gh-159037: Improve performance of str.join pre-pass

[eendebakpt]

The pre-pass of _PyUnicode_JoinArray() did more work per item than needed. In this PR we now ORs the kinds and ANDs the ASCII flags of the items; the maximum character and use_memcpy are derived from these after the loop. One separator is counted per item and the extra one is subtracted after the loop. Tests for separators and items of mixed kinds are added.

@vstinner

Benchmark	main	PR
sep '', 10 short items	63.6 ns	59.8 ns: 1.06x faster
sep '', 10**3 short items	3.13 us	2.81 us: 1.11x faster
sep '', 10**6 short items	3.12 ms	2.77 ms: 1.13x faster
sep '', 10 long items	92.7 ns	89.8 ns: 1.03x faster
sep '', 10**2 long items	432 ns	398 ns: 1.08x faster
sep '', 10**4 long items	56.4 us	51.8 us: 1.09x faster
sep '.', 10 short items	83.0 ns	80.1 ns: 1.04x faster
sep '.', 10**3 short items	4.79 us	4.39 us: 1.09x faster
sep '.', 10**6 short items	4.58 ms	4.26 ms: 1.07x faster
sep '.', 10 long items	117 ns	114 ns: 1.03x faster
sep '.', 10**2 long items	583 ns	559 ns: 1.04x faster
sep '.', 10**4 long items	60.2 us	57.6 us: 1.05x faster
Geometric mean	(ref)	1.07x faster

Benchmark	main	PR
latin1 sep '', 10 short items	77.9 ns	66.9 ns: 1.17x faster
latin1 sep '', 10**3 short items	3.39 us	2.82 us: 1.20x faster
latin1 sep '.', 10**3 short items	4.98 us	4.16 us: 1.20x faster
latin1 sep '.', 10**2 long items	596 ns	542 ns: 1.10x faster
ucs2 sep '', 10 short items	105 ns	86.9 ns: 1.21x faster
ucs2 sep '', 10**3 short items	5.79 us	5.19 us: 1.11x faster
ucs2 sep '.', 10**3 short items	10.9 us	10.6 us: 1.03x faster
ucs4 sep '', 10**3 short items	5.45 us	4.93 us: 1.10x faster
ucs4 sep '.', 10**3 short items	12.5 us	12.3 us: 1.02x faster
ucs4 sep '.', 10**2 long items	1.56 us	1.50 us: 1.04x faster
latin1 sep, ascii items, 10**3 items	4.72 us	4.69 us: 1.01x faster
ucs2 sep, ascii items, 10 items	125 ns	120 ns: 1.05x faster
mixed ascii+latin1 items, 10**3 items	4.78 us	4.29 us: 1.11x faster
mixed ascii+ucs2 items, 10**3 items	10.5 us	10.2 us: 1.04x faster
mixed latin1+ucs4 items, 10**3 items	6.73 us	5.94 us: 1.13x faster
Geometric mean (34 non-ASCII and ASCII cases)	(ref)	1.08x faster

Callgrind confirms the direction: the join function executes 8-18% fewer instructions per call, mostly from dropping the per-item kind comparison.

Benchmark scripts

First table (from the PR 158693):

import pyperf
runner = pyperf.Runner()
for sep in ('', '.'):
    for list_size in ('10', '10**3', '10**6'):
        runner.timeit(
            f'sep {sep!r}, {list_size} short items',
            setup=f"data=['x']*{list_size}; sep={sep!r}",
            stmt="sep.join(data)")

    for list_size in ('10', '10**2', '10**4'):
        runner.timeit(
            f'sep {sep!r}, {list_size} long items',
            setup=f"data=['x' * 100]*{list_size}; sep={sep!r}",
            stmt="sep.join(data)")

Second table:

import pyperf
runner = pyperf.Runner()
KINDS = {'ascii': 'x', 'latin1': '\xe9', 'ucs2': '€', 'ucs4': '\U0001f600'}
for kname, ch in KINDS.items():
    for sep in ('', '.'):
        for list_size in ('10', '10**3'):
            runner.timeit(
                f'{kname} sep {sep!r}, {list_size} short items',
                setup=f"data=[{ch!r}]*{list_size}; sep={sep!r}",
                stmt="sep.join(data)")
        runner.timeit(
            f'{kname} sep {sep!r}, 10**2 long items',
            setup=f"data=[{ch!r} * 100]*10**2; sep={sep!r}",
            stmt="sep.join(data)")
for name, sep, items in (
        ("latin1 sep, ascii items", '\xe9', "['x']"),
        ("ucs2 sep, ascii items", '€', "['x']"),
        ("mixed ascii+latin1 items", '.', "['x', '\\xe9']"),
        ("mixed ascii+ucs2 items", '.', "['x', '\\u20ac']"),
        ("mixed latin1+ucs4 items", '', "['\\xe9', '\\U0001f600']"),
        ):
    for list_size in ('10', '10**3'):
        runner.timeit(
            f'{name}, {list_size} items',
            setup=f"data=({items}*{list_size})[:{list_size}]; sep={sep!r}",
            stmt="sep.join(data)")

• Issue: Improve performance of str.join #159037

[vstinner]

Interesting optimization! I will try to review it next week.