Cannot use field twice after an is not None check?

[bennn]

On Python 3.10.5+cinder, Cinder commit 1aeff3259dc

Example program:

class Node:
    def __init__(self) -> None:
        self.wins: int = 0
        self.losses: int = 0
        self.parent: None | Node = None

    def score1(self) -> int:
        n = 0
        if self.parent is not None:
            n += self.parent.wins
        if self.parent is not None:
            n += self.parent.losses
        return n

    def score2(self) -> int:
        n = 0
        if self.parent is not None:
            n += self.parent.losses
            n += self.parent.wins # type error !?
        return n

Expected no type error.

But got a type error inside score2 accessing self.parent.wins:

  File "foo.py", line 20
    n += self.parent.wins
        ^
cinderx.compiler.errors.TypedSyntaxError: Optional[__main__.Node]: 'NoneType' object has no attribute 'wins'

If you swap that line and the previous one, the error is on self.parent.losses instead!

[bennn]

Also, it's strange that the test must be self.parent is not None. I'd rather write self.parent by itself.

[carljm]

I think this is Static Python reflecting the idea that losses could be a dynamic property which could execute arbitrary code. So once you've accessed it, any guarantees from the narrowing go out the window. That explains the seemingly odd "second access is an error" behavior.

It might be possible for SP to know that this isn't actually the case for an attribute of a given static type, though, since we can see the definition of the type. The trick here is subclassing, though -- it depends what our rules are on subclasses overriding a normal attribute with a dynamic property.

Respecting if x (and not just if x is not None) as eliminating the possibility of None in an optional type should definitely be possible, I think it just hasn't been implemented.

[bennn]

ah, that makes sense, thanks!