Make FakeInfo a more general-purpose TypeInfo placeholder #5469
Conversation
There are a lot of places in mypy currently that shove None into fields expecting TypeInfo. Since mypyc does not appreciate this kind of behavior, use FakeInfo instances instead of None. FakeInfo is generalized to allow different assertion messages and to override `__bool__` so that it can be conveniently checked against.
mypy/nodes.py
Outdated
| self.info = FUNC_NO_INFO # type: TypeInfo | ||
| self.is_property = False # type: bool | ||
| self.is_class = False # type: bool | ||
| self.is_static = False # type: bool |
There was a problem hiding this comment.
Why do you need these three bools here?
There was a problem hiding this comment.
Because referring to FUNC_NO_INFO, which is defined later in the file, causes __init__ to get second-pass typechecked, which causes these to not get inferred right or something.
Annotating the FakeInfos fixes this and the weird subtype.py assert also so I'm doing that instead.
mypy/nodes.py
Outdated
| @@ -2184,6 +2185,7 @@ def __init__(self, names: 'SymbolTable', defn: ClassDef, module_name: str) -> No | |||
| self.assuming_proper = [] | |||
| self.inferring = [] | |||
| self.add_type_vars() | |||
| self.replaced = TYPEINFO_NO_REPLACED | |||
There was a problem hiding this comment.
I would actually make this Optional, it is used in few places, so should be easy to fix. Also I think we have an issue about removing replaced altogether.
There was a problem hiding this comment.
Oh, that was easy.
mypy/nodes.py
Outdated
| if (isinstance(node, Var) and node.type is not None): | ||
| return node.type | ||
| # mypy thinks this branch is unreachable but it is wrong | ||
| elif (isinstance(node, FuncBase) and node.type is not None): |
There was a problem hiding this comment.
Hm, is this a bug then? Do we have an issue for this?
There was a problem hiding this comment.
It's the "intersection types bug". #3603. I'll point to it in the comments
mypy/subtypes.py
Outdated
| @@ -560,6 +560,7 @@ def find_node_type(node: Union[Var, FuncBase], itype: Instance, subtype: Type) - | |||
| else: | |||
| typ = signature | |||
| itype = map_instance_to_supertype(itype, node.info) | |||
| assert typ is not None | |||
There was a problem hiding this comment.
This seems to be redundant, could you please add a comment why this is needed?
There was a problem hiding this comment.
It's a knock-on effect of Var's __init__ getting typechecked in the second pass I think. But it is pretty bogus and probably a mypy bug, but I don't think I have the cycles right now to track it down. Annotating the FakeInfos fixes the problem, though. SIgh.
mypy/typeanal.py
Outdated
| @@ -413,7 +413,8 @@ def visit_callable_type(self, t: CallableType, nested: bool = True) -> Type: | |||
| variables = self.bind_function_type_variables(t, t) | |||
| ret = t.copy_modified(arg_types=self.anal_array(t.arg_types, nested=nested), | |||
| ret_type=self.anal_type(t.ret_type, nested=nested), | |||
| fallback=t.fallback or self.named_type('builtins.function'), | |||
| fallback=(t.fallback if t.fallback.type | |||
There was a problem hiding this comment.
This pattern is a bit confusing, I would add a little remainder here that FakeInfo is Falsey. (maybe also below?)
There are a lot of places in mypy currently that shove None into
fields expecting TypeInfo. Since mypyc does not appreciate this kind
of behavior, use FakeInfo instances instead of None. FakeInfo is
generalized to allow different assertion messages and to override
__bool__so that it can be conveniently checked against.