gh-133968: Add fast path to PyUnicodeWriter_WriteStr()

[vstinner]

Don't call PyObject_Str() if the input type is str.

• Issue: Using the public PyUnicodeWriter C API made the json module slower #133968

[vstinner]

Microbenchmark:

Mean +- std dev: [ref] 70.5 ns +- 1.1 ns -> [change] 58.3 ns +- 3.8 ns: 1.21x faster

from _testcapi import PyUnicodeWriter
import pyperf

range_100 = range(100)

def bench_write_str():
    writer = PyUnicodeWriter(0)
    for _ in range_100:
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")
        writer.write_str("true")

runner = pyperf.Runner()
runner.bench_func('write_str', bench_write_str, inner_loops=1_000)

[vstinner]

JSON benchmark: #133832 (comment)

Benchmark	ref	change
encode 100 booleans	9.52 us	7.18 us: 1.33x faster
encode 100 integers	13.9 us	11.6 us: 1.20x faster
encode 100 floats	25.1 us	20.6 us: 1.22x faster
encode 100 "ascii" strings	17.3 us	13.2 us: 1.31x faster
encode ascii string len=100	913 ns	902 ns: 1.01x faster
encode escaped string len=128	1.11 us	1.10 us: 1.01x faster
encode Unicode string len=100	1.09 us	1.07 us: 1.02x faster
encode 1000 booleans	59.6 us	38.9 us: 1.53x faster
encode 1000 integers	104 us	82.7 us: 1.26x faster
encode 1000 floats	210 us	166 us: 1.27x faster
encode 1000 "ascii" strings	132 us	93.2 us: 1.42x faster
encode ascii string len=1000	3.48 us	3.49 us: 1.00x slower
encode escaped string len=896	4.12 us	4.11 us: 1.00x faster
encode Unicode string len=1000	4.90 us	4.91 us: 1.00x slower
encode 10000 booleans	553 us	343 us: 1.61x faster
encode 10000 integers	1.00 ms	805 us: 1.25x faster
encode 10000 floats	2.07 ms	1.62 ms: 1.28x faster
encode 10000 "ascii" strings	1.27 ms	868 us: 1.46x faster
encode ascii string len=10000	28.4 us	28.5 us: 1.00x slower
encode escaped string len=9984	38.5 us	38.6 us: 1.00x slower
encode Unicode string len=10000	42.3 us	42.4 us: 1.00x slower
Geometric mean	(ref)	1.18x faster

Encoding booleans is now up to 1.61x faster which is quite appealing!

[miss-islington-app]

Thanks @vstinner for the PR 🌮🎉.. I'm working now to backport this PR to: 3.14.
🐍🍒⛏🤖

[bedevere-app]

GH-133971 is a backport of this pull request to the 3.14 branch.

[vstinner]

When I wrote PyUnicodeWriter_WriteStr(), I skipped this fast path since PyObject_Str() already has a fast path. But it seems like adding one in PyUnicodeWriter_WriteStr() makes a big difference on microbenchmarks!