Optimize or Wait? Using llc Fast-Prototyping Tool to Evaluate CUDA Optimizations