isinstance inference not working in closures

[ThomasHickman]

In the code:

class SuperCls:
    pass

class Cls(SuperCls):
    prop = 9

def do_something(inp):
    # type: (SuperCls) -> None
    if isinstance(inp, Cls):
        def x():
            # type: () -> None
            print(inp.prop)
        return x()

inst = Cls()
do_something(inst)

this warning is wrongly emitted:

test.py:12: error: "SuperCls" has no attribute "prop"

[gvanrossum]

This is a known issue; it's somewhat of a feature because sometimes the closure variable can change its value (and hence type) again before the function is called (e.g. consider if your code has inp = SuperCls() just before the return x()). Detecting the cases where this isn't a problem is somewhat complicated. You can work around it by passing inp as an argument to x().

[nilehmann]

Related to this. I would like to write code similar to this one

def add(m: int, n: int) -> int:
    return m + n

one: Optional[int] = 1

if one:
    map(lambda n: add(one, n), [1, 2, 3])

Maybe the case for anonymous function is easier?

[gvanrossum]

OTOH the way mypy treats lambda is different from regular functions. Until someone (not me) tries to implement this I have no idea what's easy or not in this case.

[nilehmann]

Giving it a second thought, I think is it equally easy to invalidates things for lambdas than for regular functions.

[jrideout]

I ran into this as well:

def test(a: Optional[str] = None) -> str:
    if a is None:
        a = 'abc'
    def inner() -> str:
        return a  #  <---  error: Incompatible return value type (got "Optional[str]", expected "str")
    return inner()

In this case the closure isn't going to change. I wonder if we can infer that.

[ilevkivskyi]

This is actually a duplicate of #2608 (already high priority). Yes, one of proposed solutions was to use the binder, if e.g. the inner function appears at the end of the outer function, so the closure can't change.