[Python-Dev] PyWeakref_GetObject() borrows its reference from... whom?

[Nick Coghlan]

On 11 October 2016 at 08:49, Guido van Rossum <guido at python.org> wrote:
> On Mon, Oct 10, 2016 at 1:57 PM, Larry Hastings <larry at hastings.org> wrote:
>> For now I'm going to leave the names as-is and just change the semantics
>> everywhere.  If this approach really works, and if we decide we want to
>> merge it back into trunk--and those are both still big if's--we can revisit
>> this decision.
>
> Changing something as subtle as this to an API sounds really bad even
> in an experimental branch. But it's your branch and you can do what
> you want (then again, why did you bring it up here? :-).

I'm guessing Larry was hoping someone else would see a possible
solution that he'd missed.

On that front, while I don't see an alternative to making borrowed
references real references in a GIL-free Python, I do think we may
need to keep the borrowed/new distinction for the sake of folks
writing cross-version compatible code. To explain why, consider these
scenarios:

* CPython with a GIL:
  - new references MUST be decref'ed
  - borrowed references MUST NOT be decref'ed

* CPython without a GIL:
  - new references MUST be decref'ed
  - borrowed references MUST be decref'ed

Assuming that the GILectomy does end up needing to implicitly treat
all borrowed references as new references, that suggests we're going
to need APIs like "PyBorrowed_DECREF" to provide the "only decref when
built without the GIL" semantics, and the rules for GIL-independent
source code would become:

- new references must be decref'ed using the normal APIs
- borrowed references must be decref'ed using the PyBorrowed* APIs
(with a no-op compatibility shim for older GIL-only CPython versions)

Regards,
Nick.

-- 
Nick Coghlan   |   ncoghlan at gmail.com   |   Brisbane, Australia