Hi! On Thu, May 28, 2020 at 04:22:16PM +0200, Richard Biener wrote: > For GIMPLE level transforms I don't think targets have more knowledge > than the middle-end.
Yes, certainly. > In fact GIMPLE complete unrolling is about > secondary effects, removing redundancies and abstraction. So IMHO > the correct approach is to look at individual cases and try to improve > the generic code Yep. > rather than try to get better benchmark results > on a per-target manner by magical parameter tuning. I'm no fan of that for target-specific code either. It's fine to be led by benchmarks, but usually a better justification is needed. > For what the RTL unroller does it indeed depends very heavily on > the target whether sth is beneficial or not. Yes :-( And this means it will need to remain late in the pass pipeline, or at least the decision needs to use target information (just like what ivopts does). > So I'd like to see specific cases where you think cunroll should > do "better" on powerpc only but not elsewhere. It is probably not a good idea in general to unroll 14 times, yes :-) Segher
