func.return
}

func.func private @write_chain_func(%arg0: tensor<1xf32>) -> (tensor<1xf32>) {
  %cst = "tf.Const"() <{value = dense<[[0, 1]]> : tensor<1x2xi32>}> : () -> tensor<1x2xi32>
  %0 = "tf.XlaAllReduce"(%arg0, %cst) <{mode = "CrossReplica", reduce_op = "Add"}> : (tensor<1xf32>, tensor<1x2xi32>) -> tensor<1xf32>
  return %0 : tensor<1xf32>

func.func @QuantizedCatsAddRequantsTest(%arg0: tensor<1x1xf32>, %arg1: tensor<1x1xf32>, %arg2: tensor<1x1xf32>, %arg3: tensor<1x1xf32>) -> (tensor<1x4xf32>, tensor<1x3xf32>) {
  %0 = "quantfork.stats"(%arg0) {layerStats = dense<[-0.440728068, 0.189515018]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>
  %1 = "quantfork.stats"(%arg1) {layerStats = dense<[-0.154693216, 0.26483655]> : tensor<2xf32>} : (tensor<1x1xf32>) -> tensor<1x1xf32>

// CHECK:           }) : (tensor<1x?xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, none, none, none, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<3x1xf32>, tensor<3xf32>, tensor<1x3xf32>, tensor<1x1xf32>, none, none, none, none) -> tensor<1x3xf32>
// CHECK:           [[VAL_50:%.*]] = tensor.cast [[VAL_51:%.*]] : tensor<1x3xf32> to tensor<1x?xf32>

// CHECK-DAG: %[[padding_rank_2:.*]] = "tf.Reshape"(%[[padding_rank_1]], {{.*}}) : (tensor<8xi32>, tensor<2xi64>) -> tensor<4x2xi32>
// CHECK-DAG: %[[input_padded:.*]] = "tf.PadV2"(%{{.*}}, %[[padding_rank_2]], {{.*}}) : (tensor<?x?x?x3xi8>, tensor<4x2xi32>, tensor<i8>) -> tensor<?x?x?x3xi8>

// CHECK: return %[[GATHER]] : tensor<1x300x10xf32>

// -----

module {
  func.func @propagate_xlagather(%arg0: tensor<10x2xi32>) -> tensor<1x300x10xf32> {
    %cst = "tf.Const"() {value = dense<[1, 1, 300]> : tensor<3xi64>} : () -> tensor<3xi64>

  // CHECK-NEXT:     %[[GROUP:.*]] = "tf.Const"()
  // CHECK-SAME:       [0, 1, 2, 3]
  // CHECK-NEXT:     %[[REDUCED:.*]] = "tf.XlaAllReduce"(%[[ID]], %[[GROUP]]) <{mode = "CrossReplica", reduce_op = "Add"}> : (tensor<f32>, tensor<1x4xi32>) -> tensor<f32>
  // CHECK-NEXT:     "tf.OpA"(%[[REDUCED]]) : (tensor<f32>) -> ()
  tf_device.replicate {n = 4 : i32} {
    "tf_device.cluster"() ({
      "tf.OpA"(%arg0) : (tensor<f32>) -> ()
      tf_device.return

  %one = "glue.constant"() { value = 1: i32 } : () -> i32
  %done = "glue.compare" (%one, %one) { predicate = #glue<"compare LTE"> } : (i32, i32) -> i1
  %2 = mhlo.constant dense<[[1.1]]> : tensor<1x1xf32>
  %3 = mhlo.multiply %2, %2 : tensor<1x1xf32>
  %cst = "tf.Const"() {value = dense<0.0> : tensor<f32>} : () -> tensor<f32>
  %0 = "tf.AddV2"(%arg0, %cst) {device = "/device:CPU:0"} : (tensor<f32>, tensor<f32>) -> tensor<f32>

  %elem_1 = "tf._FirstOp"() : () -> tensor<8x1xf32>
  %tl_set_item = "tf.TensorListSetItem"(%tl, %one, %elem_1) : (tensor<!tf_type.variant<tensor<?x1xf32>>>, tensor<i32>, tensor<8x1xf32>) -> tensor<!tf_type.variant<tensor<?x1xf32>>>
  %elem_2 = "tf._SecondOp"() : () -> tensor<16x1xf32>

  // Just to introduce extra usages for %cst.
  %3 = "tf.Mul"(%cst, %cst) {device = ""} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>

  return %2: tensor<1x3xi32>
}
// The output of tf.Shape is considered a compile-time constant, so the
// constant leading to tf.Shape (which transitively becomes an input to the
// first arg of tf.Fill) is not duplicated.

  // CHECK-LABEL: @xla_svd
  func.func @xla_svd(%arg0: tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>) {
    // CHECK-NOT: XlaSvd
    %s, %u, %v = "tf.XlaSvd"(%arg0) {max_iter = 1, epsilon = 1.0E-09 : f32, precision_config = ""} : (tensor<1x1xf32>) -> (tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>)
    func.return %s, %u, %v : tensor<1xf32>, tensor<1x1xf32>, tensor<1x1xf32>
  }

  func.func @identity(%arg0: f32) -> f32 {