}

func.func @mean(%arg0: tensor<2x2xf32>, %arg1: tensor<1xi32>) -> tensor<1x2xf32> {
  %0 = "tf.Mean"(%arg0, %arg1) : (tensor<2x2xf32>, tensor<1xi32>) -> tensor<1x2xf32>
  func.return %0 : tensor<1x2xf32>

// CHECK-LABEL: mean
// CHECK:  "tfl.mean"(%arg0, %arg1) <{keep_dims = false}> : (tensor<2x2xf32>, tensor<1xi32>) -> tensor<1x2xf32>
}

func.func @mean_true(%arg0: tensor<2x2xf32>, %arg1: tensor<1xi32>) -> tensor<1x2xf32> {

    %unshaped = "tf.Cast"(%arg1) : (tensor<1x2x3xf32>) -> tensor<*xf32>
    // CHECK: <{is_stateless = true}>
    %0 = "tf.IfRegion"(%arg0) <{is_stateless = true}> ({
      // CHECK: "tf.Add"{{.+}}(tensor<1x2x3xf32>, tensor<1x2x3xf32>) -> tensor<1x2x3xf32>
      // CHECK: "tf.Yield"{{.+}}(tensor<1x2x3xf32>) -> ()

  %0 = "tf.Const"() {value = dense<0.0> : tensor<1x2xf32>} : () -> tensor<1x2xf32>

  // Operand and identity shapes are broadcastable. However, we cannot fold
  // because the operand does not match the result shape.
  %1 = "tf.AddV2"(%arg0, %0) : (tensor<4x1xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>

// CHECK:           }) : (tensor<1x?xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, none, none, none, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<3x1xf32>, tensor<3xf32>, tensor<1x3xf32>, tensor<1x1xf32>, none, none, none, none) -> tensor<1x3xf32>
// CHECK:           [[VAL_50:%.*]] = tensor.cast [[VAL_51:%.*]] : tensor<1x3xf32> to tensor<1x?xf32>

  func.return %24 : tensor<1x4xf32>
}

    func.return %0 : tensor<128xf32>
  }
  func.func @_func() -> tensor<128xf32> {
    %0 = "tf.Const"() {value = dense<0.0> : tensor<128xf32>} : () -> tensor<128xf32>
    func.return %0 : tensor<128xf32>
  }
}

// -----
// CHECK-LABEL: func @return_from_host_and_tpu

// `tfl.concatenation`.

func.func @concatenate_float(%arg0: tensor<3x2xf32>, %arg1: tensor<1x2xf32>) -> tensor<4x2xf32> {
  %0 = "stablehlo.concatenate"(%arg0, %arg1) {dimension = 0 : i64} : (tensor<3x2xf32>, tensor<1x2xf32>) -> tensor<4x2xf32>
  return %0 : tensor<4x2xf32>
}
// CHECK-LABEL: concatenate_float
// CHECK-NOT: tfl.concatenation
// CHECK: stablehlo.concatenate

// -----

func.func @testPack(%arg0: tensor<4x8xf32>, %arg1: tensor<4x8xf32>) -> tensor<*xf32> {
  %0 = "tf.Pack"(%arg0, %arg1) {axis = 1 : i64} : (tensor<4x8xf32>, tensor<4x8xf32>) -> tensor<*xf32>
  func.return %0 : tensor<*xf32>
}


// -----

func.func @testPack(%arg0: tensor<4x8xf32>, %arg1: tensor<4x2xf32>) -> tensor<*xf32> {

    } else {
      // Recurse on the subtypes in the variant/resource. Basically if the input
      // were:
      //   tensor<!tf_type.variant<tensor<?x8xf32>>>
      // and:
      //   tensor<!tf_type.variant<tensor<10x8xf32>>>
      // we'll try here to refine tensor<?x8xf32> with tensor<10x8xf32>.
      auto refined_subtype = mlir::cast<TensorType>(
          TypeMeet(lhs_element_type_with_subtype.GetSubtypes().front(),