diff options
Diffstat (limited to 'llvm/lib/Target')
| -rw-r--r-- | llvm/lib/Target/Hexagon/HexagonISelLowering.cpp | 44 | ||||
| -rw-r--r-- | llvm/lib/Target/Hexagon/HexagonPatterns.td | 13 |
2 files changed, 28 insertions, 29 deletions
diff --git a/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp b/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp index 67242764d45..88fb1bee27d 100644 --- a/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp +++ b/llvm/lib/Target/Hexagon/HexagonISelLowering.cpp @@ -1286,16 +1286,6 @@ HexagonTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const { MachinePointerInfo(SV)); } -// Creates a SPLAT instruction for a constant value VAL. -static SDValue createSplat(SelectionDAG &DAG, const SDLoc &dl, EVT VT, - SDValue Val) { - EVT T = VT.getVectorElementType(); - if (T == MVT::i8 || T == MVT::i16) - return DAG.getNode(HexagonISD::VSPLAT, dl, VT, Val); - - return SDValue(); -} - static bool isSExtFree(SDValue N) { // A sign-extend of a truncate of a sign-extend is free. if (N.getOpcode() == ISD::TRUNCATE && @@ -2394,7 +2384,7 @@ HexagonTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) // Test if V1 is a SCALAR_TO_VECTOR. if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) - return createSplat(DAG, dl, VT, V1.getOperand(0)); + return DAG.getNode(HexagonISD::VSPLAT, dl, VT, V1.getOperand(0)); // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR // (and probably will turn into a SCALAR_TO_VECTOR once legalization @@ -2409,9 +2399,10 @@ HexagonTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) } } if (IsScalarToVector) - return createSplat(DAG, dl, VT, V1.getOperand(0)); + return DAG.getNode(HexagonISD::VSPLAT, dl, VT, V1.getOperand(0)); } - return createSplat(DAG, dl, VT, DAG.getConstant(Lane, dl, MVT::i32)); + return DAG.getNode(HexagonISD::VSPLAT, dl, VT, + DAG.getConstant(Lane, dl, MVT::i32)); } if (UseHVX) { @@ -2531,19 +2522,26 @@ HexagonTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const { if (Size > 64) return SDValue(); - APInt APSplatBits, APSplatUndef; - unsigned SplatBitSize; - bool HasAnyUndefs; unsigned NElts = BVN->getNumOperands(); // Try to generate a SPLAT instruction. - if ((VT.getSimpleVT() == MVT::v4i8 || VT.getSimpleVT() == MVT::v4i16) && - (BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize, - HasAnyUndefs, 0, true) && SplatBitSize <= 16)) { - unsigned SplatBits = APSplatBits.getZExtValue(); - int32_t SextVal = ((int32_t) (SplatBits << (32 - SplatBitSize)) >> - (32 - SplatBitSize)); - return createSplat(DAG, dl, VT, DAG.getConstant(SextVal, dl, MVT::i32)); + if (VT == MVT::v4i8 || VT == MVT::v4i16 || VT == MVT::v2i32) { + APInt APSplatBits, APSplatUndef; + unsigned SplatBitSize; + bool HasAnyUndefs; + if (BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize, + HasAnyUndefs, 0, false)) { + if (SplatBitSize == VT.getVectorElementType().getSizeInBits()) { + unsigned ZV = APSplatBits.getZExtValue(); + assert(SplatBitSize <= 32 && "Can only handle up to i32"); + // Sign-extend the splat value from SplatBitSize to 32. + int32_t SV = SplatBitSize < 32 + ? int32_t(ZV << (32-SplatBitSize)) >> (32-SplatBitSize) + : int32_t(ZV); + return DAG.getNode(HexagonISD::VSPLAT, dl, VT, + DAG.getConstant(SV, dl, MVT::i32)); + } + } } // Try to generate COMBINE to build v2i32 vectors. diff --git a/llvm/lib/Target/Hexagon/HexagonPatterns.td b/llvm/lib/Target/Hexagon/HexagonPatterns.td index ba98b899493..9dc423932d2 100644 --- a/llvm/lib/Target/Hexagon/HexagonPatterns.td +++ b/llvm/lib/Target/Hexagon/HexagonPatterns.td @@ -3073,6 +3073,10 @@ def: Pat<(v4i8 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrb I32:$Rs)>; // four halfwords of 64-bits destination register. def: Pat<(v4i16 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrh I32:$Rs)>; +def: Pat<(v2i32 (HexagonVSPLAT s8_0ImmPred:$s8)), + (A2_combineii imm:$s8, imm:$s8)>; +def: Pat<(v2i32 (HexagonVSPLAT I32:$Rs)), (A2_combinew I32:$Rs, I32:$Rs)>; + class VArith_pat <InstHexagon MI, SDNode Op, PatFrag Type> : Pat <(Op Type:$Rss, Type:$Rtt), @@ -3099,14 +3103,11 @@ def: VArith_pat <A2_xorp, xor, V8I8>; def: VArith_pat <A2_xorp, xor, V4I16>; def: VArith_pat <A2_xorp, xor, V2I32>; -def: Pat<(v2i32 (sra V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c), - (i32 u5_0ImmPred:$c))))), +def: Pat<(v2i32 (sra V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c)))), (S2_asr_i_vw V2I32:$b, imm:$c)>; -def: Pat<(v2i32 (srl V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c), - (i32 u5_0ImmPred:$c))))), +def: Pat<(v2i32 (srl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c)))), (S2_lsr_i_vw V2I32:$b, imm:$c)>; -def: Pat<(v2i32 (shl V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c), - (i32 u5_0ImmPred:$c))))), +def: Pat<(v2i32 (shl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c)))), (S2_asl_i_vw V2I32:$b, imm:$c)>; def: Pat<(v4i16 (sra V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c)))), |

