Apple Silicon (macOS arm64) dynarec port

[AurisDSP]

Summary

Brings the aarch64 dynarec online on macOS Apple Silicon (M1/M2/M3/M4). Resolves #561.

The host code ports cleanly with a handful of platform shims:

• Build (Makefile): enable WITH_DYNAREC=aarch64 on osx for native arm64; keep -DNO_ASM only when no dynarec backend is in use.

• Assembly (linkage_arm64.S): Apple's clang assembler treats ; as a comment delimiter, so the GAS-style cpp macros that emit multiple directives on one line don't work. Switched GLOBAL_FUNCTION / LOCAL_FUNCTION / GLOBAL_VARIABLE to assembler-side .macro definitions on Apple, with an _-prefixed Mach-O symbol and .private_extern. Added an ADRP_LO12 macro that selects between :lo12: (GAS) and @PAGE/@PAGEOFF (Apple) syntax. Internal b.ne new_dyna_stop was rerouted through a local L-label so Apple's assembler doesn't reject the conditional branch's external target.

• JIT memory (new_dynarec.c): macOS Hardened Runtime forbids RWX in a single mapping, so on Apple we mmap the cache with MAP_JIT and toggle protection per-thread via pthread_jit_write_protect_np(). To handle set_jump_target() being called from inside new_recompile_block (where the page is already in write mode), the BEGIN/END pair nests via a per-thread depth counter (`_Thread_local`); END clamps to zero so cache_flush() called from invalidation paths still lands the page in execute mode.

• W^X helpers (new_dynarec.h): `M64P_JIT_WRITE_BEGIN` / `M64P_JIT_WRITE_END` macros wrap `pthread_jit_write_protect_np` with the nesting counter on Apple, no-ops elsewhere. The counter is `_Thread_local` since the underlying syscall is per-thread.

• Cache flush (assem_arm64.c): on Apple, replaced the manual `dc civac` / `ic ivau` loop with `sys_icache_invalidate()` and a paired `M64P_JIT_WRITE_END`. `arch_init()` and `set_jump_target()` now wrap their writes with BEGIN/END.

• Portability (custom/dependencies):

  • `pngpriv.h`'s `TARGET_OS_MAC` branch pulls in `<fp.h>` (a Mac OS 9 header), so it's been excluded on modern macOS.
  • `zutil.h`'s `MACOS` branch redefines `fdopen`, which collides with Apple's stdio.h `__memcpy_chk`-style macros; skip the redefinition on `APPLE`.

The second commit is hardening that fell out of an audit while looking at the Apple AudioConverter path — same fixes apply to the non-Apple sinc-resampler branch (NULL checks after malloc, zero-frequency guard, oversized-input truncation, drain-loop break on `audio_batch_cb` returning 0).

What works

• Mario Kart 64 boots and plays with audio + video on M-series hardware (`cached_interpreter` and `dynamic_recompiler` both work; `angrylion` + HLE RSP avoids a paraLLEl-RDP / MoltenVK threading bug we'll file separately).
• Apple Silicon dylib loads cleanly with both video drivers (`glcore` and `vulkan` via MoltenVK).
• All Apple-specific code is guarded behind `APPLE` so Linux / Windows / Android paths are unchanged.

What doesn't work yet (out of scope)

Some homebrew ROMs (e.g. BikeRace '98, Wet Dreams POM '99) hit a SIGSEGV inside generated code with `LR=1`-shaped corruption. The crash isn't macOS-specific — it reproduces on this branch's tree on Linux with the same recompiler — so it's a pre-existing aarch64 codegen edge case rather than something the macOS port introduced. Worth filing separately.

Tested on

• M4 MacBook Air, macOS 26.x, RetroArch 1.22.2

Test plan

• Cached interpreter mode runs Mario Kart 64 end-to-end
• Dynamic recompiler mode runs Mario Kart 64 end-to-end (boot, audio, video, input)
• Build with `make platform=osx` succeeds clean
• Linux paths unchanged (all changes are `APPLE`-guarded)

[m4xw]

This will take some time to review