gcc 4.2.3 and MMX to mem move oddity

[Uros Bizjak]

Prakash Punnoor wrote:

>> Why is movaps (SSE, floating point data) instead of movdqa (SSE2. integer
>> data) used as store? Bug or feature? Even with -O0 compiled it is used.
>>     
>
> Testing further: The -march=k8 seems to cause this. Leaving it out, movdqa is 
> used, so I guess it is a feature.
>   

This is a feature, X86_TUNE_SSE_TYPELESS_STORES. It is faster on K8.

Regarding MMX stuff, I have just committed the patch to 4.4 that should 
fix your (and similar) problems. Unfortunatelly, it won't be backported, 
as this is kind of tricky register allocator fine-tuning.

Uros.