[rheard]

Fixes mypyc/mypyc#1213.

A compiled frozen dataclass can't be pickled or copied, so it can't be sent to a multiprocessing worker either:

import copy
import pickle
from dataclasses import dataclass

@dataclass(frozen=True)
class Config:
    py_version: str = "3"

pickle.loads(pickle.dumps(Config()))  # FrozenInstanceError (works when interpreted)
copy.copy(Config())                   # FrozenInstanceError

The __setstate__ that mypyc gives native classes restored each attribute with PyObject_SetAttr, which calls the class's __setattr__, and the one @dataclass(frozen=True) adds always raises.

CPython restores the state of an object by updating its __dict__ directly, without calling __setattr__, and for a frozen dataclass with slots=True, dataclasses adds a __setstate__ that uses object.__setattr__. Now __setstate__ sets the attributes with PyObject_GenericSetAttr, as object.__setattr__ does. The copies are still frozen.

This only changes anything for classes with their own __setattr__, since PyObject_GenericSetAttr is already the tp_setattro of the others. A native class that defines __setattr__ no longer has it called when an instance is unpickled or copied, which matches CPython for a class without __slots__.

[p-sawicki]

thanks!