Distribution: platform matrix — linux/arm64 first (per-arch helper cache, CI on an arm64 runner or QEMU), then macOS arm64; a target-triple abstraction in Linker/TargetMachine instead of x86_64 assumptions (prelinked/helper-runtime/<arch>/, lib/AOT/Linker.php, lib/JIT/Context.php)

[PurHur]

Category

Foundation: · portability · child of #36379

Why

prelinked/helper-runtime/ has exactly one directory: x86_64-linux. The object emit uses MCJIT defaults for the host (no explicit triple/CPU, #36200 landed lib gating only), the linker hardcodes crt/ld paths for Ubuntu 22.04 x86_64, and the Docker image is x86_64 only. Every developer laptop is arm64 now and Graviton is the cheapest cloud CPU; a compiler that runs on one architecture is a demo.

Deliverable

1. Target abstraction: PHP_COMPILER_TARGET=x86_64-linux|aarch64-linux|aarch64-darwin selects triple, CPU (generic), reloc model, data layout, Linker toolchain paths, and the helper-cache directory; emitted objects are deterministic per target (Build: every AOT binary links libsodium/libargon2/libbz2/libcrypto/libz/libpcre2 unconditionally, and every compiler process re-parses an 816 KB LLVM FFI header (lib/AOT/Linker.php, vendor/ircmaxell/php-llvm/ffi/llvm9.php) #36200 reproducibility item).
2. aarch64-linux: build the dev image for arm64 (docker buildx --platform linux/arm64), emit and commit/publish the helper cache for it, run aot-smoke + differential sweeps on an arm64 runner (GitHub-hosted arm64 or QEMU with a longer timeout) weekly.
3. aarch64-darwin: LLVM 9 on macOS is hard; evaluate whether this waits for the LLVM 22 migration (Foundation: LLVM 9 → 22 migration Phase 0 — prove the FFI binding can be generated against LLVM 22 headers and thread pointee types on LLVM 9 (docs/roadmap/LLVM22-MIGRATION.md) #36220) — record the answer in the ADR.

Done when

• aot-smoke.sh 9/9 on linux/arm64 in CI; helper cache published for both arches; ghcr.io/purhur/phpc is multi-arch
• Target triple, CPU and toolchain are data, not code paths; phpc doctor prints the target

[PurHur]

claim: runtime-migration-worker — starting this run (PHP_COMPILER_TARGET abstraction + helper-cache dir selection; x86_64 default stays green)

[PurHur]

Merged: #36447 (Part of #36391)

Slice landed: PHP_COMPILER_TARGET → CompileTarget selects helper-cache arch dir + Linker toolchain data; cross-link refused; doctor prints target. aot-smoke 9/9; north-star5-verify-fast OK; CompileTargetTest 5/5.

Still open on Done-when: aarch64 helper-cache emit, arm64 aot-smoke/CI, TargetMachine-driven object emit (today still host MCJIT).

[PurHur]

claim: runtime-migration-worker — continuing #36391 this run. Next slice: initialize the LLVM backend for PHP_COMPILER_TARGET (not only initializeNative), apply triple/data-layout on standalone modules, refuse host-MCJIT fallback on cross-target, prove aarch64 ELF object emit on x86_64 without linking (no full helper-runtime prelink).

[PurHur]

Merged #36506 — slice: LLVM backend + triple/layout for PHP_COMPILER_TARGET; cross-target refuses host MCJIT fallback. Verified aarch64 ELF e_machine=183 on x86_64 host (CompileTargetTest 10/10); aot-smoke 9/9; north-star5-verify-fast OK.

Still open on Done-when: committed aarch64 helper-cache units, arm64 aot-smoke CI, multi-arch ghcr.io/purhur/phpc.

[PurHur]

claim: runtime-migration-worker — continuing #36391 this run after #36506. Next slice: publish a seed aarch64 helper-runtime unit (ELF e_machine=183) + committed-tier arch gate; no full --prelink (20 min / 8g cap).

[PurHur]

Merged: #36522 (Part of #36391)

Slice: seed aarch64 helper-runtime unit + ELF e_machine gate + PLATFORMS for multi-arch docker-publish.

Verified:

• ./script/phpunit.sh --filter CompileTargetTest → 12/12 OK
• ./script/aot-smoke.sh → 9/9 + size gate OK
• php script/check-helper-runtime-prelink.php --all-arches → aarch64 1 fresh, elf ok (e_machine=183)

Still open on Done-when: full aarch64 helper corpus, arm64 aot-smoke runner, published multi-arch ghcr.io/purhur/phpc.

[PurHur]

claim: runtime-migration / build-architecture — starting this run after #36522. Done-when still open: aot-smoke on linux/arm64 (QEMU or runner), fuller aarch64 helper-cache tier, multi-arch ghcr. Next: measure master target/doctor/helper-cache gaps and land cheapest honest slice under 8g/20min (no full --prelink).