%1 = "tf.Transpose"(%arg0, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>
  %2 = "tf.Transpose"(%1, %0) : (tensor<1x8x4x4xf32>, tensor<4xi32>) -> tensor<1x4x8x4xf32>
  %3 = "tf.Transpose"(%arg1, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>
  %4 = "tf.Transpose"(%3, %0) : (tensor<1x8x4x4xf32>, tensor<4xi32>) -> tensor<1x4x8x4xf32>
  %5 = "tf.AddV2"(%2, %4) : (tensor<1x4x8x4xf32>, tensor<1x4x8x4xf32>) -> tensor<1x4x8x4xf32>

// CHECK: %[[TRANSPOSE_2:.+]] = stablehlo.transpose %[[CUSTOM_AGGREGATOR_1]], dims = [0, 3, 1, 2] : (tensor<1x4x4x8xf32>) -> tensor<1x8x4x4xf32>

  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 %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

  func.return %dequant : tensor<1x3x4x2xf32>
}

// -----

// CHECK-LABEL: @dont_merge_quantization_followed_by_quantization
// CHECK-SAME: %[[ARG_0:.*]]: tensor<1x3x4x3xf32>
func.func @dont_merge_quantization_followed_by_quantization(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32> {
  // CHECK: %[[QUANT_ARG_0:.*]] = stablehlo.uniform_quantize %[[ARG_0]]

    %1 = "tf.Conv2D"(%0, %cst) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>

      %1 = "tf.XlaCallModule"(%0, %cst_0, %cst_1) <{Sout = [#tf_type.shape<1x4x4x3>], module = "", version = 9 : i64}> {_entry_function = @composite_fn_2, _original_entry_function = "composite_fn_2", _tfl_quant_trait = "fully_quantizable"} : (tensor<1x4x4x3xf32>, tensor<3x1x1x3xf32>, tensor<3xf32>) -> tensor<1x4x4x3xf32>
      return %1 : tensor<1x4x4x3xf32>

  %1 = "tf.AddV2"(%0, %0) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  %2 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %3 = "tf.Transpose"(%1, %2) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>

  func.return %3 : tensor<1x8x4x4xf32>
}

// CHECK-LABEL: move_transpose_handle_broadcast

func.func @out_of_bounds_check(%arg0: tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32> {
  %0 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %1 = "tf.Const"() {value = dense<[0, 0x4141, 3, 1]> : tensor<4xi32>} : () -> tensor<4xi32>
  %2 = "tf.Transpose"(%arg0, %0) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>

// CHECK:  [[VAL_1:%.*]] = "tfl.reshape"(%2, %[[CST]]) {tac.device = "GPU",  tac.inference_type = "FLOAT"} : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:  [[VAL_2:%.*]] = "tfl.concatenation"([[VAL_0]], [[VAL_1]]) <{axis = 3 : i32, fused_activation_function = "NONE"}> {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<1x1x1x1xf32>, tensor<1x1x1x1xf32>) -> tensor<1x1x1x2xf32>