Description
Bug report
Bug summary
In basic_units.py, the TaggedValue has a __new__ method that does the following:
matplotlib/examples/units/basic_units.py
Lines 120 to 128 in f710ad3
However, when value_class is not object, then object.__new__(subcls) raises:
In [1]: class A:
...: def __new__(cls, value):
...: value_class = type(value)
...: subcls = type(f'A_{value_class.__name__}', (cls, value_class), {})
...: return object.__new__(subcls)
...:
In [2]: A(object())
Out[2]: <__main__.A_object at 0x7f9ce8a684c0>
In [3]: A(1.0)
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
<ipython-input-6-04f0da727658> in <module>
----> 1 A(1.0)
<ipython-input-5-3b6ef5f2aee2> in __new__(cls, value)
3 value_class = type(value)
4 subcls = type(f'A_{value_class.__name__}', (cls, value_class), {})
----> 5 return object.__new__(subcls)
6
TypeError: object.__new__(A_float) is not safe, use float.__new__()
This means that the intended subclassing there never happens, as it falls back to the generic object.__new__ in the exception handler. I came across this while trying to fix #19535 with a conditional method, but of course, that never worked because the code always ended up using the generic non-subclass.
I tried to fix this by doing super().__new__ or value_class.__new__ (and then some object.<method> needed to be super().<method> in other methods), but this runs into issues with other classes that also override __new__, namely MaskedArray.
I think it makes sense to create subclasses (and then fix it so it actually leaves out methods that shouldn't be there), but I'm not sure that multiple inheritance is the right way to do it.
Matplotlib version
- Matplotlib version (
import matplotlib; print(matplotlib.__version__)):masterat the moment - Matplotlib backend (
print(matplotlib.get_backend())): n/a - Python version: 3.7.6