// CHECK-LABEL: reduce_window_max_activation_transpose_with_padding
func.func @reduce_window_max_activation_transpose_with_padding(%arg0: tensor<16x16x4x1xf32>) -> tensor<1x4x8x8xf32> {
  %0 = stablehlo.constant dense<0xFF800000> : tensor<f32>  // -inf
  %1 = stablehlo.transpose %arg0, dims = [3, 2, 1, 0] : (tensor<16x16x4x1xf32>) -> tensor<1x4x16x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({
  ^bb0(%arg1: tensor<f32>, %arg2: tensor<f32>):

// CHECK: %cst_1 = arith.constant dense<[0, 3, 1, 2]> : tensor<4xi32>
// CHECK: %3 = "tfl.transpose"(%2, %cst_1) : (tensor<1x1x4x1xf32>, tensor<4xi32>) -> tensor<1x1x1x4xf32>
// CHECK: return %3 : tensor<1x1x1x4xf32>

  // Unsupported strides
  %2 = "tf.AvgPool"(%arg0) {T = "tfdtype$DT_FLOAT", data_format = "NCHW", ksize = [1, 1, 3, 6], padding = "VALID", strides = [1, 3, 3, 1]} : (tensor<1x16x6x6xf32>) -> tensor<1x16x1x1xf32>

  %5 = arith.addf %0, %1 : tensor<1x16x1x1xf32>
  %6 = arith.addf %2, %5 : tensor<1x16x1x1xf32>
  func.return %6 : tensor<1x16x1x1xf32>

//   %cst = arith.constant dense<1.0> : tensor<16x16x4xf32>
//   %0 = "tfl.conv_2d"...
//   %1 = "tfl.add"(%0, %cst) : (tensor<16x16x4xf32>, tensor<16x16x4xf32>)
// After this optimization:
//   %cst = arith.constant dense<1.0> : tensor<f32>
//   %0 = "tfl.conv_2d"...
//   %1 = "tfl.add"(%0, %cst) : (tensor<16x16x4xf32>, tensor<f32>)