Mixed integer/x87 islands: cross-game measurements

Historical source last updated, 2026-10-12. Runtime and implementation claims below describe that dated source; see current app notes and source for subsequent changes.

Follow-up: island coverage, live disassembly, and an isolated pure/mixed evaluator split.

2026-09-30. Follow-up to the isolated MW3 prototype. This compares the same frozen baseline and candidate, without changing the production interpreter or its defaults.

Method

Dedicated ASCII box bx_4r5uzdwv, restored onto an AMD EPYC-Rome KVM host, 4 vCPU / 8 GB, Node 24.18.1, V8 13.6.233.17-node.50. Initial load was 0.14 / 0.09 / 0.08. This is a different host from the earlier Ryzen MW3 run; absolute times must not be compared across those machines.

Frozen source revision: 837f0a74.

The existing tools/uop-game-ab.js drives fixed input routes with a pinned guest calendar, branch clock and uop tier enabled in both arms. Runs are serial, with separate processes, in ABBA order. Quake II and Moorhuhn 3 receive a second BAAB group to check the first group's apparent effect.

node tools/uop-game-ab.js --no-build --games=h2,h3,q2,mh3 \
  --arms=control,candidate,candidate2,control2 \
  --arm=control=--wasm=baseline.wasm \
  --arm=candidate=--wasm=candidate3.wasm --jobs=1 \
  --extra="--max-seconds=600 --no-threads --x87-fusion" \
  --out=build/mixed-corpus2

Repeat with --games=q2,mh3, --arms=candidate,control,control2,candidate2, and --out=build/mixed-corpus3. The numeric arm suffix repeats the same artifact. --no-threads selects cooperative guest scheduling; guest audio threads still execute.

The primary metric is whole-process user CPU seconds, including startup, JIT compilation, host rendering and cooperative guest threads. These are fixed-route timings, not browser FPS. The harness's --slice-split measures only the initial main-instance run per batch, omitting cooperative thread work and extra main interleaves; it is not an aggregate guest CPU measurement, especially for Heroes III's audio-heavy workload.

Game Route Final batch
Heroes II Enter scenario and scroll adventure map; integer control 2400
Heroes III Enter scenario and scroll adventure map; includes audio thread 5101
Quake II +set vid_ref soft +map demo1; software rendering 1400
Moorhuhn 3 Start round and shoot across field; renderer-heavy with FP audio 2270

The first attempted sweep lacked test/binaries/tlbs/stdole2.tlb and stopped before guest execution. Its timings are excluded entirely. The CLI returned zero even on that startup failure, so validity requires the expected final batch count and a rendered frame, not just exit status. Restoring the shared type-library fixture and checking startup preceded all measurements below.

Results

Positive change means the candidate consumed more CPU. Each row averages two baseline and two candidate launches; repeat groups remain separate so machine drift is visible rather than hidden in a pooled average.

Game / order Baseline CPU s Candidate CPU s Change Baseline repeat spread
Heroes II / ABBA 5.410 5.395 −0.28% 2.59%
Heroes III / ABBA 108.205 108.230 +0.02% 2.06%
Quake II / ABBA 15.230 15.800 +3.74% 2.63%
Quake II / BAAB 14.295 14.255 −0.28% 1.05%
Moorhuhn 3 / ABBA 31.260 30.950 −0.99% 10.75%
Moorhuhn 3 / BAAB 30.090 30.515 +1.41% 0.47%

Spread is the absolute difference between the two baseline launches divided by their mean, not a confidence interval. Two samples per arm per group do not support a precise estimate of a small effect on a virtual machine.

All 24 runs reached their expected final batch. API-call totals agree within each game: Heroes II 695,972; Heroes III 87,011,857; Quake II 77,469; Moorhuhn 3 492,944. No startup error or WASM trap was found. Every final frame, including repeats, is pixel-identical to its game's first baseline. Visual inspection confirms adventure maps for both Heroes games, an in-level Quake II view, and Moorhuhn's shooting field at the round's end (timer 0:00, score 250). This checks final visible output, not full state or audio parity.

No cross-game speedup is established. Heroes II/III are neutral at this resolution. Quake II's initial slowdown did not reproduce. Moorhuhn's first group has substantial drift; its tighter repeat suggests a possible small regression (+1.41%), so it must not be described as an improvement.

These measurements retain the prototype's conservative block-executor restriction and extra per-operation classification. They do not isolate which mechanism accounts for any difference, or count mixed-island coverage in these games. Keep the prototype isolated; before enabling it, investigate Moorhuhn's repeat signal and measure affected hot paths directly. The earlier MW3 projection-kernel gain is still valid, but is not a general game-speed claim.

Raw logs, frames, machine metadata and pixel comparisons are under build/lazy-games/mixed-corpus/ locally. runs/ contains ABBA, repeat/ contains BAAB, and summary.json is produced by the artifact-local summarize.cjs, which rejects missing/truncated routes, errors, differing API totals within a group, or differing final pixels. The original harness's printed candidate/control percentage compares the first pair only; the table above uses both repetitions. Final observed load during the repeat was 1.09 / 0.64 / 0.49 with one benchmark child at a time.