Hi,
this patch address problem I run into with strenghtened cunroll pass.  I made
cunroll to use loop_max_iterations bounds into an account that makes us to
occasionally produce out of bounds loop accesses in loop like:

for (i=0;i<n;i++)
{
  <something>
  if (test)
    break
  a[i]=1;
}
Here the constantly sized array appears in loop with multiple exits and we then
in the last iteration produce out of bound array (since we need to duplicate 
<something> and tree-ssa-vrp warns.

I think this is more generic problem (though appearing rarely): as soon as we
specialize the code, we can no longer warn about out of bounds accesses when we
prove array ref to be constant.  We have no idea if this code path will execute
in practice.

Seems resonable?  I think similar problems can be constructed by inlining, too.
Regtested/bootstrapped x86_64-linux.

Honza

        * tree-cfg.c (gimple_duplicate_bb): Drop NO_WARNING on array refs.
Index: tree-cfg.c
===================================================================
*** tree-cfg.c  (revision 192360)
--- tree-cfg.c  (working copy)
*************** gimple_duplicate_bb (basic_block bb)
*** 5504,5509 ****
--- 5504,5519 ----
        if (lhs && TREE_CODE (lhs) != SSA_NAME)
        {
          tree base = get_base_address (lhs);
+         tree op = lhs;
+
+         /* Silence the array bounds warning.  After specializing the code we
+            no longer know if the particular code path can execute at runtime. 
 */
+         while (handled_component_p (op))
+           {
+             if (TREE_CODE (op) == ARRAY_REF)
+               TREE_NO_WARNING (op) = 1;
+             op =  TREE_OPERAND (op, 0);
+           }
          if (base
              && (TREE_CODE (base) == VAR_DECL
                  || TREE_CODE (base) == RESULT_DECL)
*************** gimple_duplicate_bb (basic_block bb)
*** 5514,5519 ****
--- 5521,5539 ----
                  || !DECL_HAS_VALUE_EXPR_P (base)))
            DECL_NONSHAREABLE (base) = 1;
        }
+
+       /* Silence the array bounds warning.  After specializing the code we
+          no longer know if the particular code path can execute at runtime.  
*/
+       if (gimple_code (stmt) == GIMPLE_ASSIGN)
+       {
+         tree op = gimple_assign_rhs1 (stmt);
+         while (handled_component_p (op))
+           {
+             if (TREE_CODE (op) == ARRAY_REF)
+               TREE_NO_WARNING (op) = 1;
+             op =  TREE_OPERAND (op, 0);
+           }
+       }
  
        /* Create new names for all the definitions created by COPY and
         add replacement mappings for each new name.  */

Reply via email to