perf: specialize comparison bitmap packing for 8-bit inputs

[mhk197]

Summary

Speed up comparisons over i8 and u8 arrays by packing each 64-row group with the existing collect_bool_word helper. The byte representation lets the compiler vectorize the comparison and bitmap construction efficiently.

Dedicated collect_bits_dispatch and collect_zip_bits_dispatch helpers select this path from T::PTYPE, which is constant for each compiled primitive type. The byte-packing loops live separately in collect_bits_narrow and collect_zip_bits_narrow. Wider integer and floating-point types continue to use the original lane collectors, so they avoid the AVX2 regressions from applying byte packing to wider masks.

Changes

• Specialize bitmap collection for 8-bit array/array and array/constant comparisons, including constants on either side.

No benchmark source or public API changes. This primitive optimization is independent of NarrowArray.

Performance

Unchanged existing benchmarks on Apple M5 Max (aarch64), Rust 1.98.0, 8,192 rows.

Existing benchmark	Before	8-bit specialization	Speedup
compare_f32	2.833 µs	2.832 µs	1.00×
compare_float	3.791 µs	3.791 µs	1.00×
compare_int	1.958 µs	1.958 µs	1.00×
compare_int_constant	1.624 µs	1.624 µs	1.00×
compare_int_eq	1.978 µs	1.958 µs	1.01×
compare_int_nullable	2.165 µs	2.166 µs	1.00×
compare_u8	1.874 µs	0.698 µs	2.69×
compare_u64	1.979 µs	1.999 µs	0.99×

[codspeed]

Merging this PR will regress 4 benchmarks

⚠️ Unknown Walltime execution environment detected

Using the Walltime instrument on standard Hosted Runners will lead to inconsistent data.

For the most accurate results, we recommend using CodSpeed Macro Runners: bare-metal machines fine-tuned for performance measurement consistency.

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 6 improved benchmarks
❌ 4 regressed benchmarks
✅ 2198 untouched benchmarks
⏩ 293 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	mul_u64_nonnull_neon	28.9 µs	40.4 µs	-28.63%
❌	Simulation	take_fsl_u32_random[256, 10]	124.4 µs	157.9 µs	-21.22%
❌	WallTime	mul_i64_nonnull_neon	32.9 µs	38.6 µs	-14.97%
❌	WallTime	multiply_shapes_neon[(32768, PerRowPerRow)]	32.8 µs	38.4 µs	-14.53%
⚡	WallTime	compare_u8_avx2	3.5 µs	1.8 µs	+95.03%
⚡	WallTime	compare_u8_neon	3.9 µs	2.2 µs	+77.02%
⚡	WallTime	compare_u8_avx512	2.4 µs	1.7 µs	+39.79%
⚡	Simulation	take_fsl_u32_random[16, 100]	165.8 µs	124.3 µs	+33.43%
⚡	Simulation	take_fsl_nullable_random[16, 100]	187.7 µs	162.9 µs	+15.25%
⚡	WallTime	dict_canonicalize_gt_u8_neon[1000000]	547 µs	488.3 µs	+12.03%

Tip

Investigate this regression by commenting @codspeedbot fix this regression on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing mk/primitive-compare-bitmap (c215b0e) with develop (c3ade49)²}

Footnotes

1. 293 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩

2. No successful run was found on develop (ca2538d) during the generation of this report, so c3ade49 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[connortsui20]

related: #9703

[robert3005]

Let's merge this, I know we will soon replace this with rowfn