Native calls and fulfilled await performance #

Status: completed Last reviewed: 2026-09-18 Owner: theMackabu

Goal #

Recover the remaining Math.min, Array.pop, and fulfilled-promise control regressions after the exception-record migration. Preserve the 64-bit native/JIT return ABI, exception ownership, and pending-error promotion.

Outcome #

The candidate skips general number conversion for numeric Math.min arguments, uses the already-validated dense array pointer throughout pop, and queues a fulfilled await directly when the promise has no reactions, active trigger, processing batch, or parent chain. Await still takes a microtask turn and retains the awaited promise for coroutine ownership and generator state.

Only src/ant.c and src/modules/math.c changed in runtime code. Native dispatch, exception records, pending-error checks, and the 64-bit result ABI remain unchanged. Isolated ABBA/BAAB runs (four samples per binary, same existing profile) show Math.min -9.72% and pop -12.60% time. A separately trained final profile has no runtime profile mismatches.

The fulfilled-promise control improves 11.64% at the original 250,000 iterations and 11.00% at 1,000,000 iterations. Fulfilled async iteration improves 13.57%; async-next cases improve 5–6%; stream cases stay flat. Direct resume jobs also preserve an intervening microtask between two awaits of the same eligible promise, matching Node instead of batching both awaits ahead of that microtask. The final test uses an async rejection handler with explicit failure exit: it exits 1 on the pinned pre-change binary and passes on Node and final Ant.

Final qualification #

Pinned installed 15.1.59a2d6b3.1 versus this dirty tree based on 5903d763, with a separate fresh profile. Four serial processes per binary for targeted cases, in ABBA/BAAB order; two per binary for V8, in ABBA order. No builds, tests, or profilers overlapped timings. Lower time is better.

Target	Installed	Final	Time change
Math.min, ns/op	7.155	6.640	-7.20%
Array.pop, ns/op	18.730	17.360	-7.31%
Fulfilled-promise control, ms	19.269	17.300	-10.22%
Control at 4x iterations, ms	76.226	69.456	-8.88%

Map iteration is 6.77% faster, regex cases are 1.86–7.10% faster, and Game of Life is effectively unchanged. V8 Richards, DeltaBlue, Crypto, RayTrace, and NavierStokes are within 0.7%; EarleyBoyer and RegExp improve about 1%; Splay improves 28.88%. Array push retains an 8.23% gap. The microbench pass covers six rows, not all 43; no new full-microbench aggregate is claimed. Installed build provenance is incomplete, so the final comparison includes build and profile differences as well as source changes. Small deltas remain noisy.

All 4,240 specs in 102 files, 17 focused JS files, and native error-handoff and exact-GC-edge tests pass. New coverage checks numeric errors/signed zero, array fallbacks, await ordering, async-generator requests, queued-value GC with conservative C-stack roots disabled, cancellation, and reference release. Preflight, knowledge, and whitespace checks pass. No opcode, interpreter handler, or compiler emitter changed, so the stack-depth sweep was not repeated.

The checked-in user profile and index were preserved. The main build uses pgo_profile pointing into the artifact directory; all 937 vendor compile commands remain free of profile-use flags.

Artifacts: .cache/native-async-recovery-8p8y499t, including the full report, raw comparisons, source/assembly snapshots, profile, binary identities, and validation logs. The initial premature math comparison is marked INVALID and excluded; math-valid is its replacement.