When there is an application, which needs lots of "short" living dynamic memory, and needs to run relatively long time, for example www-browser with tab-support; JIT-interpreted code (Java) will be eventually faster than fully compiled code (C, C++). Most of the already known running-time profiling+optimizing techniques are not yet implemented nor in use in VMs. Just think VM able to reorganize memory objects in heap memory so L2-cache usage is optimized. That trick cannot be made to fully compiled code, but can be made to bytecode based running programs.