This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: propagate_freq enormous performance hit in 3.1
- To: Brad Lucier <lucier at math dot purdue dot edu>
- Subject: Re: propagate_freq enormous performance hit in 3.1
- From: Jan Hubicka <jh at suse dot cz>
- Date: Tue, 14 Aug 2001 12:57:34 +0200
- Cc: jh at suse dot cz, gcc at gcc dot gnu dot org
- References: <200108131643.f7DGhVI22336@banach.math.purdue.edu>
> Jan:
>
> >From examining the sources, it seems that predict.c is your baby,
> and it seems to have a serious performance problem in propagate_freq
> for large flow graphs, see
>
> http://gcc.gnu.org/ml/gcc-prs/2001-08/msg00184.html
>
> Do you have any idea of what's going on here?
It can have problems with large graphs with many irregular regions (see the
comment inside function).
I will take a look - in meantime I made code to find back edges, so I can
discover irregular regions easilly.
Can you please try to take a look at debug output and let me know if it
complains about the irregular regions?
Otherwise the time complexity should be OK. I will try to prepare patch for
better irregular regions handling overnight, but my time is somewhat limited, so
it may take longer.
Thanks for letting me know.
Honza
>
> BTW, I apologize for the large, useless inclusion in that GNAT PR. The real
> source code I tested it on is at
>
> http://www.math.purdue.edu/~lucier/all.i.gz
>
> Brad Lucier