LOAD_ATTR cache does not fully replicate PyObject_GetAttr behavior

[KevinModzelewski]

BPO	42266
Nosy	@pmp-p, @1st1, @pablogsal, @Jongy
PRs	bpo-42266: Handle monkey-patching descriptors in LOAD_ATTR cache #23157
Files	test.py: Test case with unexpected opcache behavior

^{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 = <Date 2020-11-05.09:23:52.835>
created_at = <Date 2020-11-04.21:56:10.288>
labels = ['interpreter-core', 'type-bug', '3.10']
title = 'LOAD_ATTR cache does not fully replicate PyObject_GetAttr behavior'
updated_at = <Date 2020-11-05.09:23:52.835>
user = 'https://bugs.python.org/KevinModzelewski'

bugs.python.org fields:

activity = <Date 2020-11-05.09:23:52.835>
actor = 'pablogsal'
assignee = 'none'
closed = True
closed_date = <Date 2020-11-05.09:23:52.835>
closer = 'pablogsal'
components = ['Interpreter Core']
creation = <Date 2020-11-04.21:56:10.288>
creator = 'Kevin Modzelewski'
dependencies = []
files = ['49568']
hgrepos = []
issue_num = 42266
keywords = ['patch']
message_count = 7.0
messages = ['380370', '380374', '380375', '380376', '380378', '380379', '380399']
nosy_count = 5.0
nosy_names = ['pmpp', 'yselivanov', 'Kevin Modzelewski', 'pablogsal', 'Yonatan Goldschmidt']
pr_nums = ['23157']
priority = 'normal'
resolution = 'fixed'
stage = 'resolved'
status = 'closed'
superseder = None
type = 'behavior'
url = 'https://bugs.python.org/issue42266'
versions = ['Python 3.10']

[KevinModzelewski]

The problem is that the descriptor-ness of a type-level attribute is only checked at opcache-set time, not at opcache-hit time.

$ python3.8 test.py
2

$ ./python --version
Python 3.10.0a2+
$ git rev-parse --short HEAD
789359f47c
$ ./python test.py
1

[pablogsal]

Good catch, Kevin! Would you like to submit a PR for fixing this?

[pablogsal]

Given that having attributes that are classes is quite uncommon, I think we can not optimize of the attribute itself is a class instead of checking for descriptors on every hit, hurting the performance gains

[pablogsal]

Yury, any preference here?

[pablogsal]

We could also store the tag of the type object if is a descriptor and compare against that on the cache hit to check that our assumptions are valid. The price here would be an extra pointer on the cache per opcode that may not even be used most of the time.

[pablogsal]

s/the attribute itself is a class/the attribute itself is in the class/

[pablogsal]

New changeset 80449f2 by Pablo Galindo in branch 'master':
bpo-42266: Handle monkey-patching descriptors in LOAD_ATTR cache (GH-23157)
80449f2