[c++] Are the days of passing const std::string & as a parameter over?

I heard a recent talk by Herb Sutter who suggested that the reasons to pass std::vector and std::string by const & are largely gone. He suggested that writing a function such as the following is now preferable:

std::string do_something ( std::string inval )
{
   std::string return_val;
   // ... do stuff ...
   return return_val;
}

I understand that the return_val will be an rvalue at the point the function returns and can therefore be returned using move semantics, which are very cheap. However, inval is still much larger than the size of a reference (which is usually implemented as a pointer). This is because a std::string has various components including a pointer into the heap and a member char[] for short string optimization. So it seems to me that passing by reference is still a good idea.

Can anyone explain why Herb might have said this?

std::string is not Plain Old Data(POD), and its raw size is not the most relevant thing ever. For example, if you pass in a string which is above the length of SSO and allocated on the heap, I would expect the copy constructor to not copy the SSO storage.

The reason this is recommended is because inval is constructed from the argument expression, and thus is always moved or copied as appropriate- there is no performance loss, assuming that you need ownership of the argument. If you don't, a const reference could still be the better way to go.

The problem is that "const" is a non-granular qualifier. What is usually meant by "const string ref" is "don't modify this string", not "don't modify the reference count". There is simply no way, in C++, to say which members are "const". They either all are, or none of them are.

In order to hack around this language issue, STL could allow "C()" in your example to make a move-semantic copy anyway, and dutifully ignore the "const" with regard to the reference count (mutable). As long as it was well-specified, this would be fine.

Since STL doesn't, I have a version of a string that const_casts<> away the reference counter (no way to retroactively make something mutable in a class hierarchy), and - lo and behold - you can freely pass cmstring's as const references, and make copies of them in deep functions, all day long, with no leaks or issues.

Since C++ offers no "derived class const granularity" here, writing up a good specification and making a shiny new "const movable string" (cmstring) object is the best solution I've seen.

IMO using the C++ reference for std::string is a quick and short local optimization, while using passing by value could be (or not) a better global optimization.

So the answer is: it depends on circumstances:

1. If you write all the code from the outside to the inside functions, you know what the code does, you can use the reference const std::string &.
2. If you write the library code or use heavily library code where strings are passed, you likely gain more in global sense by trusting std::string copy constructor behavior.

Almost.

In C++17, we have basic_string_view<?>, which brings us down to basically one narrow use case for std::string const& parameters.

The existence of move semantics has eliminated one use case for std::string const& -- if you are planning on storing the parameter, taking a std::string by value is more optimal, as you can move out of the parameter.

If someone called your function with a raw C "string" this means only one std::string buffer is ever allocated, as opposed to two in the std::string const& case.

However, if you don't intend to make a copy, taking by std::string const& is still useful in C++14.

With std::string_view, so long as you aren't passing said string to an API that expects C-style '\0'-terminated character buffers, you can more efficiently get std::string like functionality without risking any allocation. A raw C string can even be turned into a std::string_view without any allocation or character copying.

At that point, the use for std::string const& is when you aren't copying the data wholesale, and are going to pass it on to a C-style API that expects a null terminated buffer, and you need the higher level string functions that std::string provides. In practice, this is a rare set of requirements.

As @JDlugosz points out in the comments, Herb gives other advice in another (later?) talk, see roughly from here: https://youtu.be/xnqTKD8uD64?t=54m50s.

His advice boils down to only using value parameters for a function f that takes so-called sink arguments, assuming you will move construct from these sink arguments.

This general approach only adds the overhead of a move constructor for both lvalue and rvalue arguments compared to an optimal implementation of f tailored to lvalue and rvalue arguments respectively. To see why this is the case, suppose f takes a value parameter, where T is some copy and move constructible type:

void f(T x) {
  T y{std::move(x)};
}

Calling f with an lvalue argument will result in a copy constructor being called to construct x, and a move constructor being called to construct y. On the other hand, calling f with an rvalue argument will cause a move constructor to be called to construct x, and another move constructor to be called to construct y.

In general, the optimal implementation of f for lvalue arguments is as follows:

void f(const T& x) {
  T y{x};
}

In this case, only one copy constructor is called to construct y. The optimal implementation of f for rvalue arguments is, again in general, as follows:

void f(T&& x) {
  T y{std::move(x)};
}

In this case, only one move constructor is called to construct y.

So a sensible compromise is to take a value parameter and have one extra move constructor call for either lvalue or rvalue arguments with respect to the optimal implementation, which is also the advice given in Herb's talk.

As @JDlugosz pointed out in the comments, passing by value only makes sense for functions that will construct some object from the sink argument. When you have a function f that copies its argument, the pass-by-value approach will have more overhead than a general pass-by-const-reference approach. The pass-by-value approach for a function f that retains a copy of its parameter will have the form:

void f(T x) {
  T y{...};
  ...
  y = std::move(x);
}

In this case, there is a copy construction and a move assignment for an lvalue argument, and a move construction and move assignment for an rvalue argument. The most optimal case for an lvalue argument is:

void f(const T& x) {
  T y{...};
  ...
  y = x;
}

This boils down to an assignment only, which is potentially much cheaper than the copy constructor plus move assignment required for the pass-by-value approach. The reason for this is that the assignment might reuse existing allocated memory in y, and therefore prevent (de)allocations, whereas the copy constructor will usually allocate memory.

For an rvalue argument the most optimal implementation for f that retains a copy has the form:

void f(T&& x) {
  T y{...};
  ...
  y = std::move(x);
}

So, only a move assignment in this case. Passing an rvalue to the version of f that takes a const reference only costs an assignment instead of a move assignment. So relatively speaking, the version of f taking a const reference in this case as the general implementation is preferable.

So in general, for the most optimal implementation, you will need to overload or do some kind of perfect forwarding as shown in the talk. The drawback is a combinatorial explosion in the number of overloads required, depending on the number of parameters for f in case you opt to overload on the value category of the argument. Perfect forwarding has the drawback that f becomes a template function, which prevents making it virtual, and results in significantly more complex code if you want to get it 100% right (see the talk for the gory details).

Short answer: NO! Long answer:

• If you won't modify the string (treat is as read-only), pass it as const ref&.
  (the const ref& obviously needs to stay within scope while the function that uses it executes)
• If you plan to modify it or you know it will get out of scope (threads), pass it as a value, don't copy the const ref& inside your function body.

There was a post on cpp-next.com called "Want speed, pass by value!". The TL;DR:

Guideline: Don’t copy your function arguments. Instead, pass them by value and let the compiler do the copying.

TRANSLATION of ^

Don’t copy your function arguments --- means: if you plan to modify the argument value by copying it to an internal variable, just use a value argument instead.

So, don't do this:

std::string function(const std::string& aString){
    auto vString(aString);
    vString.clear();
    return vString;
}

do this:

std::string function(std::string aString){
    aString.clear();
    return aString;
}

When you need to modify the argument value in your function body.

You just need to be aware how you plan to use the argument in the function body. Read-only or NOT... and if it sticks within scope.

Herb Sutter is still on record, along with Bjarne Stroustroup, in recommending const std::string& as a parameter type; see https://github.com/isocpp/CppCoreGuidelines/blob/master/CppCoreGuidelines.md#Rf-in .

There is a pitfall not mentioned in any of the other answers here: if you pass a string literal to a const std::string& parameter, it will pass a reference to a temporary string, created on-the-fly to hold the characters of the literal. If you then save that reference, it will be invalid once the temporary string is deallocated. To be safe, you must save a copy, not the reference. The problem stems from the fact that string literals are const char[N] types, requiring promotion to std::string.

The code below illustrates the pitfall and the workaround, along with a minor efficiency option -- overloading with a const char* method, as described at Is there a way to pass a string literal as reference in C++.

(Note: Sutter & Stroustroup advise that if you keep a copy of the string, also provide an overloaded function with a && parameter and std::move() it.)

#include <string>
#include <iostream>
class WidgetBadRef {
public:
    WidgetBadRef(const std::string& s) : myStrRef(s)  // copy the reference...
    {}

    const std::string& myStrRef;    // might be a reference to a temporary (oops!)
};

class WidgetSafeCopy {
public:
    WidgetSafeCopy(const std::string& s) : myStrCopy(s)
            // constructor for string references; copy the string
    {std::cout << "const std::string& constructor\n";}

    WidgetSafeCopy(const char* cs) : myStrCopy(cs)
            // constructor for string literals (and char arrays);
            // for minor efficiency only;
            // create the std::string directly from the chars
    {std::cout << "const char * constructor\n";}

    const std::string myStrCopy;    // save a copy, not a reference!
};

int main() {
    WidgetBadRef w1("First string");
    WidgetSafeCopy w2("Second string"); // uses the const char* constructor, no temp string
    WidgetSafeCopy w3(w2.myStrCopy);    // uses the String reference constructor
    std::cout << w1.myStrRef << "\n";   // garbage out
    std::cout << w2.myStrCopy << "\n";  // OK
    std::cout << w3.myStrCopy << "\n";  // OK
}

OUTPUT:

const char * constructor
const std::string& constructor

Second string
Second string

Unless you actually need a copy it's still reasonable to take const &. For example:

bool isprint(std::string const &s) {
    return all_of(begin(s),end(s),(bool(*)(char))isprint);
}

If you change this to take the string by value then you'll end up moving or copying the parameter, and there's no need for that. Not only is copy/move likely more expensive, but it also introduces a new potential failure; the copy/move could throw an exception (e.g., allocation during copy could fail) whereas taking a reference to an existing value can't.

If you do need a copy then passing and returning by value is usually (always?) the best option. In fact I generally wouldn't worry about it in C++03 unless you find that extra copies actually causes a performance problem. Copy elision seems pretty reliable on modern compilers. I think people's skepticism and insistence that you have to check your table of compiler support for RVO is mostly obsolete nowadays.

In short, C++11 doesn't really change anything in this regard except for people that didn't trust copy elision.

I've copy/pasted the answer from this question here, and changed the names and spelling to fit this question.

Here is code to measure what is being asked:

#include <iostream>

struct string
{
    string() {}
    string(const string&) {std::cout << "string(const string&)\n";}
    string& operator=(const string&) {std::cout << "string& operator=(const string&)\n";return *this;}
#if (__has_feature(cxx_rvalue_references))
    string(string&&) {std::cout << "string(string&&)\n";}
    string& operator=(string&&) {std::cout << "string& operator=(string&&)\n";return *this;}
#endif

};

#if PROCESS == 1

string
do_something(string inval)
{
    // do stuff
    return inval;
}

#elif PROCESS == 2

string
do_something(const string& inval)
{
    string return_val = inval;
    // do stuff
    return return_val; 
}

#if (__has_feature(cxx_rvalue_references))

string
do_something(string&& inval)
{
    // do stuff
    return std::move(inval);
}

#endif

#endif

string source() {return string();}

int main()
{
    std::cout << "do_something with lvalue:\n\n";
    string x;
    string t = do_something(x);
#if (__has_feature(cxx_rvalue_references))
    std::cout << "\ndo_something with xvalue:\n\n";
    string u = do_something(std::move(x));
#endif
    std::cout << "\ndo_something with prvalue:\n\n";
    string v = do_something(source());
}

For me this outputs:

$ clang++ -std=c++11 -stdlib=libc++ -DPROCESS=1 test.cpp
$ a.out
do_something with lvalue:

string(const string&)
string(string&&)

do_something with xvalue:

string(string&&)
string(string&&)

do_something with prvalue:

string(string&&)
$ clang++ -std=c++11 -stdlib=libc++ -DPROCESS=2 test.cpp
$ a.out
do_something with lvalue:

string(const string&)

do_something with xvalue:

string(string&&)

do_something with prvalue:

string(string&&)

The table below summarizes my results (using clang -std=c++11). The first number is the number of copy constructions and the second number is the number of move constructions:

+----+--------+--------+---------+
|    | lvalue | xvalue | prvalue |
+----+--------+--------+---------+
| p1 |  1/1   |  0/2   |   0/1   |
+----+--------+--------+---------+
| p2 |  1/0   |  0/1   |   0/1   |
+----+--------+--------+---------+

The pass-by-value solution requires only one overload but costs an extra move construction when passing lvalues and xvalues. This may or may not be acceptable for any given situation. Both solutions have advantages and disadvantages.

See “Herb Sutter "Back to the Basics! Essentials of Modern C++ Style”. Among other topics, he reviews the parameter passing advice that’s been given in the past, and new ideas that come in with C++11 and specifically looks at the idea of passing strings by value.

The benchmarks show that passing std::strings by value, in cases where the function will copy it in anyway, can be significantly slower!

This is because you are forcing it to always make a full copy (and then move into place), while the const& version will update the old string which may reuse the already-allocated buffer.

See his slide 27: For “set” functions, option 1 is the same as it always was. Option 2 adds an overload for rvalue reference, but this gives a combinatorial explosion if there are multiple parameters.

It is only for “sink” parameters where a string must be created (not have its existing value changed) that the pass-by-value trick is valid. That is, constructors in which the parameter directly initializes the member of the matching type.

If you want to see how deep you can go in worrying about this, watch Nicolai Josuttis’s presentation and good luck with that (“Perfect — Done!” n times after finding fault with the previous version. Ever been there?)

This is also summarized as ?F.15 in the Standard Guidelines.

This highly depends on the compiler's implementation.

However, it also depends on what you use.

Lets consider next functions :

bool foo1( const std::string v )
{
  return v.empty();
}
bool foo2( const std::string & v )
{
  return v.empty();
}

These functions are implemented in a separate compilation unit in order to avoid inlining. Then :
1. If you pass a literal to these two functions, you will not see much difference in performances. In both cases, a string object has to be created
2. If you pass another std::string object, foo2 will outperform foo1, because foo1 will do a deep copy.

On my PC, using g++ 4.6.1, I got these results :

• variable by reference: 1000000000 iterations -> time elapsed: 2.25912 sec
• variable by value: 1000000000 iterations -> time elapsed: 27.2259 sec
• literal by reference: 100000000 iterations -> time elapsed: 9.10319 sec
• literal by value: 100000000 iterations -> time elapsed: 8.62659 sec

Are the days of passing const std::string & as a parameter over?

No. Many people take this advice (including Dave Abrahams) beyond the domain it applies to, and simplify it to apply to all std::string parameters -- Always passing std::string by value is not a "best practice" for any and all arbitrary parameters and applications because the optimizations these talks/articles focus on apply only to a restricted set of cases.

If you're returning a value, mutating the parameter, or taking the value, then passing by value could save expensive copying and offer syntactical convenience.

As ever, passing by const reference saves much copying when you don't need a copy.

Now to the specific example:

However inval is still quite a lot larger than the size of a reference (which is usually implemented as a pointer). This is because a std::string has various components including a pointer into the heap and a member char[] for short string optimization. So it seems to me that passing by reference is still a good idea. Can anyone explain why Herb might have said this?

If stack size is a concern (and assuming this is not inlined/optimized), return_val + inval > return_val -- IOW, peak stack usage can be reduced by passing by value here (note: oversimplification of ABIs). Meanwhile, passing by const reference can disable the optimizations. The primary reason here is not to avoid stack growth, but to ensure the optimization can be performed where it is applicable.

The days of passing by const reference aren't over -- the rules just more complicated than they once were. If performance is important, you'll be wise to consider how you pass these types, based on the details you use in your implementations.