Keep regexp.zig as the JS matcher and speed up exec

[aneryu]

Engineering Signoff

• No fixture-shaped parser, emitter, VM, or builtin shortcut was added.
• No new broad anyerror, empty catch, @ts-ignore, or test weakening was introduced.
• No public API, validation boundary, skip, or exclude changed without a failing scenario and exit criterion.
• Object/shape/GC/IC ownership rules were preserved, or the owning design note was updated.
• Durable validation evidence is recorded in this PR, the commit message, or the owning issue.

Summary

JS RegExp compile and exec use src/libs/regexp.zig (QuickJS-style bytecode). The V8 Irregexp C++ island is not in this tree or the production link.

This branch first wired Irregexp, then switched back: stripped zjs grew ~718 KiB for a matcher-only gain, while \p{} coverage the original engine already had would have been lost, and remaining bench-v8 time after matcher peeps is match-array construction.

What lands:

• JS compile/exec through regexp_adapter.zig → libs/regexp.zig
• Trusted exec reads operands without error unions; latin1 backtracker is force-inlined; 64-slot inline backtrack/undo stack
• Greedy class8 / not_class8 loops consume the whole run in a latin1 byte loop. Extra length is one greedy_shrink chain, deferred on ExecState so a successful match at greedy max does not push a frame.
• The first in-place shrink now parks the rest of that chain and clears greedy_pending. Leaving pending armed let a later * consume given-back letters as extra [a-z]+ iterations (test262 S15.10.2.8_A3_T32 / T33, ARM64 CI).
• Consecutive-char fusion (new chars8 opcode, in-arm peel, and a+ → loop_class8_g) was tried and dropped: each grew or perturbed the always-inlined latin1 backtracker enough to lose on url-exec/url-test.
• Group-0 exec match arrays skip the N-capture path; header is parsed once per exec
• test() still does not build a match object
• \p{} / \P{} work again (they failed at Irregexp compile time)
• Short latin1 slices (length 2..32) interned in a 32-slot JSRuntime cache and reused from stringSliceValue. Repeated exec on the same subject dups capture strings instead of createSlice each time. Not a sliced-string view; the 12-byte JSString layout is unchanged. lreExecBacktrack is untouched.

Follow-up after review: intern size constants are file-private so the embedder surface grows by one method (recentLatin1Slice). The JSRuntime public decl pin is 169. LIMITATIONS.md records intern vs still-eager longer/UTF-16 copies.

Validation

• zig build test-core --summary new -- "short latin1 slices intern" (1/1)
• zig build test-exec --summary new -- "interned latin1 captures" (1/1)
• zig build test-exec --summary new -- "repeated URL matches" (1/1)
• zig build test-exec --summary new -- "nested greedy class8" (1/1)
• zig build test --summary new -- "JavaScript RegExp adapter" (6/6)
• zig build test-exec --summary new (504/504)
• zig build test-embedding --summary new (18/18)
• zig build checkpoint-gate --summary all (pre-CI-fix pin commit)
• git diff --check
• ./zig-out/bin/run-test262 -c test262.conf -t 4 -d test262/test/built-ins/RegExp (0/1879)
• ./zig-out/bin/run-test262 -c test262.conf -t 4 -d test262/test 0 100000 (0/49778 errors, passed 44584)
• zig build engine-production-gate --summary all

ReleaseFast on this x86_64 host (taskset -c 0; not published ARM). A/B vs frozen intern bin /tmp/zjs-slice-intern.

Unstripped zjs: HEAD matcher 29,850,096 · intern 29,872,384 (+22 KiB debug).

Microbench medians, ms, lower better (200k url/class/long, 400k tiny):

case	old matcher	Irregexp	switchback	HEAD matcher	intern	this PR
url-exec	174	137	125	125	116	116
url-test	98	55	48	47	47	46
tiny-exec	86	81	85	87	86	86
tiny-test	51	44	50	52	50	50
class-exec-g	127	121	126	126	126	126
class-replace	85	81	84	89	88	88
long-class-x (/[a-z]+x/ on 64 as)	—	—	46	32	32	32
isolated url-exec	—	—	~116	129	120	116

url-test stays at greedy-max-no-frame. Isolated 9-run url-exec after the T32/T33 pending fix is 116 vs intern 118.

Further matcher peeps that add code to the always-inlined lreExecBacktrack currently lose on the URL microbench. Do not retry those without a new mechanism.

Rollback Notes

Revert this branch. main already uses regexp.zig for JS RegExp; this PR is exec/JS-path speedups on that engine.

  

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c24db8472f

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve named-group lengths beyond 16 bits

When a valid capture name encodes to 65,536 or more UTF-8 bytes, this cast truncates its length while the following line still appends the complete name. The Zig metadata reader then treats the first name as shorter (zero bytes at exactly 65,536), parses the remaining name bytes as subsequent records, and omits or corrupts properties in exec(...).groups and indices.groups. Either widen the blob length field or reject overlong names before emitting a malformed names section.

AGENTS.md reference: AGENTS.md:L15-L15

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Return OOM instead of aborting the process

Under memory pressure while compiling a large regular expression, zone growth reaches this allocation path and calls FATAL when malloc fails, terminating the embedding process rather than returning ZJS_IRREGEXP_OOM; the other std::vector/make_unique allocations have the same problem because the library is built with -fno-exceptions. Consequently the Zig error.OutOfMemory handling and runtime OOM boundary cannot catch native compiler exhaustion. Route shim allocations through a fallible boundary that propagates the ABI's OOM status.

AGENTS.md reference: AGENTS.md:L17-L18

Useful? React with 👍 / 👎.