Added support for the following definition of shufflevector
<result> = shufflevector <n x <ty>> <v1>, <n x <ty>> <v2>, <m x i32> <mask>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@58964 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index f8d92cb..0b6809b 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -4997,6 +4997,12 @@
SDValue ShufMask = N->getOperand(2);
unsigned NumElts = ShufMask.getNumOperands();
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+
+ assert(N0.getValueType().getVectorNumElements() == NumElts &&
+ "Vector shuffle must be normalized in DAG");
+
// If the shuffle mask is an identity operation on the LHS, return the LHS.
bool isIdentity = true;
for (unsigned i = 0; i != NumElts; ++i) {
@@ -5043,8 +5049,6 @@
}
}
- SDValue N0 = N->getOperand(0);
- SDValue N1 = N->getOperand(1);
// Normalize unary shuffle so the RHS is undef.
if (isUnary && VecNum == 1)
std::swap(N0, N1);
diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 84c6fb8..5ac7a80 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -3099,7 +3099,7 @@
}
Result = DAG.UpdateNodeOperands(Result, Tmp1, Tmp2);
-
+
switch (TLI.getOperationAction(Node->getOpcode(), Node->getValueType(0))) {
default: assert(0 && "BinOp legalize operation not supported");
case TargetLowering::Legal: break;
@@ -7210,16 +7210,34 @@
Lo = Node->getOperand(0);
Hi = Node->getOperand(1);
} else {
- SmallVector<SDValue, 8> LoOps(Node->op_begin(),
- Node->op_begin()+NewNumSubvectors);
+ SmallVector<SDValue, 8> LoOps(Node->op_begin(),
+ Node->op_begin()+NewNumSubvectors);
Lo = DAG.getNode(ISD::CONCAT_VECTORS, NewVT_Lo, &LoOps[0], LoOps.size());
- SmallVector<SDValue, 8> HiOps(Node->op_begin()+NewNumSubvectors,
+ SmallVector<SDValue, 8> HiOps(Node->op_begin()+NewNumSubvectors,
Node->op_end());
Hi = DAG.getNode(ISD::CONCAT_VECTORS, NewVT_Hi, &HiOps[0], HiOps.size());
}
break;
}
+ case ISD::EXTRACT_SUBVECTOR: {
+ SDValue Vec = Op.getOperand(0);
+ SDValue Idx = Op.getOperand(1);
+ MVT IdxVT = Idx.getValueType();
+
+ Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, NewVT_Lo, Vec, Idx);
+ ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Idx);
+ if (CIdx) {
+ Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, NewVT_Hi, Vec,
+ DAG.getConstant(CIdx->getZExtValue() + NewNumElts_Lo,
+ IdxVT));
+ } else {
+ Idx = DAG.getNode(ISD::ADD, IdxVT, Idx,
+ DAG.getConstant(NewNumElts_Lo, IdxVT));
+ Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, NewVT_Hi, Vec, Idx);
+ }
+ break;
+ }
case ISD::SELECT: {
SDValue Cond = Node->getOperand(0);
@@ -7517,7 +7535,7 @@
}
case ISD::EXTRACT_SUBVECTOR:
Result = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, NewVT, Node->getOperand(0),
- Node->getOperand(1));
+ Node->getOperand(1));
break;
case ISD::BIT_CONVERT: {
SDValue Op0 = Op.getOperand(0);
diff --git a/lib/CodeGen/SelectionDAG/LegalizeTypes.h b/lib/CodeGen/SelectionDAG/LegalizeTypes.h
index 1d1ed34..cb89eff 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeTypes.h
+++ b/lib/CodeGen/SelectionDAG/LegalizeTypes.h
@@ -476,6 +476,7 @@
void SplitVecRes_BUILD_PAIR(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_BUILD_VECTOR(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_CONCAT_VECTORS(SDNode *N, SDValue &Lo, SDValue &Hi);
+ void SplitVecRes_EXTRACT_SUBVECTOR(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_FPOWI(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_INSERT_VECTOR_ELT(SDNode *N, SDValue &Lo, SDValue &Hi);
void SplitVecRes_LOAD(LoadSDNode *N, SDValue &Lo, SDValue &Hi);
diff --git a/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 2378798..c155990 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -336,10 +336,11 @@
case ISD::SELECT_CC: SplitRes_SELECT_CC(N, Lo, Hi); break;
case ISD::UNDEF: SplitRes_UNDEF(N, Lo, Hi); break;
- case ISD::BIT_CONVERT: SplitVecRes_BIT_CONVERT(N, Lo, Hi); break;
- case ISD::BUILD_VECTOR: SplitVecRes_BUILD_VECTOR(N, Lo, Hi); break;
- case ISD::CONCAT_VECTORS: SplitVecRes_CONCAT_VECTORS(N, Lo, Hi); break;
- case ISD::FPOWI: SplitVecRes_FPOWI(N, Lo, Hi); break;
+ case ISD::BIT_CONVERT: SplitVecRes_BIT_CONVERT(N, Lo, Hi); break;
+ case ISD::BUILD_VECTOR: SplitVecRes_BUILD_VECTOR(N, Lo, Hi); break;
+ case ISD::CONCAT_VECTORS: SplitVecRes_CONCAT_VECTORS(N, Lo, Hi); break;
+ case ISD::EXTRACT_SUBVECTOR: SplitVecRes_EXTRACT_SUBVECTOR(N, Lo, Hi); break;
+ case ISD::FPOWI: SplitVecRes_FPOWI(N, Lo, Hi); break;
case ISD::INSERT_VECTOR_ELT: SplitVecRes_INSERT_VECTOR_ELT(N, Lo, Hi); break;
case ISD::LOAD: SplitVecRes_LOAD(cast<LoadSDNode>(N), Lo, Hi);break;
case ISD::VECTOR_SHUFFLE: SplitVecRes_VECTOR_SHUFFLE(N, Lo, Hi); break;
@@ -486,6 +487,32 @@
Hi = DAG.getNode(ISD::CONCAT_VECTORS, HiVT, &HiOps[0], HiOps.size());
}
+void DAGTypeLegalizer::SplitVecRes_EXTRACT_SUBVECTOR(SDNode *N, SDValue &Lo,
+ SDValue &Hi) {
+ MVT LoVT, HiVT;
+ GetSplitDestVTs(N->getValueType(0), LoVT, HiVT);
+ unsigned LoNumElts = LoVT.getVectorNumElements();
+
+ SDValue Vec = N->getOperand(0);
+ SDValue Idx = N->getOperand(1);
+ MVT IdxVT = Idx.getValueType();
+ Lo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, LoVT, Vec, Idx);
+
+ ConstantSDNode *CIdx = dyn_cast<ConstantSDNode>(Idx);
+ if (CIdx) {
+ unsigned IdxVal = CIdx->getZExtValue();
+ assert (IdxVal % LoVT.getVectorNumElements() == 0 &&
+ (IdxVal+LoNumElts) % HiVT.getVectorNumElements()==0 &&
+ "Index must be a multiple of the result type");
+ Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, HiVT, Vec,
+ DAG.getConstant(IdxVal + LoNumElts, IdxVT));
+ } else {
+ assert(LoVT == HiVT && "Low and High value type should be the same");
+ Idx = DAG.getNode(ISD::ADD, IdxVT, Idx, DAG.getConstant(LoNumElts, IdxVT));
+ Hi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, HiVT, Vec, Idx);
+ }
+}
+
void DAGTypeLegalizer::SplitVecRes_FPOWI(SDNode *N, SDValue &Lo,
SDValue &Hi) {
GetSplitVector(N->getOperand(0), Lo, Hi);
@@ -631,14 +658,19 @@
Ops.clear();
for (unsigned i = LoNumElts; i != NumElements; ++i) {
- unsigned Idx = cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
- SDValue InVec = N->getOperand(0);
- if (Idx >= NumElements) {
- InVec = N->getOperand(1);
- Idx -= NumElements;
+ SDValue Arg = Mask.getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) {
+ Ops.push_back(DAG.getNode(ISD::UNDEF, EltVT));
+ } else {
+ unsigned Idx = cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
+ SDValue InVec = N->getOperand(0);
+ if (Idx >= NumElements) {
+ InVec = N->getOperand(1);
+ Idx -= NumElements;
+ }
+ Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, EltVT, InVec,
+ DAG.getIntPtrConstant(Idx)));
}
- Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, EltVT, InVec,
- DAG.getIntPtrConstant(Idx)));
}
Hi = DAG.getNode(ISD::BUILD_VECTOR, HiVT, &Ops[0], Ops.size());
}
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 3601b6e..fb8630f 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -2697,7 +2697,8 @@
}
break;
case ISD::VECTOR_SHUFFLE:
- assert(VT == N1.getValueType() && VT == N2.getValueType() &&
+ assert(N1.getValueType() == N2.getValueType() &&
+ N1.getValueType().isVector() &&
VT.isVector() && N3.getValueType().isVector() &&
N3.getOpcode() == ISD::BUILD_VECTOR &&
VT.getVectorNumElements() == N3.getNumOperands() &&
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp
index 3e19b71..032e8bf 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuild.cpp
@@ -2288,14 +2288,180 @@
TLI.getValueType(I.getType()), InVec, InIdx));
}
+
+// Utility for visitShuffleVector - Returns true if the mask is mask starting
+// from SIndx and increasing to the element length (undefs are allowed).
+static bool SequentialMask(SDValue Mask, unsigned SIndx) {
+ unsigned NumElems = Mask.getNumOperands();
+ for (unsigned i = 0; i != NumElems; ++i) {
+ if (Mask.getOperand(i).getOpcode() != ISD::UNDEF) {
+ unsigned Idx = cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
+ if (Idx != i + SIndx)
+ return false;
+ }
+ }
+ return true;
+}
+
void SelectionDAGLowering::visitShuffleVector(User &I) {
SDValue V1 = getValue(I.getOperand(0));
SDValue V2 = getValue(I.getOperand(1));
SDValue Mask = getValue(I.getOperand(2));
- setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE,
- TLI.getValueType(I.getType()),
- V1, V2, Mask));
+ MVT VT = TLI.getValueType(I.getType());
+ MVT VT1 = V1.getValueType();
+ unsigned MaskNumElts = Mask.getNumOperands();
+ unsigned Src1NumElts = VT1.getVectorNumElements();
+
+ if (Src1NumElts == MaskNumElts) {
+ setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask));
+ return;
+ }
+
+ // Normalize the shuffle vector since mask and vector length don't match.
+ if (Src1NumElts < MaskNumElts && MaskNumElts % Src1NumElts == 0) {
+ // We can concat vectors to make the mask and input vector match.
+ if (Src1NumElts*2 == MaskNumElts && SequentialMask(Mask, 0)) {
+ // The shuffle is concatenating two vectors.
+ setValue(&I, DAG.getNode(ISD::CONCAT_VECTORS, VT, V1, V2));
+ return;
+ }
+
+ // Pad both vectors with undefs to the same size as the mask.
+ unsigned NumConcat = MaskNumElts / Src1NumElts;
+ std::vector<SDValue> UnOps(Src1NumElts,
+ DAG.getNode(ISD::UNDEF,
+ VT1.getVectorElementType()));
+ SDValue UndefVal = DAG.getNode(ISD::BUILD_VECTOR, VT1,
+ &UnOps[0], UnOps.size());
+
+ SmallVector<SDValue, 8> MOps1, MOps2;
+ MOps1.push_back(V1);
+ MOps2.push_back(V2);
+ for (unsigned i = 1; i != NumConcat; ++i) {
+ MOps1.push_back(UndefVal);
+ MOps2.push_back(UndefVal);
+ }
+ V1 = DAG.getNode(ISD::CONCAT_VECTORS, VT, &MOps1[0], MOps1.size());
+ V2 = DAG.getNode(ISD::CONCAT_VECTORS, VT, &MOps2[0], MOps2.size());
+
+ // Readjust mask for new input vector length.
+ SmallVector<SDValue, 8> MappedOps;
+ for (unsigned i = 0; i != MaskNumElts; ++i) {
+ if (Mask.getOperand(i).getOpcode() == ISD::UNDEF) {
+ MappedOps.push_back(Mask.getOperand(i));
+ } else {
+ unsigned Idx = cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
+ if (Idx < Src1NumElts) {
+ MappedOps.push_back(DAG.getConstant(Idx,
+ Mask.getOperand(i).getValueType()));
+ } else {
+ MappedOps.push_back(DAG.getConstant(Idx + MaskNumElts - Src1NumElts,
+ Mask.getOperand(i).getValueType()));
+ }
+ }
+ }
+ Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
+ &MappedOps[0], MappedOps.size());
+
+ setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask));
+ return;
+ }
+
+ if (Src1NumElts > MaskNumElts) {
+ // Resulting vector is shorter than the incoming vector.
+ if (Src1NumElts == MaskNumElts && SequentialMask(Mask,0)) {
+ // Shuffle extracts 1st vector.
+ setValue(&I, V1);
+ return;
+ }
+
+ if (Src1NumElts == MaskNumElts && SequentialMask(Mask,MaskNumElts)) {
+ // Shuffle extracts 2nd vector.
+ setValue(&I, V2);
+ return;
+ }
+
+ // Analyze the access pattern of the vector to see if we can extract each
+ // subvector and then do the shuffle. The analysis is done by calculating
+ // the range of elements the mask access on both vectors. If it is useful,
+ // we could do better by considering separate what elements are accessed
+ // in each vector (i.e., have min/max for each vector).
+ int MinRange = Src1NumElts+1;
+ int MaxRange = -1;
+ for (unsigned i = 0; i != MaskNumElts; ++i) {
+ SDValue Arg = Mask.getOperand(i);
+ if (Arg.getOpcode() != ISD::UNDEF) {
+ assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
+ int Idx = cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
+ if (Idx > (int) Src1NumElts)
+ Idx -= Src1NumElts;
+ if (Idx > MaxRange)
+ MaxRange = Idx;
+ if (Idx < MinRange)
+ MinRange = Idx;
+ }
+ }
+ // Adjust MinRange to start at an even boundary since this give us
+ // better quality splits later.
+ if ((unsigned) MinRange < Src1NumElts && MinRange%2 != 0)
+ MinRange = MinRange - 1;
+ if (MaxRange - MinRange < (int) MaskNumElts) {
+ // Extract subvector because the range is less than the new vector length
+ unsigned StartIdx = (MinRange/MaskNumElts)*MaskNumElts;
+ if (MaxRange - StartIdx < MaskNumElts) {
+ V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, VT, V1,
+ DAG.getIntPtrConstant(MinRange));
+ V2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, VT, V2,
+ DAG.getIntPtrConstant(MinRange));
+ // Readjust mask for new input vector length.
+ SmallVector<SDValue, 8> MappedOps;
+ for (unsigned i = 0; i != MaskNumElts; ++i) {
+ if (Mask.getOperand(i).getOpcode() == ISD::UNDEF) {
+ MappedOps.push_back(Mask.getOperand(i));
+ } else {
+ unsigned Idx =
+ cast<ConstantSDNode>(Mask.getOperand(i))->getZExtValue();
+ if (Idx < Src1NumElts) {
+ MappedOps.push_back(DAG.getConstant(Idx - StartIdx,
+ Mask.getOperand(i).getValueType()));
+ } else {
+ Idx = Idx - Src1NumElts - StartIdx + MaskNumElts;
+ MappedOps.push_back(DAG.getConstant(Idx,
+ Mask.getOperand(i).getValueType()));
+ }
+ }
+ }
+ Mask = DAG.getNode(ISD::BUILD_VECTOR, Mask.getValueType(),
+ &MappedOps[0], MappedOps.size());
+
+ setValue(&I, DAG.getNode(ISD::VECTOR_SHUFFLE, VT, V1, V2, Mask));
+ return;
+ }
+ }
+ }
+
+ // We can't use either concat vectors or extract subvectors so we fall back
+ // to insert and extracts.
+ MVT EltVT = VT.getVectorElementType();
+ MVT PtrVT = TLI.getPointerTy();
+ SmallVector<SDValue,8> Ops;
+ for (unsigned i = 0; i != MaskNumElts; ++i) {
+ SDValue Arg = Mask.getOperand(i);
+ if (Arg.getOpcode() == ISD::UNDEF) {
+ Ops.push_back(DAG.getNode(ISD::UNDEF, EltVT));
+ } else {
+ assert(isa<ConstantSDNode>(Arg) && "Invalid VECTOR_SHUFFLE mask!");
+ unsigned Idx = cast<ConstantSDNode>(Arg)->getZExtValue();
+ if (Idx < Src1NumElts)
+ Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, EltVT, V1,
+ DAG.getConstant(Idx, PtrVT)));
+ else
+ Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, EltVT, V2,
+ DAG.getConstant(Idx - Src1NumElts, PtrVT)));
+ }
+ }
+ setValue(&I, DAG.getNode(ISD::BUILD_VECTOR, VT, &Ops[0], Ops.size()));
}
void SelectionDAGLowering::visitInsertValue(InsertValueInst &I) {