Add `PyObject_CallFinalizerFromDealloc` to Stable ABI

[encukou]

As requested here: python/cpython#146063

I don't like that we ended up here, but, we need this for the current Stable ABI.

• @encukou
• @erlend-aasland
• @serhiy-storchaka
• @vstinner
• @zooba

[davidhewitt]

I don't like that we ended up here, but, we need this for the current Stable ABI.

Can you elaborate what the concerns are? Maybe we can find another way for PyO3...

[zooba]

Do we remember the reason we had to distinguish between calling during dealloc vs. just handling the refcnt == 0 case in PyObject_CallFinalizer? As Sam pointed out on the issue, the latter function doesn't really get much use, but it would be nicer to have that one in the stable ABI (and alias ...FromDealloc to it).

[zooba]

Maybe we can find another way for PyO3

I don't think there is another way, which is why Petr doesn't like it but is proposing it anyway. Unless we've got some way to safely call the finalizer before calling dealloc, but if we did then I assume we'd already be doing that.

[da-woods]

The only other thing you could do is just to manually let people set the finalized flag. That's the but that you can't do in the limited API. Everything else you need (check the finalized flag, mess around with the refcount, get the tp_finalize slot and call it) could be done explicitly already.

Personally I think this is worse than exposing PyObject_CallFinalizerFromDealloc (just because it lets people arbitrarily mess with the finalized flag at any time) but it is the other option

[encukou]

Can you elaborate what the concerns are?

More of a generic sigh about the state of the finalization API :)