attrs plugin error: "The erased type of self '...*' is not a supertype of its class" on class definition with a generic using a constrained type variable, but then seems to work fine after definition

[glyph]

Consider the following example:

from typing import TypeVar, Generic
import attr

class V:
    pass

class W:
    pass

T = TypeVar("T", V, W)

@attr.s(auto_attribs=True)
class A(Generic[T]):
    v: T

reveal_type(A(W()))

My expectation here would be a revealed type of A[W*], which I do get.

But I also get an error on the @attr.s definition:

The erased type of self 'V*' is not a supertype of its class 'A[T`1]'
The erased type of self 'W*' is not a supertype of its class 'A[T`1]'

if I remove the constraints on T, (i.e. T = TypeVar("T")) then the errors go away.

[gvanrossum]

Maybe the attrs plugin author (@euresti) can shed some insight?

[ilevkivskyi]

This may be related to #5263 (if not a duplicate).

[ilevkivskyi]

Just to clarify, both issues may be caused by bad interaction of patched self/cls with checkmember.bind_self().

[euresti]

Actually. I can repro with dataclasses too.

from typing import TypeVar, Generic
import attr
from dataclasses import dataclass

class V:
    pass

class W:
    pass

T = TypeVar("T", V, W)

@dataclass(order=True)
class A(Generic[T]):
    v: T

reveal_type(A(W()))

So it's related to the added __cmp-ish__ methods

[ilevkivskyi]

OK, the problem is here and here where we unconditionally grab id=1 for the self-type type variable thus causing a type variable clash (the same probably happens for attrs). I think the solution would be to check somehow the next available id instead.

Also btw by agreement ids for generic functions/methods should have negative numbers.

[euresti]

Would the solution just be to make the id -1? It seems to me like -1 should never be used at this point because negative numbers are used for methods/functions and I'm at this point creating it right? Or are the id's global in scope?

[ilevkivskyi]

After some thinking it seems to me just using -1 should be fine since the methods are auto-generated anyway. Btw, I have found that senamal_namedtuple also uses 1 instead of -1. Could you please try replacing it by -1 in both plugins and in named tuple code and play with it trying to break (try some nested classes within generic functions, maybe some scenarios with type inference)?