Unexpected name mangling behavior with generics

[gaesa]

Bug report

Bug description:

Name mangling behaves inconsistently when used with generics. Here’s the code to reproduce the issue:

class __Foo(type):
    pass


class Bar[T](metaclass=__Foo):
    pass

This raises NameError: name '_Bar__Foo' is not defined. However, removing the [T] results in the error disappearing. This suggests that the name mangling behavior is inconsistent when generics are involved. This behavior seems to be a bug, as it would be expected for the name mangling to either always occur or never occur, regardless of whether generics are used.

CPython versions tested on:

3.12

Operating systems tested on:

Linux

Linked PRs

• gh-119311: Fix name mangling with PEP 695 generic classes #119464
• [3.13] gh-119311: Fix name mangling with PEP 695 generic classes (#119464) #119643
• [3.12] gh-119311: Fix name mangling with PEP 695 generic classes (#119464) #119644
• gh-119311: Add missing magic number (3571) for 3.13.0b1 #125771

[JelleZijlstra]

I agree this is a bug. We apply name mangling to all names that appear in the type parameter scope for the generic class (and also to all code after it, which is a separate bug: #119395). In the example, this means __Foo in the bases gets mangled to _Bar__Foo.

We could instead not do any mangling in the type parameter scope. That would fix this bug and simplify the code. However, it would break cases like this:

class X[__T]:
    def foo(self, x: __T): ...

If we didn't do mangling in the type parameter scope, the second use of __T would get mangled, but the first one wouldn't, so the second use of __T would fail to resolve.

Fortunately, the set of names that can appear in a type parameter scope is very limited. I think it should be possible to mangle only the type parameter names defined in the scope but nothing else. This would make both my example and the OP's example work, but it adds some complexity to the symtable.