Fix crash in ajava-based WPI related to captures

[kelloggm]

I'm fairly confident that there is a better way to do this than catching the ClassCastException, but I was worried of the performance implications of calling asSuper on every inference of a field's type, since only very rarely will the two AnnotatedTypeMirror not actually have the same actual Java class (causing the exception).

Technically, their kind is the same - they are both typevars - so it cannot be checked that way, even though the underlying Java class of one is AnnotatedDeclaredType. I think it's probably bad that updateAtmWithLub directly casts to AnnotatedTypeVariable, but I'm not sure how else it could work.

[smillst]

Technically, their kind is the same - they are both typevars - so it cannot be checked that way, even though the underlying Java class of one is AnnotatedDeclaredType. I think it's probably bad that updateAtmWithLub directly casts to AnnotatedTypeVariable, but I'm not sure how else it could work.

Are you saying there's an AnnotatedDelcaredType that returns TYPEVAR when getKind is called? If so, that needs to be fix.

[smillst]

A local variable can't have a wildcard type. I think you when an expression has type ? extends T.

[kelloggm]

Are you saying there's an AnnotatedDelcaredType that returns TYPEVAR when getKind is called? If so, that needs to be fix.

I think that is what's happening. I've added debug code to the catch block in this PR so that you can see for yourself (by running ./gradlew ainferIndexAjavaTest -Pemit.test.debug). That prints out something like this every time the exception is caught:

    Catching a ClassCastException: java.lang.ClassCastException: class org.checkerframework.framework.type.AnnotatedTypeMirror$AnnotatedDeclaredType cannot be cast to class org.checkerframework.framework.type.AnnotatedTypeMirror$AnnotatedTypeVariable (org.checkerframework.framework.type.AnnotatedTypeMirror$AnnotatedDeclaredType and org.checkerframework.framework.type.AnnotatedTypeMirror$AnnotatedTypeVariable are in unnamed module of loader 'app')
    The cast must be one of the following two AnnotatedTypeMirrors. The type to cast to is decided by the kind of rhsATM.
    rhsATM: capture#40 extends T extends Object
    rhsATM.getKind(): TYPEVAR
    atmFromStorage: T extends Object
    atmFromStorage.getKind(): TYPEVAR

I have no idea where the bad AnnotatedTypeMirror is coming from - I'm not sure even where to start looking to debug that. My goal with this PR was to treat the symptom of this problem so that I can run WPI without a crash 🙃 However, I agree this seems like it breaks one of our invariants for the CF as a whole.

[smillst]

The problem isn't a AnnotatedDeclaredType with an underlying type of TYPEVAR, it's when updateAtmWithLub recurs on the upper bounds of the type var, they may not be the same kind of type. This happens with capture conversion, but also two type variables. See the test case I added.

Hopefully using that information you can fix the problem, rather than catching the exception.

[kelloggm]

You were exactly right - thanks for figuring this out! I'm glad there wasn't a deeper problem, and I've now written a proper fix.

[kelloggm]

@smillst I'll need you to review this again: the tests for the version you approved earlier today failed for a good reason: the check did the wrong thing if one of the arguments to updateAtmWithLub was a null type. With this improved handling of type variable bounds in updateAtmWithLub, I was also able to remove another workaround for a similar problem that I added a few months ago.