SymbolTable symbol_table(call_op->getParentOfType<ModuleOp>());
    auto func_op =
        dyn_cast_or_null<func::FuncOp>(symbol_table.lookup(func_name));
    if (!func_op) return failure();
    // The quantized fusion should have requantize and return ops at the end.
    auto return_op = dyn_cast_or_null<func::ReturnOp>(
        func_op.getRegion().getBlocks().front().getTerminator());
    if (!return_op) return failure();

  llvm::SmallVector<Type> arg_types, result_types;

  for (auto func_op : module_op.getOps<func::FuncOp>()) {
    if (!ShouldIncludeInMainFunction(func_op)) continue;

    arg_types.append(func_op.getArgumentTypes().begin(),
                     func_op.getArgumentTypes().end());
    auto& return_op = func_op.getBody().getBlocks().front().back();
    result_types.append(return_op.getOperandTypes().begin(),

};

// Returns true iff func_op has either no Region or the body has no Blocks.
bool IsFuncOpEmpty(func::FuncOp func_op) {
  return func_op->getNumRegions() == 0 || func_op.getBody().empty();
}

// Gets the GraphOp from the function op. Returns an empty op iff it doesn't
// exist.
GraphOp GetGraphOpFromFuncOp(func::FuncOp func_op) {
  if (IsFuncOpEmpty(func_op)) return {};

// tf_device.replicate.
Operation* FindCompilePredecessor(Operation* func_op,
                                  Operation* preprocess_op) {
  llvm::DenseSet<Block*> blocks = GetAllBlocksBetween(preprocess_op, func_op);

  llvm::DenseMap<Block*, Operation*> scope;
  Operation* o = preprocess_op;
  while (o && o != func_op) {
    scope[o->getBlock()] = o;
    o = o->getParentOp();
  }

  Value input = func_op.getArgument(0);
  Value output_init_state = func_op.getArgument(1);
  Value hidden_init_state = func_op.getArgument(2);
  Value weight_kernel = func_op.getArgument(3);
  Value recurrent_kernel = func_op.getArgument(4);
  Value bias = func_op.getArgument(5);

  // The func op should have 5 outputs.
  if (func_op.getNumResults() != 5) return failure();

  if (call_op_has_platform_index_arg) {
    arg_mapper.map(func_op.getArgument(0),
                   builder.create<mhlo::ConstantOp>(
                       func_op.getLoc(), builder.getI16IntegerAttr(0)));
  }
  for (auto [func_arg, operand] : llvm::zip_equal(
           func_op.getArguments().take_back(call_op.getNumOperands()),
           call_op.getOperands())) {

};

void RemoveShardingCustomCallPass::runOnOperation() {
  func::FuncOp func_op = getOperation();
  MLIRContext& ctx = getContext();

  RewritePatternSet patterns(&ctx);
  populateWithGenerated(patterns);

  FrozenRewritePatternSet frozen_patterns(std::move(patterns));
  if (failed(applyPatternsAndFoldGreedily(func_op, frozen_patterns))) {
    func_op.emitWarning() << "Failed to converge "

};

// Folds tf.IdentityOp and tf.IdentityNOp if op device and the argument devices
// from the defining ops match.
void TensorDeviceCopyConversionPass::runOnOperation() {
  func::FuncOp func_op = getOperation();

  auto should_fold_op_func = [&func_op](const Value &arg,
                                        const StringAttr &op_device) {
    // In TFRT TPU, tensor transfer is handled specifically by D2H and

void ConvertNMSPaddedFunc::RewriteFunc() {
  func_->setAttr(kTFImplements,
                 StringAttr::get(func_.getContext(), kTfNMSPadded));
  Value boxes = func_.getArgument(0);
  Value scores = func_.getArgument(1);
  Value max_output_size = func_.getArgument(2);
  Value iou_threshold = func_.getArgument(3);
  Value score_threshold = func_.getArgument(4);
  auto output_type0 = func_.getFunctionType().getResult(0);

    // Builder builder(func_op.getContext());
    func_op->setAttr(
        kTfInputShapesAttr,
        ArrayAttr::get(func_op.getContext(), updated_input_shapes_attr));
  }
}

// Validates func ops. Returns `failure` if the function is invalid.
LogicalResult ValidateFuncOp(func::FuncOp func_op) {
  auto input_shapes_attr =
      func_op->getAttrOfType<ArrayAttr>(kTfInputShapesAttr);