`==`-based narrowing of `Optional`

[tyralla]

1. Mypy's current behaviour:

x: int | None
if x == None:
 x # "Union[builtins.int, None]"
else:
 x # "Union[builtins.int, None]"

2. Pyright's current behaviour:

x: int | None
if x == None:
 x # x is None
else:
 x # x is int

3. I think the ideal solution would be:

x: int | None
if x == None:
 x # "Union[builtins.int, None]"
else:
 x # "builtins.int"

This Mypy code comment seems to favour solution 2.

I am asking because if Mypy would follow solution 2 or 3, in-based narrowing of optional types could be implemented more consistently.

Related pull requests: #15760 #17154 #17044.

[hauntsaninja]

I think both 2 and 3 are fine. If we go with 3, we should look at presence of __eq__ (int has __eq__ defined, not sure why, dates back to 2015)

[tyralla]

Did you mean we should check for __eq__ when going for solution 2?

If so, this could be an additional safety net for most cases. Then, the modified solution 2 would only result in wrong narrowing for subtypes that override __eq__ strangely:

class A: ...

class B(A):
 def __eq__(self, o: object) -> bool:
 return True

def f(x: A | None) -> None:
 if x == None:
 assert x is None

f(B()) # AssertionError

I think, for complete safety, we would have to combine looking for __eq__ with checking for final:

from typing import final

@final
class A: ...

def f(x: A | None) -> None:
 if x == None:
 reveal_type(x) # must be None