Re: PR opt/9745: Recent SPECfp2000 slowdown on mainline/3.3

[Jim Wilson <wilson at tuliptree dot org>]

On Sun, 2003-07-13 at 18:54, Roger Sayle wrote:
> I was wondering whether the following patch was described on
>gcc-patches?

There was a bit of discussion, but in hind sight I think none of the
messages were on gcc-patches.  Franz Sirl mentioned the thread on the
gcc list.  See also the PRs 9745, 10021, and 11319.

> The reason I ask is that this change seems to be responsible for a
fairly
> serious performance regression of SPECfp2000 that occurred sometime on
> Friday (11th July).

That is a shame, but unavoidable in the short term.  My patch solves a
problem where the alias info was badly corrupted.  This corruption was
causing code to be miscompiled.  It is also possible that this
corruption happened to cause a performance gain for some benchmarks by
accident.  However, since the code was bad, it had to be fixed, so we
have to live with the performance regression for now.  It looks like a
3-4% performance regression.

In one of my messages, I pointed out various limitations of
loop_regs_update that if fixed would give better performance.  If you
want to help get the performance back, you could look into this.  See
	http://gcc.gnu.org/ml/gcc/2003-07/msg00668.html
This message also gives the justification for the patch I checked in.

Meanwhile, there are still more loop aliasing bugs that need to be
fixed.  My patch fixed 9745 but it did not fix 10021 or 11319. 
Additional loop aliasing patches will be necessary.  I will try to avoid
more performance regressions but some may be unavoidable.  I think we
can get them back with improvements to loop_regs_update, but I have no
idea when I will have time to work on such improvements.

I do not have access to a copy of SPEC yet, so I am unable to look into
the performance regression.
-- 
Jim Wilson, GNU Tools Support, http://www.specifix.com