// FIXME: In order to prevent SSE instructions being expanded to MMX ones
// with -msoft-float, disable use of MMX as well.
if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
+ addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass, false);
+
+ // FIXME: Remove the rest of this stuff.
addRegisterClass(MVT::v8i8, X86::VR64RegisterClass, false);
addRegisterClass(MVT::v4i16, X86::VR64RegisterClass, false);
addRegisterClass(MVT::v2i32, X86::VR64RegisterClass, false);
case X86ISD::MOVLPD:
case X86ISD::MOVSHDUP:
case X86ISD::MOVSLDUP:
+ case X86ISD::MOVDDUP:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
case X86ISD::UNPCKLPS:
default: llvm_unreachable("Unknown x86 shuffle node");
case X86ISD::MOVSHDUP:
case X86ISD::MOVSLDUP:
+ case X86ISD::MOVDDUP:
return DAG.getNode(Opc, dl, VT, V1);
}
/// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32.
static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
- if (SV->getValueType(0).getVectorNumElements() <= 4)
- return SDValue(SV, 0);
-
EVT PVT = MVT::v4f32;
EVT VT = SV->getValueType(0);
DebugLoc dl = SV->getDebugLoc();
}
/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
-/// ones, or rewriting v4i32 / v2i32 as 2 wide ones if possible. This can be
+/// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
/// done when every pair / quad of shuffle mask elements point to elements in
/// the right sequence. e.g.
-/// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
+/// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
static
SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
- SelectionDAG &DAG,
- const TargetLowering &TLI, DebugLoc dl) {
+ SelectionDAG &DAG, DebugLoc dl) {
EVT VT = SVOp->getValueType(0);
SDValue V1 = SVOp->getOperand(0);
SDValue V2 = SVOp->getOperand(1);
unsigned NumElems = VT.getVectorNumElements();
unsigned NewWidth = (NumElems == 4) ? 2 : 4;
- EVT MaskVT = (NewWidth == 4) ? MVT::v4i16 : MVT::v2i32;
- EVT NewVT = MaskVT;
+ EVT NewVT;
switch (VT.getSimpleVT().SimpleTy) {
default: assert(false && "Unexpected!");
case MVT::v4f32: NewVT = MVT::v2f64; break;
case MVT::v16i8: NewVT = MVT::v4i32; break;
}
- if (NewWidth == 2) {
- if (VT.isInteger())
- NewVT = MVT::v2i64;
- else
- NewVT = MVT::v2f64;
- }
int Scale = NumElems / NewWidth;
SmallVector<int, 8> MaskVec;
for (unsigned i = 0; i < NumElems; i += Scale) {
return false;
}
+// FIXME: the version above should always be used. Since there's
+// a bug where several vector shuffles can't be folded because the
+// DAG is not updated during lowering and a node claims to have two
+// uses while it only has one, use this version, and let isel match
+// another instruction if the load really happens to have more than
+// one use. Remove this version after this bug get fixed.
+static bool RelaxedMayFoldVectorLoad(SDValue V) {
+ if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
+ V = V.getOperand(0);
+ if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
+ V = V.getOperand(0);
+ if (ISD::isNormalLoad(V.getNode()))
+ return true;
+ return false;
+}
+
+/// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
+/// a vector extract, and if both can be later optimized into a single load.
+/// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
+/// here because otherwise a target specific shuffle node is going to be
+/// emitted for this shuffle, and the optimization not done.
+/// FIXME: This is probably not the best approach, but fix the problem
+/// until the right path is decided.
+static
+bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
+ const TargetLowering &TLI) {
+ EVT VT = V.getValueType();
+ ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
+
+ // Be sure that the vector shuffle is present in a pattern like this:
+ // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
+ if (!V.hasOneUse())
+ return false;
+
+ SDNode *N = *V.getNode()->use_begin();
+ if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+ return false;
+
+ SDValue EltNo = N->getOperand(1);
+ if (!isa<ConstantSDNode>(EltNo))
+ return false;
+
+ // If the bit convert changed the number of elements, it is unsafe
+ // to examine the mask.
+ bool HasShuffleIntoBitcast = false;
+ if (V.getOpcode() == ISD::BIT_CONVERT) {
+ EVT SrcVT = V.getOperand(0).getValueType();
+ if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
+ return false;
+ V = V.getOperand(0);
+ HasShuffleIntoBitcast = true;
+ }
+
+ // Select the input vector, guarding against out of range extract vector.
+ unsigned NumElems = VT.getVectorNumElements();
+ unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
+ int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
+ V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
+
+ // Skip one more bit_convert if necessary
+ if (V.getOpcode() == ISD::BIT_CONVERT)
+ V = V.getOperand(0);
+
+ if (ISD::isNormalLoad(V.getNode())) {
+ // Is the original load suitable?
+ LoadSDNode *LN0 = cast<LoadSDNode>(V);
+
+ // FIXME: avoid the multi-use bug that is preventing lots of
+ // of foldings to be detected, this is still wrong of course, but
+ // give the temporary desired behavior, and if it happens that
+ // the load has real more uses, during isel it will not fold, and
+ // will generate poor code.
+ if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
+ return false;
+
+ if (!HasShuffleIntoBitcast)
+ return true;
+
+ // If there's a bitcast before the shuffle, check if the load type and
+ // alignment is valid.
+ unsigned Align = LN0->getAlignment();
+ unsigned NewAlign =
+ TLI.getTargetData()->getABITypeAlignment(
+ VT.getTypeForEVT(*DAG.getContext()));
+
+ if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
+ return false;
+ }
+
+ return true;
+}
+
static
SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
bool HasSSE2) {
return 0;
}
-SDValue
-X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
+static
+SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
+ const TargetLowering &TLI,
+ const X86Subtarget *Subtarget) {
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
- SDValue V1 = Op.getOperand(0);
- SDValue V2 = Op.getOperand(1);
EVT VT = Op.getValueType();
DebugLoc dl = Op.getDebugLoc();
- unsigned NumElems = VT.getVectorNumElements();
- bool isMMX = VT.getSizeInBits() == 64;
- bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
- bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
- bool V1IsSplat = false;
- bool V2IsSplat = false;
- bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
- bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
- bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
- MachineFunction &MF = DAG.getMachineFunction();
- bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
+ SDValue V1 = Op.getOperand(0);
+ SDValue V2 = Op.getOperand(1);
if (isZeroShuffle(SVOp))
return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
- // Promote splats to v4f32.
+ // Handle splat operations
if (SVOp->isSplat()) {
- if (isMMX || NumElems < 4)
+ // Special case, this is the only place now where it's
+ // allowed to return a vector_shuffle operation without
+ // using a target specific node, because *hopefully* it
+ // will be optimized away by the dag combiner.
+ if (VT.getVectorNumElements() <= 4 &&
+ CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
return Op;
+
+ // Handle splats by matching through known masks
+ if (VT.getVectorNumElements() <= 4)
+ return SDValue();
+
+ // Canonize all of the remaining to v4f32.
return PromoteSplat(SVOp, DAG);
}
// If the shuffle can be profitably rewritten as a narrower shuffle, then
// do it!
if (VT == MVT::v8i16 || VT == MVT::v16i8) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
if (NewOp.getNode())
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
- LowerVECTOR_SHUFFLE(NewOp, DAG));
+ return DAG.getNode(ISD::BIT_CONVERT, dl, VT, NewOp);
} else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
// FIXME: Figure out a cleaner way to do this.
// Try to make use of movq to zero out the top part.
if (ISD::isBuildVectorAllZeros(V2.getNode())) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
if (NewOp.getNode()) {
if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
DAG, Subtarget, dl);
}
} else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
- SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
+ SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
DAG, Subtarget, dl);
}
}
+ return SDValue();
+}
+
+SDValue
+X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ SDValue V1 = Op.getOperand(0);
+ SDValue V2 = Op.getOperand(1);
+ EVT VT = Op.getValueType();
+ DebugLoc dl = Op.getDebugLoc();
+ unsigned NumElems = VT.getVectorNumElements();
+ bool isMMX = VT.getSizeInBits() == 64;
+ bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
+ bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
+ bool V1IsSplat = false;
+ bool V2IsSplat = false;
+ bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
+ bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
+ bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
+ MachineFunction &MF = DAG.getMachineFunction();
+ bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
+
+ // FIXME: this is somehow handled during isel by MMX pattern fragments. Remove
+ // the check or come up with another solution when all MMX move to intrinsics,
+ // but don't allow this to be considered legal, we don't want vector_shuffle
+ // operations to be matched during isel anymore.
+ if (isMMX && SVOp->isSplat())
+ return Op;
+
+ // Vector shuffle lowering takes 3 steps:
+ //
+ // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
+ // narrowing and commutation of operands should be handled.
+ // 2) Matching of shuffles with known shuffle masks to x86 target specific
+ // shuffle nodes.
+ // 3) Rewriting of unmatched masks into new generic shuffle operations,
+ // so the shuffle can be broken into other shuffles and the legalizer can
+ // try the lowering again.
+ //
+ // The general ideia is that no vector_shuffle operation should be left to
+ // be matched during isel, all of them must be converted to a target specific
+ // node here.
+
+ // Normalize the input vectors. Here splats, zeroed vectors, profitable
+ // narrowing and commutation of operands should be handled. The actual code
+ // doesn't include all of those, work in progress...
+ SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
+ if (NewOp.getNode())
+ return NewOp;
// NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
// unpckh_undef). Only use pshufd if speed is more important than size.
if (VT != MVT::v2i64 && VT != MVT::v2f64)
return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
+ if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
+ RelaxedMayFoldVectorLoad(V1) && !isMMX)
+ return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
+
+ if (!isMMX && X86::isMOVHLPS_v_undef_Mask(SVOp))
+ return getMOVHighToLow(Op, dl, DAG);
+
+ // Use to match splats
+ if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
+ (VT == MVT::v2f64 || VT == MVT::v2i64))
+ return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
+
if (X86::isPSHUFDMask(SVOp)) {
// The actual implementation will match the mask in the if above and then
// during isel it can match several different instructions, not only pshufd
X86::getShufflePALIGNRImmediate(SVOp),
DAG);
- // MMX shuffles not already handled must be expanded.
- if (VT.getSizeInBits() == 64)
+ // Only a few shuffle masks are handled for 64-bit vectors (MMX), and
+ // 64-bit vectors which made to this point can't be handled, they are
+ // expanded.
+ if (isMMX)
return SDValue();
- // FIXME: pshufb, blends, shifts.
- if (VT.getVectorNumElements() == 2)
- return Op;
-
if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
SVOp->getSplatIndex() == 0 && V2IsUndef) {
if (VT == MVT::v2f64)
int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
MachineMemOperand *MMO =
F->getMachineMemOperand(
- PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
- MachineMemOperand::MOStore, Offset,
+ MachinePointerInfo(PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
+ Offset),
+ MachineMemOperand::MOStore,
/*Size=*/16, /*Align=*/16);
BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
.addFrameIndex(RegSaveFrameIndex)
return TargetLowering::getConstraintType(Constraint);
}
+/// Examine constraint type and operand type and determine a weight value,
+/// where: -1 = invalid match, and 0 = so-so match to 3 = good match.
+/// This object must already have been set up with the operand type
+/// and the current alternative constraint selected.
+int X86TargetLowering::getSingleConstraintMatchWeight(
+ AsmOperandInfo &info, const char *constraint) const {
+ int weight = -1;
+ Value *CallOperandVal = info.CallOperandVal;
+ // If we don't have a value, we can't do a match,
+ // but allow it at the lowest weight.
+ if (CallOperandVal == NULL)
+ return 0;
+ // Look at the constraint type.
+ switch (*constraint) {
+ default:
+ return TargetLowering::getSingleConstraintMatchWeight(info, constraint);
+ break;
+ case 'I':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
+ if (C->getZExtValue() <= 31)
+ weight = 3;
+ }
+ break;
+ // etc.
+ }
+ return weight;
+}
+
/// LowerXConstraint - try to replace an X constraint, which matches anything,
/// with another that has more specific requirements based on the type of the
/// corresponding operand.