// CHECK-DAG: [[cst_0:%.*]] = "tf.Const"() <{value = dense<2> : tensor<i32>}>
// CHECK: [[r0:%.*]] = "tf.AddV2"([[x]], [[cst]])
// CHECK: [[r1:%.*]] = "tf.AddV2"([[y]], [[r0]])
// CHECK: [[r2:%.*]] = "tf.AddV2"([[x]], [[cst_0]])
// CHECK: [[r3:%.*]] = "tf.AddV2"([[y]], [[r2]])
// CHECK: return [[r0]], [[r1]], [[r3]]

// CHECK-LABEL: func @merge_stateless

  // CHECK-DAG: %[[CAST:.*]] = "tf.Cast"(%arg0) <{Truncate = false}>
  // CHECK-DAG: %[[SHIFT:.*]] = "tf.AddV2"(%[[CAST]], %[[HALF_RANGE]])
  // CHECK-DAG: %[[DRANGE:.*]] = "tf.Sub"(%arg2, %arg1)
  // CHECK-DAG: %[[SCALE:.*]] = "tf.Div"(%[[DRANGE]], %[[C255:.*]])
  // CHECK-DAG: %[[SS:.*]] = "tf.Mul"(%[[SHIFT]], %[[SCALE]])
  // CHECK-DAG: %[[RESULT:.*]] = "tf.AddV2"(%[[SS]], %arg1)

  Output add1 =
      ops::Add(root.WithOpName("add1"), sw_0.output_false, sw_2.output_false);
  Output add2 =
      ops::Add(root.WithOpName("add2"), sw_1.output_false, sw_2.output_false);
  ops::Merge m1(root.WithOpName("m1"), {add1, add2});

  Output add3 = ops::Add(root.WithOpName("add3"), add0, m1.output);

  std::unique_ptr<DeadnessAnalysis> result;
  TF_ASSERT_OK(AnalyzeDeadness(root.graph(), &result));

  // CHECK-NOT: "tf.ReadVariableOp"
  // CHECK: %[[CONST:.*]] = "tf.Const"() <{value = dense<4.200000e+01> : tensor<f32>}>
  // CHECK: %[[ADD1:[0-9]*]] = "tf.AddV2"(%arg1, %[[CONST]])
  // CHECK: %[[ADD2:[0-9]*]] = "tf.AddV2"(%[[ADD1]], %arg1)
  // CHECK: %[[PACK:[0-9]*]] = "tf.Pack"(%[[CONST]], %[[ADD2]])
  // CHECK: return %[[PACK]]

  %0 = "tf.Const"() {value = dense<4.200000e+01> : tensor<f32>} : () -> tensor<f32>

// RUN: tf-tfrt-opt -split-input-file -tf-mlrt-assign-op-key %s | FileCheck %s

// CHECK-LABEL: func @main
// CHECK: tf.AddV2
// CHECK-SAME: {__op_key = 0 : i32}

// CHECK: tf.AddV2
// CHECK-SAME: {__op_key = 1 : i32}

// CHECK: tf.AddV2
// CHECK-SAME: {__op_key = 2 : i32}

// CHECK: tf.AddV2
// CHECK-SAME: {__op_key = 3 : i32}

// CHECK: tf.Sub
// CHECK-SAME: {__op_key = 4 : i32}

// CHECK: tf.Sub

    %16 = "tf.AddV2"(%15, %arg10) : (tensor<1x2x2x3xi32>, tensor<i32>) -> tensor<1x2x2x3xi32>
    %17 = "tf.Div"(%cst_1, %arg9) : (tensor<f32>, tensor<f32>) -> tensor<f32>
    %18 = "tf.Round"(%17) : (tensor<f32>) -> tensor<f32>
    %19 = "tf.Cast"(%18) : (tensor<f32>) -> tensor<i32>
    %20 = "tf.AddV2"(%19, %arg10) : (tensor<i32>, tensor<i32>) -> tensor<i32>

// CHECK-NEXT:   %[[RESULT_1:.*]] = "tf.AddV2"(%[[RESULT_0]], %[[RESULT_0]])
// CHECK-SAME:  device = "/job:worker/replica:0/task:1/device:CPU:0"

// Subgraph of @while_body function that is placed on worker:2
// CHECK: func @[[BODY_PARTITION_1]]() attributes {host = "/job:worker/replica:0/task:2"}
// CHECK-NEXT:   %[[RESULT_0:.*]] = tf_device.receive "key-1"
// CHECK-NEXT:   %[[RESULT_1:.*]] = "tf.AddV2"(%[[RESULT_0]], %[[RESULT_0]])

// CHECK-NEXT: = "tf.AddV2"{{.*}}"cluster1"
// CHECK-NEXT: = "tf.AddV2"{{.*}}"cluster1"
// CHECK: %[[ISLAND2:.*]], {{.*}} = tf_executor.island
// CHECK-NEXT: = "tf.Const"{{.*}}"cluster2"
// CHECK-NEXT: = "tf.Const"{{.*}}"cluster2"
// CHECK-NEXT: = "tf.AddV2"{{.*}}"cluster2"
// CHECK-NEXT: = "tf.AddV2"{{.*}}"cluster2"
// CHECK: tf_executor.island wraps "tf.AddV2"(%[[ISLAND1]], %[[ISLAND2]])

// CHECK:       tf.WhileRegion
// CHECK:       ^bb0
// CHECK:       tf.OpA
// CHECK:       ^bb0([[ARG_2:%[a-zA-Z0-9_]+]]
// CHECK-SAME:  [[ARG_3:%[a-zA-Z0-9_]+]]: tensor<i32>)
// CHECK-NEXT:  [[RES_3:%.*]] = "tf.AddV2"([[ARG_2]], [[RES_1]])
// CHECK-NEXT:  [[RES_4:%.*]] = "tf.Div"([[ARG_3]], [[RES_2]])
// CHECK:       "tf.Yield"([[RES_3]], [[RES_4]])

  // CHECK-NEXT: tf_mlrt.executeop([[y]]
  %w = "tf.AddV2"(%y, %z) {__op_key = 2: i32}: (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK-NEXT: tf_mlrt.executeop([[y]]
  %u = "tf.AddV2"(%y, %z) {__op_key = 3: i32} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  // CHECK-NEXT: tf_mlrt.executeop([[y]]
  %v = "tf.AddV2"(%y, %z) {__op_key = 4: i32}: (tensor<i32>, tensor<i32>) -> tensor<i32>