Translating code for one architecture into code for host (possibly same) architecture is what Qemu does. But Qemu does not do any significant optimizations in this process as it attempts to do this without extensive knowledge of host architecture (it does not even directly generate separate instructions).

Other thing that comes to mind is HP Dynamo (http://www.hpl.hp.com/techreports/1999/HPL-1999-78.html), which essentially is PA-RISC on PA-RISC optimizing JIT VM (that also does trace-driven optimization).