diff options
Diffstat (limited to 'llvm/lib/Target/X86/X86ISelLowering.cpp')
| -rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 10 |
1 files changed, 9 insertions, 1 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 4056b4982b0..67f2929dae7 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -32737,9 +32737,17 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, SDValue BitCast, if (!VT.isScalarInteger() || !VecVT.isSimple()) return SDValue(); + // If the input is a truncate from v16i8 or v32i8 go ahead and use a + // movmskb even with avx512. This will be better than truncating to vXi1 and + // using a kmov. This can especially help KNL if the input is a v16i8/v32i8 + // vpcmpeqb/vpcmpgtb. + bool IsTruncated = N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() && + (N0.getOperand(0).getValueType() == MVT::v16i8 || + N0.getOperand(0).getValueType() == MVT::v32i8); + // With AVX512 vxi1 types are legal and we prefer using k-regs. // MOVMSK is supported in SSE2 or later. - if (Subtarget.hasAVX512() || !Subtarget.hasSSE2()) + if (!Subtarget.hasSSE2() || (Subtarget.hasAVX512() && !IsTruncated)) return SDValue(); // There are MOVMSK flavors for types v16i8, v32i8, v4f32, v8f32, v4f64 and |
