Perf: remove shared-realm object-churn GC scaling collapse

[chrisbbreuer]

Parent: #62
Related: #44, #49, #116, #118, #119, #120, #131

The accepted July 17 structured-stack matrix remains the latest complete zig-js/JavaScriptCore publication:

• measured zig-js: 01fcf42c
• report and raw 1,540-sample evidence
• shared 1/2/4/8 object-churn medians: 190.969 / 293.039 / 683.870 / 1,651.879 ms
• shared scaling: 1.00x / 1.30x / 1.12x / 0.92x

The accepted amortized-publication A/B improved the then-current contended rows without changing the lone-worker path. The issue remains open because the strict 1.0x eight-lane scaling target is not yet met.

Tracked follow-up

• Perf: attribute eight-lane object churn to cooperative GC phases #426 — exact cooperative-GC phase attribution: sweep is 87.4% of the eight-lane pause.
• Perf: reclaim dead nursery backing in whole slab runs #427 — whole-run backing reclamation rejected and fully reverted after its exact A/B.

Acceptance

• Profile the exact shared object-churn row with collector, safepoint, allocation, backing, and lock evidence.
• Preserve object prototypes, initialized properties, indexed replacement, displaced reads, barriers, OOM behavior, exact checkpoints, and checksums.
• Materially improve the formerly sub-0.65x eight-lane result.
• Keep every shared lane at or above 1.0x scaling with repeated low-dispersion evidence.
• Avoid direct/independent-path regressions by retaining their allocation policy.
• Pass dependency normal+TSan, full zig-js units, focused TSan, and exact-parent A/B for retained changes.

Completion evidence — July 29, 2026

The final clean-tree publication measures runtime e690a866 with the steady shared harness at a7cf92c7, published by e4187c0f:

• shared object-churn scaling: 1.26x / 1.16x / 1.20x at 2/4/8 lanes;
• dispersion: 6.26% / 11.31% / 2.82% RSD at 2/4/8 lanes;
• exact-parent reserve A/B: 59,200 → 1,600 publication batches, 4.83x faster at eight lanes, identical checksums, 1.001x RSS;
• validation: zig-gc normal/ReleaseSafe/TSan, 1,460/1,460 zig-js units, focused TSan, benchmark publication gates, link check, and docs build.

Evidence: final report, 1,540 raw samples, and causal reserve A/B. This closes the sole remaining child of #49 and leaves #44 only the directly comparable independent object-churn rows.

Latest diagnosis

The #426 phase profile rules out rendezvous and tracing as the eight-lane collector cost. The #427 exact A/B then rules out backing bitmap/free-slot transitions alone: sweep improved only 0.7% while wall time regressed 0.9%, so both prototypes were reverted. Do not retry the rejected wait, threshold, unpublication, owned-enumeration, pressure-accounting, or whole-run-backing designs without new evidence. README scores remain tied to complete matrices only.

[chrisbbreuer]

Fresh exact-path sample on fdb32d8 (ReleaseFast, shared object_churn, 8 lanes) confirms the first concrete bottleneck. Across the eight worker stacks, top-of-stack samples were dominated by managed-pointer classification and allocation metadata: Heap.headerForPayload 9,523; Heap.create 8,390; __ulock_wait2 2,842; GcCellBacking.allocFn 495; Object.prototype lookup 362. The one-lane shared median is already about 4.8x the direct median because the fixed-shape allocation loop is deliberately disabled whenever parallel IC synchronization is active, forcing per-op VM dispatch, property classification, literal initialization, and generic barriers. I am testing a parallel-safe version of that existing general quick path: seqlock-only literal-shape reads, property/element lock snapshots, exact managed barriers, and generic store fallback for observable/exotic races.

[chrisbbreuer]

The VM-side slice is now published as #98 in b8a4973, with refreshed evidence in 6b27f4d.\n\nFull saved shared-realm object-churn medians improved:\n- 1 lane: 986.746 ms -> 488.792 ms\n- 8 lanes: 28,321.516 ms -> 18,472.129 ms\n\nThat is substantial absolute progress, but mode-local eight-lane scaling is now 0.21x because the one-lane path improved more. The quick path is no longer the primary blocker; the next slice remains open here and should target Heap.create / header lookup / backing allocation and collector serialization observed in the profile.

[chrisbbreuer]

Batched publication slice landed and is published.

Implementation:

• zig-gc: 65441fe, 3183d08, 179ffe2 under zig-gc#15
• zig-js consumer: 84d2f9f4
• calibrated full-matrix harness: ce652645, 3d790a60 under Bench: restore full-matrix timing-floor margin #99
• report/README/raw evidence: 3cc2a58c

Validation:

• Focused semantic/GC and TSan tests pass; the real four-Thread test proves >=2,048 object cells use batch publication.
• Full zig-js suite: 761/761, zero leaks. Comparison harness: 7/7.
• Exact-parent direct A/B: 195.410 -> 192.974 ms (-1.25%, 5/7 pair wins), exact checksum.
• Exact-parent shared 8-lane A/B: 902.841 -> 393.507 ms (-56.42%, 7/7 pair wins), exact checksum.

Validated publication (1,540 samples, commit 3d790a60):

• shared 1 lane: 475.705 ms, 1.00x
• shared 2 lanes: 959.634 ms, 0.99x
• shared 4 lanes: 4,867.850 ms, 0.39x
• shared 8 lanes: 12,368.464 ms, 0.31x

Versus the previous publication, the 8-lane median falls 33.0% (18,472.129 -> 12,368.464 ms) and scaling improves 48% (0.21x -> 0.31x). The direct row remains the sole JSC loss at 0.59x zig-js/JSC throughput. Post-change sampling moves the dominant cost away from per-cell allocation-lock acquisition toward header classification, batch/default initialization, and backing allocation.

Keeping #97 open: acceptance requires every shared lane above 1.0x, and 4/8 lanes remain below it.

Report: https://github.com/zig-utils/zig-js/blob/main/docs/.data/benchmark-comparison-2026-07-14.md
Raw: https://github.com/zig-utils/zig-js/blob/main/docs/.data/benchmark-comparison-2026-07-14.tsv

[chrisbbreuer]

Final validation update: zig-gc unit + 20x TSan is green at https://github.com/zig-utils/zig-gc/actions/runs/29364720539, and zig-js main CI is green on attempt 2 at https://github.com/zig-utils/zig-js/actions/runs/29364341824.

All implementation/validation gates for this batch-publication slice passed. #97 remains open solely because the performance acceptance target (>1x shared scaling at every lane) is not yet met.

[chrisbbreuer]

Fresh full publication from clean measured commit c98a36c4, published by e4b72407:

• report: https://github.com/zig-utils/zig-js/blob/main/docs/.data/benchmark-comparison-2026-07-14.md
• all 1,540 raw samples: https://github.com/zig-utils/zig-js/blob/main/docs/.data/benchmark-comparison-2026-07-14.tsv
• direct: zig-js wins 9/10 rows; 2.56x geometric-mean throughput lead; object churn 197.469 ms vs JSC 119.400 ms (0.60x);
• 8 warmed independent contexts: 2.63x geometric-mean lead, 5.16x zig-js scaling vs 5.10x JSC; object churn 714.208 ms vs 169.960 ms, improved from 844.149 ms;
• 8 cold lifecycles: 2.63x geometric-mean lead, 5.13x zig-js scaling vs 5.01x JSC; object churn improved from 641.918 ms to 570.335 ms;
• shared-realm geometric-mean scaling: 3.89x; object churn improved from 12,368.464 ms to 11,089.300 ms (10.3%), but its faster one-lane baseline leaves scaling at 0.29x.

The exact-parent A/B for #101 measured 12.2% faster shared8 and 3.4% faster direct with exact checksums. #101 is complete; object churn remains the sole direct JSC loss and shared lanes above 1x remains unmet, so the parent performance issues stay open.