perf: use in-place Arrays.sort for std.sort/std.set/std.uniq

[He-Pin]

Motivation

std.sort uses Scala's sortWith which creates a new sorted copy. For homogeneous arrays (all numbers or all strings), we can use Java's optimized Arrays.sort in-place on a copied array, avoiding the overhead of Scala's generic sorting.

Key Design Decision

Detect homogeneous arrays at sort time and dispatch to Arrays.sort(Double[]) or Arrays.sort(String[]) for maximum performance. Fall back to the generic sort for mixed-type arrays.

Modification

Added homogeneous array detection in SetModule.scala sortArr method. Uses Arrays.sort with primitive specialization for numeric and string arrays.

Benchmark Results

JMH Regression Suite (1 fork, 3 warmup, 1 measurement)

Benchmark	Master	This PR	Change
bench.06	0.367	0.312	-15.0%
setDiff	0.425	similar	noise
setInter	0.372	similar	noise
setUnion	0.676	similar	noise

All other benchmarks within noise margin.

Scala Native Hyperfine (-N -w4 -m20)

Benchmark	sjsonnet master	This PR	jrsonnet	vs master	vs jrsonnet
bench.06 (array sort)	8.2 ± 1.1 ms	7.4 ± 1.0 ms	7.7 ± 9.1 ms	1.10x faster	1.05x faster

Analysis

Arrays.sort uses dual-pivot Quicksort which is highly optimized for primitive arrays. The improvement is most visible on sort-heavy benchmarks. On native, sjsonnet with this PR is now slightly faster than jrsonnet on bench.06.

References

Upstream jit branch exploration at he-pin/sjsonnet@jit

Result

15% improvement on JMH bench.06, 10% on native. sjsonnet is now 1.05x faster than jrsonnet for array sorting, closing a previous 1.36x gap.

[He-Pin]

Closing: Independent verification shows a 12% regression on setUnion (the primary sort benchmark). java.util.Arrays.sort with Comparator lambdas has more overhead per comparison than Scala's sortBy, which extracts keys once via map. The allocation savings don't offset the per-comparison overhead.

Verified with Hyperfine (warmup 5, runs 20, --shell=none):

• setUnion: master 8.0±0.6ms → in-place 8.9±0.8ms = 1.12x slower
• comparison2: neutral (175.2 vs 176.1ms)
• bench.02: neutral (75.2 vs 74.8ms)