if (Subtarget->hasXMM())
setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
- // We may not have a libcall for MEMBARRIER so we should lower this.
setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
+ setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
// On X86 and X86-64, atomic operations are lowered to locked instructions.
// Locked instructions, in turn, have implicit fence semantics (all memory
setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
+ setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
+ setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
+
// Custom lower several nodes for 256-bit types.
for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
case X86ISD::PUNPCKLQDQ:
case X86ISD::UNPCKHPS:
case X86ISD::UNPCKHPD:
+ case X86ISD::VUNPCKHPSY:
+ case X86ISD::VUNPCKHPDY:
case X86ISD::PUNPCKHWD:
case X86ISD::PUNPCKHBW:
case X86ISD::PUNPCKHDQ:
case X86ISD::PUNPCKHQDQ:
- case X86ISD::VPERMIL:
+ case X86ISD::VPERMILPS:
+ case X86ISD::VPERMILPSY:
+ case X86ISD::VPERMILPD:
+ case X86ISD::VPERMILPDY:
return true;
}
return false;
case X86ISD::PSHUFD:
case X86ISD::PSHUFHW:
case X86ISD::PSHUFLW:
- case X86ISD::VPERMIL:
+ case X86ISD::VPERMILPS:
+ case X86ISD::VPERMILPSY:
+ case X86ISD::VPERMILPD:
+ case X86ISD::VPERMILPDY:
return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
}
case X86ISD::PUNPCKLQDQ:
case X86ISD::UNPCKHPS:
case X86ISD::UNPCKHPD:
+ case X86ISD::VUNPCKHPSY:
+ case X86ISD::VUNPCKHPDY:
case X86ISD::PUNPCKHWD:
case X86ISD::PUNPCKHBW:
case X86ISD::PUNPCKHDQ:
static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool hasSSSE3) {
int i, e = VT.getVectorNumElements();
+ if (VT.getSizeInBits() != 128 && VT.getSizeInBits() != 64)
+ return false;
// Do not handle v2i64 / v2f64 shuffles with palignr.
if (e < 4 || !hasSSSE3)
return true;
}
-bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
- SmallVector<int, 8> M;
- N->getMask(M);
- return ::isPALIGNRMask(M, N->getValueType(0), true);
-}
-
/// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to SHUFP*.
static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool V2IsSplat = false) {
int NumElts = VT.getVectorNumElements();
- if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
+
+ assert((VT.is128BitVector() || VT.is256BitVector()) &&
+ "Unsupported vector type for unpckh");
+
+ if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
return false;
- // Handle vector lengths > 128 bits. Define a "section" as a set of
- // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
- // sections.
- unsigned NumSections = VT.getSizeInBits() / 128;
- if (NumSections == 0 ) NumSections = 1; // Handle MMX
- unsigned NumSectionElts = NumElts / NumSections;
+ // 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;
unsigned Start = 0;
- unsigned End = NumSectionElts;
- for (unsigned s = 0; s < NumSections; ++s) {
- for (unsigned i = Start, j = s * NumSectionElts;
+ unsigned End = NumLaneElts;
+ for (unsigned s = 0; s < NumLanes; ++s) {
+ for (unsigned i = Start, j = s * NumLaneElts;
i != End;
i += 2, ++j) {
int BitI = Mask[i];
}
}
// Process the next 128 bits.
- Start += NumSectionElts;
- End += NumSectionElts;
+ Start += NumLaneElts;
+ End += NumLaneElts;
}
return true;
static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool V2IsSplat = false) {
int NumElts = VT.getVectorNumElements();
- if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
+
+ assert((VT.is128BitVector() || VT.is256BitVector()) &&
+ "Unsupported vector type for unpckh");
+
+ if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
return false;
- for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
- int BitI = Mask[i];
- int BitI1 = Mask[i+1];
- if (!isUndefOrEqual(BitI, j + NumElts/2))
- return false;
- if (V2IsSplat) {
- if (isUndefOrEqual(BitI1, NumElts))
- return false;
- } else {
- if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
+ // 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;
+
+ 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) {
+ int BitI = Mask[i];
+ int BitI1 = Mask[i+1];
+ if (!isUndefOrEqual(BitI, j))
return false;
+ if (V2IsSplat) {
+ if (isUndefOrEqual(BitI1, NumElts))
+ return false;
+ } else {
+ if (!isUndefOrEqual(BitI1, j+NumElts))
+ return false;
+ }
}
+ // Process the next 128 bits.
+ Start += NumLaneElts;
+ End += NumLaneElts;
}
return true;
}
if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
return false;
- // Handle vector lengths > 128 bits. Define a "section" as a set of
- // 128 bits. AVX defines UNPCK* to operate independently on 128-bit
- // sections.
- unsigned NumSections = VT.getSizeInBits() / 128;
- if (NumSections == 0 ) NumSections = 1; // Handle MMX
- unsigned NumSectionElts = NumElems / NumSections;
+ // 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;
- for (unsigned s = 0; s < NumSections; ++s) {
- for (unsigned i = s * NumSectionElts, j = s * NumSectionElts;
- i != NumSectionElts * (s + 1);
+ for (unsigned s = 0; s < NumLanes; ++s) {
+ for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
+ i != NumLaneElts * (s + 1);
i += 2, ++j) {
int BitI = Mask[i];
int BitI1 = Mask[i+1];
return ::isMOVLMask(M, N->getValueType(0));
}
-/// isVPERMILMask - Return true if the specified VECTOR_SHUFFLE operand
-/// specifies a shuffle of elements that is suitable for input to VPERMIL*.
-static bool isVPERMILMask(const SmallVectorImpl<int> &Mask, EVT VT) {
+/// isVPERMILPDMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
+/// Note that VPERMIL mask matching is different depending whether theunderlying
+/// type is 32 or 64. In the VPERMILPS the high half of the mask should point
+/// to the same elements of the low, but to the higher half of the source.
+/// In VPERMILPD the two lanes could be shuffled independently of each other
+/// with the same restriction that lanes can't be crossed.
+static bool isVPERMILPDMask(const SmallVectorImpl<int> &Mask, EVT VT,
+ const X86Subtarget *Subtarget) {
+ int NumElts = VT.getVectorNumElements();
+ int NumLanes = VT.getSizeInBits()/128;
+
+ if (!Subtarget->hasAVX())
+ return false;
+
+ // Match any permutation of 128-bit vector with 64-bit types
+ if (NumLanes == 1 && NumElts != 2)
+ return false;
+
+ // Only match 256-bit with 32 types
+ if (VT.getSizeInBits() == 256 && NumElts != 4)
+ return false;
+
+ // The mask on the high lane is independent of the low. Both can match
+ // any element in inside its own lane, but can't cross.
+ int LaneSize = NumElts/NumLanes;
+ for (int l = 0; l < NumLanes; ++l)
+ for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
+ int LaneStart = l*LaneSize;
+ if (!isUndefOrInRange(Mask[i], LaneStart, LaneStart+LaneSize))
+ return false;
+ }
+
+ return true;
+}
+
+/// isVPERMILPSMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a shuffle of elements that is suitable for input to VPERMILPS*.
+/// Note that VPERMIL mask matching is different depending whether theunderlying
+/// type is 32 or 64. In the VPERMILPS the high half of the mask should point
+/// to the same elements of the low, but to the higher half of the source.
+/// In VPERMILPD the two lanes could be shuffled independently of each other
+/// with the same restriction that lanes can't be crossed.
+static bool isVPERMILPSMask(const SmallVectorImpl<int> &Mask, EVT VT,
+ const X86Subtarget *Subtarget) {
unsigned NumElts = VT.getVectorNumElements();
unsigned NumLanes = VT.getSizeInBits()/128;
- // Match any permutation of 128-bit vector with 32/64-bit types
- if (NumLanes == 1) {
- if (NumElts == 4 || NumElts == 2)
- return true;
+ if (!Subtarget->hasAVX())
+ return false;
+
+ // Match any permutation of 128-bit vector with 32-bit types
+ if (NumLanes == 1 && NumElts != 4)
return false;
- }
- // Only match 256-bit with 32/64-bit types
- if (NumElts != 8 && NumElts != 4)
+ // Only match 256-bit with 32 types
+ if (VT.getSizeInBits() == 256 && NumElts != 8)
return false;
// The mask on the high lane should be the same as the low. Actually,
int HighElt = i+LaneSize;
if (Mask[i] < 0 || Mask[HighElt] < 0)
continue;
-
if (Mask[HighElt]-Mask[i] != LaneSize)
return false;
}
return true;
}
-/// getShuffleVPERMILImmediateediate - Return the appropriate immediate to shuffle
-/// the specified VECTOR_MASK mask with VPERMIL* instructions.
-static unsigned getShuffleVPERMILImmediate(SDNode *N) {
+/// getShuffleVPERMILPSImmediate - Return the appropriate immediate to shuffle
+/// the specified VECTOR_MASK mask with VPERMILPS* instructions.
+static unsigned getShuffleVPERMILPSImmediate(SDNode *N) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
EVT VT = SVOp->getValueType(0);
return Mask;
}
+/// getShuffleVPERMILPDImmediate - Return the appropriate immediate to shuffle
+/// the specified VECTOR_MASK mask with VPERMILPD* instructions.
+static unsigned getShuffleVPERMILPDImmediate(SDNode *N) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
+ EVT VT = SVOp->getValueType(0);
+
+ int NumElts = VT.getVectorNumElements();
+ int NumLanes = VT.getSizeInBits()/128;
+
+ unsigned Mask = 0;
+ int LaneSize = NumElts/NumLanes;
+ for (int l = 0; l < NumLanes; ++l)
+ for (int i = l*LaneSize; i < LaneSize*(l+1); ++i)
+ Mask |= (SVOp->getMaskElt(i)-l*LaneSize) << i;
+
+ return Mask;
+}
+
/// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
/// of what x86 movss want. X86 movs requires the lowest element to be lowest
/// element of vector 2 and the other elements to come from vector 1 in order.
/// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
-bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
- if (N->getValueType(0).getVectorNumElements() != 4)
+/// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
+bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N,
+ const X86Subtarget *Subtarget) {
+ if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
return false;
- // Expect 1, 1, 3, 3
- for (unsigned i = 0; i < 2; ++i) {
- int Elt = N->getMaskElt(i);
- if (Elt >= 0 && Elt != 1)
- return false;
- }
+ // The second vector must be undef
+ if (N->getOperand(1).getOpcode() != ISD::UNDEF)
+ return false;
+
+ EVT VT = N->getValueType(0);
+ unsigned NumElems = VT.getVectorNumElements();
- bool HasHi = false;
- for (unsigned i = 2; i < 4; ++i) {
- int Elt = N->getMaskElt(i);
- if (Elt >= 0 && Elt != 3)
+ if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
+ (VT.getSizeInBits() == 256 && NumElems != 8))
+ return false;
+
+ // "i+1" is the value the indexed mask element must have
+ for (unsigned i = 0; i < NumElems; i += 2)
+ if (!isUndefOrEqual(N->getMaskElt(i), i+1) ||
+ !isUndefOrEqual(N->getMaskElt(i+1), i+1))
return false;
- if (Elt == 3)
- HasHi = true;
- }
- // Don't use movshdup if it can be done with a shufps.
- // FIXME: verify that matching u, u, 3, 3 is what we want.
- return HasHi;
+
+ return true;
}
/// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
/// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
-bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
- if (N->getValueType(0).getVectorNumElements() != 4)
+/// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
+bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N,
+ const X86Subtarget *Subtarget) {
+ if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
return false;
- // Expect 0, 0, 2, 2
- for (unsigned i = 0; i < 2; ++i)
- if (N->getMaskElt(i) > 0)
- return false;
+ // The second vector must be undef
+ if (N->getOperand(1).getOpcode() != ISD::UNDEF)
+ return false;
+
+ EVT VT = N->getValueType(0);
+ unsigned NumElems = VT.getVectorNumElements();
- bool HasHi = false;
- for (unsigned i = 2; i < 4; ++i) {
- int Elt = N->getMaskElt(i);
- if (Elt >= 0 && Elt != 2)
+ if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
+ (VT.getSizeInBits() == 256 && NumElems != 8))
+ return false;
+
+ // "i" is the value the indexed mask element must have
+ for (unsigned i = 0; i < NumElems; i += 2)
+ if (!isUndefOrEqual(N->getMaskElt(i), i) ||
+ !isUndefOrEqual(N->getMaskElt(i+1), i))
return false;
- if (Elt == 2)
- HasHi = true;
- }
- // Don't use movsldup if it can be done with a shufps.
- return HasHi;
+
+ return true;
}
/// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
break;
case X86ISD::UNPCKHPS:
case X86ISD::UNPCKHPD:
+ case X86ISD::VUNPCKHPSY:
+ case X86ISD::VUNPCKHPDY:
DecodeUNPCKHPMask(NumElems, ShuffleMask);
break;
case X86ISD::PUNPCKLBW:
return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
Depth+1);
}
- case X86ISD::VPERMIL:
+ case X86ISD::VPERMILPS:
+ case X86ISD::VPERMILPSY:
+ // FIXME: Implement the other types
ImmN = N->getOperand(N->getNumOperands()-1);
DecodeVPERMILMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
ShuffleMask);
LDBase->getPointerInfo(),
LDBase->isVolatile(), LDBase->isNonTemporal(),
LDBase->getAlignment());
- } else if (NumElems == 4 && LastLoadedElt == 1) {
+ } else if (NumElems == 4 && LastLoadedElt == 1 &&
+ DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
NumElems/2);
// Recreate the wider vector with the lower and upper part.
- SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
- DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
- return Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32),
+ SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Lower,
+ DAG.getConstant(0, MVT::i32), DAG, dl);
+ return Insert128BitVector(Vec, Upper, DAG.getConstant(NumElems/2, MVT::i32),
DAG, dl);
}
X86::getShuffleSHUFImmediate(SVOp), DAG);
}
-static inline unsigned getUNPCKLOpcode(EVT VT, const X86Subtarget *Subtarget) {
+static inline unsigned getUNPCKLOpcode(EVT VT) {
switch(VT.getSimpleVT().SimpleTy) {
case MVT::v4i32: return X86ISD::PUNPCKLDQ;
case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
case MVT::v4f32: return X86ISD::UNPCKHPS;
case MVT::v2f64: return X86ISD::UNPCKHPD;
+ case MVT::v8f32: return X86ISD::VUNPCKHPSY;
+ case MVT::v4f64: return X86ISD::VUNPCKHPDY;
case MVT::v16i8: return X86ISD::PUNPCKHBW;
case MVT::v8i16: return X86ISD::PUNPCKHWD;
default:
return 0;
}
+static inline unsigned getVPERMILOpcode(EVT VT) {
+ switch(VT.getSimpleVT().SimpleTy) {
+ case MVT::v4i32:
+ case MVT::v4f32: return X86ISD::VPERMILPS;
+ case MVT::v2i64:
+ case MVT::v2f64: return X86ISD::VPERMILPD;
+ case MVT::v8i32:
+ case MVT::v8f32: return X86ISD::VPERMILPSY;
+ case MVT::v4i64:
+ case MVT::v4f64: return X86ISD::VPERMILPDY;
+ default:
+ llvm_unreachable("Unknown type for vpermil");
+ }
+ return 0;
+}
+
static
SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
const TargetLowering &TLI,
// 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, getSubtarget()), dl, VT, V1, V1, DAG);
+ return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
if (X86::isMOVHLPSMask(SVOp))
return getMOVHighToLow(Op, dl, DAG);
- if (X86::isMOVSHDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
+ if (X86::isMOVSHDUPMask(SVOp, Subtarget))
return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
- if (X86::isMOVSLDUPMask(SVOp) && HasSSE3 && V2IsUndef && NumElems == 4)
+ if (X86::isMOVSLDUPMask(SVOp, Subtarget))
return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
if (X86::isMOVLPMask(SVOp))
}
if (X86::isUNPCKLMask(SVOp))
- return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
- dl, VT, V1, V2, DAG);
+ return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
if (X86::isUNPCKHMask(SVOp))
return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
if (X86::isUNPCKLMask(NewSVOp))
- return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
- dl, VT, V2, V1, DAG);
+ return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
if (X86::isUNPCKHMask(NewSVOp))
return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
}
if (X86::isUNPCKL_v_undef_Mask(SVOp))
- return getTargetShuffleNode(getUNPCKLOpcode(VT, getSubtarget()),
- dl, VT, V1, V1, DAG);
+ return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
if (X86::isUNPCKH_v_undef_Mask(SVOp))
return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
+ //===--------------------------------------------------------------------===//
+ // Generate target specific nodes for 128 or 256-bit shuffles only
+ // supported in the AVX instruction set.
+ //
+
+ // Handle VPERMILPS* permutations
+ if (isVPERMILPSMask(M, VT, Subtarget))
+ return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
+ getShuffleVPERMILPSImmediate(SVOp), DAG);
+
+ // Handle VPERMILPD* permutations
+ if (isVPERMILPDMask(M, VT, Subtarget))
+ return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
+ getShuffleVPERMILPDImmediate(SVOp), DAG);
+
+ //===--------------------------------------------------------------------===//
+ // Since no target specific shuffle was selected for this generic one,
+ // lower it into other known shuffles. FIXME: this isn't true yet, but
+ // this is the plan.
+ //
+
// Handle v8i16 specifically since SSE can do byte extraction and insertion.
if (VT == MVT::v8i16) {
SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
if (NumElems == 4 && VT.getSizeInBits() == 128)
return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
- //===--------------------------------------------------------------------===//
- // Custom lower or generate target specific nodes for 256-bit shuffles.
-
- // Handle VPERMIL permutations
- if (isVPERMILMask(M, VT)) {
- unsigned TargetMask = getShuffleVPERMILImmediate(SVOp);
- if (VT == MVT::v8f32)
- return getTargetShuffleNode(X86ISD::VPERMIL, dl, VT, V1, TargetMask, DAG);
- }
-
// Handle general 256-bit shuffles
if (VT.is256BitVector())
return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
return Op;
// SHUFPS the element to the lowest double word, then movss.
- int Mask[4] = { Idx, -1, -1, -1 };
+ int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
EVT VVT = Op.getOperand(0).getValueType();
SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
DAG.getUNDEF(VVT), Mask);
return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
}
+SDValue X86TargetLowering::LowerATOMIC_FENCE(SDValue Op,
+ SelectionDAG &DAG) const {
+ DebugLoc dl = Op.getDebugLoc();
+ AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
+ cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
+ SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
+ cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
+
+ // The only fence that needs an instruction is a sequentially-consistent
+ // cross-thread fence.
+ if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
+ // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
+ // no-sse2). There isn't any reason to disable it if the target processor
+ // supports it.
+ if (Subtarget->hasSSE2() || Subtarget->is64Bit())
+ return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
+
+ SDValue Chain = Op.getOperand(0);
+ SDValue Zero = DAG.getConstant(0, MVT::i32);
+ SDValue Ops[] = {
+ DAG.getRegister(X86::ESP, MVT::i32), // Base
+ DAG.getTargetConstant(1, MVT::i8), // Scale
+ DAG.getRegister(0, MVT::i32), // Index
+ DAG.getTargetConstant(0, MVT::i32), // Disp
+ DAG.getRegister(0, MVT::i32), // Segment.
+ Zero,
+ Chain
+ };
+ SDNode *Res =
+ DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
+ array_lengthof(Ops));
+ return SDValue(Res, 0);
+ }
+
+ // MEMBARRIER is a compiler barrier; it codegens to a no-op.
+ return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
+}
+
+
SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
EVT T = Op.getValueType();
DebugLoc DL = Op.getDebugLoc();
default: llvm_unreachable("Should not custom lower this!");
case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
+ case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op,DAG);
case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
- case X86ISD::VPERMIL: return "X86ISD::VPERMIL";
+ case X86ISD::VPERMILPS: return "X86ISD::VPERMILPS";
+ case X86ISD::VPERMILPSY: return "X86ISD::VPERMILPSY";
+ case X86ISD::VPERMILPD: return "X86ISD::VPERMILPD";
+ case X86ISD::VPERMILPDY: return "X86ISD::VPERMILPDY";
case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
+ case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
}
}
// (add Y, (setne X, 0)) -> sbb -1, Y
// (sub (sete X, 0), Y) -> sbb 0, Y
// (sub (setne X, 0), Y) -> adc -1, Y
-static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
+static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
DebugLoc DL = N->getDebugLoc();
// Look through ZExts.
DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
}
+static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
+ SDValue Op0 = N->getOperand(0);
+ SDValue Op1 = N->getOperand(1);
+
+ // X86 can't encode an immediate LHS of a sub. See if we can push the
+ // negation into a preceding instruction.
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
+ uint64_t Op0C = C->getSExtValue();
+
+ // If the RHS of the sub is a XOR with one use and a constant, invert the
+ // immediate. Then add one to the LHS of the sub so we can turn
+ // X-Y -> X+~Y+1, saving one register.
+ if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
+ isa<ConstantSDNode>(Op1.getOperand(1))) {
+ uint64_t XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getSExtValue();
+ EVT VT = Op0.getValueType();
+ SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
+ Op1.getOperand(0),
+ DAG.getConstant(~XorC, VT));
+ return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
+ DAG.getConstant(Op0C+1, VT));
+ }
+ }
+
+ return OptimizeConditionalInDecrement(N, DAG);
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
- case ISD::ADD:
- case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
+ case ISD::ADD: return OptimizeConditionalInDecrement(N, DAG);
+ case ISD::SUB: return PerformSubCombine(N, DAG);
case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
case ISD::SHL:
case X86ISD::PUNPCKHQDQ:
case X86ISD::UNPCKHPS:
case X86ISD::UNPCKHPD:
+ case X86ISD::VUNPCKHPSY:
+ case X86ISD::VUNPCKHPDY:
case X86ISD::PUNPCKLBW:
case X86ISD::PUNPCKLWD:
case X86ISD::PUNPCKLDQ:
case X86ISD::PSHUFLW:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
- case X86ISD::VPERMIL:
+ case X86ISD::VPERMILPS:
+ case X86ISD::VPERMILPSY:
+ case X86ISD::VPERMILPD:
+ case X86ISD::VPERMILPDY:
case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
}
projects / oota-llvm.git / blobdiff