Mixed-island coverage and the pure-FP evaluator split

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.

Next experiment: precise FP selectors and operand preparation, separately and stacked.

2026-09-30/10-01. Follow-up to the cross-game timings. The first mixed prototype executes useful integer bridges in both Moorhuhn 3 and MW3, but also adds a classification to hundreds of millions of operations in islands containing no integers. The block executor is disabled in these runs; lost block-executor eligibility cannot explain their timing differences.

Diagnostic method

tools/bench-mixed-census.js compiles isolated diagnostic versions of frozen 837f0a74, first requiring the uninstrumented baseline and candidate to reproduce their previously measured bytes. Both evaluator bodies count absorbed ADD/INC/DEC operations and report once per completed island to a Node preload. Records aggregate by instance, guest block EIP, evaluator, length and integer counts. An additional probe counts integer rejections in both block-executor absorbed-record scans. An exported global records whether the block executor is enabled on each guest instance at exit.

These builds and their timings are diagnostic only. Counts cover whole fixed routes, including startup and cooperative threads, not just gameplay; they count H451 islands, not every FP operation or every fused x87 family. The EIP labels are guest block entries, not necessarily the first instruction of each island. Multiple islands in one block can produce distinct rows.

Runs used the frozen CLI in wa-mixed-census-837f0a74, with its assets and dependencies linked locally, on ARM64 Node 24.21.0. The local host was busy; no timing conclusions are drawn from it. Same routes/branch clock and uop settings as the corpus harness: Moorhuhn 2270 batches, MW3 1150. Both baseline/candidate pairs have identical final pixels and API counts (492,944 / 4,423,912 respectively). MW3 reaches the cockpit. All observed instances have blockExecutorEnabled=0; there are zero rejection events.

The census's positive control copies the original MW3 projection kernel: 100 vertices executed 17 times per arm give 1700 islands each. The candidate counts two ADDs and one DEC per island; the baseline counts no integers. Output/state parity passes for ordinary and special floating inputs.

Actual coverage of the first mixed prototype

Whole route Moorhuhn 3 MW3
Pure-island executions 15,447,322 8,645,768
Operations inside pure islands 408,437,190 95,301,398
Mixed-island executions 4,777,604 490,617
Operations inside mixed islands 60,098,496 8,873,912
Absorbed integer operations 11,166,124 1,422,990

Moorhuhn's busiest affected blocks:

Block EIP Mixed island Executions
0x43030b 2 FP + 2 pointer ADDs 2,864,735
0x432c76 6 FP + 2 ADDs + DEC 1,612,222
0x432d06 112 FP + ADD + DEC 299,334

The live unpacked bytes were captured separately; its packed on-disk EXE is not suitable for these addresses. 0x43030b starts with a 61-operation FP multiply/accumulate sequence, then SUB ESP,4, then:

faddp st(1), st
add   ecx, 0x40
add   ebx, 0x80
fistp dword [esp]

The SUB splits the long pure island from this short mixed island. The result is clamped to signed 16-bit PCM range and stored. 0x432c76 reads two float arrays, computes sum/difference, advances the two pointers and decrements the loop counter. 0x432d06 contains repeated coefficient-weighted FP transforms.

MW3's dominant mixed block remains 0x4fd394: 426,405 executions, each with 17 FP operations, two pointer ADDs and one DEC. Baseline and candidate visit it the same number of times. The next mixed blocks are 0x517b01 (20,952) and 0x5171e7 (19,007), both three FP operations plus ADD/DEC.

Isolated split variant

tools/bench-mixed-split.js, selected by --split-pure on the existing build and parity tools, marks genuinely mixed H451 islands with operand bit 28. The existing fields occupy bits 0..27. The marker is snapshotted at the last FP operation, so discarded trailing integers do not mark a pure island mixed.

Decode once:
  only FP                    -> unmarked
  FP ... integer ... FP      -> mixed marker

Execute once per island:
  unmarked -> original FP evaluator (no per-op integer classification)
  mixed    -> mixed evaluator

This keeps the matcher, integer flag semantics, operation order and conservative block-executor rejection unchanged. It duplicates the evaluator body instead of making pure loops pay its inner check. Module size grows from 1,646,105 baseline / 1,646,427 first candidate to 1,652,336 bytes. The code size and per-island selection are real tradeoffs, to be judged by timing.

Split SHA-256: ac551603e51942e2d3ab78fb7dffb43249d216e206fe8f46f274a918e659f30f. Validation: 1000 randomized sequences, 26,138 operations, fast == generic == unfused including GPRs/flags/FP state/exception traces; copied projection kernel ordinary/special-input parity passes. Production sources/defaults are unchanged.

Timing and artifacts

Clean timing is separate from the census. The resumed dedicated box bx_4r5uzdwv was provisioned on Intel Xeon (Skylake, IBRS, no TSX), 4 vCPU, Node 24.18.1 / V8 13.6.233.17-node.50. This is another host change; only within-sweep comparisons are meaningful. Initial valid-sweep load was 0.2 / 0.5 / 0.7. The snapshot lacked game assets and type libraries; the failed startup sweep is excluded, the fixtures were restored, and a one-batch guest preflight passed before the valid sweep.

Order per game: baseline, original mixed, split, split, original mixed, baseline. Serial processes, same frozen CLI, branch clock/uop/cooperative threads, existing routes. Values below are means of two whole-process user CPU times; they include startup and rendering and are not FPS.

Game Baseline s Original mixed s Split s Split vs baseline Baseline spread
Moorhuhn 3 39.035 39.400 38.915 −0.31% 2.13%
Quake II 18.080 18.490 18.220 +0.77% 0.88%

Split vs original mixed is −1.23% / −1.46% respectively. All 12 runs reached their expected batch counts and produced identical pixels within each game. The same API totals persist (492,944 / 77,469). The changes are small relative to repeat variation, and this sweep does not establish a general speedup. It does not prove that the earlier regression was caused by the inner check.

A separate local uninstrumented split MW3 run reaches batch 1150 / 4,423,912 API calls and matches the earlier baseline/candidate cockpit pixels exactly. Its timing is excluded: local load and diagnostic instrumentation differ.

The copied projection-kernel benchmark on this Skylake host retains a clear benefit: 138.986 → 120.000 ms guest-thread CPU (−13.66%) for 500,000 vertices, twelve warmups and ten ABBA/BAAB groups (20 samples per arm). Same-artifact control medians are 133.962 / 133.987 ms (0.019% difference). Output, GPR/FP/flag parity passes; the diagnostic post-timing census shows 1,000,000 ADD and 500,000 DEC dispatches disappear, with 500,000 islands in both arms. This is baseline versus split, not split versus the original mixed prototype, and is not comparable to absolute times on the earlier Ryzen host. Results are split-kernel-x64.json and split-kernel-control-x64.json. Final box load was 0.38 / 0.50 / 0.66.

The split has passed correctness checks and removes a demonstrated source of redundant work, but remains an experiment. No production default is enabled.

Local artifacts: build/lazy-games/mixed-census/ has diagnostic WASM, JSON, logs, identical-frame evidence, moorhuhn-hot.asm and parity results. build/lazy-games/mixed-split2/ contains the measured uninstrumented split artifacts. mixed-split3/ reproduces their exact bytes after cleaning up source extraction to avoid duplicating adjacent macro declarations. The first split build directory is an unsuccessful builder attempt; it is not a measured artifact. Remote logs/frames are in remote-runs/, with two-sample means and frame comparisons in remote-summary.json.

Reproduction:

node tools/bench-mixed-census.js build build/lazy-games/mw3-mixed3/src \
  build/lazy-games/mw3-mixed3 build/lazy-games/census-new
MIXED_CENSUS_OUT='out/%APP%-%ARM%.json' \
  node --require=/absolute/path/tools/bench-mixed-census.js test/run.js \
  --app=moorhuhn_3 --no-build --wasm=/absolute/path/census-new/candidate.wasm \
  <the existing fixed-route arguments>
node tools/bench-mw3-mixed-build.js 837f0a74 build/lazy-games/split-new \
  build/lazy-games/mw3-mixed3/baseline.wasm --split-pure
X87_FUZZ_CASES=1000 node tools/bench-mw3-mixed-parity.js \
  build/lazy-games/mw3-mixed3 --split-pure

Use the frozen runner for these old modules, with its matching region map and host imports. Census JSON records full runner arguments. MIXED_CENSUS_CODE=1 also saves 512 live guest bytes at up to 32 hot block EIPs after execution.