Couple of thoughts:
- I wonder how relatively effective this optimization would be applied to profiled, typical "old school" C, written and tuned in an era when compilers were simpler and inlining of small functions—and other abstractions we would consider "zero cost" today—could not be assumed (and thus were often replaced by macros). This is the C of the 400-line function, which surely has a few hot paths and many cold paths as well.
- Are there existing or promising potential optimizations similar to this but without the profiling requirement? Are there effective heuristics for identifying likely "hot" code paths without requiring profiling or user annotation?