RFE: @dataclasses.dataclass(slots=True) doesn't support methods using zero-arg super()

[Tinche]

BPO	46404
Nosy	@vstinner, @ericvsmith, @encukou, @ericsnowcurrently, @hynek, @Tinche, @frenzymadness, @sweeneyde

^{Note: these values reflect the state of the issue at the time it was migrated and might not reflect the current state.}

Show more details

GitHub fields:

assignee = None
closed_at = None
created_at = <Date 2022-01-16.22:48:06.738>
labels = ['interpreter-core', 'type-feature', '3.11']
title = "RFE: @dataclasses.dataclass(slots=True) doesn't support methods using closures"
updated_at = <Date 2022-03-28.15:16:37.392>
user = 'https://github.com/Tinche'

bugs.python.org fields:

activity = <Date 2022-03-28.15:16:37.392>
actor = 'vstinner'
assignee = 'none'
closed = False
closed_date = None
closer = None
components = ['Interpreter Core']
creation = <Date 2022-01-16.22:48:06.738>
creator = 'tinchester'
dependencies = []
files = []
hgrepos = []
issue_num = 46404
keywords = []
message_count = 8.0
messages = ['410730', '410745', '411462', '415291', '415300', '415303', '416169', '416174']
nosy_count = 8.0
nosy_names = ['vstinner', 'eric.smith', 'petr.viktorin', 'eric.snow', 'hynek', 'tinchester', 'frenzy', 'Dennis Sweeney']
pr_nums = []
priority = 'normal'
resolution = None
stage = None
status = 'open'
superseder = None
type = 'enhancement'
url = 'https://bugs.python.org/issue46404'
versions = ['Python 3.11']

Linked PRs

• gh-90562: Update __class__ in function closures for dataclasses with slots=True #104038
• gh-90562: Mention slots pitfall in dataclass docs #107391
• gh-90562: Fix super() without args calls for dataclasses with slots #111538
• [3.13] gh-90562: Mention slots pitfall in dataclass docs (GH-107391) #119350
• [3.12] gh-90562: Mention slots pitfall in dataclass docs (GH-107391) #119351
• gh-90562: Support zero argument super with dataclasses when slots=True #124455
• gh-90562: Improve zero argument support for super() in dataclasses when slots=True #124692

[Tinche]

We've received a report over at the attrs issue tracker about our test suite failing on Python 3.11. Here's the link: python-attrs/attrs#907

It turns out to be an issue with the no-arg super() calls in slotted classes. Here's a minimal reproducer example:

from attrs import define


@define
class A:
 pass


@define
class B(A):
 def test(self):
 super()


B().test()

Traceback (most recent call last):
 File "/Users/tintvrtkovic/pg/attrs/a01.py", line 15, in <module>
 B().test()
 ^^^^^^^^^^
 File "/Users/tintvrtkovic/pg/attrs/a01.py", line 12, in test
 super()
 ^^^^^^^
TypeError: super(type, obj): obj must be an instance or subtype of type

This is a known issue for which we have implemented workarounds. The workarounds aren't effective for 3.11 though. I have implemented a fix in attrs (python-attrs/attrs#910), but I still thought I'd post this here to maybe get the core devs opinion.

Dataclasses exhibit the exact same issue when used with slots=True, both in 3.10 when slots was added and in 3.11. I guess no one reported it or tried fixing it.

A comprehensive description of the issue follows: since it's impossible to add slotness (i.e. set __slots__) to a class after it has been created, when creating a slotted class the class decorators in attrs and dataclasses actually replace the class they are applied to with a copy of it, with slots added. This works, except in the case of the no-arg super() being used in any of the class methods (and maybe another edge case that I can't remember). When the compiler encounters the no-arg super() form, it adds some state to the function __closure__ cells. This state causes the exception shown above, since it's incorrect when the class gets replaced.

So these closure cells need to be rewritten when the class is replaced. In Python versions prior to 3.11, the closure cells were immutable so extra effort was needed to rewrite them. The functions are here: https://github.com/python-attrs/attrs/blob/9727008fd1e40bc55cdc6aee71e0f61553f33127/src/attr/_compat.py#L145.

In 3.11, our old closure cell rewriting doesn't work any more, but closure cells don't appear to be immutable either, so the fix in my attr PR linked above is simple. Still, it's another branch in the code to support a specific version.

I don't know if there's anything actionable here for Python, apart from confirming or denying if this behavior is expected.

[sweeneyde]

bisected to here:

631f993 is the first new commit
commit 631f993
Author: Eric Snow <ericsnowcurrently@gmail.com>
Date: Mon Jun 7 16:52:00 2021 -0600

bpo-43693: Add the MAKE_CELL opcode and interleave fast locals offsets. (gh-26396)

This moves logic out of the frame initialization code and into the compiler and eval loop. Doing so simplifies the runtime code and allows us to optimize it better.

https://bugs.python.org/issue43693

[encukou]

I guess at least there should be a warning about this in dataclasses docs?
The reproducer with dataclasses (which exhibits the same error on 3.10 and 3.11):

import dataclasses

@dataclasses.dataclass(slots=True)
class A:
 pass


@dataclasses.dataclass(slots=True)
class B(A):
 def test(self):
 super()

B().test()

[frenzymadness]

In my opinion, we should keep it simple for attrs and dataclasses to fix closure cells when a class is replaced and therefore it seems to be correct to have it mutable as it currently is in 3.11.

My plan is to implement the fix for dataclasses and some tests for these use cases so the behavior should not change in the future. For attrs it means one more branch in the code now but much simpler code when the support for older releases gets dropped.

Any other opinions?

[ericvsmith]

@frenzy: I'm not sure what your fix would do. You could either describe it in rough terms (if you'd like a pre-PR opinion on the approach), or I'm happy to wait to see your PR.

[frenzymadness]

We have the same problem reported in attrs here in dataclasses and because it's not tested the way to manipulate __closure__ cells changes frequently.

My plan is to implement something similar to this into dataclasses: https://github.com/python-attrs/attrs/blob/5c040f30e3e4b3c9c0f27c8ac6ff13d604c1818c/src/attr/_make.py#L895-L916

Basically, when a new dataclass is created (with slots=True), look for references to the original class and fix them.

This fixes the problem reported to attrs in dataclasses and when we fix it and add some tests for it, the future behavior should be more stable.

What do you think?

[vstinner]

I changed the issue title to focus this issue on enhance dataclasses to support @dataclasses.dataclass(slots=True) on methods using closures: it would be a new Python 3.11 feature.

I created https://bugs.python.org/issue47143 "Add functools.copy_class() which updates closures".