[proggeramlug]

Summary

Slice 4 of the string-build work (follows #13450). 30-string-build: 7.195B -> 6.291B instructions (-12.6%); the remaining gap to scriptc is split old birth and GC.

• D, split piece births without root/reload. arena_alloc_gc_old_no_collect (and the existing nursery no-collect form through string_storage_alloc_for_no_collect) serve a piece from a swept same-size hole or from the block that is already open and never reach gc_check_trigger or a block reservation: a non-null return proves nothing moved, so split_by_delimiter keeps the source and result addresses live across the piece allocation. A refusal takes the unchanged rooted, collecting allocation and refreshes both addresses from their handles, so the collection decision is the one the unconditional call would have made.
• F, string-array .length loops. for (i < a.length) s += a[i].length over a declared string[] is admitted by the indexed-owner region: ordinary unforwarded array without element descriptors, length <= capacity, one preheader scan that every slot is a heap string or an inline ASCII string (a hole, undefined, an object or a non-ASCII inline string declines the whole region before any body effect), then a call-free body (tag test + header load per element). The effect-gated entry poll stays (slice-2 lesson).
• G, join. The size/validation walk and the copy walk test the heap-string tag first, add into plain u64 totals and copy short pieces inline.
• js_string_concat_parts_2..5. Count-specialized entries for the operand counts the compiler knows; the operand classification and the copy are unrolled with constant indices (generic over N, selected by n, no per-program code). The generic entry dispatches to them.

Results (qb9, x86-64-v3, KEEP_SYMBOLS, main 88ed1e5 vs head 304a32b)

witness	Ir main	Ir head	delta
30-string-build (callgrind)	7,195,107,053	6,291,485,243	-12.56%
30-string-build (n5 median perf instr / cycles)	7,163,152,238	6,251,569,771	-12.73% / cycles -7.4% (n11: -6.6%)
24-map-set (callgrind)	3,981,178,535	3,981,182,976	+0.0001%
31-json (callgrind)	19,391,283,157	19,389,905,102	-0.007%

All other js-compiled benchmarks: instructions within 0.02% (43-generators -0.30%). n5 cycle deltas on unchanged-instruction benchmarks reach +/-9% on this shared host (n11 re-run: 24-map-set -6.2%, 32-regex +2.2%, same instruction counts), so cycles carry host noise; instructions are exact.
GC_LOG schedule (same kinds, freed/old bytes) and stdout identical on 30, 24, 31. RSS n5 median (KiB): 30-string-build 238,720 -> 238,768, 24-map-set 28,240 -> 28,296, 31-json 149,908 -> 149,880 (no allocation count, size, capacity or generation changes; timing-free).
.text: 30-string-build -1,428 B, 24-map-set +2,152 B, 31-json +27,256 B (+0.43%: the four unrolled concat entries, 2.8-5.9 KB each, replace one 6.1 KB generic entry; reported, not hidden).

Startup, exact callgrind Ir, n3 per arm, deterministic, ASLR off:

witness	raw main	raw head	d	controlled main	controlled head	d
hello	140,644	140,644	0	159,465	159,465	0
commander	21,219,728	21,221,109	+1,381	111,769,396	111,769,325	-71
commander --help	22,778,990	22,783,051	+4,061	117,988,376	117,987,253	-1,123
fastify	159,953,030	159,870,870	-82,160	2,934,775,210	2,934,694,396	-80,814
Richards	42,272,614	42,276,096	+3,482	370,216,452	370,214,586	-1,866
Controlled = the JMS address-independent-key control applied to both arms (canonical keys, path registry, shapes, perex cache) plus an address-independent memcmp/bcmp preload on both arms. Controlled head <= main on all five. The raw positives are layout artifacts: with the JMS control alone commander was +153, all of it in glibc __memcmp_avx2_movbe (13,943 calls from the same callers in both arms; the delta sits in the `(a	b)&0xfff > 0xfe0` page-offset branch at 0x1975c0, 89 more calls take it in head); concat is -70 Ir in commander (2 calls).

Tests

• runtime: string 252, json 171, array 453, arena 133 (two new: the old no-collect entry REFUSES a full open block without reaching the allocation-point trigger, and births the same header as the collecting entry); codegen lib 2589 (new: string-array region scans once, stays call-free, poll precedes the scan; declines stores/calls/rebinding), integration 23.
• test-files/test_gap_jst6_split_join.ts against Node: split of empty strings/separators/unicode/lone surrogates/very long/20k pieces, join of holey arrays and non-string elements with toString/Symbol, string-array length loops over holes, undefined, null, objects, a Proxy, an accessor element, a subclass, a shifted array, concat with 2..6 operands across every value kind. Passes on main and head.
• Sabotage: forcing a full collection in the no-collect piece path (D) and dropping the slot-scan verdict (F) both fail that test (piece 0 comes back as "p1"; join no longer equals the source).
• Node parity 22/22; gap subset 89: 83 pass, the same 6 fail on main. Windows runtime build/sidecar: see the lane report.

[coderabbitai]

Caution

Review failed

The pull request is closed.

Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

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.}