PGO seems to be incompatible with the mold linker and fails with the error: > mold: fatal: std-7f67e63eb87e9645.std.5a0cb329cb1dc408-cgu.0.rcgu.o: don't know how to handle this LTO object file because no -plugin option was given. Please make sure you added -flto not only for creating object files but also for creating the final executable. This happens right at the beginning of the first (pre profiling) stage during cargo build. Just LTO+ld.mold doesn't have this issue. Reproducible: Always Steps to Reproduce: 1. Try to build firefox with USE=pgo and -fuse-ld=mold in *FLAGS (Tested with *FLAGS="-march=native -O2 -pipe -fuse-ld=mold") 2. Wait until for cargo build to fail (on my machine after ~1 min) Actual Results: Build fails. > process didn't exit successfully: ... clang: linker command failed Expected Results: Build succeeds.
Created attachment 883435 [details] build log
Created attachment 883436 [details] emerge --info
Shame it's with +clang, mold should work better with clang indeed. There was something in the ebuild I needed to do with gcc+mold, so doubt it's going to work better either. You can try though. Oh also, always disable ccache/sccache first and try without that. It almost always solves it.
clang+mold+pgo still fails even with (s)ccache disabled. gcc+mold+pgo isn't a thing according to the ebuild (because pgo requires lto). (gcc would likely fail anyways as rust on my system is built against LLVM's libc++) However I found a workaround. Since mold wants -flto* in LDFLAGS I just readded that flag after filter-lto for clang+lto+mold. > # LTO flag was handled via configure > filter-lto > >+ # explicitly set -flto for mold bgo#923119 >+ if use clang && use lto && tc-ld-is-mold; then >+ append-ldflags "-flto=thin" >+ fi Not the prettiest solution but now mold doesn't complain about the missing flag. Since regular lto seems to work as is you might even restrict it further to "use pgo" instead of lto.
(In reply to Jonas Rakebrandt from comment #4) > > gcc+mold+pgo isn't a thing according to the ebuild (because pgo requires > lto). > (gcc would likely fail anyways as rust on my system is built against LLVM's > libc++) Ah indeed, it falls back to bfd if it detects that gcc+lto. > > However I found a workaround. Since mold wants -flto* in LDFLAGS I just > readded that flag after filter-lto for clang+lto+mold. > > > # LTO flag was handled via configure > > filter-lto > > > >+ # explicitly set -flto for mold bgo#923119 > >+ if use clang && use lto && tc-ld-is-mold; then > >+ append-ldflags "-flto=thin" > >+ fi > > Not the prettiest solution but now mold doesn't complain about the missing > flag. > Since regular lto seems to work as is you might even restrict it further to > "use pgo" instead of lto. Awesome finding! I'll try to group the logic and keep everything related in the same place.
The bug has been closed via the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=ce63d29f57ba87b19b34e24b48a126cabd6a7264 commit ce63d29f57ba87b19b34e24b48a126cabd6a7264 Author: Joonas Niilola <juippis@gentoo.org> AuthorDate: 2024-01-31 08:29:52 +0000 Commit: Joonas Niilola <juippis@gentoo.org> CommitDate: 2024-01-31 08:30:49 +0000 www-client/firefox: drop ppc64 patch, fix mold+lto+pgo error on 122.0 Closes: https://bugs.gentoo.org/923300 Closes: https://bugs.gentoo.org/923119 Signed-off-by: Joonas Niilola <juippis@gentoo.org> www-client/firefox/Manifest | 2 +- www-client/firefox/firefox-122.0.ebuild | 6 ++++-- 2 files changed, 5 insertions(+), 3 deletions(-)
Shouldn't `append-ldflags "-flto=thin"` go after `filter-lto`? The way it is now -flto=thin is set for lto+clang+mold only to get filtered out again right after.
Hmm, looks like I didn't push the final version. I combined linker, lto and elf-hack stuff so they're together. I'll commit the final final version tomorrow!
The bug has been closed via the following commit(s): https://gitweb.gentoo.org/repo/gentoo.git/commit/?id=5a5d704f98b2099104eb371dbbad16eeda3f1ec1 commit 5a5d704f98b2099104eb371dbbad16eeda3f1ec1 Author: Joonas Niilola <juippis@gentoo.org> AuthorDate: 2024-01-31 17:40:40 +0000 Commit: Joonas Niilola <juippis@gentoo.org> CommitDate: 2024-01-31 17:41:38 +0000 www-client/firefox: fast-fix for mold+lto+pgo compile issue on 122 - handling elf-hack should be grouped together with these linker options. Closes: https://bugs.gentoo.org/923119 Signed-off-by: Joonas Niilola <juippis@gentoo.org> www-client/firefox/firefox-122.0.ebuild | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-)
Well here's an untested fast-fix from laptop :P
Nice, works now. Thanks!
