I assume you would get another result running all CLBG algorithms and comparing the geometric means. I also run benchmarks including (hashed) table/field access and was virtually as fast as the same program when compiled natively, see https://github.com/rochus-keller/Oberon/blob/master/testcase.... Of course there are parts where LuaJIT 2.0 is slower than native, but there are also some opposite cases.