Differential D10748
Make ninja opt-out in cmake.mk. Authored by tcberner on May 15 2017, 7:47 PM.
Details
As kwm@ reported a quite significant build-speedup in wekbit2-gtk3 when using ninja instead of gmake, I'm interested to know whether using the ninja-generator of cmake could lead to at least some amount of speedup in other ports as well -- as cmake is used by many ports, this could be significant for the package builders. At the moment cmake.mk has an opt-in to use ninja. Make this an opt-out.
Diff Detail
Event TimelineThere are a very large number of changes, so older changes are hidden. Show Older Changes Comment Actions As there are apparently some ports breaking with ninja, make disabling ninja an argument to cmake.mk noninja instead of a variable. This way the broken ports can be fixed cleanly by just setting USES=cmake:<otherargs>,noninja. I will test-build the ports depending on ninja... to see what breaks. Comment Actions Don't use ninja for databases/mysql56-server: ninja: error: dependency cycle: sql/GenServerSource -> sql/CMakeFiles/GenServerSource -> sql/sql_builtin.cc -> cmake_order_depends_target_sql -> sql/GenDigestServerSource -> sql/CMakeFiles/GenDigestServerSource -> sql/lex_token.h -> sql/gen_lex_token -> sql/GenServerSource Comment Actions Maybe we could call noninja just "make"? USES=cmake:outsource,make # make use make instead of ninja Comment Actions Ok, I think I should have gotten more-or-less all... 5-6% of the affected ports needed noninja. Comment Actions I just realized that this change doesn't have a ports/CHANGES yet, could you prepare one? Like describe this change and how people can opt-out if ninja isn't supported. Comment Actions Sorry for chiming in so late. I think this diff also obsoletes https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=213331 in the sense that we'd no longer have USES=cmake together with USES=gmake (only USES=cmake:noninja gmake, which is fine). To complement what @kwm said, this patch is lacking documentation: right now, I don't think it's clear to other ports contributors when to add noninja at all. I tested a handful of the ports that need noninja here on Poudriere and they seem to fall into two categories:
I wonder if it makes sense to try to fix the second case automatically in Uses/cmake.mk. Comment Actions
Sure, it's worth a try Comment Actions I could not consitently reproduce the build failures of
on i386. So make them not use ninja (for now). Further drop the USES=gmake from
which was added for FreeBSD 10.x Additionally, mark ports ignored that set USES=cmake gmake. Comment Actions I've added some inline comments about ports that do not need noninja. In addition to what you added to CHANGES, I'd expect some documentation in Uses/cmake.mk as well, otherwise it's not much better than not adding anything at all. Specifically, there needs to be some documentation explaining that noninja is a stop-gap measure that's normally used by:
Comment Actions I have trouble figuring out if there are more ports using cmake that should use ninja that ports that should not use ninja. And so, I am wondering if it should not be a USES=cmake:ninja to enable it.
Comment Actions Using freshports.org/devel/cmake as reference and filtering out ports that no longer exist, there are 1317 ports that have a build-time dependency on CMake, ~70 of which don't work with ninja as far as I can see. Comment Actions Mmmm, ok, good for me :-) Next question, do we really a need a USES=ninja, so, are there ports that use USES=ninja but not USES=cmake (and that could not use USES=cmake) ? Comment Actions Drop USES=gmake from comms/uhd lang/sagittarius-scheme, now that they don't set noninja, they need to have gmake dropped. Comment Actions
Comment Actions
Comment Actions
Comment Actions lgtm from my side, thanks for working on this. In your commit message, make sure you mention why some changes were done (e.g. why MAKE_ARGS was removed from compiler-rt). | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||