func.return %0#0 : tensor<2xi32>

// CHECK-LABEL: unpack3Tensors
// CHECK: "tfl.unpack"(%arg0) <{axis = 1 : i32, num = 3 : i32}> : (tensor<2x3xi32>) -> (tensor<2xi32>, tensor<2xi32>, tensor<2xi32>)
}

func.func @unpackNegAxis(%arg0: tensor<2x3xi32>) -> tensor<2xi32> {
  %0:3 = "tf.Unpack"(%arg0) {axis = -1 : i64} : (tensor<2x3xi32>) -> (tensor<2xi32>, tensor<2xi32>, tensor<2xi32>)

  func.return %0#0 : tensor<2xi32>
}

// -----

func.func @unpack(%arg0: tensor<*xi32>) -> (tensor<*xi32>, tensor<*xi32>) {
  %0:2 = "tfl.unpack"(%arg0) {axis = 1 : i32, num = 2 : i32} : (tensor<*xi32>) -> (tensor<*xi32>, tensor<*xi32>)
  func.return %0#0, %0#1 : tensor<*xi32>, tensor<*xi32>
}

// -----

    ```mlir
      %2 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>
      %3 = "tf.B"(%2) {device = "tpu0"} : (tensor<?xi32>) -> tensor<?xi32>
      %4 = "tf.C"(%2, %3) {device = "tpu0"} : (tensor<?xi32>, tensor<?xi32>) -> tensor<?xi32>
      %5 = "tf.D"(%4) : (tensor<?xi32>) -> tensor<?xi32>
    ```

    After the pass, we will have:

    ```mlir
      %0 = "tf.A"(%arg0) : (tensor<?xi32>) -> tensor<?xi32>

  %14 = "mhlo.broadcast_in_dim"(%1) <{broadcast_dimensions = dense<> : tensor<0xi64>}> : (tensor<i32>) -> tensor<4xi32>
  %15 = mhlo.compare  LT, %13, %14,  SIGNED : (tensor<4xi32>, tensor<4xi32>) -> tensor<4xi1>
  %16 = "mhlo.broadcast_in_dim"(%0) <{broadcast_dimensions = dense<> : tensor<0xi64>}> : (tensor<i32>) -> tensor<4xi32>
  %17 = mhlo.add %13, %16 : tensor<4xi32>

    %cst = "tf.Const"() <{value = dense<[1]> : tensor<1xi32>}> : () -> tensor<1xi32>
    %0 = "tf.VarHandleOp"() <{container = "", shared_name = "Variable"}> : () -> tensor<!tf_type.resource<tensor<2x3xbf16>>>
    %1 = "tf.ResourceGather"(%0, %cst) <{batch_dims = 0 : i64, validate_indices = true}> : (tensor<!tf_type.resource<tensor<2x3xbf16>>>, tensor<1xi32>) -> tensor<1x3xbf16>
    return %1: tensor<1x3xbf16>
  }

    module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
      func.func @main() -> (tensor<2xi32>) {
        %cst = "tf.Const"() {value = dense<[524170, 523952]> : tensor<2xi32>} : () -> tensor<2xi32>
        return %cst : tensor<2xi32>
    }
  })";

  auto compilation_result = CompileMlirModule(
      kHasReturnValuesAndNoMetadataRetvals,

  // The first Operand is assumed to be a TensorType around a variant with a
  // single subtype (e.g. tensor<!tf_type.variant<tensor<2xi32>>>). We will
  // copy this type to the first result, and copy the singular variant subtype
  // to the second result (tensor<2xi32>).
  DCOMMENT_OP(op, "Inferring shape for TensorListPopBackOp.");

  auto src_list_handle_t =

//
// For example, given the following IR:
//
// %split_sizes = "tf.Const"() {value = dense<[1, -1, 3]> : tensor<3xi32>}
// %split_dim = "tf.Const"() {value = dense<1> : tensor<i32>}
// %0:3 = "tf.SplitV"(%input, %split_sizes, %split_dim) :
//                   (tensor<4x6xf32>, tensor<3xi32>, tensor<i32>) ->
//                   (tensor<4x1xf32>, tensor<4x2xf32>, tensor<4x3xf32>)
//