%4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

          tf_device.return %c, %b : tensor<f32>, tensor<i32>
        }) {device = "CPU:0"} : () -> (tensor<f32>, tensor<i32>)
        %d = "tf.opD"() : () -> tensor<i1>
        tf_executor.yield %a, %launch#0, %launch#1, %d :
                          tensor<i1>, tensor<f32>, tensor<i32>, tensor<i1>
      }
    ```

    Will be transformed into:

    ```mlir
      %island:5 = tf_executor.island {

    for (auto& use : argument.getUses()) {
      if (llvm::isa<func::ReturnOp, TF::YieldOp>(use.getOwner())) {
        erase_indices.set(use.getOperandNumber());
      }
    }
    func.getArgument(arg_index).dropAllUses();
  }
  if (llvm::isa<func::ReturnOp, TF::YieldOp>(func.front().getTerminator())) {
    terminator->eraseOperands(erase_indices);
  }
}

  This is often useful when tracking the progress of the calibration step which
  is often slow and may look stale if there's no logs being printed.

  Args:
    representative_dataset: The representative dataset.

  Yields:
    The representative samples from `representative_dataset` without any
    modification.
  """
  num_samples: Optional[int] = rd.get_num_samples(representative_dataset)
  if num_samples is None:

  ^bb0(%carg0: tensor<?x!tf_type.qint8>):
    %check = "tf.Equal"(%arg1, %one) : (tensor<i32>, tensor<i32>) -> tensor<i1>
    "tf.Yield"(%check) : (tensor<i1>) -> ()
    }, {
  ^bb0(%barg0: tensor<?x!tf_type.qint8>):
    %id = "tf.Identity"(%barg0) : (tensor<?x!tf_type.qint8>) -> tensor<?x!tf_type.qint8>
    "tf.Yield"(%id) : (tensor<?x!tf_type.qint8>) -> ()
  }) {is_stateless = false} : (tensor<?x!tf_type.qint8>) -> tensor<?x!tf_type.qint8>

   * specified, may re-sleep or yield until time elapses.
   */
  static void delay(long millis) throws InterruptedException {
    long startTime = System.nanoTime();
    long ns = millis * 1000 * 1000;
    for (; ; ) {
      if (millis > 0L) Thread.sleep(millis);
      else // too short to sleep
      Thread.yield();
      long d = ns - (System.nanoTime() - startTime);

  auto identity_op = builder.create<TF::IdentityOp>(
      name_loc, /*result_types=*/main_file_prefix_arg.getType(),
      /*input=*/main_file_prefix_arg);

  builder.create<tf_executor::YieldOp>(name_loc, identity_op.getResult());

  return wrapper_island_op;
}

// Appends `value` to the arguments of the `FetchOp` of `graph_op`.
void AppendValueToFetch(GraphOp graph_op, Value value) {

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

                hasher.putString(element.toString())
            }
        }
        return hasher.hash()
    }

    private
    suspend fun SequenceScope<Failure>.visitFailures(failure: Failure) {
        yield(failure)
        failure.suppressed.forEach { visitFailures(it) }
        failure.causes.forEach { visitFailures(it) }
    }

        newStartedThread(
            new CheckedRunnable() {
              @Override
              public void realRun() {
                while (!a.compareAndSet(0, 2.0, 3.0)) {
                  Thread.yield();
                }
              }
            });

    assertTrue(a.compareAndSet(0, 1.0, 2.0));
    awaitTermination(t);
    assertBitEquals(3.0, a.get(0));
  }