%size = "tf.Const"() {value = dense<10> : tensor<i32>} : () -> tensor<i32>
  // CHECK-NOT: tf.TensorArrayV3
  // CHECK: %[[TA_BUFFER:.*]] = "tf.MlirLocalVarOp"() : () -> tensor<!tf_type.resource<tensor<10x3xf32>>>
  // CHECK: "tf.AssignVariableOp"(%[[TA_BUFFER]]
  // CHECK-NOT: tf.TensorArrayV3

}

// CHECK-LABEL: func @shape
func.func @shape() {
  %size = "tf.Const"() {value = dense<5> : tensor<i32>} : () -> tensor<i32>
  // CHECK: tf.TensorArrayV3
  // CHECK-SAME: element_shape = #corert.shape<*>

  } : (tensor<i32>) -> (tensor<2xi32>, tensor<*xf32>)

// CHECK: "tf.TensorArrayV3"
// CHECK-SAME: : (tensor<i32>) -> (tensor<2x!tf_type.resource>, tensor<f32>)

  %1 = "tfl.cast"(%0#1) : (tensor<*xf32>) -> tensor<f32>
  func.return %1 : tensor<f32>
}

// CHECK-LABEL: TfParseExample

  %size = "tf.Const"() {device = "/device:CPU:0", value = dense<9> : tensor<i32>} : () -> (tensor<i32>)

  // CHECK: "tf.TensorArrayV3"
  %handle_12, %flow_13 = "tf.TensorArrayV3"(%max_iter) {device = "/job:localhost/replica:0/task:0/device:CPU:0", dtype = !tf_type.variant, dynamic_size = false, element_shape = #tf_type.shape<*>, identical_element_shapes = true, tensor_array_name = ""} : (tensor<i32>) -> (tensor<2x!tf_type.resource<tensor<*x!tf_type.variant>>>, tensor<f32>)
  // CHECK: "tf.TensorArrayV3"

  // CHECK: tfrt_fallback_async.executeop key({{.*}}) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.TensorArrayWriteV3"
  // CHECK: tfrt_fallback_async.executeop key({{.*}}) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.TensorArrayGatherV3"

 StackPushV2 314
 StackV2 314
 StridedSlice 12
 Sub 5
 Sum 2
 Switch 247
 TensorArrayGatherV3 8
 TensorArrayGradV3 17
 TensorArrayReadV3 9
 TensorArrayScatterV3 8
 TensorArraySizeV3 4
 TensorArrayV3 8
 TensorArrayWriteV3 9
 Unique 2
 VariableV2 164
 _Retval 5
cluster 0 size 440
 Abs 40
 AddN 1
 Any 41
 Cast 40
 ConcatV2 2
 Const 95
 ExpandDims 2
 IsInf 1
 IsNan 40
 L2Loss 40

    // unranked tensor types. Here we set them to scalar tensor types when it is
    // safe.
    if (auto tensor_array_v3_op = dyn_cast<TF::TensorArrayV3Op>(tf_op)) {
      // The "flow" in TensorArrayV3 is always a scalar float tensor.
      // https://www.tensorflow.org/api_docs/python/tf/raw_ops/TensorArrayWriteV3
      Value flow = tensor_array_v3_op.getFlow();
      Type scalar_f32_tensor_type =

// Converts tensor array ops into operations on local variables, which can later
// be removed by resource lifting. Requires known sizes and known element shapes
// (either defined in TensorArrayV3 or implied in the first write).
std::unique_ptr<OperationPass<ModuleOp>>
CreateTensorArrayOpsDecompositionPass();

// Create a pass that legalize TFG to TF dialect.
std::unique_ptr<Pass> CreateLegalizeTFGToTFEPass();

  //     of type DT_RESOURCE that are not resource variables.
  //
  // Together these imply we cannot let resource variable safety analysis
  // constrain e.g. a TensorArrayV3->TensorArrayAssignV3 edge to be in different
  // clusters: both of them will have to be clustered because of (1) and we
  // won't be able to keep the edge between the two as neither the input to the