perf: asyncpipe is 1.12x Node but NOT CPU-bound — needs event-loop/timer counters, not a sampler

[proggeramlug]

Summary

asyncpipe runs at 1.12x Node with an instruction ratio of 1.71, but it is
not CPU-bound — a sampling profiler cannot see it. Aggregating sample over 14 runs of
an ~88 ms program yields 1 usable self-time sample. The process is almost always parked,
not running.

What that means

asyncpipe is the async/microtask-heaviest row in the corpus: async/await, Promise.all
fan-out per batch, and a setTimeout(…, 0) yield at every batch boundary
(sweep-artifacts-0819b/sources/asyncpipe.ts, function tick). The wall-clock gap is
therefore in event-loop and timer scheduling latency, not in executing user code.

Perry retires 1.36 G instructions against Node's 0.79 G on this row, so there is extra work
— it is simply spread across the scheduling machinery rather than concentrated anywhere a
sampler will catch.

Suggested approach

Sampling is the wrong instrument here. Instead:

1. Count, don't sample. Instrument the timer/microtask path: how many timer registrations,
   wakeups, and microtask-queue drains happen per batch, and what does each cost? Compare that
   count against what the program logically requires (one setTimeout(0) per batch).
2. Measure the floor. Write a micro-benchmark that does nothing but await new Promise(r => setTimeout(r, 0)) N times, and compare Perry against Node. That isolates per-yield latency
   from the benchmark's real work. If Perry's per-yield cost is materially higher, that alone
   explains the row.
3. Then look at js_promise_run_microtasks, the timer wheel, and the setTimeout(0) path
   specifically — a zero-delay timer should not need a full timer-wheel round trip.

Prior work on this row, so it is not re-derived

• asyncpipe regressed +55.6% in the 2026-08-18 sweep and was fixed by perf(codegen): skip redundant async frame capture tracking #8392 (codegen was
  marking the compiler-private async step closure's whole frame as escaped, so released box
  cells could never publish — published=0 -> 192859). It is now back to its pre-regression
  standing. This ticket is about the remaining 1.12x, which predates that regression.
• fix(async_hooks): track scheduler resource lifecycles #8328 (async_hooks scheduler lifecycles) was measured against this row and is not a
  factor: +0.2%, overlapping distributions. queue_microtask_with_type only serves
  queueMicrotask()/process.nextTick(), neither of which asyncpipe uses. Do not revisit it.

Acceptance

asyncpipe beats Node (< 1.00x), or a written conclusion with counter evidence that the
remaining gap is inherent to the event-loop design. No other row regressed, all 19 byte-exact.

---

How to build and measure (shared setup)

Build with the static wrappers or you link a stale runtime and both A/B arms behave
identically — a vacuous result:

cargo build --release -p perry -p perry-runtime-static -p perry-stdlib-static
export PERRY_RUNTIME_DIR=<your target>/release

The benchmark corpus and Node-oracle expected output are staged at
/Users/amlug/projects/perry/sweep-artifacts-0819b/ (sources/*.ts, expected/*.stdout).

export PERRY_NO_AUTO_OPTIMIZE=1 PERRY_NO_CACHE=1
perry sources/<row>.ts -o /tmp/x --no-cache
cmp <(/tmp/x) expected/<row>.stdout      # verify BEFORE timing
/usr/bin/time -l /tmp/x                  # instructions retired + peak RSS

Profiling: Perry strips its output binaries, so sample attributes everything to
???. Set PERRY_KEEP_SYMBOLS=1 PERRY_DEBUG_SYMBOLS=1 at compile time — the binary keeps
~1900 symbols and stays byte-exact — then read sample's "Sort by top of stack" section.
These programs are short; aggregate across ~14 runs for a usable sample count.

Constraints that apply to any fix here

• Correctness first: all 19 corpus programs must stay byte-exact against their Node oracle.
• Report wall time AND instructions retired AND peak RSS together. The project rule is
  minimize RSS and keep best compute — never trade one for the other silently.
• cargo test --release -p perry-runtime --lib, -p perry --bin perry, and
  bash scripts/run_lint_gates.sh (all 50) must pass.
• A diff-read is a hypothesis, not a result — measure before proposing. (fix(async_hooks): track scheduler resource lifecycles #8328 looked
  obviously guilty on this exact corpus and measured +0.2%.)
• Known unrelated flake: box::release_tests::the_intrusive_free_list_round_trips_a_whole_cohort
  (flaky: the_intrusive_free_list_round_trips_a_whole_cohort asserts a delta on a process-global counter under parallel tests #8401). If it fails once and reruns clean, it is not yours.