This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Compilation performance comparison of 3.5.0 and TreeSSA treeson MICO sources as requested in: [tree-ssa] Merge status 2004-05-03
- From: Karel Gardas <kgardas at objectsecurity dot com>
- To: Andrew Pinski <pinskia at physics dot uc dot edu>
- Cc: Mike Stump <mrs at apple dot com>, Diego Novillo <dnovillo at redhat dot com>, "gcc at gcc dot gnu dot org" <gcc at gcc dot gnu dot org>
- Date: Tue, 4 May 2004 23:25:49 +0200 (CEST)
- Subject: Re: Compilation performance comparison of 3.5.0 and TreeSSA treeson MICO sources as requested in: [tree-ssa] Merge status 2004-05-03
On Tue, 4 May 2004, Andrew Pinski wrote:
>
> On May 4, 2004, at 17:13, Mike Stump wrote:
>
> > On Tuesday, May 4, 2004, at 01:50 PM, Karel Gardas wrote:
> >> MICO doesn't look good. I've compared todays source trees:
> >
> >> File 350-O0 TSSA-O0 Delta% 350-O2 TSSA-O2
> >
> >> Sum 523.2 649.36 -19.43 774.7 904.11 -14.31
> >>
> >> Conclusion: compile time regressions for both -O0 and -O2 compilation.
> >
> > Thanks. While I'll happily take the 14% hit for optimized code, the
> > 20% drop for -O0 I want to bitch about. Last time we lost 34% in one
> > tiny patch, we lost, and the work went in anyway. That was almost ok,
> > as we just reverted the work that caused the regression in our tree.
> > However, we have no ability to `revert' treessa for -O0 to gain the
> > 20% back.
> >
> > I'd like ask people to see if there is anyway to get that back. If
> > not all of it, part of it?
>
> Yes there is a way, it is called do some optimizations at -O0. The
> problem with the tree-ssa,
> the RTL which is generated for an unoptimized run is just huge compared
> to what the mainline
> was/is. Simple things like DCE will help a lot on the compile time and
> code size.
Hmm, do you think I should try -O1 test? I guess it will place somewhere
between 14-19% regression, but if you would like to see it, I will do it
ASAP.
Karel
--
Karel Gardas kgardas@objectsecurity.com
ObjectSecurity Ltd. http://www.objectsecurity.com