test_stl_caster_vs_stl_bind issue is because sq_item is not being set even though __getitem__ is set for the type VectorInt.

Maybe defining __getitem__ doesn't automatically set the value sq_item in PyPy.

Maybe defining __getitem__ doesn't automatically set the value sq_item in PyPy

It should, as long as __getitem__ is set before calling PyType_Ready

It should, as long as getitem is set before calling PyType_Ready

Ah, thanks. I think __getitem__ gets defined after PyType_Ready is called.

PyType_Ready is called by make_new_python_type which is called by generic_type::initialize method which is called by class_ constructor and the __getitem__ method is added by class_::def method.

Somebody more familiar with pybind11 internals should confirm though.

Slightly off-topic: @isuruf, could you please keep this branch around for a bit? I'm using it here.

cc @wjakob @jagerman this PR is needs some feedback on its way to add PyPy3.6 support, please.

ping

Ok, so what kind of feedback do you require? It's been ages since I looked at PyPy internals, so I don't think I could help much with that. The patch looks reasonable to me, but I suppose it does not fully work yet?

Thanks @mattip. One more issue to fix #2146 (comment). I'm afraid, you might be the best person to fix that as well.

If I use the suggestion by @mattip in https://bitbucket.org/pypy/pypy/issues/2482/cpyext-tp_basicsize-only-considers-first, I can get the test to pass, but it segfaults other tests.

https://bitbucket.org/pypy/pypy/commits/8175c5e20480 mentions that

but perhaps we should support PyBaseObject_Type too?

I'd like to briefly check in about the status of the PR -- are we stalled on landing improvements to PyPy? There are already some good improvements, so one option would be to merge part of the changes and leave the rest pending.

Yes. I've marked a couple of tests (related to multiple inheritance) as bug_in_pypy until we can fix it in PyPy. (I'm not entirely sure it's a bug in PyPy, but it looks like it and @mattip can confirm)

CI is passing (with two expected failures), so we can merge right now to avoid any regressions in the future and work on fixing the multiple inheritance issue.

Ok, will do as soon as the PR is merge-able (conflicts).

This is ready now

Merged, let's track future improvements in a separate PR.

Tests are failing on master, for some reason.

>       assert msg(exc_info.value) == "m.class_.Pet.__init__() must be called when overriding __init__"
E       assert --- actual / +++ expected
E         - Pet.__init__() must be called when overriding __init__
E         + m.class_.Pet.__init__() must be called when overriding __init__
E         ? +++++++++

So the raised exception looks different than on CPython...

Something something __qualname__?

Should we mess with __qualname__ or instead do something like this:

assert msg(exec_info.value) in [ cpython_error_string, pypy_error_string ]

Should we mess with __qualname__ or instead do something like this:

assert msg(exec_info.value) in [ cpython_error_string, pypy_error_string ]

Depends a bit on whether this is a PyPy thing, or whether it's a failure of pybind11 to not set tp_name correctly?

If I'm reading this right, pybind11, since python 3.3 doesn't touch __qualname__.