gh-140149: use PyBytesWriter in action_helpers.c's _build_concatenated_bytes; 3x faster bytes concat in the parser

[maurycy]

The issue gh-140149 provides more details.

This effectively makes bytes concatenation about 3x faster in the parser, syntax like:

x = (b'meow')
y = (b'meow' b'cow')

Benchmark

The script:

from __future__ import annotations

import ast
import pyperf


def make_src(n, chunk_len, per_line: int = 64):
    assert n > 0 and chunk_len >= 0 and per_line > 0
    chunk = "b'" + ("x" * chunk_len) + "'"
    parts = [chunk] * n
    lines = ["x = ("]
    while parts:
        group = " ".join(parts[:per_line])
        parts = parts[per_line:]
        lines.append(f"    {group}")
    lines.append(")")
    return "\n".join(lines)


def bench_compile(loops, n, chunk_len, per_line: int = 64):
    src = make_src(n, chunk_len, per_line)
    t0 = pyperf.perf_counter()
    for _ in range(loops):
        compile(src, "<bench>", "exec")
    return pyperf.perf_counter() - t0


def bench_ast_parse(loops, n, chunk_len, per_line: int = 64):
    src = make_src(n, chunk_len, per_line)
    t0 = pyperf.perf_counter()
    for _ in range(loops):
        ast.parse(src, filename="<bench>", mode="exec")
    return pyperf.perf_counter() - t0


def main():
    runner = pyperf.Runner()

    for n in (1, 2, 4, 8, 16, 32, 64, 128, 1024):
        runner.bench_time_func(
            f"compile_bytes_concat_n{n}_chunk1",
            bench_compile,
            n,
            1,
        )
        runner.bench_time_func(
            f"parse_bytes_concat_n{n}_chunk1",
            bench_ast_parse,
            n,
            1,
        )

    for n, chunk in ((256, 4), (4, 128), (4, 256)):
        runner.bench_time_func(
            f"compile_bytes_concat_n{n}_chunk{chunk}",
            bench_compile,
            n,
            chunk,
        )
        runner.bench_time_func(
            f"parse_bytes_concat_n{n}_chunk{chunk}",
            bench_ast_parse,
            n,
            chunk,
        )


if __name__ == "__main__":
    main()

The results (with --rigorous, on 9955759):

Benchmark	main	peg-pybytes-bytes-concat-single-alloc
compile_bytes_concat_n1_chunk1	9.89 us	3.68 us: 2.68x faster
parse_bytes_concat_n1_chunk1	6.97 us	2.55 us: 2.73x faster
compile_bytes_concat_n2_chunk1	10.4 us	3.72 us: 2.79x faster
parse_bytes_concat_n2_chunk1	7.33 us	2.63 us: 2.78x faster
compile_bytes_concat_n4_chunk1	10.8 us	3.90 us: 2.78x faster
parse_bytes_concat_n4_chunk1	7.88 us	2.84 us: 2.77x faster
compile_bytes_concat_n8_chunk1	11.5 us	4.14 us: 2.78x faster
parse_bytes_concat_n8_chunk1	8.51 us	3.04 us: 2.80x faster
compile_bytes_concat_n16_chunk1	13.0 us	4.71 us: 2.76x faster
parse_bytes_concat_n16_chunk1	9.89 us	3.57 us: 2.77x faster
compile_bytes_concat_n32_chunk1	15.4 us	5.59 us: 2.75x faster
parse_bytes_concat_n32_chunk1	12.6 us	4.43 us: 2.85x faster
compile_bytes_concat_n64_chunk1	20.6 us	7.15 us: 2.88x faster
parse_bytes_concat_n64_chunk1	17.6 us	5.99 us: 2.94x faster
compile_bytes_concat_n128_chunk1	30.6 us	10.0 us: 3.05x faster
parse_bytes_concat_n128_chunk1	27.6 us	8.84 us: 3.12x faster
compile_bytes_concat_n1024_chunk1	165 us	48.9 us: 3.38x faster
parse_bytes_concat_n1024_chunk1	162 us	47.8 us: 3.40x faster
compile_bytes_concat_n256_chunk4	60.2 us	18.4 us: 3.27x faster
parse_bytes_concat_n256_chunk4	57.2 us	16.9 us: 3.38x faster
compile_bytes_concat_n4_chunk128	12.7 us	5.11 us: 2.47x faster
parse_bytes_concat_n4_chunk128	9.64 us	3.83 us: 2.51x faster
compile_bytes_concat_n4_chunk256	14.1 us	5.96 us: 2.36x faster
parse_bytes_concat_n4_chunk256	10.9 us	4.57 us: 2.38x faster
Geometric mean	(ref)	2.84x faster

The environment:

% ./python -c "import sysconfig; print(sysconfig.get_config_var('CONFIG_ARGS'))"
'--enable-optimizations' '--with-lto'

sudo ./python -m pyperf system tune ensured.

[maurycy]

cc @vstinner @cmaloney

[pablogsal]

Why not using PyBytesWriter_WriteBytes here?

[vstinner]

PyBytesWriter_WriteBytes() grows the buffer if needed. It's not needed since the code already computes the total size in advance.

[vstinner]

PyBytesWriter_WriteBytes() grows the buffer if needed. It's not needed since the code already computes the total size in advance.

[vstinner]

vstinner added the skip news label

This optimization is good to have, but I don't think that users will notice since the parser is only run once at Python startup. I don't think that it's worth it to document this optimization.

[pablogsal]

vstinner added the skip news label

This optimization is good to have, but I don't think that users will notice since the parser is only run once at Python startup. I don't think that it's worth it to document this optimization.

I concur, but on the other hand it doesn't hurt

[vstinner]

LGTM