[Bug tree-optimization/60382] [4.8/4.9 Regression] ICE on valid code at -O3 on x86_64-linux-gnu (in vect_create_epilog_for_reduction, at tree-vect-loop.c:4352)

[rguenth at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60382

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue is that in vectorizable_reduction we compute double_reduc in a wrong
way:

  if (nested_cycle)
    {
      def_bb = gimple_bb (reduc_def_stmt);
      def_stmt_loop = def_bb->loop_father;
      def_arg = PHI_ARG_DEF_FROM_EDGE (reduc_def_stmt,
                                       loop_preheader_edge (def_stmt_loop));
      if (TREE_CODE (def_arg) == SSA_NAME
          && (def_arg_stmt = SSA_NAME_DEF_STMT (def_arg))
          && gimple_code (def_arg_stmt) == GIMPLE_PHI
          && flow_bb_inside_loop_p (outer_loop, gimple_bb (def_arg_stmt))
          && vinfo_for_stmt (def_arg_stmt)
          && STMT_VINFO_DEF_TYPE (vinfo_for_stmt (def_arg_stmt))
              == vect_double_reduction_def)
        double_reduc = true;
    }

that is because def_arg is a constant (and the outer loop reduction PHI
is dead).  Easiest is to not detect this as a valid reduction with
sth like

Index: gcc/tree-vect-loop.c
===================================================================
--- gcc/tree-vect-loop.c        (revision 208269)
+++ gcc/tree-vect-loop.c        (working copy)
@@ -2193,6 +2193,12 @@ vect_is_simple_reduction_1 (loop_vec_inf
               || (!check_reduction && flow_loop_nested_p (vect_loop, loop)));

   name = PHI_RESULT (phi);
+  /* ???  If there are no uses of the PHI result the inner loop reduction
+     won't be detected as possibly double-reduction by vectorizable_reduction
+     because that tries to walk the PHI arg from the preheader edge which
+     can be constant.  See PR60382.  */
+  if (has_zero_uses (name))
+    return NULL;
   nloop_uses = 0;
   FOR_EACH_IMM_USE_FAST (use_p, imm_iter, name)
     {