There is a special handling of `__hash__` set to None in the interpreter core. This is because every class inherited the `__hash__` attribute from "object", and setting `__hash__ = None` is a simple way to make it unhashable. It makes hash() raising the correct type of exception (TypeError), but with unhelpful error message "'NoneType' object is not callable". The special case was added to make the error message more relevant: "unhashable type: '{typename}'". There is similar situation with other special methods defined in "object" or other common classes. Sometimes we want to cancel the default inherited behavior. >>> dir(object) ['__class__', '__delattr__', '__dir__', '__doc__', '__eq__', '__format__', '__ge__', '__getattribute__', '__getstate__', '__gt__', '__hash__', '__init__', '__init_subclass__', '__le__', '__lt__', '__ne__', '__new__', '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', '__sizeof__', '__str__', '__subclasshook__'] I propose to support officially the idiom "__dunder__ = None" and add special cases to raise more specialized exception instead of "TypeError: 'NoneType' object is not callable" for most of special method where cancelling the default behavior makes sense (for example I do not think that we need better error message for `__repr__ = None`). The question is how to interpret value None: * Always raise TypeError (with changed message)? This is what happen currently when you set the method to None, this is the most compatible option. * Always raise an error, but allow to change it to more appropriate type (for example AttributeError for __setattr__)? * Interpret value None the same way as an absent attribute? For `__hash__` or `__class_getitem__` all three options mean the same. But absent `__mro_entries__` and `__mro_entries__ = None` currently give different results. It is even more complicated for pickling: absent `__reduce_ex__` and `__reduce_ex__ = None` mean the same in the Python implementation, but give different results in the C implementation.
It actually is documented as being supported here: https://docs.python.org/3.10/reference/datamodel.html#id2, and as mentioned there __iter__(), __reversed__() and __contains__() also have special support so they avoid trying to fallback to __getitem__ etc. Also some discussion at https://github.com/python/cpython/issues/70146. For most methods the conclusion back then makes sense, a specific exception message being raised would just potentially slow the interpreter. - Spencer On 1 May 2022, at 3:22 pm, Serhiy Storchaka <storchaka@gmail.com> wrote: There is a special handling of `__hash__` set to None in the interpreter core. This is because every class inherited the `__hash__` attribute from "object", and setting `__hash__ = None` is a simple way to make it unhashable. It makes hash() raising the correct type of exception (TypeError), but with unhelpful error message "'NoneType' object is not callable". The special case was added to make the error message more relevant: "unhashable type: '{typename}'". There is similar situation with other special methods defined in "object" or other common classes. Sometimes we want to cancel the default inherited behavior.
dir(object) ['__class__', '__delattr__', '__dir__', '__doc__', '__eq__', '__format__', '__ge__', '__getattribute__', '__getstate__', '__gt__', '__hash__', '__init__', '__init_subclass__', '__le__', '__lt__', '__ne__', '__new__', '__reduce__', '__reduce_ex__', '__repr__', '__setattr__', '__sizeof__', '__str__', '__subclasshook__']
I propose to support officially the idiom "__dunder__ = None" and add special cases to raise more specialized exception instead of "TypeError: 'NoneType' object is not callable" for most of special method where cancelling the default behavior makes sense (for example I do not think that we need better error message for `__repr__ = None`). The question is how to interpret value None: * Always raise TypeError (with changed message)? This is what happen currently when you set the method to None, this is the most compatible option. * Always raise an error, but allow to change it to more appropriate type (for example AttributeError for __setattr__)? * Interpret value None the same way as an absent attribute? For `__hash__` or `__class_getitem__` all three options mean the same. But absent `__mro_entries__` and `__mro_entries__ = None` currently give different results. It is even more complicated for pickling: absent `__reduce_ex__` and `__reduce_ex__ = None` mean the same in the Python implementation, but give different results in the C implementation. _______________________________________________ Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-leave@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/YGAK34DR... Code of Conduct: http://python.org/psf/codeofconduct/
01.05.22 13:02, Spencer Brown пише:
It actually is documented as being supported here: https://docs.python.org/3.10/reference/datamodel.html#id2 <https://docs.python.org/3.10/reference/datamodel.html#id2>, and as mentioned there __iter__(), __reversed__() and __contains__() also have special support so they avoid trying to fallback to __getitem__ etc. Also some discussion at https://github.com/python/cpython/issues/70146 <https://github.com/python/cpython/issues/70146>. For most methods the conclusion back then makes sense, a specific exception message being raised would just potentially slow the interpreter.
Thank you Spencer. It answers my questions.
On 01/05/2022 06:20, Serhiy Storchaka wrote:
The question is how to interpret value None:
* Always raise TypeError (with changed message)? This is what happen currently when you set the method to None, this is the most compatible option. * Always raise an error, but allow to change it to more appropriate type (for example AttributeError for __setattr__)? * Interpret value None the same way as an absent attribute? What about binary operators (__add__, __eq__, etc)? Should they act as if they'd returned NotImplemented? Or immediately unconditionally raise a TypeError?
Non should behave as closely as possible to it not being defined at all. So return NotImplemented. On Sun, May 1, 2022 at 09:53 Patrick Reader <_@pxeger.com> wrote:
On 01/05/2022 06:20, Serhiy Storchaka wrote:
The question is how to interpret value None:
* Always raise TypeError (with changed message)? This is what happen currently when you set the method to None, this is the most compatible option. * Always raise an error, but allow to change it to more appropriate type (for example AttributeError for __setattr__)? * Interpret value None the same way as an absent attribute? What about binary operators (__add__, __eq__, etc)? Should they act as if they'd returned NotImplemented? Or immediately unconditionally raise a TypeError?
Python-Dev mailing list -- python-dev@python.org To unsubscribe send an email to python-dev-leave@python.org https://mail.python.org/mailman3/lists/python-dev.python.org/ Message archived at https://mail.python.org/archives/list/python-dev@python.org/message/PLWKIT7F... Code of Conduct: http://python.org/psf/codeofconduct/
-- --Guido (mobile)
01.05.22 19:32, Guido van Rossum пише:
Non should behave as closely as possible to it not being defined at all. So return NotImplemented.
But, as was noted by Spencer Brown, it is not how it works for __iter__(), __reversed__() and __contains__() (which have a special path for producing explicit error), and not how it currently works for __add__() and __eq__() (although they raise TypeError implicitly). It is all documented inhttps://docs.python.org/3.10/reference/datamodel.html#id2 If we are going to reconsider this, it is a large breaking change.
participants (4)
-
Guido van Rossum -
Patrick Reader -
Serhiy Storchaka -
Spencer Brown