summaryrefslogtreecommitdiffstats
path: root/llvm/lib/CodeGen
diff options
context:
space:
mode:
authorSimon Pilgrim <llvm-dev@redking.me.uk>2017-02-17 15:14:48 +0000
committerSimon Pilgrim <llvm-dev@redking.me.uk>2017-02-17 15:14:48 +0000
commit511d788a95750ad4f534bb0cdccf7a38526803bf (patch)
tree2aca2eac9c92e71b92b76aa25aaebadc751c4a2a /llvm/lib/CodeGen
parent7b35b781603b5d7d5be866a47d84fab8d96729ab (diff)
downloadbcm5719-llvm-511d788a95750ad4f534bb0cdccf7a38526803bf.tar.gz
bcm5719-llvm-511d788a95750ad4f534bb0cdccf7a38526803bf.zip
[DAGCombine] Recognise any_extend_vector_inreg and truncation style shuffle masks
During legalization we are often creating shuffles (via a build_vector scalarization stage) that are "any_extend_vector_inreg" style masks, and also other masks that are the equivalent of "truncate_vector_inreg" (if we had such a thing). This patch is an attempt to match these cases to help undo the effects of just leaving shuffle lowering to handle it - which typically means we lose track of the undefined elements of the shuffles resulting in an unnecessary extension+truncation stage for widened illegal types. The 2011-10-21-widen-cmp.ll regression will be fixed by making SIGN_EXTEND_VECTOR_IN_REG legal in SSE instead of lowering them to X86ISD::VSEXT (PR31712). Differential Revision: https://reviews.llvm.org/D29454 llvm-svn: 295451
Diffstat (limited to 'llvm/lib/CodeGen')
-rw-r--r--llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp125
1 files changed, 125 insertions, 0 deletions
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 7997a333862..9aa9b310d49 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -7530,6 +7530,16 @@ SDValue DAGCombiner::visitSIGN_EXTEND_INREG(SDNode *N) {
return DAG.getNode(ISD::SIGN_EXTEND, SDLoc(N), VT, N00, N1);
}
+ // fold (sext_in_reg (*_extend_vector_inreg x)) -> (sext_vector_in_reg x)
+ if ((N0.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG ||
+ N0.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG ||
+ N0.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) &&
+ N0.getOperand(0).getScalarValueSizeInBits() == EVTBits) {
+ if (!LegalOperations ||
+ TLI.isOperationLegal(ISD::SIGN_EXTEND_VECTOR_INREG, VT))
+ return DAG.getSignExtendVectorInReg(N0.getOperand(0), SDLoc(N), VT);
+ }
+
// fold (sext_in_reg (zext x)) -> (sext x)
// iff we are extending the source sign bit.
if (N0.getOpcode() == ISD::ZERO_EXTEND) {
@@ -14194,6 +14204,113 @@ static SDValue combineShuffleOfScalars(ShuffleVectorSDNode *SVN,
return DAG.getBuildVector(VT, SDLoc(SVN), Ops);
}
+// Match shuffles that can be converted to any_vector_extend_in_reg.
+// This is often generated during legalization.
+// e.g. v4i32 <0,u,1,u> -> (v2i64 any_vector_extend_in_reg(v4i32 src))
+// TODO Add support for ZERO_EXTEND_VECTOR_INREG when we have a test case.
+SDValue combineShuffleToVectorExtend(ShuffleVectorSDNode *SVN,
+ SelectionDAG &DAG,
+ const TargetLowering &TLI,
+ bool LegalOperations) {
+ EVT VT = SVN->getValueType(0);
+ bool IsBigEndian = DAG.getDataLayout().isBigEndian();
+
+ // TODO Add support for big-endian when we have a test case.
+ if (!VT.isInteger() || IsBigEndian)
+ return SDValue();
+
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned EltSizeInBits = VT.getScalarSizeInBits();
+ ArrayRef<int> Mask = SVN->getMask();
+ SDValue N0 = SVN->getOperand(0);
+
+ // shuffle<0,-1,1,-1> == (v2i64 anyextend_vector_inreg(v4i32))
+ auto isAnyExtend = [&Mask, &NumElts](unsigned Scale) {
+ for (unsigned i = 0; i != NumElts; ++i) {
+ if (Mask[i] < 0)
+ continue;
+ if ((i % Scale) == 0 && Mask[i] == (i / Scale))
+ continue;
+ return false;
+ }
+ return true;
+ };
+
+ // Attempt to match a '*_extend_vector_inreg' shuffle, we just search for
+ // power-of-2 extensions as they are the most likely.
+ for (unsigned Scale = 2; Scale < NumElts; Scale *= 2) {
+ if (!isAnyExtend(Scale))
+ continue;
+
+ EVT OutSVT = EVT::getIntegerVT(*DAG.getContext(), EltSizeInBits * Scale);
+ EVT OutVT = EVT::getVectorVT(*DAG.getContext(), OutSVT, NumElts / Scale);
+ if (!LegalOperations ||
+ TLI.isOperationLegalOrCustom(ISD::ANY_EXTEND_VECTOR_INREG, OutVT))
+ return DAG.getBitcast(VT,
+ DAG.getAnyExtendVectorInReg(N0, SDLoc(SVN), OutVT));
+ }
+
+ return SDValue();
+}
+
+// Detect 'truncate_vector_inreg' style shuffles that pack the lower parts of
+// each source element of a large type into the lowest elements of a smaller
+// destination type. This is often generated during legalization.
+// If the source node itself was a '*_extend_vector_inreg' node then we should
+// then be able to remove it.
+SDValue combineTruncationShuffle(ShuffleVectorSDNode *SVN, SelectionDAG &DAG) {
+ EVT VT = SVN->getValueType(0);
+ bool IsBigEndian = DAG.getDataLayout().isBigEndian();
+
+ // TODO Add support for big-endian when we have a test case.
+ if (!VT.isInteger() || IsBigEndian)
+ return SDValue();
+
+ SDValue N0 = SVN->getOperand(0);
+ while (N0.getOpcode() == ISD::BITCAST)
+ N0 = N0.getOperand(0);
+
+ unsigned Opcode = N0.getOpcode();
+ if (Opcode != ISD::ANY_EXTEND_VECTOR_INREG &&
+ Opcode != ISD::SIGN_EXTEND_VECTOR_INREG &&
+ Opcode != ISD::ZERO_EXTEND_VECTOR_INREG)
+ return SDValue();
+
+ SDValue N00 = N0.getOperand(0);
+ ArrayRef<int> Mask = SVN->getMask();
+ unsigned NumElts = VT.getVectorNumElements();
+ unsigned EltSizeInBits = VT.getScalarSizeInBits();
+ unsigned ExtSrcSizeInBits = N00.getScalarValueSizeInBits();
+
+ // (v4i32 truncate_vector_inreg(v2i64)) == shuffle<0,2-1,-1>
+ // (v8i16 truncate_vector_inreg(v4i32)) == shuffle<0,2,4,6,-1,-1,-1,-1>
+ // (v8i16 truncate_vector_inreg(v2i64)) == shuffle<0,4,-1,-1,-1,-1,-1,-1>
+ auto isTruncate = [&Mask, &NumElts](unsigned Scale) {
+ for (unsigned i = 0; i != NumElts; ++i) {
+ if (Mask[i] < 0)
+ continue;
+ if ((i * Scale) < NumElts && Mask[i] == (i * Scale))
+ continue;
+ return false;
+ }
+ return true;
+ };
+
+ // At the moment we just handle the case where we've truncated back to the
+ // same size as before the extension.
+ // TODO: handle more extension/truncation cases as cases arise.
+ if (EltSizeInBits != ExtSrcSizeInBits)
+ return SDValue();
+
+ // Attempt to match a 'truncate_vector_inreg' shuffle, we just search for
+ // power-of-2 truncations as they are the most likely.
+ for (unsigned Scale = 2; Scale < NumElts; Scale *= 2)
+ if (isTruncate(Scale))
+ return DAG.getBitcast(VT, N00);
+
+ return SDValue();
+}
+
SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
EVT VT = N->getValueType(0);
unsigned NumElts = VT.getVectorNumElements();
@@ -14298,6 +14415,14 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
if (SDValue S = simplifyShuffleOperands(SVN, N0, N1, DAG))
return S;
+ // Match shuffles that can be converted to any_vector_extend_in_reg.
+ if (SDValue V = combineShuffleToVectorExtend(SVN, DAG, TLI, LegalOperations))
+ return V;
+
+ // Combine "truncate_vector_in_reg" style shuffles.
+ if (SDValue V = combineTruncationShuffle(SVN, DAG))
+ return V;
+
if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
Level < AfterLegalizeVectorOps &&
(N1.isUndef() ||
OpenPOWER on IntegriCloud