On Fri, Nov 11, 2016 at 7:55 PM, Jim Wilson <jim.wil...@linaro.org> wrote: > On Thu, Nov 10, 2016 at 2:53 AM, Richard Biener > <richard.guent...@gmail.com> wrote: >> The biggest "lack" of loop distribution is the ability to undo CSE so for > > I hadn't noticed this problem yet. I will have to take a look. > >> Then of course the cost model is purely modeled for STREAM (reduce the number >> of memory streams). So loop distribution is expected to pessimize code for >> the CSE case in case you are not memory bound and improve things if you >> are memory bound. > > I noticed this problem. I think loop distribution should be callable > from inside the vectorizer or vice versa. if a loop can't be > vectorized, but distributing the loop allows the sub loops to be > vectorized, then we should go ahead and ditsribute, even if that > increases the number of memory streams slightly, as the gain from > vectorizing should be greater than the loss from the additional memory > streams. We could have a cost model that tries to compute the > gain/loss here and make a better decision of when to distribute to > increase vectorization at the expense of the number of memory streams. > This looks like a major project though, and may be more work than I > have time for.
Yes. That's true for most enabling transforms (an easier one that comes to my mind is final value replacement, which, when required from the vectorizer could use a different cost model). Richard. > Jim
