[Bug tree-optimization/81365] [7/8 Regression] GCC miscompiles swap

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81365

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The optimization and thus the bug too has been introduced in r235208, i.e.
PR70171.
The check Andrew mentions has been added in r236079, i.e. PR71039.  But it
actually does something completely different than the check we need here, it
makes sure that any SSA_NAMEs used in the lhs that we move earlier is available
there already.

What we need here is something different.
We already have:
      /* Check if we can move the loads.  The def stmt of the virtual use
         needs to be in a different basic block dominating bb.  When the
         def is an edge-inserted one we know it dominates us.  */
      vuse = gimple_vuse (use_stmt);
      def_stmt = SSA_NAME_DEF_STMT (vuse);
      if (!SSA_NAME_IS_DEFAULT_DEF (vuse)
          && (gimple_bb (def_stmt) == bb
              || (gimple_bb (def_stmt)
                  && !dominated_by_p (CDI_DOMINATORS,
                                      bb, gimple_bb (def_stmt)))))
        goto next;
which makes sure we don't miscompile
int h(int k, int i1, int j1)
{
  int *f1;
  if(k)
   f1 = &i1;
  else
   f1 = &j1;
  i1 = 7;
  return *f1;
}
where the use_stmt has an earlier store that might modify it.  That is
sufficient for the reg type case, but for the aggregate, where we actually move
the aggregate store earlier - onto the edges - we also have to ensure that
either there no memory loads between the PHI and use_stmt, or that the can't
alias the lhs.