This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: vortex peak performance regression
- From: Zoltan Hidvegi <hzoli at hzoli dot 2y dot net>
- To: jh at suse dot cz (Jan Hubicka)
- Cc: hzoli at hzoli dot 2y dot net, gcc at gcc dot gnu dot org
- Date: Wed, 2 Jan 2002 19:54:34 -0600 (CST)
- Subject: Re: vortex peak performance regression
> Zoltan,
> It looks like your patch:
> + 2001-11-30 Zoltan Hidvegi <hzoli@hzoli.2y.net>
> +
> + * unroll.c (unroll_loop): Correct special exit cases.
> +
> causes regression in the spec2000 tester - most importantly the vortex
> benchmark (- see www.suse.de/~aj/SPEC
>
> It is not obvious for me why, but can you please double check
> whether the regression is unavoidable.
I have to think more about it, as I was unable to quickly produce a
testcase to show it is necessary. But it only adds a single increment
to the loop, but that also requires an extra register, that may be
just enough to run out of registers. I have no more time to think
about this today, I'll try to look at it tomorrow. Which architecture
did you run the test? It can make a difference if it has
count-regiter loop or not.
Zoli