bpo-44953: Add vectorcall for itemgetter and attrgetter

[sweeneyde]

https://bugs.python.org/issue44953

[sweeneyde]

This was inspired by GH-27782, which does the same, and more, for methodcaller.

[sweeneyde]

Benchmarks for this change:

itemgetter: Mean +- std dev: [operator_main] 45.3 ns +- 1.3 ns -> [operator_vec] 29.5 ns +- 0.7 ns: 1.54x faster
attrgetter: Mean +- std dev: [operator_main] 61.6 ns +- 1.7 ns -> [operator_vec] 43.8 ns +- 0.9 ns: 1.41x faster

[bedevere-bot]

🤖 New build scheduled with the buildbot fleet by @rhettinger for commit f9fc83a 🤖

If you want to schedule another build, you need to add the ":hammer: test-with-buildbots" label again.

[github-actions]

This PR is stale because it has been open for 30 days with no activity.

[sweeneyde]

The __vectorcalloffset__ behavior was added in GH-20026. I don't think I understand that design decision rather than adding a new PyType_Slot.slot value called something like Py_tp_vectorcall, but it appears to be what we have. The corresponding behavior of __dictoffset__ and __weaklistoffset__ goes back to 2019 in GH-16076. Are there any plans to change these, or is there anything holding this up?

[sweeneyde]

Any plans for this? It may also be relevant to bpo-46148 about optimizing Pathlib.

[sweeneyde]

Does anyone with a commit bit want to merge this? This is the kind of callable that's often called in loops.

[sweeneyde]

I re-ran some benchmarks, and they still look good. I plan to merge this in about 24 hours if there are no more objections.

from operator import itemgetter, attrgetter
from itertools import repeat
from collections import namedtuple, deque
from pyperf import Runner

class DefaultClass:
    def __init__(self, a, b):
        self.a = a
        self.b = b

class SlotsClass:
    __slots__ = "a", "b"
    def __init__(self, a, b):
        self.a = a
        self.b = b

NamedTuple = namedtuple("NT", ["a", "b"])
MAP_LOOPS = 10_000

attr_classes = {
    'DefaultClass': DefaultClass,
    'SlotsClass': SlotsClass,
    'NamedTuple': NamedTuple,
}

item_classes = {
    'tuple': tuple,
    'list': list,
    'dict': dict.fromkeys,
}

namespace = {
    'IG': itemgetter(1),
    'AG': attrgetter('a'),
    'repeat': repeat,
    'deque': deque,
} | attr_classes | item_classes

runner = Runner()

for classname in attr_classes:
    runner.timeit(
        name=f"{classname}-1",
        setup=f"obj = {classname}(11, 22)",
        stmt="AG(obj)",
        globals=namespace
    )
    runner.timeit(
        name=f"{classname}-map",
        setup=f"obj = {classname}(11, 22)",
        stmt=f"deque(map(AG, repeat(obj, {MAP_LOOPS})), maxlen=0)",
        globals=namespace,
        inner_loops=MAP_LOOPS,
    )

for classname in item_classes:
    runner.timeit(
        name=f"{classname}-1",
        setup=f"obj = {classname}((1, 2, 3, 4, 5))",
        stmt="IG(obj)",
        globals=namespace,
    )
    runner.timeit(
        name=f"{classname}-map",
        setup=f"obj = {classname}((1, 2, 3, 4, 5))",
        stmt=f"deque(map(IG, repeat(obj, {MAP_LOOPS})), maxlen=0)",
        globals=namespace,
        inner_loops=MAP_LOOPS,
    )

Results:

Benchmark	main_bench	PR-27828
tuple-map	27.6 ns	13.2 ns: 2.09x faster
dict-map	39.4 ns	24.2 ns: 1.62x faster
NamedTuple-map	40.0 ns	24.8 ns: 1.61x faster
list-map	38.2 ns	23.8 ns: 1.61x faster
SlotsClass-map	42.8 ns	28.1 ns: 1.53x faster
DefaultClass-map	45.5 ns	31.0 ns: 1.47x faster
tuple-1	56.7 ns	41.4 ns: 1.37x faster
NamedTuple-1	70.1 ns	52.7 ns: 1.33x faster
dict-1	69.4 ns	52.2 ns: 1.33x faster
SlotsClass-1	75.1 ns	56.6 ns: 1.33x faster
list-1	68.7 ns	52.1 ns: 1.32x faster
DefaultClass-1	76.9 ns	59.4 ns: 1.29x faster
Geometric mean	(ref)	1.48x faster