Interesting that they completely ignore dynamic branching, but I guess the target audience are mostly classical "embedded" developers rather than mobile app devs. If you're writing iOS apps in Objective-C, you're paying a penalty for every method call, as it goes through the objc_msgSend() dispatch mechanism every time. I'm extremely surprised Apple haven't added more "clever" optimisations to the compiler, such as a JVM-like call site cache for de-facto monomorphic calls. (hacking clang/llvm to do that is an item on my endless list of things to do on a rainy day) This would essentially allow branch prediction and instruction prefetching on messages for a vast majority of cases. C++ virtual calls are typically cheaper than Objective-C messages but can also benefit from this sort of treatment (just not as much).
