GH-108716: Turn off deep-freezing of modules.

[markshannon]

Performance impact is neutral, maybe a little positive, but noisy

Startup might be a bit slower: no-site is 2% faster, but normal startup 5% slower.

However, without deepfreeze it is much easier to implement things like: faster-cpython/ideas#566 or #99555 which should more than compensate.

• Issue: Remove deep-freezing of code objects and modules. #108716

[brandtbucher]

We may want to benchmark this on the other machines too, but in general I'm excited by this!

[markshannon]

Windows 0% slower, Mac 1% faster

[vstinner]

FYI I merged my PR #108741 to fix a race condition in make regen-all in the main branch. Backports to 3.11 and 3.12 are on-going.

[vstinner]

Windows 0% slower, Mac 1% faster

Oh right, for the comparison to the main branch, I see:

Windows:

• python_startup_no_site: 16.4 ms => 16.2 ms: 1.01x faster
• python_startup: 19.3 ms => 19.8 ms: 1.02x slower

macOS:

• python_startup_no_site: 9.40 ms => 9.15 ms: 1.03x faster
• python_startup: 11.5 ms => 11.8 ms: 1.02x slower

To be honest, it's surprising that disabling an optimization supposed to make Python startup actually... has no effect (or might be faster, but that can be noise of the benchmark?).

[vstinner]

What I can say is that it has a nice impact on build performance! :-)

Comparison main => this PR:

• gcc -O0 (pydebug): 0m20,152s => 0m13,200s
• gcc -O3 (release): 0m39,349s => 0m35,819s

Python configured with ./configure --with-pydebug or ./configure (without LTO, without PGO). I ran time make -j14 to measure build time on my laptop with 12 threads (6 CPU cores). Quick-and-dirty non scientific benchmark, I only ran make once to measure :-)

[brandtbucher]

Looks good to me!

[vstinner]

It's unclear to me if Python/deepfreeze/deepfreeze.c is still used or not. When I build Python with make, it's not built. But make-regen creates the file. Is it intended?

# Dependencies which can add and/or remove _Py_ID() identifiers:
# - deepfreeze.c
# - "make clinic"
.PHONY: regen-global-objects
regen-global-objects: $(srcdir)/Tools/build/generate_global_objects.py $(DEEPFREEZE_C) clinic
$(PYTHON_FOR_REGEN) $(srcdir)/Tools/build/generate_global_objects.py

[markshannon]

Python/deepfreeze/deepfreeze.c shouldn't be used any more. I don't want to remove the tools yet, but we probably should remove the build target.

[vstinner]

It seems like the Windows build system was not updated. deepfreeze.c is still generated on Windows, no?

This rule in PCbuild/_freeze_module.vcxproj:

 <TargetName="_RebuildDeepFrozen"AfterTargets="_RebuildFrozen"DependsOnTargets="FindPythonForBuild"Condition="$(Configuration) != 'PGUpdate'">
<!-- BEGIN deepfreeze rule -->
<ExecCommand='$(PythonForBuild) "$(PySourcePath)Tools\build\deepfreeze.py" ^ "$(PySourcePath)Python\frozen_modules\importlib._bootstrap.h:importlib._bootstrap" ^ "$(PySourcePath)Python\frozen_modules\importlib._bootstrap_external.h:importlib._bootstrap_external" ^ "$(PySourcePath)Python\frozen_modules\zipimport.h:zipimport" ^ "$(PySourcePath)Python\frozen_modules\abc.h:abc" ^ "$(PySourcePath)Python\frozen_modules\codecs.h:codecs" ^ "$(PySourcePath)Python\frozen_modules\io.h:io" ^ "$(PySourcePath)Python\frozen_modules\_collections_abc.h:_collections_abc" ^ "$(PySourcePath)Python\frozen_modules\_sitebuiltins.h:_sitebuiltins" ^ "$(PySourcePath)Python\frozen_modules\genericpath.h:genericpath" ^ "$(PySourcePath)Python\frozen_modules\ntpath.h:ntpath" ^ "$(PySourcePath)Python\frozen_modules\posixpath.h:posixpath" ^ "$(PySourcePath)Python\frozen_modules\os.h:os" ^ "$(PySourcePath)Python\frozen_modules\site.h:site" ^ "$(PySourcePath)Python\frozen_modules\stat.h:stat" ^ "$(PySourcePath)Python\frozen_modules\importlib.util.h:importlib.util" ^ "$(PySourcePath)Python\frozen_modules\importlib.machinery.h:importlib.machinery" ^ "$(PySourcePath)Python\frozen_modules\runpy.h:runpy" ^ "$(PySourcePath)Python\frozen_modules\__hello__.h:__hello__" ^ "$(PySourcePath)Python\frozen_modules\__phello__.h:__phello__" ^ "$(PySourcePath)Python\frozen_modules\__phello__.ham.h:__phello__.ham" ^ "$(PySourcePath)Python\frozen_modules\__phello__.ham.eggs.h:__phello__.ham.eggs" ^ "$(PySourcePath)Python\frozen_modules\__phello__.spam.h:__phello__.spam" ^ "$(PySourcePath)Python\frozen_modules\frozen_only.h:frozen_only" ^ "-o" "$(PySourcePath)Python\deepfreeze\deepfreeze.c"'/>
<!-- END deepfreeze rule -->
</Target>

If you want to keep the infra for now, maybe this rule should just be commented with a reference to gh-108716?

[gvanrossum]

This PR probably should also have updated Python/deepfreeze/README.txt to indicate that it is (currently) no longer used. It should also have updated various sections in Makefile.pre.in (including some comments) that still reference deepfreeze -- those are just distractions at this point (until we decide to actually deep-freeze some other objects, which doesn't look like it's going to happen for 3.13 at least). EDIT: See gh-116919