This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: A workaround for PR 19038
- From: Mostafa Hagog <MUSTAFA at il dot ibm dot com>
- To: law at redhat dot com
- Cc: gcc at gcc dot gnu dot org, gcc-patches at gcc dot gnu dot org, pinskia at physics dot uc dot edu, Roger Sayle <roger at eyesopen dot com>, Steven Bosscher <stevenb at suse dot de>
- Date: Thu, 30 Dec 2004 21:35:20 +0200
- Subject: Re: A workaround for PR 19038
Jeffrey A Law <law@redhat.com> wrote on 29/12/2004 21:22:05:
> On Wed, 2004-12-29 at 01:04 +0200, Mostafa Hagog wrote:
> >
> >
> >
> > The attached patch implements loop header copying in the RTL level
> > to work around the problem in http://gcc.gnu.org/PR19038
> > I know that this patch is controversial, but it could be added
> > at first at -O3 (or -O2) and disabled after the final fix in
> > out-of-ssa. Anyhow, I thought it would be helpful to put the
> > patch on the mailing list.
> > Currently this patch buy us improvements of 3.5% on SPEC (a 66%
> > for 176.gcc benchmark) on a G5 machine. Compile time wasn't
> > affected by this patch, following are the compile time results
> > on a G5:
> >
> > % time gcc -c -O3 generate-3.4.ii
> > 53.140u 1.590s 0:54.70 100.0% 0+0k 0+12io 0pf+0w
> > % time gcc -c -frtl-loop-hc -O3 generate-3.4.ii
> > 53.140u 1.670s 0:54.82 99.9% 0+0k 0+11io 0pf+0w
> > % time ./compile_i_files.sh gcc
> > 150.220u 15.050s 2:45.46 99.8% 0+0k 0+124io 0pf+0w
> > % time ./compile_i_files.sh gcc -frtl-loop-hc
> > 150.340u 14.850s 2:45.53 99.7% 0+0k 0+115io 0pf+0w
> > % time ./compile_i_files.sh gcc -O3
> > 656.550u 33.980s 11:30.50 100.0% 0+0k 0+296io 0pf+0w
> > % time ./compile_i_files.sh gcc -O3 -frtl-loop-hc
> > 658.510u 35.090s 11:33.77 99.9% 0+0k 0+313io 0pf+0w
> >
> > Mostafa.
> >
> >
> > (See attached file: rtl_loop_ch.patch)
> Is this still necessary/useful if the out-of-ssa doesn't create the
> bogon basic block?
Your patch doesn't help 176.gcc (of the SPECcpu). With rtl loop
header copying 176.gcc score improves by 66.4%; the above patch
doesn't improve the score at all. (the measurements are made on
a G5 machine).
Mostafa.