gh-129813, PEP 782: Use PyBytesWriter in _socket

[vstinner]

Replace PyBytes_FromStringAndSize(NULL, size) and _PyBytes_Resize() with the new public PyBytesWriter API.

• Issue: [C API] PEP 782: Add PyBytesWriter API #129813

[vstinner]

Micro-benchmark:

import pyperf
import socket

s1, s2 = socket.socketpair()
loops = 100
size = 1

def func():
    s1.sendall(b'x' * (size * loops))
    for _ in range(loops):
        s = s2.recv(size)
        if len(s) != size:
            raise Exception

runner = pyperf.Runner()
runner.bench_func('bench', func, inner_loops=loops)

s1.close()
s2.close()

Result:

• size=1: Mean +- std dev: [ref] 803 ns +- 12 ns -> [pep782] 769 ns +- 12 ns: 1.04x faster.
• size=1_000: Mean +- std dev: [ref] 1.14 us +- 0.02 us -> [pep782] 1.15 us +- 0.03 us: 1.01x slower

[vstinner]

Another micro-benchmark:

import pyperf
import socket

s1, s2 = socket.socketpair()

def func(size):
    s1.sendall(b'x' * size)
    s = s2.recv(size)
    if len(s) != size:
        raise Exception

runner = pyperf.Runner()
for size in (1, 10, 100, 1000, 10_000, 100_000):
    runner.bench_func(f'recv {size:,}B', func, size)

s1.close()
s2.close()

Result:

Benchmark	ref	pep782
recv 1B	2.29 us	2.19 us: 1.05x faster
recv 10B	2.36 us	2.29 us: 1.03x faster
recv 100B	2.41 us	2.34 us: 1.03x faster
recv 1,000B	2.85 us	2.73 us: 1.04x faster
recv 10,000B	4.89 us	4.95 us: 1.01x slower
Geometric mean	(ref)	1.02x faster

Benchmark hidden because not significant (1): recv 100,000B

The benchmark is not reliable :-( If I run it again, I get different result:

Benchmark	ref	pep782
recv 1B	2.22 us	2.17 us: 1.02x faster
recv 100B	2.27 us	2.33 us: 1.03x slower
recv 100,000B	25.0 us	25.2 us: 1.01x slower
Geometric mean	(ref)	1.00x slower

Benchmark hidden because not significant (3): recv 10B, recv 1,000B, recv 10,000B

[encukou]

This introduced some refleaks.

[vstinner]

This introduced some refleaks.

Ooops, thanks for the report! I wrote #139151 to fix it.