* Remove `@suppress_immortalization` decorator
* Make suppression flag per-thread instead of per-interpreter
* Suppress immortalization in `eval()` to avoid refleaks in three tests
(test_datetime.test_roundtrip, test_logging.test_config8_ok, and
test_random.test_after_fork).
* frozenset() is constant, but not a singleton. When run multiple times,
the test could fail due to constant interning.
Deferred reference counting is not fully implemented yet. As a temporary
measure, we immortalize objects that would use deferred reference
counting to avoid multi-threaded scaling bottlenecks.
This is only performed in the free-threaded build once the first
non-main thread is started. Additionally, some tests, including refleak
tests, suppress this behavior.
The test suite fetches the C recursion limit from the _testcapi
extension module. Test extension modules can be disabled using the
--disable-test-modules configure option.
Use support.infinite_recursion() in test_recursive_pickle() of
test_functools to prevent a stack overflow on "ARM64 Windows
Non-Debug" buildbot.
Lower Py_C_RECURSION_LIMIT to 1,000 frames on Windows ARM64.
The registry() method of functools.singledispatch() functions checks now
the first argument or the first parameter annotation and raises a TypeError if it is
not supported. Previously unsupported "types" were ignored (e.g. typing.List[int])
or caused an error at calling time (e.g. list[int]).
In order to fix a bug in the 3.9 branch in #29394, more tests were added to
``test_functools.py`` to ensure that ``singledispatchmethod`` still correctly
wrapped a target method, even if the target method had already been wrapped by
multiple other decorators. This PR brings the new tests into the 3.11 and 3.10
branches as well.
In Python 3.8 and 3.9, stacking `@functools.singledispatchmethod` on top of
`@classmethod` or `@staticmethod` caused an exception to be raised if the
method was registered using type-annotations rather than
`@method.register(int)`. This was not caught by unit tests, however, as the
tests only tested the `@method.register(int)` way of registering additional
implementations. The bug is no longer present in Python 3.10+, but
`test_functools.py` is still lacking regression tests for these cases. This
commit adds these test cases.
The hard part was making all the tests pass; there are some subtle issues here, because apparently the future import wasn't tested very thoroughly in previous Python versions.
For example, `inspect.signature()` returned type objects normally (except for forward references), but strings with the future import. We changed it to try and return type objects by calling `typing.get_type_hints()`, but fall back on returning strings if that function fails (which it may do if there are future references in the annotations that require passing in a specific namespace to resolve).
The topological sort functionality that was introduced initially in the
functools module has been moved to a new graphlib module to
better accommodate the new tools and keep the original scope of the
functools module.
Sam Gross
dgpb
Jelle Zijlstra
Serhiy Storchaka
Bénédikt Tran
Nikita Sobolev
CF Bolz-Tereick
Erlend E. Aasland
Victor Stinner
Furkan Onder
Ammar Askar
Mark Shannon
Tyler Smart
Pieter Eendebak
Carl Meyer
Irit Katriel
Jingxuan He
Thaddeus1499
Christian Heimes
Shantanu
Russel Webber
Yurii Karabas
Alex Waygood
Raymond Hettinger
Erlend Egeberg Aasland
Pablo Galindo
Hai Shi
Batuhan Taskaya