docs: record the measured #7109 timing in the #7121 changelog fragment

[proggeramlug]

Follow-up to #7121 (#7109). Documentation only — six lines in the changelog
fragment, no code.

#7121 shipped with no timing claim, because the host was at load 15–80 all
session and 08_string_concat at its committed 100 000 iterations runs in under
10 ms — below the resolution of its own Date.now() delta. Saying "we did not
measure" was the right call then; the measurement has since been made and
belongs in the release notes rather than in a PR comment.

The measurement

08_string_concat's loop, unchanged in shape, scaled to 3 000 000 iterations so
the in-program timer has something to resolve, and timed by that timer (process
startup excluded). 20 interleaved pairs — alternating the two binaries
rather than running 20 of one then 20 of the other, so a drift in host load
cannot be attributed to the change.

const ITERATIONS = 3000000;
let result = "";
const start = Date.now();
for (let i = 0; i < ITERATIONS; i++) {
    result = result + "x";
}
console.log("elapsed_ms:" + (Date.now() - start));

arm	median	min	distribution
4d3ddc9a3 (base)	74 ms	72 ms	72–79 ms
with #7109	69 ms	68 ms	68–72 ms

−6.8 % median, −5.6 % minimum. The distributions are essentially disjoint,
which is what makes this reportable at load ≈26 — the intra-arm spread is ~9 %
of the median but the arms do not interleave.

The mechanism is visible in the IR and was already in #7121: the top-level
result = result + "x" went from two js_get_string_pointer_unified calls per
iteration to Phase 3a's four-arm tag dispatch whose hot arm derives both handles
with a bare and i64 …, 0xFFFF_FFFF_FFFF.

Why it is stated narrowly

29 of the 39 workloads in #7121's object A/B compile to a byte-identical
.o. #7109 converted 321 module_init_context denials into selections, but
most of those are canonical-i32 promotions that -O3 was already achieving —
under the parallel-shadow model every LocalGet of such a local already read
the i32 slot, so the double slot was dead code.

So this is a number about the top-level += self-append idiom, which is the
one lowering that changes and survives optimization. It is deliberately not
presented as "#7109 is worth −6.8 %".

Verification

Both binaries built from the same runtime archives (libperry_runtime.a /
libperry_stdlib.a unchanged between the two commits — #7109 touches
perry-codegen only), PERRY_NO_AUTO_OPTIMIZE=1, --no-cache, and the two
linked binaries confirmed to differ by hash before timing, so this is not an
A/B of one binary against itself.

No code, no test, and no gate changes. changelog.d/ fragments are folded into
the release notes at tag time.

Summary by CodeRabbit

• Performance
  • Reported a measured improvement for a targeted module initialization scenario, reducing runtime from 74 ms to 69 ms in a 3,000,000-iteration benchmark.
  • Documented timing ranges and noted that 29 of 39 tested workloads produce byte-identical results.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 932dd45c-6534-46b7-85c4-7cc1259ea5f6

📥 Commits

Reviewing files that changed from the base of the PR and between 95988fb and d814cb6.

📒 Files selected for processing (1)

• changelog.d/7121-module-init-repsel.md

---

📝 Walkthrough

Walkthrough

The changelog adds measured performance results for the module-init canonical-representation change. It records a 3,000,000-iteration improvement from 74 ms to 69 ms and notes output identity across 39 workloads.

Changes

Module-init benchmark documentation

Layer / File(s)	Summary
Benchmark results changelog.d/7121-module-init-repsel.md	Documents the benchmark method, timing distributions, observed speedup, and the 29 of 39 workloads that produce byte-identical objects.

Estimated code review effort: 1 (Trivial) | ~2 minutes

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch docs/7123-7121-timing-followup

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[proggeramlug]

Independent reconciliation of this PR's number, from #7128 — it holds up.

Measured on a Raspberry Pi 5 at load 0.16 with perf stat -e instructions:u
(11 repeats), plus a Mac mini M1 for a wall-time sign check. Same arms as here
(4d3ddc9a3 vs #7109), built from one target dir in one session.

measurement	program	metric	result
this PR	3M, in-program timer (startup excluded), load ≈26	wall	−6.8% median
#7128, Pi, load 0.16	3M, whole process	instructions	−4.21%
#7128, Pi, load 0.16	3M, whole process	wall	−3.4%
#7128, mini, load ~4–8	3M, whole process	wall	−5.1%
#7128, Pi, load 0.16	committed 100k, whole process	instructions	−4.12%

These agree; the spread is almost entirely whether process startup is in the
denominator. The committed 100k program is ~24 ms wall on the mini and
essentially all startup, so the mini's 3M numbers net of startup are ~51 vs
~55 ms — −7.3%, bracketing the −6.8% here. Your figure is right for
loop-only wall; −4.2% is the instruction cost actually removed. Time falls
further than instruction count because the change removes a call
(js_get_string_pointer_unified) whose cost exceeds its instruction count.

Two additions:

1. The effect survives at the committed 100,000-iteration size — −4.12%
   instructions on the unmodified committed file, no scaling. Wall clock reads
   21 ms in both arms there, which is exactly why you needed the 30× scaling for
   a timer; instructions resolve it without changing the program's working set
   or cache behaviour. So the scaling was a legitimate technique and it did not
   manufacture the effect.
2. It is canonical Str and nothing else — the 4d3ddc9a3 arm and a
   PERRY_CANONICAL_STR_LOCALS=0 arm of main agree to within 900 instructions
   out of 926 million.

One caveat for anyone reusing that knob: PERRY_CANONICAL_STR_LOCALS=0 is not
scoped to Str locals — it also gates nanbox_string_ref_boxed, so 24 of 26
census workloads emit differently under it, including ones with
canonical-str = 0. The compiler-arm A/B you used is the isolation that is
actually valid here.

Cost side, reported separately from speed: canonical Str adds 2,808 bytes
of Mach-O __text on 08_string_concat.