By Scott Meyers
It isn’t hard to keep a count of all the objects allocated for a given class in C++, unless you have to deal with distractions.
Sometimes easy things are easy, but they’re still subtle. For example, suppose you have a class Widget, and you’d like to have a way to find out at run time how many Widget objects exist. An approach that’s both easy to implement and that gives the right answer is to create a static counter in Widget, increment the counter each time a Widget constructor is called, and decrement it whenever the Widget destructor is called. You also need a static member function howMany to report how many Widgets currently exist. If Widget did nothing but track how many of its type exist, it would look more or less like this:
class Widget {
public:
Widget() { ++count; }
Widget(const Widget&) { ++count; }
~Widget() { –count; }
static size_t howMany()
{ return count; }
private:
static size_t count;
};
// obligatory definition of count. This
// goes in an implementation file
size_t Widget::count = 0;
This works fine. The only mildly tricky thing is to remember to implement the copy constructor, because a compiler-generated copy constructor for Widget wouldn’t know to increment count. If you had to do this only for Widget, you’d be done, but counting objects is something you might want to implement for several classes. Doing the same thing over and over gets tedious, and tedium leads to errors. To forestall such tedium, it would be best to somehow package the above object-counting code so it could be reused in any class that wanted it. The ideal package would:
be easy to use — require minimal work on the part of class authors who want to use it. Ideally, they shouldn’t have to do more than one thing, that is, more than basically say "I want to count the objects of this type."
be efficient — impose no unnecessary space or time penalties on client classes employing the package.
be foolproof — be next to impossible to accidently yield a count that is incorrect. (We’re not going to worry about malicious clients, ones who deliberately try to mess up the count. In C++, such clients can always find a way to do their dirty deeds.)
Stop for a moment and think about how you’d implement a reusable object-counting package that satisfies the goals above. It’s probably harder than you expect. If it were as easy as it seems like it should be, you wouldn’t be reading an article about it in this magazine.
new, delete, and Exceptions
While you’re mulling over your solution to the object-counting problem, allow me to switch to what seems like an unrelated topic. That topic is the relationship between new and delete when constructors throw exceptions. When you ask C++ to dynamically allocate an object, you use a new expression, as in:
class ABCD { … }; // ABCD = "A Big Complex Datatype"
ABCD *p = new ABCD; // a new expression
The new expression — whose meaning is built into the language and whose behavior you cannot change — does two things. First, it calls a memory allocation function called operator new. That function is responsible for finding enough memory to hold an ABCD object. If the call to operator new succeeds, the new expression then invokes an ABCD constructor on the memory that operator new found. But suppose operator new throws a std::bad_alloc exception. Exceptions of this type indicate that an attempt to dynamically allocate memory has failed. In the new expression above, there are two functions that might give rise to that exception. The first is the invocation of operator new that is supposed to find enough memory to hold an ABCD object. The second is the subsequent invocation of the ABCD constructor that is supposed to turn the raw memory into a valid ABCD object. If the exception came from the call to operator new, no memory was allocated. However, if the call to operator new succeeded and the invocation of the ABCD constructor led to the exception, it is important that the memory allocated by operator new be deallocated. If it’s not, the program has a memory leak. It’s not possible for the client — the code requesting creation of the ABCD object — to determine which function gave rise to the exception. For many years this was a hole in the draft C++ language specification, but in March 1995 the C++ Standards committee adopted the rule that if, during a new expression, the invocation of operator new succeeds and the subsequent constructor call throws an exception, the runtime system must automatically deallocate the memory that operator new allocated. This deallocation is performed by operator delete, the deallocation analogue of operator new. (For details, see the sidebar on placement new and placement delete.) It is this relationship between new expressions and operator delete affects us in our attempt to automate the counting of object instantiations.
Counting Objects
In all likelihood, your solution to the object-counting problem involved the development of an object-counting class. Your class probably looks remarkably like, perhaps even exactly like, the Widget class I showed earlier:
// see below for a discussion of why
// this isn’t quite right
class Counter {
public:
Counter() { ++count; }
Counter(const Counter&) { ++count; }
~Counter() { –count; }
static size_t howMany()
{ return count; }
private:
static size_t count;
};
// This still goes in an
// implementation file
size_t Counter::count = 0;
The idea here is that authors of classes that need to count the number of objects in existence simply use Counter to take care of the bookkeeping. There are two obvious ways to do this. One way is to define a Counter object as a class data member, as in:
// embed a Counter to count objects
class Widget {
public:
….. // all the usual public
// Widget stuff
static size_t howMany()
{ return Counter::howMany(); }
private:
….. // all the usual private
// Widget stuff
Counter c;
};
The other way is to declare Counter as a base class, as in:
// inherit from Counter to count objects
class Widget: public Counter {
….. // all the usual public
// Widget stuff
private:
….. // all the usual private
// Widget stuff
};
Both approaches have advantages and disadvantages. But before we examine them, we need to observe that neither approach will work in its current form. The problem has to do with the static object count inside Counter. There’s only one such object, but we need one for each class using Counter. For example, if we want to count both Widgets and ABCDs, we need two static size_t objects, not one. Making Counter::count nonstatic doesn’t solve the problem, because we need one counter per class, not one counter per object. We can get the behavior we want by employing one of the best-known but oddest-named tricks in all of C++: we turn Counter into a template, and each class using Counter instantiates the template with itself as the template argument. Let me say that again. Counter becomes a template:
template
class Counter {
public:
Counter() { ++count; }
Counter(const Counter&) { ++count; }
~Counter() { –count; }
static size_t howMany()
{ return count; }
private:
static size_t count;
};
template
size_t
Counter ::count = 0; // this now can go in header
The first Widget implementation choice now looks like:
// embed a Counter to count objects
class Widget {
public:
…..
static size_t howMany()
{return Counter ::howMany();}
private:
…..
Counter c;
};
And the second choice now looks like:
// inherit from Counter to count objects
class Widget: public Counter {
…..
};
Notice how in both cases we replace Counter with Counter . As I said earlier, each class using Counter instantiates the template with itself as the argument. The tactic of a class instantiating a template for its own use by passing itself as the template argument was first publicized by Jim Coplien. He showed that it’s used in many languages (not just C++) and he called it "a curiously recurring template pattern" [1]. I don’t think Jim intended it, but his description of the pattern has pretty much become its name. That’s too bad, because pattern names are important, and this one fails to convey information about what it does or how it’s used. The naming of patterns is as much art as anything else, and I’m not very good at it, but I’d probably call this pattern something like "Do It For Me." Basically, each class generated from Counter provides a service (it counts how many objects exist) for the class requesting the Counter instantiation. So the class Counter counts Widgets, and the class Counter counts ABCDs. Now that Counter is a template, both the embedding design and the inheritance design will work, so we’re in a position to evaluate their comparative strengths and weaknesses. One of our design criteria was that object-counting functionality should be easy for clients to obtain, and the code above makes clear that the inheritance-based design is easier than the embedding-based design. That’s because the former requires only the mentioning of Counter as a base class, whereas the latter requires that a Counter data member be defined and that howMany be reimplemented by clients to invoke Counter’s howMany [2]. That’s not a lot of additional work (client howManys are simple inline functions), but having to do one thing is easier than having to do two. So let’s first turn our attention to the design employing inheritance.
Using Public Inheritance
The design based on inheritance works because C++ guarantees that each time a derived class object is constructed or destroyed, its base class part will also be constructed first and destroyed last. Making Counter a base class thus ensures that a Counter constructor or destructor will be called each time a class inheriting from it has an object created or destroyed. Any time the subject of base classes comes up, however, so does the subject of virtual destructors. Should Counter have one? Well-established principles of object-oriented design for C++ dictate that it should. If it has no virtual destructor, deletion of a derived class object via a base class pointer yields undefined (and typically undesirable) results:
class Widget: public Counter
{ … };
Counter *pw =
new Widget; // get base class ptr
// to derived class object
……
delete pw; // yields undefined results
// if the base class lacks
// a virtual destructor
Such behavior would violate our criterion that our object-counting design be essentially foolproof, because there’s nothing unreasonable about the code above. That’s a powerful argument for giving Counter a virtual destructor. Another criterion, however, was maximal efficiency (imposition of no unnecessary speed or space penalty for counting objects), and now we’re in trouble. We’re in trouble because the presence of a virtual destructor (or any virtual function) in Counter means each object of type Counter (or a class derived from Counter) will contain a (hidden) virtual pointer, and this will increase the size of such objects if they don’t already support virtual functions [3]. That is, if Widget itself contains no virtual functions, objects of type Widget would increase in size if Widget started inheriting from Counter . We don’t want that. The only way to avoid it is to find a way to prevent clients from deleting derived class objects via base class pointers. It seems that a reasonable way to achieve this is to declare operator delete private in Counter:
template
class Counter {
public:
…..
private:
void operator delete(void*);
…..
};
Now the delete expression won’t compile:
class Widget: public Counter { … };
Counter *pw = new Widget; ……
delete pw; // Error. Can’t call private
// operator delete
Unfortunately — and this is the really interesting part — the new expression shouldn’t compile either!
Counter *pw =
new Widget; // this should not
// compile because
// operator delete is
// private
Remember from my earlier discussion of new, delete, and exceptions that C++’s runtime system is responsible for deallocating memory allocated by operator new if the subsequent constructor invocation fails. Recall also that operator delete is the function called to perform the deallocation. But we’ve declared operator delete private in Counter, which makes it invalid to create objects on the heap via new! Yes, this is counterintuitive, and don’t be surprised if your compilers don’t yet support this rule, but the behavior I’ve described is correct. Furthermore, there’s no other obvious way to prevent deletion of derived class objects via Counter* pointers, and we’ve already rejected the notion of a virtual destructor in Counter. So I say we abandon this design and turn our attention to using a Counter data member instead.
Using a Data Member
We’ve already seen that the design based on a Counter data member has one drawback: clients must both define a Counter data member and write an inline version of howMany that calls the Counter’s howMany function. That’s marginally more work than we’d like to impose on clients, but it’s hardly unmanageable. There is another drawback, however. The addition of a Counter data member to a class will often increase the size of objects of that class type. At first blush, this is hardly a revelation. After all, how surprising is it that adding a data member to a class makes objects of that type bigger? But blush again. Look at the definition of Counter:
template
class Counter {
public:
Counter();
Counter(const Counter&);
~Counter();
static size_t howMany();
private:
static size_t count;
};
Notice how it has no nonstatic data members. That means each object of type Counter contains nothing. Might we hope that objects of type Counter have size zero? We might, but it would do us no good. C++ is quite clear on this point. All objects have a size of at least one byte, even objects with no nonstatic data members. By definition, sizeof will yield some positive number for each class instantiated from the Counter template. So each client class containing a Counter object will contain more data than it would if it didn’t contain the Counter. (Interestingly, this does not imply that the size of a class without a Counter will necessarily be bigger than the size of the same class containing a Counter. That’s because alignment restrictions can enter into the matter. For example, if Widget is a class containing two bytes of data but that’s required to be four-byte aligned, each object of type Widget will contain two bytes of padding, and sizeof(Widget) will return 4. If, as is common, compilers satisfy the requirement that no objects have zero size by inserting a char into Counter , it’s likely that sizeof(Widget) will still yield 4 even if Widget contains a Counter object. That object will simply take the place of one of the bytes of padding that Widget already contained. This is not a terribly common scenario, however, and we certainly can’t plan on it when designing a way to package object-counting capabilities.) I’m writing this at the very beginning of the Christmas season. (It is in fact Thanksgiving Day, which gives you some idea of how I celebrate major holidays…) Already I’m in a Bah Humbug mood. All I want to do is count objects, and I don’t want to haul along any extra baggage to do it. There has got to be a way.
Using Private Inheritance
Look again at the inheritance-based code that led to the need to consider a virtual destructor in Counter:
class Widget: public Counter
{ … };
Counter *pw = new Widget;
……
delete
pw; // yields undefined results
// if Counter lacks a virtual
// destructor
Earlier we tried to prevent this sequence of operations by preventing the delete expression from compiling, but we discovered that that also prohibited the new expression from compiling. But there is something else we can prohibit. We can prohibit the implicit conversion from a Widget* pointer (which is what new returns) to a Counter * pointer. In other words, we can prevent inheritance-based pointer conversions. All we have to do is replace the use of public inheritance with private inheritance:
class Widget: private Counter
{ … };
Counter *pw =
new Widget; // error! no implicit
// conversion from
// Widget* to
// Counter *
Furthermore, we’re likely to find that the use of Counter as a base class does not increase the size of Widget compared to Widget’s stand-alone size. Yes, I know I just finished telling you that no class has zero size, but — well, that’s not really what I said. What I said was that no objects have zero size. The C++ Standard makes clear that the base-class part of a more derived object may have zero size. In fact many compilers implement what has come to be known as the empty base optimization [4]. Thus, if a Widget contains a Counter, the size of the Widget must increase. The Counter data member is an object in its own right, hence it must have nonzero size. But if Widget inherits from Counter, compilers are allowed to keep Widget the same size it was before. This suggests an interesting rule of thumb for designs where space is tight and empty classes are involved: prefer private inheritance to containment when both will do. This last design is nearly perfect. It fulfills the efficiency criterion, provided your compilers implement the empty base optimization, because inheriting from Counter adds no per-object data to the inheriting class, and all Counter member functions are inline. It fulfills the foolproof criterion, because count manipulations are handled automatically by Counter member functions, those functions are automatically called by C++, and the use of private inheritance prevents implicit conversions that would allow derived-class objects to be manipulated as if they were base-class objects. (Okay, it’s not totally foolproof: Widget’s author might foolishly instantiate Counter with a type other than Widget, i.e., Widget could be made to inherit from Counter . I choose to ignore this possibility.) The design is certainly easy for clients to use, but some may grumble that it could be easier. The use of private inheritance means that howMany will become private in inheriting classes, so such classes must include a using declaration to make howMany public to their clients:
class Widget: private Counter {
public:
// make howMany public
using Counter ::howMany;
….. // rest of Widget is unchanged
};
class ABCD: private Counter {
public:
// make howMany public
using Counter ::howMany;
….. // rest of ABCD is unchanged
};
For compilers not supporting namespaces, the same thing is accomplished by replacing the using declaration with the older (now deprecated) access declaration:
class Widget: private Counter {
public:
// make howMany public
Counter ::howMany;
….. // rest of Widget is unchanged
};
Hence, clients who want to count objects and who want to make that count available (as part of their class’s interface) to their clients must do two things: declare Counter as a base class and make howMany accessible [5]. The use of inheritance does, however, lead to two conditions that are worth noting. The first is ambiguity. Suppose we want to count Widgets, and we want to make the count available for general use. As shown above, we have Widget inherit from Counter and we make howMany public in Widget. Now suppose we have a class SpecialWidget publicly inherit from Widget and we want to offer SpecialWidget clients the same functionality Widget clients enjoy. No problem, we just have SpecialWidget inherit from Counter . But here is the ambiguity problem. Which howMany should be made available by SpecialWidget, the one it inherits from Widget or the one it inherits from Counter ? The one we want, naturally, is the one from Counter , but there’s no way to say that without actually writing SpecialWidget::howMany. Fortunately, it’s a simple inline function:
class SpecialWidget: public Widget,
private Counter {
public:
…..
static size_t howMany()
{ return Counter ::howMany(); }
…..
};
The second observation about our use of inheritance to count objects is that the value returned from Widget::howMany includes not just the number of Widget objects, it includes also objects of classes derived from Widget. If the only class derived from Widget is SpecialWidget and there are five stand-alone Widget objects and three stand-alone SpecialWidgets, Widget::howMany will return eight. After all, construction of each SpecialWidget also entails construction of the base Widget part.
Summary
The following points are really all you need to remember:
Automating the counting of objects isn’t hard, but it’s not completely straightforward, either. Use of the "Do It For Me" pattern (Coplien’s "curiously recurring template pattern") makes it possible to generate the correct number of counters. The use of private inheritance makes it possible to offer object-counting capabilities without increasing object sizes.
When clients have a choice between inheriting from an empty class or containing an object of such a class as a data member, inheritance is preferable, because it allows for more compact objects.
Because C++ endeavors to avoid memory leaks when construction fails for heap objects, code that requires access to operator new generally requires access to the corresponding operator delete too.
The Counter class template doesn’t care whether you inherit from it or you contain an object of its type. It looks the same regardless. Hence, clients can freely choose inheritance or containment, even using different strategies in different parts of a single application or library. o
Notes and References
[1] James O. Coplien. "The Column Without a Name: A Curiously Recurring Template Pattern," C++ Report, February 1995.
[2] An alternative is to omit Widget::howMany and make clients call Counter ::howMany directly. For the purposes of this article, however, we’ll assume we want howMany to be part of the Widget interface.
[3] Scott Meyers. More Effective C++ (Addison-Wesley, 1996), pp. 113-122.
[4] Nathan Myers. "The Empty Member C++ Optimization," Dr. Dobb’s Journal, August 1997. Also available at http://www.cantrip.org/emptyopt.html.
[5] Simple variations on this design make it possible for Widget to use Counter to count objects without making the count available to Widget clients, not even by calling Counter ::howMany directly. Exercise for the reader with too much free time: come up with one or more such variations.
Further Reading
To learn more about the details of new and delete, read the columns by Dan Saks on the topic (CUJ January – July 1997), or Item 8 in my More Effective C++ (Addison-Wesley, 1996). For a broader examination of the object-counting problem, including how to limit the number of instantiations of a class, consult Item 26 of More Effective C++.
Acknowledgments
Mark Rodgers, Damien Watkins, Marco Dalla Gasperina, and Bobby Schmidt provided comments on drafts of this article. Their insights and suggestions improved it in several ways.
Scott Meyers authored the best-selling Effective C++, Second Edition and More Effective C++ (both published by Addison Wesley). Find out more about him, his books, his services, and his dog at http://www.aristeia.com.
The C++ equivalent of malloc is operator new, and the C++ equivalent of free is operator delete. Unlike malloc and free, however, operator new and operator delete are overloadable functions, and as such they may take different numbers and types of parameters. This has always been the case for operator new, but until relatively recently, it wasn’t valid to overload operator delete.
The "normal" signature for operator new is:
void * operator new(size_t) throw (std::bad_alloc);(To simplify things from now on, I’ll omit exception specifications. They’re not germane to the points I want to make.) Overloaded versions of operator new are limited to adding additional parameters, so an overloaded version of operator new might look like:
void * operator new(size_t, void *whereToPutObject) { return whereToPutObject; }This particular version of operator new — the one taking an extra void* argument specifying what pointer the function should return — is so commonly useful that it’s in the Standard C++ library (declared in the header ) and it has a name, "placement new." The name indicates its purpose: to allow programmers to specify where in memory an object should be created (where it should be placed). Over time, the term placement new has come to be applied to any version of operator new taking additional arguments. (This terminology is actually enshrined in the grammar for C++ in the forthcoming International Standard.) Hence, when C++ programmers talk about the placement new function, they mean the function above, the one taking the extra void* parameter specifying where an object should be placed. When they talk about a placement new function, however, they mean any version of operator new taking more than the mandatory size_t argument. That includes the function above, but it also includes a plethora of operator new functions that take more or different parameter types. In other words, when the topic is memory allocation functions, "placement new" means "a version of operator new taking extra arguments." The term can mean still other things in other contexts, but we don’t need to go down that road here, so we won’t. For details, consult the suggested reading at the end of the article. By analogy with placement new, the term "placement delete" means "a version of operator delete taking extra arguments." The "normal" signature for operator delete is:
void operator delete(void*);So any version of operator delete taking one or more arguments beyond the mandatory void* parameter is a placement delete function. Let us now revisit an issue discussed in the main article. What happens when an exception is thrown during construction of a heap object? Consider again this simple example:
class ABCD { ... }; ABCD *p = new ABCD;Suppose the attempt to create the ABCD object yields an exception. The main article pointed out that if that exception came from the ABCD constructor, operator delete would automatically be called to deallocate the memory allocated by operator new. But what if operator new is overloaded, and what if different versions of operator new (quite reasonably) allocate memory in different ways? How can operator delete know how to correctly deallocate the memory? Furthermore, what if the ABCD object being created used placement new, as in:
void *objectBuffer = getPointerToStaticBuffer(); ABCD *p = new (objectBuffer) ABCD; // create an ABCD object in a static buffer(The) placement new didn’t actually allocate any memory. It just returned the pointer to the static buffer that was passed to it in the first place. So there’s no need for any deallocation.
Clearly, the actions to be taken in operator delete to undo the actions of its corresponding operator new depend on the version of operator new that was invoked to allocate the memory.
To make it possible for programmers to indicate how the actions of particular versions of operator new can be undone, the C++ Standards committee extended C++ to allow operator delete to be overloaded, too. When an exception is thrown from the constructor for a heap object, then, the game is played a special way. The version of operator delete that’s called is the one whose extra parameter types correspond to those of the version of operator new that was invoked.
If there’s no placement delete whose extra parameters correspond to the extra parameters of the placement new that was called, no operator delete is invoked at all. So the effects of operator new are not undone. For functions like (the) placement version of operator new, this is fine, because they don’t really allocate memory. In general, however, if you create a custom placement version of operator new, you should also create the corresponding custom placement version of operator delete to go with it.
Alas, most compilers don’t yet support placement delete. With code generated from such compilers, you almost always suffer a memory leak if an exception is thrown during construction of a heap object, because no attempt is made to deallocate the memory that was allocated before the constructor was invoked. o
