return %2 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>
  }
}

// CHECK-LABEL: @add_f32(%arg0: tensor<3x3xbf16>, %arg1: tensor<3x3xbf16>) -> tensor<3x3xbf16>
func.func @add_f32(%arg0: tensor<3x3xf32>, %arg1: tensor<3x3xf32>) -> tensor<3x3xf32> {
  // CHECK-NOT: f32
  // CHECK: stablehlo.add
  %0 = stablehlo.add %arg0, %arg1: (tensor<3x3xf32>, tensor<3x3xf32>) -> tensor<3x3xf32>
  return %0 : tensor<3x3xf32>
}

// -----

// CHECK-LABEL: @add_f64(%arg0: tensor<3x3xbf16>, %arg1: tensor<3x3xbf16>) -> tensor<3x3xbf16>

    return %1 : tensor<1x3xf32>
  }

  func.func private @composite_dot_general_fn(%arg0: tensor<1x2xf32>, %arg1: tensor<2x3xf32>) -> tensor<1x3xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.dot_general %arg0, %arg1, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    return %0 : tensor<1x3xf32>
  }
}

  %3 = "tfl.dequantize"(%2) : (tensor<2x2x!quant.uniform<u8:f32, 1.0>>) -> tensor<2x2xf32>
  func.return %3 : tensor<2x2xf32>

// CHECK-NEXT: %[[Q:.*]] = "tfl.quantize"(%arg0) <{qtype = tensor<1x2x!quant.uniform<u8:f32, 1.000000e+00>>}> : (tensor<1x2xf32>) -> tensor<1x2x!quant.uniform<u8:f32, 1.000000e+00>>

  constexpr char mlir_module_string[] = R"mlir(
  module attributes {tf.versions = {bad_consumers = [], min_consumer = 0 : i32, producer = 268 : i32}} {
    func.func @main(%arg0 : tensor<2x2xf32>) -> tensor<2x2xf32> {
      %max = "tf.Const"() { value = dense<12.0> : tensor<f32> } : () -> tensor<f32>
      %min = "tf.Const"() { value = dense<-25.0> : tensor<f32> } : () -> tensor<f32>

func.func @TfBatchMatMulV2(%arg0: tensor<4x128x2xf32>, %arg1:  tensor<2x1xf32>) -> tensor<4x128x1xf32> {
  %0 = "tfl.custom"(%arg0, %arg1) {
    custom_code = "FlexBatchMatMulV2",
    custom_option = #tfl<const_bytes : "0x0D42617463684D61744D756C56320038120D42617463684D61744D756C56321A001A002A070A0154120230012A0B0A0561646A5F78120228002A0B0A0561646A5F791202280032000002493B1414042801">
  } : (tensor<4x128x2xf32>, tensor<2x1xf32>) -> tensor<4x128x1xf32>

  %3 = "tfl.pack"(%1, %2) {axis = 0 : i32, per_device_costs = {CPU = 2.0 : f32, GPU = -1.0 : f32}, values_count = 2 : i32, tac.device = "CPU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<2x1xf32>
  func.return %3 : tensor<2x1xf32>
})";
  const std::string kExpectedFB = CreateRuntimeMetadata();
  mlir::DialectRegistry registry;
  registry.insert<mlir::TFL::TensorFlowLiteDialect, mlir::arith::ArithDialect,

// CHECK-LABEL: dont_move_transpose_different_ranks
func.func @dont_move_transpose_different_ranks(%arg0:tensor<1x1x2x3xf32>, %arg1:tensor<2x3xf32>) -> tensor<1x2x1x3xf32> {
  %cst = "tf.Const"() {value = dense<[0, 2, 1, 3]> : tensor<4xi32>} : () -> tensor<4xi32>
  %0 = "tf.AddV2"(%arg0, %arg1) {device = ""} : (tensor<1x1x2x3xf32>, tensor<2x3xf32>) -> tensor<1x1x2x3xf32>
  %1 = "tf.Transpose"(%0, %cst) {device = ""} : (tensor<1x1x2x3xf32>, tensor<4xi32>) -> tensor<1x2x1x3xf32>

  %cst0 = mhlo.constant dense<[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]> : tensor<2x3xf32>
  %cst1 = mhlo.constant dense<[[[[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], [[7.0, 8.0, 9.0], [10.0, 11.0, 12.0]]]]> : tensor<1x2x2x3xf32>
  %0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<[1, 3]> : tensor<2xi64>}> : (tensor<2x3xf32>) -> tensor<1x2x2x3xf32>
  %1 = mhlo.multiply %0, %cst1 : tensor<1x2x2x3xf32>

  func.return %0#0 : tensor<2xi32>
}

// CHECK-LABEL: mirror_pad
func.func @mirror_pad(%arg0: tensor<2x3xcomplex<f64>>) -> tensor<4x7xcomplex<f64>> {
  %0 = mhlo.constant dense<[[1, 1], [2, 2]]> : tensor<2x2xi32>
  // NO_FALLBACK: tf.MirrorPad