Operation *op = *latticeElement->getValue().ops.begin();
      GlobalTensorOp globalTensor = llvm::dyn_cast<GlobalTensorOp>(op);

      if (!globalTensor)
        continue;  // happens if the name is e.g. in a VarHandleOp.

      if (globalTensor.getIsMutable()) {
        freezeable[val] = false;
        continue;
      }

      freezeable[val] = true;

      // Verify users are supported kind.

    type = op->getResult(0).getType().cast<ShapedType>();
  } else if (llvm::isa<TF::AssignVariableOp>(op)) {
    type = op->getOperand(1).getType().cast<ShapedType>();
  } else if (llvm::isa<TF::VarHandleOp>(op)) {
    type = op->getResult(0)
               .getType()
               .cast<ShapedType>()
               .getElementType()
               .cast<TF::TensorFlowTypeWithSubtype>()
               .GetSubtypes()

//

// Creates the per-device variables that represent the formatting state of each
// device.
llvm::SmallVector<TF::VarHandleOp, 4> CreateStateVars(
    const llvm::SmallDenseMap<llvm::StringRef, llvm::SmallVector<StringRef, 4>>&
        devices,
    Location loc, RankedTensorType key_type, OpBuilder* builder) {
  llvm::SmallVector<TF::VarHandleOp, 4> state_vars;

  // TODO(b/148913020): Remove this constraint once model parallelism is
  // supported.

  Output var_handle =
      ops::VarHandleOp(scope.WithOpName("Var" + id), DT_FLOAT, TensorShape({}));
  Output read =
      ops::ReadVariableOp(scope.WithOpName("Read" + id), var_handle, DT_FLOAT);
  return read.node();
}

Node* MakeWrite(const Scope& scope, const string& id) {
  Output var_handle =
      ops::VarHandleOp(scope.WithOpName("Var" + id), DT_FLOAT, TensorShape({}));
  Output value_to_write =

  static const char* kOpsToSkip[] = {
      "IdentityOp",
      "NoOp",              // b/174596063
      "OptionalHasValue",  // b/173136483
      "OptionalGetValue",  // b/173136483
      "VarHandleOp",       // b/176819198
  };
  for (const char* skip : kOpsToSkip) {
    if (absl::StartsWith(orig_op->op_name(), skip)) return absl::OkStatus();
  }

  tf_core_op_expansion_node_counter->GetCell()->IncrementBy(1);

                               result);
}

Node* MakeWrite(const Scope& scope, Output value_to_write, const string& id) {
  Output var_handle = ops::VarHandleOp(scope.WithOpName("Var_" + id), DT_FLOAT,
                                       TensorShape({}));
  ops::AssignVariableOp assign_op(scope.WithOpName("Assignee_" + id),

}

// CHECK-LABEL: func @no_set_stateless
func.func @no_set_stateless(%arg: tensor<i32>, %cond: tensor<i1>) -> (tensor<i32>, tensor<i32>) {
  %handle = "tf.VarHandleOp"() { container = "", shared_name = "var" } : () -> tensor<!tf_type.resource<tensor<i32>>>
  "tf.AssignVariableOp"(%handle, %arg) : (tensor<!tf_type.resource<tensor<i32>>>, tensor<i32>) -> ()
  // CHECK: tf.While

  auto& island_block = island_op.getBody().front();
  if (++island_block.begin() != --island_block.end()) return nullptr;

  Operation* resource_op = &island_block.front();
  if (llvm::isa<TF::VarHandleOp, TF::HashTableOp, TF::HashTableV2Op,
                TF::MutableHashTableV2Op>(resource_op)) {
    return resource_op;
  }
  return nullptr;
}

  {
    AbstractTensorHandle* var_ptr = nullptr;
    PartialTensorShape scalar_shape;
    TF_EXPECT_OK(
        PartialTensorShape::MakePartialShape<int32_t>({}, 0, &scalar_shape));
    TF_EXPECT_OK(tensorflow::ops::VarHandleOp(ctx_.get(), &var_ptr, DT_FLOAT,
                                              scalar_shape));
    var.reset(var_ptr);
  }
  // Assign a value.
  auto x = CreateScalarTensor<float, TF_FLOAT>(2.0f);
  TF_EXPECT_OK(