Optimize default set operations

[He-Pin]

Summary

Optimize default-key std.setUnion, std.setInter, and std.setDiff by adding a dedicated default keyF=id merge path.

The new path keeps the current forced key at each merge cursor, compares numbers/strings directly, and writes into a pre-sized Array[Eval] before trimming. Custom keyF keeps the previous implementation, so function-call semantics are unchanged.

This is intentionally narrower than jrsonnet's whole set implementation: the goal is a JVM/JIT-friendly hot path without the large per-type method explosion that did not improve results enough to justify the code size.

JMH

Command:

./mill -i bench.runJmh sjsonnet.bench.RegressionBenchmark.main \
  -p 'path=bench/resources/sjsonnet_suite/setUnion.jsonnet,bench/resources/sjsonnet_suite/setInter.jsonnet,bench/resources/sjsonnet_suite/setDiff.jsonnet' \
  -wi 3 -i 5 -w 1s -r 2s -f 1 -tu ms

Environment:

• JMH 1.37
• JVM: Zulu/OpenJDK 21.0.10 aarch64
• VM options from environment: --enable-native-access=ALL-UNNAMED -Xmx4G -XX:+UseG1GC

Benchmark	master	this PR	delta
setUnion.jsonnet	0.594 ± 0.039 ms/op	0.500 ± 0.036 ms/op	15.8% faster
setInter.jsonnet	0.354 ± 0.024 ms/op	0.334 ± 0.015 ms/op	5.6% faster
setDiff.jsonnet	0.416 ± 0.053 ms/op	0.365 ± 0.011 ms/op	12.3% faster

JMH GC profiler

Command:

./mill -i bench.runJmh sjsonnet.bench.RegressionBenchmark.main \
  -p 'path=bench/resources/sjsonnet_suite/setUnion.jsonnet,bench/resources/sjsonnet_suite/setInter.jsonnet,bench/resources/sjsonnet_suite/setDiff.jsonnet' \
  -wi 2 -i 3 -w 1s -r 1s -f 1 -tu ms -prof gc

Normalized allocation is effectively unchanged/slightly lower:

Benchmark	master alloc.norm	this PR alloc.norm
setUnion.jsonnet	1,512,845.660 B/op	1,510,378.717 B/op
setInter.jsonnet	1,268,889.087 B/op	1,268,595.813 B/op
setDiff.jsonnet	1,329,900.452 B/op	1,327,755.458 B/op

Native hyperfine

Command shape:

hyperfine --shell=none --warmup 20 --runs 100 \
  '<master native> -J 1 -o /dev/null bench/resources/sjsonnet_suite/setUnion.jsonnet' \
  '<this PR native> -J 1 -o /dev/null bench/resources/sjsonnet_suite/setUnion.jsonnet' \
  '<jrsonnet release> -o /dev/null bench/resources/sjsonnet_suite/setUnion.jsonnet' \
  # repeated for setInter/setDiff

Benchmark	master native	this PR native	jrsonnet release
setUnion.jsonnet	4.7 ± 0.3 ms	4.4 ± 0.1 ms	2.8 ± 0.1 ms
setInter.jsonnet	4.2 ± 0.2 ms	4.2 ± 0.3 ms	2.1 ± 0.1 ms
setDiff.jsonnet	4.2 ± 0.1 ms	4.2 ± 0.2 ms	2.4 ± 0.1 ms

Native CLI remains dominated by startup/whole-program costs on these small inputs. jrsonnet is still about 1.9-2.2x faster for the native CLI set cases.

Correctness / validation

./mill -i 'sjsonnet.jvm[3.3.7].test.testOnly' \
  sjsonnet.StdWithKeyFTests \
  sjsonnet.StdSetUnionTests \
  sjsonnet.StdSetDiffTests \
  sjsonnet.StdLibOfficialCompatibilityTests \
  sjsonnet.PreserveOrderTests

./mill -i 'sjsonnet.jvm[3.3.7]'.test
./mill -i 'sjsonnet.js[3.3.7]'.compile 'sjsonnet.native[3.3.7]'.compile '__.checkFormat'
./mill -i 'sjsonnet.native[3.3.7].nativeLink'