The exciting thing would be to see applications compiled to LLVM IR get JIT'ed. Finally C or C++ apps that could inline through function pointer or virtual function calls at runtime.
It's an interesting idea but the current LLVM JIT is pretty slow so it hurts my head to think about JIT'ing average size C++ programs. Leaving that aside, the cost of doing this for every codepath would probably outweigh the benefits except in very specific use cases (especially on a cpu with decent indirect prediction). This would be an ideal place for a tracing JIT, which LLVM does not have at the moment.
