auto parallel_execute = builder->create<tf_device::ParallelExecuteOp>(
      cluster_func.getLoc(), 1, output_types);
  cluster_func->remove();
  auto& block = parallel_execute.GetRegionBlockWithIndex(0);
  builder->setInsertionPointToEnd(&block);
  builder->insert(cluster_func);
  cluster_func.replaceAllUsesWith(parallel_execute);
  builder->create<tf_device::ReturnOp>(block.getParent()->getLoc(),

  llvm::SmallVector<llvm::SmallVector<int, 4>, 4> results_to_remove_by_region;
  results_to_remove_by_region.resize(num_regions);
  for (int i = 0; i < num_regions; ++i) {
    Block& block = parallel_execute.GetRegionBlockWithIndex(i);
    auto results = parallel_execute.GetRegionOutputs(i);
    auto& results_to_remove = results_to_remove_by_region[i];
    results_to_remove.reserve(results.size());
    Operation* terminator = block.getTerminator();

  // will be replaced.
  int cluster_idx = -1;
  for (size_t child_idx = 0;
       child_idx < old_parallel_execute.getRegions().size(); ++child_idx) {
    auto& block = old_parallel_execute.GetRegionBlockWithIndex(child_idx);
    if (cluster_func->getBlock() == &block) {
      assert(cluster_idx == -1);
      cluster_idx = child_idx;
    }
  }
  assert(cluster_idx != -1);

  // Move children that are preserved.