On 14. 08. 24 11:04, Zbigniew Jędrzejewski-Szmek wrote:
On Wed, Aug 14, 2024 at 01:22:19AM +0200, Miro Hrončok wrote:
On 13. 08. 24 21:45, Zbigniew Jędrzejewski-Szmek wrote:
On Tue, Aug 06, 2024 at 11:29:01AM +0200, Miro Hrončok wrote:
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.
I think the argument about pip is important: with -O3, the users will
get a "reasonable default". Local pip installs intended for computations
are certainly better with '-O2' or '-O3' rather than the compiler default
of -O0.
But then the setup becomes inconsistent. The flags are:
$ python -c "import sysconfig; print(sysconfig.get_config_var('CFLAGS'))"
-fno-strict-overflow -Wsign-compare -DDYNAMIC_ANNOTATIONS_ENABLED=1 -DNDEBUG
-fcf-protection -fexceptions -fcf-protection -fexceptions -fcf-protection
-fexceptions
-O3
I'd argue that the goal is to have flags that provide a reasonable
efficient default that integrates well with the cpython code provided
by the distribution,
-fno-omit-frame-pointer -mno-omit-leaf-frame-pointer
should be included too. (Those flags allow whole-system profiling of
the system, and this works best if all of the system has frame pointers,
so the extensions should have them too.)
I'd argue that whoever wants to profile their system using frame pointers is
also capable of setting the flags for the extension modules their build
themselves.
Let me restate my argument.
I think that there are two possibilities that make sense:
in one, the flags that are exposed are the minimum that is required for
ABI compatibility with the distro build.
In the other, we additionally provide flags that give the most commonly
useful and reasonable defaults.
In the first case, the user has a very clear state, but they have to
include additional flags to get "reasonable behaviour", e.g. efficiency.
In the second case, things work nicely out-of-the-box, but custom and
minimalistic builds require an additional step. I don't think it makes
sense to mix the two, i.e. provide _some_ flags but not all flags for
the common scenario.
Benchmarking with frame pointers is a goal for both Fedora and the
Python upstream. The whole system is built with FPs, and if the user
tries to profile, that'll work as expected if all code is compiled
appropriately. We can't expect the user to go back and recompile all
code once they realize that some profiling is needed. (This is the
same argument as in the discussions about distro builds.)
If that's correct, the fix is to add the flags to %extension_cflags, right?
Thus, if we go for the second case with -O3, the FP flags should be
included too.
(And if we go for the first case, at least
'-Wsign-compare -DDYNAMIC_ANNOTATIONS_ENABLED=1 -DNDEBUG' should
be dropped, because those things have no effect on the ABI.)
I've checked and it's Python that inserts those, not us.
--
Miro Hrončok
--
Phone: +420777974800
Fedora Matrix: mhroncok
--
_______________________________________________
python-devel mailing list -- python-devel@lists.fedoraproject.org
To unsubscribe send an email to python-devel-le...@lists.fedoraproject.org
Fedora Code of Conduct:
https://docs.fedoraproject.org/en-US/project/code-of-conduct/
List Guidelines: https://fedoraproject.org/wiki/Mailing_list_guidelines
List Archives:
https://lists.fedoraproject.org/archives/list/python-devel@lists.fedoraproject.org
Do not reply to spam, report it:
https://pagure.io/fedora-infrastructure/new_issue
