This is a very sound approach. The main benefit is that all JVM languages can utilize optimizations if the optimizations are in the JVM, thus logic duplication is reduced.

Also, please note that both Clang and GCC use the same approach, except in their cases the programmer typically does not see the intermediary code (which is LLVM for Clang, and a family of intermediate languages for GCC).

http://en.wikipedia.org/wiki/Intermediate_language