Hi, On Thu, 31 Mar 2011, Jeff Law wrote:
> Without trip count estimates the speculative read can introduce a > performance regression. Like any other transformation. So what are you trying to say? That therefore introducing a mode in GCC disallowing any speculative reads (even if they happen to improve performance!?) is a bright idea? I sure hope not. The right way to deal with performance concerns is to model them and (try to) avoid them. Not to introduce random restrictions onto a (timing independend) as-if model. So, please let's leave out any disguising arguments about performance in connection with the cxx mem-model, shall we? It has different goals AFAIU, and if it has to resort to performance for showing its usefulness it's bound to fail already. Ciao, Michael.
