Roberto Ierusalimschy, Luiz Henrique de Figueiredo, and Waldemar Celes have written a paper on the implementation of LUA (downloadable PDF). In it, they describe the register-based virtual machine design of LUA. Interestingly, I was watching Richard Hipp's talk on SQLite the other day, in which he says that he believes a register-based virtual machine is far easier to optimize than a stack-based design.
