[red-knot] Add __init__ arguments check when doing try_call on a class literal (#16512)

## Summary

* Addresses #16511 for simple cases where only `__init__` method is
bound on class or doesn't exist at all.
* fixes a bug with argument counting in bound method diagnostics

Caveats:
* No handling of `__new__` or modified `__call__` on metaclass.
* This leads to a couple of false positive errors in tests

## Test Plan

- A couple new cases in mdtests
- cargo nextest run -p red_knot_python_semantic --no-fail-fast

---------

Co-authored-by: Carl Meyer <carl@astral.sh>
Co-authored-by: David Peter <sharkdp@users.noreply.github.com>
This commit is contained in:
Mike Perlov
2025-04-08 23:26:20 +02:00
committed by GitHub
parent ed14dbb1a2
commit fab7d820bd
11 changed files with 853 additions and 121 deletions

View File

@@ -8,7 +8,11 @@ Currently, red-knot doesn't support `typing.NewType` in type annotations.
from typing_extensions import NewType
from types import GenericAlias
X = GenericAlias(type, ())
A = NewType("A", int)
# TODO: typeshed for `typing.GenericAlias` uses `type` for the first argument. `NewType` should be special-cased
# to be compatible with `type`
# error: [invalid-argument-type] "Object of type `NewType` cannot be assigned to parameter 2 (`origin`) of function `__new__`; expected type `type`"
B = GenericAlias(A, ())
def _(

View File

@@ -1,7 +1,325 @@
# Constructor
When classes are instantiated, Python calls the meta-class `__call__` method, which can either be
customized by the user or `type.__call__` is used.
The latter calls the `__new__` method of the class, which is responsible for creating the instance
and then calls the `__init__` method on the resulting instance to initialize it with the same
arguments.
Both `__new__` and `__init__` are looked up using full descriptor protocol, but `__new__` is then
called as an implicit static, rather than bound method with `cls` passed as the first argument.
`__init__` has no special handling, it is fetched as bound method and is called just like any other
dunder method.
`type.__call__` does other things too, but this is not yet handled by us.
Since every class has `object` in it's MRO, the default implementations are `object.__new__` and
`object.__init__`. They have some special behavior, namely:
- If neither `__new__` nor `__init__` are defined anywhere in the MRO of class (except for `object`)
\- no arguments are accepted and `TypeError` is raised if any are passed.
- If `__new__` is defined, but `__init__` is not - `object.__init__` will allow arbitrary arguments!
As of today there are a number of behaviors that we do not support:
- `__new__` is assumed to return an instance of the class on which it is called
- User defined `__call__` on metaclass is ignored
## Creating an instance of the `object` class itself
Test the behavior of the `object` class itself. As implementation has to ignore `object` own methods
as defined in typeshed due to behavior not expressible in typeshed (see above how `__init__` behaves
differently depending on whether `__new__` is defined or not), we have to test the behavior of
`object` itself.
```py
reveal_type(object()) # revealed: object
# error: [too-many-positional-arguments] "Too many positional arguments to class `object`: expected 0, got 1"
reveal_type(object(1)) # revealed: object
```
## No init or new
```py
class Foo: ...
reveal_type(Foo()) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to bound method `__init__`: expected 0, got 1"
reveal_type(Foo(1)) # revealed: Foo
```
## `__new__` present on the class itself
```py
class Foo:
def __new__(cls, x: int) -> "Foo":
return object.__new__(cls)
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of function `__new__`"
reveal_type(Foo()) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to function `__new__`: expected 1, got 2"
reveal_type(Foo(1, 2)) # revealed: Foo
```
## `__new__` present on a superclass
If the `__new__` method is defined on a superclass, we can still infer the signature of the
constructor from it.
```py
from typing_extensions import Self
class Base:
def __new__(cls, x: int) -> Self: ...
class Foo(Base): ...
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of function `__new__`"
reveal_type(Foo()) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to function `__new__`: expected 1, got 2"
reveal_type(Foo(1, 2)) # revealed: Foo
```
## Conditional `__new__`
```py
def _(flag: bool) -> None:
class Foo:
if flag:
def __new__(cls, x: int): ...
else:
def __new__(cls, x: int, y: int = 1): ...
reveal_type(Foo(1)) # revealed: Foo
# error: [invalid-argument-type] "Object of type `Literal["1"]` cannot be assigned to parameter 2 (`x`) of function `__new__`; expected type `int`"
reveal_type(Foo("1")) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of function `__new__`"
reveal_type(Foo()) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to function `__new__`: expected 1, got 2"
reveal_type(Foo(1, 2)) # revealed: Foo
```
## A descriptor in place of `__new__`
```py
class SomeCallable:
def __call__(self, cls, x: int) -> "Foo":
obj = object.__new__(cls)
obj.x = x
return obj
class Descriptor:
def __get__(self, instance, owner) -> SomeCallable:
return SomeCallable()
class Foo:
__new__: Descriptor = Descriptor()
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__call__`"
reveal_type(Foo()) # revealed: Foo
```
## A callable instance in place of `__new__`
### Bound
```py
class Callable:
def __call__(self, cls, x: int) -> "Foo":
return object.__new__(cls)
class Foo:
__new__ = Callable()
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__call__`"
reveal_type(Foo()) # revealed: Foo
```
### Possibly Unbound
```py
def _(flag: bool) -> None:
class Callable:
if flag:
def __call__(self, cls, x: int) -> "Foo":
return object.__new__(cls)
class Foo:
__new__ = Callable()
# error: [call-non-callable] "Object of type `Callable` is not callable (possibly unbound `__call__` method)"
reveal_type(Foo(1)) # revealed: Foo
# TODO should be - error: [missing-argument] "No argument provided for required parameter `x` of bound method `__call__`"
# but we currently infer the signature of `__call__` as unknown, so it accepts any arguments
# error: [call-non-callable] "Object of type `Callable` is not callable (possibly unbound `__call__` method)"
reveal_type(Foo()) # revealed: Foo
```
## `__init__` present on the class itself
If the class has an `__init__` method, we can infer the signature of the constructor from it.
```py
class Foo:
def __init__(self, x: int): ...
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__init__`"
reveal_type(Foo()) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to bound method `__init__`: expected 1, got 2"
reveal_type(Foo(1, 2)) # revealed: Foo
```
## `__init__` present on a superclass
If the `__init__` method is defined on a superclass, we can still infer the signature of the
constructor from it.
```py
class Base:
def __init__(self, x: int): ...
class Foo(Base): ...
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__init__`"
reveal_type(Foo()) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to bound method `__init__`: expected 1, got 2"
reveal_type(Foo(1, 2)) # revealed: Foo
```
## Conditional `__init__`
```py
def _(flag: bool) -> None:
class Foo:
if flag:
def __init__(self, x: int): ...
else:
def __init__(self, x: int, y: int = 1): ...
reveal_type(Foo(1)) # revealed: Foo
# error: [invalid-argument-type] "Object of type `Literal["1"]` cannot be assigned to parameter 2 (`x`) of bound method `__init__`; expected type `int`"
reveal_type(Foo("1")) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__init__`"
reveal_type(Foo()) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to bound method `__init__`: expected 1, got 2"
reveal_type(Foo(1, 2)) # revealed: Foo
```
## A descriptor in place of `__init__`
```py
class SomeCallable:
# TODO: at runtime `__init__` is checked to return `None` and
# a `TypeError` is raised if it doesn't. However, apparently
# this is not true when the descriptor is used as `__init__`.
# However, we may still want to check this.
def __call__(self, x: int) -> str:
return "a"
class Descriptor:
def __get__(self, instance, owner) -> SomeCallable:
return SomeCallable()
class Foo:
__init__: Descriptor = Descriptor()
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__call__`"
reveal_type(Foo()) # revealed: Foo
```
## A callable instance in place of `__init__`
### Bound
```py
class Callable:
def __call__(self, x: int) -> None:
pass
class Foo:
__init__ = Callable()
reveal_type(Foo(1)) # revealed: Foo
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__call__`"
reveal_type(Foo()) # revealed: Foo
```
### Possibly Unbound
```py
def _(flag: bool) -> None:
class Callable:
if flag:
def __call__(self, x: int) -> None:
pass
class Foo:
__init__ = Callable()
# error: [call-non-callable] "Object of type `Callable` is not callable (possibly unbound `__call__` method)"
reveal_type(Foo(1)) # revealed: Foo
# TODO should be - error: [missing-argument] "No argument provided for required parameter `x` of bound method `__call__`"
# but we currently infer the signature of `__call__` as unknown, so it accepts any arguments
# error: [call-non-callable] "Object of type `Callable` is not callable (possibly unbound `__call__` method)"
reveal_type(Foo()) # revealed: Foo
```
## `__new__` and `__init__` both present
### Identical signatures
A common case is to have `__new__` and `__init__` with identical signatures (except for the first
argument). We report errors for both `__new__` and `__init__` if the arguments are incorrect.
At runtime `__new__` is called first and will fail without executing `__init__` if the arguments are
incorrect. However, we decided that it is better to report errors for both methods, since after
fixing the `__new__` method, the user may forget to fix the `__init__` method.
```py
class Foo:
def __new__(cls, x: int) -> "Foo":
return object.__new__(cls)
def __init__(self, x: int): ...
# error: [missing-argument] "No argument provided for required parameter `x` of function `__new__`"
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__init__`"
reveal_type(Foo()) # revealed: Foo
reveal_type(Foo(1)) # revealed: Foo
```
### Compatible signatures
But they can also be compatible, but not identical. We should correctly report errors only for the
mthod that would fail.
```py
class Foo:
def __new__(cls, *args, **kwargs):
return object.__new__(cls)
def __init__(self, x: int) -> None:
self.x = x
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__init__`"
reveal_type(Foo()) # revealed: Foo
reveal_type(Foo(1)) # revealed: Foo
# error: [too-many-positional-arguments] "Too many positional arguments to bound method `__init__`: expected 1, got 2"
reveal_type(Foo(1, 2)) # revealed: Foo
```

View File

@@ -20,9 +20,11 @@ class C:
def _(subclass_of_c: type[C]):
reveal_type(subclass_of_c(1)) # revealed: C
# TODO: Those should all be errors
# error: [invalid-argument-type] "Object of type `Literal["a"]` cannot be assigned to parameter 2 (`x`) of bound method `__init__`; expected type `int`"
reveal_type(subclass_of_c("a")) # revealed: C
# error: [missing-argument] "No argument provided for required parameter `x` of bound method `__init__`"
reveal_type(subclass_of_c()) # revealed: C
# error: [too-many-positional-arguments] "Too many positional arguments to bound method `__init__`: expected 1, got 2"
reveal_type(subclass_of_c(1, 2)) # revealed: C
```

View File

@@ -111,6 +111,8 @@ class E[T]:
def __init__(self, x: T) -> None: ...
# TODO: revealed: E[int] or E[Literal[1]]
# TODO should not emit an error
# error: [invalid-argument-type] "Object of type `Literal[1]` cannot be assigned to parameter 2 (`x`) of bound method `__init__`; expected type `T`"
reveal_type(E(1)) # revealed: E
```
@@ -118,7 +120,8 @@ The types inferred from a type context and from a constructor parameter must be
other:
```py
# TODO: error
# TODO: the error should not leak the `T` typevar and should mention `E[int]`
# error: [invalid-argument-type] "Object of type `Literal["five"]` cannot be assigned to parameter 2 (`x`) of bound method `__init__`; expected type `T`"
wrong_innards: E[int] = E("five")
```

View File

@@ -18,7 +18,7 @@ class Number:
def __invert__(self) -> Literal[True]:
return True
a = Number()
a = Number(0)
reveal_type(+a) # revealed: int
reveal_type(-a) # revealed: int