Regression: wave-1 codegen changes made `mandelbrot` time out (0.44 s → > 60 s, 51 s of it in the kernel) and `simple` 2x slower (0.23 → 0.50 s) — the bench gate exists but has never run because master CI is red (benchmarks/mandelbrot.php, script/bench-gate.sh, lib/JIT.php)

[PurHur]

Category

Regression: · generated-code performance, introduced between 4eed6a2785 and 1edb103c34 · child of #36379

Problem (pinned image, same box, outputs verified)

program	Zend	2026-09-01 4eed6a2785	2026-09-02 1edb103c34
benchmarks/mandelbrot.php	0.28 s	0.44 s	> 60 s (timeout), user 7.6 s / sys 51.4 s, plus false warnings (#36405)
benchmarks/simple.php	0.09 s	0.23 s	0.50 s
fibo(30)	0.13 s	0.125 s	0.041 s (improved; July: 0.015 s)
Ack(3,8)	0.16 s	0.16 s	0.066 s (improved)
2M untyped calls	0.09 s	0.575 s	0.643 s

51 s of system time on a float loop means the binary is thrashing the allocator/kernel — allocating and freeing (or faulting) per iteration — which points at the new dtor/{main} ownership path (#36215/#36252, one of which was partly reverted: "Fix AOT function-scoped array append segfault — revert broken delref dtor/GC CFG") or the __value__ layout change (#36214) interacting with float boxes. simple's regression is call overhead (three 1M-call loops).

script/bench-gate.sh (#36196) would have flagged both — it is wired into compiler-gate.yml, which has not had a green run on master in its last 60 attempts (#36377, #36401). A gate that never runs is a gate that does not exist; 131 commits merged over it in 24 h.

PHP implementation target

• Bisect the two programs between the two SHAs (each build ~20 s; ~7 steps) — start with the merges of Runtime: dropping an array or object to refcount 0 frees only the header — packed values, string-key nodes and property boxes are never freed or delref'd, so every container death leaks (lib/JIT/Builtin/Refcount.php) #36215, AOT: every array or string mutation in {main} scope is O(n²) in time and memory — refcounts never decrement in main, so separateContainerForWrite copies the whole table on every store (16k appends: 6.1 s in main vs 0.011 s inside a function) (lib/JIT/Builtin/Refcount.php phpc_destruct_delref_allowed, lib/JIT/HashTableWriteLlvm.php:1694) #36252, Runtime: __value__ is 9 bytes / align 1 so every i64/double payload access is an over-aligned (UB) load, and packed arrays allocate 16 B/slot but stride 9 (lib/JIT/Builtin/Type/Value.php, lib/JIT/Builtin/Type/HashTable.php) #36214, Codegen: no LLVM pass ever runs at the default level — 40 alwaysinline runtime accessors stay real calls, MCJIT never optimises, and the per-function pass manager is mis-wired to a non-existent symbol (lib/JIT/Context.php runModuleOptimizationPasses, vendor/ircmaxell/php-llvm) #36213, Codegen: every boxed assignment open-codes a 15-block type switch with duplicated error blocks — 111k-line IR module and 51 KB helper functions for a 4-line script (lib/JIT/JitValueBox.php) #36193; name the commit in the PR.
• perf stat/strace -c on the mandelbrot binary to confirm the syscall (brk/mmap/munmap per iteration?) before fixing.
• Until CI is green, bench-gate.sh must be run locally and pasted in every PR touching lib/JIT*, lib/AOT, ext/standard (CONTRIBUTING rule, enforced by Process: Definition of Done — a PR may say "Closes #N" only when every Done-when box is checked; otherwise "Part of #N"; PR template + a scope-check gate + a weekly "closed but not done" audit (CONTRIBUTING.md, .github/PULL_REQUEST_TEMPLATE.md, script/check-issue-close-scope.php) #36400's template).

Repro

./script/docker-exec.sh -- bash -lc 'source script/php-env.sh && php bin/compile.php -o build/mb benchmarks/mandelbrot.php && time timeout 60 ./build/mb > /dev/null; php bin/compile.php -o build/sp benchmarks/simple.php && time ./build/sp'

Done when

• mandelbrot ≤ 0.3 s and simple ≤ 0.25 s under AOT in the pinned image (bench-gate rows pasted); no stderr output
• Bisected commit named; a differential case pins the shape; script/differential-sweep.sh --aot --repeat 3 unchanged

[PurHur]

The repo's own gate confirms it — script/bench-gate.sh on master 1edb103c34 in the pinned image (run 2026-09-02 19:48 UTC):

case              zend(s)     aot(s)      ratio   ir_lines    defines
fibo(30)           0.1314     0.0320     0.2432      97543        419
simple             0.0825     0.4252     5.1558     104718        424
mandelbrot         0.2409    31.3620   130.1851     248216        413
Ack(3,8)           0.0987     0.0420     0.4249     101488        425
bench-gate: PERFORMANCE REGRESSION
  mandelbrot.ratio: 130.1851 exceeds baseline 1.1192 by 11531.6% (limit +30%)
  fibo(30).ratio:   0.2432 exceeds baseline 0.1687 by 44.2% (limit +30%)

Note mandelbrot's IR is 248k lines against ~100k for the others — 2.5x the module for the same runtime, which is a lead: whatever the float loop now lowers to is both huge and allocation-heavy. The gate is red and has been since it landed; it lives in compiler-gate.yml, which fails in the patch-apply setup step before reaching it (#36377). Until that is green, anyone touching lib/JIT* must run this locally and paste it.

[PurHur]

claim: implementer-worker-lane-b — starting this run (lane overflow: #36143–#36148 closed; mandelbrot timeout regression)

[PurHur]

Lane B overflow check: mandelbrot/simple already green on master 8a07fd8 (mandelbrot 0.106s, simple 0.167s). No PR needed — claimed in error, pivoted to #36411.

[PurHur]

claim: implementer-worker-lane-c — starting this run (mandelbrot bench-gate ratio + IR bloat; timeout already green on master)

[PurHur]

claim: implementer-worker-lane-b — starting this run (lane overflow: north-star children closed; finish done-when + bench-gate ratio on mandelbrot/simple)