Metaclass is not properly inherited when using multiple inheritance with int

[gvanrossum]

Consider:

class M(type):
    def foo(self): pass
class C(metaclass=M):
    pass
print(C.__class__)
C.foo()
class D(int, C):
    pass
print(D.__class__)
D.foo()

Running this with Python 3 shows that D's metaclass is M, but when checking this with mypy, we get the error

u.py:10: error: "D" has no attribute "foo"

on the last line, showing that mypy doesn't believe so.

This seems unique to int (or perhaps builtin types?) -- when create a separate class A (without a metaclass) and use that instead of int, there is no error.

FWIW this prevents me from fixing #2305 (iterating on enums) -- I can get it to work for Enum but not for IntEnum, because it multiply inherits from int and from Enum.

[elazarg]

It does not reproduce in unit tests.

[elazarg]

In general, the class/metaclass hierarchy in typeshed does not mirror the hierarchy in Python. I'm not sure, but it could be the reason. Perhaps the special casing of ABCMeta in mypy is also a problem.

[elazarg]

Removing SupportsInt and SupportsFloat from baseclasses makes the error go away. SupportsAbs[int] has nothing to do with the error. Since the formers have ABCMeta as a metaclass, and the latter does not, It points to ABCMeta as the problem.

[elazarg]

Well I guess I'm completely wrong. Perhaps it's a simple case of choosing the most derived common metaclass. In typeshed type(int) == ABCMeta. In Python type(int) == type.

[elazarg]

A possible solution: make EnumMeta inherit ABCMeta, so it will be the most derived common metaclass.

[gvanrossum]

That mostly works, but I needed to add a # type: ignore. Thoughts? https://github.com/python/typeshed/tree/enum-iter

[elazarg]

Is there some problem with selftype? I have encountered something, but I wasn't sure.

[gvanrossum]

Without that I get

typeshed/stdlib/3.4/enum.pyi:11: error: The erased type of self 'Type[enum.Enum]' is not a supertype of its class 'enum.EnumMeta'

FWIW the test program is

import enum
class E(enum.Enum):
    a = 1
    b = 2
for i in E:
    reveal_type(i)
class I(enum.IntEnum):
    a = 1
    b = 2
for k in I:
    reveal_type(k)

and its output (less the error) is

t.py:6: error: Revealed type is 't.E*'
t.py:11: error: Revealed type is 't.I*'

which is exactly what it should be.

[elazarg]

There are additional checks to add to the subtype visitor that makes this problem go away, but something bothers me. The variable seems to be overly constrained by a class not defined yet. We should add another check for the metaclass declaration, to verify that we are constrained correctly by the type of cls.

[elazarg]

As I have mentioned in a different discussion here, it might make sense to use the form cls: Meta[T] which means precisely the type whose metaclass is a subclass of Meta (directly inheriting the generic [] operator from Type). Then we don't need to constrain further T in any way.

[elazarg]

Root cause: python/typeshed#1595

[JelleZijlstra]

Recap:

• There was no mypy bug here; the problem is because of incorrect inheritance in typeshed (fix being tracked in Deriving builtins from ABCs implies incorrect and problematic metaclass structure typeshed#1595)
• This problem usually appears in practice with enum classes, and we merged a workaround for those in Derive EnumMeta from ABCMeta typeshed#1598.