The backtrace says it is crashing in [window close] in FigureManager_dealloc. In FigureManager_new, the self->window is allocated, and in FigureManager_init it is initialized using initWithContentRect. If we raise from __init__ immediately, then the super-init is not called and the window is not initialized. Presumably an uninitialized window cannot be closed, but I'm not sure how to check if it has been.

It seems like the real issue is a misuse/disuse of tp_alloc, tp_init, tp_finalize, and tp_dealloc.

The window ought to be allocated in a custom tp_alloc. Currently, this is done in tp_init.

The tp_init should call some init method on the window.

The window should be closed in tp_finalize. This should be decoupled from releasing it. I think that can be done by setting isReleasedWhenClosed to false during tp_init.

The window should be released in tp_dealloc.

Is that all correct?

@joelfrederico I milestoned this as 3.6 as you did not label this as a regression, but if there is an easy fix to reduce segfaults, we can definitely backport it.

Checking, the state is current the same, however if I am reading this correctly we only have the issue if we raise in the __init__?