Checker for unspecified-encoding is unable to infer attributes from dataclasses

[DanielNoord]

Bug description

A small part of this issue has been fixed in #5332, but the following tests should still be added to tests/functional/u/unspecified_encoding_py38.py.

import dataclasses
from typing import Optional

FILENAME = "test_two.py"


@dataclasses.dataclass
class IOArgs:
    """Dataclass storing information about how to open a file"""

    encoding: Optional[str]
    mode: str


args_good_one = IOArgs(encoding=None, mode="wb")
args_good_two = IOArgs(encoding="utf8", mode="w")
args_bad_one = IOArgs(encoding=None, mode="w")
LOCALE_ENCODING = None
args_bad_two = IOArgs(encoding=LOCALE_ENCODING, mode="w")

file = open(FILENAME, args_good_one.mode, encoding=args_good_one.encoding)
file = open(FILENAME, args_good_two.mode, encoding=args_good_two.encoding)
file = open(FILENAME, args_bad_one.mode, encoding=args_bad_one.encoding)
file = open(FILENAME, args_bad_two.mode, encoding=args_bad_two.encoding)

Configuration

No response

Command used

pylint test.py

Pylint output

No response

Expected behavior

It has something to do with the inference of dataclasses but I couldn't figure it out easily. I only know that astroid handles this in a special way but I have no knowledge of that part of the codebase.
Any help would be appreciated!

Found while working on #5173

Pylint version

Latest commit on `main`

OS / Environment

No response

Additional dependencies

No response

[Pierre-Sassoulas]

I don't know what this Proxy class is doing, git log --pretty=short -u -L 128,136:astroid/bases.py does not bring a lot of information either. It seems like it's making everything more complicated though, shouldn't we try to remove the class and handle it in each class with proper typing ? Any idea about that @hippo91 ?

[Pierre-Sassoulas]

I think this is a duplicate of #5266

[cdce8p]

This seems like a two part issue:

1. args_good_one.encoding is inferred as <Instance of builtins.str>. It should have been a Const (string) node with value wb. Likely an issue in the dataclass brain.
2. _check_open_encoded does an unprotected attribute access to mode_arg.value. A fix could be to add an isinstance(mode_arg, nodes.Const) check. I would think this should resolve 'ClassDef' object has no attribute 'value' #5266 as well.

https://github.com/PyCQA/pylint/blob/7d2b6c9bd807dda7502e739c7b9fb2182d89f457/pylint/checkers/stdlib.py#L629

[cdce8p]

I just removed the milestone here. The crash was fixed with #5332 (closed #5266 for it). Should we leave this open for the dataclass inference issue?

[DanielNoord]

I think so! I'll update the OP and title!