54

I have a question about the return value of operator overloading in C++. Generally, I found two cases, one is return-by-value, and one is return-by-reference. So what's the underneath rule of that? Especially at the case when you can use the operator continuously, such as cout<<x<<y.

For example, when implementing a + operation "string + (string)". how would you return the return value, by ref or by val.

0

5 Answers 5

81

Some operators return by value, some by reference. In general, an operator whose result is a new value (such as +, -, etc) must return the new value by value, and an operator whose result is an existing value, but modified (such as <<, >>, +=, -=, etc), should return a reference to the modified value.

For example, cout is a std::ostream, and inserting data into the stream is a modifying operation, so to implement the << operator to insert into an ostream, the operator is defined like this:

std::ostream& operator<< (std::ostream& lhs, const MyType& rhs)
{
  // Do whatever to put the contents of the rhs object into the lhs stream
  return lhs;
}

This way, when you have a compound statement like cout << x << y, the sub-expression cout << x is evaluated first, and then the expression [result of cout << x ] << y is evaluated. Since the operator << on x returns a reference to cout, the expression [result of cout << x ] << y is equivalent to cout << y, as expected.

Conversely, for "string + string", the result is a new string (both original strings are unchanged), so it must return by value (otherwise you would be returning a reference to a temporary, which is undefined behavior).

Sign up to request clarification or add additional context in comments.

14 Comments

Which goes hand in hand with "if it's const, return a value, if it's non-const, return a reference".
Technically << does not modify the existing value. It's just a left-shift operator. It's the stream interface breaking the rules.
@GMan correct, although that doesn't cover the cases (such as operator<<) where the overload is implemented as a non-member.
@KennyTM I had no idea old C codgers were still putting up that fight after so many years. :)
@nocomprende the stream object isn't a console or file, it's a C++ construct that connects to a console or file (or other things e.g. stringstream). Its "value" is very much implementation dependent, and is not the same as the state of the resource that it connects to.
|
16

To attempt an answer to your question regarding strings, the operator+() for strings is almost always implemented as a free (non-member) function so that implicit conversions can be performed on either parameter. That is so you can say things like:

string s1 = "bar";
string s2 = "foo" + s1;

Given that, and that we can see that neither parameter can be changed, it must be declared as:

RETURN_TYPE operator +( const string & a, const string & b );

We ignore the RETURN_TYPE for the moment. As we cannot return either parameter (because we can't change them), the implementation must create a new, concatenated value:

RETURN_TYPE operator +( const string & a, const string & b ) {
    string newval = a;
    newval += b;    // a common implementation
    return newval;
}

Now if we make RETURN_TYPE a reference, we will be returning a reference to a local object, which is a well-known no-no as the local object don't exist outside the function. So our only choice is to return a value, i.e. a copy:

string operator +( const string & a, const string & b ) {
    string newval = a;
    newval += b;    // a common implementation
    return newval;
}

2 Comments

So, for freeing the allocated result, am I correct that because the invocation of the Operator is in effect a declaration, the anonymous return value goes on the stack, and so is freed when the containing function returns?
@anon: In this case, it is more concise and possibly better performing (in the case where a is constructed by the call) to simply take a by value and modify/return that, rather than needing any other intermediate string. Generally, if you need to copy an argument, take it by value: that's getting a copy for free. So those 3 lines are simplified to return a += b;. @user: The mistake there was thinking about "freeing" and "allocation" at all. The whole point of RAII and classes like std::string is to avoid the end user having to worry about freeing. Just use the result and let RAII KO it
7

If you want your operator overload to behave like the built-in operator, then the rule is pretty simple; the standard defines exactly how the built-in operators behave and will indicate if the result of a built-in is an rvalue or an lvalue.

The rule you should use is:

  • if the built-in operator returns an rvalue then your overload should return a reference
  • if the built-in returns an lvalue then your overload should return a value

However, your overload isn't required to return the same kind of result as the built-in, though that's what you should do unless you have a good reason to do otherwise.

For example, KennyTM noted in a comment to another answer that the stream overloads for the << and >> operators return a reference to the left operand, which is not how the built-ins work. But the designers of the stream interface did this so stream I/O could be chained.

2 Comments

For reference the built-in operators are in § 13.6 of the N3337 draft C++11 standard, which is the only version of the standard I have at hand.
UV for Do as the ints Do!
4

Depending on the operator you may have to return by value.

When both can be used though, like in operator+= you could consider the following:

  • If your objects are immutable it's probably better to return by value.
  • If your objects are mutable it's probably better to return by reference.

1 Comment

The purpose of operator+= is to mutate the object. It should not be called on immutable objects.
3

Usually you return by reference in an operation that changes the value of the things it's operating on, like = or +=. All other operations are return by value.

This is more a rule of thumb, though. You can design your operator either way.

1 Comment

It's a rule of thumb because in most cases, a user will expect - & quite possibly need - overloaded operators in user-defined classes to have the same semantics as the counterparts for built-in types. "Do as the ints do" is generally the best advice. Obviously, this isn't always 100% feasible (although my mind has just gone blank of examples), or sometimes you're doing something wild like ostream.operator<< or path.operator/... but in most cases, the int-style behaviour is both possible & desirable. You "can" do "either way" - but chances are you'll just annoy users if you flip them!

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.