%5 = "mhlo.broadcast_in_dim"(%4) {broadcast_dimensions = dense<[2, 3]> : tensor<2xi64>} : (tensor<6x6xf32>) -> tensor<1x3x6x6xf32>
  %6 = mhlo.divide %3, %5 : tensor<1x3x6x6xf32>
  return %6 : tensor<1x3x6x6xf32>
}

// CHECK-LABEL:   func.func @avg_pool2d_2(
// CHECK-SAME:                               %[[VAL_0:.*]]: tensor<1x3x6x6xf32>) -> tensor<*xf32> {
// CHECK:           %[[VAL_1:.*]] = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>

}
// CHECK: @main(%[[ARG:.+]]: tensor<1x8x4x4xf32>) -> tensor<1x8x4x4xf32>

// Contains the `stablehlo.transpose` op of the arg (e.g. [b, f, 0, 1] to
// [b, 0, 1, f]). The weight constant is folded into [0, 1, i, o] format.
// CHECK-DAG: %[[CST:.+]] = stablehlo.constant dense<3.000000e+00> : tensor<3x3x8x8xf32>
// CHECK: %[[TRANSPOSE_1:.+]] = stablehlo.transpose %arg0, dims = [0, 2, 3, 1] : (tensor<1x8x4x4xf32>) -> tensor<1x4x4x8xf32>

  return %2 : tensor<1x8x4x4xf32>
}

// CHECK-DAG: %[[CONST:.+]] = stablehlo.constant {{.*}} : tensor<8x8x3x3xf32>
// CHECK-DAG: %[[TRANSPOSE_0:.+]] = stablehlo.transpose %[[ARG]], dims = [0, 2, 3, 1] : (tensor<1x8x4x4xf32>) -> tensor<1x4x4x8xf32>

  return %2 : tensor<1x8x4x4xf32>
}
// CHECK-DAG: %[[CONST:.+]] = stablehlo.constant {{.*}} : tensor<3x3x8x8xf32>
// CHECK-DAG: %[[TRANSPOSE_0:.+]] = stablehlo.transpose %[[ARG]], dims = [0, 2, 3, 1] : (tensor<1x8x4x4xf32>) -> tensor<1x4x4x8xf32>

// permutated to match the new input shape.
// CHECK: (tensor<1x16x16x4xf32>, tensor<f32>) -> tensor<1x8x8x4xf32>

// Check that a `stablehlo.transpose` is added to the result to match the shape
// of the users.
// CHECK: %[[TRANSPOSE:.+]] = stablehlo.transpose %[[REDUCE_WINDOW]], dims = [0, 3, 1, 2] : (tensor<1x8x8x4xf32>) -> tensor<1x4x8x8xf32>
// CHECK: return %[[TRANSPOSE]]

// -----

// CHECK:           %[[VAL_2:.*]] = "tf.Conv2D"(%[[VAL_0]], %[[VAL_1]]) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true}> : (tensor<1x8x8x207xf32>, tensor<3x3x207x16xf32>) -> tensor<1x8x8x16xf32>
// CHECK:           return %[[VAL_2]] : tensor<1x8x8x16xf32>
// CHECK:         }

  %1 = "tf.Const"() {value = dense<[0, 2, 3, 1]> : tensor<4xi32>} : () -> tensor<4xi32>
  %2 = "tf.Transpose"(%arg0, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>
  %3 = "tf.Transpose"(%2, %1) : (tensor<1x8x4x4xf32>, tensor<4xi32>) -> tensor<1x4x4x8xf32>

  func.return %3 : tensor<1x4x4x8xf32>
  // CHECK: return %arg0
}

// CHECK-LABEL: @nonCancellableTransposeCrossRegion

  %cst = arith.constant dense<2.0> : tensor<3x5xf32>
  %shape = arith.constant dense<[1, 8, 3, 5]> : tensor<4xi32>
  %1 = "tfl.reshape"(%arg0, %shape) : (tensor<8x15xf32>, tensor<4xi32>) -> tensor<1x8x3x5xf32>
  %2 = "tfl.add"(%1, %cst) {fused_activation_function = "NONE"} : (tensor<1x8x3x5xf32>, tensor<3x5xf32>) -> tensor<1x8x3x5xf32>
  func.return %2 : tensor<1x8x3x5xf32>

  %0 = "tf.Conv2D"(%arg0, %arg1) {data_format = "NHWC", dilations = [1, 2, 3, 1], padding = "SAME", strides = [1, 4, 5, 1]} : (tensor<?x32x32x6xf32>, tensor<3x3x3x16xf32>) -> tensor<?x8x7x16xf32>
  func.return %0 : tensor<?x8x7x16xf32>
}

//===----------------------------------------------------------------------===//