Add blueMSX to the ASYNCIFY-required cores list

[LokiCode404]

blueMSX reads machine configuration XML files synchronously during retro_load_game() to initialize hardware emulation. Without ASYNCIFY=1 the core deadlocks on the Emscripten event loop and never reaches its main loop, causing a silent content load failure where only the RetroArch XMB menu appears.

Related Issues

None directly — this is part of adding MSX/MSX2/SG-1000 support to EmulatorJS via the blueMSX libretro core.

Related Pull Requests

• Add MSX, MSX2 and SG-1000 support via the blueMSX core EmulatorJS#1244 — MSX system support + system_directory fix
• Add blueMSX core to the build system build#22 — Core build entry + system files packaging
• Add Emscripten build support and C-BIOS machine fallback libretro/blueMSX-libretro#207 — Core source changes (Emscripten flags + C-BIOS fallback)

[LokiCode404]

Pushed an update with some perf improvements for the blueMSX Emscripten build.
I found that full ASYNCIFY is a pretty significant bottleneck for this core — the Z80 CPU loop makes millions of indirect calls per second, and instrumenting all of them with save/restore code adds a lot of overhead.
What this commit adds to Makefile.emulatorjs:

• A MIN_ASYNC path that uses ASYNCIFY_IGNORE_INDIRECT=1 and an ASYNCIFY_REMOVE list to exclude the CPU execution engine, VDP renderers, sound chip updates, and memory slot manager from instrumentation. Only the init/load path actually needs to yield.
• A USE_JSPI mode as a third option alongside ASYNC and MIN_ASYNC. When USE_JSPI=1 is set, it uses -sJSPI=1 instead of -sASYNCIFY, which eliminates instrumentation overhead entirely — the VM handles stack switching natively.
• JSPI added to the audio external-block exemption chain (same behavior as full ASYNC — no workarounds needed).
  And in build-emulatorjs.sh:
• For minAsync cores (blueMSX), the script now produces two variants: JSPI → bluemsx-wasm.data for modern browsers, MIN_ASYNC → bluemsx-legacy-wasm.data as fallback.
  On the audio side, I disabled HAVE_AL for the MIN_ASYNC builds because the OpenAL codepath force-overrides ASYNC=1, which defeats the whole point. Using HAVE_RWEBAUDIO=1 instead works fine with the fake-block audio path.