C++ Signatures: It’s Not Just About Performance, It’s About Ownership

In Java and Python, every object is tracked by a pointer, but the language hides it to make devs’ lives easier. These languages always “pass by value”, but since every object is a reference, at the end of the day, the observed behaviour feels like passing “by reference”.

If you want to pass something “by value”, you need to actively build a deep copy, or call a method that creates a deep copy internally. For these languages, which usually care more about usability than performance, passing “by reference” is a good generic way to do things.

You may be wondering: but passing “by reference” is always better in C++ too! You literally avoid creating a new object. Why would someone need to use a different approach? Well, the devil lies in the details.

Java, Python, and many other languages hide ownership of variables. I mean, if the ownership of an object gets hard to track, who cares? The garbage collector will clean up the mess for us after all!

For people working with C++, or even Rust, languages that delegate ownership management to the developer, passing things by reference can become a nightmare. That’s why things like smart pointers and RAII were created: to make sure that whoever allocated memory will not forget to deallocate it, or worse, deallocate it twice!

This brief article discusses some rules of thumb to help you not lose track of ownership when writing API signatures.

Rules of thumb to decide signatures

Before starting, you have to remember that C++ has and cares about rvalues and lvalues:

• rvalues: Temporary objects that do not have a persistent memory address. E.g., literals and temporary return values.
• lvalues: Objects that have an identifiable location in memory (an address) and a name. E.g., variables with names.

std::println("{}", 1); // prints an rvalue

int x = 2;
std::println("{}", x); // prints an lvalue

When writing a method, you usually don’t know (or even don’t care) if the value passed will be an rvalue or an lvalue.

That’s why these rules of thumb exist: to help you decide which signature to use, without overthinking and while avoiding strange pitfalls or corner cases related to ownership.

As a rule of thumb, there are principles you can follow when developing API signatures.

The main keyword here is OWNERSHIP, so you can divide methods into three main categories:

• owning methods
• non-owning methods
• generic methods

You will see that I’ll break these categories down into five sub-types, but the guiding principle is still ownership.

Owning Methods

Here you receive the argument and own it. In this example, you “steal” (move) the data.

Person temp_person;

void set_name(std::string str){
    temp_person.name = std::move(str);
}

set_name(std::string("Jane Doe"));      // 1

std::string name = "John Doe";
set_name(name);                         // 2

set_name(std::move(name));              // 3

For some people this looks weird, even for me when I first heard about this kind of pattern. You know, always creating a copy? sounds crazy at first. But lets break it down.

The good thing about passing by value is that it often minimizes total move/copy operations without introducing ownership pitfalls.

Let me explain the marks in the code:

1 - When passing an rvalue, the “str” variable (i.e., the param of set_name) is created via move construction, and after that it’s moved again. So there are two moves here.

2 - When passing an lvalue, the “str” variable (i.e., the param of set_name) is created by the copy construction, and after that it’s moved. So there are one move and one copy. Unfortunately, it’s impossible to do it in fewer operations, since the method does not know if it can take the ownership of an lvalue (which may be used in its source).

3 - Additionally, if the user wants to allow the moving of the lvalue, it can use std::move to cast it to an rvalue and benefit of the possible speedup of case 1.

Is it not fast enough for you?

If you know this is in the middle of your hot path, or you have measured it and it is killing your performance, one simple way to fix this is using overloads:

// accepts lvalue references
void set_name(std::string& str){
    temp_person.name = std::move(str);
}

// accepts rvalue references
void set_name(std::string&& str){
    temp_person.name = std::move(str);
}

In this case you would have only one move operation for the calls, no matter whether you pass an rvalue or an lvalue. Yet, this comes with more code to maintain and, worst of all, the semantics of the method change. Every variable passed to set_name would give its ownership to the method. This could be a problem in the following scenario.

Person temp_person;

void set_name(std::string& str){
    temp_person.name = std::move(str);
}

void set_name(std::string&& str){
    temp_person.name = std::move(str);
}

std::string my_precious_string = "password123";
set_name(my_precious_string); // you just lost the ownership of the string

std::println("My password: {}", my_precious_string); // no longer works properly

Thats why passing by copy is usally a good design choice. And if you ever need more sophisticated approaches, you implement it later.

For even more generic approaches, you should look at “Modern Effective C++. The parts talking about Universal References would make you happy, also known as Forwarding References.

Pipe Methods

This part is not based on the “Modern Effective C++”, it is more subjective to my previous experience.

This is a special case of ownership. You can see it is as a “temporary owner”, like borrowing ownership.

You pass by value, modify and then return it:

std::string add_exclamation(std::string str){
    // Taking ownership here would be OK.
    str.push_back('!');
    return str;
}

std::string phrase = "Lorem Ipsum";
phrase = add_exclamation(phrase);

This guarantees that you do not lose ownership of something you didn’t mean to.

Another way to see a “pipe” is as a “non-owner” who has access to a non-owned object. You do this by passing a reference, but be aware to NEVER take ownership:

void add_exclamation(std::string& str){
    // Taking ownership here would be a huge antipattern.
    str.push_back('!');
}

Pass by value has another pro: it is friendly to move-only objects, as std::thread and std::unique_ptr.

std::unique_ptr<int> foo(std::unique_ptr<int> ptr){ 
    // You are the owner of ptr
    // Modifications DO NOT affect the caller.
    return ptr;
}

void bar(std::unique_ptr<int>& ptr){
    // You are NOT the owner of ptr
    // Modifications affect the caller.
}

Not Owning Methods

In this case, you only receive the value, but you do not plan to modify it.

void print_name(const std::string& str){
    std::println("Name: {}", str);
}

print_name(std::string("Jane Doe"));    // 1

std::string name = "John Doe";
print_name(name);                       // 2

Here are the cases:

1 - When passing an rvalue, it binds to a const std::string&, since it is const, the language allows binding rvalue. No copies or moves happen here.

2 - When passing an lvalue, no copies or moves happen either.

The only exception here is when the type you are passing is smaller than or equal to a reference (usually 8 bytes), such as int, char, bool, and many others.

This is still up for discussion, as you might stick with const T& to maintain consistency unless this optimization is truly on a hot path.

Generic Methods

This is a huge topic, but when you want a method that is extremely generic, use forwarding references, auto return types, and perfect forwarding.

But before we start, forwarding references should be reserved for generic wrappers and factory-style utilities. Overusing them in regular APIs often leads to hard-to-debug overload resolution issues.

// forwarding references only work using templates.
template<class T>
decltype(auto) omega_foo(T&& t){ // it creates references that work for rvalues and lvalues.
    // calls another generic method; forward replicates the r-ness or l-ness of the original param
    return omega_bar(std::forward<T>(t)); // the call is usually to a constructor or to a function overloaded on rvalue and lvalue references
}

This pattern is used widely across the STL, but the best example of it is emplace_back.

std::vector<std::pair<int, int>> v;
v.emplace_back(1, 1);           // perfect-forwarded arguments

std::pair<int, int> x = {2, 2};
v.emplace_back(x);              // copy
v.emplace_back(std::move(x));   // move

In both calls it is resolved to the right (usually optimized) constructor.

You could implement some stuff to initialize objects:

template<class T, class U>
auto constructor_wrapper(U&& u){
    // builds and returns an object of type T initialized with u
    return T(std::forward<U>(u)); 
}

It’s also possible to use Forwarding References with variadic parameters.

template<class T, class... Args>
auto constructor_wrapper(Args&&... args){
    // builds and returns an object of type T initialized with whatever was passed.
    return T(std::forward<Args>(args)...);
}

One common problem with this is the fact that it accepts EVERYTHING. So if you do not want to accept some types, in C++20 you can use concepts to do it:

template <typename Arg>
concept ValidArg = !std::is_null_pointer_v<std::remove_cvref_t<Arg>>;

template <typename T, typename... Args>
requires (ValidArg<Args> && ...)
auto constructor_wrapper(Args&&... args) {
    return T(std::forward<Args>(args)...);
}

This implementation does not allow std::nullptr_t.

Attentions to Forwarding References

Some types undergo implicit casting when passed to forwarding references. For instance, passing NULL as an argument will be forwarded as 0, an integer. It also does not behave well with std::initializer_list deduction.

Another point to consider is that in the constructor_wrapper example, the object is initialized using the “parenthesis constructor” T(...). This can diverge from the result of using a “uniform constructor” (i.e., using T{...}), especially when dealing with types like std::vector.

Decision Lookup Table

Here is a quick guide to help you choose the right signature based on the categories we discussed:

Category	Recommended Signature	Why?
Owning Methods (Sink)	void f(T arg) + std::move	Efficiently takes ownership (move or copy).
Pipe Methods	T& arg or T f(T arg)	Modifies in-place or via “copy-modify-return”, depending on ownership intent.
Non-owning Methods (Read)	const T& arg	Observes the object without copying.
Non-owning Methods (Small*)	T arg	Small types are faster to pass by value.
Generic Methods	template<class T> f(T&& arg)	For perfect forwarding in wrappers/factories.

*Small types: int, double, bool, std::string_view, std::span, etc.

Disclaimer

Things are not as easy as they seem in these examples. There are many corner cases, and when you need to optimize things, you should design better signatures. But before designing, measure performance to make sure the changes are actually needed.

To better understand the concepts presented here, as well as their corner cases, read Modern Effective C++.

And ALWAYS remember:

“APIs live longer than optimizations.”