// function, which might not exist in case of multi-signature graphs. In that
// case, this pass will create a new main function, which calls signature
// functions.
//
// An already existing @main function will be renamed by attaching a numeric
// suffix like `@main_0` to avoid conflict with the newly created main function.
class InsertMainFunctionPass
    : public PassWrapper<InsertMainFunctionPass, OperationPass<ModuleOp>> {

    return builder_.create<mlir::tf_executor::SwitchOp>(loc, types, operands,
                                                        result.attributes);
  }
  if (node.IsMerge()) {
    return builder_.create<mlir::tf_executor::MergeOp>(loc, types, operands,
                                                       result.attributes);
  }
  if (node.IsNextIteration()) {

    // Finally, use the quantization parameter to create the quantize and
    // dequantize ops, and insert them between the tf.FakeQuantWithMinMaxVarsOp
    // and its users.
    Value value = tf_op.getOutputs();
    auto quantize = rewriter.create<quantfork::QuantizeCastOp>(
        tf_op.getLoc(), qtype.getValue(), value);
    auto dequantize = rewriter.create<quantfork::DequantizeCastOp>(

    auto new_func_type = builder.getFunctionType(block.getArgumentTypes(),
                                                 terminator->getOperandTypes());
    func.setType(new_func_type);
  });
}

// Creates an instance of the pass to decompose the TF ops.
std::unique_ptr<OperationPass<ModuleOp>> CreateRewriteQuantizedIOPass() {
  return std::make_unique<RewriteQuantizedIOPass>();
}

      ToNodeProperties(R"pb(
        name: 'XTimesY' op: 'XTimesY' input: 'a' input: 'b'
      )pb");
  (*(callsite->node_def.mutable_attr()))["_XlaMustCompile"] = BoolAttr(true);

  // Note: need to set attribute on the created node.
  Status status = xla_kernel_creator.CreateKernel(flr_, callsite, &kernel_);
  ASSERT_TRUE(status.ok()) << status.ToString();

  EXPECT_EQ("XTimesY", kernel_->name());

// Test to ensure that a registered graph optimization pass is only executed
// once (i.e., on the main function side) in running distributed functions.
// This test creates a cluster with two workers, create a variable on the
// second worker, and run a distributed function (VariableAddFunction) whose ops
// span the local and remote workers. If the graph optimization pass is executed

  auto graph_op = builder.create<tf_executor::GraphOp>(
      loc, func.getFunctionType().getResults());
  graph_op.getBody().push_back(new Block);
  builder.setInsertionPointToEnd(&graph_op.GetBody());
  auto island = builder.create<tf_executor::IslandOp>(
      loc, func.getFunctionType().getResults(),
      tf_executor::ControlType::get(&getContext()), ArrayRef<Value>());
  // Create Fetch.

    llvm::SmallVector<Value, 2> new_operands;
    FilterOutBlockArgControlDep(op->getOperands().drop_front(), new_operands);

    auto source_op = builder.create<tf_executor::NextIterationSourceOp>(
        op->getLoc(), op->getOperand(0).getType());
    builder.create<tf_executor::NextIterationSinkOp>(
        op->getLoc(), source_op.getToken(), /*input=*/op->getOperand(0),
        /*controlInputs=*/new_operands);

  // Create SessionInitializerOp; should reference main function.
  StringAttr func_name = main_func.getSymNameAttr();
  llvm::SmallVector<mlir::Attribute, 2> func_names = {
      mlir::SymbolRefAttr::get(builder.getContext(), func_name)};
  builder.create<tf_saved_model::SessionInitializerOp>(
      module->getLoc(), builder.getArrayAttr(func_names));
  // Create AssetOp; this holds the checkpoint_path.

// control flow op into an i1 value.
static Value LowerCondition(Location loc, Value value, OpBuilder* builder) {
  Value zero_d = builder->create<ToBoolOp>(loc, value);
  Value scalar = builder->create<tensor::ExtractOp>(loc, zero_d);
  return scalar;
}

// Calls the function `fn` with arguments provided by the given function and