Skip to content

feat(jsc): support bounded incremental private bytecode promotion - #7

Closed
steipete wants to merge 5 commits into
openclaw/mainfrom
claude/w154-incremental-bytecode
Closed

steipete wants to merge 5 commits into
openclaw/mainfrom
claude/w154-incremental-bytecode

Conversation

@steipete

@steipete steipete commented Oct 5, 2026 •

Copy link
Copy Markdown

This experimental implementation is not proposed for landing. It passes correctness and budget checks, but both measured versions trade a messaging improvement for a providers regression.

The final version serializes generated function bodies into immediately charged immutable deltas and combines them once when a weak program block needs decoding. Weak executable identities, stale-callback checks, and accounting for live payloads after eviction preserve the existing 256 MiB-per-VM bound. No decoded executable graph is retained by the cache. Public cached-data generation remains separate.

Final native comparison: one quiet AWS c7a.16xlarge, four workers on CPUs 0–7, isolated process state/cache directories, eight observations per comparison in ABBAABBA order (four per arm). The unchanged workload is OpenClaw affb994cf4193ff967b0f792b4634435d3942948. Both Bun arms include the same namespace/stack adapters and the independent cached-data repair in openclaw/bun#118.

Config Baseline wall Candidate wall Candidate / baseline Collection ratio
Providers 126.059 s 127.418 s 1.0108× (95% 1.0023–1.0194) 1.0210× (1.0088–1.0354)
Messaging 56.392 s 54.687 s 0.9698× (0.9570–0.9747) 0.9474× (0.9322–0.9522)

The independent Node 24 comparison puts the candidate at 1.3247× Node wall / 1.5464× collection for providers and 1.0886× / 1.0959× for messaging. Collection values sum overlapping worker/module durations. A prior 64-observation version copied the whole payload after each generated body; it also improved messaging while regressing providers. Those samples remain separate from the final version.

Correctness: 196 JSC configurations, 361 VM tests, 365 surrounding stack/source-map/util/WebKit tests, 47/47 fork-selection results, 273 exact position observations matching Node, lazy-body reuse after GC, update/eviction bounds, public cached-data determinism, and debugger/referrer/import.meta checks pass. Existing VM skips and TODOs are unchanged. The final engine and matching Bun adapter passed https://github.com/openclaw/WebKit/actions/runs/37258784020 on e4dead84626bf5816f934aeef2beb7fe737ac71a; scoped P2 reviews are clean.

The providers regression fails the performance goal for the target set. Closing the prototype with its measurements instead of merging it. The separate JIT-less cached-data fix from @robobun's oven-sh/bun#41769 is merged as Bun oven-sh#118. No artifact publication workflow was dispatched, and no decoded-graph retention work was started.

@steipete steipete closed this Oct 5, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant