mmintrin slower than inline asm or even plain C

mmintrin slower than inline asm or even plain C

[Jack Andrews]

hi guys,

i write to you direct because i can't find the relevant mailing list
for help with the mmintrin functions.  there's a thread at:

  http://gcc.gnu.org/ml/gcc-help/2007-04/msg00201.html

that details my problems.

i want to sum an array of longs using mmx.  i use the functions:
    _mm_set_pi32 and _m_paddd
but the resultant binary contains significantly less efficient code
than inline asm or even plain C ( for(i=0;i<n;i++)total+=a[i]; ).
here's the relevant function:

simd_mmintrin(n, is)
I *is;
{   __m64 q,r;
   I i;
   _m_empty();
   q=_m_from_int(0);
   for (i=0; i < n; i+=W) {
       r=_mm_set_pi32(is[i],is[i+1]);
       q=_m_paddd(q,r);
   }
   union {long a[2];__m64 m;}u;
   u.m=q;
   return u.a[0]+u.a[1];
}

and the rest of the code and a shell script to run it is in the thread above.

thank you,


jack

Re: mmintrin slower than inline asm or even plain C

[Richard Guenther]

On 4/22/07, Jack Andrews <effbiae@gmail.com> wrote:
> hi guys,
>
> i write to you direct because i can't find the relevant mailing list
> for help with the mmintrin functions.  there's a thread at:
>
>   http://gcc.gnu.org/ml/gcc-help/2007-04/msg00201.html
>
> that details my problems.
>
> i want to sum an array of longs using mmx.  i use the functions:
>     _mm_set_pi32 and _m_paddd
> but the resultant binary contains significantly less efficient code
> than inline asm or even plain C ( for(i=0;i<n;i++)total+=a[i]; ).
> here's the relevant function:
>
> simd_mmintrin(n, is)
> I *is;
> {   __m64 q,r;
>    I i;
>    _m_empty();
>    q=_m_from_int(0);
>    for (i=0; i < n; i+=W) {
>        r=_mm_set_pi32(is[i],is[i+1]);
>        q=_m_paddd(q,r);
>    }
>    union {long a[2];__m64 m;}u;
>    u.m=q;
>    return u.a[0]+u.a[1];
> }
>
> and the rest of the code and a shell script to run it is in the thread above.

You should do a bugreport.  I suspect that we cannot combine
_mm_set_pi32(is[i],is[i+1]) to a movq as you do in the asm and that
we have non-optimal register allocation.

Richard.