Update qiskit.utils.wrap_method for Python 3.11

[wshanks]

Summary

Fixed the behavior of qiskit.utils.wrap_method when applied to a classmethod that is on type. Code changes in Python 3.11 and 3.11.1 revealed that wrap_method did not avoid calling a classmethod desscriptor's __get__ method. wrap_method was changed to use inspect.getattr_static to avoid calling __get__.

Details and comments

The poor handling of classmethod by wrap_method was first noticed with Python 3.11.1 as described in #9291. qiskit.test.decorators.enforce_subclasses_call wraps QiskitBaseTestCase.__init_subclass__ using wrap_method. Prior to 3.11.1, this meant wrapping object.__init_subclass__ (which is a builtin method that behaves differently), but in 3.11.1 unittest.TestCase added an __init_subclass__ method. This change meant that now wrap_method was wrapping a "real" classmethod (__init_subclass__ is a special case but it behaves enough like a classmethod for this issue).

In Python 3.11.0, the behavior of the bound method descriptor changed (specifically, it was this PR and the change to funcobject.c there). The following code illustrates the difference in behavior:

class A:
    def m(self):
        pass


print(A().m.__get__(None, A))

In Python 3.10, this gives

<bound method A.m of <__main__.A object at 0x7f942bc8ace0>>

while in Python 3.11 it gives

<function A.m at 0x7f432b0daa20>

So in Python 3.10 and earlier, the bound method's descriptor returns itself. With Python 3.11 though, the descriptor returns the unbound function.

Prior to the change here, wrap_method used object.__getattribute__ and type.__getattribute__ to retrieve the descriptor for the method being wrapped so that its __get__ could be called by the wrapping descriptor. For methods that are not in dir(type), it used object.__getattribute__(cls, name) to retrieve the descriptor in a loop over every cls in the wrapped class's MRO passing on AttributeError. object.__getattribute__(cls, name) checks type(cls) (so type) for a descriptor and otherwise checks cls.__dict__ for name, raising an AttributeError if name is not found (the code is here). By doing this loop over the MRO, name is eventually found. Note, however, that object.__getattribute__ follows a different path if name is a descriptor in type(cls) (so in type). To avoid this different path, wrap_method checked name against dir(type) and used type.__getattribute__ instead. type.__getattribute__(cls, name) loops over the MRO of cls looking for name and when it finds it calls .__get__(None, cls) on it (the code is here). This call returns a bound method version of the classmethod with cls bound and this is where the Python 3.11 change in behavior comes in. In qiskit.utils.classtools._WrappedMethod, __get__ tries to delegate to the wrapped method by calling .__get__ on it but for this type case that means calling .__get__ on a bound method instead of the classmethod and so for 3.11 getting back an unbound function that expects an extra first argument containing the cls.

So in Python 3.11.1 with the addition of __init_subclass__ to unittest.TestCase, the wrap_method call around __init_subclass__ on QiskitBaseTestCase meant that the definition of QiskitTestCase(QiskitBaseTestCase) now tried to call the wrapped __init_subclass__ as an unbound function without passing it the required cls argument and this led to the error. In 3.11.0, the handling of the descriptor did not matter because object.__init_subclass__ was the method being wrapped and it has a *args, **kwargs signature and does nothing, so it did not matter that it was invoked incorrectly (see here).

Closes #9291

[qiskit-bot]

Thank you for opening a new pull request.

Before your PR can be merged it will first need to pass continuous integration tests and be reviewed. Sometimes the review process can be slow, so please be patient.

While you're waiting, please feel free to review other open PRs. While only a subset of people are authorized to approve pull requests for merging, everyone is encouraged to review open pull requests. Doing reviews helps reduce the burden on the core team and helps make the project's code better for everyone.

One or more of the the following people are requested to review this:

• @Qiskit/terra-core

[Cryoris]

Nice this seems to work! The 3.11 tests are failing due to NumPy, but that should be fixed now.

[coveralls]

Pull Request Test Coverage Report for Build 3878635934

• 1 of 1 (100.0%) changed or added relevant line in 1 file are covered.
• 6 unchanged lines in 2 files lost coverage.
• Overall coverage decreased (-0.04%) to 84.569%

---

Files with Coverage Reduction	New Missed Lines	%
qiskit/pulse/library/waveform.py	3	91.49%
src/sabre_swap/mod.rs	3	98.84%

Totals
Change from base Build 3878466191:	-0.04%
Covered Lines:	64024
Relevant Lines:	75706

---

💛 - Coveralls

[wshanks]

While this lets the tests run, I think it is not quite right in general.

[wshanks]

Now I think the change is correct.

[mtreinish]

This LGTM, I didn't know about getattr_static but that looks like exactly what we were trying to hack together before (and what was causing issues for us on 3.11.1). I'm going to hold off on automerge until @jakelishman can take a look though since he wrote all of this decorator code (and I've lost most of the context I previously had for it).

[jakelishman]

This looks great, thanks Will. I think I didn't know about getattr_static when I wrote that first implementation, which is part of why it got super hacky.