Unladen Swallow might not be usable in production, but as a research project it's pretty interesting. One of its problems was supporting C-based extensions, but that's not a requirement for Emacs Lisp, and so a similar project might be more effective for Emacs than for Python. (I also like the idea of using LLVM for dynamic language runtimes, but as time goes forward this seems less likely to work out.)
This has actually been done (possibly badly?) and turned out to be a wash in terms of performance.