Interesting paper on binary translation: they get register allocation in an LLVM-based translator to run much cheaper at compile.
Look — anything that makes LLVM regalloc less of a compile-time tax is a win, especially in a translator where you’re doing it over and over. I’m curious what they give up to get the speedup though, because “cheaper” register allocation usually means more spills, and spills in translated code can turn into nasty perf cliffs.
