Production-path VM optimization

Issue: #689

The VM now resolves each global managed type ID to its module/type key at admission and stores the reverse mapping in the shared immutable execution image. Managed-reference type lookup uses an indexed read instead of repeatedly searching module offsets. Invalid IDs still fail and module identity is retained across empty modules.

Measurement path

The opt-in guest-benchmark binary compiles the existing interpreter and verifier source modules without cfg(test) register-initialization instrumentation and invokes the existing crate-private untraced Session path. It changes no public API, artifact encoding or exported native ABI. Its manifest inputs are the same generated Guest artifacts used in the earlier Ryzen measurements. It checks ready/checksum output, normal termination, and all eight construction/hot work counters; a traced control must agree with the untraced path.

The measurements below compare the same binary harness and artifacts before/after the type-key optimization, rather than comparing diagnostic traced timing against untraced execution. This is standalone interpreter evidence; it excludes Minecraft lifecycle/scheduling and JVM/native transport costs.

Paired results

Ryzen 9 9950X3D, Linux 7.2.8-zen1-2-zen, Rust 1.99.0 / LLVM 23.1.1. Native release profile uses thin LTO, one codegen unit and debug line tables (CARGO_PROFILE_RELEASE_DEBUG=1). The governor remains powersave, without fixed affinity/frequency; unrelated desktop/agent activity was uncontrolled. No own compilation or profiling overlapped these timing runs.

Three paired rounds each use seven fresh-session samples per case, reversing baseline/candidate order in the second round. Case order also reverses inside each invocation. Each invocation warms up and checks a traced correctness control before timing. Verification, admission, start and teardown are outside timers; final scans/output text copies are included. Medians below combine 21 samples per version. The archive retains per-round medians, min/max and all raw samples.

All 11 cases improve in each paired round. Aggregate median elapsed time decreases by 6.67–8.87%; these local ratios are not portable guarantees. The first object-transformation round had higher baseline times, so its larger 13–15% ratios are not used as the aggregate claim. Every case retains identical fixed/dynamic/maintenance units, block counts, executed/retired instructions and host request/response counts across versions and rounds. There are 462 ordinary timed samples in this comparison.

Loops use a 16 KiB Guest heap for 256 elements × 300 summation rounds; transformations use 256 KiB for 1024 two-Int records × 100 rounds. No fallback heap was used. The optimized image adds eight bytes of immutable mapping payload per declared type, plus one boxed-slice field and native allocation metadata per image. Mutable per-session arenas and Guest heap budgets are unchanged. Admission pays for construction of this additional table; admission elapsed time was not isolated in this experiment.

Comparison TSV, environment, commands and binary hashes, loop samples, transformation samples.

CPU profiles

Separate profiles used perf record -e cycles:u -F 499 --call-graph dwarf and 99 samples of one selected case. These runs are excluded from elapsed-time comparisons. Profiles include verification, admission, warmup/traced control and final termination, so percentages are whole-process sampling evidence rather than exact hot-loop attribution.

Before optimization, ExecutionImage::type_key accounts for 4.48–6.12% of sampled cycles in iterator/fold/object-transform profiles. The fold profile reports 6.12%. Afterward this function is absent from the standalone-symbol report at the 1% threshold: lookup work is inlined as an indexed read, not eliminated entirely. Other measured costs, including register access and virtual dispatch, remain candidates for separate work.

Indexed loop, iterator loop, fold before, transformation, fold after. Raw perf.data and frozen benchmark binaries remain under ignored build/reports/benchmarks/issue-689/. A register-helper inlining trial gave mixed screening results and was not retained.

Reproduction and verification

Generate the unchanged input programs using the artifact-generation commands in the earlier Ryzen report. The opt-in Gradle entrypoint runs the eight production-path loop cases:

JAVA_HOME=/usr/lib/jvm/java-25-openjdk ./gradlew-sandbox-dev-parallel-summary benchmarkGuestProductionLoops

For paired or profiling runs, build and retain an independent binary for each VM revision:

CARGO_TARGET_DIR=.toolchain/build/cargo/guest-benchmark CARGO_PROFILE_RELEASE_DEBUG=1 \
  cargo build --manifest-path host/compukter-vm/Cargo.toml --release --locked --offline \
  --features guest-benchmark --bin guest-benchmark

Invoke the retained binary with ARTIFACT_DIR REPORT_DIR 7 HEAP_BYTES untraced. Use 16384 for transient-allocations or 262144 for collection-reuse; an optional final ID prefix selects a profile case. Match the three round orders and exact arguments in environment.json. Baseline core is VM 8416b99; the benchmark-only VM commit is 7cd0e87; optimized VM is 7fea2f27691bc21ffae3eadac787e039d00abdbc.

Verification passed 519 native crate tests (430 unit tests plus integrations/doc tests), clippy with warnings denied, formatting, build-script tests and the Gradle benchmark entrypoint. Six freshly executed Kotlin-to-VM scenarios passed: function values, generic interfaces, mapNotNull, mutable lists, tasks and exceptions. New native tests cover empty-module type-ID mapping, admitted ID round trips and unknown-ID rejection. Verification summary.

This is focused development verification. No full-checkout/release gate or fresh Minecraft server profile was run. The change was integrated into dev on 2026-10-06 together with the current Runtime ABI 1.13 implementation. Integration passed 613 Rust workspace tests, formatting, Clippy with warnings denied, six Kotlin-to-VM scenarios, runtime-host integration, and FFM/JNI verification. The benchmark executable also passed a traced/untraced correctness smoke check. These integration checks do not replace the dated timing measurements above.