gh-161145: Test more bytes simultaneously in find_first_nonascii (#161147) Long ASCII runs now OR eight aligned words and test the high bits once, one branch per 64 bytes (32 on 32-bit) instead of one per word. When a block contains a non-ASCII byte, the existing one-word loop finds it. It is plain C: GCC and Clang compile it to SSE2/NEON loads, other compilers to ORed word loads with the same single branch. Multiple compiler and architecture codegen sample: https://godbolt.org/z/8bM83vx3a pyperf, PGO+LTO, 2 interleaved rounds: | Benchmark | main | this | |---|---:|---:| | (b'a' * 4096).decode() | 165 ns | 108 ns: 1.53x faster | | (b'a' * 200).decode() | 28.9 ns | 26.5 ns: 1.09x faster | | b'a' * 200 + 'é' + ..., .decode() | 60.9 ns | 56.0 ns: 1.09x faster | | b'a' * 136 + 'é' + ..., .decode() | 53.8 ns | 51.3 ns: 1.05x faster | | 8-100 byte decodes, 'é' within 100 | | within 1.2 ns | ```python import pyperf runner = pyperf.Runner() tail = 'é'.encode() + b'a' * 50 for n in (8, 100, 200, 4096): runner.timeit(f"(b'a' * {n}).decode()", "s.decode()", globals={"s": b'a' * n}) for n in (16, 100, 136, 200): runner.timeit(f"(b'a' * {n} + 'é'.encode() + b'a' * 50).decode()", "s.decode()", globals={"s": b'a' * n + tail}) ``` Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Co-authored-by: Victor Stinner <vstinner@python.org>
diff --git a/Misc/NEWS.d/next/Core_and_Builtins/2026-11-01-16-02-42.gh-issue-161145.5vLdrm.rst b/Misc/NEWS.d/next/Core_and_Builtins/2026-11-01-16-02-42.gh-issue-161145.5vLdrm.rst new file mode 100644 index 0000000..5d80962 --- /dev/null +++ b/Misc/NEWS.d/next/Core_and_Builtins/2026-11-01-16-02-42.gh-issue-161145.5vLdrm.rst
@@ -0,0 +1,2 @@ +Speed up UTF-8 decoding of long ASCII runs, such as in :meth:`bytes.decode`, +by up to 1.5x.
diff --git a/Objects/unicodeobject.c b/Objects/unicodeobject.c index da1ba44..0fe765c 100644 --- a/Objects/unicodeobject.c +++ b/Objects/unicodeobject.c
@@ -5063,6 +5063,16 @@ find_first_nonascii(const unsigned char *start, const unsigned char *end) } #endif + // Long ASCII runs: test eight words per branch. + for (; end - p >= 8 * SIZEOF_SIZE_T; p += 8 * SIZEOF_SIZE_T) { + const size_t *w = (const size_t *)p; + if ((w[0] | w[1] | w[2] | w[3] | + w[4] | w[5] | w[6] | w[7]) & ASCII_CHAR_MASK) + { + break; + } + } + const unsigned char *e = end - SIZEOF_SIZE_T; while (p <= e) { size_t u = (*(const size_t *)p) & ASCII_CHAR_MASK;