gh-125196: Add a free list to PyUnicodeWriter

[vstinner]

• Issue: Use the public PyUnicodeWriter API #125196

[vstinner]

Benchmark:

importpyperfrunner=pyperf.Runner() runner.bench_func('repr([])', repr, []) runner.bench_func('repr([1,2,3])', repr, [1, 2, 3]) list3= ['abcdef']*10runner.bench_func('repr(list3)', repr, list3) list4= ['abcdef']*50runner.bench_func('repr(list4)', repr, list4)

Results, Python built with clang -flto=thin, CPU isolation, benchmark run with --rigorous:

+----------------+---------+-----------------------+ | Benchmark | main | freelist | +================+=========+=======================+ | repr([]) | 61.9 ns | 56.3 ns: 1.10x faster | -5.6 ns +----------------+---------+-----------------------+ | repr([1,2,3]) | 189 ns | 179 ns: 1.06x faster | -10 ns +----------------+---------+-----------------------+ | repr(list3) | 873 ns | 859 ns: 1.02x faster | -14 ns +----------------+---------+-----------------------+ | repr(list4) | 3.12 us | 3.02 us: 1.03x faster | -0.1 us +----------------+---------+-----------------------+ | Geometric mean | (ref) | 1.05x faster | +----------------+---------+-----------------------+ 

[vstinner]

@pitrou: That's one way to reduce the cost of the PyUnicodeWriter_Create() memory allocation.

[vstinner]

cc @serhiy-storchaka

I started with a free list of 1 item. We may enlarge it later for nested function calls, like repr() of a list which contains lists (which may contain lists as well, etc.).

[vstinner]

Another benchmark on repr(tuple) from gh-125242:

+----------------+---------+-----------------------+ | Benchmark | change | freelist | +================+=========+=======================+ | repr(()) | 55.9 ns | 57.9 ns: 1.03x slower | +2 ns +----------------+---------+-----------------------+ | repr(('abc',)) | 239 ns | 232 ns: 1.03x faster | -7 ns +----------------+---------+-----------------------+ | repr((1,2,3)) | 189 ns | 182 ns: 1.04x faster | -7 ns +----------------+---------+-----------------------+ | repr(tuple4) | 870 ns | 846 ns: 1.03x faster | -24 ns +----------------+---------+-----------------------+ | repr(tuple5) | 3.02 us | 3.00 us: 1.01x faster | -20 ns +----------------+---------+-----------------------+ | Geometric mean | (ref) | 1.01x faster | +----------------+---------+-----------------------+