[Bug tree-optimization/57517] [4.7/4.8/4.9 Regression] internal compiler error: in eliminate_temp_copies, at tree-predcom.c:1913

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=57517

Richard Biener <rguenth at gcc dot gnu.org> changed:

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

--- Comment #3 from Richard Biener <rguenth at gcc dot gnu.org> ---
We also fail to sink the loads (because appearantly we don't want to do that,
eh),
which in the end causes this havoc (not-always-executed combination).

C testcase:

int x[1024], y[1024], z[1024], w[1024];
void foo (void)
{
  int i;
  for (i = 1; i < 1024; ++i)
    {
      int a = x[i];
      int b = y[i];
      int c = x[i-1];
      int d = y[i-1];
      if (w[i])
        z[i] = (a + b) + (c + d);
    }
}

Fixed by for example the simple

Index: gcc/tree-predcom.c
===================================================================
--- gcc/tree-predcom.c  (revision 204948)
+++ gcc/tree-predcom.c  (working copy)
@@ -2068,7 +2068,11 @@ combinable_refs_p (dref r1, dref r2,

   stmt = find_common_use_stmt (&name1, &name2);

-  if (!stmt)
+  if (!stmt
+      /* A simple post-dominance check - make sure the combination
+         is executed under the same condition as the references.  */
+      || (gimple_bb (stmt) != gimple_bb (r1->stmt)
+         && gimple_bb (stmt) != gimple_bb (r2->stmt)))
     return false;

   acode = gimple_assign_rhs_code (stmt);


Zdenek?