BUG: np.array cleared errors occured in PyMemoryView_FromObject

[superbobry]

This commit changes PyMemoryView_FromObject call sites to only
clear BufferError and TypeError. Any other error will be raised
to the caller of np.array.

---

I'd be happy to add a test for this, but it requires writing an extension class
with a custom buffer interface implementation. Should I do that?

[eric-wieser]

This change isn't enough - there are two calls to GetBuffer done in succession above - you need to insert a check for an error between them else you risk hitting a SystemError.

[superbobry]

Done, apologies for the delay.

[mattip]

It does seem that this logic is sufficiently convoluted to justify a test, but it seems such a test would require a custom bf_getbuffer function and a lot of C code. If you are comfortable working with PyPy (which we test in the CI), it exposes bf_getbuffer as a __buffer__ attribute on pure python classes that inherit from __pypy__.bufferable, so theoretically you could write a pure-python test and mark it @pytest.skipif(not IS_PYPY, reason='only PyPy has bufferable class), write a ndarray subclass, override the __buffer__ attribute, and check that it asserts properly.

Let me know if you want help with that

[mattip]

Thanks @superbobry

[superbobry]

@mattip sorry for not commenting earlier. I implemented a test but discovered more calls to PyObject_GetBuffer which need to be handled. I think it might be best to revert this PR until all of the calls are covered.

[mattip]

@superbobry is there harm in what I merged? It is fine to fix more in a follow-on, as long as the current state has no regressions from the previous one.

[mattip]

Fwiw, the xref to the pandas issue is wrong, that was due to #13399 and a fix proposed in #13663

[superbobry]

No, no harm, but the commit description might be misleading :)