That's not a terrible goal if you want to make key libraries faster and more efficient.
But I'm unconvinced that it would be easy to scale up these techniques to auto-clone and improve a big application like Photoshop or Word - or even a generic CRUD app.
One advantage of peephole optimization is that pretty much all code is amenable to it, so it doesn't rely on their being one key, hot loop to optimize to see performance gains. The main disadvantage is that gains in peephole optimization tend to be quite small, so you're combing for single-digit percentage speedups.
"In addition to searching over programs, STOKE contains verification infrastructure to show the equivalence between x86-64 programs. STOKE can consider test-cases, perform bounded verification all the way to fully formal verification that shows the equivalence for all possible inputs."
Definitely an interesting though extremely challenging problem.
The original STOKE paper [1] only worked for loop free programs, which may be a program fragment for which program equivalence is decidable. Be that as it may, there is a fast heuristic for proving non-equivalence between two programs P and Q, just feed them with random input -- if they return different outputs, they are NOT equivalent, if OTOH they return the same results, they they are candidates for equivalence and you can throw a modern SMT solver at them which may actually be able to prove equivalence (or come up with a counterexample). Otherwise the SMT solver times out. You can even refine this scheme into some form of local search, in that you measure how much P and Q differ (e.g. Hamming distance). This is used in later papers, e.g. [2].
[1] Stochastic Superoptimization – ASPLOS 2013 https://raw.githubusercontent.com/StanfordPL/stoke/develop/d...
[2] Stochastic Program Optimization – CACM 2016 https://raw.githubusercontent.com/StanfordPL/stoke/develop/d...