`locate` found an entry's word by multiplication (the MAGIC table) but
still divided `64 / bits` at run time for the shift, on every block read
and write of a section stored in 5, 6 or 7 bits. A division by a value the
compiler does not know costs tens of cycles on x86: `perf annotate` of a
10-bot server on the ThinkPad put 12% of `PalettedContainer::get`, the
server thread's top function, on that `div`. The table now also holds the
bits a word's entries fill, so the shift is `i * bits - word * filled`.
A test checks `locate` against the division for every width and entry.
ThinkPad (i5-7200U), 4M random reads, 5-7 bit palettes: 28.5-36.9 ->
24.6-32.4 ns over 2,000 sections, 6.7-7.1 -> 5.4-5.8 ns within one;
4 and 8 bits (shifts, unchanged) the same.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Ludovic Marie
40c77c02f0