%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>

  %8 = "tfl.concatenation"(%2, %0) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x2xf32>
  %9 = "quantfork.stats"(%8) {layerStats = dense<[-0.488159984, 0.189515018]> : tensor<2xf32>} : (tensor<1x2xf32>) -> tensor<1x2xf32>
  %10 = "tfl.concatenation"(%9, %7) {axis = -1 : i32, fused_activation_function = "NONE"} : (tensor<1x2xf32>, tensor<1x2xf32>) -> tensor<1x4xf32>

    func.func @main() -> (tensor<1x4x4xf32>) {
      %%arg0 = "tf.Const"() {value = dense<-3.0> : tensor<1x4x2xf32>} : () -> tensor<1x4x2xf32>
      %%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>

  %2 = "tf.Add"(%1, %1) {
    _xla_compile_device_type = "TPU", _replication_info = "cluster",
    device = "/task:0/device:TPU:0", dtype = f32
  } : (tensor<1x80xf32>, tensor<1x80xf32>) -> tensor<1x80xf32>
  %3 = "tf.ResourceGatherNd"(%arg0, %0) {
    Tindices = i32
  } : (tensor<*x!tf_type.resource<tensor<80xf32>>>, tensor<i32>) -> tensor<1x80xf32>

%arg17, %arg18, %arg19, %arg20, %arg21, %arg22, %arg23) <{fused_activation_function = "NONE", time_major = false}> : (tensor<600x?x20xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<40xf32>, tensor<?x?xf32>, tensor<?xf32>, tensor<600x40xf32>, tensor<600x40xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>, tensor<?xf32>)...

  %4 = "tf.MatMul"(%arg0, %3) {device = "", transpose_a = false, transpose_b = false} : (tensor<2x3xf32>, tensor<3x4xf32>) -> tensor<2x4xf32>
  %5 = "tf.Identity"(%4) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  %6 = "tf.Identity"(%5) {device = ""} : (tensor<2x4xf32>) -> tensor<2x4xf32>
  func.return %6 : tensor<2x4xf32>

  %cst_weights = arith.constant dense<[[1.0, 1.0e38, 1.0, -1.0e38]]> : tensor<1x4xf32>
  %cst_bias = arith.constant dense<0.0> : tensor<1xf32>

  %0 = "tfl.fully_connected" (%cst_input, %cst_weights, %cst_bias) {fused_activation_function = "NONE", keep_num_dims = false, weights_format = "DEFAULT"} : (tensor<4xf32>, tensor<1x4xf32>, tensor<1xf32>) -> tensor<1xf32>

  func.return %0 : tensor<1xf32>

    // CHECK-NEXT: %[[DIV:.*]] = mhlo.divide %{{.*}}, %[[ADD]] : tensor<4x10xf32>
    // CHECK-NEXT: return %[[DIV]] : tensor<4x10xf32>
    %0 = "tf.Softsign"(%arg0) : (tensor<4x10xf32>) -> tensor<4x10xf32>
    func.return %0 : tensor<4x10xf32>
}

// -----

// CHECK-LABEL: func @softsign_grad
func.func @softsign_grad(%arg0: tensor<4x10xf32>, %arg1: tensor<4x10xf32>) -> tensor<4x10xf32> {

// CHECK:          return %[[VALUES]], %[[INDICES]] : tensor<1x4xf32>, tensor<1x4xi32>
// CHECK:        }
func.func @convert_approx_top_k_custom_call(%arg0: tensor<1x4xf32>, %arg1: tensor<1x4xi32>, %arg2: tensor<f32>, %arg3: tensor<i32>) -> (tensor<1x4xf32>, tensor<1x4xi32>) {
  %0:2 = mhlo.custom_call @ApproxTopK(%arg0, %arg1, %arg2, %arg3) {

    }
  }
  return new_arg_type;
}

// Combination of value producer and port of value produced (e.g.,
//   <value result output>:<value in output tensor>,
// so for tf.Const -> tensor<10x20xf32>, [0,2,18] would point to a unique output
// scalar value).
struct ValuePort {
  PointerUnion<Operation*, BlockArgument> producer;
  SmallVector<unsigned int, 2> port;