Re: downgrading optimization for m68k [was: Bug#328453: pbzip2_0.9.4-1(m68k/unstable/zeus): FTBFS on m68k]
- To: Joe Wreschnig <piman@debian.org>
- Cc: debian-python@lists.debian.org
- Subject: Re: downgrading optimization for m68k [was: Bug#328453: pbzip2_0.9.4-1(m68k/unstable/zeus): FTBFS on m68k]
- From: Matthias Klose <doko@cs.tu-berlin.de>
- Date: Fri, 16 Sep 2005 23:45:02 +0200
- Message-id: <17195.15582.643361.67185@gargle.gargle.HOWL>
- In-reply-to: <1126906523.5393.36.camel@localhost>
- References: <20050915122225.GJ30702@marenka.net> <20050916004050.GH9253@debianrules.debiancolombia.org> <20050916005925.GB4714@marenka.net> <432A2AB4.3080206@debian.org> <20050916024733.GC26064@tennyson.dodds.net> <1126849047.13867.9.camel@localhost.localdomain> <1126894153.5393.25.camel@localhost> <17195.5686.337464.790718@gargle.gargle.HOWL> <1126906523.5393.36.camel@localhost>
Joe Wreschnig writes:
> On Fri, 2005-09-16 at 21:00 +0200, Matthias Klose wrote:
> > Joe Wreschnig writes:
> > > I guess I should ask here, too. Does anyone know why Python is compiled
> > > with -O3 rather than -O2? Also, does anyone know the best way to
> > > override distutils on a per-arch basis to change that?
> >
> > There's only one optimization macro to build the interpreter and the
> > modules. IMO it makes sense to build the interpreter with -O3, and
> > even to build the standard modules like _sre with this optimization
> > level. What can be done, is to lower the opt level after compilation
> > in the package.
> >
> > But as you can see, even with -O2 python2.4 FTBFS on m68k.
>
> Regardless, -O3 has been historically buggy compared to -O2 on every
> arch. Bill Allombert mentioned it was broken on x86 right now too. And
> given that I've spent about 5 hours digging around trying to discover
> the cause of #328587 -- or even reproduce it -- with no luck, I'm about
> to blame -O3 for it.
>
> If there aren't numbers suggesting -O3 is a real win, I think it's a bad
> idea to use it *anywhere*. It's a bad default, and an especially bad one
> for something like Python modules that sit on top of a very complex
> layer of code. And that's the case regardless of whether we're in the
> middle of m68k breakage or not.
You oversimplify. Just reread what I did write.
There have been measurements by Gregor Hoffleit in the past, please
look back on the mailing lists.
Reply to: