diff options
| author | Sanjay Patel <spatel@rotateright.com> | 2019-07-15 18:17:23 +0000 |
|---|---|---|
| committer | Sanjay Patel <spatel@rotateright.com> | 2019-07-15 18:17:23 +0000 |
| commit | eb99165b97b79c3ccc7b5ebcd445a98b4240e171 (patch) | |
| tree | bbe973153fccc1cd6af4e679242f17357166c594 /llvm/lib | |
| parent | 032e3c468fab3ffc033db219314c2151c6bda917 (diff) | |
| download | bcm5719-llvm-eb99165b97b79c3ccc7b5ebcd445a98b4240e171.tar.gz bcm5719-llvm-eb99165b97b79c3ccc7b5ebcd445a98b4240e171.zip | |
[x86] try to keep FP casted+truncated+extracted vector element out of GPRs
inttofp (trunc (extelt X, 0)) --> inttofp (extelt (bitcast X), 0)
We have pseudo-vectorization of scalar int to FP casts, so this tries to
make that more likely by replacing a truncate with a bitcast. I didn't see
any test diffs starting from 'uitofp', so I left that as a TODO. We can't
only match the shorter trunc+extract pattern because there's an opposing
transform somewhere, so we infinite loop. Waiting to try this during
lowering is another possibility.
A motivating case is shown in PR39975 and included in the test diffs here:
https://bugs.llvm.org/show_bug.cgi?id=39975
Differential Revision: https://reviews.llvm.org/D64710
llvm-svn: 366098
Diffstat (limited to 'llvm/lib')
| -rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 39 |
1 files changed, 39 insertions, 0 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 47389f2df32..34a85806f56 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -42430,6 +42430,41 @@ static SDValue combineVectorCompareAndMaskUnaryOp(SDNode *N, return SDValue(); } +/// If we are converting a value to floating-point, try to replace scalar +/// truncate of an extracted vector element with a bitcast. This tries to keep +/// the sequence on XMM registers rather than moving between vector and GPRs. +static SDValue combineToFPTruncExtElt(SDNode *N, SelectionDAG &DAG) { + // TODO: This is currently only used by combineSIntToFP, but it is generalized + // to allow being called by any similar cast opcode. + // TODO: Consider merging this into lowering: vectorizeExtractedCast(). + SDValue Trunc = N->getOperand(0); + if (!Trunc.hasOneUse() || Trunc.getOpcode() != ISD::TRUNCATE) + return SDValue(); + + SDValue ExtElt = Trunc.getOperand(0); + if (!ExtElt.hasOneUse() || ExtElt.getOpcode() != ISD::EXTRACT_VECTOR_ELT || + !isNullConstant(ExtElt.getOperand(1))) + return SDValue(); + + EVT TruncVT = Trunc.getValueType(); + EVT SrcVT = ExtElt.getValueType(); + unsigned DestWidth = TruncVT.getSizeInBits(); + unsigned SrcWidth = SrcVT.getSizeInBits(); + if (SrcWidth % DestWidth != 0) + return SDValue(); + + // inttofp (trunc (extelt X, 0)) --> inttofp (extelt (bitcast X), 0) + EVT SrcVecVT = ExtElt.getOperand(0).getValueType(); + unsigned VecWidth = SrcVecVT.getSizeInBits(); + unsigned NumElts = VecWidth / DestWidth; + EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), TruncVT, NumElts); + SDValue BitcastVec = DAG.getBitcast(BitcastVT, ExtElt.getOperand(0)); + SDLoc DL(N); + SDValue NewExtElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, TruncVT, + BitcastVec, ExtElt.getOperand(1)); + return DAG.getNode(N->getOpcode(), DL, N->getValueType(0), NewExtElt); +} + static SDValue combineUIntToFP(SDNode *N, SelectionDAG &DAG, const X86Subtarget &Subtarget) { SDValue Op0 = N->getOperand(0); @@ -42523,6 +42558,10 @@ static SDValue combineSIntToFP(SDNode *N, SelectionDAG &DAG, return FILDChain; } } + + if (SDValue V = combineToFPTruncExtElt(N, DAG)) + return V; + return SDValue(); } |

