Adding # type: ignore to EnumClass[string_value] doesn't set the type to Any

[gvanrossum]

Example:

from typing import Dict
from enum import Enum
class E(Enum):
 V = 'V'
d = {} # type: Dict[E, int]
def f(s: str):
 e = E[s] # type: ignore
 d[e] = 0 # Still error here

We have to use # type: ignore on the line e = E[s] because of #741. But somehow the type of e, when ignoring the error, is not Any but something else. So on the last line we get

__tmp__.py: note: In function "f":
__tmp__.py:8: error: Invalid index type "E" for "dict"

The weird thing is that if e had type E, d[e] would be valid, so the error is also somehow lying...

UPDATE: When trying to repro this, don't try E['V'] -- mypy sees the literal and gives some other error, probably also because of #741.

[gvanrossum]

This was introduced by 57ff82b (PR #2378, Implement generic type aliases). @ilevkivskyi Can you look into this? It would break our internal build (if we were to update mypy).

[ilevkivskyi]

@gvanrossum
The strange error is because message simply takes the name of type. Here actual type of e is def (value: Any) -> __tmp__.E (type object), not __tmp__.E (instance), and the former is indeed an invalid type for index where the latter is expected.

The problem itself appear because ignored line is interpreted by mypy as e = E. This is because TypeAnalyserPass3 tries to recover incorrect instances instead of simply returning Any.

This is a common behaviour. For example:

def g() -> Dict[int]: # type: ignore
 ...
reveal_type(g()) # Revealed type is Dict[Any, Any]

My recent PR just brought this behaviour also to aliases (before aliases always returned Any), and e = E[s] looks like a type alias.

I think the right way to fix this is to mark such "recovered" instances as invalid, so that they could be turned into Any in checker. I do this for type aliases in PR #2419

The mentioned PR fixes the problem, but I am thinking maybe we should reconsider recovery after errors (PEP 484 seems to be silent about this): should errors always lead to Any, or to best possible inferred type? Or first option for # type: ignore and the second by default? Although probably it is not a good idea to make such changes few days before release.