This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: basic-block and profile-based optimizing (was Re: New attribute "infrequent"?)
- To: Scott A Crosby <crosby at qwes dot math dot cmu dot edu>, g at atrey dot karlin dot mff dot cuni dot cz
- Subject: Re: basic-block and profile-based optimizing (was Re: New attribute "infrequent"?)
- From: Jan Hubicka <jh at suse dot cz>
- Date: Fri, 31 Aug 2001 22:17:49 +0200
- Cc: Andreas Jaeger <aj at suse dot de>, David Edelsohn <dje at watson dot ibm dot com>, Jan Hubicka <jh at suse dot cz>, gcc at gcc dot gnu dot org, pfk at fuchs dot offl dot uni-jena dot de
- References: <u84rqop4le.fsf@gromit.moeb> <Pine.LNX.4.33.0108311311050.25863-100000@qwe3.math.cmu.edu>
> This doesn't really matter, because the application itself presumabely
> will be going through profile-based optimization.. Thus, when profiling
I really don't expect this to hapend anytime soon.
Building distributions is really dificult task even now and using profile
feedback makes it even more dificult. It is easy to do profile feedback
for gcc, where bootstrap can be just adjusted to have one training stage,
but imagine interactive programs, like gimp that should be fast too.
Some research on static prediction methods shows that they can do relativly
well (as shows our results too) and we really should take care to avoid
compiler from being depdendent on profile feedback, as it is dificult to
use for programmers/distribution builders.
Honza