Red Hat Bugzilla – Bug 492939
gcc 4.4 broke ppc64 booting again
Last modified: 2009-04-13 09:46:49 EDT
Description of problem:
Somewhere between gcc 4.4.0-0.27.ppc and 4.4.0-0.29.ppc, we stopped getting bootable ppc64 kernels, at least on the YDL PowerStation. Kernel 2.6.29-0.258.rc8.git2.fc11, which was built with gcc 4.4.0-0.26.ppc boots just fine. The F11 beta kernel, 2.6.29-0.258.1.2.rc8.git2.fc11 (or something like that), which was built with gcc 4.4.0-0.29.ppc doesn't boot, nor do any later kernels built in koji. A local rebuild using the older compiler on my PowerStation results in a kernel that boots just fine.
The last thing I get on screen is "returning from prom_init", and the system just hangs there for a few minutes before resetting itself.
Version-Release number of selected component (if applicable):
I guess please first try -0.28, then do a binary search between -0.28 and -0.29 (resp. -0.27 and -0.28) compiled objects to see which file matters.
Hrm. I had a build using gcc 4.4.0-0.30.fc11 already under way, and I let that go ahead and finish. The resulting kernel boots fine. Perhaps the '- fix -maltivec conditional vector macro (PR target/39558)' portion of the build 0.30 changelog entry is relevant here? I could keep digging, but I don't have a heck of a lot of spare time to put into this...
None of the changes between -0.29 and -0.30 should affect code generation on ppc/ppc64, all the changes are just about diagnostics or debuginfo, or, as in the case of the vector changes, whether code involving -maltivec vector is parsed or not (but neither of these is a recent regression).
Ugh. Seems possibly the blame lies somewhere else then... I have yet to produce a locally-built kernel that didn't boot. Everything out of the build system of late fails to boot (well, haven't tried anything later than 2.6.29-15.fc11). Will poke it some more tomorrow...
Try local mock, if that still fails, try to change one component at a time (gcc, binutils, glibc, ...).
Latest kernels out of koji are bootable again, on both my powerstation and on Josh Boyer's G5 (which was also failing similarly).
Don't have time to dig into it, and its working now, just going to chalk it up to rawhide being rawhide and close it.