On 02/14/2017 07:57 AM, Jan Hubicka wrote: > So it seems that the frequency of the loop itself is unrealistically scaled > down. > Before vetorizing the frequency is 8500 and predicted number of iterations is > 6.6. Now the loop is intereed via BB 8 with frequency 1148, so the loop, by > exit probability exits with 15% probability and thus still has 6.6 iterations, > but by BB frequencies its body executes fewer times than the preheader. > > Now this is a fragile area vectirizing loop should scale number of iterations > down > 8 times. However guessed CFG profiles are always very "flat". Of course > if loop iterated 6.6 times at the average vectorizing would not make any > sense. > Making guessed profiles less flat is unrealistic, because average loop > iterates few times, > but of course while vectorizing we make additional guess that the > vectorizable loops > matters and the guessed profile is probably unrealistic.
We have the same problem in the RTL loop unroller in that we'll scale the unrolled loop by the unroll factor (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68212#c3), which can result in a loop with lower frequency than surrounding code. Problem is compounded if we vectorize the loop and then unroll it. Whatever approach is decided for the case when we have guessed profile should be applied to both vectorizer and RTL loop unroller. -Pat
