perf(regexp): constructing a RegExp recompiles the pattern every time (14 KB emoji-regex: 631 µs vs 2 µs in bun, 274×) — string-width makes opencode --help cost 31 s #10179

Parent: #10106 (OpenCode startup) / #10107. Found while attributing opencode --help (31 s of CPU vs 0.4 s under bun).

The number

yargs 18 renders help through cliui → string-width@7.2.0, which for every grapheme of every line calls emojiRegex() — i.e. constructs new RegExp(<14,116-char emoji pattern>, "g") — then .test()s it. Per-operation cost, perry (linux-x64, --no-auto-optimize, 8-module probe) vs bun 1.3.14, same machine:

Net: 9,000 stringWidth() calls on three typical help lines = 340 s under perry vs 72 ms under bun (≈38 ms per call; GC share 1 %). That is the whole --help / run --help cost of the native OpenCode binary. Construction dominates (per grapheme, ~90 graphemes per line); the reused-regex .test and .replace paths are the second tier.

Repro

// inside opencode-v1.18.30/packages/opencode (paths from the bun isolated install)
import emojiRegex from "…/node_modules/.bun/emoji-regex@10.6.0/node_modules/emoji-regex/index.js"
import stripAnsi from "…/strip-ansi@7.1.2/node_modules/strip-ansi/index.js"
function t(name, n, f) { const t0 = performance.now(); for (let i = 0; i < n; i++) f(); console.log(name, ((performance.now() - t0) * 1000 / n).toFixed(1), "us/iter") }
t("construct", 200, () => { emojiRegex() })
const re = emojiRegex(); t("test reused", 2000, () => { re.test("a") })
t("stripAnsi", 2000, () => { stripAnsi("      --log-level   log level   [string] [choices: \"DEBUG\", \"INFO\"]") })

V8/JSC cache the compiled program by (source, flags), so a repeated new RegExp(sameLiteral) costs a hash lookup; perry recompiles the 14 KB pattern every time (see the earlier regex analysis: the engine's per-byte matching is fine, the cost is eager compilation plus per-call wrapper overhead).

Expected

• new RegExp(source, flags) for a source already compiled in this process returns a RegExp sharing the compiled program (per-object lastIndex stays per object) — construction of the emoji regex drops to the microsecond range.
• .test/.replace on a reused regex within ~3× of bun for short inputs (the per-call wrapper overhead noted in the regex analysis).
• opencode --help user CPU falls from 31 s to well under 1 s.

Acceptance

• The probe above under perry: construction < 20 µs, test reused < 2 µs, stripAnsi < 1 µs.
• Node parity unchanged (lastIndex, flags, source, sticky/global semantics, RegExp.prototype.compile if supported), regex test262 slice green.
• opencode --help timing posted on perf: OpenCode native binary startup must beat the bun binary — attribute and remove the module-init cost (measured 13× bun CPU on the Aug 21 build) #10106.