Repository navigation
Description
Parent: #531
Related: #532, #478, #461
Problem
After #532 removed the accessor-reader serialization hotspot, the frozen shared transition-publication row scales through four lanes but still loses aggregate throughput at eight lanes. Current shared attribution shows the partitioned 8-lane invocation publishes 8,008 independent transition misses with zero Shape writer yields, two Object property-lock contentions, and no backing or element contention. A post-change collapsed sample instead places Shape.lookupState first among engine leaves.
The ordinary named-data mutation paths resolve the same immutable Shape/name state more than once while one Object property snapshot is already held:
- OrdinarySet checks the receiver state, then setOwnUnlocked repeats it;
- named Delete checks the current slot, then deleteTransition repeats it.
At 8 lanes this adds 16,000 redundant shallow probes to 8,000 exact add/read/delete operations.
Scope
- Carry a proven Shape.LookupState through the locked OrdinarySet and named Delete mutation helpers.
- Preserve one public lookup wrapper for callers that do not already hold a proven state.
- Add opt-in lookup request/comparison attribution if needed to prove the eliminated work without charging production-disabled paths.
- Keep Shape transition keys collision-independent and exact; do not change the skip-list publication protocol.
No-workaround rules
- No benchmark recognition, prepublished keys, skipped descriptor/prototype checks, weakened Proxy or accessor semantics, GIL, larger GC threshold, hidden lifecycle exclusion, or reduced work at higher lane counts.
- Do not modify stacks, queues, jobs, Promise/microtask behavior, or Thread/Worker lifecycle.
- Do not trade OOM atomicity, delete/re-add ordering, stable slots, same-name convergence, or moving-GC safety for fewer probes.
Acceptance
- Exact attribution proves one lookup per logically required state observation and identifies every eliminated duplicate.
- Base, anti-specialization variant, and same-parent contended rows retain frozen checksums and equal timed work.
- Exact-parent 1/2/4/8 evidence reports wall, CPU/instructions, dispersion, allocation, GC, RSS, and remaining synchronization.
- OrdinarySet inherited setters/non-writable descriptors, accessors, Proxies, arrays, module namespaces, typed arrays, delete/re-add order, and OOM rollback remain exact.
- Shape, Object, Array-prototype, and Proxy focused tests plus full units, no-GIL TSan, threadfuzz, moving/parallel GC, and exact-parent test262 pass.
- Any performance claim waits for clean checked-in raw evidence and report.
Activity
chrisbbreuer commentedon Aug 13, 2026
The exact-parent diagnostic rejects this optimization, so no code is landing.
I preserved the candidate runner, rebuilt the exact
d437b5fcparent with the same ReleaseFast toolchain, and ran seven order-balanced pairs for base, anti-specialization variant, and same-parent contended controls at 1/2/4/8 lanes (1,000 jobs/lane). Every pair retained the frozen checksums.The partitioned base candidate/parent median wall ratios were
0.9879 / 1.0006 / 0.9998 / 1.0091; the variant ratios were1.0378 / 0.9876 / 0.9933 / 0.9870. Counter ratios likewise showed no consistent CPU-work reduction. The deliberately contended 8-lane row regressed to1.0450xparent wall time and won only 2/7 pairs.The proposed state threading was semantically sound in focused Shape/Object tests, but it added internal API complexity without a causal performance gain. The patch was fully reverted. This negative result also prevents the sampled
Shape.lookupStateleaf from being mistaken for the remaining >4-lane scaling cause; #531/#478 remain the appropriate parent investigations.