summaryrefslogtreecommitdiffstats
path: root/llvm/lib
diff options
context:
space:
mode:
Diffstat (limited to 'llvm/lib')
-rw-r--r--llvm/lib/Target/AArch64/AArch64ISelLowering.cpp118
-rw-r--r--llvm/lib/Target/AArch64/AArch64ISelLowering.h6
2 files changed, 95 insertions, 29 deletions
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 27277c47f39..327547c27cc 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -4153,22 +4153,76 @@ AArch64TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
return false;
}
+// Check whether a shuffle_vector could be presented as concat_vector.
+bool AArch64TargetLowering::isConcatVector(SDValue Op, SelectionDAG &DAG,
+ SDValue V0, SDValue V1,
+ const int *Mask,
+ SDValue &Res) const {
+ SDLoc DL(Op);
+ EVT VT = Op.getValueType();
+ if (VT.getSizeInBits() != 128)
+ return false;
+ if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
+ VT.getVectorElementType() != V1.getValueType().getVectorElementType())
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
+ bool isContactVector = true;
+ bool splitV0 = false;
+ if (V0.getValueType().getSizeInBits() == 128)
+ splitV0 = true;
+
+ for (int I = 0, E = NumElts / 2; I != E; I++) {
+ if (Mask[I] != I) {
+ isContactVector = false;
+ break;
+ }
+ }
+
+ if (isContactVector) {
+ int offset = NumElts / 2;
+ for (int I = NumElts / 2, E = NumElts; I != E; I++) {
+ if (Mask[I] != I + splitV0 * offset) {
+ isContactVector = false;
+ break;
+ }
+ }
+ }
+
+ if (isContactVector) {
+ EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
+ NumElts / 2);
+ if (splitV0) {
+ V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
+ DAG.getConstant(0, MVT::i64));
+ }
+ if (V1.getValueType().getSizeInBits() == 128) {
+ V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
+ DAG.getConstant(0, MVT::i64));
+ }
+ Res = DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
+ return true;
+ }
+ return false;
+}
-// Check whether a Build Vector could be presented as Shuffle Vector. If yes,
-// try to call LowerVECTOR_SHUFFLE to lower it.
+// Check whether a Build Vector could be presented as Shuffle Vector.
+// This Shuffle Vector maybe not legalized, so the length of its operand and
+// the length of result may not equal.
bool AArch64TargetLowering::isKnownShuffleVector(SDValue Op, SelectionDAG &DAG,
- SDValue &Res) const {
+ SDValue &V0, SDValue &V1,
+ int *Mask) const {
SDLoc DL(Op);
EVT VT = Op.getValueType();
unsigned NumElts = VT.getVectorNumElements();
unsigned V0NumElts = 0;
- int Mask[16];
- SDValue V0, V1;
// Check if all elements are extracted from less than 3 vectors.
for (unsigned i = 0; i < NumElts; ++i) {
SDValue Elt = Op.getOperand(i);
- if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT)
+ if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
+ Elt.getOperand(0).getValueType().getVectorElementType() !=
+ VT.getVectorElementType())
return false;
if (V0.getNode() == 0) {
@@ -4189,25 +4243,7 @@ bool AArch64TargetLowering::isKnownShuffleVector(SDValue Op, SelectionDAG &DAG,
return false;
}
}
-
- if (!V1.getNode() && V0NumElts == NumElts * 2) {
- V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
- DAG.getConstant(NumElts, MVT::i64));
- V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
- DAG.getConstant(0, MVT::i64));
- V0NumElts = V0.getValueType().getVectorNumElements();
- }
-
- if (V1.getNode() && NumElts == V0NumElts &&
- V0NumElts == V1.getValueType().getVectorNumElements()) {
- SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
- if(Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
- Res = Shuffle;
- else
- Res = LowerVECTOR_SHUFFLE(Shuffle, DAG);
- return true;
- } else
- return false;
+ return true;
}
// If this is a case we can't handle, return null and let the default
@@ -4413,9 +4449,31 @@ AArch64TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
return SDValue();
// Try to lower this in lowering ShuffleVector way.
- SDValue Shuf;
- if (isKnownShuffleVector(Op, DAG, Shuf))
- return Shuf;
+ SDValue V0, V1;
+ int Mask[16];
+ if (isKnownShuffleVector(Op, DAG, V0, V1, Mask)) {
+ unsigned V0NumElts = V0.getValueType().getVectorNumElements();
+ if (!V1.getNode() && V0NumElts == NumElts * 2) {
+ V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
+ DAG.getConstant(NumElts, MVT::i64));
+ V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V0,
+ DAG.getConstant(0, MVT::i64));
+ V0NumElts = V0.getValueType().getVectorNumElements();
+ }
+
+ if (V1.getNode() && NumElts == V0NumElts &&
+ V0NumElts == V1.getValueType().getVectorNumElements()) {
+ SDValue Shuffle = DAG.getVectorShuffle(VT, DL, V0, V1, Mask);
+ if (Shuffle.getOpcode() != ISD::VECTOR_SHUFFLE)
+ return Shuffle;
+ else
+ return LowerVECTOR_SHUFFLE(Shuffle, DAG);
+ } else {
+ SDValue Res;
+ if (isConcatVector(Op, DAG, V0, V1, Mask, Res))
+ return Res;
+ }
+ }
// If all else fails, just use a sequence of INSERT_VECTOR_ELT when we
// know the default expansion would otherwise fall back on something even
@@ -4601,6 +4659,10 @@ AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
return DAG.getNode(ISDNo, dl, VT, V1, V2);
}
+ SDValue Res;
+ if (isConcatVector(Op, DAG, V1, V2, &ShuffleMask[0], Res))
+ return Res;
+
// If the element of shuffle mask are all the same constant, we can
// transform it into either NEON_VDUP or NEON_VDUPLANE
if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.h b/llvm/lib/Target/AArch64/AArch64ISelLowering.h
index 3879663e570..149de557d08 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.h
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.h
@@ -232,7 +232,11 @@ public:
SDLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) const;
- bool isKnownShuffleVector(SDValue Op, SelectionDAG &DAG, SDValue &Res) const;
+ bool isConcatVector(SDValue Op, SelectionDAG &DAG, SDValue V0, SDValue V1,
+ const int *Mask, SDValue &Res) const;
+
+ bool isKnownShuffleVector(SDValue Op, SelectionDAG &DAG, SDValue &V0,
+ SDValue &V1, int *Mask) const;
SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
const AArch64Subtarget *ST) const;
OpenPOWER on IntegriCloud