ant
javascript for 🐜's, a tiny runtime with big ambitions antjs.org
Star
Fork
Watch
ant docs exec-plans completed async-error-performance.md
4.0 kB · 83 lines
Markdown
at master

Async Error Boundary Performance #

Status: completed Last reviewed: 2026-09-17 Owner: theMackabu

Goal #

Keep the handled-error fixes while eliminating measurable successful-operation overhead. Retain installed ant as the user's comparison baseline and use a matched HEAD build to distinguish source effects from build/profile differences.

Evidence and decisions #

  • Original candidate regressed readable sizing by roughly 7-8% versus installed Ant. A clean HEAD control built with the same flags, PGO profile, dependency objects, and generated headers confirms a readable regression from the patch.
  • Changed error control flow invalidates hot readable PGO records: enqueue loses up to 27.6 million counts and pull-if-needed loses up to 138 million. Disassembly shows changed inlining of repeated internal slot access.
  • js_to_number returns NaN when coercion throws. Use the existing invalid-size branch to consume that exception; successful numeric and object conversions need no additional pending-exception check.
  • Benchmarks stay outside PGO training. Do not remove correctness coverage or weaken exception handling to improve timings.

Outcome #

Successful size conversion uses the original error/type checks and existing invalid-size test. Pending coercion errors are consumed only after that test detects NaN. Error values remain rooted across rejection allocation, including when the size function has already changed the writable stream state.

Refreshed the checked-in Darwin ARM64 profile using env -u NO_COLOR ./meson/pgo/build.sh, unmodified. Training completed without timeouts or failed workloads. No final profile hash mismatches remain. The new microbenchmark was not part of training.

Disassembly confirms restored slot-lookup inlining: public readable enqueue had 3 js_get_slot callsites in clean HEAD, 9 in the original patch, and 3 after refresh. These are static callsites, not executed instruction counts.

Measurements #

All comparisons used pinned binaries, equal work, warmup, checked outputs, and serial ABBA blocks. Other applications remained active on the host.

Comparison Readable numeric size Readable object size
Initial patch vs installed Ant +6.9% +8.1%
Final candidate vs installed Ant +1.7% +1.6%
Final candidate vs clean current HEAD -2.8% +0.4%
Matched modified units without PGO -0.24% -0.21%

Positive means longer elapsed time. The profile-disabled experiment rebuilt the same 20 changed translation units on both sides without profile use, keeping all remaining objects identical. All six primary source controls were within 0.5%; a separate events.once comparison was 2.46% faster. This removes the asymmetric effect of rejected profiles when assessing the source changes.

The final installed-binary comparison covers all 12 original workloads. Its unchanged await control is 2.6% slower, so do not claim literal zero difference between whole binaries or assign their residual 1-2% difference solely to the patched paths. Clean HEAD uses its checked-in profile; the final candidate uses the refreshed one. The matched source experiment and restored inlining support removal of the original patch-specific slowdown.

Detailed build commands, profiles, source overlays/diffs, binaries, raw samples, assembly, and all workload results: /tmp/ant-error-perf-1lsan8qr/README.md. Final binary SHA-256: d4c5348af1fc509797bcfa7e0f565f68de72a549f591b109b81f2c08fc04bec3. Final profile SHA-256: dbbcab7bbb369394ecc9d97d82f0630fc141df92adcf42ef5a84b91982e442ac.

Validation #

  • Native build and all 16 Meson tests passed.
  • Full spec suite: 4,240 assertions and 102 files passed, zero failures.
  • Four handled-error suites, callback exception coverage, and focused fetch and writable-stream tests passed.
  • Added coercion regressions for valueOf/Symbol.toPrimitive throwing Error objects or primitives, plus nonthrowing NaN/Infinity sizes.
  • Preflight and whitespace checks passed.