https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66646
--- Comment #4 from rguenther at suse dot de <rguenther at suse dot de> --- On Wed, 20 Apr 2016, amker at gcc dot gnu.org wrote: > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66646 > > --- Comment #3 from amker at gcc dot gnu.org --- > BTW, shall we version the loop in ldist pass if # of iterations can't be > computed at compilation time. We can check the target dependent value when > deciding which version should be executed. It doesn't make sense to turn > loops > into memory functions, while trying to inline small memory functions at the > same time. Well, the idea is that the _target_ knows how to inline it (and whether to inline it) best. A stupid memset loop using chars can be very well expanded faster like with a rep; movb; Of course the issue here is that memmove doesn't have a target specific expansion way. That's something we should eventually fix. That said - heuristics are hard - w/o transforming into a call we completely peel / jump-thread the whole thing which is obviously better.
