Generic type inference incorrectly handled in edge case

[zevbo]

Bug Report

If you try to widdle down the type of a generic class in a certain manner, it only works if you use an intermediate variable. A little tough to explain without looking at the example

To Reproduce

https://mypy-play.net/?mypy=latest&python=3.10&gist=d3ebef88a57147cef8a8d311e1b33fb7

from abc import ABC, abstractmethod
from typing import Generic, TypeVar

T = TypeVar("T")

class A(ABC, Generic[T]):

    @abstractmethod
    def t(self) -> T:
        pass

def get_a_does_work(a: A[T]) -> T:
    t_value = a.t()
    if t_value is None:
        return None 
    return a.t()

def get_a_should_work(a: A[T]) -> T:
    if a.t() is None:
        return None 
    return a.t()

Expected Behavior

Both get_a_does_work and get_a_should_work should pass type checking.

Actual Behavior

get_a_should_work outouts:
main.py:20: error: Incompatible return value type (got "None", expected "T")

Your Environment

You can see on the playground

[hauntsaninja]

There's no reason for mypy to believe that a.t() is deterministic

[zevbo]

Shouldn't matter. If a.t() can return None, then None must be a subtype of T. Hence, it is valid to return None.

And mypy does know this, given that the first function does works fine.

[headtr1ck]

Might be related to #13956

[ilevkivskyi]

Now mypy shows errors in both cases which is at least more consistent. First example used to work by accident, mypy doesn't support dependencies between expressions in general (except few common cases).