// CHECK-LABEL: reduce_window_max_activation_transpose
// CHECK-SAME: (%[[ARG:.+]]: tensor<16x8xf32>) -> tensor<4x8xf32>
func.func @reduce_window_max_activation_transpose_rank2(%arg0: tensor<16x8xf32>) -> tensor<4x8xf32> {
  %0 = stablehlo.constant dense<0xFF800000> : tensor<f32>  // -inf
  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<16x8xf32>) -> tensor<8x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({

  // CHECK: return %[[RESULT]]
  func.return %0: tensor<10x2xf32>
}

// -----

func.func @select_first(%arg0: tensor<10x2xf32>, %arg1: tensor<10x2xf32>) -> tensor<10x2xf32> {
  func.return %arg0: tensor<10x2xf32>
}

// CHECK-LABEL: testMultiInputLegacyCallOp
func.func @testMultiInputLegacyCallOp(%arg0: tensor<10x2xf32>, %arg1: tensor<10x2xf32>) -> tensor<10x2xf32> {

  %8 = "tfl.concatenation"(%2, %0) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x2xf32>
  %9 = "quantfork.stats"(%8) {layerStats = dense<[-0.488159984, 0.189515018]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  %10 = "tfl.concatenation"(%9, %7) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x4xf32>

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

    ```mlir
      %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
    ```

    then running this pass with 'default-device=foobar', we get:

    ```mlir

func.func @select_float(%arg0: tensor<1x3xi1>, %arg1: tensor<1x3xf32>, %arg2: tensor<1x3xf32>) -> tensor<1x3xf32> {
  %0 = "stablehlo.select"(%arg0, %arg1, %arg2) : (tensor<1x3xi1>, tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  return %0 : tensor<1x3xf32>
}
// CHECK-LABEL: select_float
// CHECK-NOT: tfl.select_v2
// CHECK: stablehlo.select

// -----

  // CHECK: return %arg0
  func.return %0 : tensor<2x4xf32>
}

// CHECK-LABEL: func @testBroadcastToNoOp
func.func @testBroadcastToNoOp(%arg0: tensor<2x4xf32>, %arg1: tensor<2xi32>) -> tensor<2x4xf32> {
  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<2x4xf32>, tensor<2xi32>) -> tensor<2x4xf32>

  // CHECK: return %arg0
  func.return %0 : tensor<2x4xf32>
}

      // 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(),

    // CHECK-NOT: tf.TPUPartitionedOutputV2
    %partitioned_output:2 = "tf.TPUPartitionedOutputV2"(%computation) {_XlaSharding = "\08\03\1A\02\01\02\22\02\00\01", partition_dims = [1, 2]} : (tensor<3x4xf32>) -> (tensor<3x2xf32>, tensor<3x2xf32>)
    // CHECK: "tf.AssignVariableOp"(%arg0, %[[PARALLEL_EXECUTE_OUTPUT]]#0)