commit 45060f533d45ada1db15c197918f3318b53df9ef
Author: Haoxiang Fei <feihaoxiang@idea.edu.cn>
Date:
perf: store trie mid levels one index per Int
Packing two 16-bit leaf block indices per Int left a shift and a mask on
the path to the leaf load. In xlang, native General_Category on text was
4.5% slower than main's two-byte read. One index per Int makes it 7%
faster than main on native and keeps the wasm, wasm-gc and JS gains, at
48 KB more mid-level data across all tries. 16-bit leaf values stay
packed.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>