Mypy not resolving protocol class as callable in check_reverse_op_method (and possible fix)
Nobody has claimed this yet.
- Dominant language
- Python
- Stars
- 20.6k
- Forks
- 3.3k
- PR merge metrics
- PR metrics pending
Description
Bug Report
I'm writing a custom vector class for a specific type of simulation I'm working on. This class implements, amongst other operations, __add__ and __radd__ operations, with numpy floating point arguments. However, for some reason, Mypy flags a np.floating type as not having a callable __add__ method.
To Reproduce
Minimal example.py:
from __future__ import annotations
import numpy as np
class my_float:
def __radd__(self, other: np.float32) -> my_float:
return my_float()
Running mypy example.py:
example.py:6: error: Forward operator "__add__" is not callable
Found 1 error in 1 file (checked 1 source file)
More investigation
At first, I thought this was perhaps a bug in numpy, but I am not convinced. Numpy, even though all arithmetic operators are based on ufuncs written in C, provides accurate static typing stubs for them. Mypy can resolve these correctly:
a = np.float32()
b = np.float32()
reveal_type(a + b) # Revealed type is "numpy.floating[numpy.typing._32Bit]"
I believe the problem lies in that for the np.floating class and derivatives, __add__ (and other operators) are not directly defined as (overloaded) methods, but hinted as __add__: _FloatOp[_NBit1]
Here,
class _FloatOp(Protocol[_NBit1]):
@overload
def __call__(self, __other: bool) -> floating[_NBit1]: ...
@overload
def __call__(self, __other: int) -> floating[Union[_NBit1, _NBitInt]]: ...
@overload
def __call__(self, __other: float) -> floating[Union[_NBit1, _NBitDouble]]: ...
@overload
def __call__(
self, __other: complex
) -> complexfloating[Union[_NBit1, _NBitDouble], Union[_NBit1, _NBitDouble]]: ...
@overload
def __call__(
self, __other: Union[integer[_NBit2], floating[_NBit2]]
) -> floating[Union[_NBit1, _NBit2]]: ...
I had a look in more detail in how this results in mypy not resolving the operator as callable, by cloning the current mypy master and digging into it. In the end, I believe the problem lies in the mypy method check_reverse_op_method, more specifically in the call here: https://github.com/python/mypy/blob/21897149830f36fee8d7c639b8d7596ce6e24ea4/mypy/checker.py#L1240
Instead of returning some function-like variable, it only returns an instance of type:
str(forward_type) = numpy.typing._callable._FloatOp[numpy.typing._32Bit*]
This is different than for example changing in my example.py the argument to other: float, which gives:
str(forward_type) = def (builtins.float) -> builtins.float
The forward_type is used in a call to check_overlapping_op_methods, and since there is no check for isinstance(forward_item, Instance), we end up here in the case of my original example.py:
https://github.com/python/mypy/blob/21897149830f36fee8d7c639b8d7596ce6e24ea4/mypy/checker.py#L1306
resulting in the error as in my example.
One (perhaps dirty) fix would be to add a simple test for an instance right after line 1293 in check_overlapping_op_methods and test if it is callable, and resolve that:
if isinstance(forward_item, Instance):
if forward_item.has_readable_member('__call__'):
forward_item = self.expr_checker.analyze_external_member_access('__call__',
forward_item,
context)
This fixes the issue. However, I do not know enough about mypy internals to know why the analyze_external_member_access call does not resolve the (overloaded) __call__ method in the first place. I have tried to get it to return an instance with some non-protocol classes with overloaded __call__ methods, but I cannot easily reproduce this issue.
Any other suggestions? Or is this perhaps an implementation error on the part of Numpy?
Apologies if my issue submission is too long, but since I already did some digging I thought to share it as well.
Your Environment
- Mypy version used: 0.92
- Numpy version used: 1.21.4
- Python version used: 3.9.7
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Research direction
Start with the minimal example.py reproduction and inspect mypy/checker.py around check_reverse_op_method and check_overlapping_op_methods, using the linked source locations. Compare the protocol-based numpy operator path with the float path described in the issue. Done means mypy accepts the radd example without the “Forward operator add is not callable” error and the behavior is covered appropriately.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- numpy, python
- Domain
- devtools, tooling
- Issue type
- Bug
- Difficulty
- 4/5
- Estimated time
- 3-5 days
- Activity status
- Stale
- Clarity
- Mostly clear
- Newbie friendliness
- 35/100