I propose to add type slots for magic methods introduced by PEP 560. In the brief discussion on the mailing list Guido OK'd the idea: https://mail.python.org/pipermail/python-dev/2017-December/151262.html I'll submit a PR that implements this proposal by adding new 'tp_as_class' slot, that points to a PyClassMethods struct, that has two fields: `cm_getitem` and `cm_mro_entries`. Interestingly, when __class_getitem__ is implemented through the type/slots machinery it becomes a bit faster. Given the following microbenchmark: class Foo: def __class_getitem__(cls, o): return o for _ in range(3): t = time.monotonic() for i in range(10 ** 8): assert i == Foo[i] print(f'{time.monotonic() - t:.3f}s') I see these numbers for the master branch: 17.161s 17.024s 16.530s and these for the C-slots branch: 15.010s 15.693s 15.035s
This isn't so easy. You can't just access new slot, because this is binary incompatible. You should add new type flag, say Py_TPFLAGS_HAVE_CLASS, set it for all classes with tp_as_class, and guard access to tp_as_class with if (!PyType_HasFeature(tp, Py_TPFLAGS_HAVE_CLASS))
Serhiy, I've tested METH_STATIC approach and it works just fine. Thanks for pointing that out! I agree that adding slots for something that's already possible is an overkill, so let's close this issue.
