// Get total bytes transferred.
int64_t GetTransferredTensorBytes(func::CallOp from_graph,
                                  func::CallOp to_graph) {
  int64_t total_size_transferred = 0;
  for (auto input : to_graph.getOperands()) {
    Operation* input_op = input.getDefiningOp();
    if (input_op && input_op == from_graph.getOperation()) {
      auto input_type =
          mlir::dyn_cast_or_null<RankedTensorType>(input.getType());

                                       decomposed_partitioned_call_callees))) {
    return failure();
  }
  if (!signature_change) return success();
  auto new_if_operands = llvm::to_vector<8>(if_op.getOperands());
  for (auto operand : if_op.getOperands()) {
    auto it = data_var_to_size_var.find(operand);
    if (it == data_var_to_size_var.end()) continue;
    new_if_operands.push_back(it->getSecond());
  }

      liveouts.snapshot_previous_state();
      return;
    }

    reverse_subgraph.insert(op);
    defined_values.insert(op->getResults().begin(), op->getResults().end());
    operands.insert(op->getOperands().begin(), op->getOperands().end());
  };

  for (Operation* op : reverse_main_func_block_ops) {
    if (!ops_to_add.contains(op)) continue;
    // When hitting a non-StableHLO op, i.e. tf.CustomAggregatorOp, start

      // Propagate the analysis results from within the callee's body.
      PropagatePotentiallyWrittenUpFromCallee(batch_function.func().getRegion(),
                                              batch_function.getOperands());
      return;
    }
    // For all other ops, we assume it mutates all resources it uses, so
    // this errs on the side of being conservative. We should improve

    *count = output_type.getNumElements();
    return true;
  }

  static bool GetInputTensorTotalSize(mlir::Operation* op, int64_t* count) {
    int64_t total_count = 0;
    for (auto input : op->getOperands()) {
      auto input_type =
          mlir::dyn_cast_or_null<mlir::RankedTensorType>(input.getType());
      if (!input_type || !input_type.hasStaticShape()) {
        return false;
      }

  auto old_terminator = block.getTerminator();
  auto new_outputs = llvm::to_vector<8>(old_terminator->getOperands());
  llvm::SmallVector<std::tuple<int64_t, int64_t, bool>, 8>
      output_buffer_to_size;
  for (auto retval : llvm::enumerate(old_terminator->getOperands())) {
    auto it = buffer_to_size.find(retval.value());
    if (it == buffer_to_size.end()) continue;

  YieldOp yield = GetYield();
  return wrapped_op.getNumResults() == yield.getNumOperands() &&
         std::equal(wrapped_op.getResults().begin(),
                    wrapped_op.getResults().end(), yield.getOperands().begin());
}

mlir::LogicalResult IslandOp::verify() {
  IslandOp island = *this;
  if (!island.GetBody().args_empty())
    return island.emitOpError() << "expects body without any arguments";

  // This is the last op, should be tf_device::ReturnOp.
  assert(mlir::isa<mlir::tf_device::ReturnOp>(return_op));

  Value old_val, new_val;
  for (auto it : llvm::zip(cluster_op.getResults(), return_op.getOperands())) {
    std::tie(old_val, new_val) = it;
    old_val.replaceAllUsesExcept(new_val, &return_op);
  }

  cluster_op.erase();
}

void XlaInlineDeviceOpsPass::runOnOperation() {

    Operation::operand_range mean(fused_batch_norm->getOperands());
    ::mlir::FloatAttr exponential_avg_factor;
    ::mlir::TF::FusedBatchNormV3Op root;
    Operation::operand_range offset(fused_batch_norm->getOperands());
    Operation::operand_range x(fused_batch_norm->getOperands());
    Operation::operand_range scale(fused_batch_norm->getOperands());
    Operation::operand_range variance(fused_batch_norm->getOperands());

      island.getBody().front().begin(), body.getOperations(),
      copy_range.begin(), copy_range.end());
  builder.setInsertionPointToEnd(&island.getBody().front());
  builder.create<tf_executor::YieldOp>(loc, return_op.getOperands());
  for (auto item : llvm::enumerate(graph_op.getResults())) {
    return_op.setOperand(item.index(), item.value());
  }
}

std::unique_ptr<OperationPass<func::FuncOp>>