AOT: split-compilation helper-runtime cache — experimental scaffold (#15889)

[PurHur]

First working scaffold for #15889 (multiple .o merged at link time), draft / default-off — see honest status below. Includes the #16010 revert as its base (branched before the revert lands on master; rebase on merge).

What works

• script/emit-helper-runtime-object.php: reflection discovery of all 137 helper units (*HELPER_PATH/*COMPILED_HELPERS pairs incl. variants), isolated per-unit subprocess lowering — 96 unit objects, 160 helpers cached, 41 Regression: user-script AOT builds broken for core string builtins — preg_match capture assign, explode CRLF, str_contains, substr negative length (bin/compile.php) #15642-class crashers auto-skipped, 22.8 MB, one-time ~6 min
• Per-script declaration binding with exact types from unit bitcode (no module linking, no linkage mutation)
• Linker merges used unit objects with -z muldefs (script object first → ABI state single-copy)
• Hello-world builds+runs correctly with the cache enabled

Known blocker (why draft/off)

Each unit object carries its own __init__ chain (interned-constant setup for its helper bodies). muldefs keeps only the script's init, so cached helpers can read uninitialized state → silent exit for helper-using scripts. Fix design (in #15889): rename each unit's init to a unique symbol at emit time and register it in llvm.global_ctors, so ld keeps and libc runs all of them; helper-state init must be idempotent vs script init.

Also documented in-code: latent php-llvm bug (createMemoryBufferWithFile references unimported FFI), worked around via createMemoryBufferWithString.

Measured potential

The nested helper lowering this replaces is 47% of every build (#15889 profile); a mid-session broken-but-linking variant showed hello 3.0 s → 1.9 s and 006-project 5.3 s → 2.2 s — the win is real once init merging lands.

🤖 Generated with Claude Code

[PurHur]

Third commit: the cache is now incremental and resumable — per-unit fingerprints (core machinery + unit source content) and persistent per-unit crash markers.

Scenario	Wall
cold (137 units)	5 m 53 s — 96 emitted, 41 crash-marked
warm rerun	0.15 s — everything skipped
one helper edited	3.0 s — exactly 1 unit re-emitted
interrupted run (5 units missing)	13.8 s — only those re-processed

Crashed units are remembered by fingerprint and re-attempted only when their source or the compiler core changes — 'resume only the breaking parts' both within a run and across compiler evolution. The remaining blocker for enabling per-script consumption is unchanged (unit init chains, #15889 item 4).

[PurHur]

Correctness matrix (fresh per-unit cache, reverted base): cache-on output == baseline == VM on every construct plain AOT supports; the earlier silent-exit was stale pre-revert cache artifacts, not the init-chain theory. Remaining gap for real speedups: (1) hello-class scripts' hot helpers (WeakRefRegistry/ErrorSilence/DirHandle/Progress) lower via the ensureStandaloneBodies corpus, not JitVmHelperLink units — coverage extension needed; (2) the 41 crash-marked units are dominated by Array*JitHelpers failing only in isolated emission. Both are #16075 step-4 (helper ABI registry) scope.