I'll ping Mark when he's back.

IIUC, LOAD_METHOD_CLASS doesn't need to check cls.__dict__ at all. A single cls.tp_version_tag check is sufficient. So we can save on dict calculation and a branch.

🤖 New build scheduled with the buildbot fleet by @Fidget-Spinner for commit af0dd27 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

@erlend-aasland the more I look at this the more iffy my code feels.

I'm trying to think if this could cause any segfaults. E.g. it accesses a non-type object at offsetoff(PyTypeObject, tp_version_tag, but the object size is smaller than that location, and it just crashes. I will try to produce some code to test my theory.

Previously the dict check also served as an object identity check since dk_version is unique. With that, accessing tp_version_tag is always safe.

This is guaranteed to always work though if all objects are bigger than offsetoff(PyTypeObject, tp_version_tag). I dont know if thats the case though. Do you remember?

I dont know if thats the case though. Do you remember?

I do not, and unfortunately I do not have any time to investigate now :(

Thanks for the review Erlend, ultimately I decided to roll back the changes and make the checks stricter. My tests indicate we lost 0.1% of specialization hits, but the added safety is worth it. I can't prove that accessing tp_version_tag will never segfault in all cases.

Looks good, thanks.
Don't know what is up with the Windows builds 😕

FWIW, I can't repro the windows build errors locally. It seems to be affecting every PR right now.

Thanks Mark and Erlend for the reviews!