gh-156537: Optimize codec.decode by inlining the per-character write in CJK decoders

[brittanyrey]

Every CJK decoder emits output through OUTCHAR(), which called the exported _PyUnicodeWriter_WriteChar(). Since the codecs are shared extension modules, that's a cross boundary call to the exported function for each output character. Swapping to the _PyUnicodeWriter_WriteCharInline() removes the overhead of crossing the boundary speeding up decoding.

The single macro applies to all of the codec modules.

Benchmark	main	PR
decode 1.8MB shift_jis	6.87 ms	2.55 ms: 2.70x faster
decode 1.8MB euc_kr	6.34 ms	2.50 ms: 2.54x faster
decode 1.8MB euc_jp	6.36 ms	2.55 ms: 2.49x faster
decode 1.8MB big5	6.27 ms	2.56 ms: 2.45x faster
decode 1kB euc_jp	3.97 us	1.82 us: 2.18x faster
decode 1.8MB euc_jp kana	3.86 ms	3.07 ms: 1.26x faster

Benchmark script

"""CJK decoder throughput (Modules/cjkcodecs)."""
import pyperf

ASCII = ('abc' * 30 + '\n')
KANA = ('あいう' * 10 + '\n')

def dec(d, e):
    d.decode(e)

if __name__ == "__main__":
    r = pyperf.Runner()
    for enc in ("euc_jp", "shift_jis", "big5", "euc_kr"):
        r.bench_func("decode 1.8MB %s" % enc, dec, (ASCII * 20000).encode(enc), enc)
    r.bench_func("decode 1.8MB euc_jp kana", dec, (KANA * 20000).encode("euc_jp"), "euc_jp")
    r.bench_func("decode 1kB euc_jp", dec, (ASCII * 11).encode("euc_jp"), "euc_jp")

• Issue: Performance: CJK codec decoders pay a cross-module call per character #156537

[corona10]

lgtm

Thank you for doing this.