*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;
      }

        composite->getLoc(), composite->getResultTypes(),
        composite->getOperands().slice(2, 3), composite.getName(),
        CustomOption(&builder, custom_option_buffer));
  }
  return builder.create<TFL::CustomOp>(
      composite->getLoc(), composite->getResultTypes(),
      composite->getOperands(), composite.getName(),
      CustomOption(&builder, custom_option_buffer));
}

}  // namespace

// 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());

        op.getLoc(), op.getType(), op.getOperands(),
        builder.getI64ArrayAttr(partition_dims), builder.getBoolAttr(false),
        op.get_XlaShardingAttr());
    op->replaceAllUsesWith(pi);
  } else {
    auto po = builder.create<TF::TPUPartitionedOutputV2Op>(
        op.getLoc(), op.getResultTypes(), op.getOperand(),
        builder.getI64ArrayAttr(partition_dims), op.get_XlaShardingAttr());

  llvm::SmallVector<Operation*> predecessors;
  for (Value operand : op->getOperands()) {
    if (Operation* pred = operand.getDefiningOp()) {
      int result_number = llvm::cast<OpResult>(operand).getResultNumber();
      if (auto pred_island = llvm::dyn_cast<tf_executor::IslandOp>(pred)) {
        Value yield_operand = pred_island.GetYield().getOperand(result_number);
        predecessors.push_back(yield_operand.getDefiningOp());
      }

      rewriter.setInsertionPoint(current_terminator);
      SmallVector<Value, 4> scf_yield_input;
      for (auto it : llvm::zip(tf_if_region_return_type,
                               current_terminator->getOperands())) {
        scf_yield_input.push_back(rewriter.create<CastOp>(
            current_terminator->getLoc(), std::get<0>(it), std::get<1>(it)));
      }

      rewriter.replaceOpWithNewOp<scf::YieldOp>(current_terminator,

      // 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

    auto const_type = mlir::dyn_cast<ShapedType>(op.getOperand(0).getType());
    auto result_type = mlir::dyn_cast<ShapedType>(op.getResult().getType());
    if (!const_type || !const_type.getElementType().isSignlessInteger(4) ||
        !result_type || !result_type.getElementType().isSignlessInteger(8)) {
      return failure();
    }

    auto input_op =
        dyn_cast<arith::ConstantOp>(op.getOperand(0).getDefiningOp());
    if (!input_op) {

  // 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() {

      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>>