[Bug middle-end/36201] [4.4 Regression] NVR in the front-end causes missed optimization later on (<retval> thought to alias arguments)

[pinskia at gcc dot gnu dot org]

------- Comment #5 from pinskia at gcc dot gnu dot org  2008-12-31 21:11 -------
Actually this testcase is regression because of the removal of SFTs.  in
4.1-4.3, the loop did not exist.
In fact I filed this bug right after SFTs were removed.

The reason why it is considered addressable is because of:
/* Return true if T is a GIMPLE variable whose address is not needed.  */

bool
is_gimple_non_addressable (tree t)
{
  if (TREE_CODE (t) == SSA_NAME)
    t = SSA_NAME_VAR (t);

  return (is_gimple_variable (t) && ! needs_to_live_in_memory (t));
}

needs_to_live_in_memory is true because of

  return (TREE_ADDRESSABLE (t)
          || is_global_var (t)
          || (TREE_CODE (t) == RESULT_DECL
              && aggregate_value_p (t, current_function_decl)));

And aggregate_value_p is true because of:
  if (TREE_CODE (exp) == CALL_EXPR && fndecl && DECL_RESULT (fndecl)
      && DECL_BY_REFERENCE (DECL_RESULT (fndecl)))
    return 1;

DECL_BY_REFERENCE is true because this is a nonPOD.


-- 

pinskia at gcc dot gnu dot org changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
           Severity|enhancement                 |normal
      Known to fail|                            |4.4.0
      Known to work|                            |4.3.0 4.1.2
            Summary|NVR in the front-end causes |[4.4 Regression] NVR in the
                   |missed optimization later on|front-end causes missed
                   |(<retval> thought to alias  |optimization later on
                   |arguments)                  |(<retval> thought to alias
                   |                            |arguments)
   Target Milestone|---                         |4.4.0


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36201