perf: `str.replace/match/split(regexp)` is 150–210× slower than Node on short strings (spec `flags`/`@@species` Gets run through the generic property path on pristine RegExps)

[proggeramlug]

Found by the package performance audit (real npm packages compiled from source, profiled against Node 26.5.1) and
re-measured on Perry 7661bc0 (v0.5.1589), Linux x64. String.prototype.replace/match/split with an untouched
RegExp spend most of each call reading rx.flags (whose getter performs 8 more observable Gets) and, for split,
rx.constructor[@@species], each through js_reflect_get inside a setjmp trap frame. On a 3-character subject this makes
replace 181×, match 148–212× and split 190× Node, while the matching itself is ~7 % of the time.

Reproduction

bench.ts:

// RegExp-taking String methods on a pristine RegExp, tiny subject (engine work negligible)
const variant = process.argv[2] || "replace"; const N = Number(process.argv[3] || "300000");
const RE_G: any = /\+/g, RE: any = /\+/, s = "a+b", b64 = "ndKK8/0z8Qw+OyczXXftx3rfjSEDizNo8Sb5awzd7Fw=";
const ops: Record<string, () => number> = {
  flags: () => RE_G.flags.length,                       // the getter alone (8 inner Gets)
  global: () => (RE_G.global ? 1 : 0),                  // control: one flag getter
  replace: () => s.replace(RE_G, " ").length,           // reads rx.flags, then matches
  replace_empty: () => s.replace(RE_G, "").length,      // control: canonical removal path skips the flags Get
  replaceAll_str: () => s.replaceAll("+", " ").length,  // control: no RegExp
  match_g: () => s.match(RE_G)!.length,                 // reads rx.flags
  match: () => s.match(RE)!.length,                     // reads rx.flags (non-global)
  test: () => (RE.test(s) ? 1 : 0),                     // control: exec Get skipped (#10166)
  split: () => s.split(RE).length,                      // constructor + @@species + flags + new sticky RegExp
  split_str: () => s.split("+").length,                 // control: string separator
  jwt_chain: () => b64.replace(/=/g, "").replace(/\+/g, "-").replace(/\//g, "_").length,
};
const op = ops[variant];
function run(n: number): number { let acc = 0; for (let i = 0; i < n; i++) acc += op(); return acc; }
run(N / 5 | 0); const t0 = performance.now(); const cs = run(N);
console.log(`variant=${variant} checksum=${cs} ms=${(performance.now() - t0).toFixed(2)}`);

PERRY_NO_AUTO_OPTIMIZE=1 perry compile bench.ts -o bench
node bench.ts replace 200000; ./bench replace 200000     # repeat per variant

Measurements

N = 200,000. Median of 3, shared host (load average ~100+ on 64 threads; instruction counts are the load-independent
figure). "Instructions/iter" = whole-process instructions:u ÷ (1.2·N timed + warm-up iterations).

variant	Node loop ms	Perry loop ms	ratio	Perry instructions/iter	Node wall ms	Perry wall ms
re.flags	5.2	2,201.5	420×	72.5 k	132	2,639
re.global (control)	1.7	202.6	119×	7.0 k	177	282
s.replace(/\+/g, " ")	19.2	3,473.2	181×	117.6 k	148	4,059
s.replace(/\+/g, "") (control)	17.3	587.0	34×	20.1 k	166	744
s.replaceAll("+", " ") (control)	19.4	662.3	34×	26.9 k	163	892
s.match(/\+/g)	20.1	2,984.6	148×	94.7 k	185	3,577
s.match(/\+/)	12.7	2,696.1	212×	90.3 k	157	3,262
/\+/.test(s) (control)	7.2	220.4	30×	6.3 k	110	307
s.split(/\+/)	25.9	4,922.8	190×	155.3 k	175	5,964
s.split("+") (control)	12.1	308.9	26×	12.3 k	193	408
jws base64url chain	98.9	8,615.5	87×	286.6 k	287	10,369

Checksums identical. The same replace with an empty replacement (which already takes a proven-builtin path that skips
the Gets) costs 20 k instructions instead of 118 k; non-global match costs 90 k against 6 k for test on the same RegExp.

Profile (perf record -g, inclusive): replace — flags getter 55 %, get_symbol(@@replace) 5.8 %, engine
execute_output 7 %; match(/g) — flags getter 67 %, get_symbol(@@match) 7.9 %, engine 7.5 %; split — flags getter
40 %, species 22 %, perex_construct::construct 6.8 %. Top self symbols are the generic property path:
try_data_get_bytes, from_utf8, js_reflect_get, get_accessor_descriptor, js_object_get_field_by_name,
exception::try_push_with_kind.

Impact

From the audit profiles (v0.5.1587, strings and objects group reports); "RegExp protocol Gets" was a separate bucket
from regex-engine time:

• date-fns 4.x format: 38.8 % of the mode, ≈34.5 % of the date-fns loop (formatStr.match(longFormattingTokensRegExp) and
  .match(formattingTokensRegExp), both global).
• qs 6.x parse: 29 % (utils.decode: str.replace(/\+/g, ' ') per key and value).
• validator 13.x: isEmail 18.7 %, isURL 10.8 % (isByteLength: encodeURI(str).split(/%..|./)).
• dayjs format: 9.8 % (str.replace(REGEX_FORMAT, fn)).
• jsonwebtoken 9.0.3 sign: js_string_replace_js 40 % of sign, verify 13 % (jwa/jws base64url
  .replace(/=/g,'').replace(/\+/g,'-').replace(/\//g,'_'); includes engine time).
  Group mean across the four string packages: ≈16.7 % of Perry time.

Mechanism

All citations read at 7661bc0 (verified) unless marked.

• crates/perry-runtime/src/regex/perex_replace.rs:69 — RegExp.prototype[@@replace] does dispatch::get(&receiver, b"flags");
  :239 String.prototype.replace first looks up @@replace via dispatch::get_symbol. The only bypass is :218
perex_remove::try_remove, which admits an empty replacement string only (perex_remove.rs:17-35).
• crates/perry-runtime/src/regex/perex_match_search.rs:59-67 match_flags → dispatch::get(receiver, b"flags"), used for
  global and non-global match (:187); :274/:300 look up @@match first.
• crates/perry-runtime/src/regex/perex_split.rs:126-127 — match_all::species then get(flags); :155-165 constructs a new
  splitter RegExp (js_regexp_construct) on every call; :335 looks up @@split.
• crates/perry-runtime/src/regex/match_all.rs:86-101 species: Get(rx, "constructor") + Get(C, @@species).
• crates/perry-runtime/src/regex/perex_match_search.rs:29-51 — the flags getter performs 8 dispatch::gets (hasIndices,
  global, ignoreCase, multiline, dotAll, unicode, unicodeSets, sticky).
• crates/perry-runtime/src/regex/perex_dispatch.rs:52-63 get: a RuntimeHandleScope, key interning
  (canonical_key), an api::caught setjmp trap frame, and js_reflect_get; each flag getter then enters
  object/regex_proto_thunks.rs:755 regexp_get_property inside a second catch_js_throw frame.
• The proof needed to skip all of this already exists: crates/perry-runtime/src/object/regex_canonical.rs:45-84 computes a
  per-prototype-ShapeId Proof { flags: bool, … } (all nine flag accessors are the native thunks), :89-123 exec() admits
  an untouched RegExp, and :127-155 replace() adds the @@replace check — but only perex_remove and
  perex_dispatch.rs:118 consult it; flags, @@match, @@split and @@species are never answered from it.

What fast looks like

When regex_canonical::exec(rx) and its flags proof hold, read global/unicode/sticky straight from the
RegExpHeader (as perex_remove already does), answer @@replace/@@match/@@split and the species constructor from the
same shape proof, and let split reuse the cached non-sticky program without allocating a splitter object. Targets on
this microbenchmark: s.replace(/\+/g, " ") ≤ the empty-replacement control (≈20 k instructions/call), s.match(re) and
s.split(re) within 1.5× of the test / string-split controls, re.flags ≤ 2 k instructions. Getting to ≤2× Node
additionally needs the per-call engine cost tracked in #10166.

Notes

• Distinct from perf(regex): split, replace, matchAll and global exec do quadratic work under Perex — split regressed from ~7x Node to timing out #10165 (quadratic work in split/replace/matchAll), perf(regex): RegExp.prototype.test costs ~1.5 µs per call under Perex — 44x slower than the old engine even with the regex hoisted #10166 (per-call cost of test; its fix skipped only
  the exec Get) and perf(regex): replace with a string template holds ~1 KB per output piece — 545 MB RSS on a 550 KB subject, 4.5x Node #10411 (template replacement memory). Those issues do not cover the flags/@@species/symbol Gets.
• Semantics to preserve: a RegExp subclass, an own flags/exec property, a patched RegExp.prototype getter or
  Symbol.species must still be observed; the ShapeId + symbol-epoch proof in regex_canonical.rs is designed for exactly
  that.
• The string-separator split control is itself 26× here; long string splits are covered separately in perf: str.split(".") with a string separator is 345× slower than Node on a 179-char JWT (Perex KMP reads the subject one UTF-16 unit per bounded-reader call) #10519.

[proggeramlug]

Partial progress: #11543 landed and cuts host-side split / string-template replace overhead on ASCII subjects. encodeURI(s).split(/%..|./) is −43.3% instructions/iter, "a-b-c".replace(/-/g, "+") is −14.7%, and validator/sanitize is −4.2%. None of that is in the matcher.

What remains for this issue:

• The matcher side is Take package-shaped repeats and classes out of the per-character phase loop perex#3, measured on top of perf(regex): cut split and replace host overhead on ASCII subjects #11543: uuid/v4 −48%, jsonwebtoken/decode −52%, uuid/v7 −30%, uuid/v5_parse −24%, nanoid −20%. Perry adopts it once that perex release clears the 7-day min-publish-age soak.
• Keeping this open until then.

[proggeramlug]

Progress: #11548 landed perex 0.1.11 (the PerryTS/perex#3 matcher), under the owner-approved one-time publish-age override. Instructions: jsonwebtoken decode −61%, uuid v4 −48%, v7 −34%, v5_parse −28%, nanoid −30%, uuid/jws regex micro −65%/−84%; commander +0.4%; RSS flat. Together with #11543 this is the planned matcher + host work; leaving it open for the remaining gap to Node.