diff options
Diffstat (limited to 'llvm/lib')
| -rw-r--r-- | llvm/lib/Target/X86/X86ISelLowering.cpp | 44 |
1 files changed, 44 insertions, 0 deletions
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp index 77fa9ff26f2..3f25cb15441 100644 --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -3844,6 +3844,7 @@ static bool isTargetShuffleVariableMask(unsigned Opcode) { default: return false; case X86ISD::PSHUFB: case X86ISD::VPERMILPV: + case X86ISD::VPERMIL2: case X86ISD::VPPERM: return true; } @@ -25288,6 +25289,49 @@ static bool combineX86ShuffleChain(ArrayRef<SDValue> Inputs, SDValue Root, return true; } + // With XOP, binary shuffles of 128/256-bit floating point vectors can combine + // to VPERMIL2PD/VPERMIL2PS. + if ((Depth >= 3 || HasVariableMask) && Subtarget.hasXOP() && + (MaskVT == MVT::v2f64 || MaskVT == MVT::v4f64 || MaskVT == MVT::v4f32 || + MaskVT == MVT::v8f32)) { + // VPERMIL2 Operation. + // Bits[3] - Match Bit. + // Bits[2:1] - (Per Lane) PD Shuffle Mask. + // Bits[2:0] - (Per Lane) PS Shuffle Mask. + unsigned NumLanes = MaskVT.getSizeInBits() / 128; + unsigned NumEltsPerLane = NumMaskElts / NumLanes; + SmallVector<SDValue, 8> VPerm2Idx; + MVT MaskIdxSVT = MVT::getIntegerVT(MaskVT.getScalarSizeInBits()); + MVT MaskIdxVT = MVT::getVectorVT(MaskIdxSVT, NumMaskElts); + unsigned M2ZImm = 0; + for (int M : Mask) { + if (M == SM_SentinelUndef) { + VPerm2Idx.push_back(DAG.getUNDEF(MaskIdxSVT)); + continue; + } + if (M == SM_SentinelZero) { + M2ZImm = 2; + VPerm2Idx.push_back(DAG.getConstant(8, DL, MaskIdxSVT)); + continue; + } + int Index = (M % NumEltsPerLane) + ((M / NumMaskElts) * NumEltsPerLane); + Index = (MaskVT.getScalarSizeInBits() == 64 ? Index << 1 : Index); + VPerm2Idx.push_back(DAG.getConstant(Index, DL, MaskIdxSVT)); + } + V1 = DAG.getBitcast(MaskVT, V1); + DCI.AddToWorklist(V1.getNode()); + V2 = DAG.getBitcast(MaskVT, V2); + DCI.AddToWorklist(V2.getNode()); + SDValue VPerm2MaskOp = DAG.getBuildVector(MaskIdxVT, DL, VPerm2Idx); + DCI.AddToWorklist(VPerm2MaskOp.getNode()); + Res = DAG.getNode(X86ISD::VPERMIL2, DL, MaskVT, V1, V2, VPerm2MaskOp, + DAG.getConstant(M2ZImm, DL, MVT::i8)); + DCI.AddToWorklist(Res.getNode()); + DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Res), + /*AddTo*/ true); + return true; + } + // If we have 3 or more shuffle instructions or a chain involving a variable // mask, we can replace them with a single PSHUFB instruction profitably. // Intel's manuals suggest only using PSHUFB if doing so replacing 5 |
