[llvm-commits] [llvm] r56885 - in /llvm/trunk: include/llvm/Target/TargetLowering.h lib/CodeGen/SelectionDAG/SelectionDAG.cpp lib/Target/X86/X86ISelLowering.cpp lib/Target/X86/X86ISelLowering.h lib/Target/X86/X86Subtarget.cpp lib/Target/X86/X86Su

[Evan Cheng]

On Sep 30, 2008, at 4:21 PM, Dale Johannesen wrote:

>
> On Sep 30, 2008, at 4:13 PMPDT, Evan Cheng wrote:
>
>> Actually no.  It's a terrible idea. :-)
>>
>> This logic should not be in codegen at all. Users can still right
>> __builtin_memcpy and those can be optimized with -fno-builtin.
>>
>> So correct solution, take two. :-) FE should not generate calls to
>> @llvm.memset.i32 etc. when -fno-builtin is specified. Instead, it
>> should generate a simple call to memset. So the only changes required
>> are in the FE, LLVM should not know anything about -fno-builtin at
>> all.
>
> Right.
>
>> That is, unless simplifylibcall decides to turn memset into
>> @llvm.memset*.
>
> This is why simplifylibcall shouldn't be keying off names.  I don't
> think I got through before, maybe now....

This is not a simplifylibcall issue though. Even at -O0, llvm-gcc will  
emit these intrinsics. See llvm-convert.cpp:

void TreeToLLVM::EmitMemSet(Value *DestPtr, Value *SrcVal, Value *Size,
                             unsigned Align) {
   const Type *SBP = PointerType::getUnqual(Type::Int8Ty);
   const Type *IntPtr = TD.getIntPtrType();
   Value *Ops[4] = {
     BitCastToType(DestPtr, SBP),
     CastToSIntType(SrcVal, Type::Int8Ty),
     CastToSIntType(Size, IntPtr),
     ConstantInt::get(Type::Int32Ty, Align)
   };

   Intrinsic::ID IID =
     (IntPtr == Type::Int32Ty) ? Intrinsic::memset_i32 :  
Intrinsic::memset_i64;

   Builder.CreateCall(Intrinsic::getDeclaration(TheModule, IID), Ops,  
Ops+4);
}

Evan
>
>
>> Evan
>>
>> On Sep 30, 2008, at 4:08 PM, Bill Wendling wrote:
>>
>>> On Tue, Sep 30, 2008 at 3:55 PM, Evan Cheng <evan.cheng at apple.com>
>>> wrote:
>>>>
>>>> Actually, I think it'd be better to do this the right way.
>>>
>>> Bah! :-)
>>>
>>>> The correct
>>>> way to handle this option is to convert all memcpy / memset (and
>>>> other) intrinsic calls to normal calls. Perhaps in
>>>> SelectionDAGBuild.cpp? e.g.
>>>>
>>>> case Intrinsic::memcpy_i32:
>>>> case Intrinsic::memcpy_i64: {
>>>>  SDValue Op1 = getValue(I.getOperand(1));
>>>>  SDValue Op2 = getValue(I.getOperand(2));
>>>>  SDValue Op3 = getValue(I.getOperand(3));
>>>>  unsigned Align = cast<ConstantInt>(I.getOperand(4))-
>>>>> getZExtValue();
>>>>  DAG.setRoot(DAG.getMemcpy(getRoot(), Op1, Op2, Op3, Align, false,
>>>>                            I.getOperand(1), 0, I.getOperand(2),
>>>> 0));
>>>>  return 0;
>>>> }
>>>>
>>>> We can lower it to a call to memcpy with -fno-builtin.
>>>>
>>> That's basically what I suggested in another email. So I think  
>>> it's a
>>> brilliant idea. ;-)
>>>
>>> -bw
>>> _______________________________________________
>>> llvm-commits mailing list
>>> llvm-commits at cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>>
>> _______________________________________________
>> llvm-commits mailing list
>> llvm-commits at cs.uiuc.edu
>> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits