That's because the authors are reporting results for AOT scheme to c compilers.

For languages more directly supported by llvm/gcc, they already do various kinds of function versioning and specialization.

You would be highly unlikely to get 10% there from a magic bullet like this - in the end, there are no magic bullets. Just lots of hard work.

Mmm, my impression is you can get 90% there with significantly less effort then LLVM or GCC. A good register allocation strategy, some basic inlining, maybe peephole optimization? LLVM is fighting for 5-10% improvements on top of the big things

The key here is that this is meant to be used for dynamic languages that don't have a lot of typing info when doing AOT compilation