[Bug tree-optimization/33870] [4.3 Regression] miscompiles sqlite

[rguenth at gcc dot gnu dot org]

------- Comment #11 from rguenth at gcc dot gnu dot org  2007-10-23 21:52 
-------
Now it is memory paritioning again ;)

static void
rewrite_alias_set_for (tree tag, bitmap new_aliases)
{
  bitmap_iterator bi;
  unsigned i;
  tree mpt, sym;

  EXECUTE_IF_SET_IN_BITMAP (MTAG_ALIASES (tag), 0, i, bi)
    {
      sym = referenced_var (i);
      mpt = memory_partition (sym);
      if (mpt)
        bitmap_set_bit (new_aliases, DECL_UID (mpt));
      else
        bitmap_set_bit (new_aliases, DECL_UID (sym));
    }

  /* Rebuild the may-alias array for TAG.  */
  bitmap_copy (MTAG_ALIASES (tag), new_aliases);
}

This happily separates the zero-offset SFT from the other SFTs, keeping
some in the NMT and puts some other in the MPT.

So, what's the fix?  Never divorce SFTs of a single variable to different
(or no) partitions?  Duplicate the zero-offset SFT in both the MPT and
the NMT?  Install the fix to not assume the zero-offset SFT is in the
set of aliases?


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33870