llvm dev - Dec 2009 - [LLVMdev] a constant folding case causing different results w/ot optimization

Hi,

 

The optimizer folds "(unsigned int) -1.0f" to 0. As a result, the
following routine foo returns 0 with optimization, and returns -1
(0xffffffff) without optimization.

 

unsigned int foo()

{

    float myfloat = -1.0f;

    unsigned int myint = (unsigned int) myfloat;

    return myint;

}

 

INSTCOMBINE ITERATION #0 on

IC: ConstFold to: i32 0 from:   %conv = fptoui float -1.000000e+000 to
i32      ; <i32> [#uses=1]

 

 

 

While the result of "(unsigned int) -1.0f" is probably implementation
defined, both gcc and Microsoft cl produce -1.

Will you consider this an optimizer bug and fix it?

 

Thanks,

 

bixia

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20091207/b04a7a8f/attachment.html>

On Mon, Dec 7, 2009 at 3:51 PM, Zheng, Bixia <Bixia.Zheng at amd.com>
wrote:
> While the result of “(unsigned int) -1.0f” is probably implementation
> defined, both gcc and Microsoft cl produce -1.
LLVM (and C/C++) consider the result to be undefined (i.e. it can
produce anything).  And it can actually produce results other than -1
on some platforms supported by LLVM.  So I don't see any good reason
to choose your way of folding it over the current method.

-Eli

On Dec 7, 2009, at 4:50 PM, Eli Friedman wrote:
> On Mon, Dec 7, 2009 at 3:51 PM, Zheng, Bixia <Bixia.Zheng at amd.com>
wrote:
>> While the result of “(unsigned int) -1.0f” is probably implementation
>> defined, both gcc and Microsoft cl produce -1.
> 
> LLVM (and C/C++) consider the result to be undefined (i.e. it can
> produce anything).  And it can actually produce results other than -1
> on some platforms supported by LLVM.  So I don't see any good reason
> to choose your way of folding it over the current method.
Yes.  If you're interested in improving this, folding it to undef would be
better than to 0.

-Chris