I noticed this:

  if (p0 && p1) {
    Node* in1 = in(1)->uncast();
    Node* in2 = in(2)->uncast();
    AllocateNode* alloc1 = AllocateNode::Ideal_allocation(in1, NULL);
    AllocateNode* alloc2 = AllocateNode::Ideal_allocation(in2, NULL);
    if (MemNode::detect_ptr_independence(in1, alloc1, in2, alloc2, NULL)) {
      return TypeInt::CC_GT; // different pointers
    }
    ciKlass* klass0 = p0->klass();
    bool xklass0 = p0->klass_is_exact();
    ciKlass* klass1 = p1->klass();
    bool xklass1 = p1->klass_is_exact();
    int kps = (p0->isa_klassptr()?1:0) + (p1->isa_klassptr()?1:0);
    if (klass0 && klass1 &&
        kps != 1 && // both or neither are klass pointers
        klass0->is_loaded() && !klass0->is_interface() && // do not trust interfaces
        klass1->is_loaded() && !klass1->is_interface() &&
        (!klass0->is_obj_array_klass() ||
         !klass0->as_obj_array_klass()->base_element_klass()->is_interface()) &&
        (!klass1->is_obj_array_klass() ||
         !klass1->as_obj_array_klass()->base_element_klass()->is_interface())) {

in CmpPNode::sub() doesn't trigger for klass pointers comparisons since JDK-6964458 because TypeKlassPtr doesn't subclass TypeOopPtr anymore.