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

["wilson at specifix dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #22 from wilson at specifix dot com  2007-09-12 23:35 -------
Subject: Re:  [4.3 Regression]  Revision 128239
 causes libgomp failure

jakub at gcc dot gnu dot org wrote:
> 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.

This isn't a volatile register, nor a global register variable, so there 
should be no problem performing dead store elimination on it.  I think 
tree-ssa-operands.c needs to be fixed.

I'm not a tree-ssa expert, but the fix needed in tree-ssa-operands.c 
seems simple enough.  The append_vuse function does two things, it keeps 
a list of v_uses, and it keeps a list of loads.  We don't want a v_use 
if we already have a v_def, but we still need to mark it as a load.  We 
can do that with a little rearrangement of the code.  Like this.

This works for the testcase, but has had no other testing as yet.

2007-09-12  James E. Wilson  <wilson@specifix.com>

         * tree-ssa-operands.c (append_vuse): If ann->in_vdef_list true,
         then set build_loads before returning.

Index: tree-ssa-operands.c
===================================================================
--- tree-ssa-operands.c (revision 128394)
+++ tree-ssa-operands.c (working copy)
@@ -1164,8 +1164,15 @@

        /* Don't allow duplicate entries.  */
        ann = get_var_ann (var);
-      if (ann->in_vuse_list || ann->in_vdef_list)
+      if (ann->in_vuse_list)
         return;
+      else if (ann->in_vdef_list)
+       {
+         /* We don't want a vuse if we already have a vdef, but we must
+            still put this in build_loads.  */
+         bitmap_set_bit (build_loads, DECL_UID (var));
+         return;
+       }

        ann->in_vuse_list = true;
        sym = var;


-- 


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