Improvements to int->float cast code for PPC-64

llvm-svn: 16105
This commit is contained in:
Nate Begeman 2004-08-29 22:02:43 +00:00
parent 50d140ee3d
commit 8cb6bd5f3c
3 changed files with 60 additions and 61 deletions

View File

@ -2315,68 +2315,60 @@ void ISel::emitCastOperation(MachineBasicBlock *MBB,
// Handle casts from integer to floating point now... // Handle casts from integer to floating point now...
if (DestClass == cFP32 || DestClass == cFP64) { if (DestClass == cFP32 || DestClass == cFP64) {
// Emit a library call for long to float conversion
if (SrcClass == cLong) {
std::vector<ValueRecord> Args;
Args.push_back(ValueRecord(SrcReg, SrcTy));
Function *floatFn = (DestClass == cFP32) ? __floatdisfFn : __floatdidfFn;
MachineInstr *TheCall =
BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(floatFn, true);
doCall(ValueRecord(DestReg, DestTy), TheCall, Args, false);
return;
}
// Make sure we're dealing with a full 32 bits
unsigned TmpReg = makeAnotherReg(Type::IntTy);
promote32(TmpReg, ValueRecord(SrcReg, SrcTy));
SrcReg = TmpReg;
// Spill the integer to memory and reload it from there. // Spill the integer to memory and reload it from there.
// Also spill room for a special conversion constant unsigned TmpReg = makeAnotherReg(Type::DoubleTy);
int ConstantFrameIndex =
F->getFrameInfo()->CreateStackObject(Type::DoubleTy, TM.getTargetData());
int ValueFrameIdx = int ValueFrameIdx =
F->getFrameInfo()->CreateStackObject(Type::DoubleTy, TM.getTargetData()); F->getFrameInfo()->CreateStackObject(Type::DoubleTy, TM.getTargetData());
unsigned constantHi = makeAnotherReg(Type::IntTy); if (SrcClass == cLong) {
unsigned constantLo = makeAnotherReg(Type::IntTy); if (SrcTy->isSigned()) {
unsigned ConstF = makeAnotherReg(Type::DoubleTy); addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(SrcReg),
unsigned TempF = makeAnotherReg(Type::DoubleTy); ValueFrameIdx);
addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, TmpReg),
if (!SrcTy->isSigned()) { ValueFrameIdx);
BuildMI(*BB, IP, PPC::LIS, 1, constantHi).addSImm(0x4330); BuildMI(*MBB, IP, PPC::FCFID, 1, DestReg).addReg(TmpReg);
BuildMI(*BB, IP, PPC::LI, 1, constantLo).addSImm(0); } else {
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(constantHi), unsigned Scale = getReg(ConstantFP::get(Type::DoubleTy, 0x1p32));
ConstantFrameIndex); unsigned TmpHi = makeAnotherReg(Type::IntTy);
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(constantLo), unsigned TmpLo = makeAnotherReg(Type::IntTy);
ConstantFrameIndex, 4); unsigned FPLow = makeAnotherReg(Type::DoubleTy);
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(constantHi), unsigned FPTmpHi = makeAnotherReg(Type::DoubleTy);
ValueFrameIdx); unsigned FPTmpLo = makeAnotherReg(Type::DoubleTy);
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(SrcReg), int OtherFrameIdx = F->getFrameInfo()->CreateStackObject(Type::DoubleTy,
ValueFrameIdx, 4); TM.getTargetData());
addFrameReference(BuildMI(*BB, IP, PPC::LFD, 2, ConstF), BuildMI(*MBB, IP, PPC::RLDICL, 3, TmpHi).addReg(SrcReg).addImm(32)
ConstantFrameIndex); .addImm(32);
addFrameReference(BuildMI(*BB, IP, PPC::LFD, 2, TempF), ValueFrameIdx); BuildMI(*MBB, IP, PPC::RLDICL, 3, TmpLo).addReg(SrcReg).addImm(0)
BuildMI(*BB, IP, PPC::FSUB, 2, DestReg).addReg(TempF).addReg(ConstF); .addImm(32);
} else { addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(TmpHi),
unsigned TempLo = makeAnotherReg(Type::IntTy); ValueFrameIdx);
BuildMI(*BB, IP, PPC::LIS, 1, constantHi).addSImm(0x4330); addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(TmpLo),
BuildMI(*BB, IP, PPC::LIS, 1, constantLo).addSImm(0x8000); OtherFrameIdx);
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(constantHi), addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, TmpReg),
ConstantFrameIndex); ValueFrameIdx);
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(constantLo), addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, FPLow),
ConstantFrameIndex, 4); OtherFrameIdx);
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(constantHi), BuildMI(*MBB, IP, PPC::FCFID, 1, FPTmpHi).addReg(TmpReg);
ValueFrameIdx); BuildMI(*MBB, IP, PPC::FCFID, 1, FPTmpLo).addReg(FPLow);
BuildMI(*BB, IP, PPC::XORIS, 2, TempLo).addReg(SrcReg).addImm(0x8000); BuildMI(*MBB, IP, PPC::FMADD, 3, DestReg).addReg(Scale).addReg(FPTmpHi)
addFrameReference(BuildMI(*BB, IP, PPC::STW, 3).addReg(TempLo), .addReg(FPTmpLo);
ValueFrameIdx, 4); }
addFrameReference(BuildMI(*BB, IP, PPC::LFD, 2, ConstF), return;
ConstantFrameIndex);
addFrameReference(BuildMI(*BB, IP, PPC::LFD, 2, TempF), ValueFrameIdx);
BuildMI(*BB, IP, PPC::FSUB, 2, DestReg).addReg(TempF).addReg(ConstF);
} }
// FIXME: really want a promote64
unsigned IntTmp = makeAnotherReg(Type::IntTy);
if (SrcTy->isSigned())
BuildMI(*MBB, IP, PPC::EXTSW, 1, IntTmp).addReg(SrcReg);
else
BuildMI(*MBB, IP, PPC::RLDICL, 3, IntTmp).addReg(SrcReg).addImm(0)
.addImm(32);
addFrameReference(BuildMI(*MBB, IP, PPC::STD, 3).addReg(IntTmp),
ValueFrameIdx);
addFrameReference(BuildMI(*MBB, IP, PPC::LFD, 2, TmpReg),
ValueFrameIdx);
BuildMI(*MBB, IP, PPC::FCFID, 1, DestReg).addReg(TmpReg);
return; return;
} }

View File

@ -104,6 +104,7 @@ def FADD : AForm_2<"fadd", 63, 21, 0, 0, 0>;
def FADDS : AForm_2<"fadds", 59, 21, 0, 0, 0>; def FADDS : AForm_2<"fadds", 59, 21, 0, 0, 0>;
def FSUB : AForm_2<"fsub", 63, 20, 0, 0, 0>; def FSUB : AForm_2<"fsub", 63, 20, 0, 0, 0>;
def FSUBS : AForm_2<"fsubs", 59, 20, 0, 0, 0>; def FSUBS : AForm_2<"fsubs", 59, 20, 0, 0, 0>;
def FMADD : AForm_2<"fmul", 63, 29, 0, 0, 0>;
def FMUL : AForm_3<"fmul", 63, 25, 0, 0, 0>; def FMUL : AForm_3<"fmul", 63, 25, 0, 0, 0>;
def FMULS : AForm_3<"fmuls", 59, 25, 0, 0, 0>; def FMULS : AForm_3<"fmuls", 59, 25, 0, 0, 0>;
def FDIV : AForm_2<"fdiv", 63, 18, 0, 0, 0>; def FDIV : AForm_2<"fdiv", 63, 18, 0, 0, 0>;
@ -232,18 +233,24 @@ def EXTSB : XForm_11<31, 954, 0, 0, 0, (ops GPRC:$rA, GPRC:$rS),
"extsb $rA, $rS">; "extsb $rA, $rS">;
def EXTSH : XForm_11<31, 922, 0, 0, 0, (ops GPRC:$rA, GPRC:$rS), def EXTSH : XForm_11<31, 922, 0, 0, 0, (ops GPRC:$rA, GPRC:$rS),
"extsh $rA, $rS">; "extsh $rA, $rS">;
def EXTSW : XForm_11<31, 986, 0, 1, 0, (ops GPRC:$rA, GPRC:$rS),
"extsw $rA, $rS">;
def LFSX : XForm_25<31, 535, 0, 0, (ops FPRC:$dst, GPRC:$base, GPRC:$index), def LFSX : XForm_25<31, 535, 0, 0, (ops FPRC:$dst, GPRC:$base, GPRC:$index),
"lfsx $dst, $base, $index">; "lfsx $dst, $base, $index">;
def LFDX : XForm_25<31, 599, 0, 0, (ops FPRC:$dst, GPRC:$base, GPRC:$index), def LFDX : XForm_25<31, 599, 0, 0, (ops FPRC:$dst, GPRC:$base, GPRC:$index),
"lfdx $dst, $base, $index">; "lfdx $dst, $base, $index">;
def FCFID : XForm_26<63, 846, 0, 1, 0, (ops FPRC:$frD, FPRC:$frB),
"fcfid $frD, $frB">;
def FCTIDZ : XForm_26<63, 815, 0, 1, 0, (ops FPRC:$frD, FPRC:$frB),
"fctidz $frD, $frB">;
def FCTIWZ : XForm_26<63, 15, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB),
"fctiwz $frD, $frB">;
def FMR : XForm_26<63, 72, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB), def FMR : XForm_26<63, 72, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB),
"fmr $frD, $frB">; "fmr $frD, $frB">;
def FNEG : XForm_26<63, 80, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB), def FNEG : XForm_26<63, 80, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB),
"fneg $frD, $frB">; "fneg $frD, $frB">;
def FRSP : XForm_26<63, 12, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB), def FRSP : XForm_26<63, 12, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB),
"frsp $frD, $frB">; "frsp $frD, $frB">;
def FCTIWZ : XForm_26<63, 15, 0, 0, 0, (ops FPRC:$frD, FPRC:$frB),
"fctiwz $frD, $frB">;
def STFSX : XForm_28<31, 663, 0, 0, (ops FPRC:$frS, GPRC:$rA, GPRC:$rB), def STFSX : XForm_28<31, 663, 0, 0, (ops FPRC:$frS, GPRC:$rA, GPRC:$rB),
"stfsx $frS, $rA, $rB">; "stfsx $frS, $rA, $rB">;
def STFDX : XForm_28<31, 727, 0, 0, (ops FPRC:$frS, GPRC:$rA, GPRC:$rB), def STFDX : XForm_28<31, 727, 0, 0, (ops FPRC:$frS, GPRC:$rA, GPRC:$rB),

View File

@ -1,10 +1,10 @@
TODO: TODO:
* switch to auto-generated asm writer * switch to auto-generated asm writer
* use stfiwx in float->int * fix rlwimi generation to be use-and-def
* implement scheduling info * implement scheduling info
* implement powerpc-64 for darwin * implement powerpc-64 for darwin
* implement powerpc-64 for aix * implement powerpc-64 for aix
* fix rlwimi generation to be use-and-def * use stfiwx in float->int
* should hint to the branch select pass that it doesn't need to print the * should hint to the branch select pass that it doesn't need to print the
second unconditional branch, so we don't end up with things like: second unconditional branch, so we don't end up with things like:
b .LBBl42__2E_expand_function_8_674 ; loopentry.24 b .LBBl42__2E_expand_function_8_674 ; loopentry.24