psxrecomp/docs/ASYNC_OVERLAY_COMPILE.md

Async background overlay compilation — implementation plan

Status: LARGELY ALREADY BUILT + VALIDATED under CPS (2026-06-18). The "planned" feature below turned out to mostly exist: overlay_capture.c (overlay_autocapture_tick — autocapture on interp pressure, ON whenever the cache is on) + autocompile.c (background compile: probes for a C compiler on PATH, spawns cmd.exe /C <configured cmd> on a watcher thread, applies on the emu thread) + the code_provider abstraction (gcc vs sljit). It is FILE-BASED exactly as planned — the worker writes DLLs; the dispatch thread loads them. Config-driven via [runtime] overlay_autocompile_cmd.

What this session ADDED to make it work under CPS + match the user's gcc-only preference:

1. compile_overlays.py --cps (committed, §25.5) — the autocompile cmd passes --cps.
2. overlay_loader.c overlay_loader_apply_live_policy — when the active backend is gcc, sljit-live is forced OFF (gcc>interp; toolchain-less stays sljit>interp) — "mutually exclusive" per user.
3. The dev/soak config (tomba_soak_a.toml, untracked): overlay_cache=true + overlay_autocompile_cmd = python ../psxrecomp/tools/compile_overlays.py … --cps. The PRODUCTION game.toml overlay_autocompile_cmd needs --cps appended when CPS ships to master.

VALIDATED: autocapture fired (triggers≥1) → background compile_overlays --cps produced CPS overlay DLLs (new CRCs, post-cache-clear) → loader loaded them (loads>0) → dispatch_native rises, gcc box shows 0 sljit shards. ce_profile stays flat. As you PLAY, the cache auto-builds and each scene goes native after its autocapture cycle (~60s cooldown) + compile.

REMAINING / FUTURE: broader capture coverage is just play-time (each scene's overlays auto-compile on first sustained visit); optionally raise the autocapture cadence / cap; the in-process sljit tier (toolchain-less users) is still leaf-only under CPS (call-fragments interp) until the per-fragment entry-switch codegen lands. The DETAILED design below is retained as reference.

---

(original plan — retained for reference; most is already implemented as above)

1. Goal (user's words)

"I would like it to be (relatively) fast (via interp), and async compile in the background, and when it's in a healthy spot, the next time it's called, it finds the compiled shard and does that instead. So there shouldn't be visible user lag."

"gcc for users that have it. for those who do not, sljit. Always async. if the user has gcc, no need to also do sljit. mutually exclusive basically."

Classic JIT tiering: interp tier (immediate, no stall) → background compiler → native tier (picked up on the next dispatch). gcc and sljit are mutually exclusive backends, chosen by toolchain availability. Always off the dispatch (guest) thread.

2. The measured problem this fixes

Title screen under CPS: dispatch_interp_fallback ≈ 12.2M vs dispatch_native ≈ 14.8K — almost entirely interpreted, because the title's overlays are NOT gcc-compiled (only the kernel

• the village 0xE7000 overlay were, from a short capture). Two slownesses:

1. Sync compile hitch: try_sljit_region (overlay_loader.c) JITs a region's shards synchronously on the dispatch thread on first miss → a visible spike.
2. Steady interp: call-heavy overlays decline the scoped sljit (§25.5) → interp forever, and uncompiled regions have no gcc DLL → interp. In-game (village) feels fine because 0xE7000 is gcc-native. So: get overlays native, off-thread.

3. Architecture — FILE-BASED HANDOFF (the key safety property)

The worker produces ARTIFACT FILES (a .dll or a .sljit blob in the cache). The dispatch (guest-fiber) thread does ALL candidate registration via the EXISTING load paths (try_load_region for gcc DLLs, overlay_sljit_deserialize for sljit blobs). The worker NEVER touches s_cand / idx_* / the sljit in-memory registry. This keeps the candidate table single-threaded — no locks on the dispatch hot path, no torn reads.

guest fiber (dispatch)                     worker thread (background)
─────────────────────                      ──────────────────────────
overlay_loader_dispatch(addr):
  head<0 & in-window & not-done?
    → interp now (fast)                ──enqueue {phys, crc, snapshot}──▶ work queue
    → (try_load_region already                                          pop → compile:
       re-checks the cache each miss)                                     gcc:  compile_overlays pipeline → <crc>.dll
  next miss of same region:                                              sljit: JIT from SNAPSHOT → serialize <crc>.sljit
    try_load_region finds the new   ◀──────────── file in cache ───────  (atomic rename into place)
    .dll/.sljit → loads + registers
    (on THIS thread) → native

4. Components to build

4a. Toolchain probe (startup, once)

• gcc tier requires: the recompiler binary (psxrecomp-game.exe), a C compiler (gcc), and the compile_overlays.py driver + a Python interpreter — OR reimplement the driver in C (heavier). Probe: gcc --version exits 0 AND the recompiler + script paths exist.
• If gcc tier unavailable → sljit tier.
• Expose paths via a [overlay_compile] block in game.toml (recompiler, script, python, runtime_include, cache_dir) OR auto-derive relative to the detected project root (memcard_paths already resolves a project root via .git/CMakeLists.txt — reuse that). Shipped builds: no toolchain present → sljit tier automatically.
• Store the chosen tier in a global; the worker branches on it.

4b. On-miss enqueue (overlay_loader.c, overlay_loader_dispatch)

• Where: the head < 0 && s_active && overlay_cache_window_contains(phys) block, alongside the existing try_load_region. Replace the SYNCHRONOUS try_sljit_region with: enqueue the region for async compile + fall through to interp (return 0 → dirty_ram interp).
• Region key = the overlay CRC (per-variant; same model as the offline cache — same addr can be different code in different scenes → key by CRC, additive cache).
• Dedup: a per-CRC state map { QUEUED, COMPILING, DONE, FAILED }. Don't re-enqueue QUEUED/ COMPILING/DONE. FAILED → backoff (retry after N misses, or never).
• Snapshot the overlay bytes AT ENQUEUE (memcpy the region out of guest RAM on the dispatch thread) so the worker never reads live guest RAM (avoids the RAM race). The existing capture (overlay_capture.c → overlay_captures.json) already snapshots — reuse/extend it.

4c. Worker thread

• One std::thread (or pthread) started at init when a tier is active. Condition-variable wait on the queue. Clean shutdown on exit (join; the runtime already has a shutdown path).
• gcc tier: write a SINGLE-region captures.json (from the snapshot) to a temp, run python compile_overlays.py --captures <tmp> --cps --out-dir <cache> --recompiler ... --game-toml ... --runtime-include ... (the --cps flag + codegen_ver=4 namespace already exist). The pipeline writes <crc>.dll into cache/<game>/gcc/<arch-abi>/cg4/. Mark DONE.
  • Incremental is built in (compile_overlays skips existing DLLs unless --force), so it's safe to run per-region or batched.
  • Atomicity: compile_overlays writes the .dll directly; if the loader could observe a partial file, compile to a temp + rename. (Check whether compile_overlays already does; if not, add a rename step — DLL load of a half-written file would be bad.)
• sljit tier: compile from the SNAPSHOT (not live RAM). NOTE the current overlay_sljit_try_compile reads cpu RAM (fetch_word) — add a variant that takes a byte buffer + base addr, and a SERIALIZE-ONLY mode (write the .sljit blob, do NOT call register_sljit_candidate). The dispatch thread deserializes + registers on next miss (overlay_sljit_deserialize path in overlay_loader.c). Confirm the sljit compiler instance is per-call (it is — sljit_create_compiler per fragment) so worker-thread use is safe.

4d. Loader pickup (mostly already works)

• try_load_region(phys) already scans the cache index and loads matching DLLs on miss → it will pick up worker-produced DLLs with no change, IF the cache index is refreshed. CHECK: does the loader cache its directory listing once at init, or re-scan on miss? If cached, add a "cache dirty" flag the worker sets (via an atomic) so the dispatch thread re-scans.
• For sljit blobs: ensure there's a deserialize-on-miss path keyed by CRC (the persist cache already deserializes at init — extend to deserialize-on-miss for blobs the worker just wrote).

5. Thread-safety checklist

• s_cand / idx_* / sljit in-memory registry: touched ONLY by the dispatch thread (file handoff).
• Guest RAM: worker reads ONLY the enqueue-time snapshot, never live RAM.
• Queue + per-CRC state map: guarded by one mutex (cold path — only on overlay miss + worker pop).
• File writes: atomic (temp + rename) so a partial artifact is never loaded.
• Worker lifetime: started after overlay_loader_init, joined on shutdown_runtime.
• The 4-second starvation watchdog: the worker is a separate thread, so a long gcc compile won't starve the guest fiber — but DON'T block the dispatch thread waiting on the worker.

6. Files

• runtime/src/overlay_loader.c — enqueue on miss (replace sync try_sljit_region), per-CRC state, cache re-scan flag.
• runtime/src/overlay_compile_worker.c (NEW) — the worker thread, queue, tier dispatch, subprocess spawn (gcc) / snapshot-sljit (sljit).
• runtime/src/overlay_sljit.c — add compile-from-buffer + serialize-only variant.
• runtime/src/overlay_capture.c — reuse/extend the snapshot for single-region captures.
• config_loader (+ game.toml) — optional [overlay_compile] paths; else auto-derive.
• Windows: subprocess via CreateProcess (or _popen); POSIX via fork/exec. There may already be a subprocess helper; otherwise keep it minimal + Windows-first (the dev target).

7. Validation

• Title screen: after a few seconds of background compile, dispatch_interp_fallback stops climbing and dispatch_native rises; the title visibly smooths out. No frame hitch at the moment of compile (it's off-thread).
• ce_profile stays FLAT (the leak fix must not regress).
• No crash through dwarf→overworld (the continuation-routing path) with overlays going native mid-scene.
• Toggle: force sljit tier (hide gcc) → still no lag, leaf shards native, calls interp.
• Shipped-like (no toolchain) → sljit tier auto-selected, no errors.

8. Risks / gotchas

• Running compile_overlays.py per-region is slow (Python + recompiler + gcc, seconds each). Batch or rate-limit; it's async so latency is OK, but don't spawn dozens concurrently — cap to 1 in-flight compile.
• Per-CRC variant explosion (the §coverage-campaign long tail): many scene variants. Cap the queue; LRU-evict the per-CRC DONE set if needed (the on-disk cache is the source of truth).
• compile_overlays needs the SAME --cps + codegen_ver as the runtime (cg4). If they drift, the loader's codegen_ver check rejects the DLL → wasted compile. Keep PSX_OVERLAY_CODEGEN_VER in lockstep (it already gates both).
• The captured snapshot must match the live bytes the loader will CRC-check at load (cand_crc) — capture at the same moment / validate CRC before enqueue, else the DLL is rejected on load.
• Do not regress the committed CPS leak fix: everything here is additive + gated on the active tier; legacy (non-CPS) overlays still use the existing sync paths.