Catch TypeVisitor subclasses with unimplemented abstract methods

[JamesGuthrie]

Issue #638 was a pain to find because the root cause is an unimplemented abstract method which silently returns None and somewhere down the line causes type checking to fail. It would be better to check this explicitly. The following shows that this is done for one of the abstract methods:
https://github.com/JukkaL/mypy/blob/bbb2dd4f07f995a4b28b8044fa91aa97960fe192/mypy/types.py#L532. There are two potential solutions: 1) use abc's @AbstractMethod decorator, 2) raise an error directly in TypeVisitor.

1. Ensures that future implementations and changes catch subclasses which do not implement the abstract methods, but requires adding implementations for all of the abstract methods to existing subclasses which are missing them. As there are a few, it would probably be simplest to raise an error there instead of providing the complete implementation, which raises the question: why not just use 2) straight off the bat?

@JukkaL Do you have a preferred solution?

[o11c]

I am suspicious of any visitors that can meaningfully work without full implementations.

Perhaps there is a need for more specific visitors?

[JamesGuthrie]

It appears as though there is a common pattern of unimplemented abstract methods (visit_star_type, visit_overloaded, visit_error_type, visit_erased_type), I figured that maybe they had been added after the initial implementation of all visitors, and don't appear in the test cases and as such haven't been caught. In total there are about 25 subclasses which do not implement all abstract methods, mostly it is either visit_star_type or visit_overloaded which are unimplemented.

[JukkaL]

The current implementation is just sloppy, and is inherited from early pre-mypy days when there were no abstract methods.

Using @abstractmethod is probably the best option, as missing implementations can then be caught statically. Trying to catch all of these errors using test cases is an uphill battle.

[JamesGuthrie]

I've opened a PR to get the ball rolling.

As mentioned in the PR, Subclasses with incomplete implementations are:

EraseTypeVisitor (mypy/erasetype.py: 26)
    visit_ellipsis_type()
    visit_star_type()

SameTypeVisitor(mypy/sametype.py: 29)
    visit_overloaded()
    visit_ellipsis_type()
    visit_star_type()

ExpandTypeVisitor(mypy/expandtype.py: 35)
    visit_ellipsis_type()
    visit_star_type()

SubtypeVisitor(mypy/subtypes.py: 44)
    visit_ellipsis_type()
    visit_star_type()

TypeAnalyser(mypy/typeanal.py: 58)
    visit_error_type()
    visit_overloaded()
    visit_ellipsis_type()
    visit_star_type()

TypeJoinVisitor(mypy/join.py: 70)
    visit_overloaded()
    visit_ellipsis_type()
    visit_star_type()

TypeMeetVisitor(mypy/meet.py: 77)
    visit_ellipsis_type()
    visit_star_type()

ConstraintBuilderVisitor(mypy/constraints.py: 125)
    visit_error_type()
    visit_overloaded()
    visit_ellipsis_type()
    visit_star_type()
    visit_type_list()

TypeQuery(mypy/types.py: 760)
    visit_overloaded()
    visit_ellipsis_type()

subclasses of TypeQuery:
    HasAnyQuery(mypy/stats.py: 223)
    HasAnyQuery2(mypy/stats.py: 241)
    ArgInferSecondpassQuery(mypy/checkexpr.py: 1375)
    HasTypeVarQuery(mypy/checkexpr.py: 1389)
    HasErasedComponentsQuery(mypy/checkexpr.py: 1402)

TypeAnalyserPass3(mypy/typeanal.py: 234)
    visit_error_type()
    visit_overloaded()
    visit_ellipsis_type()
    visit_star_type()
    visit_type_list()

[JamesGuthrie]

Actually I rethought this, while it makes sense to move towards using the @abstractmethod annotation on all abstract methods, forcing it now means that it will have to wait until all subclasses have all necessary abstract methods implemented, which may take a while, and doesn't add any benefit until then.

I have restructured my original PR #733 with @abstractmethod annotations only on methods which are implemented in all subtypes. Additionally I added NotImplementedErrors to the remaining abstract methods so that any possible errors caused by missing implementations are made evident.

The NotImplementedErrors caused a number of tests to fail, so I provided implementations as best I could. The tests now all pass, but I'm not sure that the implementations are completely correct, so I would appreciate any feedback you can provide.

It would also be cool to add tests which fail when all currently subtypes implement each of the abstract methods, these would indicate that the @abstractmethod annotation can be added in the abstract base class. Is there an existing appropriate file to put these in, or should I add one?

[jbsiddall]

Is there any update on this? mypy 0.580 is stil showing this error for code: for item in {int}: pass

[JelleZijlstra]

Are you sure you're on the right issue? This one is talking about an internal problem in mypy as far as I can tell.