Missing conformance tests around nonlocal rebinding of type params from annotation scopes

[samwgoldman]

The language reference mentions this restriction:

Names defined in annotation scopes cannot be rebound with nonlocal statements in inner scopes. This includes only type parameters, as no other syntactic elements that can appear within annotation scopes can introduce new names.

I noticed that Pyright doesn't implement this restriction, while Mypy does.

def f[Ty](x: Ty) -> Ty:
    nonlocal Ty
    y: Ty = x
    Ty: int = 0
    return y

[Pyright Playground] [Mypy Playground]

[samwgoldman]

Another example where checkers differ is the following, where the type param introduced in the annotation scope is shadowed by the function body.

def f[T](x: T) -> T:
    y: T = x   # should be an error, value used as a type
    T: int = 0
    return y

[Pyright Playground] [Mypy Playground]

Mypy accepts this code and Pyright rejects it. I think that Pyright's behavior is correct here, since any reference to a variable bound in a scope refers to that binding -- even if the binding is not yet initialized.

[erictraut]

To be fair, this restriction isn't mentioned in the typing spec. That explains why it's not included in the conformance tests currently.

If we add it to the conformance tests (which I think is reasonable), we should probably also add it to the typing spec as well.

[samwgoldman]

If we add it to the conformance tests (which I think is reasonable), we should probably also add it to the typing spec as well.

Thanks for this guidance. I was not sure whether it was appropriate to include in the spec, but I agree that it's reasonable to add it.

[JelleZijlstra]

I feel this is a runtime behavior and doesn't need to be mentioned in the typing spec. It's a general rule that type checkers should implement the semantics of the language.

[erictraut]

That's reasonable, but I don't know if such a test belongs in the conformance test then. We haven't included tests for other runtime behaviors that are not discussed in the typing spec. We could do so, but that would be a large undertaking.

[samwgoldman]

Okay, let's leave it alone then. For runtime behaviors, the semantics should be described in the language spec and conformance is up to type checkers individually to test. Issues should be reported to type checkers.

I think it would be valuable to have a suite of tests that type checkers can share, but I understand that conformance tests are more narrowly scoped, and increasing that scope would increase the maintenance burden on the maintainers.

Thanks for the discussion! Closing

[JelleZijlstra]

I do think it could be valuable to have shared tests of runtime behavior (e.g., scoping) too, and the conformance tests in this repo would be a good place for such tests.