[rb-general] got a reproducible gcc
Bernhard M. Wiedemann
bernhardout at lsmod.de
Thu Jun 28 11:08:49 CEST 2018
over the past weeks, I made great progress towards reproducible gcc-8 builds
Apart from this 8 year old libtool patch
and embedded mtime timestamps in ada-related files
that could be avoided with a single touch -d at 1
There is the issue with 'profiledbootstrap'
described in general at
(including a surprisingly simple example)
and our related bug
One interesting problem with PGO is that you can avoid irreproducibility
by disabling it, but then you lose 8% performance (measured for gcc8).
That is a hard trade-off.
But now I found that disabling ASLR together with non-parallelized
profiling and gcc CFLAGS: -fprofile-update=atomic -fprofile-generate
still gives a reproducible build with PGO enabled.
maybe the -fprofile-update=atomic CFLAGS is not actually needed.
I hope in the future our (SUSE (maybe with upstream)) compiler-guys will
find ways to make the profiling code even more deterministic.
E.g. there seems to be strange merging code for profile counters that is
not commutative so a+b != b+a so that ordering matters.
All my testing is using our highly normalized build-env
(same user, build-path, umask, locale, timezone)
and likely it requires this patch https://savannah.gnu.org/bugs/?52076
to keep the make order stable
With all that, it is already possible to build bit-identical
profile-optimized gcc8 packages.
More information about the rb-general