diff options
Diffstat (limited to 'llvm/lib/Target')
| -rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 9 |
1 files changed, 5 insertions, 4 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index e033c211943..df0009b3fd7 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -14605,10 +14605,11 @@ static SDValue lowerShuffleWithUndefHalf(const SDLoc &DL, MVT VT, SDValue V1, if (NumUpperHalves == 1) { // AVX2 has efficient 32/64-bit element cross-lane shuffles. if (Subtarget.hasAVX2()) { - // extract128 + vunpckhps, is better than vblend + vpermps. - // TODO: Refine to account for unary shuffle, splat, and other masks? - if (EltWidth == 32 && NumLowerHalves && - HalfVT.is128BitVector() && !is128BitUnpackShuffleMask(HalfMask)) + // extract128 + vunpckhps/vshufps, is better than vblend + vpermps. + if (EltWidth == 32 && NumLowerHalves && HalfVT.is128BitVector() && + !is128BitUnpackShuffleMask(HalfMask) && + (!isSingleSHUFPSMask(HalfMask) || + Subtarget.hasFastVariableShuffle())) return SDValue(); // If this is a unary shuffle (assume that the 2nd operand is // canonicalized to undef), then we can use vpermpd. Otherwise, we |
