Date: Mon, 05 Jul 2004 16:00:02 -0600 From: Scott Long <scottl@samsco.org> To: obrien@FreeBSD.org Cc: Kris Kennaway <kris@obsecurity.org> Subject: Re: cvs commit: src/share/examples/etc make.conf Message-ID: <40E9CF62.5080400@samsco.org> In-Reply-To: <20040705194829.GA3743@dragon.nuxi.com> References: <200407030941.i639fwt8078389@repoman.freebsd.org> <20040704032139.GA93138@VARK.homeunix.com> <20040704051607.GA78676@xor.obsecurity.org> <20040704205648.GA1617@dragon.nuxi.com> <20040704232050.GA90994@xor.obsecurity.org> <20040705194829.GA3743@dragon.nuxi.com>
next in thread | previous in thread | raw e-mail | index | archive | help
David O'Brien wrote: > On Sun, Jul 04, 2004 at 04:20:50PM -0700, Kris Kennaway wrote: > >>On Sun, Jul 04, 2004 at 01:56:48PM -0700, David O'Brien wrote: >> >>>On Sat, Jul 03, 2004 at 10:16:07PM -0700, Kris Kennaway wrote: >>> >>>>On Sat, Jul 03, 2004 at 08:21:39PM -0700, David Schultz wrote: >>>> >>>>>FWIW, I've been compiling most things with -O2 for a while (to >>>>>find -O2 bugs, not for speed) and haven't noticed many problems. >>>>>The only significant one I know of is that -O2 breaks >>>>>floating-point exceptions in libm because gcc doesn't support the >>>>>FENV_ACCESS pragma. I think for some routines like rint(3), it >>>>>may even give the wrong answer due to incorrect optimizations, but >>>>>I'd have to check that again. >>>>> >>>>>AFAIK, the necessary functionality to make gcc's optimizer treat >>>>>floating-point code in a sane manner isn't on the horizon, so >>>>>maybe -O2 should be automatically turned off while compiling libm >>>>>(and perhaps libalias as well). That would make it more >>>>>easily justifiable to make -O2 the default at some future point. >>>> >>>>I don't think we can ever make it the default since there's likely to >>>>be a lot of software in ports that would be broken too. >>> >>>99% of the ports that "may break" build with -O2 on Linux (as -O2 is >>>their default). What is different about us vs. Linux for these ports? >> >>We care about not introducing instability into our packages? >> >>If we have >=2 -O2 bugs in our source tree alone, why should you think >>that none of the 11000 ports are affected? > > > Because most everything in the ports collection was developed on Linux > using -O2. The bugs are in our code, not gcc's -O2. > Making -O2 be the global default is a bad idea right before 5.3. While 90% of the ports might be fine under linux with -O2, it just adds too much risk this late in the game. Let's revisit it shortly after 5.3. Scott
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?40E9CF62.5080400>