diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp --- a/llvm/lib/Target/X86/X86ISelLowering.cpp +++ b/llvm/lib/Target/X86/X86ISelLowering.cpp @@ -28135,7 +28135,19 @@ EVT SetCCResultType = TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT); + unsigned BitWidth = VT.getScalarSizeInBits(); if (Opcode == ISD::USUBSAT && !TLI.isOperationLegal(ISD::UMAX, VT)) { + // Handle a special-case with a bit-hack instead of cmp+select: + // usubsat X, SMIN --> (X ^ SMIN) & (X s>> BW-1) + ConstantSDNode *C = isConstOrConstSplat(Y, true); + if (C && C->getAPIntValue().isSignMask()) { + SDValue SignMask = DAG.getConstant(C->getAPIntValue(), DL, VT); + SDValue ShiftAmt = DAG.getConstant(BitWidth - 1, DL, VT); + SDValue Xor = DAG.getNode(ISD::XOR, DL, VT, X, SignMask); + SDValue Sra = DAG.getNode(ISD::SRA, DL, VT, X, ShiftAmt); + return DAG.getNode(ISD::AND, DL, VT, Xor, Sra); + } + // usubsat X, Y --> (X >u Y) ? X - Y : 0 SDValue Sub = DAG.getNode(ISD::SUB, DL, VT, X, Y); SDValue Cmp = DAG.getSetCC(DL, SetCCResultType, X, Y, ISD::SETUGT); @@ -28148,7 +28160,6 @@ if ((Opcode == ISD::SADDSAT || Opcode == ISD::SSUBSAT) && (!VT.isVector() || VT == MVT::v2i64)) { - unsigned BitWidth = VT.getScalarSizeInBits(); APInt MinVal = APInt::getSignedMinValue(BitWidth); APInt MaxVal = APInt::getSignedMaxValue(BitWidth); SDValue Zero = DAG.getConstant(0, DL, VT); diff --git a/llvm/test/CodeGen/X86/psubus.ll b/llvm/test/CodeGen/X86/psubus.ll --- a/llvm/test/CodeGen/X86/psubus.ll +++ b/llvm/test/CodeGen/X86/psubus.ll @@ -29,6 +29,7 @@ } ; This is logically equivalent to the above. +; usubsat X, (1 << (BW-1)) <--> (X ^ (1 << (BW-1))) & (ashr X, (BW-1)) define <8 x i16> @ashr_xor_and(<8 x i16> %x) nounwind { ; SSE-LABEL: ashr_xor_and: @@ -127,14 +128,14 @@ ret <4 x i32> %res } +; usubsat X, (1 << (BW-1)) <--> (X ^ (1 << (BW-1))) & (ashr X, (BW-1)) + define <4 x i32> @usubsat_custom(<4 x i32> %x) nounwind { ; SSE2OR3-LABEL: usubsat_custom: ; SSE2OR3: # %bb.0: -; SSE2OR3-NEXT: movdqa {{.*#+}} xmm1 = [2147483648,2147483648,2147483648,2147483648] -; SSE2OR3-NEXT: pxor %xmm0, %xmm1 -; SSE2OR3-NEXT: pxor %xmm2, %xmm2 -; SSE2OR3-NEXT: pcmpgtd %xmm2, %xmm1 -; SSE2OR3-NEXT: psubd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 +; SSE2OR3-NEXT: movdqa %xmm0, %xmm1 +; SSE2OR3-NEXT: psrad $31, %xmm1 +; SSE2OR3-NEXT: pxor {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0 ; SSE2OR3-NEXT: pand %xmm1, %xmm0 ; SSE2OR3-NEXT: retq ;