Repository navigation
Wrong inference of the target of __new__ #9405
Description
🐛 Bug Report
The target of an invocation of __new__ is an explicit class, but mypy erroneously believes it is another class.
To Reproduce
Create a file test.py with this content:
from typing import Any
class C(str):
def __new__(cls, value: str) -> Any:
return str.__new__(cls, value)
Then run mypy test.py
Side note # 1: this is a stripped down version to reproduce the issue, the original case has more real interest.
Side note # 2: __new__ should return C but the name C is not defined during class creation, so __new__ is declared to return Any, instead.
Expected Behavior
There should be no error, as one can check this way:
>>> import test
>>> s = test.C("abc")
# No exception raised
>>> type(s)
<class 'test.C'>
Actual Behavior
mypy complains:
test.py:4: error: Too many arguments for "__new__" of "object"
Found 1 error in 1 file (checked 1 source file)
It erroneously believes the target of __new__ is object while it is str
Your Environment
- Mypy version used: mypy 0.782
- Mypy command-line flags: none (just the file name)
- Mypy configuration options from
mypy.ini(and other config files): none (no config file) - Python version used: Python 3.8.2
Activity
[-]Wrong inference of the type of the target of `__new__`[/-][+]Wrong inference of the target of `__new__`[/+]gvanrossum commentedon Sep 3, 2020
I think this could actually be fixed in typeshed -- the problem seems to be that no
__new__method is declared for classstr.But another interpretation would be that we ought to automatically derive a signature for
__new__if one is given for__init__. That would not be a simple feature to add, and I suspect there's already some other issue open for it. (Maybe #5647 is related?)Hibou57 commentedon Sep 3, 2020
I’m not good at Python insides, anyway, my understanding is that the two issues are not that much related. The other issue as I understand it, is about
__new__vs__init__to infer the type returned by a class constructor (I agree__new__is the reference for that).I did not knew about typeshed, so just had a look: https://github.com/python/typeshed/blob/master/stdlib/2and3/builtins.pyi#L443
Indeed, no
__new__is defined. But I was believing the arguments to__new__and__init__, except forclsandself, are typically the same. If I’m not too wrong, when no__new__is defined, what about relying on__init__to infer its expected arguments? In that case,__new__(cls, o: object)could be inferred and would match the invocation.Hibou57 commentedon Sep 6, 2020
That said, I wonder if the error is finally in typeshed: I believe that’s a
__new__which should be declared forstr, not an__init__(and if so, probably the same for others likeint,float,bool). There seems to be no__init__forstr.gvanrossum commentedon Sep 6, 2020
JelleZijlstra commentedon Sep 6, 2020
We tried this before in python/typeshed#1352 and terrible things happened (see python/typeshed#1464). Perhaps the enum issue I found back then got solved some other way since then, though.
AlexWaygood commentedon Mar 25, 2022
Cannot be reproduced on mypy 0.800+