Add containing union and offending type information when reporting on missing attributes for unions

[Nurdok]

This is an attempt to address #2561.

For example, for the following code:

from typing import Union

class A: x = 1
class B: pass

v : Union[A, B] = A()
v.x

The previous error was:

Some element of union has no attribute "x"

With this PR, the error is:

Element "B" of "Union[A, B]" has no attribute "x"

---

Upade: I locally fixed the following None issue in the latest commit.

When dealing with None, mypy reports object. For the example in the above issue:

def foo(s: Optional[str]) -> None:
    print(s.count('a'))

The previous error was:

error: Some element of union has no attribute "count"

With this PR, the error becomes:

Element "object" of "Optional[str]" has no attribute "count"

I guess this happens because mypy searches up the inhertiance tree for NoneType and I'm not sure if there's a way to get the correct type in the code - suggestions welcome.

[ilevkivskyi]

Thanks for the PR! Just a small comment: I think the situation with None is simple: PEP 484 specifies None as a shorthand for NoneType. I also think it is important to special-case the situation when you have a union of exactly two types and one of them is exactly None in messages.py: you could just print such type as Optional[...].

[Nurdok]

I used the format helper to print nicer type representation, which took care of the Optional thing. I still get "object" when I expect "None" .

[JukkaL]

Thanks for the PR! Looks pretty good. A few quick comments:

• I'd prefer Item "foo" of "Union[...]" ... instead of Element "foo" ....
• Can you add a test where 2 out of 3 items are missing an attribute, and another ones where all 2/2 or 3/3 union items have missing attributes?
• It would be good to have more tests for things like __getitem__ or __contains__ with a union type, where some or all items don't have the method.

[Nurdok]

@JukkaL I addressed your first two points. Regarding magic methods, at least the ones you mentioned result in specialized errors.

Doing 1 in u results in

Unsupported right operand type for in ("Union[C, D]")

And doing u[0] results in

 Value of type "Union[C, D]" is not indexable

It would be desirable for these cases to also result in more specific errors (e.g., Item "C" in "Union[C, D]" is not indexable), but it probably belongs in a separate PR.

[JukkaL]

This PR is almost there! I'd like to just see one docstring tweak.

[JukkaL]

Document original_type in the docstring.

[ilevkivskyi]

Looks good to me!