mypyc: dataclasss __annotations__ are wrong

Bug Report

The __annotations__ of a compiled dataclass are not the same as its pure-python counterpart.

To Reproduce

1. Create a file with a dataclass

from dataclasses import dataclass

@dataclass
class Person:
    age : int
    name : str

2. check it's annotations

python -c "import myfile;print(myfile.Person.__annotations__)"
{'age': <class 'int'>, 'name':  <class 'str'>}

3. Compile it

4. Check the compiled annotations:

python -c "import myfile;print(myfile.Person.__annotations__)"
{'age': <class 'type'>, 'name':  <class 'type'>}

Expected Behavior

The annotations for the compiled class should be {'age': <class 'int'>, 'name': <class 'str'>} but instead they are all <class 'type'>.

I see this in the code:

def add_non_ext_class_attr(builder: IRBuilder,
                           non_ext: NonExtClassInfo,
                           lvalue: NameExpr,
                           stmt: AssignmentStmt,
                           cdef: ClassDef,
                           attr_to_cache: List[Tuple[Lvalue, RType]]) -> None:
    """Add a class attribute to __annotations__ of a non-extension class.

    If the attribute is initialized with a value, also add it to __dict__.
    """
    # We populate __annotations__ because dataclasses uses it to determine
    # which attributes to compute on.
    # TODO: Maybe generate more precise types for annotations
    key = builder.load_str(lvalue.name)
    typ = builder.add(LoadAddress(type_object_op.type, type_object_op.src, stmt.line))
    builder.call_c(dict_set_item_op, [non_ext.anns, key, typ], stmt.line)

So it looks like it's just hard-wiring type_object_op.

It's important to have accurate type annotations, because dataclasses are often inspected dynamically in order to automatically serialize them, generate protobufs, create UIs, etc based on their field data types.

Your Environment

• Mypy version used: 0.920 master
• Python version used: 3.8
• Operating system and version: mac

まず dataclass の例で不一致を再現し、次に mypyc の add_non_ext_class_attr 関数と、そこでの type_object_op の使用方法を調べます。アノテーション値がどのように選択されるかを追跡し、コンパイルされたクラスが pure-Python 版と同じ int および str のアノテーションを公開することを確認します。