...and that despite not being anywhere near as aggressive with exploiting UB as gcc or clang, which shows that backend-based optimisations like instruction selection, scheduling, and register allocation are far more valuable (and predictable).