> That's less efficient than retargeting the LuaJIT VM

That's like saying LLVM is inefficient because all the frontends compile to LLVM IR.

E.g. the Oberon compiler directly generates efficient LuaJIT bytecode also using FFI and even reuses the VM features for it's source-level debugger. Then benchmark written in Oberon has nearly the same performance like compiled to native when it runs on LuaJIT