perf: pack primitive comparison results in fixed-size chunks

[mhk197]

Summary

Primitive comparisons currently pack predicate results with a shift-and-OR reduction over 64 rows. In the ARM assembly inspected for this implementation, this expands narrow comparison masks into wider integer lanes and makes bitmap construction expensive.

Use the existing collect_bool_word packer for fixed 64-element chunks in primitive array/array and array/constant comparisons. Full chunks let the compiler prove safe indexing is in bounds and vectorize the predicates before byte-to-bit packing. Partial words use the existing scalar tail path.

Changes

• Add private bitmap collectors in the primitive comparison module
• Add 396 parameterized cases covering all 11 primitive types, all six comparison operators, empty inputs and bitmap boundaries, nulls, and constants on either side.

This is independent of NarrowArray and changes no benchmark source or public API.

Performance

Existing comparison benchmarks, unchanged, on Apple M5 Max (aarch64), Rust 1.98.0, 8,192 rows.

Median of three run medians per benchmark, 200 samples per run.

Existing benchmark	Before	After	Speedup
compare_f32	4.249 µs	3.083 µs	1.38×
compare_float	5.499 µs	5.165 µs	1.06×
compare_int	3.083 µs	2.832 µs	1.09×
compare_int_constant	2.416 µs	2.082 µs	1.16×
compare_int_eq	3.124 µs	2.791 µs	1.12×
compare_int_nullable	3.374 µs	3.082 µs	1.09×
compare_u8	2.707 µs	1.040 µs	2.60×
compare_u64	3.124 µs	2.791 µs	1.12×

[codspeed]

Merging this PR will degrade performance by 6.7%

⚠️ 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

⚡ 13 improved benchmarks
❌ 15 regressed benchmarks
✅ 2225 untouched benchmarks
⏩ 218 skipped benchmarks¹

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	WallTime	compare_int_constant_avx2	3 µs	6.4 µs	-52.63%
❌	WallTime	compare_int_avx2	3.5 µs	6.3 µs	-44.54%
❌	WallTime	compare_int_nullable_avx2	3.9 µs	6.7 µs	-41.53%
❌	WallTime	arrow_checked_add_u32_neon[16384]	12.3 µs	20.3 µs	-39.52%
❌	WallTime	compare_u64_avx2	4 µs	6.3 µs	-37.03%
❌	WallTime	compare_float_avx2	5.1 µs	8 µs	-36.02%
❌	WallTime	compare_int_eq_avx2	3.5 µs	5.4 µs	-35.47%
❌	Simulation	decode_primitives[f32, (1000, 512)]	42 µs	64.3 µs	-34.68%
❌	WallTime	lt_i64_nullable_avx2	2.9 µs	4.2 µs	-31.34%
❌	WallTime	mul_u64_nonnull_neon	15.4 µs	20.8 µs	-26.12%
❌	WallTime	mul_i64_nonnull_neon	17.3 µs	20.4 µs	-15.05%
❌	WallTime	multiply_shapes_neon[(16384, PerRowPerRow)]	17.3 µs	20.3 µs	-14.65%
❌	WallTime	dbp_assemble_kernel_avx2[(I256, 8192)]	8.5 µs	9.8 µs	-13.33%
❌	WallTime	filtered_owned_i64_avx2[OneNullInEight]	22.3 µs	25.7 µs	-13.24%
❌	WallTime	mul_i32_nonnull_avx512	7.1 µs	7.9 µs	-10.25%
⚡	WallTime	compare_u8_avx2	3.4 µs	1.8 µs	+92.08%
⚡	WallTime	compare_u8_neon	3.9 µs	2.1 µs	+86.39%
⚡	WallTime	compare_f32_avx2	4.6 µs	3.2 µs	+43.82%
⚡	WallTime	compare_u8_avx512	2.4 µs	1.7 µs	+38.72%
⚡	WallTime	compare_f32_avx512	3.7 µs	2.7 µs	+37.49%
...	...	...	...	...	...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

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 (0528bdb) with develop (c6678f3)}

Footnotes

1. 218 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. ↩