Re: Patch: x86 sse performance

[Andrew Pinski <pinskia at physics dot uc dot edu>]

On Sep 8, 2005, at 9:33 PM, Richard Henderson wrote:

> On Thu, Sep 08, 2005 at 05:46:40PM -0700, Dale Johannesen wrote:
> > -  return _mm_set_epi8 (__A, __A, __A, __A, __A, __A, __A, __A,
> > -		       __A, __A, __A, __A, __A, __A, __A, __A);
> ...
> > +    __m128i temp, temp2, temp3, temp4;
> > +    temp = _mm_cvtsi32_si128 ((int)__A);
> > +    temp2 = _mm_unpacklo_epi8 (temp, temp);
> > +    temp3 = _mm_unpacklo_epi8 (temp2, temp2);
> > +    temp4 = _mm_shuffle_epi32 (temp3, 0);
>
> Not ok.
>
> If you don't like the code generated by 
> ix86_expand_vector_init_duplicate,
> you should fix it rather than avoiding it.

Something like this?
I did not test this but I get the same code gen as what the above
would get.  Also I cannot really test this fully.

Dale could you have a look?

Thanks,
Andrew Pinski

Attachment: t.diff.txt
Description: Text document