Perf: outperform JavaScriptCore across the verified benchmark matrix

[chrisbbreuer]

Goal

Reach and sustain higher zig-js throughput than the system JavaScriptCore on every workload and directly comparable execution mode in the repository benchmark matrix, without weakening semantics, checksums, GC safety, or no-GIL correctness.

Completed workstreams

• benchmark symmetry and statistical rigor (Bench: make zig-js and JSC comparison symmetric and statistically robust #45)
• arithmetic and property-access throughput (Perf: beat JSC on arithmetic and property-access kernels #46)
• array allocation, growth, traversal, and GC throughput (Perf: beat JSC on array allocation, growth, traversal, and GC #47)
• recursive and function-call throughput (Perf: beat JSC on recursive and general function-call throughput #48)
• baseline native-code tier and symbolized attribution (Perf: add a baseline native-code tier for hot bytecode functions #52; broader coverage in JIT: lower guarded property regions into the native tier #132 and JIT: lower verified packed-array regions into the native tier #133)
• independent-context parallel throughput (Perf: beat JSC on arithmetic and property-access kernels #46, Perf: beat JSC on array allocation, growth, traversal, and GC #47, Perf: beat JSC on recursive and general function-call throughput #48, Bench: add bounded object-allocation churn vs JSC #62, Perf: split rare object state to reach the 256-byte GC slab #120, Perf: fit ordinary objects in the 128-byte GC slab #131)
• shared-realm no-GIL object-churn scaling (Perf: remove no-GIL scaling bottlenecks in the comparison workloads #49, Perf: remove shared-realm object-churn GC scaling collapse #97)
• controlled regression history and generated README scorecards (Perf: add benchmark regression gates and historical evidence #50)
• independent object-churn finalization and precise-nursery throughput (Perf: outperform JavaScriptCore across the verified benchmark matrix #44)

Accepted completion baseline

The July 29, 2026 matrix measures clean runtime commit 14d94ec and is published by 018517d with all 1,540 raw samples:

• Direct warmed context: 10/10 wins and 2.32x geometric-mean throughput.
• Eight independent steady contexts: 10/10 wins, 2.53x throughput, and 5.35x scaling versus JSC 4.96x.
• Eight independent cold lifecycles: 10/10 wins, 2.52x throughput, and 5.67x scaling versus JSC 4.97x.
• Shared realm: every workload remains above 1.0x scaling at every lane; eight-lane geometric scaling is 4.84x.

Independent object churn now wins every directly comparable row:

• steady 1/2/4/8 lanes: 106.180/116.422/145.601/207.214 ms versus JSC 134.892/139.490/156.091/208.272 ms;
• cold 1/2/4/8 lanes: 109.688/117.079/135.536/216.272 ms versus JSC 137.150/143.667/155.324/220.921 ms.

Evidence: report, raw samples, and README scorecard.

Completion contract

• Each engine executes identical workload and invocation bytes for directly compared rows.
• Single-context and independent-context parallel modes compare equivalent programming models.
• zig-js shared-realm results are reported separately unless a semantically equivalent public JSC mode is available.
• Every timed row is long enough and sampled enough to report medians plus dispersion.
• On the documented reference host, zig-js wins every direct and independent-context row.
• Shared-realm no-GIL scaling has no workload below 1.0x at any reported lane count (Perf: remove shared-realm object-churn GC scaling collapse #97).
• Accepted performance wins retain focused semantic, GC, no-GIL, full-unit, and TSan coverage.
• The README and dated raw report were generated from a clean tracked worktree and reference the exact measured commit.
• Controlled like-for-like history and generated README scorecard publication are complete (Perf: add benchmark regression gates and historical evidence #50).

Performance claims are scoped to the recorded workload matrix and host; they are not universal claims about arbitrary JavaScript applications.

[chrisbbreuer]

Performance program progress: #45 fairness protocol slice landed (82207c2/e70eeff7/6fb3b289), and #51 delivered the first #48 call-path optimization in eda34f5. The fairer single-context matrix now reports JSC leads of 3.92x arithmetic, 4.31x properties, 11.08x arrays, and 8.15x Fibonacci. These are the remaining targets, not a completion claim; #44–#50 stay open.

[chrisbbreuer]

Baseline native-tier foundation is now live under #52 through 2a88b8a: W^X/MAP_JIT code memory, ARM64 emitter and stable ABI, race-safe hotness/publication, context-owned teardown, and the first real bytecode-to-native constant-function path. Unsupported chunks cache fallback and the 88-group quick comparison matrix remains checksum-clean. This is infrastructure progress, not a replacement performance baseline; numeric/local/control lowering is next.

[chrisbbreuer]

Opened #59 to track the remaining symmetric cold-lifecycle gap separately from warmed execution and shared scaling: #59.

The quick matrix after c1bd949 passes all modes/checksums. Warmed Fibonacci now wins isolated and shared/parallel comparisons, while cold startup remains a real 3x to 8x deficit on most quick rows. The cold timer boundary stays unchanged; #59 owns reducing actual context/global/parser/compiler/teardown cost before the full publication report.

[chrisbbreuer]

Cold comparison progress: 1c720b36 tiers numeric candidates on their only cold invocation, and eb449b6d replaces the independent runner page allocator with a documented process-wide production allocator. Targeted full-work cold endpoints now beat JSC at both 1 and 8 lanes for all four workloads. Full-matrix rerun/publication remains pending after the remaining shared-scaling audit.

[chrisbbreuer]

Shared property gap closed in 3c4e0e43: the synchronized property workload now measures ~30.8 ms at one lane and ~44.6 ms at eight lanes, below recent JSC independent-context wall times. The complete quick matrix already showed zig-js winning every symmetric direct/steady/cold row; after this commit the separate shared panel also beats JSC endpoint wall times for the property workload. Final clean full matrix and report publication are next.