I'm one of the Souper developers and can offer a few pointers that may be more helpful than the github repo. Here's a somewhat recent blog post showing things that Souper can do:

  http://blog.regehr.org/archives/1252

Here's a talk I gave at the University of Washington last winter (before Souper did synthesis):

  https://www.youtube.com/watch?v=Ux0YnVEaI6A

Recently we've been teaching Souper to use dataflow facts such as LLVM's known bits and demanded bits. Perhaps interestingly, Souper can also teach LLVM how to compute these facts more precisely:

  http://lists.llvm.org/pipermail/llvm-dev/2015-September/089904.html

One really common question about Souper is "why operate on LLVM IR instead of instructions?" One of the main answers is "so that we can interact with dataflow analyses." This isn't something that previous superoptimizers have done. It seems to be working out really well.