[Bug tree-optimization/52091] [4.7 Regression] ICE: verify_gimple failed

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52091

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

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|ASSIGNED                    |NEW
                 CC|                            |irar at gcc dot gnu.org
         AssignedTo|jakub at gcc dot gnu.org    |unassigned at gcc dot
                   |                            |gnu.org

--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-02-02 12:06:27 UTC ---
Executable testcase:

/* PR tree-optimization/52091 */

int b, c, d, f;
unsigned h;
extern void abort (void);

int
main ()
{
  d = -1;
  h = 65;
  asm volatile ("" : : : "memory");
  for (f = 0; f < 4; f++)
    {
      h &= (unsigned short) d;
      for (b = 0; b <= 1; b++)
    {
      c = 0;
      d &= 1;
    }
    }
  asm volatile ("" : : : "memory");
  if (b != 2 || c != 0 || d != 1 || f != 4 || h != 1)
    abort ();
  return 0;
}

Following patch fixes the ICE:
--- gcc/tree-vect-stmts.c.jj    2012-02-01 10:33:58.000000000 +0100
+++ gcc/tree-vect-stmts.c    2012-02-02 11:50:42.623948066 +0100
@@ -1294,7 +1294,11 @@ vect_get_vec_def_for_operand (tree op, g

         /* Get the def before the loop  */
         op = PHI_ARG_DEF_FROM_EDGE (def_stmt, loop_preheader_edge (loop));
-        return get_initial_def_for_reduction (stmt, op, scalar_def);
+        vec_oprnd = get_initial_def_for_reduction (stmt, op, scalar_def);
+        if (!is_gimple_val (vec_oprnd))
+      vec_oprnd = vect_init_vector (stmt, vec_oprnd,
+                    TREE_TYPE (vec_oprnd), NULL);
+    return vec_oprnd;
      }

     /* Case 5: operand is defined by loop-header phi - induction.  */
but just turns ice-on-valid-code into wrong-code, h is 65 when vectorizing.

Ira, could you please look at this?  Thanks.

OT, it is strange that we are creating a reduction for a loop which loops
exactly as many times as there are units in the vector, that doesn't seem to be
profitable.

My other attempt to fix the ICE was:
--- gcc/tree-vect-loop.c.jj    2011-12-16 08:37:45.000000000 +0100
+++ gcc/tree-vect-loop.c    2012-02-02 11:57:25.354607843 +0100
@@ -3370,7 +3370,10 @@ get_initial_def_for_reduction (gimple st
         if (TREE_CONSTANT (init_val))
           init_def = build_vector (vectype, t);
         else
-          init_def = build_constructor_from_list (vectype, t);
+      {
+        init_def = build_constructor_from_list (vectype, t);
+        init_def = vect_init_vector (stmt, init_def, vectype, NULL);
+      }

         break;
but that ICEs on the testcase, because the other caller of
get_initial_def_for_reduction is calling vect_init_vector immediately after it,
with a different stmt than was passed to get_initial_def_for_reduction, and
this resulted in definition not dominating the use.  Wonder if that isn't
another bug.