Improve performance by using reference stealing methods

[eendebakpt]

Feature or enhancement

Proposal:

We can improve performance of cpython by using stealing methods (_PyList_AppendTakeRef, _PyDict_SetItem_Take2, PyTuple_FromPairSteal etc.) instead of the normal ones. Using stealing methods we typically avoid an incref/decref pair and sometimes error handling is simplified. We should only make changes where it makes a performance impact, the code is not more complex and the modules or methods affected are performance critical.

A list of possible cases (there might be more).

Location	Module / context	Helper
Modules/_csv.c:688	csv reader, per field	_PyList_AppendTakeRef
Modules/pyexpat.c:472	pyexpat StartElement, per attribute	_PyDict_SetItem_Take2
Modules/_elementtree.c:3351	ElementTree C accel attrib dict	_PyDict_SetItem_Take2
Modules/_io/bytesio.c:577	io.BytesIO.readlines	_PyList_AppendTakeRef
Modules/_io/iobase.c:756	io.IOBase.readlines (generic)	_PyList_AppendTakeRef
Objects/frameobject.c:639 & 658	frame f_locals / vars() / tracing	_PyList_AppendTakeRef
Modules/_pickle.c:5138	pickle Pickler memo copy	_PyDict_SetItem_Take2
Modules/_io/textio.c:2337 & 2343	io.TextIOWrapper readline (multi-chunk)	_PyList_AppendTakeRef
Modules/_asynciomodule.c:1213	asyncio future done-callback compaction	_PyList_AppendTakeRef

Update: Most cases already have an open PR, for new ones please open a separate issue.

Linked PRs

• gh-150942: Speed up re.findall and re.sub/subn result building #150943
• gh-150942: Speed up json.loads array and object decoding #150945
• gh-150942: Speed up csv.reader row building #150995
• gh-150942: Speed up frame local item collection #151002
• gh-151059: Use PyObject_CallMethodOneArg in datetime call_tzinfo_method #151062
• gh-150942: Speed up ElementTree attribute parsing #151209
• gh-150942: Optimize PicklerMemoProxy.copy() with _PyDict_SetItem_Take2 #151608
• gh-150942: Speed up asyncio Future.remove_done_callback #153013
• gh-150942: Speed up pyexpat attribute parsing #154860
• gh-150942: Speed up BytesIO.readlines() using reference-stealing append #155171
• gh-150942: Speed up IOBase.readlines #155304
• gh-150942: Optimize stringlib split/splitlines with _PyList_AppendTakeRef #155922
• gh-150942: Speed up re.split result building #155940
• gh-150942: Improve performance of csv reader by using _PyList_AppendT… #155957
• gh-150942: Speed up FrameLocalsProxy.values() #156055
• gh-150942: Speed up Py_BuildValue dict construction #156199

[omkar-334]

hi @eendebakpt @StanFromIreland can i work on a module , maybe csv or textio

[eendebakpt]

hi @eendebakpt @StanFromIreland can i work on a module , maybe csv or textio

Sure! Try out, benchmark, test. If benchmarks show significant improvements you can open a PR, if not just report here briefly.

[omkar-334]

hi @eendebakpt , i've opned a PR for csv reader, since that has 1.08x speedup, matches your PR speedups....

however, the benchmarks for io.BytesIO.readlines show that there is not a significant improvements

Benchmark	main	my branch	speedup
bytesio_readlines_1m_8b	15.15 ms	15.29 ms	0.99x
bytesio_readlines_500k_24b	7.66 ms	7.45 ms	1.03x
bytesio_readlines_100k_256b	3.24 ms	3.20 ms	1.01x
bytesio_readlines_2m_0b	16.33 ms	16.54 ms	0.99x
bytesio_readlines_2m_8b	32.58 ms	31.97 ms	1.02x
geomean			1.01x

Benchmark script

"""Micro-benchmark for io.BytesIO.readlines result-list building."""

from __future__ import annotations

import argparse
import io
import json
import statistics
import sys
import time
from pathlib import Path


def make_data(lines: int, payload_size: int) -> bytes:
    payload = b"x" * payload_size
    return b"".join(payload + b"\n" for _ in range(lines))


def time_readlines(data: bytes) -> float:
    t0 = time.perf_counter_ns()
    io.BytesIO(data).readlines()
    return (time.perf_counter_ns() - t0) / 1e9


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("--lines", type=int, default=200_000)
    ap.add_argument("--payload-size", type=int, default=24)
    ap.add_argument("--trials", type=int, default=15)
    ap.add_argument("--warmup", type=int, default=3)
    ap.add_argument("--label", default="run")
    ap.add_argument("--outdir", default="results/bytesio-readlines")
    args = ap.parse_args()

    data = make_data(args.lines, args.payload_size)

    for _ in range(args.warmup):
        time_readlines(data)

    trials = [time_readlines(data) for _ in range(args.trials)]

    outdir = Path(args.outdir)
    outdir.mkdir(parents=True, exist_ok=True)
    summary = {
        "label": args.label,
        "python": sys.executable,
        "lines": args.lines,
        "payload_size": args.payload_size,
        "trials": args.trials,
        "warmup": args.warmup,
        "input_bytes": len(data),
        "seconds": {
            "min": min(trials),
            "median": statistics.median(trials),
            "mean": statistics.fmean(trials),
            "stdev": statistics.pstdev(trials),
            "all": trials,
        },
        "lines_per_sec_at_min": args.lines / min(trials),
    }
    out = outdir / f"bench_{args.label}.json"
    out.write_text(json.dumps(summary, indent=2))

    s = summary["seconds"]
    print(
        f"{args.label}: min={s['min']*1000:.2f}ms median={s['median']*1000:.2f}ms "
        f"mean={s['mean']*1000:.2f}ms stdev={s['stdev']*1000:.2f}ms "
        f"lines/s@min={summary['lines_per_sec_at_min']:.2e}"
    )
    print(f"wrote {out}")
    return 0


if __name__ == "__main__":
    raise SystemExit(main())

do you think i should still make a pr for this

[tangyuan0821]

Happy to try _asynciomodule.c — swap PyList_Append for
_PyList_AppendTakeRef in future_add_done_callback.
I'll benchmark first.