 0)
- continue;
- if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
- return false;
- }
- for (int i = QuarterSize*3; i < NumElems; ++i) {
- if (!isUndefOrInRange(Mask[i], NumElems+HalfSize, NumElems*2))
- return false;
- int FstHalfIdx = i-HalfSize;
- if (NumElems == 4)
- continue;
- // VSHUFPSY handling
- if (Mask[FstHalfIdx] < 0)
- continue;
- if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
- return false;
- }
-
- return true;
-}
-
-/// isCommutedVSHUFP() - Returns true if the shuffle mask is exactly
-/// the reverse of what x86 shuffles want. x86 shuffles requires the lower
-/// half elements to come from vector 1 (which would equal the dest.) and
-/// the upper half to come from vector 2.
-static bool isCommutedVSHUFPY(ShuffleVectorSDNode *N, bool HasAVX) {
- EVT VT = N->getValueType(0);
+ bool HasAVX, bool Commuted = false) {
int NumElems = VT.getVectorNumElements();
- SmallVector<int, 8> Mask;
- N->getMask(Mask);
if (!HasAVX || VT.getSizeInBits() != 256)
return false;
//
// DST => Y3..Y2, X3..X2, Y1..Y0, X1..X0
//
- int QuarterSize = NumElems/4;
- int HalfSize = QuarterSize*2;
- for (int i = 0; i < QuarterSize; ++i)
- if (!isUndefOrInRange(Mask[i], NumElems, NumElems+HalfSize))
- return false;
- for (int i = QuarterSize; i < QuarterSize*2; ++i)
- if (!isUndefOrInRange(Mask[i], 0, HalfSize))
- return false;
-
- // For VSHUFPSY, the mask of the second half must be the same as the first
- // but with the appropriate offsets. This works in the same way as
- // VPERMILPS works with masks.
- for (int i = QuarterSize*2; i < QuarterSize*3; ++i) {
- if (!isUndefOrInRange(Mask[i], NumElems+HalfSize, NumElems*2))
- return false;
- if (NumElems == 4)
- continue;
- // VSHUFPSY handling
- int FstHalfIdx = i-HalfSize;
- if (Mask[FstHalfIdx] < 0)
- continue;
- if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
- return false;
- }
- for (int i = QuarterSize*3; i < NumElems; ++i) {
- if (!isUndefOrInRange(Mask[i], HalfSize, NumElems))
- return false;
- if (NumElems == 4)
- continue;
- // VSHUFPSY handling
- int FstHalfIdx = i-HalfSize;
- if (Mask[FstHalfIdx] < 0)
- continue;
- if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
- return false;
+ unsigned QuarterSize = NumElems/4;
+ unsigned HalfSize = QuarterSize*2;
+ for (unsigned l = 0; l != 2; ++l) {
+ unsigned LaneStart = l*HalfSize;
+ for (unsigned s = 0; s != 2; ++s) {
+ unsigned QuarterStart = s*QuarterSize;
+ unsigned Src = (Commuted) ? (1-s) : s;
+ unsigned SrcStart = Src*NumElems + LaneStart;
+ for (unsigned i = 0; i != QuarterSize; ++i) {
+ int Idx = Mask[i+QuarterStart+LaneStart];
+ if (!isUndefOrInRange(Idx, SrcStart, SrcStart+HalfSize))
+ return false;
+ // For VSHUFPSY, the mask of the second half must be the same as the first
+ // but with the appropriate offsets. This works in the same way as
+ // VPERMILPS works with masks.
+ if (NumElems == 4 || l == 0 || Mask[i+QuarterStart] < 0)
+ continue;
+ if (!isUndefOrEqual(Idx, Mask[i+QuarterStart]+HalfSize))
+ return false;
+ }
+ }
}
return true;
/// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
/// the two vector operands have swapped position.
-static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
- unsigned NumElems = VT.getVectorNumElements();
+static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask,
+ unsigned NumElems) {
for (unsigned i = 0; i != NumElems; ++i) {
int idx = Mask[i];
if (idx < 0)
/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to 128-bit
-/// SHUFPS and SHUFPD.
-static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
- int NumElems = VT.getVectorNumElements();
+/// SHUFPS and SHUFPD. If Commuted is true, then it checks for sources to be
+/// reverse of what x86 shuffles want.
+static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool Commuted = false) {
+ unsigned NumElems = VT.getVectorNumElements();
if (VT.getSizeInBits() != 128)
return false;
if (NumElems != 2 && NumElems != 4)
return false;
- int Half = NumElems / 2;
- for (int i = 0; i < Half; ++i)
- if (!isUndefOrInRange(Mask[i], 0, NumElems))
+ unsigned Half = NumElems / 2;
+ unsigned SrcStart = Commuted ? NumElems : 0;
+ for (unsigned i = 0; i != Half; ++i)
+ if (!isUndefOrInRange(Mask[i], SrcStart, SrcStart+NumElems))
return false;
- for (int i = Half; i < NumElems; ++i)
- if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
+ SrcStart = Commuted ? 0 : NumElems;
+ for (unsigned i = Half; i != NumElems; ++i)
+ if (!isUndefOrInRange(Mask[i], SrcStart, SrcStart+NumElems))
return false;
return true;
return ::isSHUFPMask(M, N->getValueType(0));
}
-/// isCommutedSHUFPMask - Returns true if the shuffle mask is exactly
-/// the reverse of what x86 shuffles want. x86 shuffles requires the lower
-/// half elements to come from vector 1 (which would equal the dest.) and
-/// the upper half to come from vector 2.
-static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
- int NumElems = VT.getVectorNumElements();
-
- if (NumElems != 2 && NumElems != 4)
- return false;
-
- int Half = NumElems / 2;
- for (int i = 0; i < Half; ++i)
- if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
- return false;
- for (int i = Half; i < NumElems; ++i)
- if (!isUndefOrInRange(Mask[i], 0, NumElems))
- return false;
- return true;
-}
-
-static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
- SmallVector<int, 8> M;
- N->getMask(M);
- return isCommutedSHUFPMask(M, N->getValueType(0));
-}
-
/// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVHLPS.
bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
/// specifies a shuffle of elements that is suitable for input to UNPCKL.
static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool HasAVX2, bool V2IsSplat = false) {
- int NumElts = VT.getVectorNumElements();
+ unsigned NumElts = VT.getVectorNumElements();
assert((VT.is128BitVector() || VT.is256BitVector()) &&
"Unsupported vector type for unpckh");
unsigned NumLanes = VT.getSizeInBits()/128;
unsigned NumLaneElts = NumElts/NumLanes;
- unsigned Start = 0;
- unsigned End = NumLaneElts;
- for (unsigned s = 0; s < NumLanes; ++s) {
- for (unsigned i = Start, j = s * NumLaneElts;
- i != End;
+ for (unsigned l = 0; l != NumLanes; ++l) {
+ for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
+ i != (l+1)*NumLaneElts;
i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
return false;
}
}
- // Process the next 128 bits.
- Start += NumLaneElts;
- End += NumLaneElts;
}
return true;
/// specifies a shuffle of elements that is suitable for input to UNPCKH.
static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool HasAVX2, bool V2IsSplat = false) {
- int NumElts = VT.getVectorNumElements();
+ unsigned NumElts = VT.getVectorNumElements();
assert((VT.is128BitVector() || VT.is256BitVector()) &&
"Unsupported vector type for unpckh");
unsigned NumLanes = VT.getSizeInBits()/128;
unsigned NumLaneElts = NumElts/NumLanes;
- unsigned Start = 0;
- unsigned End = NumLaneElts;
for (unsigned l = 0; l != NumLanes; ++l) {
- for (unsigned i = Start, j = (l*NumLaneElts)+NumLaneElts/2;
- i != End; i += 2, ++j) {
+ for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
+ i != (l+1)*NumLaneElts; i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
if (!isUndefOrEqual(BitI, j))
return false;
}
}
- // Process the next 128 bits.
- Start += NumLaneElts;
- End += NumLaneElts;
}
return true;
}
/// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
/// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
/// <0, 0, 1, 1>
-static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
- int NumElems = VT.getVectorNumElements();
- if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
+static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool HasAVX2) {
+ unsigned NumElts = VT.getVectorNumElements();
+
+ assert((VT.is128BitVector() || VT.is256BitVector()) &&
+ "Unsupported vector type for unpckh");
+
+ if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
+ (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
return false;
// For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
// FIXME: Need a better way to get rid of this, there's no latency difference
// between UNPCKLPD and MOVDDUP, the later should always be checked first and
// the former later. We should also remove the "_undef" special mask.
- if (NumElems == 4 && VT.getSizeInBits() == 256)
+ if (NumElts == 4 && VT.getSizeInBits() == 256)
return false;
// Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
// independently on 128-bit lanes.
- unsigned NumLanes = VT.getSizeInBits() / 128;
- unsigned NumLaneElts = NumElems / NumLanes;
+ unsigned NumLanes = VT.getSizeInBits()/128;
+ unsigned NumLaneElts = NumElts/NumLanes;
- for (unsigned s = 0; s < NumLanes; ++s) {
- for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
- i != NumLaneElts * (s + 1);
+ for (unsigned l = 0; l != NumLanes; ++l) {
+ for (unsigned i = l*NumLaneElts, j = l*NumLaneElts;
+ i != (l+1)*NumLaneElts;
i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
return true;
}
-bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
+bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N, bool HasAVX2) {
SmallVector<int, 8> M;
N->getMask(M);
- return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
+ return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0), HasAVX2);
}
/// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
/// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
/// <2, 2, 3, 3>
-static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
- int NumElems = VT.getVectorNumElements();
- if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
+static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool HasAVX2) {
+ unsigned NumElts = VT.getVectorNumElements();
+
+ assert((VT.is128BitVector() || VT.is256BitVector()) &&
+ "Unsupported vector type for unpckh");
+
+ if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8 &&
+ (!HasAVX2 || (NumElts != 16 && NumElts != 32)))
return false;
- for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
- int BitI = Mask[i];
- int BitI1 = Mask[i+1];
- if (!isUndefOrEqual(BitI, j))
- return false;
- if (!isUndefOrEqual(BitI1, j))
- return false;
+ // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
+ // independently on 128-bit lanes.
+ unsigned NumLanes = VT.getSizeInBits()/128;
+ unsigned NumLaneElts = NumElts/NumLanes;
+
+ for (unsigned l = 0; l != NumLanes; ++l) {
+ for (unsigned i = l*NumLaneElts, j = (l*NumLaneElts)+NumLaneElts/2;
+ i != (l+1)*NumLaneElts; i += 2, ++j) {
+ int BitI = Mask[i];
+ int BitI1 = Mask[i+1];
+ if (!isUndefOrEqual(BitI, j))
+ return false;
+ if (!isUndefOrEqual(BitI1, j))
+ return false;
+ }
}
return true;
}
-bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
+bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N, bool HasAVX2) {
SmallVector<int, 8> M;
N->getMask(M);
- return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
+ return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0), HasAVX2);
}
/// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
/// getShuffleVPERM2X128Immediate - Return the appropriate immediate to shuffle
/// the specified VECTOR_MASK mask with VPERM2F128/VPERM2I128 instructions.
-static unsigned getShuffleVPERM2X128Immediate(SDNode *N) {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
+static unsigned getShuffleVPERM2X128Immediate(ShuffleVectorSDNode *SVOp) {
EVT VT = SVOp->getValueType(0);
int HalfSize = VT.getVectorNumElements()/2;
/// getShuffleVPERMILPImmediate - Return the appropriate immediate to shuffle
/// the specified VECTOR_MASK mask with VPERMILPS/D* instructions.
-static unsigned getShuffleVPERMILPImmediate(SDNode *N) {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
+static unsigned getShuffleVPERMILPImmediate(ShuffleVectorSDNode *SVOp) {
EVT VT = SVOp->getValueType(0);
int NumElts = VT.getVectorNumElements();
/// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to 256-bit
/// version of MOVDDUP.
-static bool isMOVDDUPYMask(ShuffleVectorSDNode *N,
- const X86Subtarget *Subtarget) {
- EVT VT = N->getValueType(0);
+static bool isMOVDDUPYMask(const SmallVectorImpl<int> &Mask, EVT VT,
+ bool HasAVX) {
int NumElts = VT.getVectorNumElements();
- bool V2IsUndef = N->getOperand(1).getOpcode() == ISD::UNDEF;
- if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256 ||
- !V2IsUndef || NumElts != 4)
+ if (!HasAVX || VT.getSizeInBits() != 256 || NumElts != 4)
return false;
for (int i = 0; i != NumElts/2; ++i)
- if (!isUndefOrEqual(N->getMaskElt(i), 0))
+ if (!isUndefOrEqual(Mask[i], 0))
return false;
for (int i = NumElts/2; i != NumElts; ++i)
- if (!isUndefOrEqual(N->getMaskElt(i), NumElts/2))
+ if (!isUndefOrEqual(Mask[i], NumElts/2))
return false;
return true;
}
/// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
/// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
-unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
- ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
- EVT VVT = N->getValueType(0);
- unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
+static unsigned getShufflePALIGNRImmediate(ShuffleVectorSDNode *SVOp) {
+ EVT VT = SVOp->getValueType(0);
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits() >> 3;
int Val = 0;
unsigned i, e;
- for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
+ for (i = 0, e = VT.getVectorNumElements(); i != e; ++i) {
Val = SVOp->getMaskElt(i);
if (Val >= 0)
break;
DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
ShuffleMask);
break;
- case X86ISD::PUNPCKH:
- DecodePUNPCKHMask(NumElems, ShuffleMask);
- break;
- case X86ISD::UNPCKHP:
- DecodeUNPCKHPMask(VT, ShuffleMask);
+ case X86ISD::UNPCKH:
+ DecodeUNPCKHMask(VT, ShuffleMask);
break;
- case X86ISD::PUNPCKL:
- DecodePUNPCKLMask(VT, ShuffleMask);
- break;
- case X86ISD::UNPCKLP:
- DecodeUNPCKLPMask(VT, ShuffleMask);
+ case X86ISD::UNPCKL:
+ DecodeUNPCKLMask(VT, ShuffleMask);
break;
case X86ISD::MOVHLPS:
DecodeMOVHLPSMask(NumElems, ShuffleMask);
DAG);
} else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
- assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
- EVT MiddleVT = MVT::v4i32;
+ unsigned NumBits = VT.getSizeInBits();
+ assert((NumBits == 128 || NumBits == 256) &&
+ "Expected an SSE or AVX value type!");
+ EVT MiddleVT = NumBits == 128 ? MVT::v4i32 : MVT::v8i32;
Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
Item = getShuffleVectorZeroOrUndef(Item, 0, true,
Subtarget->hasXMMInt(), DAG);
// from X.
if (NumHi == 3) {
// Normalize it so the 3 elements come from V1.
- CommuteVectorShuffleMask(PermMask, VT);
+ CommuteVectorShuffleMask(PermMask, 4);
std::swap(V1, V2);
}
X86::getShuffleSHUFImmediate(SVOp), DAG);
}
-static inline unsigned getUNPCKLOpcode(EVT VT, bool HasAVX2) {
- switch(VT.getSimpleVT().SimpleTy) {
- case MVT::v32i8:
- case MVT::v16i8:
- case MVT::v16i16:
- case MVT::v8i16:
- case MVT::v4i32:
- case MVT::v2i64: return X86ISD::PUNPCKL;
- case MVT::v8i32:
- case MVT::v4i64:
- if (HasAVX2) return X86ISD::PUNPCKL;
- // else use fp unit for int unpack.
- case MVT::v8f32:
- case MVT::v4f32:
- case MVT::v4f64:
- case MVT::v2f64: return X86ISD::UNPCKLP;
- default:
- llvm_unreachable("Unknown type for unpckl");
- }
- return 0;
-}
-
-static inline unsigned getUNPCKHOpcode(EVT VT, bool HasAVX2) {
- switch(VT.getSimpleVT().SimpleTy) {
- case MVT::v32i8:
- case MVT::v16i8:
- case MVT::v16i16:
- case MVT::v8i16:
- case MVT::v4i32:
- case MVT::v2i64: return X86ISD::PUNPCKH;
- case MVT::v4i64:
- case MVT::v8i32:
- if (HasAVX2) return X86ISD::PUNPCKH;
- // else use fp unit for int unpack.
- case MVT::v8f32:
- case MVT::v4f32:
- case MVT::v4f64:
- case MVT::v2f64: return X86ISD::UNPCKHP;
- default:
- llvm_unreachable("Unknown type for unpckh");
- }
- return 0;
-}
-
static
SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
const TargetLowering &TLI,
bool V1IsSplat = false;
bool V2IsSplat = false;
bool HasXMMInt = Subtarget->hasXMMInt();
+ bool HasAVX = Subtarget->hasAVX();
bool HasAVX2 = Subtarget->hasAVX2();
MachineFunction &MF = DAG.getMachineFunction();
bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
// NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
// unpckh_undef). Only use pshufd if speed is more important than size.
- if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
- return getTargetShuffleNode(getUNPCKLOpcode(VT, HasAVX2), dl, VT, V1, V1,
- DAG);
- if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
- return getTargetShuffleNode(getUNPCKHOpcode(VT, HasAVX2), dl, VT, V1, V1,
- DAG);
+ if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp, HasAVX2))
+ return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
+ if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp, HasAVX2))
+ return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
if (X86::isMOVDDUPMask(SVOp) && Subtarget->hasSSE3orAVX() &&
V2IsUndef && RelaxedMayFoldVectorLoad(V1))
// Use to match splats
if (HasXMMInt && X86::isUNPCKHMask(SVOp, HasAVX2) && V2IsUndef &&
(VT == MVT::v2f64 || VT == MVT::v2i64))
- return getTargetShuffleNode(getUNPCKHOpcode(VT, HasAVX2), dl, VT, V1, V1,
- DAG);
+ return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
if (X86::isPSHUFDMask(SVOp)) {
// The actual implementation will match the mask in the if above and then
Commuted = true;
}
- if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
+ SmallVector<int, 32> M;
+ SVOp->getMask(M);
+
+ if (isCommutedMOVLMask(M, VT, V2IsSplat, V2IsUndef)) {
// Shuffling low element of v1 into undef, just return v1.
if (V2IsUndef)
return V1;
return getMOVL(DAG, dl, VT, V2, V1);
}
- if (X86::isUNPCKLMask(SVOp, HasAVX2))
- return getTargetShuffleNode(getUNPCKLOpcode(VT, HasAVX2), dl, VT, V1, V2,
- DAG);
+ if (isUNPCKLMask(M, VT, HasAVX2))
+ return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
- if (X86::isUNPCKHMask(SVOp, HasAVX2))
- return getTargetShuffleNode(getUNPCKHOpcode(VT, HasAVX2), dl, VT, V1, V2,
- DAG);
+ if (isUNPCKHMask(M, VT, HasAVX2))
+ return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V2, DAG);
if (V2IsSplat) {
// Normalize mask so all entries that point to V2 points to its first
ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
if (X86::isUNPCKLMask(NewSVOp, HasAVX2))
- return getTargetShuffleNode(getUNPCKLOpcode(VT, HasAVX2), dl, VT, V2, V1,
- DAG);
+ return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V2, V1, DAG);
if (X86::isUNPCKHMask(NewSVOp, HasAVX2))
- return getTargetShuffleNode(getUNPCKHOpcode(VT, HasAVX2), dl, VT, V2, V1,
- DAG);
+ return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V2, V1, DAG);
}
// Normalize the node to match x86 shuffle ops if needed
- if (!V2IsUndef && (isCommutedSHUFP(SVOp) ||
- isCommutedVSHUFPY(SVOp, Subtarget->hasAVX())))
+ if (!V2IsUndef && (isSHUFPMask(M, VT, /* Commuted */ true) ||
+ isVSHUFPYMask(M, VT, HasAVX, /* Commuted */ true)))
return CommuteVectorShuffle(SVOp, DAG);
// The checks below are all present in isShuffleMaskLegal, but they are
// inlined here right now to enable us to directly emit target specific
// nodes, and remove one by one until they don't return Op anymore.
- SmallVector<int, 16> M;
- SVOp->getMask(M);
if (isPALIGNRMask(M, VT, Subtarget->hasSSSE3orAVX()))
return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
- X86::getShufflePALIGNRImmediate(SVOp),
+ getShufflePALIGNRImmediate(SVOp),
DAG);
if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
SVOp->getSplatIndex() == 0 && V2IsUndef) {
- if (VT == MVT::v2f64)
- return getTargetShuffleNode(X86ISD::UNPCKLP, dl, VT, V1, V1, DAG);
- if (VT == MVT::v2i64)
- return getTargetShuffleNode(X86ISD::PUNPCKL, dl, VT, V1, V1, DAG);
+ if (VT == MVT::v2f64 || VT == MVT::v2i64)
+ return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
}
if (isPSHUFHWMask(M, VT))
return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
X86::getShuffleSHUFImmediate(SVOp), DAG);
- if (X86::isUNPCKL_v_undef_Mask(SVOp))
- return getTargetShuffleNode(getUNPCKLOpcode(VT, HasAVX2), dl, VT, V1, V1,
- DAG);
- if (X86::isUNPCKH_v_undef_Mask(SVOp))
- return getTargetShuffleNode(getUNPCKHOpcode(VT, HasAVX2), dl, VT, V1, V1,
- DAG);
+ if (isUNPCKL_v_undef_Mask(M, VT, HasAVX2))
+ return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V1, DAG);
+ if (isUNPCKH_v_undef_Mask(M, VT, HasAVX2))
+ return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
//===--------------------------------------------------------------------===//
// Generate target specific nodes for 128 or 256-bit shuffles only
//
// Handle VMOVDDUPY permutations
- if (isMOVDDUPYMask(SVOp, Subtarget))
+ if (V2IsUndef && isMOVDDUPYMask(M, VT, HasAVX))
return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
// Handle VPERMILPS/D* permutations
- if (isVPERMILPMask(M, VT, Subtarget->hasAVX()))
+ if (isVPERMILPMask(M, VT, HasAVX))
return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1,
getShuffleVPERMILPImmediate(SVOp), DAG);
// Handle VPERM2F128/VPERM2I128 permutations
- if (isVPERM2X128Mask(M, VT, Subtarget->hasAVX()))
+ if (isVPERM2X128Mask(M, VT, HasAVX))
return getTargetShuffleNode(X86ISD::VPERM2X128, dl, VT, V1,
V2, getShuffleVPERM2X128Immediate(SVOp), DAG);
// Handle VSHUFPS/DY permutations
- if (isVSHUFPYMask(M, VT, Subtarget->hasAVX()))
+ if (isVSHUFPYMask(M, VT, HasAVX))
return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
getShuffleVSHUFPYImmediate(SVOp), DAG);
LLVMContext *Context = DAG.getContext();
// Build some magic constants.
- std::vector<Constant*> CV0;
+ SmallVector<Constant*,4> CV0;
CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
Constant *C0 = ConstantVector::get(CV0);
SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
- std::vector<Constant*> CV1;
+ SmallVector<Constant*,2> CV1;
CV1.push_back(
ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
CV1.push_back(
EVT EltVT = VT;
if (VT.isVector())
EltVT = VT.getVectorElementType();
- std::vector<Constant*> CV;
+ SmallVector<Constant*,4> CV;
if (EltVT == MVT::f64) {
Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
- CV.push_back(C);
- CV.push_back(C);
+ CV.assign(2, C);
} else {
Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
- CV.push_back(C);
- CV.push_back(C);
- CV.push_back(C);
- CV.push_back(C);
+ CV.assign(4, C);
}
Constant *C = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
EVT EltVT = VT;
- if (VT.isVector())
+ unsigned NumElts = VT == MVT::f64 ? 2 : 4;
+ if (VT.isVector()) {
EltVT = VT.getVectorElementType();
- std::vector<Constant*> CV;
+ NumElts = VT.getVectorNumElements();
+ }
+ SmallVector<Constant*,8> CV;
if (EltVT == MVT::f64) {
Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
- CV.push_back(C);
- CV.push_back(C);
+ CV.assign(NumElts, C);
} else {
Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
- CV.push_back(C);
- CV.push_back(C);
- CV.push_back(C);
- CV.push_back(C);
+ CV.assign(NumElts, C);
}
Constant *C = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
MachinePointerInfo::getConstantPool(),
false, false, false, 16);
if (VT.isVector()) {
+ MVT XORVT = VT.getSizeInBits() == 128 ? MVT::v2i64 : MVT::v4i64;
return DAG.getNode(ISD::BITCAST, dl, VT,
- DAG.getNode(ISD::XOR, dl, MVT::v2i64,
- DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
+ DAG.getNode(ISD::XOR, dl, XORVT,
+ DAG.getNode(ISD::BITCAST, dl, XORVT,
Op.getOperand(0)),
- DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
+ DAG.getNode(ISD::BITCAST, dl, XORVT, Mask)));
} else {
return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
}
// type, and that won't be f80 since that is not custom lowered.
// First get the sign bit of second operand.
- std::vector<Constant*> CV;
+ SmallVector<Constant*,4> CV;
if (SrcVT == MVT::f64) {
CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
return DAG.getNode(ISD::MUL, dl, VT, Op, R);
}
if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
+ assert((Subtarget->hasSSE2() || Subtarget->hasAVX()) &&
+ "Need SSE2 for pslli/pcmpeq.");
+
// a = a << 5;
Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
Op.getOperand(1), DAG.getConstant(5, MVT::i32));
- ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
- ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
+ // Turn 'a' into a mask suitable for VSELECT
+ SDValue VSelM = DAG.getConstant(0x80, VT);
+ SDValue OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
+ OpVSel = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ DAG.getConstant(Intrinsic::x86_sse2_pcmpeq_b, MVT::i32),
+ OpVSel, VSelM);
- std::vector<Constant*> CVM1(16, CM1);
- std::vector<Constant*> CVM2(16, CM2);
- Constant *C = ConstantVector::get(CVM1);
- SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
- SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(),
- false, false, false, 16);
+ SDValue CM1 = DAG.getConstant(0x0f, VT);
+ SDValue CM2 = DAG.getConstant(0x3f, VT);
- // r = pblendv(r, psllw(r & (char16)15, 4), a);
- M = DAG.getNode(ISD::AND, dl, VT, R, M);
+ // r = VSELECT(r, psllw(r & (char16)15, 4), a);
+ SDValue M = DAG.getNode(ISD::AND, dl, VT, R, CM1);
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(4, MVT::i32));
- R = DAG.getNode(ISD::VSELECT, dl, VT, Op, R, M);
+ R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
+
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
+ OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
+ OpVSel = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ DAG.getConstant(Intrinsic::x86_sse2_pcmpeq_b, MVT::i32),
+ OpVSel, VSelM);
- C = ConstantVector::get(CVM2);
- CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
- M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
- MachinePointerInfo::getConstantPool(),
- false, false, false, 16);
-
- // r = pblendv(r, psllw(r & (char16)63, 2), a);
- M = DAG.getNode(ISD::AND, dl, VT, R, M);
+ // r = VSELECT(r, psllw(r & (char16)63, 2), a);
+ M = DAG.getNode(ISD::AND, dl, VT, R, CM2);
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(2, MVT::i32));
- R = DAG.getNode(ISD::VSELECT, dl, VT, Op, R, M);
+ R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel, M, R);
+
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
-
- // return pblendv(r, r+r, a);
- R = DAG.getNode(ISD::VSELECT, dl, VT, Op,
- R, DAG.getNode(ISD::ADD, dl, VT, R, R));
+ OpVSel = DAG.getNode(ISD::AND, dl, VT, VSelM, Op);
+ OpVSel = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
+ DAG.getConstant(Intrinsic::x86_sse2_pcmpeq_b, MVT::i32),
+ OpVSel, VSelM);
+
+ // return VSELECT(r, r+r, a);
+ R = DAG.getNode(ISD::VSELECT, dl, VT, OpVSel,
+ DAG.getNode(ISD::ADD, dl, VT, R, R), R);
return R;
}
case X86ISD::ANDNP: return "X86ISD::ANDNP";
case X86ISD::PSIGN: return "X86ISD::PSIGN";
case X86ISD::BLENDV: return "X86ISD::BLENDV";
+ case X86ISD::HADD: return "X86ISD::HADD";
+ case X86ISD::HSUB: return "X86ISD::HSUB";
case X86ISD::FHADD: return "X86ISD::FHADD";
case X86ISD::FHSUB: return "X86ISD::FHSUB";
case X86ISD::FMAX: return "X86ISD::FMAX";
case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
- case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
case X86ISD::MOVSD: return "X86ISD::MOVSD";
case X86ISD::MOVSS: return "X86ISD::MOVSS";
- case X86ISD::UNPCKLP: return "X86ISD::UNPCKLP";
- case X86ISD::UNPCKHP: return "X86ISD::UNPCKHP";
- case X86ISD::PUNPCKL: return "X86ISD::PUNPCKL";
- case X86ISD::PUNPCKH: return "X86ISD::PUNPCKH";
+ case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
+ case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
case X86ISD::VPERMILP: return "X86ISD::VPERMILP";
case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
EVT VT) const {
// Very little shuffling can be done for 64-bit vectors right now.
if (VT.getSizeInBits() == 64)
- return isPALIGNRMask(M, VT, Subtarget->hasSSSE3orAVX());
+ return false;
// FIXME: pshufb, blends, shifts.
return (VT.getVectorNumElements() == 2 ||
isPALIGNRMask(M, VT, Subtarget->hasSSSE3orAVX()) ||
isUNPCKLMask(M, VT, Subtarget->hasAVX2()) ||
isUNPCKHMask(M, VT, Subtarget->hasAVX2()) ||
- isUNPCKL_v_undef_Mask(M, VT) ||
- isUNPCKH_v_undef_Mask(M, VT));
+ isUNPCKL_v_undef_Mask(M, VT, Subtarget->hasAVX2()) ||
+ isUNPCKH_v_undef_Mask(M, VT, Subtarget->hasAVX2()));
}
bool
return (isMOVLMask(Mask, VT) ||
isCommutedMOVLMask(Mask, VT, true) ||
isSHUFPMask(Mask, VT) ||
- isCommutedSHUFPMask(Mask, VT));
+ isSHUFPMask(Mask, VT, /* Commuted */ true));
}
return false;
}
/// set to A, RHS to B, and the routine returns 'true'.
/// Note that the binary operation should have the property that if one of the
/// operands is UNDEF then the result is UNDEF.
-static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool isCommutative) {
+static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
// Look for the following pattern: if
// A = < float a0, float a1, float a2, float a3 >
// B = < float b0, float b1, float b2, float b3 >
// If A and B occur in reverse order in RHS, then "swap" them (which means
// rewriting the mask).
if (A != C)
- for (unsigned i = 0; i != NumElts; ++i) {
- unsigned Idx = RMask[i];
- if (Idx < NumElts)
- RMask[i] += NumElts;
- else if (Idx < 2*NumElts)
- RMask[i] -= NumElts;
- }
+ CommuteVectorShuffleMask(RMask, NumElts);
// At this point LHS and RHS are equivalent to
// LHS = VECTOR_SHUFFLE A, B, LMask
// RHS = VECTOR_SHUFFLE A, B, RMask
// Check that the masks correspond to performing a horizontal operation.
for (unsigned i = 0; i != NumElts; ++i) {
- unsigned LIdx = LMask[i], RIdx = RMask[i];
+ int LIdx = LMask[i], RIdx = RMask[i];
// Ignore any UNDEF components.
- if (LIdx >= 2*NumElts || RIdx >= 2*NumElts ||
- (!A.getNode() && (LIdx < NumElts || RIdx < NumElts)) ||
- (!B.getNode() && (LIdx >= NumElts || RIdx >= NumElts)))
+ if (LIdx < 0 || RIdx < 0 ||
+ (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
+ (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
continue;
// Check that successive elements are being operated on. If not, this is
// not a horizontal operation.
unsigned Src = (i/HalfLaneElts) % 2; // each lane is split between srcs
unsigned LaneStart = (i/NumLaneElts) * NumLaneElts;
- unsigned Index = 2*(i%HalfLaneElts) + NumElts*Src + LaneStart;
+ int Index = 2*(i%HalfLaneElts) + NumElts*Src + LaneStart;
if (!(LIdx == Index && RIdx == Index + 1) &&
- !(isCommutative && LIdx == Index + 1 && RIdx == Index))
+ !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
return false;
}
case X86ISD::SHUFPS: // Handle all target specific shuffles
case X86ISD::SHUFPD:
case X86ISD::PALIGN:
- case X86ISD::PUNPCKH:
- case X86ISD::UNPCKHP:
- case X86ISD::PUNPCKL:
- case X86ISD::UNPCKLP:
+ case X86ISD::UNPCKH:
+ case X86ISD::UNPCKL:
case X86ISD::MOVHLPS:
case X86ISD::MOVLHPS:
case X86ISD::PSHUFD:
// X86 Inline Assembly Support
//===----------------------------------------------------------------------===//
+// Helper to match a string separated by whitespace.
+static bool END_WITH_NULL matchAsm(StringRef s, ...) {
+ va_list ap;
+ va_start(ap, s);
+ s = s.substr(s.find_first_not_of(" \t")); // Skip leading whitespace.
+
+ while (const char *p = va_arg(ap, const char *)) {
+ StringRef piece(p);
+ if (!s.startswith(piece)) { // Check if the piece matches.
+ va_end(ap);
+ return false;
+ }
+
+ s = s.substr(piece.size());
+ StringRef::size_type i = s.find_first_not_of(" \t");
+ if (i == 0) { // We matched a prefix.
+ va_end(ap);
+ return false;
+ }
+ s = s.substr(i);
+ }
+
+ va_end(ap);
+ return s.empty();
+}
+
bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
std::string AsmStr = IA->getAsmString();
+ IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
+ if (!Ty || Ty->getBitWidth() % 16 != 0)
+ return false;
+
// TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
SmallVector<StringRef, 4> AsmPieces;
SplitString(AsmStr, AsmPieces, ";\n");
switch (AsmPieces.size()) {
default: return false;
case 1:
- AsmStr = AsmPieces[0];
- AsmPieces.clear();
- SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
-
// FIXME: this should verify that we are targeting a 486 or better. If not,
- // we will turn this bswap into something that will be lowered to logical ops
- // instead of emitting the bswap asm. For now, we don't support 486 or lower
- // so don't worry about this.
+ // we will turn this bswap into something that will be lowered to logical
+ // ops instead of emitting the bswap asm. For now, we don't support 486 or
+ // lower so don't worry about this.
// bswap $0
- if (AsmPieces.size() == 2 &&
- (AsmPieces[0] == "bswap" ||
- AsmPieces[0] == "bswapq" ||
- AsmPieces[0] == "bswapl") &&
- (AsmPieces[1] == "$0" ||
- AsmPieces[1] == "${0:q}")) {
+ if (matchAsm(AsmPieces[0], "bswap", "$0", NULL) ||
+ matchAsm(AsmPieces[0], "bswapl", "$0", NULL) ||
+ matchAsm(AsmPieces[0], "bswapq", "$0", NULL) ||
+ matchAsm(AsmPieces[0], "bswap", "${0:q}", NULL) ||
+ matchAsm(AsmPieces[0], "bswapl", "${0:q}", NULL) ||
+ matchAsm(AsmPieces[0], "bswapq", "${0:q}", NULL)) {
// No need to check constraints, nothing other than the equivalent of
// "=r,0" would be valid here.
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
return IntrinsicLowering::LowerToByteSwap(CI);
}
+
// rorw $$8, ${0:w} --> llvm.bswap.i16
if (CI->getType()->isIntegerTy(16) &&
- AsmPieces.size() == 3 &&
- (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
- AsmPieces[1] == "$$8," &&
- AsmPieces[2] == "${0:w}" &&
- IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
+ IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
+ (matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}", NULL) ||
+ matchAsm(AsmPieces[0], "rolw", "$$8,", "${0:w}", NULL))) {
AsmPieces.clear();
const std::string &ConstraintsStr = IA->getConstraintString();
SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
AsmPieces[0] == "~{cc}" &&
AsmPieces[1] == "~{dirflag}" &&
AsmPieces[2] == "~{flags}" &&
- AsmPieces[3] == "~{fpsr}") {
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
- return IntrinsicLowering::LowerToByteSwap(CI);
- }
+ AsmPieces[3] == "~{fpsr}")
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
break;
case 3:
if (CI->getType()->isIntegerTy(32) &&
- IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
- SmallVector<StringRef, 4> Words;
- SplitString(AsmPieces[0], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
- Words[2] == "${0:w}") {
- Words.clear();
- SplitString(AsmPieces[1], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
- Words[2] == "$0") {
- Words.clear();
- SplitString(AsmPieces[2], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
- Words[2] == "${0:w}") {
- AsmPieces.clear();
- const std::string &ConstraintsStr = IA->getConstraintString();
- SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
- std::sort(AsmPieces.begin(), AsmPieces.end());
- if (AsmPieces.size() == 4 &&
- AsmPieces[0] == "~{cc}" &&
- AsmPieces[1] == "~{dirflag}" &&
- AsmPieces[2] == "~{flags}" &&
- AsmPieces[3] == "~{fpsr}") {
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
- return IntrinsicLowering::LowerToByteSwap(CI);
- }
- }
- }
- }
+ IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
+ matchAsm(AsmPieces[0], "rorw", "$$8,", "${0:w}", NULL) &&
+ matchAsm(AsmPieces[1], "rorl", "$$16,", "$0", NULL) &&
+ matchAsm(AsmPieces[2], "rorw", "$$8,", "${0:w}", NULL)) {
+ AsmPieces.clear();
+ const std::string &ConstraintsStr = IA->getConstraintString();
+ SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
+ std::sort(AsmPieces.begin(), AsmPieces.end());
+ if (AsmPieces.size() == 4 &&
+ AsmPieces[0] == "~{cc}" &&
+ AsmPieces[1] == "~{dirflag}" &&
+ AsmPieces[2] == "~{flags}" &&
+ AsmPieces[3] == "~{fpsr}")
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
if (CI->getType()->isIntegerTy(64)) {
Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
// bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
- SmallVector<StringRef, 4> Words;
- SplitString(AsmPieces[0], Words, " \t");
- if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
- Words.clear();
- SplitString(AsmPieces[1], Words, " \t");
- if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
- Words.clear();
- SplitString(AsmPieces[2], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
- Words[2] == "%edx") {
- IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
- return IntrinsicLowering::LowerToByteSwap(CI);
- }
- }
- }
+ if (matchAsm(AsmPieces[0], "bswap", "%eax", NULL) &&
+ matchAsm(AsmPieces[1], "bswap", "%edx", NULL) &&
+ matchAsm(AsmPieces[2], "xchgl", "%eax,", "%edx", NULL))
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
}
break;
projects / oota-llvm.git / blobdiff