From 6c72c3551978ef2eabbe9727bf24fd2fcf385318 Mon Sep 17 00:00:00 2001 From: Fernando Endo Date: Sat, 15 Oct 2016 14:58:45 -0500 Subject: cpu, arm: Distinguish Float* and SimdFloat*, create FloatMem* opClass Modify the opClass assigned to AArch64 FP instructions from SimdFloat* to Float*. Also create the FloatMemRead and FloatMemWrite opClasses, which distinguishes writes to the INT and FP register banks. Change the latency of (Simd)FloatMultAcc to 5, based on the Cortex-A72, where the "latency" of FMADD is 3 if the next instruction is a FMADD and has only the augend to destination dependency, otherwise it's 7 cycles. Signed-off-by: Jason Lowe-Power --- src/arch/arm/isa/insts/fp64.isa | 91 +++++++++++++++++++++-------------------- src/arch/isa_parser.py | 16 +++++++- 2 files changed, 60 insertions(+), 47 deletions(-) (limited to 'src/arch') diff --git a/src/arch/arm/isa/insts/fp64.isa b/src/arch/arm/isa/insts/fp64.isa index 95dec5062..706f77fb0 100644 --- a/src/arch/arm/isa/insts/fp64.isa +++ b/src/arch/arm/isa/insts/fp64.isa @@ -52,7 +52,7 @@ let {{ ''' fmovImmSIop = InstObjParams("fmov", "FmovImmS", "FpRegImmOp", { "code": fmovImmSCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegImmOpDeclare.subst(fmovImmSIop); decoder_output += FpRegImmOpConstructor.subst(fmovImmSIop); exec_output += BasicExecute.subst(fmovImmSIop); @@ -65,7 +65,7 @@ let {{ ''' fmovImmDIop = InstObjParams("fmov", "FmovImmD", "FpRegImmOp", { "code": fmovImmDCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegImmOpDeclare.subst(fmovImmDIop); decoder_output += AA64FpRegImmOpConstructor.subst(fmovImmDIop); exec_output += BasicExecute.subst(fmovImmDIop); @@ -78,7 +78,7 @@ let {{ ''' fmovRegSIop = InstObjParams("fmov", "FmovRegS", "FpRegRegOp", { "code": fmovRegSCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovRegSIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovRegSIop); exec_output += BasicExecute.subst(fmovRegSIop); @@ -91,7 +91,7 @@ let {{ ''' fmovRegDIop = InstObjParams("fmov", "FmovRegD", "FpRegRegOp", { "code": fmovRegDCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovRegDIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovRegDIop); exec_output += BasicExecute.subst(fmovRegDIop); @@ -104,7 +104,7 @@ let {{ ''' fmovCoreRegWIop = InstObjParams("fmov", "FmovCoreRegW", "FpRegRegOp", { "code": fmovCoreRegWCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovCoreRegWIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovCoreRegWIop); exec_output += BasicExecute.subst(fmovCoreRegWIop); @@ -117,7 +117,7 @@ let {{ ''' fmovCoreRegXIop = InstObjParams("fmov", "FmovCoreRegX", "FpRegRegOp", { "code": fmovCoreRegXCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovCoreRegXIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovCoreRegXIop); exec_output += BasicExecute.subst(fmovCoreRegXIop); @@ -128,7 +128,7 @@ let {{ ''' fmovUCoreRegXIop = InstObjParams("fmov", "FmovUCoreRegX", "FpRegRegOp", { "code": fmovUCoreRegXCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovUCoreRegXIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovUCoreRegXIop); exec_output += BasicExecute.subst(fmovUCoreRegXIop); @@ -138,7 +138,7 @@ let {{ ''' fmovRegCoreWIop = InstObjParams("fmov", "FmovRegCoreW", "FpRegRegOp", { "code": fmovRegCoreWCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovRegCoreWIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovRegCoreWIop); exec_output += BasicExecute.subst(fmovRegCoreWIop); @@ -148,7 +148,7 @@ let {{ ''' fmovRegCoreXIop = InstObjParams("fmov", "FmovRegCoreX", "FpRegRegOp", { "code": fmovRegCoreXCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovRegCoreXIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovRegCoreXIop); exec_output += BasicExecute.subst(fmovRegCoreXIop); @@ -158,7 +158,7 @@ let {{ ''' fmovURegCoreXIop = InstObjParams("fmov", "FmovURegCoreX", "FpRegRegOp", { "code": fmovURegCoreXCode, - "op_class": "SimdFloatMiscOp" }, []) + "op_class": "FloatMiscOp" }, []) header_output += FpRegRegOpDeclare.subst(fmovURegCoreXIop); decoder_output += AA64FpRegRegOpConstructor.subst(fmovURegCoreXIop); exec_output += BasicExecute.subst(fmovURegCoreXIop); @@ -270,16 +270,16 @@ let {{ decoder_output += AA64FpRegRegRegRegOpConstructor.subst(iop) exec_output += BasicExecute.subst(iop) - buildTernaryFpOp("FMAdd", "SimdFloatMultAccOp", + buildTernaryFpOp("FMAdd", "FloatMultAccOp", "fplibMulAdd(cOp3, cOp1, cOp2, fpscr)", "fplibMulAdd(cOp3, cOp1, cOp2, fpscr)" ) - buildTernaryFpOp("FMSub", "SimdFloatMultAccOp", + buildTernaryFpOp("FMSub", "FloatMultAccOp", "fplibMulAdd(cOp3, fplibNeg(cOp1), cOp2, fpscr)", "fplibMulAdd(cOp3, fplibNeg(cOp1), cOp2, fpscr)" ) - buildTernaryFpOp("FNMAdd", "SimdFloatMultAccOp", + buildTernaryFpOp("FNMAdd", "FloatMultAccOp", "fplibMulAdd(fplibNeg(cOp3), fplibNeg(cOp1), cOp2, fpscr)", "fplibMulAdd(fplibNeg(cOp3), fplibNeg(cOp1), cOp2, fpscr)" ) - buildTernaryFpOp("FNMSub", "SimdFloatMultAccOp", + buildTernaryFpOp("FNMSub", "FloatMultAccOp", "fplibMulAdd(fplibNeg(cOp3), cOp1, cOp2, fpscr)", "fplibMulAdd(fplibNeg(cOp3), cOp1, cOp2, fpscr)" ) @@ -304,31 +304,31 @@ let {{ decoder_output += constructorTempl.subst(iop) exec_output += BasicExecute.subst(iop) - buildBinFpOp("fadd", "FAdd", "FpRegRegRegOp", "SimdFloatAddOp", + buildBinFpOp("fadd", "FAdd", "FpRegRegRegOp", "FloatAddOp", "fplibAdd(cOp1, cOp2, fpscr)", "fplibAdd(cOp1, cOp2, fpscr)") - buildBinFpOp("fsub", "FSub", "FpRegRegRegOp", "SimdFloatAddOp", + buildBinFpOp("fsub", "FSub", "FpRegRegRegOp", "FloatAddOp", "fplibSub(cOp1, cOp2, fpscr)", "fplibSub(cOp1, cOp2, fpscr)") - buildBinFpOp("fdiv", "FDiv", "FpRegRegRegOp", "SimdFloatDivOp", + buildBinFpOp("fdiv", "FDiv", "FpRegRegRegOp", "FloatDivOp", "fplibDiv(cOp1, cOp2, fpscr)", "fplibDiv(cOp1, cOp2, fpscr)") - buildBinFpOp("fmul", "FMul", "FpRegRegRegOp", "SimdFloatMultOp", + buildBinFpOp("fmul", "FMul", "FpRegRegRegOp", "FloatMultOp", "fplibMul(cOp1, cOp2, fpscr)", "fplibMul(cOp1, cOp2, fpscr)") - buildBinFpOp("fnmul", "FNMul", "FpRegRegRegOp", "SimdFloatMultOp", + buildBinFpOp("fnmul", "FNMul", "FpRegRegRegOp", "FloatMultOp", "fplibNeg(fplibMul(cOp1, cOp2, fpscr))", "fplibNeg(fplibMul(cOp1, cOp2, fpscr))") - buildBinFpOp("fmin", "FMin", "FpRegRegRegOp", "SimdFloatCmpOp", + buildBinFpOp("fmin", "FMin", "FpRegRegRegOp", "FloatCmpOp", "fplibMin(cOp1, cOp2, fpscr)", "fplibMin(cOp1, cOp2, fpscr)") - buildBinFpOp("fmax", "FMax", "FpRegRegRegOp", "SimdFloatCmpOp", + buildBinFpOp("fmax", "FMax", "FpRegRegRegOp", "FloatCmpOp", "fplibMax(cOp1, cOp2, fpscr)", "fplibMax(cOp1, cOp2, fpscr)") - buildBinFpOp("fminnm", "FMinNM", "FpRegRegRegOp", "SimdFloatCmpOp", + buildBinFpOp("fminnm", "FMinNM", "FpRegRegRegOp", "FloatCmpOp", "fplibMinNum(cOp1, cOp2, fpscr)", "fplibMinNum(cOp1, cOp2, fpscr)") - buildBinFpOp("fmaxnm", "FMaxNM", "FpRegRegRegOp", "SimdFloatCmpOp", + buildBinFpOp("fmaxnm", "FMaxNM", "FpRegRegRegOp", "FloatCmpOp", "fplibMaxNum(cOp1, cOp2, fpscr)", "fplibMaxNum(cOp1, cOp2, fpscr)") @@ -354,7 +354,7 @@ let {{ decoder_output += constructorTempl.subst(iop) exec_output += BasicExecute.subst(iop) - buildUnaryFpOp("fsqrt", "FSqrt", "FpRegRegOp", "SimdFloatSqrtOp", + buildUnaryFpOp("fsqrt", "FSqrt", "FpRegRegOp", "FloatSqrtOp", "fplibSqrt(cOp1, fpscr)", "fplibSqrt(cOp1, fpscr)") def buildSimpleUnaryFpOp(name, Name, base, opClass, singleOp, @@ -383,29 +383,29 @@ let {{ decoder_output += constructorTempl.subst(iop) exec_output += BasicExecute.subst(iop) - buildSimpleUnaryFpOp("fneg", "FNeg", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("fneg", "FNeg", "FpRegRegOp", "FloatMiscOp", "fplibNeg(cOp1)", "fplibNeg(cOp1)") - buildSimpleUnaryFpOp("fabs", "FAbs", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("fabs", "FAbs", "FpRegRegOp", "FloatMiscOp", "fplibAbs(cOp1)", "fplibAbs(cOp1)") - buildSimpleUnaryFpOp("frintn", "FRIntN", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("frintn", "FRIntN", "FpRegRegOp", "FloatMiscOp", "fplibRoundInt(cOp1, FPRounding_TIEEVEN, false, fpscr)", "fplibRoundInt(cOp1, FPRounding_TIEEVEN, false, fpscr)") - buildSimpleUnaryFpOp("frintp", "FRIntP", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("frintp", "FRIntP", "FpRegRegOp", "FloatMiscOp", "fplibRoundInt(cOp1, FPRounding_POSINF, false, fpscr)", "fplibRoundInt(cOp1, FPRounding_POSINF, false, fpscr)") - buildSimpleUnaryFpOp("frintm", "FRIntM", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("frintm", "FRIntM", "FpRegRegOp", "FloatMiscOp", "fplibRoundInt(cOp1, FPRounding_NEGINF, false, fpscr)", "fplibRoundInt(cOp1, FPRounding_NEGINF, false, fpscr)") - buildSimpleUnaryFpOp("frintz", "FRIntZ", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("frintz", "FRIntZ", "FpRegRegOp", "FloatMiscOp", "fplibRoundInt(cOp1, FPRounding_ZERO, false, fpscr)", "fplibRoundInt(cOp1, FPRounding_ZERO, false, fpscr)") - buildSimpleUnaryFpOp("frinta", "FRIntA", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("frinta", "FRIntA", "FpRegRegOp", "FloatMiscOp", "fplibRoundInt(cOp1, FPRounding_TIEAWAY, false, fpscr)", "fplibRoundInt(cOp1, FPRounding_TIEAWAY, false, fpscr)") - buildSimpleUnaryFpOp("frinti", "FRIntI", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("frinti", "FRIntI", "FpRegRegOp", "FloatMiscOp", "fplibRoundInt(cOp1, FPCRRounding(fpscr), false, fpscr)", "fplibRoundInt(cOp1, FPCRRounding(fpscr), false, fpscr)") - buildSimpleUnaryFpOp("frintx", "FRIntX", "FpRegRegOp", "SimdFloatMiscOp", + buildSimpleUnaryFpOp("frintx", "FRIntX", "FpRegRegOp", "FloatMiscOp", "fplibRoundInt(cOp1, FPCRRounding(fpscr), true, fpscr)", "fplibRoundInt(cOp1, FPCRRounding(fpscr), true, fpscr)") }}; @@ -451,8 +451,8 @@ let {{ instName = "Fcvt%s%sIntFp%s" %(regL, us, "D" if isDouble else "S") mnem = "%scvtf" %(us.lower()) fcvtIntFpDIop = InstObjParams(mnem, instName, "FpRegRegOp", - { "code": fcvtIntFpDCode, - "op_class": "SimdFloatCvtOp" }, []) + { "code": fcvtIntFpDCode, + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegOpDeclare.subst(fcvtIntFpDIop); decoder_output += AA64FpRegRegOpConstructor.subst(fcvtIntFpDIop); exec_output += BasicExecute.subst(fcvtIntFpDIop); @@ -491,7 +491,7 @@ let {{ mnem = "fcvt%s%s" %(rmode, "s" if isSigned else "u") fcvtFpIntIop = InstObjParams(mnem, instName, "FpRegRegOp", { "code": fcvtFpIntCode, - "op_class": "SimdFloatCvtOp" }, []) + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegOpDeclare.subst(fcvtFpIntIop); decoder_output += FpRegRegOpConstructor.subst(fcvtFpIntIop); exec_output += BasicExecute.subst(fcvtFpIntIop); @@ -514,7 +514,7 @@ let {{ ''' fcvtFpSFpDIop = InstObjParams("fcvt", "FCvtFpSFpD", "FpRegRegOp", { "code": fcvtFpSFpDCode, - "op_class": "SimdFloatCvtOp" }, []) + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegOpDeclare.subst(fcvtFpSFpDIop); decoder_output += AA64FpRegRegOpConstructor.subst(fcvtFpSFpDIop); exec_output += BasicExecute.subst(fcvtFpSFpDIop); @@ -531,7 +531,7 @@ let {{ ''' fcvtFpDFpSIop = InstObjParams("fcvt", "FcvtFpDFpS", "FpRegRegOp", {"code": fcvtFpDFpSCode, - "op_class": "SimdFloatCvtOp" }, []) + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegOpDeclare.subst(fcvtFpDFpSIop); decoder_output += AA64FpRegRegOpConstructor.subst(fcvtFpDFpSIop); exec_output += BasicExecute.subst(fcvtFpDFpSIop); @@ -563,7 +563,7 @@ let {{ instName = "FcvtFpHFp%s" %("D" if isDouble else "S") fcvtFpHFpIop = InstObjParams("fcvt", instName, "FpRegRegOp", { "code": code, - "op_class": "SimdFloatCvtOp" }, []) + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegOpDeclare.subst(fcvtFpHFpIop); decoder_output += AA64FpRegRegOpConstructor.subst(fcvtFpHFpIop); exec_output += BasicExecute.subst(fcvtFpHFpIop); @@ -586,7 +586,7 @@ let {{ instName = "FcvtFp%sFpH" %("D" if isDouble else "S") fcvtFpFpHIop = InstObjParams("fcvt", instName, "FpRegRegOp", { "code": code, - "op_class": "SimdFloatCvtOp" }, []) + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegOpDeclare.subst(fcvtFpFpHIop); decoder_output += AA64FpRegRegOpConstructor.subst(fcvtFpFpHIop); exec_output += BasicExecute.subst(fcvtFpFpHIop); @@ -626,7 +626,7 @@ let {{ fcmpIop = InstObjParams("fcmp%s" %("" if isQuiet else "e"), instName, "FpReg%sOp" %(typeName), {"code": fcmpCode, - "op_class": "SimdFloatCmpOp"}, []) + "op_class": "FloatCmpOp"}, []) declareTemp = eval("FpReg%sOpDeclare" %(typeName)); constructorTemp = eval("AA64FpReg%sOpConstructor" %(typeName)); @@ -673,7 +673,7 @@ let {{ fccmpIop = InstObjParams("fccmp%s" %("" if isQuiet else "e"), instName, "FpCondCompRegOp", {"code": fccmpCode, - "op_class": "SimdFloatCmpOp"}, []) + "op_class": "FloatCmpOp"}, []) header_output += DataXCondCompRegDeclare.subst(fccmpIop); decoder_output += DataXCondCompRegConstructor.subst(fccmpIop); exec_output += BasicExecute.subst(fccmpIop); @@ -718,7 +718,7 @@ let {{ mnem = "fcvtz%s" %("s" if isSigned else "u") fcvtFpFixedIop = InstObjParams(mnem, instName, "FpRegRegImmOp", { "code": fcvtFpFixedCode, - "op_class": "SimdFloatCvtOp" }, []) + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegImmOpDeclare.subst(fcvtFpFixedIop); decoder_output += AA64FpRegRegImmOpConstructor.subst(fcvtFpFixedIop); exec_output += BasicExecute.subst(fcvtFpFixedIop); @@ -759,7 +759,7 @@ let {{ mnem = "%scvtf" %("s" if isSigned else "u") fcvtFixedFpIop = InstObjParams(mnem, instName, "FpRegRegImmOp", { "code": fcvtFixedFpCode, - "op_class": "SimdFloatCvtOp" }, []) + "op_class": "FloatCvtOp" }, []) header_output += FpRegRegImmOpDeclare.subst(fcvtFixedFpIop); decoder_output += FpRegRegImmOpConstructor.subst(fcvtFixedFpIop); exec_output += BasicExecute.subst(fcvtFixedFpIop); @@ -804,7 +804,8 @@ let {{ ''' iop = InstObjParams("fcsel", "FCSel%s" %("D" if isDouble else "S"), - "FpCondSelOp", code) + "FpCondSelOp", { "code": code, + "op_class": "FloatCvtOp" }) header_output += DataXCondSelDeclare.subst(iop) decoder_output += DataXCondSelConstructor.subst(iop) exec_output += BasicExecute.subst(iop) diff --git a/src/arch/isa_parser.py b/src/arch/isa_parser.py index 749eaf88d..8d609ae5f 100755 --- a/src/arch/isa_parser.py +++ b/src/arch/isa_parser.py @@ -1130,9 +1130,21 @@ class InstObjParams(object): # These are good enough for most cases. if not self.op_class: if 'IsStore' in self.flags: - self.op_class = 'MemWriteOp' + # The order matters here: 'IsFloating' and 'IsInteger' are + # usually set in FP instructions because of the base + # register + if 'IsFloating' in self.flags: + self.op_class = 'FloatMemWriteOp' + else: + self.op_class = 'MemWriteOp' elif 'IsLoad' in self.flags or 'IsPrefetch' in self.flags: - self.op_class = 'MemReadOp' + # The order matters here: 'IsFloating' and 'IsInteger' are + # usually set in FP instructions because of the base + # register + if 'IsFloating' in self.flags: + self.op_class = 'FloatMemReadOp' + else: + self.op_class = 'MemReadOp' elif 'IsFloating' in self.flags: self.op_class = 'FloatAddOp' else: -- cgit v1.2.3