Const correctness

Part of C++ FQA Lite. To see the original answers, follow the FAQ links.

This is about the const keyword, which makes you write your program twice (or more, depending on your luck).

[18.1] What is "const correctness"?

FAQ: Oh, that's a great thing. You declare an object as const and prevent it from being modified.

For example, if you pass a std::string object to a function by const pointer or reference, the function won't modify it and it won't be copied the way it happens when the object is passed by value.

FQA: Interesting. What about vector of pointers to objects? Let's see. If the vector itself is declared const (as in const std::vector<T*>), then you can't modify the vector, but you can modify the objects. If the pointers are declared const (as in std::vector<const T*>), then you can modify the vector, but not the objects. Now suppose you have a vector of non-const objects, and you want to pass them to a function that accepts a const vector of const objects. Oops, can't do that - the vectors are two different unrelated types as far as the compiler is concerned (no, C++ weenies, it actually doesn't make sense, think again). You'll have to create a temporary vector of the right type and copy the pointers (which will compile just fine since T* is silently convertible to const T*).

That's what const correctness is all about: increasing the readability of your program and protecting you from errors without any performance penalties.

By the way, not using const in C++ is quite likely not a very good idea. Of all questionable C++ features, const probably does the most visible damage when avoided. Any piece of code using const forces every other piece touching it to use const, or else you won't be able to work with the objects it gives you. And here's the funny part: every piece of code not using const forces every other piece touching it to not use it, either - or else you won't be able to pass objects to it. So if you have one const particle and one anti-const particle, there's a big shiny explosion of const_cast in your code. Since const is hard-wired into the language (no way to pass a temporary to a function that gets a non-const reference, for example) and the standard library (iterator and const_iterator), using const is usually a safer bet than avoiding it.

[18.2] How is "const correctness" related to ordinary type safety?

FAQ: It's one form of type safety. You can think about const std::string as an almost separate type that lacks certain operations of a string, like assignment. If you find type safety useful, you'll also like const correctness.

FQA: It's related to ordinary type safety in a pretty complicated way. const and volatile are special cases in the type system - they are "type qualifiers". The relation between a qualified type and a non-qualified type is different from any other relation between types in the language. This is one of the very many complications in the implicit conversion and overload resolution rules.

It works smoothly for the simple cases, especially if there are no other complications. It breaks pretty hard whenever you have a pointer-like object. Pointers can get two const qualifiers, one for the pointer and one for the pointed values. This is awkward and unreadable and when you have pointers to pointers and three const qualifiers, you may need cryptic explicit casts, but at least there's syntax for all the levels of constness. With "smart pointers" (the things you should stuff into every possible hole because pointers are evil), there's no such syntax. That's why we have iterator and const_iterator - saying const iterator says that the iterator is immutable, but not what it points to. Exercise: implement a vector-like class that can get the storage pointer from the user, in a const-correct way that supports attaching both to constant and mutable storage areas.

And of course a vector of const pointers is compiled to a different bulk of (identical) assembly code than a vector of mutable pointers. At least here the compiler is writing the same program twice, which is better than having to do this yourself. Which also happens - const_iterator is one family of examples.

[18.3] Should I try to get things const correct "sooner" or "later"?

FAQ: As soon as possible, because when you add const to a piece of code, it triggers changes in every place related to it.

FQA: That's right. Since you can't get out of the tar pit, the best strategy is to climb right into the middle and make yourself comfortable from the beginning. No kidding, I actually agree with the FAQ. See also the advice above about not avoiding const.

[18.4] What does "const Fred* p" mean?

FAQ: A pointer to a const Fred object. The object can't be changed, for example, you can't call methods not qualified as const.

FQA: Right. But don't count on it when you debug code. const can be cast away in a snap (pretty much like everything else in C++), and there's the mutable keyword for creating members that can be modified by methods qualified as const. And of course someone can have another, non-const pointer to the same Fred object. And the const-qualified methods may modify data pointed by its member pointers and references or by pointers kept in things pointed by its member pointers, etc.

The pointer aliasing issue is one reason that the compiler can't really optimize code because it sees a pointer declared as const. And when it can figure out there are no aliases, it doesn't need your const declarations to help it. You can explain the difference between data flow analysis and type qualification to the next pseudo-performance-aware person who advocates declaring every local integer as const. See also a correct FAQ answer to this question below.

[18.5] What's the difference between "const Fred* p", "Fred* const p" and "const Fred* const p"?

FAQ: In the first example, the object is immutable. In the second example, the pointer is immutable. In the third example, both are immutable.

FQA: Um, right. Remember: const makes your programs readable.

[18.6] What does "const Fred& x" mean?

FAQ: It's a reference to an immutable Fred object.

FQA: Which is similar to a pointer to an immutable Fred object. However, the FAQ holds the "references are NOT pointers" religion (specifically, it belongs to the "pointers are evil" faction), so it dutifully replicates the explanation given in the answer about the pointer case, replacing "pointer" with "reference".

[18.7] Does "Fred& const x" make any sense?

FAQ: No. It says that you can change the object, but not the reference. But you can never change a reference anyway, it will always refer to a single object.

Don't write such declarations, it confuses people.

FQA: So why does this compile? Wait, I don't really want to know. Whether it's because they didn't bother to disallow it, or because some special case of template instantiation (like when you add a const to a parameter type which is in fact a reference) would fail or anything like that - that's just another lame excuse as far as a language user is concerned.

[18.8] What does "Fred const& x" mean?

FAQ: It's equivalent to "const Fred& x". The question is - which form should you use? Nobody can answer this for your organization without understanding your needs. The are lots of business scenarios, some producing the need for one form, others for the other. The discussion takes a full screen of text.

FQA: Yawn. Another stupid duplication in C++.

Come on. What "business scenarios" can "produce needs" for a form of a const declaration? If your organization employs people who can't memorize both forms or check what a declaration means, you have to admit they will drown in C++ even if you somehow know for sure that in general they can program. C++ has megatons of syntax.

[18.9] What does "Fred const* x" mean?

FAQ: The FAQ replicates the previous answer, replacing "reference" with "pointer". Probably the same religion thing again.

FQA: Yawn. Another stupid duplication in the C++ FAQ.

[18.10] What is a "const member function"?

FAQ: It's declared by appending const to the prototype of a class member function. Only this kind of methods may be called when you have a const object. Errors are caught at compile time, without any speed or space penalty.

FQA: As discussed above, this breaks into little pieces when your class is similar to a pointer in the sense that a user can change both the state of your object and use your object to change some other state. As to the speed and space penalty, you may check the symbol table of your program if you want to know how many functions were generated twice from a single template because of const and non-const template parameters. Then you can count the functions having identical code except for extra const qualifiers in some versions, but not others.

[18.11] What's the relationship between a return-by-reference and a const member function?

FAQ: const member functions returning references to class members should use a const reference. Many times the compiler will catch you when you violate this rule. Sometimes it won't. You have to think.

FQA: Of course you have to think about this complete and utter nonsense! Somebody has to, and the C++ designers didn't.

As usual with const, the compiler becomes dumb when levels of indirection appear. For instance, if you keep a member reference, the thing it points to is not considered part of the object. It's up to you to decide whether you want to return this reference as const or non-const from your const member accessor. Both choices may eventually lead to const_cast.

[18.12] What's the deal with "const-overloading"?

FAQ: You can have two member functions with a single difference in the prototype - the trailing const. The compiler will select the function based on whether the this argument is const or not. For instance, a class with a subscript operator may have two versions: one returning mutable objects and one returning const objects - in the latter version, this is also qualified as const.

FQA: Please try to avoid this feature. In particular, get and set functions having the same name (const-overloaded) are not a very good idea.

Having to replicate the code of a subscript operator just to add two const qualifiers - for the return value and the this argument - is yet another example of const forcing you to write your program twice. Well, most likely there are more problems where this one came from. A class with a subscript operator is a "smart array", probably not unlike std::vector, so you'll end up with much more replicated code - an iterator and a const_iterator or some such.

[18.13] What do I do if I want a const member function to make an "invisible" change to a data member?

FAQ: There are legitimate use cases for this - that would be when a user doesn't see the change using the interface of the class. For example, a set class may cache the last look-up to possibly speed up the next look-up.

Use the mutable keyword when you declare the members you want to change this way. If your compiler doesn't have mutable, use const_cast<Set*>(this) or the like. However, try to avoid this because it can lead to undefined behavior if the object was originally declared as const.

FQA: Most compilers probably support mutable today, but you may still need const_cast because the tweaks to the type system supporting const break in many cases. Which may be a problem combined with the fact that const_cast is not designed to work with objects declared as const (as opposed to those declared non-constant but passed to a function by const reference or pointer).

There are numerous reasons making const_cast safer in practice than in theory. People rarely declare objects as const. Compiler writers are unlikely to add special cases to their compiler to support const objects of C++ classes because it's hard work that is unlikely to pay off. For example, allocating a global const C-style structure with an aggregate initializer in read-only memory is easy. Doing the same for a C++ class with a constructor is hard because the constructor must be able to write to the object upon initialization. So you'd have to use writable memory you later make readable, which is not typically supported by object file formats and operating systems. Or you could translate C++ code to a C-style aggregate initializer, spending lots of effort to only handle the cases when the compiler can inline the code of the constructors.

However, there is no simple rule making it "very close to impossible" for a compiler writer to use the opportunity provided by the standard and implement const objects differently from other objects. In the case of optimizing the dereferencing of const pointers, such a rule does exist (there might be other, non-const pointers to the same location). But in the case of objects declared const, there should be no such pointers.

The fact that the compiler is allowed to work under this assumption could be great if it were more likely to actually yield performance benefits, and if const worked to an extent making the use of const_cast unnecessary. The current state of affairs just creates another source of uncertainty for the programmer.

[18.14] Does const_cast mean lost optimization opportunities?

FAQ: Theoretically, yes. In practice, no. If a compiler knows that a value used in a piece of code is not modified, it can optimize the access to that value, for example, fetch it to a machine register. However, the compiler can't be sure that a value pointed by a const pointer is not modified because of the aliasing problem: there can be other pointers to the same object, not necessarily constant. Proving the opposite is almost always impossible.

And when the compiler can't prove it, it can't speed up the access to the value. For example, if it uses the value it fetched to a register after someone modified it in the original memory location, the compiler changes the meaning of the program because it uses an outdated value.

FQA: Three cheers to the FAQ! No, really, this time the answer describes the actual state of affairs. You understand why I'm so deeply touched if you met some of the many C++ users who give a different answer to this question, and know their ways to reason about performance.

I think the answer is "no" in theory, too, because basically the compiler has to figure out which parts of the code modify the data, and if it knows the data won't get modified while this piece of code is running, it can use this fact, and otherwise it can't, so what difference do your const-qualifications make? But let's say I'm not really sure, and anyway, it's not the time to argue about the exact phrasing when once in a lifetime a realistic statement appears about the performance of C++ code.

The thing that is worth noting is that aliasing problems impede the performance of "generic" libraries, not just compilers. For example, consider vector<T>::push_back(). The vector will try to append the object to its storage. If there's no free space left, it will allocate a larger chunk, copy the old objects, and then append the new one. The new object is thus copied once (from wherever it was into the vector storage), right?

But what if the object is itself located in the old chunk, and vector tries to be efficient and use realloc, which may or may not free the old memory? The object may be wiped out by realloc. Very good, then, vector creates another copy. Hop 1 - from the parameter to a temporary, hop 2 - from a temporary to the vector storage and everyone is happy - after a certain amount of clock cycles.

"Theoretically, yes. In practice, no." Quite a nice summary of C++.

[18.15] Why does the compiler allow me to change an int after I've pointed at it with a const int*?

FAQ: Because when you point to something with a const pointer, this only means that you can't use that pointer to change the object. It doesn't mean the object can't be changed at all, because it can be accessible in other ways, not only through this pointer.

FQA: Exactly; it's related to the previous FAQ. Surprisingly for C++, it even makes sense. After all, when the object is declared, it seems like the right place to say what can be done with it. But if anyone can take an existing object and point to it and thus redefine what can be done with it, it's just weird, and can't really work, because it could be done from many places in incompatible ways.

[18.16] Does "const Fred* p" mean that *p can't change?

FAQ: No, for example, it could change through a non-constant Fred* q pointing to the same object.

FQA: This question is just like the previous one.

[18.17] Why am I getting an error converting a Foo** to const Foo**?

FAQ: It would be "invalid and dangerous"! Suppose Foo** points to an array of pointers, which can be used to modify the pointed objects. Suppose C++ would allow to pass this array to a function which expects const Foo** - an array of pointers which can't be used to modify the pointed objects.

Now suppose that this function fills the array with a bunch of pointers to constant objects. This looks perfectly good in that context, because that's what the function was passed - an array of pointers to constant objects. But what we've got now is an array of pointers which can be used to modify those const objects, because the declaration of the array does allow such modifications!

It's a good thing we get a compile time error instead. Don't use casts to work around this!

FQA: Wonderful. But why are we discussing evil built-in pointers anyway? Let's talk about the smart C++ array classes. Why can't I pass std::vector<T*>& to a function expecting const std::vector<const T*>&? The function clearly says that it's not going to modify anything: neither the vector nor the pointed objects. Why can't I pass my mutable vector of mutable objects to a function that promises not to modify either of those?

Copyright © 2007-2009 Yossi Kreinin
revised 17 October 2009