[robert3005]

TiledArray repeats its child end to end count times. Execution decodes
the child once and appends it per tile via Chunked; scalar_at wraps the
index; validity tiles the child's validity. Reduce rules collapse empty,
single-count, constant-child and nested tiles, and slices within one tile
or over whole tiles stay cheap. It is in-memory only and not registered
for serialization.

[codspeed]

Merging this PR will degrade performance by 0.06%

2 warnings

⚠️ Walltime results may be unreliable

Some benchmarks used the Walltime instrument outside CodSpeed Macro Runners, which can make the results inconsistent.

Learn more about Macro Runners

See affected benchmarks

⚠️ Base and head ran on different runner types

Some benchmarks with significant performance changes were compared across different runner types, so the changes might not come from the code. Run base and head on the same runner type, or use CodSpeed Macro Runners for more stable results.

Learn more about Macro Runners

Open the report in CodSpeed to investigate

⚡ 11 improved benchmarks
❌ 8 regressed benchmarks
✅ 2287 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	Walltime	mul_u64_nonnull_neon	20.3 µs	39.2 µs	-48.25%
❌	CPU Simulation	filter_powerlaw_by_random[10000]	142.2 ns	267.5 ns	-46.83%
❌	CPU Simulation	missing_stats[16]	66.8 µs	85.6 µs	-22.01%
❌	CPU Simulation	bench_compare_varbin[(50000, 2)]	133.2 µs	162.5 µs	-17.99%
❌	CPU Simulation	bench_compare_varbin[(50000, 64)]	135 µs	164.3 µs	-17.85%
❌	CPU Simulation	bench_compare_sliced_dict_varbinview[(7500, 10000)]	200.6 µs	228.4 µs	-12.21%
❌	CPU Simulation	bench_compare_sliced_dict_varbinview[(1000, 10000)]	78.1 µs	88.8 µs	-12.09%
❌	CPU Simulation	bench_compare_sliced_dict_varbinview[(10000, 10000)]	206.7 µs	234.6 µs	-11.87%
⚡	Walltime	threshold_neon[64]	4.4 µs	2.6 µs	+68.8%
⚡	CPU Simulation	canonicalize_sparse_fixed_size_list[(1024, 17, 8)]	427.4 µs	300.6 µs	+42.19%
⚡	Walltime	filter_neon[64]	2.7 µs	2.1 µs	+27.59%
⚡	CPU Simulation	canonicalize_sparse_fixed_size_list[(512, 7, 4)]	461.7 µs	364.9 µs	+26.51%
⚡	Walltime	filter_neon[80]	3.3 µs	2.7 µs	+20.39%
⚡	Walltime	threshold_neon[80]	4.2 µs	3.5 µs	+18.38%
⚡	CPU Simulation	take_filter_primitive_small_uncached_random_mask_random_indices[4096, 10]	24.3 µs	20.8 µs	+16.62%
⚡	CPU Simulation	test_repeated_scan_scalar_at[int_str_float]	481.6 µs	414.5 µs	+16.19%
⚡	Walltime	add_constant_shapes_neon[(32768, ConstantPerRow)]	15.7 µs	14.2 µs	+10.42%
⚡	Walltime	add_constant_shapes_neon[(32768, PerRowConstant)]	15.7 µs	14.2 µs	+10.41%
⚡	Walltime	scalar_subtract_neon	11.3 µs	10.3 µs	+10.15%

Tip

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

---

_{Comparing rk/tiled-array (10fc9ce) with develop (d4d9234)¹}

Footnotes

1. No successful run was found on develop (7e70afc) during the generation of this report, so d4d9234 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[robert3005]

CodSpeed Performance Analysis is red, but the 8 regressions don't come from this PR. Every other check on 10fc9ce is green.

• Code path: none of the regressed benchmarks (dict_compare, filter_bool, zone_map_prune, neon mul_u64) build fixed-size lists or RepeatedArray. The PR only changes the repeated fixed-size-list builder append and constant fixed-size-list canonicalization. The benchmarks that do hit that path improved, as expected: canonicalize_sparse_fixed_size_list is 26–42% faster.

• Local check: base 40e0b1c vs head, divan medians on the same machine:

  Benchmark	Base	Head
  bench_compare_varbin[(50000, 2)]	8.339 µs	8.325 µs
  bench_compare_varbin[(50000, 64)]	8.433 µs	8.439 µs
  filter_powerlaw_by_random[10000]	1.13 µs	1.138 µs
  filter_powerlaw_by_random[100000]	7.299 µs	7.307 µs

• CodSpeed's own warning: base and head ran on different runner types.

There's no code fix to make. The regressions need acknowledging in CodSpeed to clear the check.

---

Generated by Claude Code