I've used Souper a bunch of times on hand-written, performance-critical LLVM code, hoping for some magic superoptimal version of the function. Even saving a few

I've used Souper a bunch of times on hand-written, performance-critical LLVM code, hoping for some magic superoptimal version of the function. Even saving a few instructions would help in these cases. Unfortunately, in my experience Souper is the identity function: it has literally not once changed my LLVM code in any way. If I'm doing something wrong, I'd love to know what. Perhaps it's not expected that hand-written LLVM will have much room for improvement? Is the intended target of optimization only expected to be fairly obviously redundant machine-generated code?