I'm not following you. Clearly in a pure JIT environment or with custom calling conventions you can do better than a typical "blind" C call restricted by the pl

I'm not following you. Clearly in a pure JIT environment or with custom calling conventions you can do better than a typical "blind" C call restricted by the platform ABI - but we're talking about FFI calling into a "plain" shared object, right? Such functions obey the platform ABI, so the caller needs to save exactly every callee-clobbered register they are using. Also, if the external C function wants to use any callee-saved registers, it needs to save them. It doesn't know that it's being called by a VM that allows FFI functions to clobber everything (if that's what you're saying). So yeah, there is all kinds of great stuff you can do with respect to function calls within a VM, or when you are working outside the platform C ABI (e.g., if you invented your own ABI), but it isn't clear how that relates to making fast FFI calls from higher level languages.