StringRef getDescription() const final {
    return "Renames the entrypoint in SavedModel to `main`";
  }

  void runOnOperation() override {
    auto fail = [&](Operation* op, std::string message) {
      op->emitError(message);
      signalPassFailure();
    };

    DenseMap<StringRef, func::FuncOp> entrypoints;
    auto module = getOperation();
    module.walk([&](func::FuncOp op) {

  if (keyword == "future") return FutureType::get(getContext());
  if (keyword == "promise") return PromiseType::get(getContext());
  if (keyword == "async_handle") return AsyncHandleType::get(getContext());

  parser.emitError(parser.getNameLoc(), "unknown type: ") << keyword;
  return mlir::Type();
}

// Print a type registered to this dialect.
void MlrtDialect::printType(mlir::Type type,

    }
    for (auto operand_type : op->getOperandTypes()) {
      if (!mlir::tf_type::HasCompatibleElementTypes(
              operand_type, type, /*may_ignore_ref_type_lhs=*/true))
        return op->emitError() << "requires all operands and results to have "
                                  "compatible element types";
    }
    return success();
  }
};

namespace detail {

    mlir::DialectAsmParser &parser) const {
  llvm::StringRef keyword;
  if (parser.parseKeyword(&keyword)) return mlir::Type();

  if (keyword == "tensor") return TFTensorType::get(getContext());

  parser.emitError(parser.getNameLoc(), "unknown type: ") << keyword;
  return mlir::Type();
}

// Print a type registered to this dialect.
void TensorflowMlrtDialect::printType(mlir::Type type,

  ModuleOp module_op = getOperation();

  for (auto init_func_op : GetInitializerFunctions(module_op)) {
    if (!init_func_op) return;

    if (init_func_op.getBlocks().size() != 1) {
      init_func_op.emitError("expects exactly one block in the MLIR function");
      return signalPassFailure();
    }

    auto var_handle_ops =
        init_func_op.getBlocks().front().getOps<VarHandleOp>();

  auto input_shapes_attr =
      func_op->getAttrOfType<ArrayAttr>(kTfInputShapesAttr);
  if (!input_shapes_attr) return success();

  if (input_shapes_attr.size() != func_op.getNumArguments()) {
    return func_op->emitError(
               "Number of arguments and 'tf._input_shapes' "
               "attribute size do not match. ")
           << "Num args: " << func_op.getNumArguments()

  int32_t container_size =
      get_resource_op->getAttrOfType<mlir::ArrayAttr>("container").size();

  if (!(indices_size == shared_name_size &&
        shared_name_size == container_size)) {
    return get_resource_op->emitError()
           << "length of attribute arrays do not match. indices = "
           << indices_size << ", shared_name = " << shared_name_size
           << ", container = " << container_size;
  }

  return success();
}

// failure.
template <typename OpT>
LogicalResult GetOp(Region* region, OpT* result) {
  *result = {};
  for (auto op : region->getOps<OpT>()) {
    if (*result) return op.emitError("should be unique within a function");
    *result = op;
  }
  return success();
}

LogicalResult RunOnRegion(Region* region) {
  RecvTPUEmbeddingActivationsOp recv_op;

  patterns.add<RemoveTpuOp>(ctx);
  patterns.add<RemoveIdentity>(ctx);

  if (failed(applyPatternsAndFoldGreedily(module_op, std::move(patterns)))) {
    module_op.emitError() << "quant-convert-tpu-model-to-cpu pattern "
                             "conversion did not converge.";
    signalPassFailure();
    return;
  }
}

}  // namespace

  RewritePatternSet patterns(ctx);
  func::FuncOp func = getOperation();

  patterns.add<ConvertCustomAggregationOpToQuantStats>(ctx);
  if (failed(applyPatternsAndFoldGreedily(func, std::move(patterns)))) {
    func.emitError()
        << "quant-convert-tf-custom-aggregator-op-to-quant-stats failed.";
    signalPassFailure();
  }
}

}  // namespace

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