The use of dynamically allocated memory must be managed, because it has a tremendous effect on the performance of your programs. The current trend in handling dynamic memory seems to be shifting toward an automated approach. While C++ uses the manual approach for managing dynamic memory, this does not mean that it can't be automated in that language -- thus giving the C++ programmer the best of both worlds. This article explains how to do it. It is excerpted from chapter two of The Art of C++, written by Herbert Schildt (McGraw-Hill/Osborne, 2004; ISBN: 0072255129).
A Simple Garbage Collector for C++ - Choosing a Garbage Collection Algorithm (Page 3 of 12 )
Before implementing a garbage collector for C++, it is necessary to decide what garbage collection algorithm to use. The topic of garbage collection is a large one, having been the focus of serious academic study for many years. Because it presents an intriguing problem for which there is a variety of solutions, a number of different garbage collection algorithms have been designed. It is far beyond the scope of this book to examine each in detail. However, there are three archetypal approaches: reference counting, mark and sweep, and copying. Before choosing an approach, it will be useful to review these three algorithms.
In reference counting, each dynamically allocated piece of memory has associated with it a reference count. This count is incremented each time a reference to the memory is added and decremented each time a reference to the memory is removed. In C++ terms, this means that each time a pointer is set to point to a piece of allocated memory, the reference count associated with that memory is incremented. When the pointer is set to point elsewhere, the reference count is decremented. When the reference count drops to zero, the memory is unused and can be released.
The main advantage of reference counting is simplicity—it is easy to understand and implement. Furthermore, it places no restrictions on the organization of the heap because the reference count is independent of an object’s physical location. Reference counting adds overhead to each pointer operation, but the collection phase is relatively low cost. The main disadvantage is that circular references prevent memory that is otherwise unused from being released. A circular reference occurs when two objects point to each other, either directly or indirectly. In this situation, neither object’s reference count ever drops to zero. Some solutions to the circular reference problem have been devised, but all add complexity and/or overhead.
Mark and Sweep
Mark and sweep involves two phases. In the first phase, all objects in the heap are set to their unmarked state. Then, all objects directly or indirectly accessible from program variables are marked as “in-use.” In phase two, all of allocated memory is scanned (that is, a sweep of memory is made), and all unmarked elements are released.
There are two main advantages of mark and sweep. First, it easily handles circular references. Second, it adds virtually no runtime overhead prior to collection. It has two main disadvantages. First, a considerable amount of time might be spent collecting garbage because the entire heap must be scanned during collection. Thus, garbage collection may cause unacceptable runtime characteristics for some programs. Second, although mark and sweep is simple conceptually, it can be tricky to implement efficiently.
The copying algorithm organizes free memory into two spaces. One is active (holding the current heap), and the other is idle. During garbage collection, in-use objects from the active space are identified and copied into the idle space. Then, the roles of the two spaces are reversed, with the idle space becoming active and the active space becoming idle. Copying offers the advantage of compacting the heap in the copy process. It has the disadvantage of allowing only half of free memory to be in use at any one time.
Given that there are advantages and disadvantages to all of the three classical approaches to garbage collection, it might seem hard to choose one over the other. However, given the constraints enumerated earlier, there is a clear choice: reference counting. First and most importantly, reference counting can be easily “layered onto” the existing C++ dynamic allocation system. Second, it can be implemented in a straightforward manner and in a way that does not affect preexisting code. Third, it does not require any specific organization or structuring of the heap, thus the standard allocation system provided by C++ is unaffected.
The one drawback to using reference counting is its difficulty in handling circular references. This isn’t an issue for many programs because intentional circular references are not all that common and can usually be avoided. (Even things that we call circular, such as a circular queue, don’t necessarily involve a circular pointer reference.) Of course, there are cases in which circular references are needed. It is also possible to create a circular reference without knowing you have done so, especially when working with third-party libraries. Therefore, the garbage collector must provide some means to gracefully handle a circular reference, should one exist.
To handle the circular reference problem, the garbage collector developed in this chapter will release any remaining allocated memory when the program exits. This ensures that objects involved in a circular reference will be freed and their destructors called. It is important to understand that normally there will be no allocated objects remaining at program termination. This mechanism is explicitly for those objects that can’t be released because of a circular reference. (You might want to experiment with other means of handling the circular reference problem. It presents an interesting challenge.)
Implementing the Garbage Collector
To implement a reference counting garbage collector, there must be some way to keep track of the number of pointers that point to each piece of dynamically allocated memory. The trouble is that C++ has no built-in mechanism that enables one object to know when another object is pointing to it. Fortunately, there is a solution: you can create a new pointer type that supports garbage collection. This is the approach used by the garbage collector in this chapter.
To support garbage collection, the new pointer type must do three things. First, it must maintain a list of reference counts for active dynamically allocated objects. Second, it must keep track of all pointer operations, increasing an object’s reference count each time a pointer is pointed to that object and decreasing the count each time a pointer is redirected to another object. Third, it must recycle those objects whose reference counts drop to zero. Aside from supporting garbage collection, the new pointer type will look and feel just like a normal pointer. For example, all pointer operations, such as * and –>, are supported.
In addition to being a convenient way to implement a garbage collector, the creation of a garbage collection pointer type satisfies the constraint that the original C++ dynamic allocation system must be unaffected. When garbage collection is desired, garbage collection-enabled pointers are used. When garbage collection is not desired, normal C++ pointers are available. Thus, both types of pointers can be used within the same program.
To Multithread or Not?
Another consideration when designing a garbage collector for C++ is whether it should be single-threaded or multithreaded. That is, should the garbage collector be designed as a background process, running in its own thread and collecting garbage as CPU time permits? Or, should the garbage collector run in the same thread as the process that uses it, collecting garbage when certain program conditions occur? Both approaches have advantages and disadvantages.
The main advantage to creating a multithreaded garbage collector is efficiency. Garbage can be collected when idle CPU cycles are available. The disadvantage is, of course, that C++ does not provide any built-in support for multithreading. This means that any multithreaded approach will depend upon operating system facilities to support the multitasking. This makes the code nonportable.
The main advantage to using a single-threaded garbage collector is portability. It can be used in situations that do not support multithreading or in cases in which the price of multithreading is too high. The main disadvantage is that the rest of the program stops when garbage collection takes place.
In this chapter, the single-threaded approach is used because it works in all C++ environments. Thus, it can be used by all readers of this book. However, for those readers wanting a multithreaded solution, one is given in Chapter 3, which deals with the techniques needed to successfully multithread a C++ program in a Windows environment.
When to Collect Garbage?
One final question that needs to be answered before a garbage collector can be implemented: When is garbage collected? This is less of an issue for a multithreaded garbage collector, which can run continuously as a background task, collecting garbage whenever CPU cycles are available, than it is for a single-threaded garbage collector, such as that developed in this chapter, which must stop the rest of the program to collect garbage.
In the real world, garbage collection usually takes place only when there is sufficient reason to do so, such as the case of memory running low. This makes sense for two reasons. First, with some garbage collection algorithms, such as mark and sweep, there is no way to know that a piece of memory is unused without actually performing the collection. (That is, sometimes there is no way to know that there is garbage to be collected without actually collecting it!) Second, collecting garbage is a time-consuming process which should not be performed needlessly.
However, waiting for memory to run low before initiating garbage collection is not suitable for the purposes of this chapter because it makes it next to impossible to demonstrate the collector. Instead, the garbage collector developed in this chapter will collect garbage more frequently so that its actions can easily be observed. As the collector is coded, garbage is collected whenever a pointer goes out of scope. Of course, this behavior can easily be changed to fit your applications.
One last point: when using reference-count based garbage collection, it is technically possible to recycle unused memory as soon as its reference count drops to zero, rather than using a separate garbage collection phase. However, this approach adds overhead to every pointer operation. The method used by this chapter is to simply decrement the memory’s reference count each time a pointer to that memory is redirected and let the collection process handle the actual recycling of memory at more convenient times. This reduces the runtime overhead associated with pointer operations, which one typically wants to be as fast as possible.