[Bug c++/64191] [4.9/5 Regression] -march=native messes up dead code elimination in loop calling dtor

["rguenth at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64191

--- Comment #6 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Richard Biener from comment #3)
> It also seems that DCE cannot remove the empty loop with the clobber, because
> it's marked useful and thus its SSA requirements are marked useful (we
> explicitely exclude the VDEF but _not_ the SSA uses on the LHS of
> *ssa_2 = {} clobbers).  Even if we don't make that SSA use necessary we
> still have made the clobber itself necessary and thus marked control
> dependent
> edges necessary.
> 
> Thus IMHO we shouldn't make clobbers necessary at all but only keep those
> live that we can keep live (all uses still live).  That makes cddce1
> kill bar_dtor_loop:
> 
> void bar_dtor_loop(Bar*, unsigned int) (struct Bar * p, unsigned int n)
> {
>   struct Bar * e;
> 
>   <bb 2>:
>   return;
> 
> }
> 
> but of course it probably will DCE all indirect clobbers on an address
> that is otherwise unused.  Jakub - do you think this will be a problem?
> 
> Index: gcc/tree-ssa-dce.c
> ===================================================================
> --- gcc/tree-ssa-dce.c  (revision 218479)
> +++ gcc/tree-ssa-dce.c  (working copy)
> @@ -292,8 +292,7 @@ mark_stmt_if_obviously_necessary (gimple
>        break;
>  
>      case GIMPLE_ASSIGN:
> -      if (TREE_CODE (gimple_assign_lhs (stmt)) == SSA_NAME
> -         && TREE_CLOBBER_P (gimple_assign_rhs1 (stmt)))
> +      if (gimple_clobber_p (stmt))
>         return;
>        break;
>  
> @@ -813,8 +812,9 @@ propagate_necessity (bool aggressive)
>                 }
>             }
>  
> -         FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
> -           mark_operand_necessary (use);
> +         if (!gimple_clobber_p (stmt))
> +           FOR_EACH_SSA_TREE_OPERAND (use, stmt, iter, SSA_OP_USE)
> +             mark_operand_necessary (use);
>  
>           use = gimple_vuse (stmt);
>           if (!use)
> @@ -1362,6 +1362,24 @@ eliminate_unnecessary_stmts (void)
>           /* If GSI is not necessary then remove it.  */
>           if (!gimple_plf (stmt, STMT_NECESSARY))
>             {
> +             /* Keep clobbers that we can keep live live.  */
> +             if (gimple_clobber_p (stmt))
> +               {
> +                 ssa_op_iter iter;
> +                 use_operand_p use_p;
> +                 bool dead = false;
> +                 FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
> +                   {
> +                     gimple def = SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p));
> +                     if (!gimple_plf (def, STMT_NECESSARY))
> +                       {
> +                         dead = true;
> +                         break;
> +                       }
> +                   }
> +                 if (!dead)
> +                   continue;
> +               }
>               if (!is_gimple_debug (stmt))
>                 something_changed = true;
>               remove_dead_stmt (&gsi, bb);

Or rather

              /* Keep clobbers that we can keep live live.  */
              if (gimple_clobber_p (stmt))
                {
                  ssa_op_iter iter;
                  use_operand_p use_p;
                  bool dead = true;
                  FOR_EACH_SSA_USE_OPERAND (use_p, stmt, iter, SSA_OP_USE)
                    {
                      gimple def = SSA_NAME_DEF_STMT (USE_FROM_PTR (use_p));
                      if (gimple_nop_p (def)
                          || gimple_plf (def, STMT_NECESSARY))
                        {
                          dead = false;
                          break;
                        }
                    }

so we keep the indirect clobbers in destructors (with use of parameter
SSA default defs).