> Is the behavior of tp_clear the key to this bug?
Once func_clear(my_func) is called, calling my_func() will crash: my_func() is unsuable.
Because of a complex dance involving borrowed references, the function is called *after* it's cleared.
Pablo's PR 15645 works around the problem by detecting this very specific case: call a function "after it's cleared" (while it's being garbage collected in practice).
Honestly, I'm still not sure how *exactly* the bug is triggered. I tried but failed to reproduce the initial cffi crash. reproducer.tar.gz is still a giant piece of code. I failed to simplify it to a few objects.
According to Pablo, gc_crash.py is a *different* (but similar) bug.
I would only be confident in approving a fix once I would be confident that I understood *exactly* how the initial bug (reproducer.tar.gz) occurred.