I would simply design the CPU to run the LLVM IR instead of requiring a second step of compilation. reduce wasted time