[llvm-commits] [llvm] r41967 - in /llvm/trunk: include/llvm/ADT/APInt.h include/llvm/ADT/FoldingSet.h include/llvm/Constants.h lib/CodeGen/SelectionDAG/LegalizeDAG.cpp lib/CodeGen/SelectionDAG/SelectionDAG.cpp lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp lib/Support/APInt.cpp lib/Support/FoldingSet.cpp lib/Target/CBackend/CBackend.cpp lib/Target/X86/X86ISelLowering.cpp lib/Target/X86/X86RegisterInfo.cpp lib/Transforms/Scalar/InstructionCombining.cpp lib/VMCore/Constants.cpp

[Dale Johannesen]

On Sep 14, 2007, at 11:12 PM, Chris Lattner wrote:
>> @@ -4722,9 +4737,11 @@
>>      if (DestVT == MVT::f64) {
>>        // do nothing
>>        Result = Sub;
>> -    } else {
>> +    } else if (DestVT == MVT::f32) {
>>       // if f32 then cast to f32
>>        Result = DAG.getNode(ISD::FP_ROUND, MVT::f32, Sub);
>> +    } else if (DestVT == MVT::f80) {
>> +      Result = DAG.getNode(ISD::FP_EXTEND, MVT::f80, Sub);
>>      }
>
> How about "if DestVT < f32 use round.  If DestVT > f64, use  
> FP_EXTEND"?

Comparisons other than equality against an enum?  Ick.  IMO, that  
sort of thing
is a bad misuse of enums.  Obviously you disagree, we should talk?

>> ===================================================================== 
>> =
>> ========
>> --- llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp (original)
>> +++ llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp Fri Sep 14
>> 17:26:36 2007
>> @@ -3724,9 +3714,15 @@
>>    if (const ConstantSDNode *CSDN = dyn_cast<ConstantSDNode>(this)) {
>>      cerr << "<" << CSDN->getValue() << ">";
>>    } else if (const ConstantFPSDNode *CSDN =
>> dyn_cast<ConstantFPSDNode>(this)) {
>> -    cerr << "<" << (&CSDN->getValueAPF().getSemantics()
>> ==&APFloat::IEEEsingle ?
>> -                    CSDN->getValueAPF().convertToFloat() :
>> -                    CSDN->getValueAPF().convertToDouble()) << ">";
>> +    if (&CSDN->getValueAPF().getSemantics()==&APFloat::IEEEsingle)
>> +      cerr << "<" << CSDN->getValueAPF().convertToFloat() << ">";
>> +    else if (&CSDN->getValueAPF().getSemantics()
>> ==&APFloat::IEEEdouble)
>> +      cerr << "<" << CSDN->getValueAPF().convertToDouble() << ">";
>> +    else {
>> +      cerr << "<APFloat(";
>> +      CSDN->getValueAPF().convertToAPInt().dump();
>> +      cerr << ")>";
>> +    }
>
> APFloat really needs a "convertToString" method. :)

Yeah.  That would help CBackend too.  It's not trivial.

>> ===================================================================== 
>> =
>> ========
>> --- llvm/trunk/lib/Target/X86/X86RegisterInfo.cpp (original)
>> +++ llvm/trunk/lib/Target/X86/X86RegisterInfo.cpp Fri Sep 14
>> 17:26:36 2007
>> @@ -168,7 +168,11 @@
>>      Opc = X86::MOV32_mr;
>>    } else if (RC == &X86::GR16_RegClass) {
>>      Opc = X86::MOV16_mr;
>> +  } else if (RC == &X86::RFP80RegClass) {
>> +    Opc = X86::ST_FpP80m;   // pops
>>    } else if (RC == &X86::RFP64RegClass || RC == &X86::RSTRegClass) {
>> +    /// FIXME spilling long double values as 64 bit does not work.
>> +    /// We need RST80, unfortunately.
>
> The FP Stack has 80 bit load and store instructions, what is the
> issue? or is it just a todo?

I may be missing something, but I think  when we've gotten to the RST  
regclass,
we don't know what size the value in the register is, so we'd need to  
use the 80-bit
insns for all spills.  That's possible, but slower and more  
consumptive of stack space.
I haven't looked recently, but IIRC the time difference is significant.
(There's a difference in precision of the result also, but that's  
randomly
optimization-dependent anyway.)