[Dr-Emann]

The array store vector ops now pass lane masks around as mask16x8<S> instead of turning them into u8 bitmasks at every step.

[Kerollmops]

Thanks for bumping the fearless_simd, I didn't saw that before.

It seems that the CI is unhappy about this change 👀

[Dr-Emann]

Comparing this branch vs main (with and without simd):
M1 Pro (aarch64), nightly 1.101.0 (2026-10-24).

bench	dataset	main	main + simd	stable_simd_compress
and/len	census1881	106.5 µs (1.00×)	59.5 µs (1.79×)	60.3 µs (1.76×)
and/len	census1881_srt	19.2 µs (1.00×)	15.9 µs (1.20×)	16.0 µs (1.19×)
and/len	wikileaks-noquotes	255.0 µs (1.00×)	100.5 µs (2.54×)	101.5 µs (2.51×)
and/ref_ref	census1881	184.2 µs (1.00×)	149.0 µs (1.24×)	147.8 µs (1.25×)
and/ref_ref	census1881_srt	111.0 µs (1.00×)	108.4 µs (1.02×)	114.7 µs (0.97×)
and/ref_ref	wikileaks-noquotes	366.8 µs (1.00×)	243.1 µs (1.51×)	249.7 µs (1.47×)
or/ref_ref	census1881	468.2 µs (1.00×)	534.8 µs (0.88×)	461.7 µs (1.01×)
or/ref_ref	census1881_srt	305.1 µs (1.00×)	383.1 µs (0.80×)	322.6 µs (0.95×)
or/ref_ref	wikileaks-noquotes	552.3 µs (1.00×)	565.4 µs (0.98×)	466.9 µs (1.18×)
sub/ref_ref	census1881	333.6 µs (1.00×)	294.0 µs (1.13×)	298.8 µs (1.12×)
sub/ref_ref	census1881_srt	214.5 µs (1.00×)	210.2 µs (1.02×)	218.0 µs (0.98×)
sub/ref_ref	wikileaks-noquotes	461.4 µs (1.00×)	319.6 µs (1.44×)	325.0 µs (1.42×)
xor/ref_ref	census1881	475.9 µs (1.00×)	510.1 µs (0.93×)	471.3 µs (1.01×)
xor/ref_ref	census1881_srt	310.7 µs (1.00×)	316.6 µs (0.98×)	317.1 µs (0.98×)
xor/ref_ref	wikileaks-noquotes	547.0 µs (1.00×)	543.6 µs (1.01×)	489.5 µs (1.12×)

Geometric Means

bench	main	main + simd	stable_simd_compress
and/len	1.00×	1.76×	1.74×
and/ref_ref	1.00×	1.24×	1.21×
or/ref_ref	1.00×	0.88×	1.04×
sub/ref_ref	1.00×	1.19×	1.16×
xor/ref_ref	1.00×	0.97×	1.03×
all	1.00×	1.17×	1.21×

Results varied 2% between passes for the median benchmark and 6.4% at the 90th percentile, so differences of a few percent are probably noise.

• SIMD helps and/sub are faster with either SIMD version, up to 2.5× on wikileaks-noquotes and/len. On census1881_srt neither version gains much, because its containers hold about 72 elements on average.
• or/xor: main's simd feature makes or slower than without SIMD (0.88×). stable_simd_compress is at least as fast as plain main on or/xor everywhere except census1881_srt or, where it's 0.95×.
• Overall: stable_simd_compress is 1.21× faster than main without SIMD, against 1.17× for main + simd.

CI's fixed

[Kerollmops]

Hey @Dr-Emann 👋

Thank you very much for the performance measurements. Performances are impressive and it's faster than the SIMD implementation we had. I'm ready to merge all of your PR's in one go.

Thank you very much and be ready 🌵