// CHECK-NOT: "tf.VarHandleOp"
  // CHECK-NOT: "tf.ReadVariableOp"
  // CHECK: %[[CONST:.*]] = "tf.Const"()
  // CHECK: %[[ADD:[0-9]*]] = "tf.AddV2"(%arg1, %[[CONST]])
  // CHECK: %[[PACK:[0-9]*]] = "tf.Pack"(%[[CONST]], %[[ADD]])
  // CHECK: return %[[PACK]]
  %0 = "tf.Const"() {value = dense<4.200000e+01> : tensor<f32>} : () -> tensor<f32>

    // CHECK: tf_executor.island
    %island = tf_executor.island {
      // expected-remark@below {{Result #0, ID 0 : 0, 1, 2}}
      %vh0 = "tf.VarHandleOp"() {container = "c", shared_name = "v0"} : () -> !tf_res
      // expected-remark@below {{Result #0, ID 1 : 0, 1, 2}}
      %vh1 = "tf.VarHandleOp"() {container = "c", shared_name = "v0"} : () -> !tf_res
      // expected-remark@below {{Result #0, ID 2 : 0, 1, 2}}

  // CHECK-NOT: tf.VarHandleOp
  // CHECK-NOT: tf.ReadVariableOp
  // CHECK: [[v:%.*]] = "tf._TfrtGetResource"() <{container = [""], indices = [0], shared_name = [""]}> {device = "/CPU:0"} : () -> tensor<i32>
  // CHECK: [[r:%.*]] = "tf.AddV2"({{.*}}, [[v]]) {device = "/CPU:0"} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK: return [[r]]

    %outputs_7, %control_8 = tf_executor.island wraps "tf.Const"() {value = dense<"MatMul/b_0"> : tensor<1x!tf_type.string>} : () -> tensor<1x!tf_type.string>
    %outputs_9, %control_10 = tf_executor.island wraps "tf.VarHandleOp"() {container = "", shared_name = "MatMul/b_0"} : () -> tensor<!tf_type.resource<tensor<20x4096xf32>>>

  %arg0: !tf_res {tf.device = "/TPU:0"},
  %arg1: !tf_res {tf.device = "/TPU:1"}) {
  tf_executor.graph {
    // CHECK: tf_executor.island
    %island = tf_executor.island {
      // CHECK-NEXT: "tf.VarHandleOp"
      %var_handle = "tf.VarHandleOp"() {container = "c", shared_name = "v0", device = "/CPU:0"}
        : () -> !tf_res
      // CHECK-NEXT: "tf.Identity"
      // CHECK-SAME: {device = "/TPU:0"}
      %id0 = "tf.Identity"(%arg0) : (!tf_res)

// readonly.

// CHECK-LABEL: readvariableop_is_hoisted_if_readonly2
// CHECK:       [[CST_0:%.*]] = "tf.Const"
// CHECK:       [[VAR_1:%.*]] = "tf.VarHandleOp"
// CHECK-SAME:  "shared_name_var1"
// CHECK:       [[VAR_2:%.*]] = "tf.VarHandleOp"
// CHECK-SAME:  "shared_name_var2"
// CHECK:       [[CST_1:%.*]] = "tf.Const"
// CHECK:       [[VAR_VAL:%.*]] = "tf.ReadVariableOp"([[VAR_1]])

}

// Returns the name of the VarHandleOp that provides input (possibly indirectly)
// to node with node_name. A typical indirect chain of nodes (that can occur due
// to graph inlining) is the following: VarHandleOp -> Identity -> Identity ->
// ReadVariableOp. Calling the function on any of these nodes would return the
// name of the VarHandleOp.
StatusOr<string> GetVarHandleName(

  func.return %3 : tensor<*xi32>
}

}

// -----

module attributes {tf_saved_model.semantics} {

// CHECK-LABEL: func private @reorder_execute_arg_defining_ops
// CHECK: tf.VarHandleOp
// CHECK-NEXT: tf.ReadVariableOp
// CHECK-NEXT: tf.TPUCompileMlirAndExecute
func.func private @reorder_execute_arg_defining_ops(%arg0: tensor<1x3xf32> {tf.device = "/CPU:0"}) -> (tensor<1x1xf32> {tf.device = "/TPU:0"}) {

        %0 = "tf.VarHandleOp"() {container = "", device = "/device:CPU:0", shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
        "tf.AssignVariableOp"(%0, %cst) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()

        %cst_0 = "tf.Const"() {device = "", value = dense<[3, 4, 5, 6]> : tensor<4xi32>} : () -> tensor<4xi32>

      "/job:localhost/replica:0/task:0/device:CPU:1"};
  ParallelDevice parallel_device(std::move(devices));
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> handle_op(
      TFE_NewOp(context.get(), "VarHandleOp", status.get()), TFE_DeleteOp);
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  TFE_OpSetAttrType(handle_op.get(), "dtype", TF_FLOAT);