Benchmark would not be relevant here. Souper is intended to detect missing optimizations that could have applied to LLVM code already generated by, say, Clang, and to point out new patterns to recognize and transform to the LLVM developers. The improvements brought by Souper are distilled over a long period and mixed with other unrelated LLVM improvements. On the other hand, all improvements suggested by Souper have looked like a good idea to an LLVM developer who let themselves be convinced at one point, and many of them simply remove instructions and at least reduce the size of the generated binary, even if nothing else.

From the github page:

>Alternatively, you may immediately let Souper modify the bitcode and let it apply the missed optimization opportunties by using the Souper llvm opt pass.

So it seems like a benchmark of the program before and after this process would be relevant.