False positive of incompatible types (about generic protocol) since 1.1.337

[Azureblade3808]

Describe the bug

See sample code below.

In function f, v1 can be assigned to v0 and v2 can be assigned to v1, but v2 cannot be directly assigned to v0.

Code or Screenshots

from typing import Generic, Protocol, TypeVar

V_co = TypeVar("V_co", covariant=True)

class C(Generic[V_co]):
    def f(self, /) -> V_co: ...

class P(Protocol[V_co]):
    def f(self, /) -> V_co: ...

def f(
    v0: C[C[object]],
    v1: C[P[object]],
    v2: P[P[object]],
):
    _ = v0, v1, v2

    if 0:
        v1 = v0
    elif 0:
        v2 = v1
    elif 0:
        v2 = v0  # False report.

Pyright Playground

VS Code extension or command-line

This issue appears since version 1.1.337. It doesn't appear with version from 1.1.292 through 1.1.336.

[Azureblade3808]

If the sample code is modified as below --

from typing import Generic, Protocol, TypeVar

V_co = TypeVar("V_co", covariant=True)

class C(Generic[V_co]):
    def f(self, /) -> V_co: ...

class P0(Protocol[V_co]):
    def f(self, /) -> V_co: ...

class P1(Protocol[V_co]):
    def f(self, /) -> V_co: ...

def f(
    v0: C[C[object]],
    v1: C[P0[object]],
    v2: P1[P0[object]],
):
    _ = v0, v1, v2

    if 0:
        v1 = v0
    elif 0:
        v2 = v1
    elif 0:
        v2 = v0

Pyright Playground

No error is reported.

So I suspict there is some data pollution among generic types of a same protocol.

[erictraut]

Thanks, I really appreciate the work you put into writing great bug reports.

I've narrowed the regression to this commit. The issue is related to the constraint solver, which is conflating the type variable of the "inner" P with that of the "outer" P in the annotation P[P[object]].

[Azureblade3808]

I'm glad to help.

[erictraut]

This is addressed in pyright 1.1.364.