methods not assignable to `Protocol` with `__self__` property

[jorenham]

After reading https://discuss.python.org/t/when-should-we-assume-callable-types-are-method-descriptors/92938/9, I was curious to see if I could write structural method type, dual to the nominal ones in the types stdlib. But I ran into a bit of an issue:

from typing import Protocol

class HasSelf(Protocol):
    @property
    def __self__(self, /) -> object: ...

class A:
    def f(self) -> None: ...

reveal_type(A.f.__self__)   # ✅ type[A]
reveal_type(A().f.__self__) # ✅ A
f1: HasSelf = A.f           # ❌ reportAssignmentType
f2: HasSelf = A().f         # ❌ reportAssignmentType

• Type of "A.f.__self__" is "type[A]"
  

• Type of "A().f.__self__" is "A"
  

• Type "(self: A) -> None" is not assignable to declared type "HasSelf"
    "FunctionType" is incompatible with protocol "HasSelf"
      "__self__" is not present  (reportAssignmentType)
  

• Type "() -> None" is not assignable to declared type "HasSelf"
    "FunctionType" is incompatible with protocol "HasSelf"
      "__self__" is not present  (reportAssignmentType)
  

pyright playground

And I'm guessing that "__self__" is not present is not meant in a spiritual-philosophical sense :)

[erictraut]

Pyright has some hard-coded special casing for __self__ when accessed through a . operator. I added this special casing many years ago to work around the fact that __self__ isn't defined in the types.FunctionType class (at the time, the builtins.function class) in typeshed. I'd be happy to remove this special-casing if you want to to add __self__ to FunctionType in typeshed. That would definitely make this more consistent. If you want to pursue this route, let me know when __self__ is added, and I'll pull the updated typeshed into pyright.

In the meantime, I'm going to close this because I think pyright's doing the best it can do here given the fact that the typeshed definition is incomplete.

[erictraut]

I just noticed that types.MethodType does have __self__ defined. So maybe the right answer here is to use MethodType rather than FunctionType in this case when doing prototype matching. This could also allow me to remove the special casing.

Reopening.

[jorenham]

Haha I was just about to comment exactly that, after I noticed the __self__ in the typeshed code.

[jorenham]

Would it help if I try to make MethodType generic for __self__?

---

edit:

Something like this, for example:
python/typeshed@main...jorenham:typeshed:generic-types.MethodType

[jorenham]

Hmm, it seems that only A().f is a MethodType, but that A.f is a FunctionType.

>>> import types
>>> class A:
...     def f(self): ...
...     
>>> isinstance(A.f, types.FunctionType)
True
>>> isinstance(A.f, types.MethodType)
False
>>> isinstance(A().f, types.FunctionType)
False
>>> isinstance(A().f, types.MethodType)
True

And apparently A.f.__self__ isn't a thing; only A().f.__self__ exists:

>>> A.f.__self__
Traceback (most recent call last):
  File "<python-input-16>", line 1, in <module>
    A.f.__self__
AttributeError: 'function' object has no attribute '__self__'. Did you mean: '__call__'?
>>> A().f.__self__
<__main__.A object at 0x775c242cbb10>

So the A.f should stay a FunctionType, but without the __self__ special-casing.
Turning A().f into a MethodType looks like the right call.

[erictraut]

This is addressed in pyright 1.1.402.