This new paper covers some state-of-the-art techniques:
https://www.eecs.berkeley.edu/~mangpo/www/papers/lens-asplos...
https://www.eecs.berkeley.edu/~mangpo/www/papers/lens-asplos...
If my understanding of Souper is correct, it's using synthesis to improve optimization rules inside LLVM. Lens, on the other hand, uses synthesis to optimize code directly with some neat algorithmic improvements to prune the search space. It targets ARM and, of all things, GreenArrays (a weird, extremely low-energy Forth-based architecture).
To be clear: I think both projects are interesting, as are other efforts in the field (ie Aiken at Stanford has done some cool stuff). Program synthesis is a topic I'm definitely interested in. But if you're interested more in LLVM and less in synthesis in general, this paper is going to be less interesting.
I had no idea Lens targets it. That's awesome.
I imagine ARM has more immediate practical applications and useful benchmarks, but it came second—I assume out of a collaboration with Qualcomm Research.