%arg8: tensor<2048x640xf32>,
  %arg9: tensor<2048xf32>,
  %arg10: tensor<2048xf32>,
  %arg11: tensor<2048xf32>,
  %arg12: tensor<2048xf32>,
  %arg13: tensor<640x2048xf32>,
  %arg14: tensor<640xf32>,
  %arg15: tensor<2048xf32>,
  %arg16: tensor<2048xf32>,
  %arg17: tensor<2048xf32>,
  %arg18: tensor<2048xf32>,
  %arg19: tensor<1x640xf32>,
  %arg20: tensor<1x2048xf32>
) -> tensor<1x640xf32> {

  // CHECK-LABEL: test_batch_matmul_to_einsum
  // CHECK: "tf.Einsum"(%arg0, %arg1) <{equation = "...mk,...kn->...mn"}> : (tensor<1x2x3xf32>, tensor<3x4xf32>) -> tensor<1x2x4xf32>
  %0 = "tf.BatchMatMul"(%arg0, %arg1) {adj_x = false, adj_y = false} : (tensor<1x2x3xf32>, tensor<3x4xf32>) -> tensor<1x2x4xf32>
  func.return %0: tensor<1x2x4xf32>
}

      tf_executor.yield %1, %1 : tensor<2048xf32>, tensor<2048xf32>
    }
    tf_executor.fetch %outputs#0, %outputs#1 : tensor<2048xf32>, tensor<2048xf32>
  }
  func.return %0#0, %0#1 : tensor<2048xf32>, tensor<2048xf32>
}

// CHECK: func @tpu_load_embedding_ops_sink_controls
// CHECK: {{%.+}}, [[READ0:%.+]] = tf_executor.island wraps "tf.ReadVariableOp"

      tf_executor.yield %1, %1 : tensor<2048xf32>, tensor<2048xf32>
    }
    tf_executor.fetch %outputs#0, %outputs#1 : tensor<2048xf32>, tensor<2048xf32>
  }
  func.return %0#0, %0#1 : tensor<2048xf32>, tensor<2048xf32>
}
// CHECK-LABEL: func @single_op_island_duplicate_result
// CHECK: %[[CONST:.*]], %{{.*}} = tf_executor.island wraps "tf.Const"
// CHECK: tf_executor.fetch %[[CONST]], %[[CONST]]

func.func @single_op_island_forward_block_arg(%arg0: tensor<?x?x?x?xbf16>) -> (tensor<2048xf32>, tensor<?x?x?x?xbf16>) {
  %0:2 = tf_executor.graph {
    %outputs, %control = tf_executor.island wraps "tf.Const"() {value = dense<0.000000e+00> : tensor<2048xf32>} : () -> tensor<2048xf32>
    tf_executor.fetch %outputs, %arg0 : tensor<2048xf32>, tensor<?x?x?x?xbf16>
  }
  return %0#0, %0#1 : tensor<2048xf32>, tensor<?x?x?x?xbf16>
}

// RUN: tf-quant-opt %s -split-input-file -quant-merge-duplicate-resource-ops | FileCheck %s

func.func @merge_duplicate_variable(%arg0: tensor<1x20xf32>, %arg1: tensor<!tf_type.string>) -> (tensor<20x4096xf32>) {
  %0 = tf_executor.graph {
    %outputs_5, %control_6 = tf_executor.island wraps "tf.Const"() {value = dense<""> : tensor<1x!tf_type.string>} : () -> tensor<1x!tf_type.string>

      %%arg1 = "tf.Const"() {value = dense<-3.0> : tensor<1x2x4xf32>} : () -> tensor<1x2x4xf32>

      %%1 = "tf.%s"(%%arg0, %%arg1) {T = f32, adj_x = false, adj_y = false, grad_x = false, grad_y = false, device = ""} : (tensor<1x4x2xf32>, tensor<1x2x4xf32>) -> tensor<1x4x4xf32>

      func.return %%1 : tensor<1x4x4xf32>
    }
  })";
  std::string mat_mul_method =

    %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[-2.0, 1.0]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %cell_input = arith.constant dense<1.0> : tensor<1x20xf32>
    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %0 = "tfl.unidirectional_sequence_lstm"(%arg0,

    %recurrent_stats = "quantfork.stats"(%recurrent_input) {layerStats = dense<[-2.0, 1.0]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>
    %cell_input = arith.constant dense<1.0> : tensor<1x20xf32>
    %cell_stats = "quantfork.stats"(%cell_input) {layerStats = dense<[-2.73090601, 7.94872093]> : tensor<2xf32>} : (tensor<1x20xf32>) -> tensor<1x20xf32>

  %3 = "tf.Const"() {device = "", dtype = f32, value = dense<0.000000e+00>: tensor<1x28xf32>} : () -> tensor<1x28xf32>
  %4:2 = "tf.UnidirectionalSequenceRnn"(%arg, %1, %1, %2, %3) {_tflite_input_indices = [0, 1, 2, 3, 4], device = ""} : (tensor<28x1x28xf32>, tensor<28x28xf32>, tensor<28x28xf32>, tensor<28xf32>, tensor<1x28xf32>) -> (tensor<*xf32>, tensor<28x1x28xf32>)
  func.return %4#1 : tensor<28x1x28xf32>
}