Allow instantiation of Type[A], if A is abstract

[ilevkivskyi]

Fixes #1843 (It was also necessary to fix few minor things to make this work correctly)

The rules are simple, assuming we have:

class A:
    @abstractmethod
    def m(self) -> None: pass
class C(A):
    def m(self) -> None:
        ...

then

def fun(cls: Type[A]):
    cls() # OK
fun(A) # Error
fun(C) # OK

The same applies to variables:

var: Type[A]
var() # OK
var = A # Error
var = C # OK

Also there is an option for people who want to pass abstract classes around: type aliases, they work as before. For non-abstract A, Type[A] also works as before.

---

@gvanrossum You could be interested in this.
My intuition why you opened #1843 is when someone writes annotation Type[A] with an abstract A, then most probably one wants a class object that implements a certain protocol, not just inherits from A.

EDIT: as discussed in python/peps#224 this behaviour is good for both protocols and usual ABCs.

[elazarg]

#2430 depends on this PR (although the test should be removed there)

[gvanrossum]

Thanks! I finally got to looking at this. I like the general idea.

[gvanrossum]

I take it this is only because of the var: Type[A] below? If this behavior is supposed to work for all Python versions maybe you can rewrite that part of the test using a function? (Also it's a rather long test; maybe it would be better split up.)

[ilevkivskyi]

Since recently --python-version flag is not needed. Also I split the test into three smaller and added a test case for @classmethod.

[gvanrossum]

For what, exactly?

[ilevkivskyi]

I added a detailed comment here.

[gvanrossum]

I'm not keen on this formatting.

[gvanrossum]

"Cannot only" -> "Cannot" (I presume).

[ilevkivskyi]

Oh, sorry, it should be "Can only" (this is probably a worst kind of mistakes).

[gvanrossum]

Maybe also except static methods?

I found a case in our codebase that goes basically like this:

class Logger:
    @staticmethod
    def log(a: Type[C]): ...
class C:
    @classmethod
    def action(cls) -> None:
        Logger.log(cls)  #E: Only non-abstract class can be given where 'Type[C]' is expected
    @abstractmethod
    ...

(I'm far from positive that it's because of this line though.)

[ilevkivskyi]

This code looks reasonable, I added a corresponding exception to check_arg.

[gvanrossum]

could -> can (again :-)

[ilevkivskyi]

@gvanrossum
Thank you very much for review!
I applied your comments in new commits.

[gvanrossum]

LGTM. We're holding off merging until we've sorted out some issues with recent mypy and typeshed commits.

[ilevkivskyi]

@gvanrossum Could this be merged soon?

Sorry for bothering you again, this will be helpful to avoid merge conflicts with my work on protocols.

[gvanrossum]

NP. We're finally at a point where we can merge PRs with confidence again.