The simple explanation why is that both problems are very hard to solve in isolation. Combining the problems makes it even harder. That said, of course it is b

The simple explanation why is that both problems are very hard to solve in isolation. Combining the problems makes it even harder. That said, of course it is better to try to solve the real problem and not only the decomposition. I think that the focus on cheap compiler passes is a bit obsolete now. A cheap debug-build with some optimizations is good, but for a real production build where speed is needed I am willing to wait quite some extra time to get the last few drops of performance. I've seen some researchers looking into solving the combined problem which feels promising. That kind of interest combined with the nice modularity of LLVM makes me quite optimistic that there will be nice results that are relevant both academically and industrially.