Re: Help wanted testing gcc-3.3 / gcc-snapshot
Aaron Lehmann writes:
> On Sat, Mar 01, 2003 at 08:03:30PM +0100, Matthias Klose wrote:
> > For the upcoming GCC 3.3 release, the compiler collection should be
> > tested with real word applications according to the release
> > criteria. As many of the real-world applications mentioned on
> >
> > http://gcc.gnu.org/gcc-3.3/criteria.html
> >
> > are part of Debian, it may be useful for Debian to check these even
> > before GCC 3.3 is released and becomes the default compiler at least
> > on some platforms.
>
> How about XFree86? 3d acceleration has been *broken* on the radeon
well, it's no _gcc_ test criteria.
> family of cards on i386 since xfree86 -5, which was compiled with gcc
> 3.2. (#178242). -6 is also affected. About a month ago I was able to
> duplicate the problem with gcc-snapshot. I then compiled the package
> with mainline gcc and GL acceleration began to work. I posted to
> gcc@gcc.gnu.org asking for suggestions for tracking down this bug so I
> could backport the fix to 3.3, but since I'm not familiar with what
> exactly the miscompilation is, nobody was able to help me. IMHO this is
> a very severe bug, even though apparently XFree86 4.3 (which is not yet
> officially packed) does not cause this miscompilation.
Try to build with the 20030301 snapshot package, first with standard
CFLAGS (-O2), then -O1, then -O0. Maybe it's sufficient to compile
only the Radeon specific files. Try to reproduce the bug after each
build.
If it's not reproduceable, then try to replace the some object files
with object files compiled with a gcc from HEAD.
Maybe here you can stop and ask the xfree86 maintainer to insert a
file specific rule to produce the correct object file.
If you have identified the file, which is wrongly compiled, split it
up and search the function, which get's miscompiled...
I know, that's tedious...
Matthias
Reply to: