C++ "Best" Parameter Passing Method

Question

I was coding up a C++ class today, and I wrote a function that took an argument as a reference rather than a pointer, something I rarely ever do. I've always passed by pointers. So I was about to change it back, and then I realized - I have no idea if I should, or if it even matters.

So I turn to you guys. I have three ways of passing parameters about:

//1: By pointer
Object* foo(Object* bar) {…}

//2: By reference
Object& foo(Object& bar) {…}

//3: By value (just for completeness)
Object foo(Object bar) {…}

Assuming #3's out for performance reasons (yes, I know compilers have gotten pretty good at this, but still), the other two are more or less equivalent.

So: What's the "best" method? Pointers? References? Some combination of the two? Or does it even matter? Technical reasons are the best, but stylistic reasons are just as good.

Update: I've accepted YeenFei's answer, since it deals with the difference that clinched it for me (even if I then pointedly ignored his advice - I like having NULL as an option...). But everyone made good points - especially GMan (in the comments), and Nemo, in the answer dealing with performance and passing by value. If you're here for answers, check them all!

Answer 1

I would suggest to pass your argument by reference if it is expected to be valid. This would be a by-design optimization and save you from defensive programming.

Reference cannot be null while pointer can. If you are dealing with pointer, you will need to verify whether given pointer is valid (non-null) regardless it is in raw form or wrapped in managed container (shared_ptr), before using them.

Answer 2

So I am going to make the case for choice #3. Consider the following code:

struct Foo {
    int x;
    int y;
};

Foo
add(Foo a, Foo b)
{
    Foo result;
    result.x = a.x + b.x;
    result.y = a.y + b.y;
    return result;
}

Foo
add2(Foo &a, Foo &b)
{
    Foo result;
    result.x = a.x + b.x;
    result.y = a.y + b.y;
    return result;
}

Try examining the generated assembly. Notice how add is almost entirely register operations, nicely scheduled. Notice how add2 is lots of memory accesses without any reordering.

I wrote a main that called each of these functions 10 billion times. Result? add took 22 seconds, while add2 took 26 seconds. Even for this trivial example, that's 10-20% better performance for the pass-by-value version.

OK, so the structure is trivial. But so is the function. The more complex the function, the more likely the pass-by-value version is to be faster, because the compiler knows that the two arguments do not "overlap". This is a huge benefit to optimization.

Of course, this decision should primarily be based on the semantics of the function: Do you need NULL to be a legal value? If so, obviously you need a pointer. Do you need to modify the objects? Then use a pointer or a reference.

But if you do not need to modify the objects, prefer to pass them by value unless the objects are large and/or have a non-trivial copy constructor (e.g. std::string). If by-value really is too slow, pass by reference-of-const or pointer-to-const.

But do not underestimate the potential speed advantages of passing by value, which derive from the advantages of registers vs. memory and instruction reordering. And note that these advantages become more pronounced with every generation of CPU.

Answer 3

Passing by pointer and by reference are really the same, except in syntax. I prefer passing by pointer, because it makes things explicit:

Object bar;
ptr_foo(&bar); // bar may change

ref_foo(bar); // can bar change? Now I need to go look at the prototype...

val_foo(bar); // bar cannot change. (Unless you use references here and there)

The only technical preference between passing values and pointers, as you have touched on is if the class is large enough to make its passing slow.

Answer 4

Use:

1. const reference if the object is not modified
2. pointer if the object is modified or can be null
3. value if the object is small and you care about performance or if you need a copy of the object inside the function. This allows the compiler to pick the best way to copy/move the argument.
4. std::unique_ptr if ownership is transferred to the function.

Answer 5

You can take a look at https://www.boost.org/doc/libs/1_51_0/libs/utility/call_traits.htm library, it converts the type to the best arguments type automatically.

Answer 6

References any day, if you're designing everything yourself. Idiomatic modern C++ should almost never have raw pointers sticking out anywhere. Dynamically allocated objects should travel in resource managing containers (shared_ptr or unique_ptr, or weak_ptr if applicable), but for most operations passing by (const) reference is the primary way to pass arguments that need to be modified or that are of a heavy-weight type. Don't forget that passing by value may be a viable option if you have movable types.