php-in-php: delete JitFileGetContentsKernel — always JitVmHelperLink + FileGetContentsJitHelper PHP (#1492)

[PurHur]

Category

php-in-php · runtime-shrink (#1492, finish #19279 / Unlink pattern #19186)

Problem

lib/JIT/Builtin/StringFileGetContents.php still routes UserScriptAotDeferNestedJit::shouldDefer() through ext/standard/JitFileGetContentsKernel.php (~185 LOC hand-written libc open/read LLVM). Embed already uses FileGetContentsJitHelper → VmFs::fileGetContents().

Committed helper-runtime unit ext_standard_FileGetContentsJitHelper_php already exists with runtime_safe: true — same prelink that made Unlink (#19186) drop its libc defer. Kernel can go.

php-src reference

• php/php-src ext/standard/file.c / streamsfuncs.c — PHP_FUNCTION(file_get_contents) / php_stream_copy_to_mem

PHP implementation target

• lib/JIT/Builtin/StringFileGetContents.php — always JitVmHelperLink::ensureBridge (drop defer → kernel)
• Delete ext/standard/JitFileGetContentsKernel.php
• Update test/unit/FileGetContentsRuntimeShrinkTest.php + spine require list
• SSOT remains ext/standard/FileGetContentsJitHelper.php → VmFs::fileGetContents()

Repro

./script/docker-exec.sh -- bash -lc 'source script/php-env.sh
vendor/bin/phpunit test/unit/FileGetContentsRuntimeShrinkTest.php
tmpdir=$(mktemp -d); echo hi > "$tmpdir/x"
PHP_COMPILER_AOT_USER_SCRIPT=1 php bin/compile.php -o "$tmpdir/fgc" -r "echo file_get_contents(\"$tmpdir/x\");"
"$tmpdir/fgc"
make bootstrap-selfhost-vm-driver-execute-probe
'

Done when

• JitFileGetContentsKernel.php deleted
• UserScriptAotDeferNestedJit removed from StringFileGetContents bridge path
• User-script AOT file_get_contents smoke prints file contents
• FileGetContentsRuntimeShrinkTest green
• make bootstrap-selfhost-vm-driver-execute-probe green
• PR body cites lines removed

Refs #1492, #15309, #19186, #19279

[PurHur]

claim: PHP Runtime → PHP — starting this run.

[PurHur]

Blocker (this run)

Tried Unlink-style always JitVmHelperLink::ensureBridge + delete JitFileGetContentsKernel (committed unit is runtime_safe: true).

User-script AOT regresses:

• PHP_COMPILER_AOT_USER_SCRIPT=1 php bin/compile.php … file_get_contents($path) → compiles (helper-runtime cache hit) but execute prints empty (expected hi)
• missing-path binary segfaults after c:main_before_php
• AotTest file_get_contents / file_get_contents_offset fixtures fail (empty vs payload)

Same class as rename (#19215 / #19216): prelinked helper binds at compile time but user-script AOT execute is wrong/unsafe vs thin libc kernel.

Reverted experimental tree. Leaving issue open for when helper-unit __init__ / #16075 makes NestedJit helpers honest under PHP_COMPILER_AOT_USER_SCRIPT. Not merging a red AOT path.

Pivoting this run to a kernel-move that keeps AOT green (bin2hex Builtin → ext/standard/JitBin2hexKernel, same shape as #19311).

[PurHur]

Related ship this run: #19344 / #19345 (bin2hex kernel move). #19339 remains open for honest Unlink-style fgc when #16075 helper-unit init is fixed.

[PurHur]

claim: PHP Runtime → PHP — starting this run.

Pre-flight: inventory OK 5040/5040 · vm-driver-probe OK · release-readiness user_release_ready: no (spine missing 10 socket_* units after #19404/#19405 — Lane A/B claimed that trust fix).

Plan: retry Unlink-style always JitVmHelperLink::ensureBridge + delete JitFileGetContentsKernel (same shape as #19186). Prior empty/segfault may have been helper coerce / null-result path — will verify AOT smoke + FileGetContentsRuntimeShrinkTest before merge.

[PurHur]

Blocker confirmation (this run) — not merging

Reproduced prior empty/segfault under always-PHP-bridge:

Mode	Result
Master + JitFileGetContentsKernel + PHP_COMPILER_AOT_USER_SCRIPT=1	[hi] OK
Always FileGetContentsJitHelper bridge (kernel deleted)	empty output
Same + missing path	segfault after c:main_before_php
Helper switched to host @\file_get_contents (Unlink-shaped)	still blocked (helper TU rebuild / empty)

Root cause (narrower than #16075 handwave): user-script / ensureMinimalUserStandaloneBodies path needs the libc kernel today. Plain JitVmHelperLink::ensureBridge is also unsafe for nullable __string__* (no null-result branch — would explain missing-path segfault). Embed-style PHP bridge is not what bin/compile.php uses when defer is on.

Not shipping a red AOT path. Leaving kernel in place.

Next migration this run: quarantine another *UserScriptLlvm stub into ext/…/Jit*Kernel (pattern #19399 / #19389) — net shrink of lib/JIT/Builtin/ without deleting a working AOT kernel.

[PurHur]

claim: compile-link-worker — continuing blocker fix (refresh helper prelink init_via_global_ctor + drop kernel defer)

[PurHur]

Blocker confirmation (compile-link-worker run)

Pre-flight (master): user_release_ready: yes · inventory OK 5076/5076 · bootstrap-selfhost-link OK · north-star5-verify-fast OK

Repro matrix (this run)

Config	echo file_get_contents($existing)
Master + JitFileGetContentsKernel defer	hi OK
Kernel deleted + PHP bridge + prelink refresh (init_via_global_ctor)	empty
Master StringUnlink PHP bridge (no kernel)	fail, file not deleted

Root cause (narrower)

1. User-script AOT bridge calls prelinked FileGetContentsJitHelper → VmFs::fileGetContents → VmFsReadPure → host @\fopen/@\fread.
2. Helper units are emitted under PHP_COMPILER_AOT_USER_SCRIPT=1 (script/emit-helper-runtime-object.php L40–42). That matches minimal standalone init where host stream calls inside the linked helper TU do not work (same class as prior empty-output report).
3. Clearing PHP_COMPILER_AOT_USER_SCRIPT during unit emit (mirroring JitVmHelperLink nested compile) fails LLVM verify for this unit: Terminator found in the middle of a basic block! label %standalone_main — so we cannot refresh a working full-init helper TU today.
4. init_via_global_ctor alone on the committed prelink is insufficient (tested).

Conclusion: JitFileGetContentsKernel libc open/read remains required for PHP_COMPILER_AOT_USER_SCRIPT=1 until either (a) VmFsReadPure gains a non-host read path usable from helper TUs (#8950 / stream JIT), or (b) helper-unit emit can lower this unit without user-script stubs without verify failure.

Not merging a red user-script AOT path. Reverted local WIP.

Next: child issue for helper-unit emit verify failure on full-init FileGetContentsJitHelper, or unblock VmFsReadPure AOT read before re-attempting kernel deletion.

[PurHur]

claim: PHP Runtime → PHP — starting this run.

Pre-flight: bootstrap-selfhost-link OK · release-readiness north-star5-fast OK · inventory OK (via release-readiness gate).

Plan: Route user-script AOT through existing implementPhpBridge (nullable __string__* null branches) + delete JitFileGetContentsKernel — Unlink-shaped helper-runtime prelink, not plain ensureBridge.

[PurHur]

Blocker confirmed (this run) — kernel still required

Reproduced under current master + attempted defer→implementPhpBridge one-liner:

Case	Master (kernel defer)	defer→PHP bridge
bin/compile.php temp script echo file_get_contents($path)	hi	(empty)
Missing path var_export(file_get_contents(...))	safe false	segfault after c:main_before_php

Root cause (narrower): bin/compile.php sets PHP_COMPILER_AOT_USER_SCRIPT=1 for normal user scripts (phpc_compile_is_user_script_aot), so standalone AOT always hits the defer path today. The PHP bridge (implementPhpBridge with nullable __string__* null branches) is not equivalent to the libc kernel under helper-runtime prelink in that mode.

Broader signal: UnlinkBuiltinTest::testAotNativeBinaryMatchesPhpSubset also fails on master (ok → fail for first unlink) even though StringUnlink uses JitVmHelperLink::ensureBridge without defer — suggests helper-runtime init for file builtins in user-script AOT needs a separate fix before any kernel delete (#19339, unlink parity).

Not merging kernel delete until user-script AOT smoke is green without libc open/read.

Next eligible migration (filed): mechanical quarantine DomDocumentMethodUserScriptLlvm → ext/dom/JitDomDocumentMethodKernel (pattern #19471, no behavior change).

[PurHur]

claim: PHP Runtime → PHP — starting this run.

Pre-flight: user_release_ready: yes · inventory OK 5121/5121 · north-star5-fast OK · vm-driver-probe OK (via ./script/release-readiness.sh --json).

Plan: Unblock kernel delete via Rename/Fpow pattern — FileGetContentsJitHelper → phpc_file_get_contents_kernel Internal whose call() lowers through thin libc JitFileGetContentsKernel (same as RenameJitHelper → phpc_rename_kernel). Then always implementPhpBridge (drop UserScriptAotDeferNestedJit), refresh FileGetContents helper-runtime unit, verify user-script AOT smoke prints contents.

[PurHur]

Blocker confirmed (this run) — Rename/Fpow Internal pattern does not unblock user-script AOT

Pre-flight: user_release_ready: yes · inventory OK 5121/5121 · north-star5-fast OK · vm-driver-probe OK.

Attempt

1. Added phpc_file_get_contents_kernel Internal (call() → JitFileGetContentsKernel::invoke libc leaf) — same shape as RenameJitHelper → phpc_rename_kernel.
2. Pointed FileGetContentsJitHelper at that Internal; removed UserScriptAotDeferNestedJit from StringFileGetContents (always implementPhpBridge).
3. Re-emitted ext_standard_FileGetContentsJitHelper_php under default PHP_COMPILER_AOT_USER_SCRIPT=1 emit and refreshed prelink.

Results

Config	User-script AOT echo file_get_contents($path)
Master (kernel defer)	hi
Always PHP bridge + new helper unit	empty
Emit helper without AOT_USER_SCRIPT	LLVM verify: Terminator found in the middle of a basic block! label %standalone_main

Narrower root cause

• User-script-emitted helper TU for readPathArgv begins with movabs $0x0 + call *%reg (null indirect calls) — same class as other helper TUs; embed/full-init nested compile is what makes Rename/Fpow helpers usable, but user-script still defers to a direct kernel for those builtins.
• Full-init helper-unit emit for this unit is still blocked by the %standalone_main terminator verify failure (prior report stands).
• Therefore deleting JitFileGetContentsKernel / removing defer remains unsafe until helper-unit emit produces a runtime-safe read leaf under user-script link.

Not merging. Reverted local WIP. Continuing on mechanical Builtin→ext kernel quarantine: #19555 (EnvironLibcWalkJit).