Re: Optimization for lower(), upper(), casefold() functions.

12.03.2025 19:55, Alexander Borisov wrote:

[...]

>> A couple questions:
>>
>> * Is there a reason the fast-path for codepoints < 0x80 is in
>> unicode_case.c rather than unicode_case_func.h?
> 
> Yes, this is an important optimization, below are benchmarks that

[...]

I forgot to add the benchmark:

Benchmark

Anatation:
Patch v3j_0001: patch v3j_0001 without any changes.
Patch v3j_0001 + static: patch v3j_0001 with adding static to casemap().
Patch v3j_0001 + static + fast path: patch v3j_0001 with adding static
     to casemap() and fast path for codepoint < 0x80.
v4_0001: v3j_0001 + static + fast path
v4_0001 + v3j_0002: v3j_0002 patch unchanged, but with static for
     casemap() (inherited from v4_0001 patch)
v4_0001 + v4_0002: changed fast path for codepoint < 0x80. Made fast
     return to avoid unnecessary checks.
Without: current version of the code without any patches.

All results are in tps.

* macOS 15.1 (Apple M3 Pro) (Apple clang version 16.0.0)

ASCII:
Repeated characters (700kb) in the range from 0x20 to 0x7E.
Patch v3j_0001: 201.029609
Patch v3j_0001 + static: 247.155438
Patch v3j_0001 + static + fast path: 267.941047
Patch v4_0001 + v3j_0002: 260.737601
Patch v4_0001 + v4_0002: 268.913658
Without: 260.437508

Cyrillic:
Repeated characters (1MB) in the range from 0x0410 to 0x042F.
Patch v3j_0001: 48.130350
Patch v3j_0001 + static: 49.365897
Patch v3j_0001 + static + fast path: 48.891842
Patch v4_0001 + v3j_0002: 88.833061
Patch v4_0001 + v4_0002: 88.329603
Without: 47.869687

Unicode:
A query consisting of all Unicode characters from 0xA0 to 0x2FA1D
(excluding 0xD800..0xDFFF).
Patch v3j_0001: 91.333557
Patch v3j_0001 + static: 92.464786
Patch v3j_0001 + static + fast path: 91.359428
Patch v4_0001 + v3j_0002: 103.390609
Patch v4_0001 + v4_0002: 102.819763
Without: 92.279652


Regards,
Alexander Borisov