Fix crash from NamedTuple placeholder

[hauntsaninja]

Fixes #17396

I'm having trouble writing a regression test, but the following reproduces the issue nicely:

rm -rf repro
mkdir repro
echo 'from mod2 import UniqueAllResult' > repro/mod1.pyi
echo '
from typing import Generic, NamedTuple, TypeVar
from mod1 import does_not_exist

_SCT = TypeVar("_SCT", bound=does_not_exist)

class UniqueAllResult(NamedTuple, Generic[_SCT]):
    values: int
' > repro/mod2.pyi

MYPYPATH=repro mypy -c 'import mod1'

[ilevkivskyi]

This looks a bit hacky, but I guess it may be a necessary evil. Two things:

• Is there the same issue with generic TypedDicts? If yes, please fix that as well (and maybe try adding a test)
• Can you verify that a pattern like this still works? (And maybe add a corresponding test if we don't have one):

T = TypeVar("T", bound="NT")
class NT(NamedTuple, Generic[T]):
    parent: T
    item: int

[hauntsaninja]

Thanks for the review!

• Yeah, I thought about this also but wasn't able to reproduce the issue with TypedDict
• Good point, it no longer works. I opened Add regression test for NamedTuple with recursive bound #18399 for regression test, will figure out a way forward
• I also realised why I can't repro in unit tests. There's some interaction with install location (so e.g. my original repro in PR comment may not even repro for you depending on how you have mypy installed

[ilevkivskyi]

Yeah, I thought about this also but wasn't able to reproduce the issue with TypedDict

Hm, I think this requires further investigation. NamedTuples and TypedDicts are almost identical w.r.t. deferrals etc. This may indicate that there is a simpler/better fix, or there is something deeper going on. I would recommend adding some debugging prints in a repro that mirrors the NmaedTuple one, and see what types appear there (it may be that there are still some unresolved placeholders, but they don't cause a crash in case of TypedDicts for some reason).

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[sterliakov]

@hauntsaninja I think I see the non-hackish solution here!

I don't know, however, whether this should apply to HasPlaceholders specifically or to TypeQuery and BoolTypeQuery in general. This placeholder is only retained because HasPlaceholders query visitor never reaches partial_fallback of a tuple's self type (and so it works as intended for TypedDict's that have no self type at all). Any thoughts?

diff --git a/mypy/semanal_shared.py b/mypy/semanal_shared.py
index bdd01ef6a..9d3d4b2eb 100644
--- a/mypy/semanal_shared.py
+++ b/mypy/semanal_shared.py
@@ -364,6 +364,9 @@ class HasPlaceholders(BoolTypeQuery):
     def __init__(self) -> None:
         super().__init__(ANY_STRATEGY)
 
+    def visit_tuple_type(self, t: TupleType, /) -> bool:
+        return super().visit_tuple_type(t) or t.partial_fallback.accept(self)
+
     def visit_placeholder_type(self, t: PlaceholderType) -> bool:
         return True

Your repro in this PR doesn't work for me, but the repro you posted on original issue still does (and passes with the patch above applied). It also heals #18582 crash and does not fail testGenericNamedTupleRecursiveBound!