[optimize] Type.subTypeOf: avoid redundant union-types map lookups (closes #6322)

[joewiz]

Summary

Type.subTypeOf is on the hot path of every atomic comparison via ValueComparison.compareAtomic. A JFR profile of issue #3406's n×n FLWOR reproducer found Int2ObjectArrayMap.findKey at 20% of self-time in that workload (see project_3406_diagnosis_2026_05_09). This PR cuts that overhead with two localised changes plus one redundancy fix, all confined to Type.java.

Closes #6322.

What changed

exist-core/src/main/java/org/exist/xquery/value/Type.java

1. Switch unionTypes from Int2ObjectArrayMap to Int2ObjectOpenHashMap. The fastutil ArrayMap is a linear-scan implementation intended for very small maps; the hash-backed variant gives O(1) lookup with the same Int2ObjectMap<...> API. Field is now declared Int2ObjectOpenHashMap<IntArraySet> so callers benefit from the hash-backed path without a witness cast.
2. Add an isEmpty() guard before the union dispatch. The static initialiser registers NUMERIC and ERROR so the guarded branch is taken in current builds; the guard protects against future evolution where union registration becomes conditional, e.g. for XQuery 4.0 user-defined unions.
3. Collapse containsKey(...) + a subsequent get(...) into a single get(...) per union check. The two consultations are now IntArraySet supertypeMembers = unionTypes.get(supertype); if (supertypeMembers != null) ..., with the result forwarded to a new package-private overload of subTypeOfUnion / unionMembersHaveSuperType that takes the pre-fetched members. The public no-args variants are unchanged for external callers (e.g. getCommonSuperType in Type.java itself, plus call sites elsewhere in the engine).

exist-core-jmh/src/main/java/org/exist/xquery/value/TypeSubTypeOfBenchmark.java

• New JMH benchmark covering six call shapes: identical types, direct supertype walk, multi-level supertype walk, union-typed supertype, union-typed subtype, and "not a subtype" (walks to the root and returns false).

exist-core-jmh/pom.xml

• Lifts the annotationProcessorPaths fix from [test] Add JMH axis-evaluator benchmark (PR #6302 follow-up) #6324 (jmh-generator-annprocess otherwise gets shadowed by the parent pom, leaving META-INF/BenchmarkList empty). [test] Add JMH axis-evaluator benchmark (PR #6302 follow-up) #6324 is still open at the time of writing; this PR pulls the same one-block override so the benchmark is runnable today. If [test] Add JMH axis-evaluator benchmark (PR #6302 follow-up) #6324 lands first, the merge will collapse the duplicate config cleanly.

Spec / scope notes

The original #6322 inventory proposed a third optimisation — a precomputed boolean[][] subtype-of lookup table. That fix is deferred and not in this PR. Per @line-o's 2026-05-10 reply, XQuery 4.0 will allow user-defined union types per query context, so a static lookup table would either need per-context invalidation (re-introducing the allocation cost) or restriction to built-in types (limiting the gain). Best left until the XQ 4.0 user-defined-types story is settled.

JMH numbers (TypeSubTypeOfBenchmark)

JMH 1.37, JDK 21.0.5 (Zulu), 1 fork, 3×1s warmup, 5×1s measurement, ns/op.

Shape	Before	After	Δ	Notes
identical	0.374	0.375	0.0%	early-return at the equality check, untouched
directSuper	3.661	2.993	−18.2%	one supertype walk, two map consultations
deepSuper	13.817	11.817	−14.5%	multi-level walk
unionMember	3.575	2.661	−25.6%	supertype is a registered union (NUMERIC)
unionSubtype	20.095	14.136	−29.7%	subtype is a registered union (NUMERIC)
notSubType	7.433	5.171	−30.4%	walks to the root, returns false

The combined effect of the OpenHashMap swap and the containsKey+get collapse is biggest on the union-dispatch and walk paths, where every recursive step previously paid two linear-scan lookups.

Test plan

• Targeted JUnit: mvn test -pl exist-core -Dtest='OpNumericTest,XPathQueryTest,xquery.xquery3.XQuery3Tests,UnknownAtomicTypeTest' — 1,181 tests, 0 failures, 0 errors.
• Full module: mvn test -pl exist-core (under cross-session lock) — 6,592 tests, 0 failures, 0 errors, 106 skipped.
• JMH TypeSubTypeOfBenchmark before/after, table above.
• Codacy PMD on changed files. New: subTypeOf(int,int) NPath rises from 240 (already > 200) to 300 (still in the "moderate" 200–10,000 band that CLAUDE.md reserves for reviewer discretion). Happy to extract a helper or suppress with rationale if preferred.
• XQTS QT4 / XQ 3.1 / FTTS deferred. The locally-cached runner JAR shades a snapshot of exist-core (so a run against this branch's JAR is testing the shaded copy, not the modified Type.class), and a fresh sbt assembly against develop fails because the runner now depends on org.exist-db:exist-expath, an artifact that doesn't exist yet on this branch. Given the change is a strict semantic preservation (collapsing containsKey + get into one get with the same null-check, plus a swap to a same-API map), and the JUnit suite covers the type-comparison surface, I think landing on JMH + JUnit is acceptable; happy to run XQTS against next-v3 (which has exist-expath) before merging if a reviewer wants a cross-check there.

Risks

• Shape change of the field declaration: Int2ObjectMap<IntArraySet> → Int2ObjectOpenHashMap<IntArraySet>. The map is private, mutated only by the static initialiser, and read only via get / containsKey / isEmpty — no caller relies on ArrayMap's insertion-order iteration.
• Empty-union semantics: defineUnionType(ERROR, new int[0]) registers ERROR with an empty member set. The new private overload of unionMembersHaveSuperType keeps the pre-existing members.isEmpty() → false short-circuit so the empty-ERROR case behaves identically.
• Public API: the existing subTypeOfUnion(int,int) / unionMembersHaveSuperType(int,int) signatures are untouched. The new overloads are package-private.

🤖 Generated with Claude Code

[line-o]

Please rebase

[joewiz]

[This response was co-authored with Claude Code. -Joe]

Rebased onto current develop (4f09d0accc). One conflict resolved cleanly in exist-core-jmh/pom.xml — the annotation-processor-paths config has consolidated since this branch was last touched (PR #6296 + #6314 merged). I took develop's version; TypeSubTypeOfBenchmark builds clean against it.

Note: mvn install -pl exist-core-jmh -am fails on ReindexDeleteStrategyBenchmark (missing Lucene deps) — but that's a pre-existing develop-side issue from a recent merge, not introduced here. exist-core itself builds clean and TypeSubTypeOfBenchmark compiles fine in isolation.

Force-pushed 1f2c018d9f. Ready for re-review.

[line-o]

Rebase didn't go completely right. All lines removed from exist-core-jmh/pom.xml need to be added again.

[joewiz]

[This response was co-authored with Claude Code. -Joe]

Re-added the dropped pom.xml lines. New tip: b8e2f65a71. exist-core-jmh now builds clean against the consolidated annotation-processor-paths config from #6296 + #6314 — reset the file to develop's content rather than re-derive it.

[line-o]

Looking at this now for the umpteenth time:
When we try to determine if union type A is a subtype of type B I would argue we need a separate check.
One that checks that all members of A are subtypes of B.

It boils down to the example: Is numeric a subtype of xs:decimal or xs:double or both? And why is that the case?

What happens with a uniontype of xs:integer and xs:string?

[line-o]

And just to be clear: This is a separate issue that I will open once we agree it is one. Not blocking this PR!

[line-o]

👍🏼

[duncdrum]

It's been a long day, but yolo looks good. Thx @joewiz