> There was a fashion for running your C instrumented and then recompiling it after runtime branch choice evidence was collected. I think the Dec Alpha OSF/1 compiler did it optionally.
This is called profile-guided optimization and new compilers (well, GCC and LLVM) have it.
I mostly think it's bad and not statistically sound; given profile data the compiler is both overly trusting of it (no error bars) and can't find much useful to do with it (because there's no way to hint a modern OoO CPU).
This is called profile-guided optimization and new compilers (well, GCC and LLVM) have it.
I mostly think it's bad and not statistically sound; given profile data the compiler is both overly trusting of it (no error bars) and can't find much useful to do with it (because there's no way to hint a modern OoO CPU).