// CHECK: return %[[output]] : tensor<6x3xi8>
  %0 = "tf.ConcatV2"(%arg0, %arg1, %axis) : (tensor<3x3xf32>, tensor<3x3xf32>, tensor<i64>) -> tensor<6x3xf32>
  %1 = "tf.UniformQuantize"(%0, %scales, %zps) {
    quantization_axis = -1 : i64, quantization_min_val = -128 : i64, quantization_max_val = 127 : i64
  } : (tensor<6x3xf32>, tensor<f32>, tensor<i32>) -> tensor<6x3x!tf_type.qint8>
  func.return %1 : tensor<6x3x!tf_type.qint8>
}

  func.func @concat_v2(%arg0: tensor<3x3xf32>, %arg1: tensor<3x3xf32>) -> tensor<6x3xf32> {
    // CHECK: "mhlo.concatenate"({{.*}}) <{dimension = 0 : i64}> : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<6x3xf32>
    %axis = "tf.Const"() { value = dense<0> : tensor<i64> } : () -> tensor<i64>
    %1 = "tf.ConcatV2"(%arg0, %arg1, %axis) : (tensor<3x3xf32>, tensor<3x3xf32>, tensor<i64>) -> tensor<6x3xf32>
    func.return %1 : tensor<6x3xf32>
  }

    %9 = "tfl.fully_connected"(%8, %cst_10, %cst_0) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<4x5xf32>, tensor<8x5xf32>, tensor<8xf32>) -> tensor<4x8xf32>
    %10:4 = "tfl.split"(%cst_5, %9) {num_splits = 4 : i32} : (tensor<i32>, tensor<4x8xf32>) -> (tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>, tensor<4x2xf32>)

  // CHECK-DAG: %[[EXP:.*]] = "tf.Exp"(%[[SHIFTED]]) : (tensor<2x3xf32>) -> tensor<2x3xf32>
  // CHECK-DAG: %[[SUM:.*]] = "tf.Sum"(%[[EXP]], %[[AXIS]]) <{keep_dims = true}> : (tensor<2x3xf32>, tensor<1xi64>) -> tensor<2x1xf32>
  // CHECK-DAG: %[[RESULT:.*]] = "tf.Div"(%[[EXP]], %[[SUM]]) : (tensor<2x3xf32>, tensor<2x1xf32>) -> tensor<2x3xf32>
  // CHECK: return %[[RESULT]]

  %outputs =  "tf.Const"() {device = "/job:localhost/replica:0/task:0/device:CPU:0", value = dense<0> : tensor<i32>} : () -> tensor<i32>...

  %1 = "tf.ReadVariableOp"(%arg1) {device = "/device:CPU:0"} : (tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<1x3xf32> 
  %2 = "tf.AddV2"(%0, %1) {device = "/device:CPU:0"} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  %3 = "tf.Identity"(%2) {device = "/device:CPU:0"} : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %3 : tensor<1x3xf32>
}

// CHECK-NEXT:    %6 = "tf.Select"(%4, %3, %arg4) : (tensor<i1>, tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
// CHECK-NEXT:   %7 = "tf.Identity"(%4) : (tensor<i1>) -> tensor<i1>
// CHECK-NEXT:    %8 = "tf.MatMul"(%5, %arg7) : (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
// CHECK-NEXT:    %9 = "tf.Select"(%7, %8, %arg7) : (tensor<i1>, tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>

  func.return %1 : tensor<i32>
}

func.func private @xla_func_0(%arg0: tensor<1x3xf32>, %arg1: tensor<1x3xf32>) -> tensor<1x3xf32> attributes {tf._XlaMustCompile = true, tf._noinline = true, tf._original_func_name = "should_not_be_used"} {
  %1 = "tf.AddV2"(%arg0, %arg1) : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>
  %2 = "tf.Const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>

    return %1 : tensor<1x3xf32>
  }
}

// -----

// Do not merge when function is not quantized function.

module attributes {tf_saved_model.semantics} {
  // CHECK-LABEL: func.func private @merge_relu_fusion
  func.func private @merge_relu_fusion(%arg0: tensor<1x4xf32>) -> tensor<1x3xf32> {

  return %2 : tensor<1x3xf32>
}
func.func private @composite_dot_general_fn_1(%arg0: tensor<1x1024xf32>, %arg1: tensor<1024x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
  %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
  return %0 : tensor<1x3xf32>
}
// CHECK: func.func @main