%1 = "quantfork.qcast"(%0) {volatile} : (tensor<1x3xf32>) -> tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
      %2 = "quantfork.dcast"(%1) : (tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<1x3xf32>
      %3 = stablehlo.reshape %2 : (tensor<1x3xf32>) -> tensor<3x1xf32>

  // CHECK-SAME: ({{%.+}}: tensor<1x2x3xf32>)
  // CHECK-SAME: -> (tensor<1x8x3xf32>, tensor<1x8x3xf32>)
  func.func @while_shape_invariant_different_dims(%arg0: tensor<1x2x3xf32>) -> (tensor<1x8x3xf32>, tensor<1x8x3xf32>) {
    // CHECK: "tf.While"
    // CHECK-SAME: (tensor<1x2x3xf32>)
    // CHECK-SAME: -> tensor<1x8x3xf32>

      %cst = stablehlo.constant dense<1.0> : tensor<1x3xf32>
      %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [] : (tensor<1x1024xf32>, tensor<1024x3xf32>) -> tensor<1x3xf32>
      %1 = stablehlo.subtract %cst, %0 : tensor<1x3xf32>
      %2 = stablehlo.add %0, %cst : tensor<1x3xf32>
      return %2 : tensor<1x3xf32>
    }
  }
)mlir";

    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>
    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2

  %1 = "tfl.neg"(%arg0) : (tensor<2x3xf32>) -> tensor<2x3xf32>
  %2 = "tfl.relu"(%1) : (tensor<2x3xf32>) -> tensor<2x3xf32>
  %3 = "tfl.mul"(%alpha, %2) {fused_activation_function = "NONE"} : (tensor<f32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  %4 = "tfl.add"(%0, %3) {fused_activation_function = "NONE"} : (tensor<2x3xf32>, tensor<2x3xf32>) -> tensor<2x3xf32>
  func.return %4 : tensor<2x3xf32>

      %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<1x3xf32>) -> tensor<1x3xf32>
      return %2 : tensor<1x3xf32>
    }

    func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {

    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> {

  func.func private @composite_and_select(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>, %arg2: tensor<1x3xi1>, %arg3: tensor<1x3xf32>) -> tensor<1x3xf32> {
    // CHECK: %[[Q1:.*]] = "quantfork.qcast"(%[[ARG0]]) {volatile} : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<i8:f32, 5.000000e-03>>

// CHECK-LABEL: func private @batched_function
func.func private @batched_function(%arg0: tensor<1x3xf32>) -> tensor<1x3xf32> {
  %2 = "tf.Identity"(%arg0) : (tensor<1x3xf32>) -> tensor<1x3xf32>
  func.return %2 : tensor<1x3xf32>
}

// CHECK-LABEL: func @main
func.func @main(%arg0: tensor<1x3xf32>) -> tensor<*xf32> {
  // CHECK:  "tf.BatchFunction"
  // CHECK-SAME: allowed_batch_sizes = [6]

  func.return %2 : tensor<1x3xf32>
}

// CHECK-LABEL: func @main
func.func @main(%arg0: tensor<1x3xf32>) -> tensor<*xf32> attributes {tf.entry_function = {control_outputs = "", inputs = "input:0", outputs = "batch/BatchFunction:0"}} {
  %0 = "tf.VarHandleOp"() {device = "/device:CPU:0", container = "", shared_name = "variable"} : () -> tensor<!tf_type.resource<tensor<1x3xf32>>>