On 06. 08. 24 1:20, Miro Hrončok wrote:
Hello Pythonistas.
For years, the CFLAGS embedded in Python sysconfig contained -O2 in Fedora. This was how Python was built and by default, all flags used to build Python were embedded.
Later, the flag was removed in Fedora 39 via this change: https://fedoraproject.org/wiki/Changes/Python_Extension_Flags_Reduction
We wanted to remove as many flags as possible, with this motivation:
""" Python developers will get more upstream-like experience when building Python extension modules and also closer to building vanilla C programs. """
Note that removing -O2 specifically was not the primary motivation, but the removal was intentional at that time.
Now we build Python 3.13 with -O3 via this change: https://fedoraproject.org/wiki/Changes/Python_built_with_gcc_O3
The change proposal said:
""" Other Python extension modules will remain bulidng as before, e.g. in RPM packages, they will still be built with -O2... """
However, I made a mistake and the -O3 flag leaked into sysconfig CFLAGS.
The good news is this does not seem to affect RPM packages, they are still being built with -O2, like this:
<flags embedded in sysconfig> <$CFLAGS from automatic %set_build_flags>
E.g. ... -fcf-protection -fexceptions -O3 -O2 -flto=auto ...
The latter flag wins.
OTOH users building their own extension modules will get -O3.
This is not what was intended. However, "Python developers will get more upstream-like experience" is more true now, because upstream Python builds use -O3:
$ podman run --rm -ti python:3.12 /usr/bin/bash # python
import sysconfig sysconfig.get_config_var('CFLAGS')
'-fno-strict-overflow -Wsign-compare -DNDEBUG -g -O3 -Wall'
So the question is:
Do we keep -O3 for user-built extension modules for speed and upstream-like experience? (I would update the -O3 change proposal.)
Or do we loose the flag, as currently documented?
Alternatively, do something else entirely (e.g. embed -O2, or other flag...)?
I slept on it and I support keeping the -O3 flag in sysconfig CFLAGS.
The original motivation for Python_Extension_Flags_Reduction was:
""" Python developers will get more upstream-like experience when building Python extension modules and also closer to building vanilla C programs. """
Keeping -O3 supports "more upstream-like experience". Getting rid of it supports "closer to building vanilla C programs".
When choosing from the two, I prefer the first one.
Every time somebody pip installs something without a wheel, or even builds their own extension module code, on CI etc. they would benefit from the added speed. If we want Fedora to succeed on the CI field, we should be competitive.
The benchmarks used to measure the -O2 -> -O3 transition were also measured with -O3 in sysconfig CFLAGS, so if they build some custom C code, it might have impacted the results.
If nobody speaks up against keeping the flag, I'll amend the change proposal and submit it for FESCo re-approval.