The jemalloc ebuild configures jemalloc with: --with-jemalloc-prefix=j This makes it impossible to use LD_PRELOAD=/usr/lib/libjemalloc.so.1 on existing applications (as a glibc malloc replacement). Please add an option to IUSE. Reproducible: Always
Thanks for reporting jemalloc-3.2.0 is in tree with modificiations.
This broke ebuilds that rely on building against a prefixed jemalloc (for instance, dev-db/redis). USE=prefix or the other way around? Going from prefixed to non-prefixe is a pretty major step for a library since it will break stuff relying on it.
Here are some other potentially affected packages: app-office/libreoffice dev-db/mariadb dev-db/mysql net-fs/cvmfs The trouble with having a USE flag is that affected packages would need to block in situations where the user has enabled the jemalloc flag but also chosen not to use a prefix. In Linux, the jemalloc build system doesn't define a prefix by default. I think it would be simpler if all of these packages didn't build in such a way that they required a prefix. I wonder if that approach is feasible though?
I'm going to re-open this issue since we're introducing gentoo tree breakage here. I've been involved in dev-db/{mysql,mariadb} and dev-db/redis ebuilds, and I wouldn't have a big problem moving to this general approach. I just want others to chip in before hand.
I've now updated redis ebuilds in bug 444796 and bug 439354.
No more breakage as far as I can see. Closing (and sorry for somewhat abusing this ticket).
Thanks, Johan. Good job.
