diff options
author | Andrea Di Biagio <Andrea_DiBiagio@sn.scee.net> | 2014-06-26 10:45:21 +0000 |
---|---|---|
committer | Andrea Di Biagio <Andrea_DiBiagio@sn.scee.net> | 2014-06-26 10:45:21 +0000 |
commit | 817812f61caa50ba3ce67ceb5a033cb69a3eea5c (patch) | |
tree | 3246ef4217ca3ed21a10573b856cc18a085035d4 /lib/Target/X86 | |
parent | c2b11baf5fdb26c2bf2c9b02dae9f0f0f61e9f1b (diff) | |
download | llvm-817812f61caa50ba3ce67ceb5a033cb69a3eea5c.tar.gz llvm-817812f61caa50ba3ce67ceb5a033cb69a3eea5c.tar.bz2 llvm-817812f61caa50ba3ce67ceb5a033cb69a3eea5c.tar.xz |
[X86] Improve the selection of SSE3/AVX addsub instructions.
This patch teaches the backend how to canonicalize a shuffle vectors
according to the rule:
- (shuffle (FADD A, B), (FSUB A, B), Mask) ->
(shuffle (FSUB A, -B), (FADD A, -B), Mask)
Where 'Mask' is:
<0,5,2,7> ;; for v4f32 and v4f64 shuffles.
<0,3> ;; for v2f64 shuffles.
<0,9,2,11,4,13,6,15> ;; for v8f32 shuffles.
In general, ISel only knows how to pattern-match a canonical
'fadd + fsub + blendi' dag node sequence into an ADDSUB instruction.
This new rule allows to convert a non-canonical dag sequence into a
canonical one that will be matched by a single ADDSUB at ISel stage.
The idea of converting a non-canonical ADDSUB into a canonical one by
swapping the first two operands of the shuffle, and then negating the
second operand of the FADD and FSUB, was originally proposed by Hal Finkel.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@211771 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib/Target/X86')
-rw-r--r-- | lib/Target/X86/X86ISelLowering.cpp | 43 |
1 files changed, 43 insertions, 0 deletions
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp index cde413f455..ebacc425df 100644 --- a/lib/Target/X86/X86ISelLowering.cpp +++ b/lib/Target/X86/X86ISelLowering.cpp @@ -18014,6 +18014,49 @@ static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG, SDValue N1 = N->getOperand(1); EVT VT = N->getValueType(0); + // Canonicalize shuffles that perform 'addsub' on packed float vectors + // according to the rule: + // (shuffle (FADD A, B), (FSUB A, B), Mask) -> + // (shuffle (FSUB A, -B), (FADD A, -B), Mask) + // + // Where 'Mask' is: + // <0,5,2,7> -- for v4f32 and v4f64 shuffles; + // <0,3> -- for v2f64 shuffles; + // <0,9,2,11,4,13,6,15> -- for v8f32 shuffles. + // + // This helps pattern-matching more SSE3/AVX ADDSUB instructions + // during ISel stage. + if (N->getOpcode() == ISD::VECTOR_SHUFFLE && + ((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) || + (Subtarget->hasAVX() && (VT == MVT::v8f32 || VT == MVT::v4f64))) && + N0->getOpcode() == ISD::FADD && N1->getOpcode() == ISD::FSUB && + // Operands to the FADD and FSUB must be the same. + ((N0->getOperand(0) == N1->getOperand(0) && + N0->getOperand(1) == N1->getOperand(1)) || + // FADD is commutable. See if by commuting the operands of the FADD + // we would still be able to match the operands of the FSUB dag node. + (N0->getOperand(1) == N1->getOperand(0) && + N0->getOperand(0) == N1->getOperand(1))) && + N0->getOperand(0)->getOpcode() != ISD::UNDEF && + N0->getOperand(1)->getOpcode() != ISD::UNDEF) { + + ShuffleVectorSDNode *SV = cast<ShuffleVectorSDNode>(N); + unsigned NumElts = VT.getVectorNumElements(); + ArrayRef<int> Mask = SV->getMask(); + bool CanFold = true; + + for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) + CanFold = Mask[i] == (i & 1) ? i + NumElts : i; + + if (CanFold) { + SDValue Op0 = N1->getOperand(0); + SDValue Op1 = DAG.getNode(ISD::FNEG, dl, VT, N1->getOperand(1)); + SDValue Sub = DAG.getNode(ISD::FSUB, dl, VT, Op0, Op1); + SDValue Add = DAG.getNode(ISD::FADD, dl, VT, Op0, Op1); + return DAG.getVectorShuffle(VT, dl, Sub, Add, Mask); + } + } + // Don't create instructions with illegal types after legalize types has run. const TargetLowering &TLI = DAG.getTargetLoweringInfo(); if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType())) |