(Mul64 x y) =>
	(Int64Make
		(Add32 <typ.UInt32>
			(Mul32 <typ.UInt32> (Int64Lo x) (Int64Hi y))
			(Add32 <typ.UInt32>
				(Mul32 <typ.UInt32> (Int64Hi x) (Int64Lo y))
				(Select0 <typ.UInt32> (Mul32uhilo (Int64Lo x) (Int64Lo y)))))
		(Select1 <typ.UInt32> (Mul32uhilo (Int64Lo x) (Int64Lo y))))

(And64 x y) =>
	(Int64Make
		(And32 <typ.UInt32> (Int64Hi x) (Int64Hi y))
		(And32 <typ.UInt32> (Int64Lo x) (Int64Lo y)))

  // Just to introduce an extra use for %cst.
  %1 = "tf.AddV2"(%feature_group_count, %feature_group_count) {device = ""} : (tensor<i32>, tensor<i32>) -> tensor<i32>
  %2 = "tf.AddV2"(%lhs_dilation, %lhs_dilation) {device = ""} : (tensor<3xi32>, tensor<3xi32>) -> tensor<3xi32>
  %3 = "tf.AddV2"(%rhs_dilation, %rhs_dilation) {device = ""} : (tensor<3xi32>, tensor<3xi32>) -> tensor<3xi32>

  // CHECK-NEXT: %[[ADDN2:.*]] = "tf.AddN"(%[[ADDN]], %[[ZEROS_LIKE]]) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10xf32>
  %addn2 = "tf.AddN"(%addn, %zeros-like) : (tensor<!tf_type.variant<tensor<f32>>>, tensor<!tf_type.variant<tensor<f32>>>) -> tensor<!tf_type.variant<tensor<f32>>>
  %stack = "tf.TensorListStack"(%addn2, %elem_shape) : (tensor<!tf_type.variant<tensor<f32>>>, tensor<0xi32>) -> tensor<10xf32>

// CHECK:    %[[ADD3:.*]], %[[ADD3_control:.*]] = tf_executor.island(%[[ISLAND1]], %[[ADD2_control]]) wraps "tf.Add"(%[[ADD2]], %[[ADD2]])
// CHECK:    %[[PRINT2:.*]], %[[PRINT2_control:.*]] = tf_executor.island wraps "tf.Print"(%[[ADD3]]) <{message = "add result"}>
// CHECK:    tf_executor.fetch %[[ADD2]], %[[MUL]], %[[PRINT1_control]], %[[PRINT2_control:.*]] :
// CHECK:  }

// CHECK: [[PROMISE:%.*]], [[FUTURE:%.*]] = "tf_mlrt.allocate_futures"
// CHECK: tf.AddV2
// CHECK: tf.AddV2
// CHECK: tf.AddV2
// CHECK: [[VALUE:%.*]] = "tf.AddV2"
// CHECK: [[HANDLE:%.*]] = mlrt.async([[VALUE]], {{%.*}}, [[PROMISE]])
// CHECK-SAME: callee = @main_stream_{{[0-9]*}}
// CHECK: tf.AddV2
// CHECK: tf.AddV2
// CHECK: tf.AddV2
// CHECK: [[x:%.*]] = "tf.AddV2"
// CHECK: [[y:%.*]] = "tf_mlrt.tf_await"([[FUTURE]])

  // 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>

// 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>

		addr1 := v1.UnsafeAddr()
		addr2 := v2.UnsafeAddr()
		if addr1 > addr2 {
			// Canonicalize order to reduce number of entries in visited.
			addr1, addr2 = addr2, addr1
		}

		// Short circuit if references are identical ...
		if addr1 == addr2 {
			return true
		}

		// ... or already seen
		typ := v1.Type()
		v := visit{addr1, addr2, typ}
		if visited[v] {
			return true
		}