PPC/PPC64 disabled in Koji for dist-f13
Kevin Kofler
kevin.kofler at chello.at
Thu Oct 1 00:58:15 UTC 2009
Jeff Garzik wrote:
> I would rather the problem be approached in a logical, scalable fashion:
> by distributing the workload across the package maintainers who have
> firsthand knowledge. ie. how things worked before.
>
> But you're dodging the larger point -- Fedora has, de facto, demoted big
> endian support in its entirety to a second-hand effort, rather than
> distributed the workload much more widely. Given M package maintainers
> and N secondary-platform volunteers, it is clear M > N by orders of
> magnitude.
I think it is only fair that the people who actually care get to do the
work. Package maintainers can still fix their packages for PPC, they'll even
get e-mail reports from the secondary arch Koji if the builds fail. (It
already happens for the other secondary architectures.) They just won't be
required to do it anymore. You can't force volunteers (and many Fedora
developers are volunteers; even those paid by RH are paid primarily to work
on RHEL, not Fedora, and often do Fedora work in their own free time) to
work on something they're not interested in.
> Was ppc really such a burden?
Yes. It slowed down builds, and it often triggered bizarre build failures
which were NOT bugs in the program, but in the toolchain or in some core
library like glibc, which in turn delayed important updates to the affected
packages.
In fact, my "favorite" ppc64 issue was a problem with OpenBabel hitting a
limitation in the ppc64 toolchain: there's a "table of contents" which can
grow only to a small fixed size, so large compilation units just don't
compile on ppc64, while being perfectly valid C/C++ and compiling fine on
all other architectures. (And that's already with the "minimal TOC". Without
it, the limit is for the whole executable!) OpenBabel's SWIG-generated
bindings exceeded that limit. We were the only ones hitting it as no other
distribution is insane enough to build their packages for ppc64. (The
binaries don't even get actually used as ppc32 is the preferred multilib on
64-bit PowerPC!) I played around with both compiler and SWIG flags to reduce
the amount of TOC entries, which worked for 2 beta releases (requiring
additional tweaking for the second one), but then it overflowed again. This
was a big annoyance because the new OpenBabel betas were required for new
kdeedu betas in Rawhide, so this stalled our KDE work. In the end, upstream
removed some things from their bindings to get them to build, a quite
suboptimal solution. IMHO the ppc64 ABI is just completely broken and needs
to be redesigned from scratch.
Kevin Kofler
More information about the fedora-devel-list
mailing list