Make Enum Iterable (etc.) only structurally

[elazarg]

Should fix python/mypy#3210 and python/mypy#3349

Remove Iterable, Reversible and Container from being (nominal) baseclasses of Enum.

[elazarg]

How should I add a test for this?

[ilevkivskyi]

There is one small warning, pytype and PyCharm may be not ready for this. (In principle it should be possible to fix this in mypy by modifying nominal constrain inference to make it more like structural, but I am not really proposing this since I don't have time to implement this in nearest future.)

[JelleZijlstra]

@elazarg typeshed doesn't currently have tests for specific stubs. There is an issue for adding it. In the meantime, could you share files you tested with mypy and their output?

@matthiaskramm @vlasovskikh are you OK with this change (which depends on protocols) for mypy and pycharm?

[elazarg]

[updated with bad tests]

from enum import Enum
from typing import Container, Iterable, Reversible, TypeVar
T = TypeVar('T')

class E(Enum):
    one = 1

reveal_type(list(E))  # E: Revealed type is 'builtins.list[E*]'
reveal_type(iter(E))  # E: Revealed type is 'typing.Iterator[E*]'
reveal_type(reversed(E))  # E: Revealed type is 'typing.Iterator[E*]'
reveal_type(x for x in E)  # E: Revealed type is 'typing.Iterator[E*]'
for a in E: reveal_type(a)  # E: Revealed type is 'E*'
x: Container[E] = E
xx: Container[str] = E  # doesn't fail - bad
y: Iterable[E] = E
yy: Iterable[str] = E  # doesn't fail - bad
z: Reversible[E] = E
zz: Reversible[str] = E  # doesn't fail - bad

def c(x: Container[T]) -> T: raise Exception
def it(x: Iterable[T]) -> T: raise Exception
def re(x: Reversible[T]) -> T: raise Exception
reveal_type(c(E))  # E: Revealed type is '<nothing>'  # Bad; see below
reveal_type(it(E))  # E: Revealed type is 'E'
reveal_type(re(E))  # E: Revealed type is 'E'

Container does not seem to be suitable for inference as a protocol, since the signature of __contains__ accepts Any. Maybe I should keep it as an explicit base class.

(And maybe Container should not be generic in the first place?)

[ilevkivskyi]

Situation with Container is a topic of a separate issue: mypy infers UninhabitedType if there are no constraints, see python/mypy#3032

[elazarg]

@JelleZijlstra given the bad tests in the updated tests, this PR should not be merged yet. I'd like to keep this open as a WIP though. I wonder why the assignments does not fail.

[elazarg]

@ilevkivskyi do you have any hint as to why x: Iterable[A] = MyEnum will not fail?

[ilevkivskyi]

This may be a problem in bind_self. I also realized that there is only one small test in mypy for protocols that use self-types of the form Type[T]. It makes sense to open an issue for this on mypy tracker.

[ilevkivskyi]

@elazarg Yes, it looks like the fix is very simple, bind_self assumes that Type[T] can appear only in class methods, replacing is_classmethod with is_classmethod or isinstance(original_type, Instance) and original_type.type.is_metaclass() fixes the problem, I get:

Incompatible types in assignment (expression has type "Type[E]", variable has type "Iterable[str]")

[ilevkivskyi]

Hm, but then I get other spurious errors, so the actual fix will be a bit more complex, but I think this is the right direction.

[elazarg]

Thanks. I think I will open an issue about this.

[vlasovskikh]

@JelleZijlstra PyCharm doesn't support protocols yet, but we will support them eventually, so this change is OK.