Optimizing LLVM Back-End: Global Instruction Selection