Repair the faithful-adopt-caner merge: 3 dropped lifter fixes + 6 dangling runtime refs

[sp00nznet]

Fixes for regressions introduced by the faithful-adopt-caner merge itself. Targets #82's branch so that PR isn't rewritten — merge or cherry-pick, whichever suits.

The merge resolution for tools/ppu_lifter.py was not a superset

git diff sagemono/spu/faithful-adopt-caner HEAD -- tools/ppu_lifter.py   # 28 lines

It took sagemono's file wholesale and discarded ~835 lines of the ydkj side. Three of the losses are silent miscompiles:

dropped	evidence in a fresh flOw lift
ctx->lr = <ret> on direct bl	every lr= field in the runtime logs printed 0x00000000; ctx->lr writes 140k → 29k
TOCFIX ld r2,N(r1) → literal module TOC	14,146 sites reloading r2 from uninitialised stack → garbage TOC → OPD loads read code-as-data
_mem_base() rA=0 D-form literal-zero base	loads/stores basing on ctx->gpr[0] (the _xea() X-form twin survived)

After relifting: 14146 TOCFIX / 0 stack-TOC reads / 120101 ctx->lr writes / 0 gpr[0] D-form bases, and the lr= diagnostics match the known-good ps3-draw lift exactly.

Still dropped, not restored here (flOw's EBOOT doesn't use them, so I can't verify them): VRSAVE mfspr/mtspr, stvlx/stvrx(l), vrfin/vrfiz/vrfip/vrfim, vexptefp, vlogefp. Worth a follow-up.

The runtime didn't link for any title

Undefined symbols in a .lib are only diagnosed when a title links it, so the archive built clean and this went unnoticed.

The SPU interpreter was not "consolidated into spu_channels.c" as 6dd8524 says — it was deleted, while both call sites in lv2_register.c stayed. Restored, along with spu_run_interp_job.

The function registry genuinely was consolidated, so restoring spu_fn_registry.c wholesale duplicates spu_register_function/spu_lookup/spu_begin_image; only spu_lifted_lookup was missing and it's re-added as the 3-line wrapper it always was.

lbp_hle_complete_pending, g_taskset_policy_bytes/size and g_vm_page_bitmap are title symbols referenced from shared runtime code. Each gets a default definition in its own one-symbol TU, per the existing spu_tsp_weak.c pattern — a port supplying the real one never pulls the archive member. They must stay separate files; a port can supply one without the others. g_vm_page_bitmap storage moves out of lbp/main.cpp, which is precisely why every non-LBP port failed.

Gotcha: the runtime CMakeLists globs runtime/*.c without CONFIGURE_DEPENDS — a new TU needs cmake -S . -B build re-run or it's silently never compiled.

Two more dropped/missing behaviours

• cellSaveData: ps3-draw returned CELL_OK on CBRESULT_ERR_NODATA with the comment "an ERROR return leaves the title parked in MODE_AUTO_LOAD". The merge restored the spec-correct CELL_SAVEDATA_ERROR_NODATA and flOw parked exactly as predicted — 0 flips, never reaching m_InitEntityHierarchy. Applied to both entry points: flOw calls the old non-_2 cellSaveDataAutoLoad, which 9db188d added as a separate function, so patching only AutoLoad2 does nothing. Per the SDK the error return is correct; that flOw can't handle it is an unexplained bug in its own state machine, and the code now says so.
• sys_ppu_thread: guest threads never got SetThreadStackGuarantee, so on a stack overflow the handler faults while reporting and the process dies with a bare AV inside the handler, losing the backtrace. main() already did this for the main thread.

Status

flOw builds and boots substantially further against this branch (209 → 306 log lines: savedata clears, thr_auto_load() completes), then hits a stack overflow before the app loop. FLOW_NO_SPU_WL=1 reproduces it, so it's the PPU lift. Lead for the follow-up: this lifter emits 7,915 g_trampoline_fn sites vs ps3-draw's 19,171, and 16 mid-function tail-entry wrappers vs 1,141, at near-identical function counts (38,287 vs 38,421) — so it's the branch routing, not the function list. The trampoline exists specifically to stop backward branches re-entering a prologue and growing the stack.

🤖 Generated with Claude Code