There are times squeezing every last bit of performance is great. I am saying is in general there is no need to run -O3 on a 500 line function which is run once

There are times squeezing every last bit of performance is great. I am saying is in general there is no need to run -O3 on a 500 line function which is run once at initialization compared to a compute kernel. Our compilers and languages don't have the granularity to specify this currently. I do want more specialized optimization tools however. Things like automatic super optimizers [1] which can be targeted by programmers with special directives. [1] http://theory.stanford.edu/~sbansal/superoptimizer.html http://theory.stanford.edu/~sbansal/superoptimizer.html