Do not intersect types in isinstance checks if at least one is final.

[tyralla]

Fixes #15148

I think it also fixes the initial bug reported in #12163 (this is why I added a TypeVar test case) but not this bug reported later in the same issue.

[github-actions]

Diff from mypy_primer, showing the effect of this PR on open source code:

CPython (Argument Clinic) (https://github.com/python/cpython)
+ Tools/clinic/clinic.py:5830: error: Subclass of "Sentinels" and "bool" cannot exist: "bool" is final  [unreachable]
+ Tools/clinic/clinic.py:5830: error: Subclass of "Sentinels" and "NoneType" cannot exist: "NoneType" is final  [unreachable]
+ Tools/clinic/clinic.py:5830: error: Subclass of "Null" and "bool" cannot exist: "bool" is final  [unreachable]
+ Tools/clinic/clinic.py:5830: error: Subclass of "Null" and "NoneType" cannot exist: "NoneType" is final  [unreachable]
+ Tools/clinic/clinic.py:5831: error: Statement is unreachable  [unreachable]

sphinx (https://github.com/sphinx-doc/sphinx)
+ sphinx/util/typing.py:205: error: Unused "type: ignore" comment  [unused-ignore]

[tyralla]

Regarding the Mypy primer:

The (correct) new error reports for clinic.py show that this commit somehow partially overlaps with #16315.

For sphinx, I am unsure why the affected branch was not considered unreachable in the first place, like in the following example (where I apply Mypy 1.6):

from typing import Any, Type, _SpecialForm

cls: Type[Any]

if isinstance(cls, _SpecialForm):
    cls.asdf  # error: Statement is unreachable

[tyralla]

The sphinx issue caused me some headaches, but I think I've collected some hints.

The argument cls of function restify is annotated with type | None but should be annotated (at least) with str | type | None. Mypy should complain about this in the implemented if isinstance(cls, str): check but doesn't because it assumes the intersection "sphinx.util.typing.<subclass of "type" and "str">" which in my opinion is impossible because of instance layout conflicts (and I cannot think of a way this makes sense).

Interestingly, Mypy 1.6 agrees with my reasoning when annotating with Type instead of type:

from typing import Type

t1: type
t2: Type

if isinstance(t1, str):
    reveal_type(t1)  # note: Revealed type is "temp.<subclass of "type" and "str">"
if isinstance(t2, str):
    reveal_type(t2)  # error: Statement is unreachable

If I change the annotation of cls from type | None to Type | None, Mypy 1.6 also does not require the type: ignore[attr-defined] comment anymore because it considers the affected line unreachable.

So, everything seems clear. However, when applying Mypy 1.6 on the following example code involving _SpecialForm instead of the complete sphinx code, it finds out about the unreachability no matter if annotating with type or Type:

from typing import Type, _SpecialForm

t1: type
t2: Type

if isinstance(t1, _SpecialForm):  # error: Subclass of "type" and "<typing special form>" cannot exist: would have incompatible method signatures
    reveal_type(t1)  # error: Statement is unreachable
if isinstance(t2, _SpecialForm):
    reveal_type(t2)  # error: Statement is unreachable

To conclude, strange things might be going on, but they are not related to this pull request.

[KotlinIsland]

LGTM!

[tyralla]

@KotlinIsland: Thanks for looking over it!

[KotlinIsland]

@KotlinIsland: Thanks for looking over it!

@tyralla All good, I had started implementing this in basedmypy, but now I don't have to, thanks!

[hauntsaninja]

Nice feature, thank you!