%1 = "tfl.abs"(%arg2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    %2 = stablehlo.add %1, %arg2 : tensor<5xf32>
    %3 = "tfl.abs"(%2) {fused_activation_function = "NONE"} : (tensor<5xf32>) -> tensor<5xf32>
    "stablehlo.return"(%3) : (tensor<5xf32>) -> ()
  }) {dimensions = array<i64: 0>} : (tensor<7x5xf32>, tensor<5xf32>) -> tensor<5xf32>
  func.return %0: tensor<5xf32>
}

// -----

      // Some loop invariant math
      %li0 = "tf.Div"(%a, %b) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
      %cst = arith.constant dense<22.0> : tensor<f32>
      %li1 = "tf.Mul"(%li0, %cst) : (tensor<4xf32>, tensor<f32>) -> tensor<4xf32>

      %final = "tf.Add"(%add, %li1) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
      "tf.Yield"(%final, %sub) : (tensor<4xf32>, tensor<i32>) -> ()
    }

// CHECK: tf.MyCustomOp(tensor<4xf32>, tensor<4xf32>) -> (tensor<4xf32>, tensor<3xf32>) : {name = "MyCustomOp"}

func.func @main(tensor<4xf32>) -> tensor<4xf32> {
^bb0(%arg0: tensor<4xf32>):
  %0 = "tfl.pseudo_const" () {name = "Const", value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32>
  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE", name = "mul"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>

  // CHECK: %{{.*}} = "tfl.reshape"(%{{.*}}, %[[cst]]) : (tensor<4xf32>, tensor<2xi32>) -> tensor<2x2xf32>
  %cst = arith.constant dense<[2, 2]> : tensor<2xi32>
  %0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.reshape" (%0, %cst) : (tensor<4xf32>, tensor<2xi32>) -> tensor<2x2xf32> loc("reshape")

    "tf.AssignVariableOp"(%2, %arg0) : (tensor<!tf_type.resource<tensor<?xf32>>>, tensor<?xf32>) -> ()
    return
  }
}

// -----

// CHECK-LABEL: module
module attributes {tf_saved_model.semantics} {
  // CHECK: ml_program.global{{.*}}[[V:@.+]]({{.*}}) : tensor<3xf32>
  // CHECK: func.func @read_inserts_cast_if_necessary(%arg0: tensor<?xf32>
  // CHECK: [[V1:%.*]] = ml_program.global_load

module {
func.func @main(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> attributes {tf.entry_function = {inputs = "input0,input1,input2,input3", outputs = "output"}} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  %1 = "tfl.mul"(%0, %arg2) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

// CHECK-NEXT:   "tf.LayerNorm"(%arg4, %arg5, %arg6, %arg7) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
// CHECK-NEXT:   "tfl.yield"
// CHECK-NEXT: }) {_tfl_quant_trait = "fully_quantizable", device = ""} :

    %0 = tf_executor.graph {
      %outputs, %control = tf_executor.island wraps "tf.Const"() {device = "", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>
      %outputs_0, %control_1 = tf_executor.island wraps "tf.Const"() {device = "", dtype = f32, value = dense<[]> : tensor<0xf32>} : () -> tensor<0xf32>

// -----

func.func @testAddReluPack(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>) {
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU6"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
   // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"
  %1 = "tfl.add"(%arg0, %0) {fused_activation_function = "RELU"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

  %0:2 = "tf.NonMaxSuppressionV4"(%arg0, %arg1, %max_size, %arg2, %arg3) {pad_to_max_output_size = true}: (tensor<3x4xf32>, tensor<3xf32>, tensor<i32>, tensor<f32>, tensor<f32>) -> (tensor<2xi32>, tensor<i32>)
  func.return %0#0 : tensor<2xi32>
}

// CHECK-LABEL: mirror_pad
func.func @mirror_pad(%arg0: tensor<2x3xcomplex<f64>>) -> tensor<4x7xcomplex<f64>> {