[llvm-commits] [llvm] r61552 - in /llvm/trunk: lib/Transforms/IPO/FunctionAttrs.cpp test/Transforms/FunctionAttrs/2008-12-31-NoCapture.ll

[Nick Lewycky]

Duncan Sands wrote:
> Hi Nick,
>
>   
>> +      // Does not escape if the callee is readonly and doesn't return a
>> +      // copy through its own return value.
>> +      if (CS.onlyReadsMemory() && I->getType() == Type::VoidTy)
>> +        continue;
>>     
>
> such function calls are trivially dead, so shouldn't occur.  Does this turn
> up in practice?
>   
FunctionAttrs just went through and added the readnone/readonly 
attributes. Nobody's had the chance to eliminate dead calls yet.

I had started with a patch that did "CS.onlyReadsMemory() && 
!isa<PointerType>(I->getType())" which doubled the number of nocaptures, 
but before committing, your testcase showed an example of leaking a bit 
of pointer out a non-pointer return type.

I haven't redone the metrics with the new change. The answer is 
"probably not much" but I don't think it hurts to leave it in.
>> +    // If the function is readonly and doesn't return any value, we 
>> +    // know that the pointer value can't escape. Mark all of its pointer 
>> +    // arguments nocapture.
>>     
>
> Because of this, the above nocapture logic enhancement shouldn't be needed.
>   
Yep, we can remove this one. Those functions will just be eliminated anyway.
>   
>> -        NumNoCapture++;
>> +        ++NumNoCapture;
>>     
>
> If this is an improvement, please change NumReadOnly++ etc.
>   
Ok!

Nick