https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102383
Kewen Lin <linkw at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |linkw at gcc dot gnu.org
--- Comment #4 from Kewen Lin <linkw at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #2)
> The issue is that we tame PRE because it tends to inhibit vectorization.
>
> /* Inhibit the use of an inserted PHI on a loop header when
> the address of the memory reference is a simple induction
> variable. In other cases the vectorizer won't do anything
> anyway (either it's loop invariant or a complicated
> expression). */
> if (sprime
> && TREE_CODE (sprime) == SSA_NAME
> && do_pre
> && (flag_tree_loop_vectorize || flag_tree_parallelize_loops > 1)
> && loop_outer (b->loop_father)
> && has_zero_uses (sprime)
> && bitmap_bit_p (inserted_exprs, SSA_NAME_VERSION (sprime))
> && gimple_assign_load_p (stmt))
>
> the heuristic would either need to become much more elaborate (do more
> checks whether vectorization is likely) or we could make the behavior
> depend on the cost model as well, for example exclude very-cheap here.
> That might have an influence on the performance benefit seen from
> -O2 default vectorization though.
>
> IIRC we suggested to enable predictive commoning at -O2 but avoid
> unroll factors > 1 when it was not explicitely enabled.
>
Yeah, it's PR100794. I also collected some data for different approaches at
that time. Recently I opened another issue PR102054 which is also related to
that we restrict PRE due to loop-vect.
