diff options
Diffstat (limited to 'llvm/lib')
| -rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 23 |
1 files changed, 21 insertions, 2 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 67f2929dae7..f643482884e 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -32743,7 +32743,8 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, SDValue BitCast, // vpcmpeqb/vpcmpgtb. bool IsTruncated = N0.getOpcode() == ISD::TRUNCATE && N0.hasOneUse() && (N0.getOperand(0).getValueType() == MVT::v16i8 || - N0.getOperand(0).getValueType() == MVT::v32i8); + N0.getOperand(0).getValueType() == MVT::v32i8 || + N0.getOperand(0).getValueType() == MVT::v64i8); // With AVX512 vxi1 types are legal and we prefer using k-regs. // MOVMSK is supported in SSE2 or later. @@ -32799,12 +32800,30 @@ static SDValue combineBitcastvxi1(SelectionDAG &DAG, SDValue BitCast, case MVT::v32i1: SExtVT = MVT::v32i8; break; + case MVT::v64i1: + // If we have AVX512F, but not AVX512BW and the input is truncated from + // v64i8 checked earlier. Then split the input and make two pmovmskbs. + if (Subtarget.hasAVX512() && !Subtarget.hasBWI()) { + SExtVT = MVT::v64i8; + break; + } + return SDValue(); }; SDLoc DL(BitCast); SDValue V = DAG.getNode(ISD::SIGN_EXTEND, DL, SExtVT, N0); - if (SExtVT == MVT::v16i8 || SExtVT == MVT::v32i8) { + if (SExtVT == MVT::v64i8) { + SDValue Lo, Hi; + std::tie(Lo, Hi) = DAG.SplitVector(V, DL); + Lo = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Lo); + Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Lo); + Hi = DAG.getNode(X86ISD::MOVMSK, DL, MVT::i32, Hi); + Hi = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, Hi); + Hi = DAG.getNode(ISD::SHL, DL, MVT::i64, Hi, + DAG.getConstant(32, DL, MVT::i8)); + V = DAG.getNode(ISD::OR, DL, MVT::i64, Lo, Hi); + } else if (SExtVT == MVT::v16i8 || SExtVT == MVT::v32i8) { V = getPMOVMSKB(DL, V, DAG, Subtarget); } else { if (SExtVT == MVT::v8i16) |
