Regression: AOT password_hash(PASSWORD_ARGON2ID) LLVM verify fail — dominate uses (ext/standard/password)

[PurHur]

Category

bug · php-src-strict · ext/standard/password · user-script AOT · pillar 4

Problem

password_hash(..., PASSWORD_ARGON2ID, [...]) compiles and runs on VM/JIT (matches Zend) but AOT aborts at LLVM module verify with a dominance error on the options-cost phi. Probed 2026-08-02 @ b5715a3e2.

Repro	Zend 8.2+ / VM / JIT	AOT (2026-08-02)
password_hash('secret', PASSWORD_ARGON2ID, memory_cost=8,time_cost=1,threads=1) + verify	ok / ok	build: Instruction does not dominate all uses! on %pw_cost_* phi (Module->verify)

Open Reflection-only siblings (#24827, #25479) do not cover this AOT IR bug.

php-src reference

• php/php-src ext/standard/password.c — php_password_hash / argon2
• php/php-src ext/standard/password.stub.php — password_hash

PHP implementation target

• ext/standard / lib/JIT password_hash options lowering for AOT (cost/time/threads phi wiring)
• PHP-in-PHP; no new runtime/*.c

Repro

./script/docker-exec.sh -- bash -lc 'source script/php-env.sh
cat > /tmp/pw.php <<'"'"'PHP'"'"'
<?php
$hash = password_hash("secret", PASSWORD_ARGON2ID, [
  "memory_cost" => 8, "time_cost" => 1, "threads" => 1,
]);
echo password_verify("secret", $hash) ? "ok\n" : "bad\n";
echo password_verify("no", $hash) ? "bad\n" : "ok\n";
PHP
php /tmp/pw.php
php bin/vm.php /tmp/pw.php
php bin/jit.php /tmp/pw.php
./phpc build -o /tmp/pw /tmp/pw.php && /tmp/pw
'

Also: test/repro/maintainer_gap_aot_password_hash_argon2id.php

Done when

• AOT build succeeds and repro prints ok / ok
• VM/JIT unchanged
• AOT fixture or compliance guard
• php-src-strict; no php-compiler-strict shortcut

[PurHur]

claim: lane-B implementer — starting this run

Pre-flight (this workspace):

Gate	Result
bootstrap-inventory (release-readiness)	ok 6728/6728
north-star5-verify --fast	ok
vm-driver-probe	ok
helper-runtime-prelink --strict	fail — stale ext/dom helpers (Ext DOM lane; not this issue)
user_release_ready	no (prelink stale)
gen-0 provenance	stale (known #23468 / #26756)

Implementing AOT password_hash ARGON2ID LLVM dominate-uses fix (php-src-strict).

[PurHur]

Progress (lane-B) — PR #26780

Compile / Module->verify: FIXED on this branch.

./script/docker-exec.sh -- bash -lc 'source script/php-env.sh
./phpc build -o /tmp/pw test/repro/maintainer_gap_aot_password_hash_argon2id.php'
# exit 0 — no "Instruction does not dominate all uses" / PHI predecessor mismatch

VM/JIT still ok/ok.

AOT runtime: still red — binary prints bad/ok (hash is false). Root cause appears to be NestedJIT helper-unit string return into __compiler_password_hash bridge (not the options PHI). Debug during nest-compile showed hashArgv can produce a non-false result while user code still sees boolean false.

Landed in #26780

• JitPasswordBcryptCost PHI predecessors + __value__* for readHashtable
• extractStringPtrFromHelperResult i64 path
• PasswordJitHelper::hashArgv — no bcrypt-only !== gate; string return; helper-runtime unit refresh
• compile-only fixture password_hash_argon2id_options.php

Next (do not merge as Closes until green)

1. Dump LLVM return type of passwordjithelper__hashargv from the helper .o / nest-compile IR
2. Align bridge null/empty handling with that ABI
3. Confirm argon2 path in standalone (FFI / deps) once string returns work
4. Re-run issue repro → expect ok/ok, then squash-merge Compiler: AOT password_hash ARGON2ID NestedJIT thin kernels (#26773) #26780

Not merging this run — repro runtime still fails.

[PurHur]

claim: PHP Compiler STD Lib — continuing PR #26780 (AOT runtime)

Pre-flight: docker ok; inventory OK 6728/6728; release-readiness still running.

Status: compile/verify green on branch; AOT runtime still false for bcrypt and argon2id (gettype boolean). Root cause: VmPasswordPure only uses __compiler_libcrypt when NestedJitCompileScope::isActive(), which is false at AOT runtime; NestedJIT'd \crypt() does not call libc crypt(3) (circular PasswordJitHelper). Restoring thin crypt(3) + always-route; then argon2 thin ABI for the issue repro.

[PurHur]

Maintainer: #26780 compile/dominate-uses fix is useful WIP but do not close until AOT runtime returns a real hash ( on the issue repro). Hold-merge commented on the PR.

[PurHur]

Maintainer follow-up: #26780 compile/dominate-uses fix is useful WIP but do not close this issue until AOT runtime returns a real hash (ok/ok on the issue repro). Hold-merge already commented on the PR.

[PurHur]

handoff: PHP Compiler STD Lib — AOT runtime still red after #26780 push

Compile/verify: green (PHI dominate-uses).
VM/JIT: ok/ok.
AOT run: still false hash (bad/ok).

Landed: isActive NestedJIT fold + thin phpc_libcrypt_kernel (see PR comment). Next is honest NestedJIT closure for VmPasswordNative/Pure (or argon2 thin ABI) without pulling broken VmSscanf into the password bundle.

[PurHur]

claim: PHP Compiler STD Lib — continuing PR #26780 (AOT runtime → ok/ok)

Pre-flight: docker ok; inventory OK 6738/6738; hello VM smoke ok.

Focus: NestedJIT/helper string ABI or argon2 thin path so AOT repro prints ok/ok (compile already green).