ENH: add support for PEP 803 `abi3t` with Python 3.15.0b2+ by mgorny · Pull Request #856 · mesonbuild/meson-python

mgorny · 2026-06-23T14:22:41Z

Add tag/extension support and initial tests for abi3t as specified in PEP 803 and implemented in Python 3.15.0b2. This is currently limited to building abi3t extensions from freethreading builds of Python, since Meson does not support forcing abi3t builds explicitly right now. See mesonbuild/meson#15637 for the relevant discussion.

The additional test case utilizes the new API introduced in PEP 793 and PEP 820, and therefore requires Python 3.15.0b2. The relevant test tests wheel correctness both with GIL-enabled and free-threading builds of Python 3.15.0b2.

The behavior for older versions of Python remains unchanged.

Add tag/extension support and initial tests for `abi3t` as specified in PEP 803 and implemented in Python 3.15.0b2. This is currently limited to building `abi3t` extensions from freethreading builds of Python, since Meson does not support forcing `abi3t` builds explicitly right now. See mesonbuild/meson#15637 for the relevant discussion. The additional test case utilizes the new API introduced in PEP 793 and PEP 820, and therefore requires Python 3.15.0b2. The relevant test tests wheel correctness both with GIL-enabled and free-threading builds of Python 3.15.0b2. The behavior for older versions of Python remains unchanged. Signed-off-by: Michał Górny <mgorny@quansight.com>

mgorny · 2026-06-23T14:23:32Z

I didn't touch the test matrix, since testing on 3.15 is already being addressed by #853.

Signed-off-by: Michał Górny <mgorny@quansight.com>

dnicolodi · 2026-06-23T16:15:18Z



-@pytest.mark.skipif(NOGIL_BUILD and CYTHON_VERSION < (3, 1, 0),
+@pytest.mark.skipif(mesonpy._is_free_threaded() and CYTHON_VERSION < (3, 1, 0),


I think the NOGIL_BUILD was more self explanatory. Reading the function call mesonpy._is_free_threaded() I would imagine that what has the property of being free threaded is mesonpy, which does not make much sense.

I would keep the old constant.

Do you think it would be better to put it as a constant in mesonpy? I was torn between using a function and a constant, given this is fixed value.

Constant seems fine to me. Slight preference for FREE_THREADED_BUILD over NOGIL_BUILD (the latter name will fade over time I think), but I don't care that much either way.

@rgommers IIRC it was you that introduced the NOGIL_BUILD constant 🙂

Guilty:) I guess that was early days, before the Steering Council pronounced that that was a negative name and they wanted everyone to use free-threading/free-threaded everywhere.

For now, I've focused on the changes relevant to the PR. Do you want me to change the constant name, and possibly move it to mesonpy/__init__.py here, or separately?

For now, I've focused on the changes relevant to the PR.

That sounds fine to me - small diffs are nice

Co-authored-by: Daniele Nicolodi <daniele@grinta.net>

This reverts commit a23fe90. Signed-off-by: Michał Górny <mgorny@quansight.com>

Signed-off-by: Michał Górny <mgorny@quansight.com>

rgommers · 2026-06-24T10:58:23Z

Downstream testing, and releasing after rc1 in early August, is currently blocked on this, so I added a v0.21.0 tag - we should do that next month I think.

This reverts commit 0d43c65. Signed-off-by: Michał Górny <mgorny@quansight.com>

Signed-off-by: Michał Górny <mgorny@quansight.com>

rgommers

A couple of small issues in the test extension. The rest looks right; I refreshed my memory on Py_TARGET_ABI3T and that should be added automatically in Python.h for a free-threaded interpreter if Py_LIMITED_API is defined - so all good there.

rgommers · 2026-06-24T13:40:35Z

+PyABIInfo_VAR(abi_info);
+
+static PySlot module_slots[] = {
+    PySlot_STATIC_DATA(Py_mod_name, "plat"),


"plat" should be "module" right, to match the filename and the PyModExport_module below?

To be honest, I've copied the name from the limited-api test. I have no clue why it's "plat" there, so I've assumed there must be a reason :-).

Changed. I suspect I'm going to follow this up with "misc. fixes" PR xD.

You changed it in the old test, but not here yet

Uh, sorry about that. Should be correct now.

dnicolodi · 2026-06-24T14:05:25Z

+    if '__pypy__' in sys.builtin_module_names:
+        abi = ABI
+    elif NOGIL_BUILD and sys.version_info >= (3, 15):
+        abi = 'abi3.abi3t'


I think that, if this works, it does by chance. The fake data passed to the wheel builder uses ABI3SUFFIX which is defined at the top of this module. That that resolves to abi3t because of the orders in which the suffixes are listed in importlib.machinery.EXTENSION_SUFFIXES. Is the order guaranteed or should we try to make this more robust? If the order is guaranteed, I would still add a comment in the get_abi3_suffix() explaining what it is expected and why it works.

A comment would be great. My understanding is that it is guaranteed - and Meson relies on it as well, which came up in the PEP 803 discussions. (that said, future-proofing might be wise).

Yes, it is guaranteed, CPython is driving module loading on it. For GIL-enabled, it's:

>>> importlib.machinery.EXTENSION_SUFFIXES ['.cpython-315-x86_64-linux-gnu.so', '.abi3.so', '.abi3t.so', '.so']

It's ordered from most specific (i.e. presumably most performant) to least specific, to .so legacy fallback. I'll add a comment.

Doesn't this mean that get_abi3_suffix() on Python 3.15t returns .abi3.so and this test fails because when building on Python 3.15t the extension modules should be suffixed with .abi3t.so?

Doesn't this mean that get_abi3_suffix() on Python 3.15t returns .abi3.so and this test fails because when building on Python 3.15t the extension modules should be suffixed with .abi3t.so?

Nope, EXTENSION_SUFFIXES is different on 3.15t:

Python 3.15.0b2 free-threading build (main, Jun 4 2026, 14:23:26) [Clang 21.0.0 (clang-2100.1.1.101)] on darwin Type "help", "copyright", "credits" or "license" for more information. >>> import importlib.machinery >>> importlib.machinery.EXTENSION_SUFFIXES ['.cpython-315t-darwin.so', '.abi3t.so', '.so']

This is actually explicitly specified in PEP 803: https://peps.python.org/pep-0803/#the-abi3t-wheel-and-filename-tags

And yes, abi3.so is included on 3.14t and 3.13t and that's probably a bug but it was only noticed in the discussion around PEP 803 after 3.14t came out. Arguably we probably should have fixed this in the lead-up to 3.13t but no one noticed. I have no idea if removing .abi3.so from EXTENSION_SUFFIXES on 3.14t would have fallout.

Signed-off-by: Michał Górny <mgorny@quansight.com>

rgommers · 2026-06-24T18:43:08Z

This is looking pretty good now. Are you expecting it to be squash-merged, or do you want to rewrite history? (we usually do the latter, and rebase-merge)

mgorny commented Jun 23, 2026

View reviewed changes

Comment thread mesonpy/__init__.py Outdated

mgorny added 2 commits June 23, 2026 17:14

Move the version check to a helper function

0d43c65

Signed-off-by: Michał Górny <mgorny@quansight.com>

Replace the inline free-threaded checks with a function as well

a23fe90

Signed-off-by: Michał Górny <mgorny@quansight.com>

dnicolodi requested changes Jun 23, 2026

View reviewed changes

mgorny and others added 4 commits June 23, 2026 19:33

Apply suggestions from code review

bf0fb07

Co-authored-by: Daniele Nicolodi <daniele@grinta.net>

Revert "Replace the inline free-threaded checks with a function as well"

c20b470

This reverts commit a23fe90. Signed-off-by: Michał Górny <mgorny@quansight.com>

Rename the variable to abi3t

27e23c8

Signed-off-by: Michał Górny <mgorny@quansight.com>

Wrap the long text

1a16140

Signed-off-by: Michał Górny <mgorny@quansight.com>

dnicolodi reviewed Jun 24, 2026

View reviewed changes

Comment thread mesonpy/__init__.py Outdated

Rewrap message

66ebdd1

Signed-off-by: Michał Górny <mgorny@quansight.com>

dnicolodi reviewed Jun 24, 2026

View reviewed changes

Comment thread mesonpy/__init__.py Outdated

rgommers mentioned this pull request Jun 24, 2026

TST: use PyModExport hook in one test package for Python >=3.15 #833

Closed

rgommers added the enhancement New feature or request label Jun 24, 2026

rgommers added this to the v0.21.0 milestone Jun 24, 2026

mgorny added 2 commits June 24, 2026 13:43

Revert "Move the version check to a helper function"

f18d1ca

This reverts commit 0d43c65. Signed-off-by: Michał Górny <mgorny@quansight.com>

Shorten the version check to just (3, 15)

e56474b

Signed-off-by: Michał Górny <mgorny@quansight.com>

rgommers requested changes Jun 24, 2026

View reviewed changes

dnicolodi reviewed Jun 24, 2026

View reviewed changes

mgorny added 4 commits June 24, 2026 18:24

Correct PySlot_DATA usage

ede9744

Signed-off-by: Michał Górny <mgorny@quansight.com>

Add a comment to get_abi3_suffix()

a0aa5e1

Signed-off-by: Michał Górny <mgorny@quansight.com>

Fix module name

1657f0b

Signed-off-by: Michał Górny <mgorny@quansight.com>

Fix module name in the correct test

f816e00

Signed-off-by: Michał Górny <mgorny@quansight.com>



		@pytest.mark.skipif(NOGIL_BUILD and CYTHON_VERSION < (3, 1, 0),
		@pytest.mark.skipif(mesonpy._is_free_threaded() and CYTHON_VERSION < (3, 1, 0),

Uh oh!

Conversation

mgorny commented Jun 23, 2026

Uh oh!

mgorny commented Jun 23, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

rgommers commented Jun 24, 2026

Uh oh!

rgommers left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

rgommers commented Jun 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants