Speed up JS script evaluation

[rubensworks]

Optimizes the JavaScript script evaluation pipeline, based on a new benchmark and JFR profiling. Each optimization is a separate commit with its own measurement.

Benchmarking

The existing BenchmarkValueTranslators only covers value translation in isolation, so this PR first adds BenchmarkScriptEvaluation, which covers the paths scripts actually go through at runtime: context creation, script instantiation, reading a script member value, calling a JS function as an Integrated Dynamics operator, calling an Integrated Dynamics operator from JS, and unwrapping values that were translated to Graal before. Both benchmarks run from test (./gradlew benchmark benchmarkScriptEvaluation).

The benchmark harness also had to be made trustworthy first: it now warms up before measuring, measures with System.nanoTime, and reports the fastest of several rounds. The previous millisecond-precision single-shot measurements had ~50% run-to-run variance, enough to hide or invent any of the changes below.

Profiling with JFR showed that translation cost is dominated by HostToGuestRootNode.execute and the thread-local context enter/leave around it — i.e. the number of Value API calls crossing the host boundary, not the Java-side work. Every optimization below reduces that number.

Commits

Lazily populate idContext.ops in script contexts — every created context eagerly translated all 276 global interact operators, which took longer than the rest of script instantiation combined, while many scripts use only a handful of them or none. ops is now a self-replacing lazy getter, so after the first access it is a plain data property again and repeated idContext.ops.x accesses stay as fast as before. Exposing ops as a host proxy object instead would have made every access cross the host boundary, which measured ~9% slower per operator call.

Don't overwrite the exports binding when translating an NBT end tag — a correctness fix rather than an optimization: this path evaluated exports = { 'nbt_end': true }, which assigns to the global exports binding and silently discarded whatever the script itself had exported. It now uses a proxy object, which also removes a JS parse+eval from that path.

Unwrap Graal proxies with instanceof instead of ClassCastException — the translators cast the proxy and caught the ClassCastException to detect a mismatch, so translating any proxy threw and caught up to seven exceptions before reaching its own translator. Proxy-unwrapNbt 2.67us → 1.10us (-59%).

Dispatch object value translators on their member key — each of the six object translators materialized the value's member keys to compare against its own single key. Translators now report that key (IValueTranslator#getGraalValueMemberKey, defaulting to null), so the registry materializes the set once. FromGraal-nbt 10.63us → 7.56us (-29%), FromGraal-item 5.10us → 3.73us (-27%). This one regresses Proxy-unwrapItem 0.59us → 1.17us, since the object translators used to detect their own proxies before looking at member keys — the next commit more than makes up for it.

Resolve value translators for round-tripped Graal proxies directly — proxies wrapping an Integrated Dynamics value now report their value type through IValueProxy, so their translator is found in two calls instead of a scan. Proxy-unwrapItem 1.28us → 0.54us (-58%), Operator-callJsFromId-item 1.71us → 1.03us (-40%), Proxy-unwrapNbt 0.93us → 0.56us (-39%).

Net result

Median of 3 alternating runs of the whole branch against its base:

benchmark	before	after	change
Script-instantiate	474us	206us	-56%
Proxy-unwrapNbt	2.10us	0.54us	-74%
FromGraal-nbt	11.6us	8.07us	-31%
Context-createPopulated	449us	319us	-29%
FromGraal-item	5.35us	4.03us	-25%
Operator-callJsFromId-item	1.05us	0.91us	-13%
Proxy-unwrapItem	0.62us	0.56us	-10%
Script-instantiate-useOps	423us	448us	+6%

Instantiating a script that does use idContext.ops is the one regression: it now pays for the lazy getter on top of building the ops object. That is a one-time ~25us cost per instantiation, against ~270us saved for scripts that don't touch ops.

Caveats on the numbers, so they aren't read as more precise than they are: absolute timings shift by ~20% between measurement sessions on this machine, so only the large relative changes above should be taken as solid. The sub-microsecond rows (Operator-callJsFromId-*, Script-memberValue, Operator-roundtrip-int) have measured anywhere between -5% and -25% across sessions; they improve, but I would not put a specific figure on them. The primitive FromGraal-*/ToGraal-* benchmarks run below ~100ns per operation and move by more than ±100% between runs of identical code, so no conclusions are drawn from those at all.

Remaining cost in FromGraal-item is mostly ValueObjectTypeItemStack.deserialize running the vanilla ItemStack codec, which is outside this repo.

Testing

• ./gradlew build passes (50 unit tests, including two new ones covering the lazy ops resolution and the exports fix).
• ./gradlew runGameTestServer passes (21/21).
• Every commit compiles on its own, so the history stays bisectable.

🤖 Generated with Claude Code

https://claude.ai/code/session_014MtPuLF94baQNqedrqSY63

[coveralls]

Coverage Report for CI Build 33327925285

Coverage increased (+0.4%) to 46.243%

Details

• Coverage increased (+0.4%) from the base build.
• Patch coverage: No coverable lines changed in this PR.
• 44 coverage regressions across 8 files.

Uncovered Changes

No uncovered changes found.

Coverage Regressions

44 previously-covered lines in 8 files lost coverage.

File	Lines Losing Coverage	Coverage
org/cyclops/integratedscripting/evaluate/translation/translator/ValueTranslatorOperator.java	13	59.65%
org/cyclops/integratedscripting/evaluate/translation/translator/ValueTranslatorObjectAdapter.java	9	65.08%
org/cyclops/integratedscripting/evaluate/ScriptHelpers.java	5	88.16%
org/cyclops/integratedscripting/evaluate/translation/translator/NbtCompoundTagProxyObject.java	5	64.71%
org/cyclops/integratedscripting/evaluate/translation/translator/ValueObjectProxyObject.java	5	79.49%
org/cyclops/integratedscripting/evaluate/translation/ValueTranslatorRegistry.java	3	89.16%
org/cyclops/integratedscripting/evaluate/translation/translator/OperatorProxyExecutable.java	2	86.67%
org/cyclops/integratedscripting/evaluate/translation/translator/ValueTranslatorNbt.java	2	92.96%

---

Coverage Stats

Relevant Lines:	2574
Covered Lines:	1309
Line Coverage:	50.85%
Relevant Branches:	860
Covered Branches:	279
Branch Coverage:	32.44%
Branches in Coverage %:	Yes
Coverage Strength:	22.74 hits per line

---

💛 - Coveralls