func.func @packNegInputAxis2(%arg0: tensor<1x4xi32>, %arg1: tensor<1x4xi32>) -> tensor<1x2x4xi32> {
  // CHECK: "tfl.pack"(%arg0, %arg1) <{axis = -2 : i32, values_count = 2 : i32}>
  %0 = "tfl.pack"(%arg0, %arg1) {axis = -2 : i32, values_count = 2 : i32} : (tensor<1x4xi32>, tensor<1x4xi32>) -> tensor<1x2x4xi32>
  func.return %0 : tensor<1x2x4xi32>
}

func.func @packNegInputAxis3(%arg0: tensor<1x4xi32>, %arg1: tensor<1x4xi32>) -> tensor<2x1x4xi32> {

    %1 = "tf.Const"() {value = dense<2> : tensor<2x2xi32>} : () -> tensor<2x2xi32>
    %2 = "tf.Identity"(%1) {device = ""} : (tensor<2x2xi32>) -> tensor<2x2xi32>
    // CHECK: SpaceToBatchND{{.*}}-> tensor<4x98x130x128xf32>
    %3 = "tf.SpaceToBatchND"(%arg0, %0, %2) {device = ""} : (tensor<1x192x256x128xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x?x?x128xf32>
    func.return
  }

  %values, %indices = "tf.ApproxTopK"(%arg0) {aggregate_to_topk = true, is_max_k = true, k = 4 : i64, recall_target = 8.500000e-01 : f32, reduction_dimension = 1 : i64, reduction_input_size_override = -1 : i64} : (tensor<1x4xf32>) -> (tensor<1x4xf32>, tensor<1x4xi32>)
  func.return %values, %indices: tensor<1x4xf32>, tensor<1x4xi32>

// -----

// Valid slice operation.
func.func @testSlice(%arg0: tensor<3x4xi32>, %arg1: tensor<2xi64>) -> tensor<1x4xi32> {
  %sizes = "tf.Const"() {value = dense<[1, 4]> : tensor<2xi64>} : () -> (tensor<2xi64>)
  %0 = "tf.Slice"(%arg0, %arg1, %sizes) : (tensor<3x4xi32>, tensor<2xi64>, tensor<2xi64>) -> tensor<1x4xi32>
  func.return %0 : tensor<1x4xi32>
}

// -----

    %0 = "tf.BatchMatMulV2"(%arg3, %arg1) {adj_x = false, adj_y = false} : (tensor<1x3xf32>, tensor<3x4xf32>) -> tensor<1x4xf32>
    %1 = arith.constant dense<[[2.3, 3.4, 4.5, 5.5]]> : tensor<1x4xf32>
    %2 = "tf.Add"(%0, %1) : (tensor<1x4xf32>, tensor<1x4xf32>) -> tensor<1x4xf32>
    %3 = tensor.cast %2 : tensor<1x4xf32> to tensor<1x?xf32>
    func.return %3 : tensor<1x?xf32>
}

  %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>
  %2 = "tf.AddV2"(%0, %arg0) : (tensor<1x2xf32>, tensor<4x1xf32>) -> tensor<4x2xf32>

      tf_device.return %5 : tensor<512xi32>
    }) {device = "/job:localhost/replica:0/task:0/device:CPU:0"} : () -> (tensor<512xi32>, tensor<1024xi32>, tensor<1024xi32>, tensor<1024xf32>)
    // CHECK: %[[COMPILE_OUTPUT:[0-9]*]]:4 = "tf_device.launch"
    // CHECK: "tf._TPUCompileMlir"()
    // CHECK: is_bounded_dynamic_dim: true