https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77536
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2016-09-14
CC| |hubicka at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think the reason of the current behavior is that without a profile we predict
loops to run for 10 iterations (or so), thus after vectorization with a
vectorization factor of two you'd get two vector iterations out of that guess.
Honza fixed the profile updating but the real issue is probably the frequencies
we use for the extra tests the vectorizer performs which seem to be all
66% / 33%:
/* There are many aspects to how likely the first loop is going to be
executed.
Without histogram we can't really do good job. Simply set it to
2/3, so the first loop is not reordered to the end of function and
the hot path through stays short. */
int first_guard_probability = 2 * REG_BR_PROB_BASE / 3;
int second_guard_probability = 2 * REG_BR_PROB_BASE / 3;
there isn't any attempt to distinguish the different cases. Even with that
fixed the low number of predicted iterations for the scalar loop make
coming up with a "hot" profile for the vectorized loop difficult.
