[patch] enabling vectorization by default at -O3

Dorit Nuzman DORIT@il.ibm.com
Sat Sep 15 00:03:00 GMT 2007


> Dorit Nuzman wrote:
>
> > SpecCPU2006 on powerpc970 shows 1.5% on specINT (mostly due to a 14%
> > improvement on hmmer), and a -0.4% on specFP. This was comparing '-O3'
with
> > '-O3 -ftree-vectorize -maltivec'. The costmodel wasn't on, cause  it is
> > still undergoing some changes, and also it is not tuned yet for
powerpc970,
> > so I hope that with the costmodel on we'll be able to get rid of the
small
> > degradations we have here and there later on.
>
> I was hoping for a bit more of an overall win.  The negative result on
> specFP worries me;

if I use --param min-vect-loop-bound=2 (which limits vectorization to loops
whose vectorized loop bound is at least 2) the -0.4% degradation becomes a
0.15% improvement. However, I'm using a single run and the differences
between single runs of individual fp benchmarks on my system are up to
3.7%, so as Diego pointed out this is well within the noise range.

> although the overall scores are higher, it would be
> nice if we could eliminate that negative.
>

Agreed. I believe that the cost model has been getting a lot of attention
recently, and will continue to.

dorit

> Do any of the other optimization maintainers have a feeling about this?
>  Should we turn on vectorization at -O3, or are we not quite there yet?
>
> Thanks,
>
> --
> Mark Mitchell
> CodeSourcery
> mark@codesourcery.com
> (650) 331-3385 x713



More information about the Gcc-patches mailing list