#1011 · zig-utils/zig-js

Measured problem

The production representative_json invocation at current main takes 2,216,853 conservative-interior ownership bucket locks out of 4,615,990 total slab-lock acquisitions. Most scanned machine words are ordinary non-heap values, but classifyConservativeInterior currently probes every size class before it can reject them.

Candidate

Maintain a monotonic, ever-issued slab-address envelope in GcCellBacking. A coherent lock-free read can reject an address outside that envelope before any bucket lock. The envelope intentionally never shrinks when tail chunks are trimmed: stale inclusion only falls back to the exact existing lookup, while stale exclusion would be unsafe.

A ReleaseFast prototype preserves the exact checksum, cell/allocation counts, and collection counts while moving:

conservative-interior locks: 2,216,853 → 11,277 (-99.5%)

ownership locks: 2,233,415 → 27,839 (-98.8%)

total slab locks: 4,615,990 → 2,410,414 (-47.8%)

A three-pair exact-parent screen is positive on the sustained rows:

representative_json: 0.9931x instructions, 0.9688x CPU, 0.9700x wall

anti-specialization variant: 0.9982x instructions, 0.9717x CPU

escaped strings: 0.9962x instructions, 0.9804x CPU

shallow stringify: 1.0001x instructions, 0.9737x CPU

reviver/depth controls remain near instruction parity; the 1-job cold controls remain checksum controls below the timing floor

Definition of done

Prove coherent concurrent expansion and no false-negative rejection; chunk trimming may only leave a conservative superset.

Pin the zero-bucket-lock rejection path and the in-range exact fallback in focused tests, including TSan.

Preserve JSON test262 results and required VM/tree agreement.

Preserve deep-input, OOM, no-GIL, and GC relocation/ownership behavior.

Record seven interleaved exact-parent samples with identical checksums and no hidden control regression.

Land as one conventional commit on main, close this issue with raw evidence, and update #473.

Landed and pushed to main as 3abc2fc.

Exact-source attribution

ReleaseFast binary SHA-256: 2952b7c5fd2be7ca63cc2a9c968a20d5e2a220f2a7f8d95bfdfe7ea2a8d8797b

On the exact representative_json invocation, compared with parent 934d306:

checksum: 324952086 → 324952086
GC cells: 2,010,721 → 2,010,721
fresh/reused/freed cells and collection counts: identical
conservative-interior locks: 2,216,853 → 16,863 (-99.24%)
all ownership locks: 2,233,415 → 33,425 (-98.50%)
total slab locks: 4,615,990 → 2,416,000 (-47.66%)

The earlier prototype repeat reached 11,277 conservative locks; address-layout variance changes how many ordinary words fall inside the deliberately broad envelope, but both exact runs remove more than 99% of this lock source.

Seven interleaved exact-parent pairs

Raw files: /private/tmp/zig-js-json-prod.83FAYc/pairs-1011-final-7/

row	candidate / parent instructions	parent/candidate instruction RSD	CPU	wall
representative_json	0.99824x	0.41% / 0.48%	0.99936x	1.00350x
anti-specialization variant	0.99327x	0.68% / 0.40%	0.98919x	0.98813x
reviver source	0.98940x	0.64% / 0.61%	1.00808x	1.01233x
escaped strings	0.99540x	0.13% / 0.17%	0.98601x	0.98421x
stringify depth 4096	0.99956x	0.30% / 0.33%	0.99467x	0.99628x
stringify shallow 4096	1.00034x	0.29% / 0.29%	1.00740x	1.00977x
cold plain 4096 control	1.00166x	0.95% / 5.25%	unscored below timing floor	unscored below timing floor
cold Unicode 4096 control	1.00133x	0.22% / 0.42%	unscored below timing floor	unscored below timing floor

All checksums are exact and all Darwin thermal readings are nominal → nominal. The scored sustained rows stay within the control gate; short cold rows remain semantic/instruction controls rather than throughput claims.

Correctness and safety

GC cell backing: 31/31 Debug
coherent envelope expansion: TSan pass
conservative moving-nursery/native-stack integration: 5/5
JSON/OOM/moving-GC units: 34/34
shared JSON no-GIL callers under TSan: 4/4
test262 test/built-ins/JSON: 165 pass, 0 fail
required tree/VM agreement: 165 agree, 0 diverge
50,000-level object and array depth probes: OK true in tree and VM
peak compile/link footprint: 2,142 MB or less for all final gates

The envelope is monotonic across trimming, its writer is serialized by the existing backing inner lock, and readers use a sequence-protected coherent min/max snapshot. Out-of-range values take no size-class lock; in-range values still use the existing exact chunk indexes and publication bitmaps.