Repository navigation
Adding # type: ignore to EnumClass[string_value] doesn't set the type to Any #2416
Description
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 hereWe 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.
Activity
gvanrossum commentedon Nov 8, 2016
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 commentedon Nov 8, 2016
@gvanrossum
The strange error is because message simply takes the name of type. Here actual type of
eisdef (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 becauseTypeAnalyserPass3tries to recover incorrect instances instead of simply returningAny.This is a common behaviour. For example:
My recent PR just brought this behaviour also to aliases (before aliases always returned
Any), ande = 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
Anyin checker. I do this for type aliases in PR #2419The 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: ignoreand the second by default? Although probably it is not a good idea to make such changes few days before release.Fix the problem with ignored potential type alias (#2419)