To improve JVM start-up, it may make sense to pre-generate the bytecode interpreter. I've measured the time spent in generating the different parts of the interpreter:

http://hg.openjdk.java.net/jdk/jdk/file/2fbc66ef1a1d/src/hotspot/share/runtime/init.cpp#l108

Time spent in stubRoutines_init1() = 0.380ms (539365 instructions)
Time spent ininterpreter_init() = 1.343ms (3628038 instructions)
Time spent in SharedRuntime::generate_stubs() = 0.173ms (313958 instructions)
total = 1.896ms (4481362 instructions)

Total time of "java -XX:-UsePerfData -cp . HelloWorld" = 43.117ms

1.896 / 43.117 = 4.4%

Because the contents of the interpreter depends on VM options, we would need to pre-generate several versions for popular VM option combinations.