For the testcase in PR108782 we're rejecting to vectorize the nested cycle
because

      if (!flow_bb_inside_loop_p (loop, gimple_bb (use_stmt)))
        {
          if (dump_enabled_p ())
            dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location,
                             "intermediate value used outside loop.\n");

          return NULL;
        }

but really all inner PHIs should be trivially vectorizable without
further restrictions.  The only thing we need to do is detect double
reductions, that is, when there's an outer reduction feeding the inner
loop PHI.