[connortsui20]

Tracking Issue: #10701

Stacked on #11004. #11006, #10905 and #10995 build on it. No release gate.

Summary

Old decimal nodes and decoded DecimalByteParts can store values at another width than their dtype's canonical width. They need one conversion rule, since a child precision that misses a value would make later comparison shortcuts incorrect.

Changes

Canonical storage reuses its buffer without a scan, including device buffers. Other storage reads on the host and becomes a Widen over the smallest valid child, or the canonical array when no child is narrower. Encoding shares the same child selection. Known ranges can avoid a scan, and exact extrema are cached on the child so later aggregates reuse that work.

API Changes

Adds WidenArray::from_stored_decimal(parts, known_range, ctx). A supplied range must contain every valid value. Exact ranges become exact child extrema, while inexact ranges remain bounds.

[codspeed]

Merging this PR will regress 2 benchmarks

1 warning

⚠️ 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

⚡ 13 improved benchmarks
❌ 2 regressed benchmarks
✅ 2345 untouched benchmarks

Warning

Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
❌	CPU Simulation	filter_powerlaw_by_random[10000]	348.4 ns	1,626.4 ns	-78.58%
❌	Walltime	mul_u64_nonnull_neon	20.5 µs	39.1 µs	-47.6%
⚡	Walltime	threshold_neon[80]	5.7 µs	3.6 µs	+61.22%
⚡	Walltime	threshold_neon[64]	3.7 µs	2.6 µs	+42.21%
⚡	Walltime	filter_neon[64]	2.4 µs	1.8 µs	+33.92%
⚡	Walltime	filter_neon[80]	3.4 µs	2.6 µs	+28.03%
⚡	CPU Simulation	bench_compare_varbin[(50000, 2)]	162.1 µs	132 µs	+22.76%
⚡	CPU Simulation	bench_compare_varbin[(50000, 64)]	164.2 µs	133.8 µs	+22.73%
⚡	CPU Simulation	bench_compare_sliced_dict_varbinview[(1000, 10000)]	88.8 µs	77.9 µs	+14.05%
⚡	CPU Simulation	bench_compare_sliced_dict_varbinview[(7500, 10000)]	227.5 µs	199.6 µs	+14%
⚡	CPU Simulation	bench_compare_sliced_dict_varbinview[(10000, 10000)]	233.7 µs	205.8 µs	+13.59%
⚡	Walltime	add_constant_shapes_neon[(32768, ConstantPerRow)]	15.9 µs	14.2 µs	+11.92%
⚡	Walltime	add_constant_shapes_neon[(32768, PerRowNullableConstant)]	16.5 µs	14.8 µs	+11.9%
⚡	Walltime	add_constant_shapes_neon[(32768, PerRowConstant)]	15.9 µs	14.3 µs	+10.79%
⚡	Walltime	scalar_subtract_neon	11.3 µs	10.2 µs	+10.55%

Tip

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

---

_{Comparing ct/decimal-widen-stored (b3b006c) with ct/decimal-widen-aggregates (5d992ff)}