VM: the 1M-iteration loop still takes 218 s and `fibo(30)` still times out at 300 s after the "VM hot path" PRs; boot is 0.99 s after the opcache flag — the O(scope²) scans and lazy builtin registration did not land (lib/VM.php, lib/Block.php, lib/ExtensionRegistry.php) (re-#36207, re-#36206)

[PurHur]

Category

Regression: · closed-but-not-done, the interpreter · child of #36379

Problem (master 1edb103c34, pinned image, source script/php-env.sh so the new opcache flags apply)

	2026-09-01	2026-09-02	Zend
php bin/vm.php hello.php	1.15 s	0.99 s	0.05 s
1M-iteration for loop	207.6 s	218.4 s	0.01 s
fibo(30)	> 300 s	> 300 s	0.13 s

#36207 closed via three PRs ("skip opcode-loop preamble work on hot paths", "storage reverse-index, argRecv cache", "memoize Block scope-slot lookups at runtime") whose combined effect on the issue's own probe is zero (within noise). #36206 closed with the opcache flag (1.15 → 0.99 s) — the lazy registration of 2,629 builtin objects, the harness spawn changes and the stream_select drain were not delivered. The compliance suite is therefore still ~4 h, and the VM is still unusable as a benchmark or fuzzing oracle.

PHP implementation target

Profile before touching code: php -d opcache.enable_cli=1 -r with pcntl_alarm sampling (the recipe in docs/AGENT-HANDOVER-2026-08-17.md) on the 1M loop; the top frames must be named in the PR. The audit's candidates remain: Block::isNamedVariableSlot/operandForScopeSlot per operand per op (lib/Block.php:1260/:1555), Context::globalNameForStorage per slot write (lib/VM/Context.php:877), MemoryAccounting::noteBytes getenv per string write, the per-op preamble (lib/VM.php:4640-4659), frame-per-block (Block::getFrame). Acceptance is the probe, not the PR count.

Done when

• 1M loop < 10 s and fibo(30) < 60 s under bin/vm.php (numbers pasted); boot < 0.3 s with lazy registration
• script/compliance-baseline.sh --diff --suite=VMTest no regressions; serial VMTest wall time recorded before/after

[PurHur]

claim: implementer-worker-lane-b — starting this run (lane overflow: #36407 already green on master 8a07fd8; VM 1M-loop/fibo timeout)

[PurHur]

Lane B run — partial progress (PR #36421, not merged — 1M loop still red vs <10s target)

Gates (branch 4073c2413d)

• make dev-verify-fast OK (aot-smoke 9/9, VM diff tier-0 15/15)
• make north-star5-verify-fast OK

Repro

	wall
master 8a07fd8	60.4s
PR #36421	44.8s (~25% faster)
Zend	0.017s

Output: Done 1000000 (correct).

Changes

Early inc/dec fast path for initialized simple locals; int-CV-vs-literal compare fast path; ephemeral-slot JUMPIF dead-temp scan; skip redundant copyFrom/markScopeSlotInitialized.

Next step

Profile releaseVmDeadScopeSlot on loop JUMPIF (temp null+recreate each iter?) and defer dead-temp release on back-edges; boot lazy registration (#36206) for fibo timeout.

[PurHur]

claim: implementer-worker-lane-b — continuing PR #36421 (hot loop: continue-in-while, skip dead inc/dec result copy, fast int temp release)

[PurHur]

claim: implementer-worker-lane-c — continuing PR #36421 (counted for-loop fast path)

[PurHur]

Merged via #36421 — 1M-loop criterion green.

Verify (Docker, eb57a4938e):

• test/repro/vm_1m_loop.php: 2.4s user (was 60s master / 46s prior commit); output Done 1000000 matches Zend
• make dev-verify-fast → OK (aot-smoke 9/9, VM diff tier-0 15/15)
• make north-star5-verify-fast → OK

Still open on this issue: fibo(30) < 60s not re-run; recursive call hot path is a separate follow-up.

[PurHur]

Merged #36425 — native PHP ints in counted for-loop fast path.

test/repro/vm_1m_loop.php: 2.4s → 0.38s (Zend ~0.02s); Done 1000000 matches Zend
make dev-verify-fast OK; make north-star5-verify-fast OK

1M loop <10s done-when met. fibo(30) still >300s — follow-up needed.