Index: llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h =================================================================== --- llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h +++ llvm/include/llvm/CodeGen/GlobalISel/LegalizerHelper.h @@ -275,6 +275,7 @@ LegalizeResult lowerFMinNumMaxNum(MachineInstr &MI); LegalizeResult lowerFMad(MachineInstr &MI); LegalizeResult lowerIntrinsicRound(MachineInstr &MI); + LegalizeResult lowerFFloor(MachineInstr &MI); LegalizeResult lowerUnmergeValues(MachineInstr &MI); LegalizeResult lowerShuffleVector(MachineInstr &MI); LegalizeResult lowerDynStackAlloc(MachineInstr &MI); Index: llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp =================================================================== --- llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp +++ llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp @@ -2194,6 +2194,8 @@ } case TargetOpcode::G_FMAD: return lowerFMad(MI); + case TargetOpcode::G_FFLOOR: + return lowerFFloor(MI); case TargetOpcode::G_INTRINSIC_ROUND: return lowerIntrinsicRound(MI); case TargetOpcode::G_ATOMIC_CMPXCHG_WITH_SUCCESS: { @@ -4646,6 +4648,39 @@ LegalizerHelper::LegalizeResult LegalizerHelper::lowerIntrinsicRound(MachineInstr &MI) { + Register DstReg = MI.getOperand(0).getReg(); + Register X = MI.getOperand(1).getReg(); + const unsigned Flags = MI.getFlags(); + const LLT Ty = MRI.getType(DstReg); + const LLT CondTy = Ty.changeElementSize(1); + + // round(x) => + // t = trunc(x); + // d = fabs(x - t); + // o = copysign(1.0f, x); + // return t + (d >= 0.5 ? o : 0.0); + + auto T = MIRBuilder.buildIntrinsicTrunc(Ty, X, Flags); + + auto Diff = MIRBuilder.buildFSub(Ty, X, T, Flags); + auto AbsDiff = MIRBuilder.buildFAbs(Ty, Diff, Flags); + auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); + auto One = MIRBuilder.buildFConstant(Ty, 1.0); + auto Half = MIRBuilder.buildFConstant(Ty, 0.5); + auto SignOne = MIRBuilder.buildFCopysign(Ty, One, X); + + auto Cmp = MIRBuilder.buildFCmp(CmpInst::FCMP_OGE, CondTy, AbsDiff, Half, + Flags); + auto Sel = MIRBuilder.buildSelect(Ty, Cmp, SignOne, Zero, Flags); + + MIRBuilder.buildFAdd(DstReg, T, Sel, Flags); + + MI.eraseFromParent(); + return Legalized; +} + +LegalizerHelper::LegalizeResult +LegalizerHelper::lowerFFloor(MachineInstr &MI) { Register DstReg = MI.getOperand(0).getReg(); Register SrcReg = MI.getOperand(1).getReg(); unsigned Flags = MI.getFlags(); @@ -4656,8 +4691,8 @@ // if (src < 0.0 && src != result) // result += -1.0. - auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); auto Trunc = MIRBuilder.buildIntrinsicTrunc(Ty, SrcReg, Flags); + auto Zero = MIRBuilder.buildFConstant(Ty, 0.0); auto Lt0 = MIRBuilder.buildFCmp(CmpInst::FCMP_OLT, CondTy, SrcReg, Zero, Flags); @@ -4666,7 +4701,7 @@ auto And = MIRBuilder.buildAnd(CondTy, Lt0, NeTrunc); auto AddVal = MIRBuilder.buildSITOFP(Ty, And); - MIRBuilder.buildFAdd(DstReg, Trunc, AddVal); + MIRBuilder.buildFAdd(DstReg, Trunc, AddVal, Flags); MI.eraseFromParent(); return Legalized; } Index: llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir =================================================================== --- llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir +++ llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-intrinsic-round.mir @@ -11,35 +11,41 @@ ; GFX6-LABEL: name: test_intrinsic_round_s32 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 - ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[COPY]] - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s32), [[C]] - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]] - ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] + ; GFX6: [[FSUB:%[0-9]+]]:_(s32) = G_FSUB [[COPY]], [[INTRINSIC_TRUNC]] + ; GFX6: [[FABS:%[0-9]+]]:_(s32) = G_FABS [[FSUB]] + ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX6: $vgpr0 = COPY [[FADD]](s32) ; GFX8-LABEL: name: test_intrinsic_round_s32 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 - ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[COPY]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s32), [[C]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] + ; GFX8: [[FSUB:%[0-9]+]]:_(s32) = G_FSUB [[COPY]], [[INTRINSIC_TRUNC]] + ; GFX8: [[FABS:%[0-9]+]]:_(s32) = G_FABS [[FSUB]] + ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX8: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] ; GFX8: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX8: $vgpr0 = COPY [[FADD]](s32) ; GFX9-LABEL: name: test_intrinsic_round_s32 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 - ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[COPY]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s32), [[C]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] + ; GFX9: [[FSUB:%[0-9]+]]:_(s32) = G_FSUB [[COPY]], [[INTRINSIC_TRUNC]] + ; GFX9: [[FABS:%[0-9]+]]:_(s32) = G_FABS [[FSUB]] + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] ; GFX9: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX9: $vgpr0 = COPY [[FADD]](s32) %0:_(s32) = COPY $vgpr0 @@ -55,36 +61,42 @@ ; GFX6-LABEL: name: test_intrinsic_round_s32_flags ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 - ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = nsz G_INTRINSIC_TRUNC [[COPY]] - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(olt), [[COPY]](s32), [[C]] - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]] - ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX6: [[FSUB:%[0-9]+]]:_(s32) = nsz G_FSUB [[COPY]], [[INTRINSIC_TRUNC]] + ; GFX6: [[FABS:%[0-9]+]]:_(s32) = nsz G_FABS [[FSUB]] + ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = nsz G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX6: [[FADD:%[0-9]+]]:_(s32) = nsz G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX6: $vgpr0 = COPY [[FADD]](s32) ; GFX8-LABEL: name: test_intrinsic_round_s32_flags ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 - ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = nsz G_INTRINSIC_TRUNC [[COPY]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(olt), [[COPY]](s32), [[C]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX8: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX8: [[FSUB:%[0-9]+]]:_(s32) = nsz G_FSUB [[COPY]], [[INTRINSIC_TRUNC]] + ; GFX8: [[FABS:%[0-9]+]]:_(s32) = nsz G_FABS [[FSUB]] + ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX8: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = nsz G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX8: [[FADD:%[0-9]+]]:_(s32) = nsz G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX8: $vgpr0 = COPY [[FADD]](s32) ; GFX9-LABEL: name: test_intrinsic_round_s32_flags ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 - ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = nsz G_INTRINSIC_TRUNC [[COPY]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(olt), [[COPY]](s32), [[C]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(one), [[COPY]](s32), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX9: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX9: [[FSUB:%[0-9]+]]:_(s32) = nsz G_FSUB [[COPY]], [[INTRINSIC_TRUNC]] + ; GFX9: [[FABS:%[0-9]+]]:_(s32) = nsz G_FABS [[FSUB]] + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = nsz G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = nsz G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX9: [[FADD:%[0-9]+]]:_(s32) = nsz G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX9: $vgpr0 = COPY [[FADD]](s32) %0:_(s32) = COPY $vgpr0 %1:_(s32) = nsz G_INTRINSIC_ROUND %0 @@ -99,57 +111,66 @@ ; GFX6-LABEL: name: test_intrinsic_round_s64 ; GFX6: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1 - ; GFX6: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 ; GFX6: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](s64) - ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 20 - ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 11 - ; GFX6: [[INT:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C1]](s32), [[C2]](s32) - ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1023 - ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[INT]], [[C3]] - ; GFX6: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 -2147483648 - ; GFX6: [[AND:%[0-9]+]]:_(s32) = G_AND [[UV1]], [[C4]] - ; GFX6: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4503599627370495 - ; GFX6: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 - ; GFX6: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C6]](s32), [[AND]](s32) - ; GFX6: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[C5]], [[SUB]](s32) - ; GFX6: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1 - ; GFX6: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[ASHR]], [[C7]] + ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 20 + ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 11 + ; GFX6: [[INT:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C]](s32), [[C1]](s32) + ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1023 + ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[INT]], [[C2]] + ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 -2147483648 + ; GFX6: [[AND:%[0-9]+]]:_(s32) = G_AND [[UV1]], [[C3]] + ; GFX6: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4503599627370495 + ; GFX6: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 + ; GFX6: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C5]](s32), [[AND]](s32) + ; GFX6: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[C4]], [[SUB]](s32) + ; GFX6: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1 + ; GFX6: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[ASHR]], [[C6]] ; GFX6: [[AND1:%[0-9]+]]:_(s64) = G_AND [[COPY]], [[XOR]] - ; GFX6: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 51 - ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB]](s32), [[C6]] - ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB]](s32), [[C8]] + ; GFX6: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 51 + ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB]](s32), [[C5]] + ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB]](s32), [[C7]] ; GFX6: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[MV]], [[AND1]] ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[ICMP1]](s1), [[COPY]], [[SELECT]] ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[COPY]] - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s64), [[C]] - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s64), [[INTRINSIC_TRUNC]] - ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C9:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00 - ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND2]](s1), [[C9]], [[C]] - ; GFX6: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT1]] - ; GFX6: $vgpr0_vgpr1 = COPY [[FADD]](s64) + ; GFX6: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX6: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[COPY]], [[FNEG]] + ; GFX6: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[FADD]] + ; GFX6: [[C8:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 + ; GFX6: [[C9:%[0-9]+]]:_(s64) = G_FCONSTANT double 5.000000e-01 + ; GFX6: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 4607182418800017408 + ; GFX6: [[COPY1:%[0-9]+]]:_(s64) = COPY [[C10]](s64) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s64), [[C9]] + ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C8]] + ; GFX6: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT1]] + ; GFX6: $vgpr0_vgpr1 = COPY [[FADD1]](s64) ; GFX8-LABEL: name: test_intrinsic_round_s64 ; GFX8: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1 - ; GFX8: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[COPY]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s64), [[C]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s64), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX8: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] - ; GFX8: $vgpr0_vgpr1 = COPY [[FADD]](s64) + ; GFX8: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX8: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[COPY]], [[FNEG]] + ; GFX8: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[FADD]] + ; GFX8: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 + ; GFX8: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double 5.000000e-01 + ; GFX8: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 4607182418800017408 + ; GFX8: [[COPY1:%[0-9]+]]:_(s64) = COPY [[C2]](s64) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s64), [[C1]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX8: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX8: $vgpr0_vgpr1 = COPY [[FADD1]](s64) ; GFX9-LABEL: name: test_intrinsic_round_s64 ; GFX9: [[COPY:%[0-9]+]]:_(s64) = COPY $vgpr0_vgpr1 - ; GFX9: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[COPY]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[COPY]](s64), [[C]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[COPY]](s64), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX9: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] - ; GFX9: $vgpr0_vgpr1 = COPY [[FADD]](s64) + ; GFX9: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX9: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[COPY]], [[FNEG]] + ; GFX9: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[FADD]] + ; GFX9: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 + ; GFX9: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double 5.000000e-01 + ; GFX9: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 4607182418800017408 + ; GFX9: [[COPY1:%[0-9]+]]:_(s64) = COPY [[C2]](s64) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s64), [[C1]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX9: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX9: $vgpr0_vgpr1 = COPY [[FADD1]](s64) %0:_(s64) = COPY $vgpr0_vgpr1 %1:_(s64) = G_INTRINSIC_ROUND %0 $vgpr0_vgpr1 = COPY %1 @@ -164,57 +185,66 @@ ; GFX6-LABEL: name: test_intrinsic_round_v2s32 ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1 ; GFX6: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>) - ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV]] - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s32), [[C]] - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s32), [[INTRINSIC_TRUNC]] - ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] + ; GFX6: [[FSUB:%[0-9]+]]:_(s32) = G_FSUB [[UV]], [[INTRINSIC_TRUNC]] + ; GFX6: [[FABS:%[0-9]+]]:_(s32) = G_FABS [[FSUB]] + ; GFX6: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX6: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX6: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV1]] - ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s32), [[C]] - ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s32), [[INTRINSIC_TRUNC1]] - ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX6: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[AND1]](s1), [[C1]], [[C]] + ; GFX6: [[FSUB1:%[0-9]+]]:_(s32) = G_FSUB [[UV1]], [[INTRINSIC_TRUNC1]] + ; GFX6: [[FABS1:%[0-9]+]]:_(s32) = G_FABS [[FSUB1]] + ; GFX6: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s32), [[C1]] + ; GFX6: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C]] ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FADD]](s32), [[FADD1]](s32) ; GFX6: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>) ; GFX8-LABEL: name: test_intrinsic_round_v2s32 ; GFX8: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1 ; GFX8: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>) - ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s32), [[C]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s32), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] + ; GFX8: [[FSUB:%[0-9]+]]:_(s32) = G_FSUB [[UV]], [[INTRINSIC_TRUNC]] + ; GFX8: [[FABS:%[0-9]+]]:_(s32) = G_FABS [[FSUB]] + ; GFX8: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX8: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX8: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] ; GFX8: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV1]] - ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s32), [[C]] - ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s32), [[INTRINSIC_TRUNC1]] - ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX8: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[AND1]](s1), [[C1]], [[C]] + ; GFX8: [[FSUB1:%[0-9]+]]:_(s32) = G_FSUB [[UV1]], [[INTRINSIC_TRUNC1]] + ; GFX8: [[FABS1:%[0-9]+]]:_(s32) = G_FABS [[FSUB1]] + ; GFX8: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s32), [[C1]] + ; GFX8: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C]] ; GFX8: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FADD]](s32), [[FADD1]](s32) ; GFX8: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>) ; GFX9-LABEL: name: test_intrinsic_round_v2s32 ; GFX9: [[COPY:%[0-9]+]]:_(<2 x s32>) = COPY $vgpr0_vgpr1 ; GFX9: [[UV:%[0-9]+]]:_(s32), [[UV1:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[COPY]](<2 x s32>) - ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s32), [[C]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s32), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float -1.000000e+00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[AND]](s1), [[C1]], [[C]] + ; GFX9: [[FSUB:%[0-9]+]]:_(s32) = G_FSUB [[UV]], [[INTRINSIC_TRUNC]] + ; GFX9: [[FABS:%[0-9]+]]:_(s32) = G_FABS [[FSUB]] + ; GFX9: [[C:%[0-9]+]]:_(s32) = G_FCONSTANT float 0.000000e+00 + ; GFX9: [[C1:%[0-9]+]]:_(s32) = G_FCONSTANT float 5.000000e-01 + ; GFX9: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1065353216 + ; GFX9: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s32), [[C1]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s32) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] ; GFX9: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[UV1]] - ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s32), [[C]] - ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s32), [[INTRINSIC_TRUNC1]] - ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX9: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[AND1]](s1), [[C1]], [[C]] + ; GFX9: [[FSUB1:%[0-9]+]]:_(s32) = G_FSUB [[UV1]], [[INTRINSIC_TRUNC1]] + ; GFX9: [[FABS1:%[0-9]+]]:_(s32) = G_FABS [[FSUB1]] + ; GFX9: [[COPY2:%[0-9]+]]:_(s32) = COPY [[C2]](s32) + ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s32), [[C1]] + ; GFX9: [[SELECT1:%[0-9]+]]:_(s32) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C]] ; GFX9: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s32>) = G_BUILD_VECTOR [[FADD]](s32), [[FADD1]](s32) ; GFX9: $vgpr0_vgpr1 = COPY [[BUILD_VECTOR]](<2 x s32>) @@ -232,91 +262,106 @@ ; GFX6-LABEL: name: test_intrinsic_round_v2s64 ; GFX6: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX6: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>) - ; GFX6: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 ; GFX6: [[UV2:%[0-9]+]]:_(s32), [[UV3:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV]](s64) - ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 20 - ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 11 - ; GFX6: [[INT:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C1]](s32), [[C2]](s32) - ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 1023 - ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[INT]], [[C3]] - ; GFX6: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 -2147483648 - ; GFX6: [[AND:%[0-9]+]]:_(s32) = G_AND [[UV3]], [[C4]] - ; GFX6: [[C5:%[0-9]+]]:_(s64) = G_CONSTANT i64 4503599627370495 - ; GFX6: [[C6:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 - ; GFX6: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C6]](s32), [[AND]](s32) - ; GFX6: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[C5]], [[SUB]](s32) - ; GFX6: [[C7:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1 - ; GFX6: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[ASHR]], [[C7]] + ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 20 + ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 11 + ; GFX6: [[INT:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C]](s32), [[C1]](s32) + ; GFX6: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 1023 + ; GFX6: [[SUB:%[0-9]+]]:_(s32) = G_SUB [[INT]], [[C2]] + ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 -2147483648 + ; GFX6: [[AND:%[0-9]+]]:_(s32) = G_AND [[UV3]], [[C3]] + ; GFX6: [[C4:%[0-9]+]]:_(s64) = G_CONSTANT i64 4503599627370495 + ; GFX6: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 + ; GFX6: [[MV:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C5]](s32), [[AND]](s32) + ; GFX6: [[ASHR:%[0-9]+]]:_(s64) = G_ASHR [[C4]], [[SUB]](s32) + ; GFX6: [[C6:%[0-9]+]]:_(s64) = G_CONSTANT i64 -1 + ; GFX6: [[XOR:%[0-9]+]]:_(s64) = G_XOR [[ASHR]], [[C6]] ; GFX6: [[AND1:%[0-9]+]]:_(s64) = G_AND [[UV]], [[XOR]] - ; GFX6: [[C8:%[0-9]+]]:_(s32) = G_CONSTANT i32 51 - ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB]](s32), [[C6]] - ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB]](s32), [[C8]] + ; GFX6: [[C7:%[0-9]+]]:_(s32) = G_CONSTANT i32 51 + ; GFX6: [[ICMP:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB]](s32), [[C5]] + ; GFX6: [[ICMP1:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB]](s32), [[C7]] ; GFX6: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[ICMP]](s1), [[MV]], [[AND1]] ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[ICMP1]](s1), [[UV]], [[SELECT]] ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV]] - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s64), [[C]] - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s64), [[INTRINSIC_TRUNC]] - ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C9:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00 - ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND2]](s1), [[C9]], [[C]] - ; GFX6: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT1]] + ; GFX6: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX6: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[UV]], [[FNEG]] + ; GFX6: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[FADD]] + ; GFX6: [[C8:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 + ; GFX6: [[C9:%[0-9]+]]:_(s64) = G_FCONSTANT double 5.000000e-01 + ; GFX6: [[C10:%[0-9]+]]:_(s64) = G_CONSTANT i64 4607182418800017408 + ; GFX6: [[COPY1:%[0-9]+]]:_(s64) = COPY [[C10]](s64) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s64), [[C9]] + ; GFX6: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C8]] + ; GFX6: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT1]] ; GFX6: [[UV4:%[0-9]+]]:_(s32), [[UV5:%[0-9]+]]:_(s32) = G_UNMERGE_VALUES [[UV1]](s64) - ; GFX6: [[INT1:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C1]](s32), [[C2]](s32) - ; GFX6: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[INT1]], [[C3]] - ; GFX6: [[AND3:%[0-9]+]]:_(s32) = G_AND [[UV5]], [[C4]] - ; GFX6: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C6]](s32), [[AND3]](s32) - ; GFX6: [[ASHR1:%[0-9]+]]:_(s64) = G_ASHR [[C5]], [[SUB1]](s32) - ; GFX6: [[XOR1:%[0-9]+]]:_(s64) = G_XOR [[ASHR1]], [[C7]] - ; GFX6: [[AND4:%[0-9]+]]:_(s64) = G_AND [[UV1]], [[XOR1]] - ; GFX6: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB1]](s32), [[C6]] - ; GFX6: [[ICMP3:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB1]](s32), [[C8]] - ; GFX6: [[SELECT2:%[0-9]+]]:_(s64) = G_SELECT [[ICMP2]](s1), [[MV1]], [[AND4]] + ; GFX6: [[INT1:%[0-9]+]]:_(s32) = G_INTRINSIC intrinsic(@llvm.amdgcn.ubfe), [[C]](s32), [[C1]](s32) + ; GFX6: [[SUB1:%[0-9]+]]:_(s32) = G_SUB [[INT1]], [[C2]] + ; GFX6: [[AND2:%[0-9]+]]:_(s32) = G_AND [[UV5]], [[C3]] + ; GFX6: [[MV1:%[0-9]+]]:_(s64) = G_MERGE_VALUES [[C5]](s32), [[AND2]](s32) + ; GFX6: [[ASHR1:%[0-9]+]]:_(s64) = G_ASHR [[C4]], [[SUB1]](s32) + ; GFX6: [[XOR1:%[0-9]+]]:_(s64) = G_XOR [[ASHR1]], [[C6]] + ; GFX6: [[AND3:%[0-9]+]]:_(s64) = G_AND [[UV1]], [[XOR1]] + ; GFX6: [[ICMP2:%[0-9]+]]:_(s1) = G_ICMP intpred(slt), [[SUB1]](s32), [[C5]] + ; GFX6: [[ICMP3:%[0-9]+]]:_(s1) = G_ICMP intpred(sgt), [[SUB1]](s32), [[C7]] + ; GFX6: [[SELECT2:%[0-9]+]]:_(s64) = G_SELECT [[ICMP2]](s1), [[MV1]], [[AND3]] ; GFX6: [[SELECT3:%[0-9]+]]:_(s64) = G_SELECT [[ICMP3]](s1), [[UV1]], [[SELECT2]] ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV1]] - ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s64), [[C]] - ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s64), [[INTRINSIC_TRUNC1]] - ; GFX6: [[AND5:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX6: [[SELECT3:%[0-9]+]]:_(s64) = G_SELECT [[AND5]](s1), [[C9]], [[C]] - ; GFX6: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT3]] - ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD]](s64), [[FADD1]](s64) + ; GFX6: [[FNEG1:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX6: [[FADD2:%[0-9]+]]:_(s64) = G_FADD [[UV1]], [[FNEG1]] + ; GFX6: [[FABS1:%[0-9]+]]:_(s64) = G_FABS [[FADD2]] + ; GFX6: [[COPY2:%[0-9]+]]:_(s64) = COPY [[C10]](s64) + ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s64), [[C9]] + ; GFX6: [[SELECT3:%[0-9]+]]:_(s64) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C8]] + ; GFX6: [[FADD3:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT3]] + ; GFX6: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD1]](s64), [[FADD3]](s64) ; GFX6: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>) ; GFX8-LABEL: name: test_intrinsic_round_v2s64 ; GFX8: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX8: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>) - ; GFX8: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s64), [[C]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s64), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX8: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX8: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX8: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[UV]], [[FNEG]] + ; GFX8: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[FADD]] + ; GFX8: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 + ; GFX8: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double 5.000000e-01 + ; GFX8: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 4607182418800017408 + ; GFX8: [[COPY1:%[0-9]+]]:_(s64) = COPY [[C2]](s64) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s64), [[C1]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX8: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV1]] - ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s64), [[C]] - ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s64), [[INTRINSIC_TRUNC1]] - ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX8: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND1]](s1), [[C1]], [[C]] - ; GFX8: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] - ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD]](s64), [[FADD1]](s64) + ; GFX8: [[FNEG1:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX8: [[FADD2:%[0-9]+]]:_(s64) = G_FADD [[UV1]], [[FNEG1]] + ; GFX8: [[FABS1:%[0-9]+]]:_(s64) = G_FABS [[FADD2]] + ; GFX8: [[COPY2:%[0-9]+]]:_(s64) = COPY [[C2]](s64) + ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s64), [[C1]] + ; GFX8: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C]] + ; GFX8: [[FADD3:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX8: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD1]](s64), [[FADD3]](s64) ; GFX8: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>) ; GFX9-LABEL: name: test_intrinsic_round_v2s64 ; GFX9: [[COPY:%[0-9]+]]:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 ; GFX9: [[UV:%[0-9]+]]:_(s64), [[UV1:%[0-9]+]]:_(s64) = G_UNMERGE_VALUES [[COPY]](<2 x s64>) - ; GFX9: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV]](s64), [[C]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV]](s64), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double -1.000000e+00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX9: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX9: [[FNEG:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX9: [[FADD:%[0-9]+]]:_(s64) = G_FADD [[UV]], [[FNEG]] + ; GFX9: [[FABS:%[0-9]+]]:_(s64) = G_FABS [[FADD]] + ; GFX9: [[C:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 + ; GFX9: [[C1:%[0-9]+]]:_(s64) = G_FCONSTANT double 5.000000e-01 + ; GFX9: [[C2:%[0-9]+]]:_(s64) = G_CONSTANT i64 4607182418800017408 + ; GFX9: [[COPY1:%[0-9]+]]:_(s64) = COPY [[C2]](s64) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s64), [[C1]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s64) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX9: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s64) = G_INTRINSIC_TRUNC [[UV1]] - ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[UV1]](s64), [[C]] - ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[UV1]](s64), [[INTRINSIC_TRUNC1]] - ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX9: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[AND1]](s1), [[C1]], [[C]] - ; GFX9: [[FADD1:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] - ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD]](s64), [[FADD1]](s64) + ; GFX9: [[FNEG1:%[0-9]+]]:_(s64) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX9: [[FADD2:%[0-9]+]]:_(s64) = G_FADD [[UV1]], [[FNEG1]] + ; GFX9: [[FABS1:%[0-9]+]]:_(s64) = G_FABS [[FADD2]] + ; GFX9: [[COPY2:%[0-9]+]]:_(s64) = COPY [[C2]](s64) + ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s64), [[C1]] + ; GFX9: [[SELECT1:%[0-9]+]]:_(s64) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C]] + ; GFX9: [[FADD3:%[0-9]+]]:_(s64) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX9: [[BUILD_VECTOR:%[0-9]+]]:_(<2 x s64>) = G_BUILD_VECTOR [[FADD1]](s64), [[FADD3]](s64) ; GFX9: $vgpr0_vgpr1_vgpr2_vgpr3 = COPY [[BUILD_VECTOR]](<2 x s64>) %0:_(<2 x s64>) = COPY $vgpr0_vgpr1_vgpr2_vgpr3 %1:_(<2 x s64>) = G_INTRINSIC_ROUND %0 @@ -332,50 +377,60 @@ ; GFX6-LABEL: name: test_intrinsic_round_s16 ; GFX6: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX6: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32) - ; GFX6: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]] ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32) + ; GFX6: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC]] ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C]](s16) - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]] - ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]] - ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C1]], [[C]] + ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG]](s16) + ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT1]], [[FPEXT2]] + ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) + ; GFX6: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC1]] + ; GFX6: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX6: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C2]](s16) + ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[FABS]](s16) + ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT3]](s32), [[FPEXT4]] + ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16) - ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] - ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) - ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC1]](s16) + ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] + ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32) + ; GFX6: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FPTRUNC2]](s16) ; GFX6: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX8-LABEL: name: test_intrinsic_round_s16 ; GFX8: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX8: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32) - ; GFX8: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] - ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16) + ; GFX8: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX8: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX8: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX8: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C2]](s16) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C1]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX8: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) ; GFX8: $vgpr0 = COPY [[ANYEXT]](s32) ; GFX9-LABEL: name: test_intrinsic_round_s16 ; GFX9: [[COPY:%[0-9]+]]:_(s32) = COPY $vgpr0 ; GFX9: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[COPY]](s32) - ; GFX9: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C1]], [[C]] - ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] - ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16) + ; GFX9: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX9: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX9: [[C:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX9: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C2]](s16) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C1]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C]] + ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) ; GFX9: $vgpr0 = COPY [[ANYEXT]](s32) %0:_(s32) = COPY $vgpr0 %1:_(s16) = G_TRUNC %0 @@ -397,40 +452,47 @@ ; GFX6: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 ; GFX6: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32) ; GFX6: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32) - ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]] ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32) + ; GFX6: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC]] ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]] - ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]] - ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] + ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG]](s16) + ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT1]], [[FPEXT2]] + ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) + ; GFX6: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC1]] + ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX6: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX6: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[FABS]](s16) + ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT3]](s32), [[FPEXT4]] + ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16) - ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] - ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) + ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] + ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32) ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT7]] - ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32) + ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32) + ; GFX6: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC3]] ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) - ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT8]](s32), [[FPEXT9]] - ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) - ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16) - ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT10]](s32), [[FPEXT11]] - ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16) + ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG1]](s16) + ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT8]], [[FPEXT9]] + ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32) + ; GFX6: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC4]] + ; GFX6: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[FABS1]](s16) + ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT10]](s32), [[FPEXT11]] + ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC3]](s16) ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT1]](s16) - ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]] - ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32) - ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16) - ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16) + ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]] + ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32) + ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC2]](s16) + ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16) ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32) ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]] ; GFX6: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32) @@ -442,22 +504,27 @@ ; GFX8: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 ; GFX8: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32) ; GFX8: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32) - ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX8: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX8: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX8: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX8: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C2]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] + ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]] - ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]] - ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]] - ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] - ; GFX8: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD]](s16) - ; GFX8: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16) + ; GFX8: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX8: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[TRUNC1]], [[FNEG1]] + ; GFX8: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FADD2]] + ; GFX8: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s16), [[C2]] + ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX8: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX8: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16) + ; GFX8: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD3]](s16) ; GFX8: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32) ; GFX8: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]] ; GFX8: [[BITCAST1:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32) @@ -469,22 +536,27 @@ ; GFX9: [[C:%[0-9]+]]:_(s32) = G_CONSTANT i32 16 ; GFX9: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST]], [[C]](s32) ; GFX9: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32) - ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX9: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX9: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX9: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C2]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] + ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]] - ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]] - ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]] - ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] - ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16) - ; GFX9: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) + ; GFX9: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX9: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[TRUNC1]], [[FNEG1]] + ; GFX9: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FADD2]] + ; GFX9: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s16), [[C2]] + ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX9: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) + ; GFX9: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD3]](s16) ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32) ; GFX9: $vgpr0 = COPY [[BUILD_VECTOR_TRUNC]](<2 x s16>) %0:_(<2 x s16>) = COPY $vgpr0 @@ -512,61 +584,71 @@ ; GFX6: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>) ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32) ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32) - ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]] ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32) + ; GFX6: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC]] ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]] - ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]] - ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] + ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG]](s16) + ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT1]], [[FPEXT2]] + ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) + ; GFX6: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC1]] + ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX6: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX6: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[FABS]](s16) + ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT3]](s32), [[FPEXT4]] + ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16) - ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] - ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) + ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] + ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32) ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT7]] - ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32) + ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32) + ; GFX6: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC3]] ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) - ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT8]](s32), [[FPEXT9]] - ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) - ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16) - ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT10]](s32), [[FPEXT11]] - ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16) + ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG1]](s16) + ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT8]], [[FPEXT9]] + ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32) + ; GFX6: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC4]] + ; GFX6: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[FABS1]](s16) + ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT10]](s32), [[FPEXT11]] + ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC3]](s16) ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT1]](s16) - ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]] - ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32) + ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]] + ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32) ; GFX6: [[FPEXT14:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16) ; GFX6: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT14]] - ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32) + ; GFX6: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32) + ; GFX6: [[FNEG2:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC6]] ; GFX6: [[FPEXT15:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16) - ; GFX6: [[FPEXT16:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT15]](s32), [[FPEXT16]] - ; GFX6: [[FPEXT17:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16) - ; GFX6: [[FPEXT18:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16) - ; GFX6: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT17]](s32), [[FPEXT18]] - ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]] - ; GFX6: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]] - ; GFX6: [[FPEXT19:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16) + ; GFX6: [[FPEXT16:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG2]](s16) + ; GFX6: [[FADD4:%[0-9]+]]:_(s32) = G_FADD [[FPEXT15]], [[FPEXT16]] + ; GFX6: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD4]](s32) + ; GFX6: [[FABS2:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC7]] + ; GFX6: [[COPY3:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT17:%[0-9]+]]:_(s32) = G_FPEXT [[FABS2]](s16) + ; GFX6: [[FPEXT18:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT17]](s32), [[FPEXT18]] + ; GFX6: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[FCMP2]](s1), [[COPY3]], [[C1]] + ; GFX6: [[FPEXT19:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16) ; GFX6: [[FPEXT20:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT2]](s16) - ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT19]], [[FPEXT20]] - ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32) - ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16) - ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16) + ; GFX6: [[FADD5:%[0-9]+]]:_(s32) = G_FADD [[FPEXT19]], [[FPEXT20]] + ; GFX6: [[FPTRUNC8:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD5]](s32) + ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC2]](s16) + ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16) ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32) ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]] ; GFX6: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32) - ; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16) - ; GFX6: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 - ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[C3]], [[C]](s32) + ; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC8]](s16) + ; GFX6: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 + ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[C4]], [[C]](s32) ; GFX6: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]] ; GFX6: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32) ; GFX6: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST2]](<2 x s16>), [[BITCAST3]](<2 x s16>) @@ -587,34 +669,41 @@ ; GFX8: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>) ; GFX8: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32) ; GFX8: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32) - ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX8: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX8: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX8: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX8: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C2]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] + ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]] - ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]] - ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]] - ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX8: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX8: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[TRUNC1]], [[FNEG1]] + ; GFX8: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FADD2]] + ; GFX8: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s16), [[C2]] + ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX8: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] ; GFX8: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]] - ; GFX8: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]] - ; GFX8: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]] - ; GFX8: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]] - ; GFX8: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] - ; GFX8: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD]](s16) - ; GFX8: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16) + ; GFX8: [[FNEG2:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC2]] + ; GFX8: [[FADD4:%[0-9]+]]:_(s16) = G_FADD [[TRUNC2]], [[FNEG2]] + ; GFX8: [[FABS2:%[0-9]+]]:_(s16) = G_FABS [[FADD4]] + ; GFX8: [[COPY3:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS2]](s16), [[C2]] + ; GFX8: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[FCMP2]](s1), [[COPY3]], [[C1]] + ; GFX8: [[FADD5:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] + ; GFX8: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16) + ; GFX8: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD3]](s16) ; GFX8: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32) ; GFX8: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]] ; GFX8: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32) - ; GFX8: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FADD2]](s16) - ; GFX8: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 - ; GFX8: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[C3]], [[C]](s32) + ; GFX8: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FADD5]](s16) + ; GFX8: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 0 + ; GFX8: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[C4]], [[C]](s32) ; GFX8: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]] ; GFX8: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32) ; GFX8: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BITCAST2]](<2 x s16>), [[BITCAST3]](<2 x s16>) @@ -635,30 +724,37 @@ ; GFX9: [[BITCAST1:%[0-9]+]]:_(s32) = G_BITCAST [[UV1]](<2 x s16>) ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32) ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32) - ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX9: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX9: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX9: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C2]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] + ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]] - ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]] - ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]] - ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX9: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX9: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[TRUNC1]], [[FNEG1]] + ; GFX9: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FADD2]] + ; GFX9: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s16), [[C2]] + ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX9: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] ; GFX9: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]] - ; GFX9: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]] - ; GFX9: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]] - ; GFX9: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]] - ; GFX9: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] - ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16) - ; GFX9: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) + ; GFX9: [[FNEG2:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC2]] + ; GFX9: [[FADD4:%[0-9]+]]:_(s16) = G_FADD [[TRUNC2]], [[FNEG2]] + ; GFX9: [[FABS2:%[0-9]+]]:_(s16) = G_FABS [[FADD4]] + ; GFX9: [[COPY3:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS2]](s16), [[C2]] + ; GFX9: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[FCMP2]](s1), [[COPY3]], [[C1]] + ; GFX9: [[FADD5:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] + ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) + ; GFX9: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD3]](s16) ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32) - ; GFX9: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD2]](s16) + ; GFX9: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD5]](s16) ; GFX9: [[DEF1:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[DEF1]](s32) ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) @@ -691,75 +787,88 @@ ; GFX6: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32) ; GFX6: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32) ; GFX6: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32) - ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX6: [[FPEXT:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) ; GFX6: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT]] ; GFX6: [[FPTRUNC:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC]](s32) + ; GFX6: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC]] ; GFX6: [[FPEXT1:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT1]](s32), [[FPEXT2]] - ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC]](s16) - ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) - ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT3]](s32), [[FPEXT4]] - ; GFX6: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] + ; GFX6: [[FPEXT2:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG]](s16) + ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT1]], [[FPEXT2]] + ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) + ; GFX6: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC1]] + ; GFX6: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX6: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX6: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX6: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT3:%[0-9]+]]:_(s32) = G_FPEXT [[FABS]](s16) + ; GFX6: [[FPEXT4:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT3]](s32), [[FPEXT4]] + ; GFX6: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] ; GFX6: [[FPEXT5:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC]](s16) ; GFX6: [[FPEXT6:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT]](s16) - ; GFX6: [[FADD:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] - ; GFX6: [[FPTRUNC1:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD]](s32) + ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT5]], [[FPEXT6]] + ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32) ; GFX6: [[FPEXT7:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) ; GFX6: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT7]] - ; GFX6: [[FPTRUNC2:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32) + ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC1]](s32) + ; GFX6: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC3]] ; GFX6: [[FPEXT8:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) - ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT8]](s32), [[FPEXT9]] - ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC1]](s16) - ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16) - ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT10]](s32), [[FPEXT11]] - ; GFX6: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC2]](s16) + ; GFX6: [[FPEXT9:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG1]](s16) + ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT8]], [[FPEXT9]] + ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32) + ; GFX6: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC4]] + ; GFX6: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT10:%[0-9]+]]:_(s32) = G_FPEXT [[FABS1]](s16) + ; GFX6: [[FPEXT11:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT10]](s32), [[FPEXT11]] + ; GFX6: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX6: [[FPEXT12:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC3]](s16) ; GFX6: [[FPEXT13:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT1]](s16) - ; GFX6: [[FADD1:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]] - ; GFX6: [[FPTRUNC3:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD1]](s32) + ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = G_FADD [[FPEXT12]], [[FPEXT13]] + ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32) ; GFX6: [[FPEXT14:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16) ; GFX6: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT14]] - ; GFX6: [[FPTRUNC4:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32) + ; GFX6: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC2]](s32) + ; GFX6: [[FNEG2:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC6]] ; GFX6: [[FPEXT15:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16) - ; GFX6: [[FPEXT16:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT15]](s32), [[FPEXT16]] - ; GFX6: [[FPEXT17:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC2]](s16) - ; GFX6: [[FPEXT18:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16) - ; GFX6: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT17]](s32), [[FPEXT18]] - ; GFX6: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]] - ; GFX6: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]] - ; GFX6: [[FPEXT19:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC4]](s16) + ; GFX6: [[FPEXT16:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG2]](s16) + ; GFX6: [[FADD4:%[0-9]+]]:_(s32) = G_FADD [[FPEXT15]], [[FPEXT16]] + ; GFX6: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD4]](s32) + ; GFX6: [[FABS2:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC7]] + ; GFX6: [[COPY3:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT17:%[0-9]+]]:_(s32) = G_FPEXT [[FABS2]](s16) + ; GFX6: [[FPEXT18:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT17]](s32), [[FPEXT18]] + ; GFX6: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[FCMP2]](s1), [[COPY3]], [[C1]] + ; GFX6: [[FPEXT19:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16) ; GFX6: [[FPEXT20:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT2]](s16) - ; GFX6: [[FADD2:%[0-9]+]]:_(s32) = G_FADD [[FPEXT19]], [[FPEXT20]] - ; GFX6: [[FPTRUNC5:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD2]](s32) + ; GFX6: [[FADD5:%[0-9]+]]:_(s32) = G_FADD [[FPEXT19]], [[FPEXT20]] + ; GFX6: [[FPTRUNC8:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD5]](s32) ; GFX6: [[FPEXT21:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16) ; GFX6: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s32) = G_INTRINSIC_TRUNC [[FPEXT21]] - ; GFX6: [[FPTRUNC6:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC3]](s32) + ; GFX6: [[FPTRUNC9:%[0-9]+]]:_(s16) = G_FPTRUNC [[INTRINSIC_TRUNC3]](s32) + ; GFX6: [[FNEG3:%[0-9]+]]:_(s16) = G_FNEG [[FPTRUNC9]] ; GFX6: [[FPEXT22:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16) - ; GFX6: [[FPEXT23:%[0-9]+]]:_(s32) = G_FPEXT [[C1]](s16) - ; GFX6: [[FCMP6:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[FPEXT22]](s32), [[FPEXT23]] - ; GFX6: [[FPEXT24:%[0-9]+]]:_(s32) = G_FPEXT [[TRUNC3]](s16) - ; GFX6: [[FPEXT25:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16) - ; GFX6: [[FCMP7:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[FPEXT24]](s32), [[FPEXT25]] - ; GFX6: [[AND3:%[0-9]+]]:_(s1) = G_AND [[FCMP6]], [[FCMP7]] - ; GFX6: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[AND3]](s1), [[C2]], [[C1]] - ; GFX6: [[FPEXT26:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC6]](s16) + ; GFX6: [[FPEXT23:%[0-9]+]]:_(s32) = G_FPEXT [[FNEG3]](s16) + ; GFX6: [[FADD6:%[0-9]+]]:_(s32) = G_FADD [[FPEXT22]], [[FPEXT23]] + ; GFX6: [[FPTRUNC10:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD6]](s32) + ; GFX6: [[FABS3:%[0-9]+]]:_(s16) = G_FABS [[FPTRUNC10]] + ; GFX6: [[COPY4:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX6: [[FPEXT24:%[0-9]+]]:_(s32) = G_FPEXT [[FABS3]](s16) + ; GFX6: [[FPEXT25:%[0-9]+]]:_(s32) = G_FPEXT [[C2]](s16) + ; GFX6: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FPEXT24]](s32), [[FPEXT25]] + ; GFX6: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[FCMP3]](s1), [[COPY4]], [[C1]] + ; GFX6: [[FPEXT26:%[0-9]+]]:_(s32) = G_FPEXT [[FPTRUNC9]](s16) ; GFX6: [[FPEXT27:%[0-9]+]]:_(s32) = G_FPEXT [[SELECT3]](s16) - ; GFX6: [[FADD3:%[0-9]+]]:_(s32) = G_FADD [[FPEXT26]], [[FPEXT27]] - ; GFX6: [[FPTRUNC7:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD3]](s32) - ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC1]](s16) - ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC3]](s16) + ; GFX6: [[FADD7:%[0-9]+]]:_(s32) = G_FADD [[FPEXT26]], [[FPEXT27]] + ; GFX6: [[FPTRUNC11:%[0-9]+]]:_(s16) = G_FPTRUNC [[FADD7]](s32) + ; GFX6: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC2]](s16) + ; GFX6: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16) ; GFX6: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32) ; GFX6: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]] ; GFX6: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32) - ; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC5]](s16) - ; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC7]](s16) + ; GFX6: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC8]](s16) + ; GFX6: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FPTRUNC11]](s16) ; GFX6: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32) ; GFX6: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]] ; GFX6: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32) @@ -777,39 +886,48 @@ ; GFX8: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32) ; GFX8: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32) ; GFX8: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32) - ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX8: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]] - ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX8: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX8: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX8: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX8: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX8: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX8: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX8: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX8: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C2]] + ; GFX8: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] + ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX8: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]] - ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]] - ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]] - ; GFX8: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX8: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX8: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[TRUNC1]], [[FNEG1]] + ; GFX8: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FADD2]] + ; GFX8: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s16), [[C2]] + ; GFX8: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX8: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] ; GFX8: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]] - ; GFX8: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]] - ; GFX8: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]] - ; GFX8: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]] - ; GFX8: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] + ; GFX8: [[FNEG2:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC2]] + ; GFX8: [[FADD4:%[0-9]+]]:_(s16) = G_FADD [[TRUNC2]], [[FNEG2]] + ; GFX8: [[FABS2:%[0-9]+]]:_(s16) = G_FABS [[FADD4]] + ; GFX8: [[COPY3:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS2]](s16), [[C2]] + ; GFX8: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[FCMP2]](s1), [[COPY3]], [[C1]] + ; GFX8: [[FADD5:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] ; GFX8: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC3]] - ; GFX8: [[FCMP6:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC3]](s16), [[C1]] - ; GFX8: [[FCMP7:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC3]](s16), [[INTRINSIC_TRUNC3]] - ; GFX8: [[AND3:%[0-9]+]]:_(s1) = G_AND [[FCMP6]], [[FCMP7]] - ; GFX8: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[AND3]](s1), [[C2]], [[C1]] - ; GFX8: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC3]], [[SELECT3]] - ; GFX8: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD]](s16) - ; GFX8: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16) + ; GFX8: [[FNEG3:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC3]] + ; GFX8: [[FADD6:%[0-9]+]]:_(s16) = G_FADD [[TRUNC3]], [[FNEG3]] + ; GFX8: [[FABS3:%[0-9]+]]:_(s16) = G_FABS [[FADD6]] + ; GFX8: [[COPY4:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX8: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS3]](s16), [[C2]] + ; GFX8: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[FCMP3]](s1), [[COPY4]], [[C1]] + ; GFX8: [[FADD7:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC3]], [[SELECT3]] + ; GFX8: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[FADD1]](s16) + ; GFX8: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[FADD3]](s16) ; GFX8: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C]](s32) ; GFX8: [[OR:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL]] ; GFX8: [[BITCAST2:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR]](s32) - ; GFX8: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FADD2]](s16) - ; GFX8: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FADD3]](s16) + ; GFX8: [[ZEXT2:%[0-9]+]]:_(s32) = G_ZEXT [[FADD5]](s16) + ; GFX8: [[ZEXT3:%[0-9]+]]:_(s32) = G_ZEXT [[FADD7]](s16) ; GFX8: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[ZEXT3]], [[C]](s32) ; GFX8: [[OR1:%[0-9]+]]:_(s32) = G_OR [[ZEXT2]], [[SHL1]] ; GFX8: [[BITCAST3:%[0-9]+]]:_(<2 x s16>) = G_BITCAST [[OR1]](s32) @@ -827,37 +945,46 @@ ; GFX9: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[BITCAST1]](s32) ; GFX9: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[BITCAST1]], [[C]](s32) ; GFX9: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR1]](s32) - ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 ; GFX9: [[INTRINSIC_TRUNC:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC]] - ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC]](s16), [[C1]] - ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC]](s16), [[INTRINSIC_TRUNC]] - ; GFX9: [[AND:%[0-9]+]]:_(s1) = G_AND [[FCMP]], [[FCMP1]] - ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xHBC00 - ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[AND]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] + ; GFX9: [[FNEG:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC]] + ; GFX9: [[FADD:%[0-9]+]]:_(s16) = G_FADD [[TRUNC]], [[FNEG]] + ; GFX9: [[FABS:%[0-9]+]]:_(s16) = G_FABS [[FADD]] + ; GFX9: [[C1:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH0000 + ; GFX9: [[C2:%[0-9]+]]:_(s16) = G_FCONSTANT half 0xH3800 + ; GFX9: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 15360 + ; GFX9: [[COPY1:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS]](s16), [[C2]] + ; GFX9: [[SELECT:%[0-9]+]]:_(s16) = G_SELECT [[FCMP]](s1), [[COPY1]], [[C1]] + ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC]], [[SELECT]] ; GFX9: [[INTRINSIC_TRUNC1:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC1]] - ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC1]](s16), [[C1]] - ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC1]](s16), [[INTRINSIC_TRUNC1]] - ; GFX9: [[AND1:%[0-9]+]]:_(s1) = G_AND [[FCMP2]], [[FCMP3]] - ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[AND1]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD1:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] + ; GFX9: [[FNEG1:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC1]] + ; GFX9: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[TRUNC1]], [[FNEG1]] + ; GFX9: [[FABS1:%[0-9]+]]:_(s16) = G_FABS [[FADD2]] + ; GFX9: [[COPY2:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP1:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS1]](s16), [[C2]] + ; GFX9: [[SELECT1:%[0-9]+]]:_(s16) = G_SELECT [[FCMP1]](s1), [[COPY2]], [[C1]] + ; GFX9: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC1]], [[SELECT1]] ; GFX9: [[INTRINSIC_TRUNC2:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC2]] - ; GFX9: [[FCMP4:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC2]](s16), [[C1]] - ; GFX9: [[FCMP5:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC2]](s16), [[INTRINSIC_TRUNC2]] - ; GFX9: [[AND2:%[0-9]+]]:_(s1) = G_AND [[FCMP4]], [[FCMP5]] - ; GFX9: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[AND2]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD2:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] + ; GFX9: [[FNEG2:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC2]] + ; GFX9: [[FADD4:%[0-9]+]]:_(s16) = G_FADD [[TRUNC2]], [[FNEG2]] + ; GFX9: [[FABS2:%[0-9]+]]:_(s16) = G_FABS [[FADD4]] + ; GFX9: [[COPY3:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP2:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS2]](s16), [[C2]] + ; GFX9: [[SELECT2:%[0-9]+]]:_(s16) = G_SELECT [[FCMP2]](s1), [[COPY3]], [[C1]] + ; GFX9: [[FADD5:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC2]], [[SELECT2]] ; GFX9: [[INTRINSIC_TRUNC3:%[0-9]+]]:_(s16) = G_INTRINSIC_TRUNC [[TRUNC3]] - ; GFX9: [[FCMP6:%[0-9]+]]:_(s1) = G_FCMP floatpred(olt), [[TRUNC3]](s16), [[C1]] - ; GFX9: [[FCMP7:%[0-9]+]]:_(s1) = G_FCMP floatpred(one), [[TRUNC3]](s16), [[INTRINSIC_TRUNC3]] - ; GFX9: [[AND3:%[0-9]+]]:_(s1) = G_AND [[FCMP6]], [[FCMP7]] - ; GFX9: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[AND3]](s1), [[C2]], [[C1]] - ; GFX9: [[FADD3:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC3]], [[SELECT3]] - ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD]](s16) - ; GFX9: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) + ; GFX9: [[FNEG3:%[0-9]+]]:_(s16) = G_FNEG [[INTRINSIC_TRUNC3]] + ; GFX9: [[FADD6:%[0-9]+]]:_(s16) = G_FADD [[TRUNC3]], [[FNEG3]] + ; GFX9: [[FABS3:%[0-9]+]]:_(s16) = G_FABS [[FADD6]] + ; GFX9: [[COPY4:%[0-9]+]]:_(s16) = COPY [[C3]](s16) + ; GFX9: [[FCMP3:%[0-9]+]]:_(s1) = G_FCMP floatpred(oge), [[FABS3]](s16), [[C2]] + ; GFX9: [[SELECT3:%[0-9]+]]:_(s16) = G_SELECT [[FCMP3]](s1), [[COPY4]], [[C1]] + ; GFX9: [[FADD7:%[0-9]+]]:_(s16) = G_FADD [[INTRINSIC_TRUNC3]], [[SELECT3]] + ; GFX9: [[ANYEXT:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD1]](s16) + ; GFX9: [[ANYEXT1:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD3]](s16) ; GFX9: [[BUILD_VECTOR_TRUNC:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT]](s32), [[ANYEXT1]](s32) - ; GFX9: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD2]](s16) - ; GFX9: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD3]](s16) + ; GFX9: [[ANYEXT2:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD5]](s16) + ; GFX9: [[ANYEXT3:%[0-9]+]]:_(s32) = G_ANYEXT [[FADD7]](s16) ; GFX9: [[BUILD_VECTOR_TRUNC1:%[0-9]+]]:_(<2 x s16>) = G_BUILD_VECTOR_TRUNC [[ANYEXT2]](s32), [[ANYEXT3]](s32) ; GFX9: [[CONCAT_VECTORS:%[0-9]+]]:_(<4 x s16>) = G_CONCAT_VECTORS [[BUILD_VECTOR_TRUNC]](<2 x s16>), [[BUILD_VECTOR_TRUNC1]](<2 x s16>) ; GFX9: $vgpr0_vgpr1 = COPY [[CONCAT_VECTORS]](<4 x s16>) Index: llvm/unittests/CodeGen/GlobalISel/LegalizerHelperTest.cpp =================================================================== --- llvm/unittests/CodeGen/GlobalISel/LegalizerHelperTest.cpp +++ llvm/unittests/CodeGen/GlobalISel/LegalizerHelperTest.cpp @@ -2341,4 +2341,36 @@ // Check EXPECT_TRUE(CheckMachineFunction(*MF, CheckStr)) << *MF; } + +// Test lowering of G_FFLOOR +TEST_F(GISelMITest, LowerFFloor) { + setUp(); + if (!TM) + return; + + // Declare your legalization info + DefineLegalizerInfo(A, {}); + // Build Instr + auto Floor = B.buildFFloor(LLT::scalar(64), Copies[0], MachineInstr::MIFlag::FmNoInfs); + AInfo Info(MF->getSubtarget()); + DummyGISelObserver Observer; + LegalizerHelper Helper(*MF, Info, Observer, B); + // Perform Legalization + EXPECT_EQ(LegalizerHelper::LegalizeResult::Legalized, + Helper.lower(*Floor, 0, LLT())); + + auto CheckStr = R"( + CHECK: [[COPY:%[0-9]+]]:_(s64) = COPY + CHECK: [[TRUNC:%[0-9]+]]:_(s64) = ninf G_INTRINSIC_TRUNC [[COPY]] + CHECK: [[ZERO:%[0-9]+]]:_(s64) = G_FCONSTANT double 0.000000e+00 + CHECK: [[CMP0:%[0-9]+]]:_(s1) = ninf G_FCMP floatpred(olt), [[COPY]]:_(s64), [[ZERO]]:_ + CHECK: [[CMP1:%[0-9]+]]:_(s1) = ninf G_FCMP floatpred(one), [[COPY]]:_(s64), [[TRUNC]]:_ + CHECK: [[AND:%[0-9]+]]:_(s1) = G_AND [[CMP0]]:_, [[CMP1]]:_ + CHECK: [[ITOFP:%[0-9]+]]:_(s64) = G_SITOFP [[AND]] + = ninf G_FADD [[TRUNC]]:_, [[ITOFP]]:_ + )"; + + // Check + EXPECT_TRUE(CheckMachineFunction(*MF, CheckStr)) << *MF; +} } // namespace