In recent years, it seems to have become somewhat unfashionable to build implementations of logic programming languages by translating them to Prolog. Instead, implementors appear, more and more, to be designing their own abstract instruction sets and indulging in a great deal of slow and painful low-level hacking. This paper argues that in many cases, it is preferable to build systems by translating programs to Prolog, using a good underlying Prolog system, and using dataflow analysis and high-level optimizations to reduce performance overheads. In support of our arguments, we compare two sequential implementations of committed choice languages: QD-Janus, which translates to Prolog, and FCP(:), which compiles to a low-level byte-coded instruction set. Even though QD-Janus took significantly less time to implement, its performance is significantly better than that of FCP(:).