Fix memory leak during argument parsing.

[jamadden]

The bug was on Python 2, or PyPy < 7.2

Fixes #3578

[auvipy]

any tests need to be updated?

[jamadden]

Yeah, ideally it could use tests; it didn't seem to have any before, though, and I'm not sure how to go about adding any.

Possibly even more ideally this would be integrated, somehow, with the refnanny system, which would probably just make tests happen. But that's beyond my familiarity with this code base. So this is just the minimal fix for now.

[da-woods]

refnanny is mainly for the Cython generated code. It assumes that the utility code is right. So it wouldn't be the right thing to use here. I think it's probably OK for this to be test-less.

[da-woods]

From the test failures I think PyObject_GetItem was right because it'll raise exceptions (and the PyDict version won't. I think the you probably want a decref to make the reference "borrowed" instead.

[scoder]

Oh well… #3091

[scoder]

I think it can be ok to add the refnanny also to some non-trivial utility functions, but this isn't a candidate.

[scoder]

I think the you probably want a decref to make the reference "borrowed" instead.

PyPy won't be happy about that. It's generally not happy about borrowed references, but this would be even worse.

I think the best fix would be to step away from borrowed references in the kwargs parsing code, but that's probably not for today any more.

[scoder]

I think the best fix would be to step away from borrowed references in the kwargs parsing code

Or, drop Py2 support. :)

[scoder]

Thanks for the analysis and the proposed fix. I found a more correct implementation in #3593.