gh-132042: Speed up slot lookup during class creation - #159479

fixup_slot_dispatchers() resolves the special-method slots of a new class with one dict lookup per slot name and per class in the MRO: about 92 names times the MRO length, ~180 lookups for class A: pass and ~650 for a MutableMapping subclass. This PR scans each dict along the MRO once instead, matches its dunder keys against a small static table of slot names, and records the first hit per slot. The rest of update_one_slot() is unchanged and the result is identical to find_name_in_mro() for every slot. The per-name lookup remains as the fallback for dicts that cannot be scanned (non-string keys, more than 1000 entries, or on the free-threaded build a dict whose mutex is held by another thread, since the scan never blocks while holding the type lock) and for update_slot().

Compared to earlier work on this problem, there is no caching and no per-type memory (#4902 (gh-76527) cached the slot descriptors ; #132618 builds a merged MRO dict per class creation (allocation plus a PyDict_Merge per base) and had to change test_descr). A regression test covers class creation racing with setattr on a base on the free-threaded build.

@sergey-miryanov

Benchmarks

-O3 non-PGO build on Linux, pyperf pinned to one core:

The benchmark script from #132156 / #132618 (1000-iteration cases):

Notes: #132618 no longer applies to main, so for a direct comparison its approach was re-implemented on top of current main. On 14 of the cases above, run back to back on the same machine, this PR is 1.78x faster than main (geometric mean) and the #132618 approach 1.39x.

Whole-process effect: python -I -c pass executes 8% fewer instructions (47.6M to 43.8M, callgrind), and importing class-heavy stdlib modules (asyncio, typing, unittest.mock, email.message, argparse, inspect, decimal, tkinter) is 4-8% faster in wall time.

import pyperf
runner = pyperf.Runner()
runner.timeit("class A: pass", stmt="class A: pass")
runner.timeit("class E(D) depth-5 chain", setup="class A: pass\nclass B(A): pass\nclass C(B): pass\nclass D(C): pass", stmt="class E(D): pass")
runner.timeit("class T(MagicMock)", setup="from unittest.mock import MagicMock", stmt="class T(MagicMock): pass")
runner.timeit("class T(MutableMapping)", setup="from cohis PR
runner.timeit("types.new_class", setup="import types", stmt="types.new_class('A')")
runner.timeit("typing.Protocol", setup="from typing import Protocol", stmt="class P(Protocol):\n    def f(self) -> int: ...")
runner.timeit("typing.TypedDict", setup="from typing import TypedDict", stmt="class P(TypedDict):\n    x: int\n    y: int")
runner.timeit("namedtuple", setup="from collections import namedtuple", stmt="namedtuple('P', 'x y z')")
runner.timeit("Enum", setup="from enum import Enum", stmt="class Color(Enum):\n    RED = 1\n    GREEN = 2")
runner.timeit("dataclass", setup="from dataclasses import dataclass", stmt="@dataclass\nclass P:\n    x: int\n    y: int")
runner.timeit("setattr slot dunder", setup="class A: pass\nf = lambda s: 0", stmt="A.__len__ = f")
runner.timeit("setattr non-slot dunder", setup="class A: pass", stmt="A.__tablename__ = 'x'")

Generated with help from Claude Code

• Issue: Optimize class creation #132042