annotationlib: namespace can be non-`dict`

[eliegoudout]

Hello!

I think the following, made so that get_annotate_function is usable in metaclasses' __new__, should check for Mapping since __prepare__ can provide any mapping (see picture from PEP3115.

cpython/Lib/annotationlib.py

Line 630 in a81232c

if isinstance(obj, dict):

Linked PRs

• gh-132426: get_annotate_function vs. mapping namespaces #132427
• gh-132426: Add get_annotate_from_class_namespace replacing get_annotate_function #132490

[eliegoudout]

However, I now realize that this is a major issue since there might be a conflict between the Mapping subtype's own annotations and the annotations it contains as a mapping for another type's namespace.

Shouldn't we separate get_annotate_function into get_annotate_function and get_namespace_annotations? Or maybe add a keyword get_annotate_function(obj, *, is_namespace: bool)?

[picnixz]

cc @JelleZijlstra

[JelleZijlstra]

Good point! I think what I'd like to do is drop get_annotate_function completely (it's now equivalent to a simple getattr call and doesn't need to be a function) and add a new function get_annotate_from_class_namespace.

Your proposed change is a bit dubious because "mapping" at the C level means something a little different (it basically just checks for the mp_subscr slot). Not sure if that will cause practical problems but better to be safe.

However, I now realize that this is a major issue since there might be a conflict between the Mapping subtype's own annotations and the annotations it contains as a mapping for another type's namespace.

Since instances no longer have accessible annotations by default, this will only be a problem in exotic cases like a Mapping that is also an instance of type. But again, better to be safe.

[JelleZijlstra]

Fixed by #132490, thanks for the report!