%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>)

    return %3 : tensor<1x3xf32>
  }
  func.func private @composite_matmul_with_bias_fn_2(%arg0: tensor<1x4xf32>, %arg1: tensor<4x3xf32>, %arg2: tensor<3xf32>) -> tensor<1x3xf32> attributes {tf_quant.composite_function} {
    %0 = "tf.MatMul"(%arg0, %arg1) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> {attr_map = "0:transpose_a,1:transpose_b", device = ""} : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>

    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<1x4xf32>, tensor<4x3xf32>) -> tensor<1x3xf32>
    %1 = stablehlo.add %0, %arg2 : tensor<1x3xf32>
    return %1 : tensor<1x3xf32>
  }

}
func.func @_func(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> tensor<2x2xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01"} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1 = "tf.Identity"(%0) : (tensor<2x2xf32>) -> tensor<2x2xf32>
  return %1 : tensor<2x2xf32>
}

// -----
// The following op sharding is used in the following test case:

  module {
    func.func @constant_add() -> (tensor<3x2xf32>) {
      %cst1 = stablehlo.constant dense<2.4> : tensor<3x2xf32>
      %cst2 = stablehlo.constant dense<5.7> : tensor<3x2xf32>
      %add = stablehlo.add %cst1, %cst2 : (tensor<3x2xf32>, tensor<3x2xf32>) -> tensor<3x2xf32>
      func.return %add : tensor<3x2xf32>
    }
  }
)mlir";

    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 @fakeQuantPerChannelForActivation(%arg0: tensor<8x4xf32>) -> (tensor<8x4xf32>) {
  %arg1 = arith.constant dense<[0.0, -1.0, 1.0, 0.0]> : tensor<4xf32>
  %arg2 = arith.constant dense<[255.0, 254.0, 256.0, 1.0e-9]> : tensor<4xf32>
  %0 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg0, %arg1, %arg2) {num_bits = 5, narrow_range = false} : (tensor<8x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<8x4xf32>
  func.return %0 : tensor<8x4xf32>

        return %2 : tensor<?x2xf32>
      }
      func.func private @composite_fn_1(%arg0: tensor<?x2xf32>, %arg1: tensor<2x2xf32>, %arg2: tensor<2xf32>) -> tensor<?x2xf32> attributes {_from_xla_call_module, tf_quant.composite_function} {
        return %arg0 : tensor<?x2xf32>
      }
    }
  )mlir";

func.func @fakeQuantPerChannelForActivation(%arg0: tensor<8x4xf32>) -> (tensor<8x4xf32>) {
  %arg1 = arith.constant dense<[0.0, -1.0, 1.0, 0.0]> : tensor<4xf32>
  %arg2 = arith.constant dense<[15.0, 14.0, 16.0, 1.0e-9]> : tensor<4xf32>
  %0 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg0, %arg1, %arg2) {num_bits = 3, narrow_range = false} : (tensor<8x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<8x4xf32>
  func.return %0 : tensor<8x4xf32>

}

func.func @main2(%arg0: tensor<2x4xf32>, %arg1: tensor<2x4xf32>) -> tensor<2x4xf32> {
  %0 = "tfl.quantize"(%arg0) {qtype = tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>} : (tensor<2x4xf32>) -> tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>
  %1 = "tfl.quantize"(%arg1) {qtype = tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>} : (tensor<2x4xf32>) -> tensor<2x4x!quant.uniform<u8:f32, 0.49803921568627452>>