Strange error for overloaded get() method and Tuple key value

[gvanrossum]

This came up when I merged python/typeshed#223.
I think I have a complete repo that doesn't depend on the stubs. Consider this:

from typing import TypeVar, Generic, Tuple, overload

KT = TypeVar('KT')
VT = TypeVar('VT')

class A(Generic[KT, VT]):
    def get(self, arg: KT) -> VT:
        pass

a = A()  # type: A[Tuple[int, int], int]
a.get((0, 0))  # OK

class B(Generic[KT, VT]):
    @overload
    def get(self, arg: KT) -> VT:
        pass
    @overload
    def get(self, arg: KT, default: VT) -> VT:
        pass

b1 = B()  # type: B[int, int]
b1.get(0)  # OK
b2 = B()  # type: B[Tuple[int, int], int]
b2.get((0, 0))  # error here

I get an error on the very last line (b2.get((0, 0)):

x.py:24: error: No overload variant of "get" of "B" matches argument types [Tuple[builtins.int, builtins.int]]

But there's no error two lines above, on b1.get(0). Also note that the same code with class A which does not overload get()) has no errors. I've got a feeling that there's a subtle bug in the handling of type variables that only shows up when a Tuple is matched against an overloaded method.

@rwbarton Any thoughts? Or @JukkaL ?

[JukkaL]

I looked at this and noticed that the fallback of tuple types are inconsistent. One of the tuple types it is builtins.tuple[builtins.int], and the other it's just builtins.tuple. We should always have the first form. In addition, erasing tuple types was broken, as the fallback didn't get erased.