This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Compilation performance comparison of gcc3.4.1 and gcc3.5.0 2004-08-30 on MICO sources
- From: Steven Bosscher <stevenb at suse dot de>
- To: Karel Gardas <kgardas at objectsecurity dot com>,Paolo Bonzini <bonzini at gnu dot org>
- Cc: GCC Mailing List <gcc at gcc dot gnu dot org>
- Date: Tue, 31 Aug 2004 12:43:11 +0200
- Subject: Re: Compilation performance comparison of gcc3.4.1 and gcc3.5.0 2004-08-30 on MICO sources
- Organization: SUSE Labs
- References: <Pine.LNX.4.43.0408311227330.1006-100000@thinkpad.gardas.net>
On Tuesday 31 August 2004 12:28, Karel Gardas wrote:
> On Tue, 31 Aug 2004, Paolo Bonzini wrote:
> > >>>1) typecode.cc: 40% regression on O1 while 7% speedup on O2
> > >>
> > >>Can you show us the time report for the 40% regression?
> >
> > Also for 3.4.1?
>
> Sure!
Hmm... No obvious hot spots eh?
Looks like the tree optimizers are to blame. We spend roughly the same
amount of time in the post-GIMPLE passes, and we spend >7.5s in the tree
optimizers. The total slowdown you measured was ~8.9s. The other 1.4s
are spent in expand as shown in the previous message:
3.4.1: expand : 0.79 ( 6%) usr 0.03 ( 1%) sys 0.78 ( 5%) wall
3.5.0: expand : 2.08 ( 9%) usr 0.07 ( 4%) sys 2.51 (10%) wall
Hmm, we should probably disable at least flag_thread_jumps and
flag_loop_optimize at -O1, and perhaps consider disabling some
of the more expensive (parts of the) tree optimizers... And
of course see if it makes sense to disable a few RTL optimizers.
So, looks like a tuning problem to me, not really a slowdown that
indicates something algorithmic being really wrong.
Gr.
Steven