quickjs-jit

This is Longbridge's JIT-enabled distribution of rquickjs. The published package is named quickjs-jit, while its Rust library name remains rquickjs for source compatibility:

rquickjs = { package = "quickjs-jit", version = "=0.12.2" }
rquickjs-jit = { package = "quickjs-jit-runtime", version = "=0.12.2", features = ["compiler"] }

The distribution and JIT runtime must use the same patch version. The JIT ABI includes internal QuickJS structure layouts, so mixing runtime, core, or sys crate patch versions may compile but will be rejected during JIT startup. quickjs-jit-runtime 0.12.3 has not been published yet. Consumers testing the 0.12.3 source must pin the distribution and runtime to the same Git revision and use Cargo source patches from that revision for both quickjs-jit-core and quickjs-jit-sys. Do not combine the published 0.12.2 runtime with 0.12.3 core or sys crates.

This library is a high level bindings of the QuickJS-NG JavaScript engine, a fork of the QuickJS Javascript engine. Its goal is to be an easy to use, and safe wrapper similar to the rlua library.

QuickJS is a small and embeddable JavaScript engine. It supports the ES2020 specification including modules, asynchronous generators, proxies and BigInt. It optionally supports mathematical extensions such as big decimal floating point numbers (BigDecimal), big binary floating point numbers (BigFloat) and operator overloading.

Main features of QuickJS

Small and easily embeddable: just a few C files, no external dependency, 210 KiB of x86 code for a simple hello world program.
Fast interpreter with very low startup time: runs the 75000 tests of the ECMAScript Test Suite in about 100 seconds on a single core of a desktop PC. The complete life cycle of a runtime instance completes in less than 300 microseconds.
Almost complete ES2020 support including modules, asynchronous generators and full Annex B support (legacy web compatibility).
Passes nearly 100% of the ECMAScript Test Suite tests when selecting the ES2020 features. A summary is available at Test262 Report.
Can compile JavaScript sources to executables with no external dependency.
Garbage collection using reference counting (to reduce memory usage and have deterministic behavior) with cycle removal.
Mathematical extensions: BigDecimal, BigFloat, operator overloading, bigint mode, math mode.
Command line interpreter with contextual colorization implemented in JavaScript.
Small built-in standard library with C library wrappers.

Features provided by this crate

Full integration with async Rust
- The ES6 Promises can be handled as Rust futures and vice versa
- Easy integration with almost any async runtime or executor
Flexible data conversion between Rust and JS
- Many widely used Rust types can be converted to JS and vice versa
Support for user-defined allocators
- The Runtime can be created using custom allocator
- Using Rust's global allocator is also fully supported
Support for user-defined module resolvers and loaders which also can be combined to get more flexible solution for concrete case
Support for bundling JS modules as a bytecode using embed macro
Support for deferred calling of JS functions
Full support of ES6 classes
- Rust data types can be represented as JS classes
- Data fields can be accessed via object properties
- Both static and instance members is also supported
- The properties can be defined with getters and setters
- Support for constant static properties
- Support for holding references to JS objects (Data type which holds refs should implement Trace trait to get garbage collector works properly)
- Support for extending defined classes by JS

Experimental JIT performance

Every JIT optimization must include benchmarks against QuickJS, Bun, and quickjs-jit, with the complete per-scenario comparison in this README. Bun default is the external performance target; the previous JIT revision tracks regressions, and the interpreter establishes whether native execution is profitable. See the repository rules and next optimization targets.

Complete 35-scenario Linux x86_64 matrix, measured 2026-09-30: 0.12.11 gpui-shell tiering fixes.

The measured candidate is revision 8fa2a48 (branch fix/shell-tiering-regression). It fixes the 0.12.10 regression in real gpui-shell renders: exhausted Tier2 profitability retries no longer return long-running Baselines to the interpreter, the funded Tier2 trial is compiled first, retries follow their backoff without host polling, and automatic tiering again keeps the opcodes Tier 1 gained in roadmap P2 (closures, generic operators, exception regions, for-in/for-of) interpreted until profitability can measure the interpreter. The previous-JIT baseline is the v0.12.10 runtime (889f7c1, the same frozen binary as the historical roadmap matrix below). QuickJS is the candidate binary with the JIT detached, so the profitability comparison uses the same QuickJS-NG fd0a0210b7be00957751871e7e01b8291268fc29 runtime and benchmark harness. Bun 1.4.0 runs the same JavaScript from a file wrapper with default flags.

Relative to the previous automatic JIT, 8 scenarios are significantly faster, 11 are significantly slower, and 16 are statistically tied. The large losses are the deliberate P2 deferral: exceptions-sync runs at 0.21x and for-of-array at 0.36x the speed of 0.12.10, back to same-version QuickJS speed (0.98x and 1.00x). The other nine slower scenarios are between 0.5% and 3.6% slower; for strings-regexp, methods-dynamic and calls-closures the unchanged interpreter moved by a similar amount. Relative to same-version QuickJS, 23 are faster, 5 are slower (collections, strings-json, json-codec, map-set-bigint, exceptions-promises-async), and 7 are tied. Relative to Bun, all 35 are slower: the closest are fibonacci-recursive (0.73x Bun speed), generic-call-entry (0.63x), float64-dense (0.62x) and quickjs-bitops (0.58x); the widest gaps are methods-dynamic (0.012x), calls-recursion-closures (0.020x), arrays-typed (0.024x) and objects-polymorphic (0.026x). The tables retain every losing, fallback-only and inconclusive scenario.

How to read the table:

QuickJS / Bun / quickjs-jit: latency in ms per ten workload calls (lower is better) — the median over processes of each process's upper-median of 16 timed batches.
vs QuickJS / vs Bun: quickjs-jit speed as a multiple of that engine's speed (that engine's latency / quickjs-jit latency, paired geometric mean). 49.2x (faster) means quickjs-jit runs 49.2 times as fast; 0.53x (slower) means it runs at 0.53 times that engine's speed. faster / slower means the 95% confidence interval lies entirely above / below 1.00x; tied means it crosses 1.00x.
QuickJS is the candidate binary with the JIT detached; quickjs-jit uses production automatic tiering. Ratios use one decimal place from 1.1x and two decimals below it.

Workload	QuickJS	Bun	quickjs‑jit	vs QuickJS	vs Bun
scalar‑control‑flow	2.13	0.0238	0.0434	49.2x (faster)	0.53x (slower)
scalar‑expressions	2.16	0.0201	0.0456	47.4x (faster)	0.43x (slower)
host‑compute	2.16	0.0287	0.0512	42.3x (faster)	0.55x (slower)
mixed‑quotes	2.77	0.0588	0.648	4.3x (faster)	0.091x (slower)
quickjs‑int‑arith	5.8	0.0477	0.319	18.3x (faster)	0.15x (slower)
quickjs‑bitops	1.19	0.0159	0.0281	42.6x (faster)	0.58x (slower)
quickjs‑fibonacci	0.947	0.0132	0.377	2.5x (faster)	0.035x (slower)
numeric	0.539	0.0115	0.0211	26.6x (faster)	0.56x (slower)
scalar‑loop	0.539	0.0114	0.0211	26.5x (faster)	0.56x (slower)
call‑heavy	1.29	0.0207	0.0388	33.1x (faster)	0.51x (slower)
generic‑call‑entry	0.972	0.017	0.026	36.4x (faster)	0.63x (slower)
generic‑call‑fallback	1.29	0.0199	0.205	6.3x (faster)	0.099x (slower)
property‑heavy	1.07	0.0095	0.0281	37.8x (faster)	0.44x (slower)
fibonacci‑iterative	35.6	0.181	0.762	45.8x (faster)	0.22x (slower)
fibonacci‑recursive	9.91	0.266	0.361	27.1x (faster)	0.73x (slower)
collections	1.58	0.0399	1.6	0.98x (slower)	0.026x (slower)
for‑of‑array	0.459	0.0423	0.464	1.00x (tied)	0.067x (slower)
strings‑json	1.91	0.0713	2	0.96x (slower)	0.038x (slower)
calls‑closures	3.24	0.0938	3.21	1.00x (tied)	0.03x (slower)
adversarial	0.937	0.0361	0.951	1.00x (tied)	0.038x (slower)
float64‑dense	2.3	0.0655	0.106	22.2x (faster)	0.62x (slower)
strings‑regexp	19.7	1.03	20.3	1.02x (tied)	0.053x (slower)
arrays‑typed	4.17	0.0454	1.9	2.2x (faster)	0.024x (slower)
packed‑array‑traversal	0.827	0.0151	0.0303	26.7x (faster)	0.48x (slower)
int32array‑traversal	1.03	0.015	0.0279	37.1x (faster)	0.53x (slower)
float64array‑traversal	1.52	0.0083	0.0568	26.9x (faster)	0.15x (slower)
typed‑convert‑traversal	1.51	0.00812	0.0569	26.9x (faster)	0.14x (slower)
objects‑polymorphic	6.26	0.122	4.79	1.3x (faster)	0.026x (slower)
property‑polymorphic	1.86	0.0234	0.565	3.3x (faster)	0.041x (slower)
calls‑recursion‑closures	6.47	0.126	6.61	1.00x (tied)	0.02x (slower)
json‑codec	78.6	7.77	81.5	0.96x (slower)	0.098x (slower)
map‑set‑bigint	17.7	0.838	18.1	0.99x (slower)	0.047x (slower)
exceptions‑promises‑async	2.35	0.207	2.42	0.98x (slower)	0.086x (slower)
methods‑dynamic	13.6	0.151	13.4	1.02x (tied)	0.012x (slower)
exceptions‑sync	2.93	0.366	2.94	0.98x (tied)	0.12x (slower)

95% confidence intervals and statistically tied results

Each cell is quickjs-jit's speed ratio followed by its 95% confidence interval. vs previous compares with the v0.12.10 automatic JIT. Exact values, including the interpreter-control intervals, are in the machine-readable comparisons.

Workload	vs QuickJS	vs Bun	vs previous
scalar‑control‑flow	49.2x [49.1, 49.2]	0.53x [0.49, 0.55]	1.01x [1.00, 1.01]
scalar‑expressions	47.4x [47.4, 47.4]	0.43x [0.40, 0.44]	0.99x [0.99, 0.99]
host‑compute	42.3x [42.0, 42.5]	0.55x [0.54, 0.56]	1.04x [1.01, 1.07]
mixed‑quotes	4.3x [4.3, 4.3]	0.091x [0.091, 0.092]	0.99x [0.99, 0.99]
quickjs‑int‑arith	18.3x [18.2, 18.5]	0.15x [0.15, 0.15]	1.00x [1.00, 1.00]
quickjs‑bitops	42.6x [42.5, 42.7]	0.58x [0.57, 0.60]	1.00x [0.99, 1.00]
quickjs‑fibonacci	2.5x [2.5, 2.5]	0.035x [0.035, 0.036]	0.99x [0.99, 1.00]
numeric	26.6x [26.0, 27.4]	0.56x [0.52, 0.59]	1.04x [1.01, 1.07]
scalar‑loop	26.5x [25.8, 27.3]	0.56x [0.55, 0.58]	1.03x [1.00, 1.06]
call‑heavy	33.1x [33.0, 33.2]	0.51x [0.47, 0.53]	1.00x [1.00, 1.01]
generic‑call‑entry	36.4x [35.8, 37.0]	0.63x [0.57, 0.67]	0.98x [0.96, 1.00]
generic‑call‑fallback	6.3x [6.3, 6.3]	0.099x [0.088, 0.11]	1.00x [1.00, 1.01]
property‑heavy	37.8x [37.5, 38.1]	0.44x [0.37, 0.53]	1.00x [0.99, 1.01]
fibonacci‑iterative	45.8x [43.8, 47.0]	0.22x [0.21, 0.23]	1.01x [0.97, 1.03]
fibonacci‑recursive	27.1x [26.9, 27.4]	0.73x [0.72, 0.74]	1.00x [0.99, 1.02]
collections	0.98x [0.98, 0.99]	0.026x [0.025, 0.027]	1.01x [0.99, 1.04]
for‑of‑array	1.00x [0.95, 1.05]	0.067x [0.056, 0.078]	0.36x [0.35, 0.38]
strings‑json	0.96x [0.95, 0.96]	0.038x [0.037, 0.039]	1.00x [0.99, 1.01]
calls‑closures	1.00x [0.98, 1.01]	0.03x [0.029, 0.032]	0.99x [0.97, 1.00]
adversarial	1.00x [0.97, 1.03]	0.038x [0.037, 0.038]	1.01x [0.99, 1.03]
float64‑dense	22.2x [21.5, 22.9]	0.62x [0.61, 0.63]	0.97x [0.95, 0.99]
strings‑regexp	1.02x [0.98, 1.08]	0.053x [0.051, 0.056]	0.96x [0.92, 0.99]
arrays‑typed	2.2x [2.2, 2.2]	0.024x [0.024, 0.025]	1.00x [0.99, 1.00]
packed‑array‑traversal	26.7x [26.3, 27.0]	0.48x [0.47, 0.50]	0.98x [0.97, 1.00]
int32array‑traversal	37.1x [36.6, 37.6]	0.53x [0.52, 0.54]	1.02x [1.01, 1.03]
float64array‑traversal	26.9x [26.7, 27.0]	0.15x [0.15, 0.15]	1.01x [1.01, 1.02]
typed‑convert‑traversal	26.9x [26.4, 27.9]	0.14x [0.14, 0.15]	1.01x [0.99, 1.02]
objects‑polymorphic	1.3x [1.2, 1.4]	0.026x [0.025, 0.027]	0.98x [0.93, 1.01]
property‑polymorphic	3.3x [3.3, 3.3]	0.041x [0.041, 0.041]	1.00x [1.00, 1.01]
calls‑recursion‑closures	1.00x [0.98, 1.03]	0.02x [0.019, 0.021]	1.00x [0.99, 1.03]
json‑codec	0.96x [0.95, 0.97]	0.098x [0.095, 0.10]	1.01x [0.99, 1.04]
map‑set‑bigint	0.99x [0.98, 1.00]	0.047x [0.047, 0.048]	0.97x [0.97, 0.98]
exceptions‑promises‑async	0.98x [0.97, 0.98]	0.086x [0.085, 0.086]	1.02x [1.00, 1.06]
methods‑dynamic	1.02x [1.00, 1.04]	0.012x [0.011, 0.013]	0.98x [0.96, 0.99]
exceptions‑sync	0.98x [0.94, 1.00]	0.12x [0.12, 0.12]	0.21x [0.20, 0.22]

Statistically tied: quickjs-int-arith vs previous JIT (between 0.4% slower and 0.2% faster); quickjs-bitops vs previous JIT (between 0.7% slower and 0.2% faster); call-heavy vs previous JIT (between 0.0% slower and 0.6% faster); generic-call-entry vs previous JIT (between 3.5% slower and 0.3% faster); property-heavy vs previous JIT (between 0.7% slower and 1.4% faster); fibonacci-iterative vs previous JIT (between 3.5% slower and 3.2% faster); fibonacci-recursive vs previous JIT (between 1.2% slower and 1.8% faster); collections vs previous JIT (between 1.2% slower and 3.7% faster); for-of-array vs QuickJS (between 4.8% slower and 4.7% faster); strings-json vs previous JIT (between 1.1% slower and 0.8% faster); calls-closures vs QuickJS (between 1.8% slower and 0.8% faster); adversarial vs QuickJS (between 2.7% slower and 2.9% faster); adversarial vs previous JIT (between 1.4% slower and 3.1% faster); strings-regexp vs QuickJS (between 2.5% slower and 8.4% faster); arrays-typed vs previous JIT (between 0.7% slower and 0.5% faster); typed-convert-traversal vs previous JIT (between 0.9% slower and 2.2% faster); objects-polymorphic vs previous JIT (between 6.6% slower and 1.0% faster); property-polymorphic vs previous JIT (between 0.2% slower and 0.9% faster); calls-recursion-closures vs QuickJS (between 1.8% slower and 2.6% faster); calls-recursion-closures vs previous JIT (between 1.3% slower and 2.6% faster); json-codec vs previous JIT (between 1.3% slower and 4.0% faster); methods-dynamic vs QuickJS (between 0.4% slower and 3.7% faster); exceptions-sync vs QuickJS (between 5.7% slower and 0.3% faster).

Statistically tied interpreter controls: scalar-control-flow (between 0.2% slower and 0.0% faster); host-compute (between 0.1% slower and 0.6% faster); mixed-quotes (between 0.3% slower and 0.2% faster); quickjs-int-arith (between 1.5% slower and 0.2% faster); quickjs-bitops (between 0.2% slower and 0.0% faster); quickjs-fibonacci (between 0.2% slower and 3.1% faster); numeric (between 0.3% slower and 0.3% faster); scalar-loop (between 0.4% slower and 0.0% faster); call-heavy (between 0.4% slower and 0.5% faster); generic-call-entry (between 0.2% slower and 0.1% faster); property-heavy (between 0.2% slower and 0.1% faster); fibonacci-iterative (between 0.5% slower and 0.1% faster); fibonacci-recursive (between 0.1% slower and 0.5% faster); collections (between 0.3% slower and 0.8% faster); for-of-array (between 6.2% slower and 1.0% faster); strings-json (between 1.8% slower and 0.2% faster); float64-dense (between 2.9% slower and 3.1% faster); arrays-typed (between 1.4% slower and 0.1% faster); packed-array-traversal (between 0.2% slower and 0.2% faster); int32array-traversal (between 1.5% slower and 1.1% faster); typed-convert-traversal (between 5.9% slower and 1.3% faster); objects-polymorphic (between 5.1% slower and 7.5% faster); calls-recursion-closures (between 3.5% slower and 0.3% faster); json-codec (between 1.3% slower and 1.2% faster); exceptions-promises-async (between 0.5% slower and 0.7% faster); methods-dynamic (between 3.7% slower and 0.4% faster); exceptions-sync (between 0.7% slower and 0.4% faster).

gpui-shell host regression. The matrix protocol polls the JIT after every batch and warms 640 calls before timing, so it cannot observe tier-up latency or host-call overhead in an embedder that never polls. The regression and its fix are measured with gpui-shell's own emit_one_jit_acceptance_sample (gpui-kit 2c2feb76): each fresh process renders 64 warmup frames, then the mean of 200 measured renders (µs per render, including the snapshot check). Ten retained process pairs per configuration (two discarded), alternating order; 1 core pins the render thread and the JIT worker to CPU 8, 2 cores to CPUs 8 and 10. Checksums agree in every process; the candidate recorded no fallbacks or deopts. These host-inclusive timings are never divided by a Bun timing; host-compute above is the equivalent pure-JavaScript kernel.

Workload	Cores	0.12.9	0.12.10	quickjs‑jit	vs 0.12.9	vs 0.12.10
compute	1	15.1	127.1	10.3	1.5x (faster)	12.3x (faster)
compute	2	16.3	62.4	10.2	1.6x (faster)	6.6x (faster)
mixed	1	178.5	363.3	92.7	1.9x (faster)	3.9x (faster)
mixed	2	179.2	180.1	92.4	1.9x (faster)	2.0x (faster)
panel	1	812.0	981.4	803.7	1.03x (faster)	1.3x (faster)
panel	2	806.9	987.9	805.5	1.00x (tied)	1.2x (faster)

Statistically tied: panel on 2 cores vs 0.12.9 (between 1.1% slower and 0.4% faster).

Protocol shared-js-multibatch-v3: each fresh process performs one initial call, 64 untimed ten-call warmup batches, then 16 timed ten-call batches; the process latency is the upper median of those batches. Result consumption and the checksum are outside the timed boundary, and inputs, scripts and driver (60408ad6…) are equivalent across engines; every checksum agrees. Each workload/configuration has five discarded and 30 retained processes, and configuration order reverses on alternating pairs. All 6,125 process records are archived; timing samples are not filtered by native readiness and outlier batches are counted, never dropped. Confidence intervals use 10,000 deterministic paired percentile-bootstrap resamples (seed 20260909). v3 timings are not comparable with the historical v2 matrices below.

Regression and diagnostics — vs previous: quickjs-jit speed as a multiple of the v0.12.10 automatic JIT. Interpreter: candidate / previous interpreter speed (identical interpreter code, so this is the measurement's noise control). Entries: native entries in the first timed batch (counts include the shared driver). Quiet: processes out of 30 with no compilation or installation during the timed window. Outliers: timed batches slower than 5x the median (affected processes).

Workload	vs previous	Interpreter	Entries	Quiet	Outliers
scalar‑control‑flow	1.01x (faster)	1.00x (tied)	20	30	0 (0)
scalar‑expressions	0.99x (slower)	1.00x (slower)	10–20	29	0 (0)
host‑compute	1.04x (faster)	1.00x (tied)	20	30	0 (0)
mixed‑quotes	0.99x (slower)	1.00x (tied)	6990	30	0 (0)
quickjs‑int‑arith	1.00x (tied)	1.00x (tied)	20	30	0 (0)
quickjs‑bitops	1.00x (tied)	1.00x (tied)	20	30	0 (0)
quickjs‑fibonacci	0.99x (slower)	1.01x (tied)	20	30	0 (0)
numeric	1.04x (faster)	1.00x (tied)	10–20	9	10 (10)
scalar‑loop	1.03x (faster)	1.00x (tied)	10–20	13	9 (9)
call‑heavy	1.00x (tied)	1.00x (tied)	10	27	3 (3)
generic‑call‑entry	0.98x (tied)	1.00x (tied)	10	17	3 (3)
generic‑call‑fallback	1.00x (faster)	1.00x (faster)	10	30	0 (0)
property‑heavy	1.00x (tied)	1.00x (tied)	10	28	1 (1)
fibonacci‑iterative	1.01x (tied)	1.00x (tied)	20	30	0 (0)
fibonacci‑recursive	1.00x (tied)	1.00x (tied)	20	30	0 (0)
collections	1.01x (tied)	1.00x (tied)	10	30	0 (0)
for‑of‑array	0.36x (slower)	0.98x (tied)	10	30	0 (0)
strings‑json	1.00x (tied)	0.99x (tied)	10	30	0 (0)
calls‑closures	0.99x (slower)	0.99x (slower)	10	30	0 (0)
adversarial	1.01x (tied)	0.98x (slower)	10	30	0 (0)
float64‑dense	0.97x (slower)	1.00x (tied)	20	4	29 (26)
strings‑regexp	0.96x (slower)	0.94x (slower)	10	30	0 (0)
arrays‑typed	1.00x (tied)	0.99x (tied)	30	30	0 (0)
packed‑array‑traversal	0.98x (slower)	1.00x (tied)	10	17	7 (7)
int32array‑traversal	1.02x (faster)	1.00x (tied)	10	28	1 (1)
float64array‑traversal	1.01x (faster)	0.99x (slower)	10	30	0 (0)
typed‑convert‑traversal	1.01x (tied)	0.98x (tied)	10	30	0 (0)
objects‑polymorphic	0.98x (tied)	1.01x (tied)	20	30	0 (0)
property‑polymorphic	1.00x (tied)	1.00x (faster)	10	30	0 (0)
calls‑recursion‑closures	1.00x (tied)	0.99x (tied)	10	30	0 (0)
json‑codec	1.01x (tied)	1.00x (tied)	10	30	0 (0)
map‑set‑bigint	0.97x (slower)	0.99x (slower)	10	30	0 (0)
exceptions‑promises‑async	1.02x (faster)	1.00x (tied)	10	30	0 (0)
methods‑dynamic	0.98x (slower)	0.98x (tied)	10	30	0 (0)
exceptions‑sync	0.21x (slower)	1.00x (tied)	10	30	0 (0)

Host: 13th Gen Intel Core i7-13700KF, Linux 7.2.5 x86_64, CPU 8 pinned, powersave governor. Toolchain: rustc 1.98.1 / LLVM 22.1.8. Bun 1.4.0, default flags. Candidate binary 5283ef08…, previous binary d0ba09cf….

All raw process records and metadata, machine-readable comparisons, collection metadata, gpui-shell comparisons and raw samples, and verified hashes and provenance.

Historical 35-scenario Linux x86_64 matrix, measured 2026-09-30: V8/Bun roadmap P0–P4 (v0.12.10). These measurements apply only to the 889f7c1 (v0.12.10) runtime. Its automatic tiering regressed real gpui-shell renders, which the current matrix above fixes.

The measured candidate is main revision 889f7c19adb46976e98ff00729d606d77cd933d2 (PR #30). The previous-JIT regression baseline is the 82d3808 (v0.12.9) runtime built with the identical shared-js-multibatch-v3 worker harness (cfd5667); the later harness changes only add workloads and tests. QuickJS is the candidate binary with the JIT detached, so the profitability comparison uses the same QuickJS-NG fd0a0210b7be00957751871e7e01b8291268fc29 runtime and benchmark harness. Bun 1.4.0 runs the same JavaScript from a file wrapper with default flags.

Relative to the previous automatic JIT, 28 scenarios are significantly faster, 2 are significantly slower (quickjs-fibonacci, calls-recursion-closures), and 5 are statistically tied. Relative to same-version QuickJS, 25 are faster, 5 are slower (strings-json, adversarial, strings-regexp, calls-recursion-closures, exceptions-promises-async), and 5 are tied. Relative to Bun, all 35 are slower: the closest are fibonacci-recursive (0.72x Bun speed), generic-call-entry (0.65x), float64-dense (0.65x), quickjs-bitops (0.61x) and exceptions-sync (0.57x); the widest gaps are methods-dynamic (0.013x), calls-recursion-closures (0.019x), arrays-typed (0.024x) and objects-polymorphic (0.025x). The tables retain every losing, fallback-only and inconclusive scenario.

How to read the table:

QuickJS / Bun / quickjs-jit: latency in ms per ten workload calls (lower is better) — the median over processes of each process's upper-median of 16 timed batches.
vs QuickJS / vs Bun: quickjs-jit speed as a multiple of that engine's speed (that engine's latency / quickjs-jit latency, paired geometric mean). 48.8x (faster) means quickjs-jit runs 48.8 times as fast; 0.54x (slower) means it runs at 0.54 times that engine's speed. faster / slower means the 95% confidence interval lies entirely above / below 1.00x; tied means it crosses 1.00x.
QuickJS is the candidate binary with the JIT detached; quickjs-jit uses production automatic tiering. Ratios use one decimal place from 1.1x and two decimals below it.

Workload	QuickJS	Bun	quickjs‑jit	vs QuickJS	vs Bun
scalar‑control‑flow	2.13	0.0238	0.0437	48.8x (faster)	0.54x (slower)
scalar‑expressions	2.16	0.0201	0.0452	47.6x (faster)	0.45x (slower)
host‑compute	2.16	0.028	0.0515	41.6x (faster)	0.52x (slower)
mixed‑quotes	2.76	0.0593	0.644	4.3x (faster)	0.092x (slower)
quickjs‑int‑arith	5.82	0.0481	0.319	18.3x (faster)	0.15x (slower)
quickjs‑bitops	1.19	0.016	0.0278	42.8x (faster)	0.61x (slower)
quickjs‑fibonacci	0.947	0.0132	0.373	2.5x (faster)	0.034x (slower)
numeric	0.538	0.0115	0.021	25.6x (faster)	0.54x (slower)
scalar‑loop	0.538	0.0114	0.021	25.6x (faster)	0.56x (slower)
call‑heavy	1.28	0.0207	0.039	33.0x (faster)	0.49x (slower)
generic‑call‑entry	0.971	0.017	0.026	36.5x (faster)	0.65x (slower)
generic‑call‑fallback	1.29	0.0238	0.206	6.2x (faster)	0.11x (slower)
property‑heavy	1.07	0.00818	0.0282	37.7x (faster)	0.37x (slower)
fibonacci‑iterative	35.5	0.181	0.78	45.5x (faster)	0.22x (slower)
fibonacci‑recursive	9.9	0.267	0.361	26.5x (faster)	0.72x (slower)
collections	1.58	0.0425	1.6	0.99x (tied)	0.029x (slower)
for‑of‑array	0.459	0.0423	0.17	2.6x (faster)	0.17x (slower)
strings‑json	1.9	0.0709	1.99	0.95x (slower)	0.037x (slower)
calls‑closures	3.2	0.0941	3.2	1.00x (tied)	0.032x (slower)
adversarial	0.937	0.0359	0.951	0.98x (slower)	0.038x (slower)
float64‑dense	2.28	0.0655	0.101	22.4x (faster)	0.65x (slower)
strings‑regexp	19.2	0.997	19.9	0.97x (slower)	0.05x (slower)
arrays‑typed	4.12	0.0446	1.89	2.2x (faster)	0.024x (slower)
packed‑array‑traversal	0.826	0.0151	0.0301	27.4x (faster)	0.47x (slower)
int32array‑traversal	1.03	0.015	0.0281	36.5x (faster)	0.53x (slower)
float64array‑traversal	1.51	0.0083	0.0571	26.4x (faster)	0.14x (slower)
typed‑convert‑traversal	1.49	0.00811	0.0571	26.0x (faster)	0.14x (slower)
objects‑polymorphic	6.12	0.121	4.8	1.3x (faster)	0.025x (slower)
property‑polymorphic	1.86	0.0234	0.564	3.3x (faster)	0.041x (slower)
calls‑recursion‑closures	6.45	0.126	6.48	0.99x (slower)	0.019x (slower)
json‑codec	79.4	7.69	80.2	0.98x (tied)	0.095x (slower)
map‑set‑bigint	17.6	0.865	17.6	1.01x (tied)	0.049x (slower)
exceptions‑promises‑async	2.32	0.208	2.42	0.97x (slower)	0.086x (slower)
methods‑dynamic	12.9	0.151	12.8	1.00x (tied)	0.013x (slower)
exceptions‑sync	2.92	0.36	0.62	4.7x (faster)	0.57x (slower)

95% confidence intervals and statistically tied results

Each cell is quickjs-jit's speed ratio followed by its 95% confidence interval. vs previous compares with the 82d3808 automatic JIT. Exact values, including the interpreter-control intervals, are in the machine-readable comparisons.

Workload	vs QuickJS	vs Bun	vs previous
scalar‑control‑flow	48.8x [48.7, 48.9]	0.54x [0.54, 0.54]	1.3x [1.2, 1.3]
scalar‑expressions	47.6x [47.5, 47.8]	0.45x [0.44, 0.46]	1.1x [1.1, 1.1]
host‑compute	41.6x [40.9, 42.1]	0.52x [0.48, 0.55]	1.4x [1.3, 1.4]
mixed‑quotes	4.3x [4.3, 4.3]	0.092x [0.091, 0.093]	4.0x [4.0, 4.0]
quickjs‑int‑arith	18.3x [18.3, 18.3]	0.15x [0.15, 0.15]	4.3x [4.2, 4.5]
quickjs‑bitops	42.8x [42.4, 43.0]	0.61x [0.57, 0.67]	2.2x [2.2, 2.3]
quickjs‑fibonacci	2.5x [2.5, 2.5]	0.034x [0.031, 0.035]	0.97x [0.96, 0.97]
numeric	25.6x [25.6, 25.6]	0.54x [0.50, 0.56]	1.4x [1.4, 1.4]
scalar‑loop	25.6x [25.6, 25.7]	0.56x [0.55, 0.57]	1.4x [1.4, 1.4]
call‑heavy	33.0x [32.9, 33.0]	0.49x [0.43, 0.53]	1.6x [1.6, 1.6]
generic‑call‑entry	36.5x [35.9, 37.0]	0.65x [0.58, 0.70]	1.2x [1.2, 1.2]
generic‑call‑fallback	6.2x [6.1, 6.3]	0.11x [0.10, 0.11]	43.3x [42.1, 44.0]
property‑heavy	37.7x [37.4, 38.0]	0.37x [0.32, 0.44]	3.5x [3.5, 3.5]
fibonacci‑iterative	45.5x [45.3, 45.6]	0.22x [0.21, 0.22]	1.3x [1.3, 1.3]
fibonacci‑recursive	26.5x [25.5, 27.2]	0.72x [0.69, 0.74]	26.4x [25.3, 27.1]
collections	0.99x [0.96, 1.02]	0.029x [0.026, 0.035]	1.01x [0.98, 1.04]
for‑of‑array	2.6x [2.5, 2.7]	0.17x [0.15, 0.21]	3.0x [2.9, 3.2]
strings‑json	0.95x [0.95, 0.95]	0.037x [0.036, 0.038]	1.01x [1.00, 1.02]
calls‑closures	1.00x [1.00, 1.00]	0.032x [0.03, 0.034]	1.07x [1.07, 1.08]
adversarial	0.98x [0.98, 0.99]	0.038x [0.037, 0.038]	1.01x [0.99, 1.02]
float64‑dense	22.4x [22.4, 22.5]	0.65x [0.64, 0.65]	1.3x [1.3, 1.3]
strings‑regexp	0.97x [0.96, 0.98]	0.05x [0.05, 0.051]	1.01x [1.01, 1.01]
arrays‑typed	2.2x [2.2, 2.2]	0.024x [0.024, 0.025]	1.6x [1.6, 1.6]
packed‑array‑traversal	27.4x [27.3, 27.5]	0.47x [0.43, 0.50]	1.6x [1.6, 1.6]
int32array‑traversal	36.5x [36.1, 36.9]	0.53x [0.53, 0.54]	1.8x [1.7, 1.8]
float64array‑traversal	26.4x [26.3, 26.5]	0.14x [0.14, 0.15]	19.7x [19.6, 19.7]
typed‑convert‑traversal	26.0x [25.9, 26.1]	0.14x [0.14, 0.14]	19.5x [19.1, 20.3]
objects‑polymorphic	1.3x [1.3, 1.3]	0.025x [0.025, 0.026]	5.0x [4.9, 5.0]
property‑polymorphic	3.3x [3.3, 3.3]	0.041x [0.041, 0.042]	3.3x [3.3, 3.3]
calls‑recursion‑closures	0.99x [0.99, 1.00]	0.019x [0.019, 0.019]	0.99x [0.99, 1.00]
json‑codec	0.98x [0.96, 1.01]	0.095x [0.092, 0.098]	0.99x [0.97, 1.00]
map‑set‑bigint	1.01x [0.99, 1.03]	0.049x [0.048, 0.05]	0.98x [0.95, 1.01]
exceptions‑promises‑async	0.97x [0.96, 0.97]	0.086x [0.086, 0.087]	1.5x [1.5, 1.5]
methods‑dynamic	1.00x [0.98, 1.01]	0.013x [0.012, 0.014]	1.00x [0.98, 1.02]
exceptions‑sync	4.7x [4.6, 4.8]	0.57x [0.56, 0.58]	4.7x [4.6, 4.8]

Statistically tied: collections vs QuickJS (between 3.6% slower and 2.2% faster); collections vs previous JIT (between 1.9% slower and 4.2% faster); calls-closures vs QuickJS (between 0.5% slower and 0.4% faster); adversarial vs previous JIT (between 0.5% slower and 2.3% faster); json-codec vs QuickJS (between 4.1% slower and 1.0% faster); json-codec vs previous JIT (between 3.4% slower and 0.3% faster); map-set-bigint vs QuickJS (between 0.7% slower and 3.4% faster); map-set-bigint vs previous JIT (between 4.6% slower and 0.9% faster); methods-dynamic vs QuickJS (between 1.9% slower and 0.9% faster); methods-dynamic vs previous JIT (between 1.5% slower and 1.8% faster).

Statistically tied interpreter controls: scalar-control-flow (between 0.1% slower and 0.2% faster); scalar-expressions (between 0.0% slower and 0.4% faster); mixed-quotes (between 0.1% slower and 0.4% faster); quickjs-int-arith (between 0.3% slower and 0.3% faster); quickjs-bitops (between 0.2% slower and 0.2% faster); numeric (between 0.0% slower and 0.2% faster); scalar-loop (between 0.1% slower and 0.2% faster); property-heavy (between 0.1% slower and 0.4% faster); fibonacci-iterative (between 0.1% slower and 0.0% faster); arrays-typed (between 0.1% slower and 0.2% faster); packed-array-traversal (between 0.1% slower and 0.1% faster); int32array-traversal (between 0.9% slower and 0.4% faster); property-polymorphic (between 0.2% slower and 0.1% faster); json-codec (between 1.9% slower and 4.2% faster); map-set-bigint (between 2.0% slower and 2.1% faster); methods-dynamic (between 0.6% slower and 2.4% faster); exceptions-sync (between 1.4% slower and 0.5% faster).

Protocol shared-js-multibatch-v3: each fresh process performs one initial call, 64 untimed ten-call warmup batches, then 16 timed ten-call batches; the process latency is the upper median of those batches. Result consumption and the checksum are outside the timed boundary, and inputs, scripts and driver (60408ad6…) are equivalent across engines; every checksum agrees. Each workload/configuration has five discarded and 30 retained processes, and configuration order reverses on alternating pairs. All 6,125 process records are archived; timing samples are not filtered by native readiness and outlier batches are counted, never dropped. Confidence intervals use 10,000 deterministic paired percentile-bootstrap resamples (seed 20260909). v3 timings are not comparable with the historical v2 matrices below.

Regression and diagnostics — vs previous: quickjs-jit speed as a multiple of the 82d3808 automatic JIT. Interpreter: candidate / previous interpreter speed, showing that some previous-JIT gains on calls-closures (1.1x) and for-of-array (1.1x) come from runtime changes that also speed up the interpreter. Entries: native entries in the first timed batch (identical in every process; counts include the shared driver). Quiet: processes out of 30 with no compilation or installation during the timed window. Outliers: timed batches slower than 5x the median (affected processes).

Workload	vs previous	Interpreter	Entries	Quiet	Outliers
scalar‑control‑flow	1.3x (faster)	1.00x (tied)	20	30	0 (0)
scalar‑expressions	1.1x (faster)	1.00x (tied)	20	14	8 (8)
host‑compute	1.4x (faster)	1.00x (faster)	20	30	0 (0)
mixed‑quotes	4.0x (faster)	1.00x (tied)	6990	30	0 (0)
quickjs‑int‑arith	4.3x (faster)	1.00x (tied)	20	30	0 (0)
quickjs‑bitops	2.2x (faster)	1.00x (tied)	20	30	0 (0)
quickjs‑fibonacci	0.97x (slower)	1.00x (faster)	20	30	0 (0)
numeric	1.4x (faster)	1.00x (tied)	20	30	0 (0)
scalar‑loop	1.4x (faster)	1.00x (tied)	20	30	0 (0)
call‑heavy	1.6x (faster)	1.00x (slower)	10	0	6 (6)
generic‑call‑entry	1.2x (faster)	0.99x (slower)	10	0	6 (6)
generic‑call‑fallback	43.3x (faster)	0.99x (slower)	10	30	0 (0)
property‑heavy	3.5x (faster)	1.00x (tied)	10	0	9 (9)
fibonacci‑iterative	1.3x (faster)	1.00x (tied)	20	30	0 (0)
fibonacci‑recursive	26.4x (faster)	0.99x (slower)	20	30	0 (0)
collections	1.01x (tied)	1.02x (faster)	10	30	0 (0)
for‑of‑array	3.0x (faster)	1.09x (faster)	20	5	19 (18)
strings‑json	1.01x (faster)	1.01x (faster)	10	30	0 (0)
calls‑closures	1.07x (faster)	1.08x (faster)	10	30	0 (0)
adversarial	1.01x (tied)	1.01x (faster)	10	30	0 (0)
float64‑dense	1.3x (faster)	1.00x (faster)	20	12	15 (13)
strings‑regexp	1.01x (faster)	1.01x (faster)	10	30	0 (0)
arrays‑typed	1.6x (faster)	1.00x (tied)	30	30	0 (0)
packed‑array‑traversal	1.6x (faster)	1.00x (tied)	10	0	4 (4)
int32array‑traversal	1.8x (faster)	1.00x (tied)	10	0	5 (5)
float64array‑traversal	19.7x (faster)	1.02x (faster)	10	0	24 (24)
typed‑convert‑traversal	19.5x (faster)	1.02x (faster)	10	0	22 (22)
objects‑polymorphic	5.0x (faster)	1.03x (faster)	20	30	0 (0)
property‑polymorphic	3.3x (faster)	1.00x (tied)	10	30	0 (0)
calls‑recursion‑closures	0.99x (slower)	1.00x (slower)	10	30	0 (0)
json‑codec	0.99x (tied)	1.01x (tied)	10	30	0 (0)
map‑set‑bigint	0.98x (tied)	1.01x (tied)	10	30	0 (0)
exceptions‑promises‑async	1.5x (faster)	1.01x (faster)	10	30	0 (0)
methods‑dynamic	1.00x (tied)	1.01x (tied)	10	30	0 (0)
exceptions‑sync	4.7x (faster)	1.00x (tied)	20	30	0 (0)

Host: 13th Gen Intel Core i7-13700KF, Linux 7.2.5 x86_64, CPU 8 pinned, powersave governor. Toolchain: rustc 1.98.1 / LLVM 22.1.8. Bun 1.4.0, default flags. Candidate binary d0ba09cf…, previous binary 9660bcbf…; the two worktrees used their own untracked Cargo.lock files (hashes recorded in the manifest). host-compute is the equivalent pure-JavaScript render kernel; it excludes GPUI allocation and snapshot work, so no Bun result is divided by a host-inclusive timing.

All raw process records and metadata, machine-readable comparisons, collection metadata, and verified hashes and provenance.

Historical 30-scenario Linux x86_64 matrix, measured 2026-09-16: bounded facts, effects, and exact recovery. These measurements apply only to the 6165383 candidate and use the retired shared-js-fixed-warmup-v2 protocol. Its Bun column was launched with bun -e and timed one batch that hit a deterministic JSC stall, inflating Bun latency by 1.2x–32x; do not compare its Bun ratios or absolute timings with the current v3 matrix.

The measured candidate is implementation revision 6165383e01ba971ff99c70b1c4d6949b38a6349c. The previous-JIT regression baseline is e174b52eeaccefcf5891fe28fa5bbc960742846e. The branch contains later test-only and warning-cleanup commits; they were not substituted for the frozen candidate binary. QuickJS is the candidate binary with JIT detached, so the profitability comparison uses the same QuickJS-NG fd0a0210b7be00957751871e7e01b8291268fc29 runtime and benchmark harness. Bun 1.4.0 runs the same JavaScript with default flags.

Relative to the previous automatic JIT, 9 scenarios are significantly faster, 12 are significantly slower, and 9 are statistically tied. Relative to same-version QuickJS, 19 are faster, 6 are slower, and 5 are tied. Relative to Bun, 7 are faster, 22 are slower, and one is tied. These counts and the table retain every losing, unsupported, fallback-only, and inconclusive scenario.

Timings are medians in ms per ten workload calls. Ratios are paired geometric means of reference latency / candidate latency with 95% confidence intervals; values above 1.00x mean quickjs-jit is faster. An interval crossing 1.00x is explicitly described as statistically tied. Production automatic tiering is used for both JIT columns.

Workload	Previous JIT ms	QuickJS ms	Bun ms	quickjs-jit ms	JIT / previous speed [95% CI]	JIT / QuickJS speed [95% CI]	JIT / Bun speed [95% CI]
scalar-control-flow	0.045778	2.133281	0.091552	0.054922	0.8386x [0.8231, 0.8541]	38.5227x [38.0405, 38.8607]	1.6688x [1.6424, 1.6960]
scalar-expressions	0.044714	2.163232	0.092583	0.050669	0.9007x [0.8650, 0.9483]	42.3554x [41.4480, 42.9199]	1.8282x [1.7845, 1.8649]
host-compute	0.053196	2.172640	0.246817	0.070956	0.7801x [0.7283, 0.8427]	30.6812x [29.7415, 31.5500]	3.5209x [3.3911, 3.6786]
mixed-quotes	2.698116	2.771889	0.152693	2.548821	1.0365x [0.9996, 1.0620]; statistically tied (between 0.04% slower and 6.20% faster)	1.0734x [1.0386, 1.0991]	0.0591x [0.0569, 0.0608]
quickjs-int-arith	1.287012	5.813427	0.132201	1.320630	0.9735x [0.9647, 0.9831]	4.4197x [4.3659, 4.5015]	0.1048x [0.1002, 0.1108]
quickjs-bitops	0.057992	1.193420	0.149943	0.062194	0.9279x [0.9182, 0.9374]	18.9681x [18.7344, 19.1685]	2.4599x [2.4036, 2.5245]
quickjs-fibonacci	0.360930	0.948122	0.147681	0.365482	1.0619x [0.9785, 1.2347]; statistically tied (between 2.15% slower and 23.47% faster)	2.6676x [2.5993, 2.7438]	0.4542x [0.4262, 0.4886]
numeric	0.021639	0.538170	0.130202	0.029413	0.7351x [0.7326, 0.7375]	18.4386x [18.2196, 18.8014]	4.4610x [4.4228, 4.5046]
scalar-loop	0.021606	0.537613	0.130402	0.029525	0.7300x [0.7269, 0.7330]	18.1336x [17.9761, 18.2676]	4.7422x [4.4129, 5.3736]
call-heavy	0.065863	1.277821	0.006313	0.059645	1.1076x [1.0942, 1.1227]	21.5486x [21.2857, 21.8386]	0.1057x [0.1047, 0.1065]
generic-call-entry	0.038561	0.962674	0.004927	0.032415	1.1790x [1.1633, 1.1920]	29.3967x [29.0059, 29.7046]	0.1512x [0.1490, 0.1531]
generic-call-fallback	3.694141	1.278822	0.008122	9.120734	0.4141x [0.4049, 0.4317]	0.1403x [0.1400, 0.1407]	0.0009x [0.0009, 0.0009]
property-heavy	0.291567	1.060369	0.007750	0.097573	2.9656x [2.9310, 2.9957]	10.7877x [10.6629, 10.8779]	0.0792x [0.0783, 0.0799]
fibonacci-iterative	0.790981	35.577327	1.011399	0.982645	0.8838x [0.8327, 0.9478]	37.9064x [36.7933, 39.1683]	1.0650x [1.0162, 1.1267]
fibonacci-recursive	9.900345	9.804779	0.551945	9.835726	1.0025x [0.9987, 1.0055]; statistically tied (between 0.13% slower and 0.55% faster)	0.9962x [0.9917, 1.0013]; statistically tied (between 0.83% slower and 0.13% faster)	0.0588x [0.0563, 0.0621]
collections	1.614670	1.679045	0.166383	1.644529	1.0499x [0.9764, 1.1314]; statistically tied (between 2.36% slower and 13.14% faster)	1.0935x [1.0402, 1.1527]	0.1059x [0.0971, 0.1159]
strings-json	2.031920	1.912065	0.173625	2.008080	1.0313x [1.0046, 1.0714]	0.9542x [0.9446, 0.9661]	0.0881x [0.0856, 0.0913]
calls-closures	3.542016	3.245382	0.192532	3.225958	1.1310x [1.0776, 1.2094]	0.9947x [0.9823, 1.0050]; statistically tied (between 1.77% slower and 0.50% faster)	0.0593x [0.0582, 0.0606]
adversarial	0.919529	0.937453	0.125707	0.949674	1.0345x [0.9904, 1.0892]; statistically tied (between 0.96% slower and 8.92% faster)	0.9870x [0.9855, 0.9886]	0.1355x [0.1323, 0.1404]
float64-dense	0.123468	2.283089	0.149039	0.128683	0.9176x [0.8328, 0.9671]	16.8169x [15.2283, 17.7104]	1.1031x [0.9966, 1.1654]; statistically tied (between 0.34% slower and 16.54% faster)
strings-regexp	20.227821	19.605459	1.423642	20.402175	0.9939x [0.9773, 1.0087]; statistically tied (between 2.27% slower and 0.87% faster)	0.9631x [0.9506, 0.9727]	0.0719x [0.0693, 0.0755]
arrays-typed	2.586654	4.163713	0.150749	2.989603	0.8696x [0.8519, 0.8927]	1.3920x [1.3740, 1.4087]	0.0516x [0.0496, 0.0543]
packed-array-traversal	0.246330	0.825657	0.004896	0.048205	5.0431x [4.9706, 5.1073]	17.0117x [16.7616, 17.2438]	0.1040x [0.0994, 0.1114]
int32array-traversal	0.296837	1.037577	0.005695	0.049892	5.8171x [5.6125, 6.0133]	20.9835x [19.9947, 22.0130]	0.1114x [0.1077, 0.1141]
float64array-traversal	1.704780	1.544923	0.008223	1.103959	1.5465x [1.5016, 1.5944]	1.3713x [1.3022, 1.4221]	0.0074x [0.0070, 0.0077]
objects-polymorphic	6.970870	6.348050	0.241782	24.704628	0.2908x [0.2658, 0.3185]	0.2502x [0.2345, 0.2661]	0.0103x [0.0092, 0.0116]
calls-recursion-closures	6.717968	6.513333	0.213999	6.534107	1.0590x [1.0053, 1.1182]	1.0202x [0.9665, 1.0835]; statistically tied (between 3.35% slower and 8.35% faster)	0.0350x [0.0314, 0.0411]
json-codec	80.670948	78.422380	7.876545	81.808247	0.9921x [0.9550, 1.0317]; statistically tied (between 4.50% slower and 3.17% faster)	0.9756x [0.9352, 1.0214]; statistically tied (between 6.48% slower and 2.14% faster)	0.1050x [0.0950, 0.1208]
map-set-bigint	17.947786	17.583268	1.035041	17.035912	1.0372x [0.9727, 1.1027]; statistically tied (between 2.73% slower and 10.27% faster)	1.0586x [0.9841, 1.1397]; statistically tied (between 1.59% slower and 13.97% faster)	0.0593x [0.0557, 0.0630]
exceptions-promises-async	3.516657	2.341826	0.284759	3.593704	1.0661x [0.8863, 1.2762]; statistically tied (between 11.37% slower and 27.62% faster)	0.7186x [0.5913, 0.8698]	0.0778x [0.0658, 0.0959]

Protocol shared-js-fixed-warmup-v2: each fresh process performs one initial call, 64 untimed ten-call warmup batches, then one timed ten-call batch. Result consumption and the checksum are outside the timed boundary and inputs, scripts, and driver are equivalent across engines. Each workload/configuration has five discarded processes and 30 retained processes. Configuration order reverses on alternating pairs. All 5,250 process records are archived; timing samples are not filtered by native readiness. Confidence intervals use 10,000 deterministic paired percentile-bootstrap resamples (seed 20260909).

Host: 13th Gen Intel Core i7-13700KF, Linux 7.2.3 x86_64, CPU 8 pinned, powersave governor. Toolchain: rustc 1.98.1 / LLVM 22.1.8. Bun 1.4.0, default flags. Candidate and previous binaries use the identical Cargo.lock c025c52c3d35545657afc32cb5e7bde9ba293f8d12adb26fe47ee713754c1588. host-compute is the equivalent pure-JavaScript render kernel; it excludes GPUI allocation and snapshot work, so no Bun result is divided by a host-inclusive timing.

All raw process records and metadata, machine-readable comparisons, collection metadata, and verified hashes and provenance.

Historical 27-scenario macOS matrix, 2026-09-10: bounded semantic inlining candidate. These measurements apply only to the archived 432afdf bounded-inlining candidate and predate scalar poll/store/induction changes.

This frozen candidate adds effect-free monomorphic inline regions to semantic SSA: caller ValueId argument binding, literal-boolean branch folding, checked Int32 add/subtract, and ordinary-call fallback when compilation budgets are exhausted. Cold recovery replays the original pure call from its caller state; this is not general inline-frame reconstruction. Loop guard hoisting, lazy frame synchronization, heap facts/LICM, property forwarding/sinking, array range elimination and the required Bun performance gates remain unfinished.

The previous JIT column is the frozen callee-retention candidate 92e1d5c90f1f30b15f65c8dd41afa6020683bd297c23fa56d4224da681312333. The original 023a220 baseline is also measured and retained in the comparison JSON and interpreter control below. Source snapshots and binary hashes identify all measured revisions; no later work is implied by these timings.

The numeric-foundation no-regression gate was unmet in this historical run. Automatic runtime slowdowns relative to 023a220 with intervals entirely below parity occur in: arrays-typed. Interpreter controls are included below; these observations alone do not isolate the cause. Relative to the immediately preceding retention candidate, host-compute and strings-json also have intervals entirely below parity. The latter has zero native entries in this window, so its slowdown is not evidence of slower JIT machine code.

Timings are medians in ms per ten workload calls. Speed is reference latency / candidate latency with paired geometric means and 95% confidence intervals; values above 1x are faster. QuickJS uses the candidate binary with JIT detached; quickjs-jit uses production automatic tiering; Bun uses defaults. All scenarios, including losing, fallback-only and inconclusive results, remain in the matrix.

Workload	Previous JIT ms	QuickJS ms	Bun ms	quickjs-jit ms	JIT / previous speed [95% CI]	JIT / QuickJS speed [95% CI]	JIT / Bun speed [95% CI]
scalar-control-flow	0.047750	1.175792	0.078271	0.047709	0.9853x [0.9280, 1.0232]; statistically tied (7.20% slower to 2.32% faster)	23.9686x [22.6382, 24.7696]	1.6098x [1.5093, 1.6789]
scalar-expressions	0.044125	0.817542	0.068312	0.044313	0.9884x [0.9727, 1.0035]; statistically tied (2.73% slower to 0.35% faster)	18.1951x [17.8991, 18.4695]	1.5752x [1.5340, 1.6255]
host-compute	0.058396	1.213834	0.197792	0.059354	0.9521x [0.8951, 0.9937]	20.8922x [19.3012, 22.4833]	3.2452x [3.0252, 3.4199]
mixed-quotes	2.142396	1.544396	0.112334	2.142438	1.0532x [0.9925, 1.1770]; statistically tied (0.75% slower to 17.70% faster)	0.7232x [0.7168, 0.7298]	0.0532x [0.0519, 0.0549]
quickjs-int-arith	1.732520	3.391000	0.113729	1.729000	1.0323x [1.0036, 1.0826]	1.9795x [1.9605, 2.0029]	0.0669x [0.0657, 0.0684]
quickjs-bitops	0.205167	0.540999	0.128145	0.207167	0.9810x [0.9454, 1.0069]; statistically tied (5.46% slower to 0.69% faster)	2.5906x [2.4972, 2.6658]	0.6159x [0.5969, 0.6316]
quickjs-fibonacci	0.473542	0.645188	0.129479	0.478375	0.9666x [0.9318, 1.0002]; statistically tied (6.82% slower to 0.02% faster)	1.3202x [1.2791, 1.3646]	0.2794x [0.2583, 0.3074]
numeric	0.024500	0.397479	0.129833	0.026083	0.9638x [0.8708, 1.0600]; statistically tied (12.92% slower to 6.00% faster)	15.9114x [14.3101, 17.6350]	4.9788x [4.3793, 5.7124]
scalar-loop	0.024521	0.394896	0.137479	0.024688	1.0165x [0.9247, 1.1130]; statistically tied (7.53% slower to 11.30% faster)	15.8441x [14.9814, 16.7318]	5.5678x [4.9140, 6.3275]
call-heavy	0.322375	0.995979	0.012875	0.289959	1.0952x [1.0457, 1.1354]	3.3787x [3.2140, 3.5186]	0.0438x [0.0424, 0.0452]
generic-call-entry	0.251812	0.815083	0.006041	0.214479	1.1719x [1.1012, 1.2358]	3.7952x [3.6470, 3.9572]	0.0275x [0.0264, 0.0286]
generic-call-fallback	2.972542	1.100104	0.012291	3.122937	0.9724x [0.8946, 1.0731]; statistically tied (10.54% slower to 7.31% faster)	0.3514x [0.3152, 0.3977]	0.0037x [0.0034, 0.0039]
property-heavy	0.304667	0.918938	0.011291	0.304833	0.9619x [0.8574, 1.0320]; statistically tied (14.26% slower to 3.20% faster)	2.9072x [2.4591, 3.2594]	0.0350x [0.0306, 0.0382]
fibonacci-iterative	0.867396	21.294833	0.977417	0.874645	0.9989x [0.9662, 1.0506]; statistically tied (3.38% slower to 5.06% faster)	25.3709x [24.0964, 27.2051]	1.1513x [1.0718, 1.2688]
fibonacci-recursive	5.520396	5.882750	0.495667	5.557437	0.9878x [0.9686, 1.0035]; statistically tied (3.14% slower to 0.35% faster)	1.0625x [1.0391, 1.0865]	0.0892x [0.0869, 0.0914]
collections	1.198521	1.188104	0.172125	1.166250	1.0778x [0.9879, 1.1808]; statistically tied (1.21% slower to 18.08% faster)	1.0185x [0.9616, 1.0826]; statistically tied (3.84% slower to 8.26% faster)	0.1446x [0.1290, 0.1620]
strings-json	2.252625	2.274958	0.176020	2.269646	0.9846x [0.9709, 0.9960]	1.1025x [0.9841, 1.3633]; statistically tied (1.59% slower to 36.33% faster)	0.0804x [0.0766, 0.0850]
calls-closures	3.942312	3.968145	0.157500	3.949771	0.9955x [0.9722, 1.0233]; statistically tied (2.78% slower to 2.33% faster)	1.0272x [0.9742, 1.0975]; statistically tied (2.58% slower to 9.75% faster)	0.0405x [0.0384, 0.0430]
adversarial	0.671167	0.684334	0.140166	0.669854	0.8897x [0.6396, 1.1120]; statistically tied (36.04% slower to 11.20% faster)	0.9023x [0.6658, 1.1178]; statistically tied (33.42% slower to 11.78% faster)	0.1728x [0.1282, 0.2163]
float64-dense	0.531980	2.376459	0.139812	0.544709	0.9889x [0.9752, 1.0030]; statistically tied (2.48% slower to 0.30% faster)	4.6261x [4.3361, 5.1820]	0.2653x [0.2565, 0.2751]
strings-regexp	25.301603	25.816605	1.342375	25.314812	1.0096x [0.9462, 1.0836]; statistically tied (5.38% slower to 8.36% faster)	1.0188x [0.9554, 1.0896]; statistically tied (4.46% slower to 8.96% faster)	0.0510x [0.0474, 0.0552]
arrays-typed	2.185376	3.746750	0.154729	2.176020	1.0033x [0.9899, 1.0165]; statistically tied (1.01% slower to 1.65% faster)	1.7542x [1.7078, 1.8296]	0.0730x [0.0702, 0.0759]
objects-polymorphic	6.406084	6.442104	0.229813	6.505584	1.0051x [0.9758, 1.0409]; statistically tied (2.42% slower to 4.09% faster)	1.0272x [0.9834, 1.1009]; statistically tied (1.66% slower to 10.09% faster)	0.0400x [0.0355, 0.0468]
calls-recursion-closures	4.838770	5.132376	0.235958	4.868229	0.9931x [0.9196, 1.0794]; statistically tied (8.04% slower to 7.94% faster)	1.0073x [0.9431, 1.0643]; statistically tied (5.69% slower to 6.43% faster)	0.0490x [0.0447, 0.0542]
json-codec	96.847521	96.966708	7.374313	97.100667	1.0144x [0.9324, 1.1193]; statistically tied (6.76% slower to 11.93% faster)	0.9550x [0.8988, 1.0054]; statistically tied (10.12% slower to 0.54% faster)	0.0724x [0.0681, 0.0762]
map-set-bigint	21.849355	21.723250	0.902333	21.780625	0.9832x [0.9535, 1.0077]; statistically tied (4.65% slower to 0.77% faster)	1.0006x [0.9564, 1.0490]; statistically tied (4.36% slower to 4.90% faster)	0.0421x [0.0398, 0.0448]
exceptions-promises-async	3.919375	2.485375	0.264833	3.936500	0.9973x [0.9868, 1.0084]; statistically tied (1.32% slower to 0.84% faster)	0.6406x [0.6260, 0.6643]	0.0683x [0.0668, 0.0699]

Protocol shared-js-fixed-warmup-v2: one initial call, 64 ten-call warmup batches, then one timed ten-call batch, with result consumption/checksum outside timing. Scripts, inputs and driver match across engines. Every workload has five discarded and 30 retained fresh processes per configuration; six configurations alternate in reversed order. All 5,670 process records are archived. Bootstrap uses 10,000 paired percentile resamples, seed 20260909, sorted indices 249/9749. Timing samples are not filtered by native readiness.

Host: Apple M3; macOS-27.0-arm64-arm-64bit-Mach-O. Bun 1.4.0, default flags. Toolchain: rustc 1.98.1 (48a229cea 2026-09-01); LLVM version: 22.1.8. Release build, no CPU pinning; power policy not recorded. host-compute measures the equivalent pure-JS render kernel, excluding GPUI allocation and snapshots. No Linux, host-reload or throughput result is inferred.

Automatic native coverage and the paired interpreter control:

Workload	Fixed-batch native entries (min–max)	Compilation quiet samples / 30	Candidate / 023a220 interpreter speed [95% CI]
scalar-control-flow	10–10	30	1.0005x [0.9945, 1.0071]; statistically tied (0.55% slower to 0.71% faster)
scalar-expressions	10–10	30	1.0040x [0.9896, 1.0184]; statistically tied (1.04% slower to 1.84% faster)
host-compute	10–10	30	0.9692x [0.8984, 1.0456]; statistically tied (10.16% slower to 4.56% faster)
mixed-quotes	6990–6990	30	0.9955x [0.9861, 1.0043]; statistically tied (1.39% slower to 0.43% faster)
quickjs-int-arith	10–10	30	1.0270x [0.9807, 1.1151]; statistically tied (1.93% slower to 11.51% faster)
quickjs-bitops	10–10	30	0.9724x [0.9557, 0.9879]
quickjs-fibonacci	10–10	30	1.0297x [0.9711, 1.1054]; statistically tied (2.89% slower to 10.54% faster)
numeric	10–10	19	0.9367x [0.8407, 1.0469]; statistically tied (15.93% slower to 4.69% faster)
scalar-loop	10–10	22	0.9758x [0.8878, 1.0768]; statistically tied (11.22% slower to 7.68% faster)
call-heavy	10–10	30	1.0221x [0.9890, 1.0588]; statistically tied (1.10% slower to 5.88% faster)
generic-call-entry	10–10	30	0.9950x [0.9535, 1.0382]; statistically tied (4.65% slower to 3.82% faster)
generic-call-fallback	20000–20000	30	1.0077x [0.9049, 1.1089]; statistically tied (9.51% slower to 10.89% faster)
property-heavy	10–10	30	0.9824x [0.9144, 1.0491]; statistically tied (8.56% slower to 4.91% faster)
fibonacci-iterative	10–10	30	0.9466x [0.8811, 0.9922]
fibonacci-recursive	0–0	30	0.9802x [0.9616, 0.9925]
collections	0–0	30	0.9791x [0.9104, 1.0502]; statistically tied (8.96% slower to 5.02% faster)
strings-json	0–0	30	0.8938x [0.7252, 0.9993]
calls-closures	0–0	30	0.9806x [0.9359, 1.0146]; statistically tied (6.41% slower to 1.46% faster)
adversarial	0–0	30	0.9247x [0.8538, 0.9854]
float64-dense	10–10	30	0.9490x [0.8459, 1.0107]; statistically tied (15.41% slower to 1.07% faster)
strings-regexp	0–0	30	1.0352x [0.9629, 1.1343]; statistically tied (3.71% slower to 13.43% faster)
arrays-typed	20–20	30	0.9890x [0.9496, 1.0144]; statistically tied (5.04% slower to 1.44% faster)
objects-polymorphic	0–0	30	0.9552x [0.8811, 0.9987]
calls-recursion-closures	0–0	30	1.0085x [0.9822, 1.0356]; statistically tied (1.78% slower to 3.56% faster)
json-codec	0–0	30	0.9928x [0.9796, 1.0050]; statistically tied (2.04% slower to 0.50% faster)
map-set-bigint	0–0	30	1.0068x [0.9561, 1.0711]; statistically tied (4.39% slower to 7.11% faster)
exceptions-promises-async	0–0	30	0.9876x [0.9521, 1.0100]; statistically tied (4.79% slower to 1.00% faster)

All raw records and source/build inputs, all configuration comparisons, verified hashes, validation and provenance, and remaining architecture and performance gates.

Historical 27-scenario macOS matrix, 2026-09-10: semantic scalar SSA candidate. These measurements apply only to the archived semantic-bitwise candidate and predate bounded semantic inlining.

This frozen candidate adds CFG ValueIds/Phi, semantic arithmetic, comparisons, increment/decrement and binary bitwise operations, pre-effect FrameState bindings, explicit checks, CFG type facts and cyclic Int32 representation proofs to 023a220. The archived source snapshot and binary hash identify the measured candidate. Proven numeric guards are removed before machine lowering. General inlining, heap facts/LICM, array range elimination, inline frame chains and full lazy materialization remain unfinished.

The numeric-foundation no-regression gate was unmet in this historical run. Automatic runtime slowdowns with intervals entirely below parity occur in: call-heavy, generic-call-fallback, adversarial. Interpreter controls are included below; these observations alone do not isolate the cause.

Timings are medians in ms per ten workload calls. Speed is reference latency / candidate latency with paired geometric means and 95% confidence intervals; values above 1x are faster. QuickJS uses the candidate binary with JIT detached; quickjs-jit uses production automatic tiering; Bun uses defaults. All scenarios, including losing, fallback-only and inconclusive results, remain in the matrix.

Workload	Previous JIT ms	QuickJS ms	Bun ms	quickjs-jit ms	JIT / previous speed [95% CI]	JIT / QuickJS speed [95% CI]	JIT / Bun speed [95% CI]
scalar-control-flow	0.050167	1.504479	0.130479	0.049562	1.0330x [0.8129, 1.2728]; statistically tied (18.71% slower to 27.28% faster)	30.2170x [23.2157, 37.9195]	2.8112x [2.1330, 3.7216]
scalar-expressions	0.073833	0.876730	0.088708	0.046146	1.5487x [1.3653, 1.7199]	19.2585x [16.6154, 21.9501]	2.0760x [1.6485, 2.6887]
host-compute	0.062833	1.289230	0.243125	0.063687	0.9973x [0.8401, 1.2016]; statistically tied (15.99% slower to 20.16% faster)	21.5916x [18.5598, 25.1055]	4.0384x [3.1145, 5.3765]
mixed-quotes	2.326312	1.648354	0.162833	2.479855	0.9423x [0.8495, 1.0398]; statistically tied (15.05% slower to 3.98% faster)	0.6512x [0.5886, 0.7160]	0.0683x [0.0585, 0.0801]
quickjs-int-arith	1.894625	3.452250	0.141021	1.849104	1.0325x [1.0002, 1.0713]	1.8593x [1.7949, 1.9334]	0.0766x [0.0704, 0.0845]
quickjs-bitops	0.235437	0.535771	0.136187	0.224750	1.0629x [1.0186, 1.1393]	2.3730x [2.3202, 2.4288]	0.6024x [0.5853, 0.6224]
quickjs-fibonacci	0.502604	0.662750	0.137999	0.514270	0.9747x [0.9247, 1.0275]; statistically tied (7.53% slower to 2.75% faster)	1.3299x [1.2647, 1.3983]	0.2802x [0.2564, 0.3108]
numeric	0.023729	0.391125	0.113834	0.023646	1.0485x [0.9860, 1.1215]; statistically tied (1.40% slower to 12.15% faster)	16.4621x [16.2655, 16.6660]	5.1627x [4.6933, 5.7545]
scalar-loop	0.023230	0.388417	0.116229	0.023500	1.0084x [0.9780, 1.0409]; statistically tied (2.20% slower to 4.09% faster)	16.8610x [16.2424, 17.6919]	5.3067x [4.9156, 5.8183]
call-heavy	0.310812	0.994854	0.013000	0.325709	0.9621x [0.9414, 0.9877]	3.0668x [3.0105, 3.1245]	0.0395x [0.0385, 0.0406]
generic-call-entry	0.249541	0.814354	0.006041	0.255167	1.0783x [0.9935, 1.1870]; statistically tied (0.65% slower to 18.70% faster)	3.3439x [3.1984, 3.5209]	0.0234x [0.0229, 0.0240]
generic-call-fallback	2.981104	1.095250	0.011813	2.985333	0.9818x [0.9631, 0.9988]	0.3670x [0.3592, 0.3748]	0.0040x [0.0039, 0.0042]
property-heavy	0.327541	0.928709	0.011291	0.323854	1.0793x [0.9873, 1.1848]; statistically tied (1.27% slower to 18.48% faster)	2.7572x [2.5867, 2.9200]	0.0357x [0.0328, 0.0395]
fibonacci-iterative	0.893228	21.526021	1.073354	0.892687	0.9687x [0.9034, 1.0227]; statistically tied (9.66% slower to 2.27% faster)	23.4207x [21.8899, 24.8480]	1.1802x [1.0819, 1.2915]
fibonacci-recursive	5.763708	6.284833	0.578959	5.815062	0.9946x [0.8774, 1.1087]; statistically tied (12.26% slower to 10.87% faster)	1.0476x [0.9887, 1.1021]; statistically tied (1.13% slower to 10.21% faster)	0.1006x [0.0854, 0.1209]
collections	1.163021	1.173167	0.154584	1.148083	1.0364x [0.9958, 1.0882]; statistically tied (0.42% slower to 8.82% faster)	1.0418x [1.0109, 1.0908]	0.1424x [0.1311, 0.1572]
strings-json	2.308521	2.294146	0.202604	2.335604	0.9640x [0.8851, 1.0331]; statistically tied (11.49% slower to 3.31% faster)	0.9887x [0.9078, 1.0731]; statistically tied (9.22% slower to 7.31% faster)	0.0801x [0.0733, 0.0867]
calls-closures	4.227520	4.224437	0.190812	4.142583	1.0858x [0.9616, 1.2414]; statistically tied (3.84% slower to 24.14% faster)	1.0233x [0.8892, 1.1814]; statistically tied (11.08% slower to 18.14% faster)	0.0460x [0.0398, 0.0533]
adversarial	0.696646	0.887750	0.156479	0.821000	0.8439x [0.7290, 0.9714]	1.0235x [0.8320, 1.2758]; statistically tied (16.80% slower to 27.58% faster)	0.1911x [0.1562, 0.2374]
float64-dense	0.591938	3.442521	0.213521	0.568875	0.9878x [0.7831, 1.2651]; statistically tied (21.69% slower to 26.51% faster)	4.7212x [3.7338, 5.9358]	0.3296x [0.2668, 0.4035]
strings-regexp	25.614916	25.992854	1.311438	25.341249	1.0039x [0.9898, 1.0171]; statistically tied (1.02% slower to 1.71% faster)	1.0277x [1.0033, 1.0576]	0.0525x [0.0509, 0.0543]
arrays-typed	2.091146	3.797083	0.151188	2.109708	0.9917x [0.9767, 1.0078]; statistically tied (2.33% slower to 0.78% faster)	1.8014x [1.7752, 1.8270]	0.0722x [0.0697, 0.0752]
objects-polymorphic	6.257688	6.330771	0.214750	6.266896	0.9982x [0.9853, 1.0109]; statistically tied (1.47% slower to 1.09% faster)	0.9992x [0.9875, 1.0107]; statistically tied (1.25% slower to 1.07% faster)	0.0356x [0.0338, 0.0380]
calls-recursion-closures	4.708750	5.050916	0.205042	4.725562	1.0058x [0.9893, 1.0263]; statistically tied (1.07% slower to 2.63% faster)	1.0621x [1.0510, 1.0723]	0.0444x [0.0433, 0.0456]
json-codec	97.112083	97.673229	7.242417	97.531042	0.9940x [0.9622, 1.0205]; statistically tied (3.78% slower to 2.05% faster)	0.9797x [0.9465, 1.0059]; statistically tied (5.35% slower to 0.59% faster)	0.0736x [0.0710, 0.0758]
map-set-bigint	22.298000	21.883353	0.908771	21.943000	1.0591x [1.0037, 1.1629]	1.0046x [0.9921, 1.0164]; statistically tied (0.79% slower to 1.64% faster)	0.0446x [0.0419, 0.0485]
exceptions-promises-async	3.930188	2.487624	0.264230	3.936104	1.0069x [0.9947, 1.0211]; statistically tied (0.53% slower to 2.11% faster)	0.6357x [0.6283, 0.6438]	0.0668x [0.0655, 0.0683]

Protocol shared-js-fixed-warmup-v2: one initial call, 64 ten-call warmup batches, then one timed ten-call batch, with result consumption/checksum outside timing. Scripts, inputs and driver match across engines. Every workload has five discarded and 30 retained fresh processes per configuration; five configurations alternate in reversed order. All 4,725 process records are archived. Bootstrap uses 10,000 paired percentile resamples, seed 20260909, sorted indices 249/9749. Timing samples are not filtered by native readiness.

Host: Apple M3; macOS-27.0-arm64-arm-64bit-Mach-O. Bun 1.4.0, default flags. Toolchain: rustc 1.98.1 (48a229cea 2026-09-01); LLVM version: 22.1.8. Release build, no CPU pinning; power policy not recorded. host-compute measures the equivalent pure-JS render kernel, excluding GPUI allocation and snapshots. No Linux, host-reload or throughput result is inferred.

Automatic native coverage and the paired interpreter control:

Workload	Fixed-batch native entries (min–max)	Compilation quiet samples / 30	Current / previous interpreter speed [95% CI]
scalar-control-flow	10–10	27	1.1283x [0.8911, 1.4263]; statistically tied (10.89% slower to 42.63% faster)
scalar-expressions	10–10	29	0.9501x [0.8608, 1.0485]; statistically tied (13.92% slower to 4.85% faster)
host-compute	10–10	22	0.9112x [0.7745, 1.0558]; statistically tied (22.55% slower to 5.58% faster)
mixed-quotes	6990–6990	30	1.0424x [0.9449, 1.1649]; statistically tied (5.51% slower to 16.49% faster)
quickjs-int-arith	10–10	30	1.0195x [0.9725, 1.0672]; statistically tied (2.75% slower to 6.72% faster)
quickjs-bitops	10–10	30	1.0024x [0.9750, 1.0339]; statistically tied (2.50% slower to 3.39% faster)
quickjs-fibonacci	10–10	30	1.0151x [0.9468, 1.0961]; statistically tied (5.32% slower to 9.61% faster)
numeric	10–10	30	1.0409x [0.9492, 1.1814]; statistically tied (5.08% slower to 18.14% faster)
scalar-loop	10–10	30	0.9930x [0.9586, 1.0240]; statistically tied (4.14% slower to 2.40% faster)
call-heavy	10–10	30	1.0188x [0.9699, 1.0855]; statistically tied (3.01% slower to 8.55% faster)
generic-call-entry	10–10	30	0.9813x [0.9273, 1.0436]; statistically tied (7.27% slower to 4.36% faster)
generic-call-fallback	20000–20000	30	1.0151x [0.9784, 1.0656]; statistically tied (2.16% slower to 6.56% faster)
property-heavy	10–10	30	1.0084x [0.9596, 1.0671]; statistically tied (4.04% slower to 6.71% faster)
fibonacci-iterative	10–10	30	1.0550x [0.9783, 1.1649]; statistically tied (2.17% slower to 16.49% faster)
fibonacci-recursive	0–0	30	1.1250x [1.0216, 1.2639]
collections	0–0	30	0.9737x [0.9281, 1.0096]; statistically tied (7.19% slower to 0.96% faster)
strings-json	0–0	30	0.9942x [0.9153, 1.0706]; statistically tied (8.47% slower to 7.06% faster)
calls-closures	0–0	30	0.9399x [0.8396, 1.0190]; statistically tied (16.04% slower to 1.90% faster)
adversarial	0–0	30	0.9975x [0.8423, 1.1742]; statistically tied (15.77% slower to 17.42% faster)
float64-dense	10–10	30	0.9836x [0.8690, 1.1273]; statistically tied (13.10% slower to 12.73% faster)
strings-regexp	0–0	30	0.9765x [0.9506, 0.9961]
arrays-typed	20–20	30	1.0233x [0.9860, 1.0882]; statistically tied (1.40% slower to 8.82% faster)
objects-polymorphic	0–0	30	1.0185x [0.9941, 1.0527]; statistically tied (0.59% slower to 5.27% faster)
calls-recursion-closures	0–0	30	1.0101x [0.9973, 1.0242]; statistically tied (0.27% slower to 2.42% faster)
json-codec	0–0	30	1.0048x [0.9955, 1.0148]; statistically tied (0.45% slower to 1.48% faster)
map-set-bigint	0–0	30	1.0439x [0.9959, 1.1260]; statistically tied (0.41% slower to 12.60% faster)
exceptions-promises-async	0–0	30	0.9865x [0.9741, 0.9981]

All raw records and source/build inputs, all configuration comparisons, verified hashes, validation and provenance, and remaining architecture and performance gates.

Historical 27-scenario macOS matrix, 2026-09-10: CFG semantic SSA candidate. These measurements predate cyclic representation proofs, semantic updates, guard-branch elimination and semantic bitwise operations. Results below apply only to the archived CFG snapshot.

This frozen candidate adds CFG ValueIds/Phi, numeric and comparison operands, pre-effect FrameState bindings, explicit numeric checks and CFG type facts to 023a220. The archived source snapshot and binary hash identify the measured candidate. General inlining, heap facts/LICM, array range elimination, inline frame chains and full lazy materialization remain unfinished.

The numeric-foundation no-regression gate was unmet in this historical run. Automatic runtime slowdowns with intervals entirely below parity occur in: host-compute, numeric, scalar-loop, fibonacci-iterative. Interpreter controls are included below; these observations alone do not isolate the cause.

Timings are medians in ms per ten workload calls. Speed is reference latency / candidate latency with paired geometric means and 95% confidence intervals; values above 1x are faster. QuickJS uses the candidate binary with JIT detached; quickjs-jit uses production automatic tiering; Bun uses defaults. All scenarios, including losing, fallback-only and inconclusive results, remain in the matrix.

Workload	Previous JIT ms	QuickJS ms	Bun ms	quickjs-jit ms	JIT / previous speed [95% CI]	JIT / QuickJS speed [95% CI]	JIT / Bun speed [95% CI]
scalar-control-flow	0.047625	1.160646	0.077291	0.061937	0.8383x [0.7571, 1.0136]; statistically tied (24.29% slower to 1.36% faster)	19.8512x [18.6734, 22.1377]	1.2708x [1.2452, 1.2993]
scalar-expressions	0.073875	0.839563	0.080312	0.059875	1.3634x [1.2484, 1.5204]	14.4407x [13.8821, 15.1259]	1.4089x [1.3345, 1.4956]
host-compute	0.060874	1.209938	0.211334	0.087208	0.7316x [0.7016, 0.7868]	14.4300x [13.9294, 15.1862]	2.6348x [2.4572, 2.8484]
mixed-quotes	2.239729	1.605792	0.132916	2.249625	1.0203x [0.9809, 1.0638]; statistically tied (1.91% slower to 6.38% faster)	0.7132x [0.6920, 0.7355]	0.0624x [0.0568, 0.0699]
quickjs-int-arith	1.819750	3.496104	0.123730	1.817584	0.9973x [0.9523, 1.0341]; statistically tied (4.77% slower to 3.41% faster)	1.9082x [1.8223, 1.9880]	0.0685x [0.0656, 0.0723]
quickjs-bitops	0.242020	0.539917	0.135667	0.247188	0.9821x [0.9105, 1.0601]; statistically tied (8.95% slower to 6.01% faster)	2.1240x [1.9988, 2.2299]	0.5678x [0.5204, 0.6261]
quickjs-fibonacci	0.484729	0.656166	0.128876	0.486000	0.9594x [0.8732, 1.0286]; statistically tied (12.68% slower to 2.86% faster)	1.3052x [1.1883, 1.3910]	0.2651x [0.2401, 0.2975]
numeric	0.023416	0.385708	0.114417	0.034146	0.6605x [0.5227, 0.8187]	9.9530x [8.1828, 11.2544]	3.1892x [2.5819, 3.7439]
scalar-loop	0.023667	0.384604	0.110750	0.034146	0.6691x [0.6182, 0.7085]	10.7932x [10.0249, 11.2975]	3.2011x [2.9338, 3.4108]
call-heavy	0.313292	0.985875	0.012562	0.322166	1.0273x [0.9781, 1.0994]; statistically tied (2.19% slower to 9.94% faster)	3.1727x [3.0786, 3.3021]	0.0389x [0.0383, 0.0396]
generic-call-entry	0.249396	0.806771	0.005959	0.245438	1.0313x [1.0054, 1.0632]	3.3086x [3.2103, 3.4208]	0.0253x [0.0241, 0.0266]
generic-call-fallback	2.904542	1.077395	0.011813	2.968459	1.0176x [0.9588, 1.1240]; statistically tied (4.12% slower to 12.40% faster)	0.3779x [0.3591, 0.4076]	0.0040x [0.0039, 0.0043]
property-heavy	0.318625	0.879250	0.010958	0.314854	0.9904x [0.9238, 1.0522]; statistically tied (7.62% slower to 5.22% faster)	2.6964x [2.5242, 2.8221]	0.0361x [0.0318, 0.0432]
fibonacci-iterative	0.876229	21.794958	1.067980	1.623417	0.5540x [0.5345, 0.5801]	13.3025x [13.0494, 13.5606]	0.6728x [0.6430, 0.7067]
fibonacci-recursive	5.661437	6.019083	0.540125	5.624375	0.9973x [0.9667, 1.0316]; statistically tied (3.33% slower to 3.16% faster)	1.0432x [1.0157, 1.0690]	0.0933x [0.0901, 0.0968]
collections	1.201938	1.164208	0.163771	1.162292	1.0012x [0.9576, 1.0422]; statistically tied (4.24% slower to 4.22% faster)	0.9842x [0.9332, 1.0363]; statistically tied (6.68% slower to 3.63% faster)	0.1386x [0.1297, 0.1498]
strings-json	2.305521	2.310125	0.203896	2.283229	1.0194x [0.9959, 1.0449]; statistically tied (0.41% slower to 4.49% faster)	1.0424x [1.0020, 1.0898]	0.0889x [0.0837, 0.0951]
calls-closures	4.076250	4.014083	0.164480	4.015146	1.0245x [0.9965, 1.0565]; statistically tied (0.35% slower to 5.65% faster)	0.9996x [0.9720, 1.0268]; statistically tied (2.80% slower to 2.68% faster)	0.0425x [0.0390, 0.0472]
adversarial	0.649042	0.656312	0.109250	0.649541	0.9932x [0.9731, 1.0072]; statistically tied (2.69% slower to 0.72% faster)	1.0122x [0.9882, 1.0313]; statistically tied (1.18% slower to 3.13% faster)	0.1762x [0.1669, 0.1888]
float64-dense	0.529229	2.311041	0.129416	0.523146	1.0148x [0.9957, 1.0365]; statistically tied (0.43% slower to 3.65% faster)	4.3765x [4.3025, 4.4295]	0.2502x [0.2460, 0.2557]
strings-regexp	25.365938	25.918208	1.301354	25.483291	1.0029x [0.9894, 1.0173]; statistically tied (1.06% slower to 1.73% faster)	1.0271x [1.0096, 1.0500]	0.0513x [0.0504, 0.0524]
arrays-typed	2.098750	3.706729	0.149812	2.083333	1.0157x [1.0015, 1.0350]	1.7830x [1.7712, 1.7948]	0.0736x [0.0712, 0.0765]
objects-polymorphic	6.219521	6.236604	0.201750	6.203583	1.0149x [1.0023, 1.0293]	1.0087x [0.9988, 1.0197]; statistically tied (0.12% slower to 1.97% faster)	0.0357x [0.0337, 0.0383]
calls-recursion-closures	4.642895	4.910125	0.197480	4.656230	0.9969x [0.9905, 1.0034]; statistically tied (0.95% slower to 0.34% faster)	1.0531x [1.0470, 1.0595]	0.0423x [0.0421, 0.0426]
json-codec	99.073521	97.773729	7.214917	97.680063	1.0483x [0.9737, 1.1425]; statistically tied (2.63% slower to 14.25% faster)	0.9983x [0.9478, 1.0632]; statistically tied (5.22% slower to 6.32% faster)	0.0792x [0.0711, 0.0919]
map-set-bigint	21.941520	21.835354	0.894354	21.704854	1.0067x [0.9936, 1.0210]; statistically tied (0.64% slower to 2.10% faster)	1.0157x [0.9979, 1.0387]; statistically tied (0.21% slower to 3.87% faster)	0.0418x [0.0408, 0.0432]
exceptions-promises-async	3.974105	2.495208	0.268896	4.012416	0.9832x [0.9506, 1.0057]; statistically tied (4.94% slower to 0.57% faster)	0.6541x [0.6250, 0.6984]	0.0704x [0.0662, 0.0764]

Protocol shared-js-fixed-warmup-v2: one initial call, 64 ten-call warmup batches, then one timed ten-call batch, with result consumption/checksum outside timing. Scripts, inputs and driver match across engines. Every workload has five discarded and 30 retained fresh processes per configuration; five configurations alternate in reversed order. All 4,725 process records are archived. Bootstrap uses 10,000 paired percentile resamples, seed 20260909, sorted indices 249/9749. Timing samples are not filtered by native readiness.

Host: Apple M3; macOS-27.0-arm64-arm-64bit-Mach-O. Bun 1.4.0, default flags. Toolchain: rustc 1.98.1 (48a229cea 2026-09-01); LLVM version: 22.1.8. Release build, no CPU pinning; power policy not recorded. host-compute measures the equivalent pure-JS render kernel, excluding GPUI allocation and snapshots. No Linux, host-reload or throughput result is inferred.

Automatic native coverage and the paired interpreter control:

Workload	Fixed-batch native entries (min–max)	Compilation quiet samples / 30	Current / previous interpreter speed [95% CI]
scalar-control-flow	10–10	30	0.9738x [0.8623, 1.0633]; statistically tied (13.77% slower to 6.33% faster)
scalar-expressions	10–10	30	0.9997x [0.9665, 1.0348]; statistically tied (3.35% slower to 3.48% faster)
host-compute	10–10	30	0.9880x [0.9352, 1.0343]; statistically tied (6.48% slower to 3.43% faster)
mixed-quotes	6990–6990	30	0.9948x [0.9677, 1.0218]; statistically tied (3.23% slower to 2.18% faster)
quickjs-int-arith	10–10	30	0.9931x [0.9616, 1.0241]; statistically tied (3.84% slower to 2.41% faster)
quickjs-bitops	10–10	30	0.9887x [0.9621, 1.0115]; statistically tied (3.79% slower to 1.15% faster)
quickjs-fibonacci	10–10	30	0.9811x [0.9423, 1.0188]; statistically tied (5.77% slower to 1.88% faster)
numeric	10–10	29	1.0096x [0.9950, 1.0258]; statistically tied (0.50% slower to 2.58% faster)
scalar-loop	10–10	30	1.0189x [0.9993, 1.0406]; statistically tied (0.07% slower to 4.06% faster)
call-heavy	10–10	30	0.9806x [0.9449, 1.0138]; statistically tied (5.51% slower to 1.38% faster)
generic-call-entry	10–10	30	1.0020x [0.9665, 1.0357]; statistically tied (3.35% slower to 3.57% faster)
generic-call-fallback	20000–20000	30	0.9576x [0.8855, 1.0077]; statistically tied (11.45% slower to 0.77% faster)
property-heavy	10–10	30	1.0077x [0.9905, 1.0255]; statistically tied (0.95% slower to 2.55% faster)
fibonacci-iterative	10–10	30	1.0036x [0.9842, 1.0303]; statistically tied (1.58% slower to 3.03% faster)
fibonacci-recursive	0–0	30	0.9978x [0.9822, 1.0138]; statistically tied (1.78% slower to 1.38% faster)
collections	0–0	30	1.0072x [0.9731, 1.0381]; statistically tied (2.69% slower to 3.81% faster)
strings-json	0–0	30	0.9809x [0.9370, 1.0237]; statistically tied (6.30% slower to 2.37% faster)
calls-closures	0–0	30	0.9989x [0.9817, 1.0161]; statistically tied (1.83% slower to 1.61% faster)
adversarial	0–0	30	1.0110x [1.0003, 1.0228]
float64-dense	10–10	30	1.0112x [0.9991, 1.0249]; statistically tied (0.09% slower to 2.49% faster)
strings-regexp	0–0	30	0.9865x [0.9584, 1.0115]; statistically tied (4.16% slower to 1.15% faster)
arrays-typed	20–20	30	1.0146x [1.0086, 1.0207]
objects-polymorphic	0–0	30	0.9998x [0.9875, 1.0137]; statistically tied (1.25% slower to 1.37% faster)
calls-recursion-closures	0–0	30	1.0044x [0.9984, 1.0101]; statistically tied (0.16% slower to 1.01% faster)
json-codec	0–0	30	1.0297x [0.9563, 1.1066]; statistically tied (4.37% slower to 10.66% faster)
map-set-bigint	0–0	30	0.9848x [0.9553, 1.0066]; statistically tied (4.47% slower to 0.66% faster)
exceptions-promises-async	0–0	30	0.9564x [0.8961, 1.0019]; statistically tied (10.39% slower to 0.19% faster)

All raw records and source/build inputs, all configuration comparisons, verified hashes, validation and provenance, and remaining architecture and performance gates.

Historical 26-scenario macOS matrix, 2026-09-10: first intra-block numeric step. This older candidate contains intra-block expression reuse only; it predates the CFG/Phi, FrameState, numeric-check and comparison changes measured above. Its timings and conclusions apply only to that archived source snapshot.

The no-regression gate was unmet in that run. collections, arrays-typed and calls-recursion-closures are 2.39%, 1.37% and 0.62% slower than the previous binary in this run; their intervals exclude equal speed. collections and calls-recursion-closures have no timed native entries. The interpreter control also slows for collections (0.9774x [0.9707, 0.9848]) and arrays (0.9900x [0.9869, 0.9935]); the source of these differences remains unresolved. The new scalar workload is statistically tied with the previous JIT, between 9.84% slower and 3.29% faster. No end-to-end gain is attributed to this change. The three main targets remain below 0.5x Bun, and generic fallback remains slower than its own interpreter.

Timings are medians in ms per ten workload calls. Ratios are paired geometric means of reference latency / candidate latency, with 95% confidence intervals; values above 1x are faster. QuickJS 0.15.1 runs with JIT detached in the candidate binary; quickjs-jit uses production automatic tiering. Bun 1.4.0 uses default flags. Every row, including fallback-only and slower results, is retained.

Workload	Previous JIT ms	QuickJS ms	Bun ms	quickjs-jit ms	JIT / previous speed [95% CI]	JIT / QuickJS speed [95% CI]	JIT / Bun speed [95% CI]
scalar-expressions	0.072042	0.834854	0.075979	0.071938	0.9691x [0.9016, 1.0329]; statistically tied (9.84% slower to 3.29% faster)	11.4345x [10.5485, 12.3586]	1.1152x [1.0146, 1.2901]
host-compute	0.059916	1.212730	0.191792	0.059604	0.9799x [0.9357, 1.0167]; statistically tied (6.43% slower to 1.67% faster)	20.4279x [19.3558, 21.6708]	3.1904x [3.0398, 3.3336]
mixed-quotes	2.213104	1.585625	0.126750	2.199521	1.0214x [0.9718, 1.0811]; statistically tied (2.82% slower to 8.11% faster)	0.7122x [0.6897, 0.7334]	0.0613x [0.0565, 0.0675]
quickjs-int-arith	1.798854	3.426979	0.122605	1.801084	0.9895x [0.9732, 1.0046]; statistically tied (2.68% slower to 0.46% faster)	1.9273x [1.8698, 1.9943]	0.0704x [0.0668, 0.0747]
quickjs-bitops	0.237208	0.528375	0.132729	0.233666	1.0183x [0.9786, 1.0574]; statistically tied (2.14% slower to 5.74% faster)	2.2472x [2.1766, 2.3083]	0.5644x [0.5424, 0.5851]
quickjs-fibonacci	0.471459	0.629875	0.118792	0.471208	1.0085x [0.9953, 1.0274]; statistically tied (0.47% slower to 2.74% faster)	1.3407x [1.3313, 1.3503]	0.2558x [0.2528, 0.2591]
numeric	0.023625	0.385563	0.112354	0.023688	0.9907x [0.9781, 1.0024]; statistically tied (2.19% slower to 0.24% faster)	16.5390x [16.1446, 17.0668]	4.8225x [4.6808, 4.9852]
scalar-loop	0.023479	0.386605	0.111812	0.023708	1.0013x [0.9841, 1.0226]; statistically tied (1.59% slower to 2.26% faster)	16.4648x [16.2284, 16.7598]	4.7744x [4.6781, 4.8888]
call-heavy	0.307042	0.968938	0.012375	0.311208	0.9775x [0.9492, 1.0003]; statistically tied (5.08% slower to 0.03% faster)	3.0648x [2.9831, 3.1319]	0.0397x [0.0384, 0.0411]
generic-call-entry	0.240458	0.790833	0.005916	0.240541	1.0023x [0.9706, 1.0313]; statistically tied (2.94% slower to 3.13% faster)	3.2422x [3.1513, 3.3091]	0.0253x [0.0239, 0.0275]
generic-call-fallback	2.966208	1.092854	0.011792	2.924708	1.0033x [0.9605, 1.0482]; statistically tied (3.95% slower to 4.82% faster)	0.4208x [0.3585, 0.5509]	0.0041x [0.0038, 0.0044]
property-heavy	0.321146	0.890709	0.011000	0.320875	0.9950x [0.9660, 1.0225]; statistically tied (3.40% slower to 2.25% faster)	2.7735x [2.7034, 2.8384]	0.0342x [0.0331, 0.0351]
fibonacci-iterative	0.864792	21.059604	0.873542	0.856750	1.0029x [0.9943, 1.0089]; statistically tied (0.57% slower to 0.89% faster)	24.7274x [24.2600, 25.4422]	1.0461x [1.0204, 1.0765]
fibonacci-recursive	5.457667	5.750792	0.471042	5.382834	1.0107x [1.0060, 1.0152]	1.0654x [1.0608, 1.0700]	0.0882x [0.0873, 0.0895]
collections	1.119354	1.170083	0.127688	1.145250	0.9761x [0.9708, 0.9811]	1.0157x [1.0099, 1.0210]	0.1142x [0.1098, 0.1201]
strings-json	2.234958	2.244709	0.163771	2.251145	0.9979x [0.9894, 1.0075]; statistically tied (1.06% slower to 0.75% faster)	0.9996x [0.9922, 1.0074]; statistically tied (0.78% slower to 0.74% faster)	0.0733x [0.0725, 0.0744]
calls-closures	3.897166	3.881625	0.126375	3.870709	1.0058x [0.9966, 1.0187]; statistically tied (0.34% slower to 1.87% faster)	1.0019x [0.9945, 1.0101]; statistically tied (0.55% slower to 1.01% faster)	0.0333x [0.0326, 0.0340]
adversarial	0.643084	0.655646	0.107375	0.642813	1.0001x [0.9966, 1.0036]; statistically tied (0.34% slower to 0.36% faster)	1.0195x [1.0157, 1.0234]	0.1673x [0.1662, 0.1686]
float64-dense	0.528292	2.292542	0.128875	0.528438	1.0021x [0.9982, 1.0065]; statistically tied (0.18% slower to 0.65% faster)	4.3521x [4.3326, 4.3728]	0.2443x [0.2436, 0.2451]
strings-regexp	24.327521	24.535333	1.233499	24.321209	0.9985x [0.9929, 1.0038]; statistically tied (0.71% slower to 0.38% faster)	1.0107x [1.0025, 1.0218]	0.0510x [0.0503, 0.0522]
arrays-typed	2.076042	3.749459	0.145646	2.111626	0.9863x [0.9827, 0.9905]	1.7787x [1.7737, 1.7844]	0.0702x [0.0688, 0.0718]
objects-polymorphic	6.137041	6.162750	0.197020	6.173625	0.9968x [0.9929, 1.0003]; statistically tied (0.71% slower to 0.03% faster)	1.0014x [0.9960, 1.0066]; statistically tied (0.40% slower to 0.66% faster)	0.0321x [0.0319, 0.0325]
calls-recursion-closures	4.644209	4.936104	0.196146	4.659541	0.9938x [0.9901, 0.9972]	1.0616x [1.0537, 1.0730]	0.0421x [0.0419, 0.0424]
json-codec	93.563875	93.337458	6.957896	93.361750	1.0021x [0.9957, 1.0083]; statistically tied (0.43% slower to 0.83% faster)	1.0005x [0.9947, 1.0060]; statistically tied (0.53% slower to 0.60% faster)	0.0745x [0.0739, 0.0754]
map-set-bigint	21.420750	21.432750	0.848958	21.434583	0.9999x [0.9980, 1.0021]; statistically tied (0.20% slower to 0.21% faster)	1.0012x [0.9988, 1.0043]; statistically tied (0.12% slower to 0.43% faster)	0.0398x [0.0395, 0.0403]
exceptions-promises-async	3.866312	2.459729	0.248438	3.865896	0.9977x [0.9928, 1.0030]; statistically tied (0.72% slower to 0.30% faster)	0.6346x [0.6310, 0.6383]	0.0640x [0.0637, 0.0644]

Protocol shared-js-fixed-warmup-v2: one initial call, 64 ten-call warmup batches, then one timed ten-call batch, with result consumption/checksum outside timing. All engines use identical scripts, inputs and driver. Each workload has five discarded and 30 retained fresh processes in each of five configurations (previous/current automatic, previous/current interpreter, default Bun), with alternating reversed order. All 4,550 process records are archived. Bootstrap: 10,000 paired percentile resamples, seed 20260909, sorted indices 249/9749.

Apple M3, macOS ARM64 / Darwin 27.0.0; Rust 1.98.1 / LLVM 22.1.8; release build, no CPU pinning, power policy not recorded. host-compute measures the equivalent pure-JS render kernel, excluding GPUI allocation and snapshots. No Linux, host reload or throughput result is inferred from these timings.

Automatic native coverage is reported without conditioning the timing samples:

Workload	Fixed-batch native entries (min–max)	Compilation quiet samples / 30
scalar-expressions	10–10	30
host-compute	10–10	30
mixed-quotes	6990–6990	30
quickjs-int-arith	10–10	30
quickjs-bitops	10–10	30
quickjs-fibonacci	10–10	30
numeric	10–10	30
scalar-loop	10–10	30
call-heavy	10–10	30
generic-call-entry	10–10	30
generic-call-fallback	20000–20000	30
property-heavy	10–10	30
fibonacci-iterative	10–10	30
fibonacci-recursive	0–0	30
collections	0–0	30
strings-json	0–0	30
calls-closures	0–0	30
adversarial	0–0	30
float64-dense	10–10	30
strings-regexp	0–0	30
arrays-typed	20–20	30
objects-polymorphic	0–0	30
calls-recursion-closures	0–0	30
json-codec	0–0	30
map-set-bigint	0–0	30
exceptions-promises-async	0–0	30

All raw records and source/build inputs, all configuration comparisons, verified hashes and machine provenance, and remaining architecture and performance gates.

Historical Linux 24-scenario matrix: candidate4, measured 2026-09-10. Host compute regression remains unresolved: automatic hot reload is 0.9444x baseline speed [0.9373, 0.9518], and interpreter steady-state is 0.9329x [0.9303, 0.9354]. The host regression budget has not passed; this is measured candidate evidence, not an acceptance claim. See the implementation, paired controls, and host findings.

Measured source snapshot: candidate4, binary SHA-256 b758218bdd1d746083f34693a78d42607bc663086fa3e64e2d2f98384d17c2a3. QuickJS and automatic use that same binary; QuickJS means JIT detached. Bun 1.4.0 uses default flags [], executable /home/jason/.bun/bin/bun, SHA-256 33d56b070be6a9e3da0ab013038b43d1645d0534ca811ecdba4472599117eb4b. Host: 13th Gen Intel(R) Core(TM) i7-13700KF; Linux-7.1.9-arch1-2-x86_64-with-glibc2.44; pinned CPU 2; governor powersave; rustc 1.98.1 (48a229cea 2026-09-01). Exact compiler/build, source revision/patch manifests, and host details belong to the linked archive; this renderer does not infer unrecorded source revisions from binary names.

Protocol shared-js-fixed-warmup-v2: one initial call, then 64 batches of ten calls, followed by one measured ten-call batch. All engines run the same driver, input policy, sequential Promise completion, and result storage; checksums run after timing. Driver SHA-256 60408ad6d28c58e7f20161c17a5db11fc142685fdd07ec7dbf35d7bb02a8f7c5. Each row has five discarded processes/config and 30 interleaved retained fresh-process pairs/config. Absolute timings are medians in ms per ten workload calls. Speed ratios are geometric means of paired reference-latency/automatic-latency ratios, with deterministic 10,000 paired percentile-bootstrap resamples (seed 20260909); they are not ratios of the displayed medians. Above 1x is faster than the named baseline. Intervals crossing 1x are statistically tied, with the slower/faster range shown explicitly.

These are fixed-warmup latencies. Samples are not excluded because they lack native entries, compile during the window, fall back, or run slowly. QuickJS includes one outer Rust call/Promise bridge; both engines include shared driver array/closure work. Fixed counters surround the batch before checksum execution and aggregate driver plus workload, not a named function. The quiet count means zero pending jobs/snapshot bytes at both endpoints and no observed installation or compile-time increment, not proof of an engine's peak tier. Native counts do not establish Bun's tier.

Readiness diagnostics are collected later only for QuickJS; their timings and cumulative counters must not replace fixed-window evidence or be divided into Bun's differently conditioned latency. This latency matrix contains no JS throughput windows, matched settled-state Bun samples, per-call P99, or host snapshot measurement. Those requirements remain separate; mixed-quotes is a pure JS kernel, not GPUI snapshot timing. Legacy gate reporting rejects this protocol. Raw evidence and machine-readable summary.

Workload	QuickJS ms	Bun ms	automatic ms	automatic / QuickJS speed [95% CI]	automatic / Bun speed [95% CI]
mixed-quotes	2.349774	0.147462	2.736590	0.8595x [0.8568, 0.8627]	0.0542x [0.0536, 0.0549]
quickjs-int-arith	5.357147	0.131634	1.634172	3.2921x [3.2831, 3.3018]	0.0831x [0.0811, 0.0855]
quickjs-bitops	1.144004	0.152575	0.205012	5.5814x [5.5525, 5.6111]	0.7575x [0.7399, 0.7778]
quickjs-fibonacci	0.839358	0.144318	0.359949	2.3896x [2.3486, 2.4372]	0.4116x [0.4054, 0.4184]
numeric	0.488411	0.129795	0.021608	22.4911x [22.2385, 22.7115]	6.1432x [5.9995, 6.3035]
scalar-loop	0.489399	0.132878	0.021566	22.2350x [21.4209, 22.7450]	6.0842x [5.8316, 6.3037]
call-heavy	1.261803	0.006315	0.312781	4.0333x [3.9902, 4.0812]	0.0203x [0.0202, 0.0205]
generic-call-entry	0.934287	0.004954	0.231675	4.0356x [4.0086, 4.0622]	0.0225x [0.0213, 0.0251]
generic-call-fallback	1.274439	0.008116	3.786595	0.3376x [0.3360, 0.3391]	0.0022x [0.0021, 0.0022]
property-heavy	1.029746	0.007804	0.325760	3.1635x [3.1323, 3.1971]	0.0243x [0.0238, 0.0252]
fibonacci-iterative	29.467614	0.936226	0.878952	33.7586x [33.3822, 34.1502]	1.0773x [1.0613, 1.0940]
fibonacci-recursive	9.865694	0.549300	9.895509	0.9992x [0.9965, 1.0023]; statistically tied; between 0.3% slower and 0.2% faster	0.0555x [0.0549, 0.0560]
collections	1.575133	0.137655	1.576190	1.0005x [0.9968, 1.0055]; statistically tied; between 0.3% slower and 0.5% faster	0.0877x [0.0871, 0.0883]
strings-json	1.972826	0.187261	2.025572	0.9753x [0.9704, 0.9801]	0.0901x [0.0879, 0.0923]
calls-closures	3.489562	0.228020	3.519299	0.9897x [0.9845, 0.9946]	0.0619x [0.0596, 0.0640]
adversarial	0.914932	0.125981	0.916971	1.0023x [0.9906, 1.0145]; statistically tied; between 0.9% slower and 1.4% faster	0.1394x [0.1361, 0.1434]
float64-dense	2.281055	0.148484	0.362700	6.3050x [6.2579, 6.3595]	0.4153x [0.4097, 0.4221]
strings-regexp	19.591037	1.309532	20.241909	0.9663x [0.9622, 0.9690]	0.0648x [0.0643, 0.0652]
arrays-typed	4.160535	0.138155	2.534226	1.6407x [1.6359, 1.6455]	0.0548x [0.0544, 0.0552]
objects-polymorphic	6.268702	0.202083	6.652939	0.9425x [0.9385, 0.9462]	0.0304x [0.0303, 0.0306]
calls-recursion-closures	6.464855	0.206337	6.552003	0.9888x [0.9854, 0.9927]	0.0317x [0.0315, 0.0321]
json-codec	77.302303	7.727146	78.995999	0.9843x [0.9758, 0.9981]	0.0978x [0.0974, 0.0981]
map-set-bigint	17.486868	0.971720	16.617827	1.0430x [1.0259, 1.0590]	0.0591x [0.0584, 0.0598]
exceptions-promises-async	2.323206	0.288878	3.487680	0.6624x [0.6530, 0.6694]	0.0825x [0.0811, 0.0838]

Separate QuickJS execution diagnostics

Workload	quiet fixed batches / 30	native entry delta	Tier2 entry delta	later readiness median ms
mixed-quotes	30	6990	6980	2.752615
quickjs-int-arith	30	10	10	1.618791
quickjs-bitops	30	10	10	0.204897
quickjs-fibonacci	30	10	0	0.371977
numeric	30	10	10	0.021836
scalar-loop	30	10	10	0.021813
call-heavy	30	10	10	0.313397
generic-call-entry	30	10	10	0.225981
generic-call-fallback	30	20000	20000	3.785854
property-heavy	30	10	10	0.319246
fibonacci-iterative	30	10	10	0.880027
fibonacci-recursive	30	0	0	9.921852
collections	30	0	0	1.586696
strings-json	30	0	0	2.031513
calls-closures	30	0	0	3.521440
adversarial	30	0	0	0.935643
float64-dense	30	10	10	0.364465
strings-regexp	30	0	0	20.256235
arrays-typed	30	20	20	2.535597
objects-polymorphic	30	0	0	6.687356
calls-recursion-closures	30	0	0	6.534747
json-codec	30	0	0	79.607234
map-set-bigint	30	0	0	16.940369
exceptions-promises-async	30	0	0	3.538708

All 24 scenarios are included. Zero-entry rows measure attached-runtime fallback; aggregate native/Tier2 counts do not identify the tier of an individual callee. The original branch-leaf call probe can now use compiled caller plus direct ABI; generic-call-fallback preserves effectful non-direct calls and native entries. It does not prove a compiled caller invokes the C CALL helper on every iteration.

Historical focused baseline comparison, 2026-09-10 (Apple M3, macOS ARM64): Eight pure-JS controls compare old runtime 928a8f2 with merged #26 023a220, built with the identical Cargo.lock and benchmark harness. host-compute reproduces the archived host render kernel and text result. That baseline comparison did not change runtime code. This is a different host from the historical 24-scenario matrix above. The complete current macOS matrix is above; native Linux host regression acceptance remains deferred.

Workload	Old JIT ms	QuickJS ms	Bun ms	automatic ms	automatic / old JIT speed [95% CI]	automatic / QuickJS speed [95% CI]	automatic / Bun speed [95% CI]
host-compute	0.061167	1.217812	0.200042	0.060688	1.0184x [0.9965, 1.0416]; statistically tied (0.35% slower to 4.16% faster)	20.2439x [19.8452, 20.6762]	3.7467x [3.3598, 4.2957]
generic-call-entry	5.170479	0.788021	0.005834	0.246062	19.4033x [17.4142, 21.0883]	2.9601x [2.6310, 3.2293]	0.0219x [0.0195, 0.0239]
generic-call-fallback	3.470271	1.067395	0.011792	2.886375	1.2263x [1.1896, 1.2789]	0.3644x [0.3586, 0.3695]	0.0041x [0.0039, 0.0045]
call-heavy	0.311833	0.967584	0.012500	0.310917	0.9989x [0.9835, 1.0131]; statistically tied (1.65% slower to 1.31% faster)	3.1262x [3.0649, 3.2045]	0.0402x [0.0395, 0.0409]
property-heavy	2.457979	0.884292	0.010937	0.318604	7.5748x [7.2831, 7.7730]	2.7268x [2.6169, 2.8091]	0.0347x [0.0330, 0.0363]
arrays-typed	6.687042	3.786563	0.159188	2.086562	3.2116x [3.1561, 3.2903]	1.8141x [1.7870, 1.8406]	0.0766x [0.0732, 0.0811]
scalar-loop	0.023959	0.385354	0.112833	0.023645	1.0018x [0.9853, 1.0181]; statistically tied (1.47% slower to 1.81% faster)	16.2007x [15.8834, 16.5010]	4.8451x [4.6780, 5.0425]
mixed-quotes	2.319021	1.574021	0.116854	2.190834	1.0756x [1.0452, 1.1177]	0.7293x [0.7007, 0.7766]	0.0551x [0.0530, 0.0575]

Units are ms per ten calls, using shared-js-fixed-warmup-v2: one initial call, 64 ten-call warmup batches, then one timed ten-call batch with checksum outside timing. Each of five configurations has five discarded and 30 retained fresh processes per workload; all 1,400 process records are archived, including slower samples. Order reverses on alternating pairs. Bun 1.4.0 uses default flags; QuickJS uses the candidate binary with JIT detached, and the JIT column uses production automatic tiering. Ratios are paired geometric means with 10,000 percentile-bootstrap resamples, seed 20260909 (sorted indices 249 and 9749). Rust 1.98.1 / LLVM 22.1.8, release build, shared Apple M3 on Darwin 27.0.0, no CPU pinning, power policy not recorded for this run. No throughput or host snapshot/reload result is inferred from these kernel latencies.

The host-compute interpreter comparison is statistically tied with the old runtime: 1.0616x [0.9898, 1.1874], between 1.02% slower and 18.74% faster. This does not resolve the native Linux regression. Generic fallback and mixed quotes remain slower than their own interpreter; the three target workloads remain below 0.5x Bun. All raw samples and build inputs, all configuration summaries, source/binary hashes and initial probe history, and remaining architecture work.

Historical matrix: the 2026-09-06 47aeb11 results used different warmup and host timing boundaries and must not be divided into these v2 timings to claim an optimization gain. The historical full table and tier diagnostics, raw samples, versions, flags, and reproduction, and derived intervals remain available. Benchmark commands and protocol details.

Community development

This crate doesn't aim to provide system and web APIs. The QuickJS library is close to V8 in that regard. If you need APIs from WinterGC or Node, then you can take a look at the follow community projects:

AWS LLRT Modules: Collection of modules that micmic some of the Node APIs in pure Rust
Rquickjs Extra: Collection of modules that complement AWS LLRT Modules in pure Rust

The community has also built various utilities which might be relevant to you:

Rquickjs Serde: Serde serializer and deserializer for rquickjs Value

Development status

This bindings is feature complete, mostly stable and ready to use. The error handling is only thing which may change in the future. Some experimental features like parallel may not works as expected. Use it for your own risk.

Supported platforms

Rquickjs needs to compile a C-library which has it's own limitation on supported platforms, furthermore it needs to generate bindings for that platform. As a result rquickjs might not compile on all platforms which rust supports. In general you can allways try to compile rquickjs with the bindgen feature, this should work for most platforms. Rquickjs ships bindings for a limited set of platforms, for these platforms you don't have to enable the bindgen feature. See below for a list of supported platforms.

platform	shipped bindings	tested	supported by quickjs

x86_64-unknown-linux-gnu	✅	✅	✅
i686-unknown-linux-gnu	✅	✅	✅
aarch64-unknown-linux-gnu	✅	✅	✅
loongarch64-unknown-linux-gnu	✅	✅	✅
x86_64-unknown-linux-musl	✅	✅	✅
aarch64-unknown-linux-musl	✅	✅	✅
loongarch64-unknown-linux-musl	✅	✅	✅
x86_64-pc-windows-gnu	✅	✅	✅
i686-pc-windows-gnu	✅	✅	✅
x86_64-pc-windows-msvc	✅	✅	❌ experimental!
aarch64-pc-windows-msvc	✅	❌	❌ experimental!
x86_64-apple-darwin	✅	✅	✅
aarch64-apple-darwin	✅	❌	✅
wasm32-wasip1	✅	✅	✅
wasm32-wasip2	✅	✅	✅
other	❌	❌	Unknown

License

This library is licensed under the MIT License

Name		Name	Last commit message	Last commit date
.github		.github
.superpowers/sdd/2026-08-29-quickjs-tiered-jit		.superpowers/sdd/2026-08-29-quickjs-tiered-jit
benchmarks		benchmarks
core		core
docs		docs
examples		examples
integrations/gpui-component		integrations/gpui-component
jit		jit
macro		macro
scripts		scripts
src		src
sys		sys
tests		tests
.clippy.toml		.clippy.toml
.gitignore		.gitignore
.gitmodules		.gitmodules
.rustfmt.toml		.rustfmt.toml
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
Cargo.toml		Cargo.toml
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
SECURITY.md		SECURITY.md
codecov.yml		codecov.yml

Navigation Menu

Repository files navigation

quickjs-jit

Main features of QuickJS

Features provided by this crate

Experimental JIT performance

Separate QuickJS execution diagnostics

Community development

Development status

Supported platforms

License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

quickjs-jit

Main features of QuickJS

Features provided by this crate

Experimental JIT performance

Separate QuickJS execution diagnostics

Community development

Development status

Supported platforms

License

About

Resources

Security policy

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages