That issue is pretty rare—a tight 'loop' where you don't know where you're coming from (aka the previous instruction) or where the next 'loop' iteration goes (aka the next instruction). Because this code block is bounded by indirect jumps, it's virtually impossible to register allocate without agreeing on a mini ABI for this scenario. Not surprisingly, this is unusual enough it's inexpressible in C. However, there are things that provide a middle ground, like computed GOTOs, which both the ocaml and python interpreter use now for bytecode evaluation.

Of course, I suspect if you pass luajit an unusual program (say, an unusual distribution of instructions), it would actually perform worse. Register allocation is np-complete, so ultimately with modern programs it's dependent on really good heuristics, and C wasn't written to be a bytecode evaluator.

> Of course, I suspect if you pass luajit an unusual program (say, an unusual distribution of instructions), it would actually perform worse.

I think that's unlikely. The LuaJIT interpreter keeps all important state in registers; this is not affected by the sequencing of bytecodes.

> The LuaJIT interpreter keeps all important state in registers; this is not affected by the sequencing of bytecodes.

Except on register-starved architectures like, say, i686.