[Bug tree-optimization/33389] [4.3 Regression] Revision 128239 causes libgomp failure

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

------- Comment #20 from jakub at gcc dot gnu dot org  2007-09-12 18:26 -------
append_vuse has:
static inline void
append_vuse (tree var)
{
  tree sym;

  if (TREE_CODE (var) != SSA_NAME)
    {
      tree mpt;
      var_ann_t ann;

      /* If VAR belongs to a memory partition, use it instead of VAR.  */
      mpt = memory_partition (var);
      if (mpt)
        var = mpt;

      /* Don't allow duplicate entries.  */
      ann = get_var_ann (var);
      if (ann->in_vuse_list || ann->in_vdef_list)
        return;

      ann->in_vuse_list = true;
      sym = var;
    }
  else
    sym = SSA_NAME_VAR (var);

  VEC_safe_push (tree, heap, build_vuses, var);
  bitmap_set_bit (build_loads, DECL_UID (sym));
}

For VAR_DECL with DECL_HARD_REGISTER set, if it is only in the input list of
asm,
it is added as vuse/load, if it is only in output list of am, it is added as
vdef/store.  But the above /* Don't allow duplicate entries.  */ means
that if there is "=r" (x) : "r" (x) with VAR_DECL DECL_HARD_REGISTER x, it is
only marked as store/vdef, not as load/vuse.

No matter whether it is right or wrong in tree-ssa-operands.c, I think
sdse should just avoid removing VAR_DECL DECL_HARD_REGISTER store elimination.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33389