While the Java implementation is suboptimal, there is really no need for it to be that way. I think the ideal way to go about it would be to run the compiler optimisations and whatnot then generate something semantically similar to C89 as output. Then you invoke a simple compiler with few optimisations on the target machine the first time the program is run, and cache the results somewhere. On all subsequent runs, you've got something which can actually run faster than pre-compiled C code because the compiler doesn't need to preserve the ABI so can do stuff like inlining dynamic library calls.
Sadly not. Wasm is attempting something similar, but it lacks certain things that would be important for this (the to specify in one module a type of unknown size and then query its size in another module at link time).
