Description of problem: When building Firefox with PGO, gcc crashes in feedback build with: 22:02.22 ../TestJSImplInheritanceGenBinding.cpp: In function ‘_GLOBAL__sub_I_TestJSImplInheritanceGenBinding.cpp’: 22:02.22 ../TestJSImplInheritanceGenBinding.cpp:2911:1: warning: ‘/home/komat/src/objdir-pgo/dom/bindings/test/TestJSImplInheritanceGenBinding.gcda’ profile count data file not found [-Wmissing-profile] 22:02.22 2911 | } // namespace mozilla 22:02.22 | ^ 22:02.26 mfbt/tests/TestFunctionTypeTraits 22:02.45 during RTL pass: ira 22:02.48 /home/komat/src/objdir-pgo/dist/include/mozilla/Casting.h: In function ‘BitwiseCast’: 22:02.48 /home/komat/src/objdir-pgo/dist/include/mozilla/Casting.h:64:1: internal compiler error: in to_frequency, at profile-count.c:273 22:02.48 64 | } 22:02.48 | ^ 22:02.56 Please submit a full bug report, 22:02.56 with preprocessed source if appropriate. 22:02.56 See <http://bugzilla.redhat.com/bugzilla> for instructions. 22:02.58 make[5]: *** [/tmp/ccCKwxhi.mk:2: /tmp/TestFloatingPoint.sifv8g.ltrans0.ltrans.o] Error 1 22:02.58 lto-wrapper: fatal error: make returned 2 exit status 22:02.58 compilation terminated. 22:02.58 /usr/bin/ld: error: lto-wrapper failed 22:02.58 collect2: error: ld returned 1 exit status 22:02.58 gmake[4]: *** [/home/komat/src/config/rules.mk:517: TestFloatingPoint] Error 1 22:02.58 gmake[4]: *** Waiting for unfinished jobs.... Version-Release number of selected component (if applicable): gcc-10.1.1-1.fc32.x86_64 I'm not sure how to extract any useful debuginfo data here. Mozilla uses clang to build PGO/LTO FF.
Does it reproduce with current F32 package git src.rpm, or any tweaks are needed? Does it happen in F33 too (which has slightly newer gcc)?
It's latest nightly only, I haven't reproduced it with Firefox 77 from Fedora due to other build issues. May be affected as well, I'll post update here. I haven't tested Fedora 33.
Hi Jakub, so we hit the gcc bug in recent Firefox at Fedora 32/33. The affected builds are here: https://koji.fedoraproject.org/koji/taskinfo?taskID=48015864 fails with: /builddir/build/BUILD/firefox-79.0/objdir/dist/include/mozilla/Casting.h: In function 'BitwiseCast': /builddir/build/BUILD/firefox-79.0/objdir/dist/include/mozilla/Casting.h:64:1: internal compiler error: in to_frequency, at profile-count.c:273 64 | } | ^ Please submit a full bug report, with preprocessed source if appropriate. See <http://bugzilla.redhat.com/bugzilla> for instructions. make[2]: *** [/tmp/cc97Ic0E.mk:2: /tmp/TestFloatingPoint.oQNWgG.ltrans0.ltrans.o] Error 1 lto-wrapper: fatal error: make returned 2 exit status all you need to reproduce is just to rebuild Firefox srpm package from https://koji.fedoraproject.org/koji/taskinfo?taskID=48015953 Thanks.
Proposing as prioritized bug, see https://pagure.io/fesco/issue/2020#comment-671672
Just a note, Jakub is on PTO this week, so I wouldn't expect any progress until he returns.
I have tried to reproduce this, but couldn't, firefox-79.0-1 builds just fine on x86_64 when LTO is not on and PGO is on. With LTO on one runs into the write_type https://gcc.gnu.org/PR96690 ICE.
I have made further 2 attempts and didn't reproduce in either of those.
Okay, I'll try new builds.
Since reproducibility is unclear, we deferred decision on accepting as a Prioritized Bug until the next meeting (9 September): https://meetbot.fedoraproject.org/fedora-meeting/2020-08-26/fedora_prioritized_bugs_and_issues.2020-08-26-15.00.log.html#l-38
Martin, can you provide more information on reproducibility? We're still unsure about whether or not to accept this as a Prioritized Bug: https://meetbot.fedoraproject.org/fedora-meeting/2020-09-09/fedora_prioritized_bugs_and_issues.2020-09-09-15.00.log.html#l-24
I think we can remove it from Prioritized bugs as PGO works. LTO is still broken. Thanks.
Depends on which LTO bug you're talking about, the ICE in write_type should be fixed already in f34 and in f33 is stuck, waiting for the testing -> stable push for a week already. Or do you encounter some other LTO issue (e.g. the known ARM Neon problem)?
> be fixed already in f34 and in f33 is stuck, waiting for the testing -> > stable push for a week already. We're in freeze for beta, so that is expected, but you can tag gcc and any other associated bits (annobin etc) into override, or you can request a freeze exception via https://qa.fedoraproject.org/blockerbugs/
Can you please submit the gcc lto fix update for Fedora 32 too? Thanks.
(In reply to Jakub Jelinek from comment #12) > Depends on which LTO bug you're talking about, the ICE in write_type should > be fixed already in f34 and in f33 is stuck, waiting for the testing -> > stable push for a week already. > Or do you encounter some other LTO issue (e.g. the known ARM Neon problem)? LTO seems to build fine on my x86_64/Fedora 33/gcc-10.2.1-3.fc33 so this one can be closed.