Detect attribute deletion that would raise AttributeError

[p-sawicki]

Native classes don't have __dict__ so dynamic attributes have to be implemented using a custom dictionary, eg.

class Test:
    _attributes: dict[str, object] = {}

    def __setattr__(self, key: str, val: object) -> None:
        self._attributes[key] = val
    
    def __getattr__(self, key: str) -> object:
        return self._attributes.get(key)

One difference between using a custom dictionary and __dict__ is that with a custom dictionary, __delattr__ needs to be defined if the class should support deleting the dynamic attributes, eg.

i = Test()
i.new_attr = 1
del i.new_attr

By default, del deletes from __dict__ so this raises an AttributeError.

It's technically correct and the same as interpreted python but when someone converts a class with dynamic attributes into a native one, they might forget about __delattr__ and not learn about it until they try to execute a del statement.

It could be helpful if mypyc reported this at compile time since it can see that Test does not have __dict__ and new_attr is not defined as a static attribute so __delattr__ is needed.

Also consider if this could be extended to delattr(i, "new_attr") and object.__delattr__(i, "new_attr")

[BobTheBuidler]

Does it make more sense to just add a new mypyc attr for has_dict=True or dict=True which adds the __dict__ member the same way it does for a native class that inherits from an interpreted class?

I was working on something similar a while back, albeit for __weakref__ and not __dict__, but it worked fine at the time. (The PR was paused for unrelated segfaults)

Not sure if that's actually helpful here

[BobTheBuidler]

How do you indicate to mypyc that "this class is a native class that will support arbitrary attributes"? Is it the presence of this _attributes member within the struct?

[p-sawicki]

Does it make more sense to just add a new mypyc attr for has_dict=True or dict=True which adds the dict member the same way it does for a native class that inherits from an interpreted class?

yeah that could be a useful feature. i've initially tried to add __dict__ to native classes by default when they override __getattr__ or __setattr__ but i ran into some issues related to keeping the attribute values in __dict__ and in the native class struct consistent. sometimes writing in one place would not update the other. so that might take a bit of work.

How do you indicate to mypyc that "this class is a native class that will support arbitrary attributes"? Is it the presence of this _attributes member within the struct?

i think just having __setattr__ would indicate it. since without it expressions like instance.new_attr are rejected by mypy.