[ty] Ignore descriptor class-level declarations for purposes of finding instance attributes, variant 3

This commit is contained in:
David Peter 2025-05-26 15:12:29 +02:00
parent 2c717c9f5e
commit cdafb3d81c
3 changed files with 17 additions and 23 deletions

View file

@ -193,7 +193,8 @@ reveal_type(C2().attr) # revealed: Unknown | Literal["non-data", "normal"]
C2().attr = 1
```
This situation does not change if the attribute is declared on the class body:
This situation changes if the class attribute is declared. Here, we should error if we see an
incompatible attribute assignment, but otherwise just follow the declared type:
```py
class C3:
@ -201,11 +202,12 @@ class C3:
def f(self):
# TODO: we should ideally emit an error here. We are overwriting the
# non-data descriptor with a string, which is not compatible with the
# declared type.
self.attr = "normal"
# non-data descriptor with an integer, which is not compatible with
# the `__get__` return type of `NonDataDescriptor` when called on an
# instance.
self.attr = 1
reveal_type(C3().attr) # revealed: Literal["non-data", "normal"] | Unknown
reveal_type(C3().attr) # revealed: Literal["non-data"]
```
The scenario above is similar to a use case where a method on a class is dynamically replaced.
@ -221,20 +223,15 @@ class C4:
def switch(self):
# Similar to the `C3` example, we are overwriting a non-data descriptor (the
# function `C4.f`) with something (a bound method) that is not compatible with
# the (implicitly) declared type of `C4.f`, which is a function literal type:
# `def f(self) -> None`. Strictly speaking, this we should also emit an error
# here.. or we should not consider the function definition to be a declaration.
# the return type of `__get__` of `C4.f` (a different bound method). Strictly
# speaking, we should also emit an error in this case.
self.f = self.replacement
reveal_type(C4.f) # revealed: def f(self) -> None
c4 = C4()
# call c4.switch() or not
# TODO: This should reveal the following type, as soon as we understand the type of self:
# `(bound method C4.f() -> None) | (bound method C4.replacement() -> None) | Unknown`
reveal_type(c4.f) # revealed: (bound method C4.f() -> None) | Unknown
reveal_type(c4.f) # revealed: bound method C4.f() -> None
# As a regression test for https://github.com/astral-sh/ty/issues/350, make sure that no
# error is emitted when calling `c4.f()`:

View file

@ -1676,7 +1676,9 @@ def _(r: Recursive):
reveal_type(r.direct) # revealed: Recursive
reveal_type(r.union) # revealed: None | Recursive
reveal_type(r.intersection1) # revealed: C & Recursive
reveal_type(r.intersection2) # revealed: C & ~Recursive
# TODO: no error, and a revealed type of `C & ~Recursive`
# error: [unresolved-attribute] "Type `Recursive` has no attribute `intersection2`"
reveal_type(r.intersection2) # revealed: Unknown
reveal_type(r.t) # revealed: tuple[int, tuple[str, Recursive]]
reveal_type(r.callable1) # revealed: (int, /) -> Recursive
reveal_type(r.callable2) # revealed: (Recursive, /) -> int

View file

@ -1771,15 +1771,10 @@ impl<'db> ClassLiteral<'db> {
// attribute is a non-data descriptor, it can not possibly be the
// correct type of the implicit instance attribute. If there are any
// attribute assignments in methods of this class, they would overwrite
// the non-data descriptor. In this case, we just return the type
// inferred from attribute assignments in methods. The descriptor
// protocol implementation in `Type::invoke_descriptor_protocol` will
// take care of unioning with the non-data descriptor type (because we
// account for the fact that the methods containing these assignments
// might never be called).
if !implicit.is_unbound() {
return implicit.into();
}
// the non-data descriptor. If they do so in a non-compatible way, we
// should emit an error elsewhere. Here, we simply return `Unbound`,
// to signal that there is no instance attribute of this name.
return Symbol::Unbound.into();
}
let bindings = use_def.public_bindings(symbol_id);