Update the behaviour of the type function for metaclasses

[TRManderson]

Calling type on a class A yields Type[A], even if A has a metaclass other than type. This leads to incorrectly typechecking interactions with type(A) when when attempting to use metaclass attributes.

[TRManderson]

Somewhat related to #2655

[elazarg]

Can you give a concrete example?
The interaction with other features might be interesting. Maybe it should work something like this

class M(Generic(T), Type[T]): pass  # maybe M(type) is a shorthand for this

T1 = TypeVar('T1', bound='A')

class A(metaclass=M):
    @classmethod
    def foo(cls: M[T1]) -> T1:
        return cls()

reveal_type(A)  # E: Revealed type is M[A]

class B(A): pass

reveal_type(B)  # E: Revealed type is M[B]
reveal_type(B.foo()) # E: Revealed type is B

All of this should be how things work right now, except that M[B] is written Type[B]. Where does this fail?

[elazarg]

I agree it should be possible to write M[T] for "a type whose metaclass is M".

[TRManderson]

class Meta(type):
    x = 1

class A(metaclass=Meta): pass

print(type(A).x) # test2826.py:6: error: Type[] has no attribute "x"

type(A) is Meta which definitely has an attribute x. Do I need to make Meta subclass Type[T] for it to work?

[elazarg]

Thanks.

The culprit in my example seems to be that the type variable T is quantified over type constructors (TypeInfo) and not over types. Perhaps there is a workaround.
(Just thinking out loud).

[elazarg]

I have a fix for this particular issue, but the WiFi here is not working.

[gvanrossum]

..., but the WiFi here is not working.

Sounds like an updated version of Fermat's claim about his last theorem. :-)

[elazarg]

:D