[Bug tree-optimization/63666] [5.0 regression] FAIL: gcc.dg/vect/pr45752.c (internal compiler error)

[glisse at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63666

--- Comment #12 from Marc Glisse <glisse at gcc dot gnu.org> ---
(I am traveling this week, so I can only post a short comment)

(In reply to Richard Biener from comment #11)
> -         if (need_mask_canon && arg2 == op2)
> +         if (need_mask_canon && arg2 == op2
> +             && can_vec_perm_p (TYPE_MODE (type), false, sel))

seems reasonable, though if we didn't have can_vec_perm_p before the
transformation (which implies we are before vector lowering) we should probably
still transform. Also, I believe that targets should try to support single
argument permutations when they support a 2 argument version that degenerates
to it when both arguments are the same.

By the way, from a quick look, I don't see where we protect against
TREE_SIDE_EFFECTS(op0) when doing op0=op1 or op1=op0 in this code, I don't
remember if there is a reason why it isn't needed.