Regression: AOT password_hash(PASSWORD_BCRYPT) segfaults after build — VM/JIT match Zend (re-#26773, ext/standard/password.c)

[PurHur]

Category

bug · php-src-strict · stdlib · user-script AOT · pillar 4

Problem

password_hash(..., PASSWORD_BCRYPT, ['cost' => 4]) + password_verify match Zend on VM and JIT (ok / ok), but user-script AOT builds then segfaults at run time (fatal signal after c:main_before_php, exit 139). Reproduced 3/3 runs on the same binary — not intermittent.

Prior closed AOT gap #26773 was PASSWORD_ARGON2ID LLVM verify / dominate-uses; this is the bcrypt path segfaulting after a successful link.

Probed 2026-08-02 @ d72ac27fd.

Repro	Zend / VM / JIT	AOT (2026-08-02)
password_hash('secret', PASSWORD_BCRYPT, ['cost'=>4]) + verify	ok / ok	build OK (~25 MB); run exit 139 ×3

php-src reference

• php/php-src ext/standard/password.c — php_password_hash / bcrypt
• php/php-src ext/standard/password.stub.php

PHP implementation target

• ext/standard/ password helpers + lib/JIT/Builtin/ password lowering — AOT must not leave a null/broken crypto trampoline; prefer PHP-in-PHP / existing VmPassword path over new runtime/*.c
• Confirm thin AOT execute links the same helper as JIT (no VmClosureCall-style stub)

Repro

./script/docker-exec.sh -- bash -lc '
cat > /tmp/issue_pwd_aot.php <<'"'"'PHP'"'"'
<?php
$h = password_hash("secret", PASSWORD_BCRYPT, ["cost" => 4]);
echo password_verify("secret", $h) ? "ok" : "bad", "\n";
echo password_verify("wrong", $h) ? "bad" : "ok", "\n";
PHP
php /tmp/issue_pwd_aot.php
php bin/vm.php /tmp/issue_pwd_aot.php
php bin/jit.php /tmp/issue_pwd_aot.php
php bin/compile.php -o /tmp/issue_pwd_aot /tmp/issue_pwd_aot.php
/tmp/issue_pwd_aot; echo exit:$?
'

Done when

• AOT binary prints ok / ok matching Zend/VM/JIT (no segfault)
• Re-run binary ≥3× (or script/differential-sweep.sh --aot --repeat 3 equivalent) — all match
• Compliance or unit guard under test/ covering bcrypt password_hash+password_verify AOT execute
• php-src-strict; no php-compiler-strict shortcut

[PurHur]

claim: lane-C language/compiler worker — starting this run

TRY_FIRST:

Gate	Status
docker info	ok
bootstrap-inventory --check	ok (6777/6777 via release-readiness)
release-readiness --json	user_release_ready: yes; north-star5-fast ok; gen-0 provenance stale (known #26756)
north-star5-verify-fast	ok (via release-readiness)

Lane note: #26756 (only non-stdlib Lane C implementation-ready) has open PR #26778 + recent lane-A claims — skipped. Taking Pillar 4 AOT regression #26861 (password_hash bcrypt segfault; VM/JIT green).

[PurHur]

Progress (lane C) — #26861 WIP handoff (not merge-ready)

TRY_FIRST: user_release_ready: yes; north-star5-fast ok; gen-0 provenance stale (known).

Root cause of AOT segfault (confirmed)

PasswordJitHelper::bcryptEncodeSalt22 used $out .= $itoa[...] + while (strlen($out) < 22). Under NestedJIT/AOT, strlen($out) stays 0 after single-char concat, so the loop never terminates and $rnd16[$i++] eventually segfaults (gdb: bcryptEncodeSalt22).

Repro bisect: same loop as a user script AOT binary also hangs/SEGV; count($chars)/implode avoids the infinite loop but NestedJIT bit-shifts still produce wrong itoa64 chars (Zend .OC/… vs AOT .O..…).

Current branch tip behavior

Path	Result
Zend / VM / JIT	ok / ok
AOT (before)	exit 139
AOT (after WIP)	no SEGV; prints bad / ok (hash still false / crypt fails)

WIP direction on agent/issue-26861-aot-password-hash

• Unrolled / LLVM phpc_bcrypt_hash leaf (encode+setting+crypt(3) with NUL via dupCstrFromString)
• JitLibcryptKernel NUL-dup for crypt(3) (__string__ is length-prefixed only)
• Allowlist phpc_bcrypt_hash in Context::isPreRegisterModuleNestedJitKernel

Blocker / next step

AOT still returns false hash (password_verify('secret', $h) → bad). Likely remaining: NestedJIT of phpc_bcrypt_hash / cost-two string / salt bytes still not feeding libcrypt a valid $2y$04$… setting (DES crypt("secret","aa") AOT also yields *0 while container crypt(3) works). Next: finish/verify the combined LLVM leaf end-to-end (force-rebuild helper unit with JitLibcryptKernel in deps), then refresh prelinked/helper-runtime/…/PasswordJitHelper + unit guard; only then merge.

Do not merge until AOT prints ok/ok ≥3×.

[PurHur]

claim: lane-C language/compiler worker — continuing #26861 / PR #26902 (own in-flight)

TRY_FIRST (this run):

Gate	Status	Symptom
php script/bootstrap-inventory.php --check	green	OK 6788/6788
./script/release-readiness.sh --json	(awaiting / prior green on master)	gen-0 provenance still stale (#26756)
docker	ok

This run: finish AOT bcrypt so password_hash+password_verify print ok/ok (currently bad/ok after SEGV fix); then unit guard + merge.

[PurHur]

claim: PHP Runtime → PHP — continuing this run (PR #26902)

TRY_FIRST:

Gate	Status
docker info	ok
bootstrap-inventory --check	OK 6788/6788
release-readiness --json	user_release_ready: yes; north-star5-fast ok; gen-0 provenance stale (known #26756)
vm-driver-probe	ok

Continuing WIP: AOT no longer SEGV but hash still bad/ok. Aiming for PHP-in-PHP NestedJIT-safe encode (soundex pattern #26912) rather than growing C.

[PurHur]

Merged via #26902

Verification

Check	Result
AOT password_hash+verify (cost 4) ×3	ok/ok (HELPER_RUNTIME_O=0 and =1)
AOT crypt("secret","aa")	aa3kxsKzvuQMc (matches Zend)
PHPUnit PasswordCrypto + LibcNameCollision	OK
make bootstrap-selfhost-vm-driver-execute-probe	OK

Cause / fix

1. AOT exported PHP strspn/strcspn/… under libc names → libxcrypt interposed → crypt(3) returned *0. Now __compiler_* ABI + link version-script hide.
2. NestedJIT bcryptEncodeSalt22 strlen($out) after .= never grew → SEGV. Now track $n.

No new runtime/*.c.