ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
- assert(Subtarget->hasDQI() && "We can only lower v8i64 with AVX-512-DQI");
// FIXME: Implement direct support for this type!
return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
ArrayRef<int> Mask = SVOp->getMask();
assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
- assert(Subtarget->hasDQI() && "We can only lower v16i32 with AVX-512-DQI!");
// FIXME: Implement direct support for this type!
return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
assert(Subtarget->hasAVX512() &&
"Cannot lower 512-bit vectors w/ basic ISA!");
+ // Check for being able to broadcast a single element.
+ if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(VT.SimpleTy, DL, V1,
+ Mask, Subtarget, DAG))
+ return Broadcast;
+
// Dispatch to each element type for lowering. If we don't have supprot for
// specific element type shuffles at 512 bits, immediately split them and
// lower them. Each lowering routine of a given type is allowed to assume that
case MVT::v16f32:
return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
case MVT::v8i64:
- if (Subtarget->hasDQI())
- return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
- break;
+ return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
case MVT::v16i32:
- if (Subtarget->hasDQI())
- return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
- break;
+ return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
case MVT::v32i16:
if (Subtarget->hasBWI())
return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
EVT VT = Op.getValueType();
// SSE1 has rsqrtss and rsqrtps.
- // TODO: Add support for AVX (v8f32) and AVX512 (v16f32).
+ // TODO: Add support for AVX512 (v16f32).
// It is likely not profitable to do this for f64 because a double-precision
// rsqrt estimate with refinement on x86 prior to FMA requires at least 16
// instructions: convert to single, rsqrtss, convert back to double, refine
// (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
// along with FMA, this could be a throughput win.
- if (Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) {
+ if ((Subtarget->hasSSE1() && (VT == MVT::f32 || VT == MVT::v4f32)) ||
+ (Subtarget->hasAVX() && VT == MVT::v8f32)) {
RefinementSteps = 1;
UseOneConstNR = false;
return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
}
-/// \brief Return (and \p Op, \p Mask) for compare instructions or
-/// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
+/// \brief Return (vselect \p Mask, \p Op, \p PreservedSrc) along with the
/// necessary casting for \p Mask when lowering masking intrinsics.
static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
SDValue PreservedSrc, SelectionDAG &DAG) {
SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
DAG.getNode(ISD::BITCAST, dl, BitcastVT, Mask),
DAG.getIntPtrConstant(0));
-
- switch (Op.getOpcode()) {
- default: break;
- case X86ISD::PCMPEQM:
- case X86ISD::PCMPGTM:
- case X86ISD::CMPM:
- case X86ISD::CMPMU:
- return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
- }
-
return DAG.getNode(ISD::VSELECT, dl, VT, VMask, Op, PreservedSrc);
}
// (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
// (i8 (bitcast
// (v8i1 (insert_subvector undef,
- // (v2i1 (and (PCMPEQM %a, %b),
- // (extract_subvector
- // (v8i1 (bitcast %mask)), 0))), 0))))
+ // (v2i1 (vselect (extract_subvector
+ // (v8i1 (bitcast %mask)), 0),
+ // (PCMPEQM %a, %b), 0))))))
EVT VT = Op.getOperand(1).getValueType();
EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
VT.getVectorNumElements());
/// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
/// specific shuffle of a load can be folded into a single element load.
/// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
-/// shuffles have been customed lowered so we need to handle those here.
+/// shuffles have been custom lowered so we need to handle those here.
static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI) {
if (DCI.isBeforeLegalizeOps())
if (!isa<ConstantSDNode>(EltNo))
return SDValue();
- EVT VT = InVec.getValueType();
+ EVT OriginalVT = InVec.getValueType();
if (InVec.getOpcode() == ISD::BITCAST) {
// Don't duplicate a load with other uses.
if (!InVec.hasOneUse())
return SDValue();
EVT BCVT = InVec.getOperand(0).getValueType();
- if (BCVT.getVectorNumElements() != VT.getVectorNumElements())
+ if (BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
return SDValue();
InVec = InVec.getOperand(0);
}
+ EVT CurrentVT = InVec.getValueType();
+
if (!isTargetShuffle(InVec.getOpcode()))
return SDValue();
SmallVector<int, 16> ShuffleMask;
bool UnaryShuffle;
- if (!getTargetShuffleMask(InVec.getNode(), VT.getSimpleVT(), ShuffleMask,
- UnaryShuffle))
+ if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
+ ShuffleMask, UnaryShuffle))
return SDValue();
// Select the input vector, guarding against out of range extract vector.
- unsigned NumElems = VT.getVectorNumElements();
+ unsigned NumElems = CurrentVT.getVectorNumElements();
int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
SDLoc dl(N);
// Create shuffle node taking into account the case that its a unary shuffle
- SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(VT) : InVec.getOperand(1);
- Shuffle = DAG.getVectorShuffle(InVec.getValueType(), dl,
+ SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
+ : InVec.getOperand(1);
+ Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
InVec.getOperand(0), Shuffle,
&ShuffleMask[0]);
- Shuffle = DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
+ Shuffle = DAG.getNode(ISD::BITCAST, dl, OriginalVT, Shuffle);
return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
EltNo);
}
projects / oota-llvm.git / blobdiff