An Introduction to Garbage Collection Part II—A Look Under the Hood
by Richard Gillam
Advisory Software Engineer
IBM Center for Java Technology-Silicon Valley

Last month, I talked about how there seem to be a lot of misconceptions about garbage collection floating around in the C++ community, part of what seems to be a big backlash against the Java fad (or "Java revolution," if you want to be more positive about it). I promised I’d take a look at what really goes on inside a garbage collector and try to dispel some of the mystery involved and inject some reality into the debate.

Then I spent all my time talking about how regular C++ memory management works, looked at some of the common programmer tricks, and examined reference counting as an avenue for automatic memory management. We needed that background to really look objectively at garbage collection. Now that we have it, I can make good on my original promises.

The term "garbage collection" is usually used to refer collectively to all techniques for automatic memory management, and therefore, reference counting can be thought of as a form of garbage collection. In fact, many implementations (especially early ones) of so-called "garbage-collected languages" such as Lisp and Smalltalk actually used reference counting as the basis of their memory-management subsystems.

As we saw last time, the biggest problem with reference counting is its inability to handle self-referential data structures. If you want to handle cyclical structures correctly, you can’t rely on just twiddling a reference count every time an object picks up or drops a reference to another object. Instead, you have to effectively generate the reference counts on the fly at a given moment in time and then use that to decide which objects to delete. (For reference counting schemes that use small reference counts that peg the count at a maximum value, such as one-byte or one-bit schemes, this is exactly what you have to do to handle cycles and un-stick the reference counts.)

The basic approach is to start from every object that is statically accessible (it’s on the stack or in static space, in C++ terms) and follow all of their references, then follow all of their references, and so on. The objects which are statically accessible are called the root set. Every object you can get to by following pointers from somewhere in the root set is live. Those that you can’t get to from the root set can’t be accessed from the running program; they’re garbage. This general approach, following every reference from the root set until you’ve seen every live object is called tracing garbage collection (because you’re "tracing" all the possible reference paths). The collective term "garbage collection" is also often used to refer just to tracing techniques, rather than to all automatic memory-management techniques.

There are two primary categories of tracing garbage collection, with dozens of variations and hybrids.

Mark-sweep garbage collection

The oldest and simplest for of tracing collection is mark-sweep garbage collection. The basic approach is quite simple: assume each object on the heap has a "mark" bit in its block header. Walk the whole heap and clear every object’s mark bit. Then, go back and set the mark bit in every object that’s referred to from the root set. Then recursively walk the whole reference tree. For every object you’re just marked that refers to other objects, mark every object the object refers to that isn’t already marked and then repeat the process again with each of the objects the object you’re working on refers to. Eventually, each chain of recursion will hit an object that contains no references to other objects at all, or contains only references to objects that have already been marked, and will terminate. At the end of this pass, only the objects that are reachable from the root set will be marked. You can now walk the heap again from beginning to end and delete any objects that aren’t marked (this is the "sweep" part).

In practice, the first pass, where you clear all the mark bits, can "execute" in zero time. This is because at the end of a garbage-collection pass, everything that has its mark bit set a certain way is garbage and is deleted. The objects that are left will all have their mark bits set the same way, so you can consider yourself to have "cleared" their mark bits just by switching your definition of which value means "marked" and which means "not marked." So mark-sweep garbage collection takes two passes.

This is pretty nice. It’s simple and obvious, it’ll do the job with no intervention from the user, and it handles cycles in the reference graph correctly. However, mark-sweep garbage collection is still generally based on the traditional new/delete implementations I described above and will inherit their limitations, such as fragmentation and locality of reference problems (actually, the sweep pass can be optimized so that it’s quite a bit faster than just doing a lot of deletes on the garbage objects would normally be). In addition, the recursion takes up more memory, even through you generally don’t do a garbage-collection pass until you’re starting to run out of memory. And, of course, the program grinds to a halt while garbage collection happens, which can take a while if there are a lot of objects on the heap, especially if a lot of them turn to garbage.

Despite these limitations, there are a lot of pretty successful runtimes based on mark-sweep garbage collection. It also works nicely as an adjunct to other forms of memory management, especially one-bit reference counting, lends itself well to incremental implementations, and is the best algorithm to use for certain types of memory-management situations.

Copying garbage collection

The other classical garbage collection algorithm is called copying garbage collection. It’s a little bit more complicated.

Start by imagining that instead of one heap, you have two. In other words, the available heap space is partitioned into two segments of equal size. At any given time, one of these is active—as far as the program is concerned, it’s "the heap." When you run a garbage-collection pass, you follow links from the root set, just like in mark-sweep collection. However, for each new live object you hit, instead of just setting a mark bit, you copy it into the other heap and adjust the pointer referring to it so that it points to the copied object in its new location. At the end of a garbage collection pass, all the live objects have been copied into the inactive heap, which is now considered the active heap. The other heap, which is now inactive, contains only garbage and can be overwritten with impunity during the next garbage collection pass.

Of course, all the pointers have to be adjusted properly. This can be accomplished quite simply: when each object is copied into the new heap, a forwarding address is left where the object used to be (remember that each block is guaranteed to be big enough to hold the forwarding address), and its block header is overwritten with a sentinel value that says the block has already been copied. After the block has been copied, every new object you find that refers to it will be pointing to the forwarding address. You just stamp the forwarding address over the old address and you’re done. When you copy a forwarding address, you know that object has already been seen by the garbage collector, so you don’t need to follow the link.

Copying garbage collection is cool. Here’s why: objects are allocated from one end of the heap to the other, and a whole sub-heap is reclaimed at once. This means that the only state the memory manager has to maintain is a pointer to where the next object to be allocated will be. Newing up an object is cheap: all you have to do is increment the next-object pointer by an appropriate amount. This is just about as cheap as stack allocation. You also know immediately when the sub-heap is full. Deleting an object effectively costs nothing: an entire sub-heap is declared to be free memory at the end of a garbage-collection pass. In other words, instead of paying to create and delete an object, you pay to keep it around. Objects that come and go quickly cost practically nothing. (Well, because they cause the collector to fire more often, they don’t really come for free, but it can be made to be pretty close.)

Copying collection also means you have no problems with fragmentation: after any garbage-collection pass, all the live objects are packed down at one end of the currently active heap. Copying collection can also give you much better locality of reference. Because objects are allocated at sequential addresses, objects allocated at the same time are guaranteed to occupy a contiguous range of memory. With conventional memory-management techniques, you can’t guarantee this because of potential fragmentation. Furthermore, because a garbage collection pass copies objects that refer to each other sequentially, you will wind up with objects that weren’t allocated at the same time, but which point to each other, near each other in memory after a garbage collection pass.

So newing and deleting objects is really cheap: effectively, heap allocation with copying garbage collection is as cheap as stack allocation, meaning we can actually get rid of stack allocation altogether. C++ programmers tend not to think this way: we’re trained to see heap allocation as expensive and therefore something to be avoided when possible. With a copying collector, this isn’t true. Newing up an object costs you no more than allocating it on the stack would. And on top of this, you don’t have to worry about fragmentation and you get greatly improved locality of reference.

Of course, copying collection has some serious drawbacks, too. The first is the most obvious: at any given time, only half of the memory in the program’s address space is actually available to store program data. This can be a huge drawback, both because it means you have less space to store stuff, and because it means the collector has to run more frequently. Also, instead of paying to create and delete objects, you pay to keep them around. Thus, the bigger the objects are, or the longer they stay around, the more you pay. Copying garbage collection works best when the ratio of live blocks to garbage blocks at any given time (the program’s residency) is relatively low. The more blocks persist across collection passes, the worse copying collection’s performance becomes. In fact, copying collection’s performance degrades exponentially as residency increases; mark-sweep collection’s performance degrades linearly with residency. So there’s a point where mark-sweep does a better job.

Hybrid approaches

One way to deal with the problem is through a hybrid approach. We know that there are going to be a lot of objects that are created during program initialization and which persist for the entire duration of the program (in C++, there are pointers declared const; in Java, they’re object references declared final). The compiler knows which objects these are and can shunt them to a separate area of memory that isn’t subject to garbage collection (basically, an extension of C++ static space). You can also reduce the impact of large objects by moving them off into a special memory space of their own. The so-called "large object space" would be managed separately by mark-sweep garbage collection. These two techniques can go a long way toward reducing the residency of the spaces controlled by copying collection. Using these techniques increases complexity, and further reduces the space available for normal objects, but we can improve on the amount of memory available, and the other changes don’t have a big impact on performance and can be transparent to the programmer.

Mark-compact garbage collection

The biggest problem with copying garbage collection is the fact that at any given time, half of the available memory is designated as garbage and unavailable for newly-allocated objects. One approach to solving this problem is a hybrid of mark-sweep and copying collection know as mark-compact garbage collection. The idea here is that instead of copying each live object to a new heap on every garbage-collection pass, you compact all the live objects down to one end of the heap. In mark-compact collection, you make three passes. On the first pass, you do the same thing you’d do in mark-sweep collection, except that you also keep track of where each object will be located after the collection pass and write a forwarding address into the object’s block header when you mark it. (Yes, this means adding extra space to every object’s block header.) On the second pass, you go through every live object and update all of its pointers to other objects based on the forwarding addresses in the block headers (if you know an object contains no pointers to other objects, you can skip it, so there’s usually a way for the collector to detect this easily). Finally, on the third pass, you move each live object to its new location. At the end of the pass, all the live objects are gathered at one end of the heap and the rest of the heap is available for new objects.

Obviously, this approach will tend to be slower then both mark-sweep and copying collection, but, as with copying collection, newing and deleting objects is very cheap and keeping them around is expensive, and it can be designed to give good locality of reference (although not as good as pure copying collection). The compacting step’s performance can be improved by trying to avoid moving most live objects and copying only some of them into the holes left by garbage objects (this is called "two-finger compaction"). This has some drawbacks, however: it will tend to waste memory due to fragmentation, it tends to damage locality of reference, and it places a greater burden on the pointer-updating pass.

One common arrangement is a garbage collector that uses straight copying collection most of the time and switches automatically to compacting collection when the heap’s residency grows too large to maintain two heaps. Again, this increases complexity (and can hurt collection performance), but all of the extra complexity can be hidden from the programmer.

Generational garbage collection

Another approach that has gained a lot of favor recently for improving the performance of the traditional techniques is something called generational garbage collection.

To understand how generational collection works, imagine that the heap is divided into three partitions instead of two. Two are the same size; the third need not be—it can be bigger or smaller. We’ll call this third partition the ephemeral partition, and we’ll call the other two the persistent partitions. All objects start out in the ephemeral partition. When the ephemeral partition is collected, all the objects in it that are still live are copied into one of the persistent partitions. With every garbage collection pass, every object in the ephemeral partition is either copied out, or is garbage. The two persistent partitions work the same way they do in normal copying collection. Notice that this technique increases the amount of memory available for new objects— only half of the persistent space, not half of the whole heap, is now off-limits.

Of course, we increase the frequency of garbage-collection passes, which can hurt us. But we can take advantage of a common property of programs: objects tend to die young. As a general rule of thumb, an object will either die when the function that created it terminates, or it’ll stick around for a while. Generally, the longer the object sticks around, the longer we can expect it to continue to stick around.

So we can take advantage of these facts by collecting the ephemeral partition more often then we collect the persistent partition. And since the ephemeral partition will tend to have a much lower residency (live-objects-to-garbage ratio) than the persistent partitions, a garbage collection pass that just covers it can be much quicker (remember that we only pay when an object survives a collection).

This is the principle behind generational garbage collection. Each time an object survives a collection, it is promoted to a higher generation. Older generations are collected less often than younger generations, since we assume most of the garbage is in the younger generations.

Now there’s a problem here. Let’s go back to our simple example where we have only two generations, the ephemeral and persistent generations. You create an object, and it sits around for a while, long enough to survive a collection and get promoted into the persistent generation. Now you call a setter method on it, passing that method a newly-created object. You’ve now fixed things so that an object in the persistent generation points to an object in the ephemeral generation. This is an example of an inter-generational reference. Inter-generational references are the bane of generational garbage collection. This is because what you really want to do is treat each generation as a completely independent heap. Otherwise, you have to trace through every object on the heap with every garbage collection pass, which means the only time you save is the actual copying time, a small fraction of the time you can save if you don’t have to trace through the objects in other generations.

It’s probably reasonable to declare that when you collect any generation, you also have to collect all the younger generations. This takes care of half the problem: you no longer have to worry about younger-to-older references. But older-to-younger references, such as in our example above, are still a problem.

There are a number of ways to deal with inter-generational references, but they all impose a cost on pointer assignment. The basic idea is that any reference from an older generation to a younger one is treated as part of the younger generation’s root set, or that an older-to-younger pointer automatically causes the object in the younger generation to be promoted. One way to accomplish this is for each generation (except the oldest) to maintain a list of objects in older generations that point into it, and to consider these to be part of the generation’s root set the next time you collect it. Another way, which will occasionally cause objects to persist longer than they need to, is to have each object’s block header contain a flag that says "promote me no matter what." (This approach saves memory, but also means you’d have to make an extra pass through the generation to make sure you catch all the "promote me" objects that weren’t also referred to from inside the generation.)

In either case, though, you have to catch the inter-generational reference when it’s first set up. The easiest way to do this is for every pointer assignment to do a quick check to see whether the new value points into a younger generation. If it does, then you add the object being changed to the younger generation’s root set. Performing this check all the time hurts, however. It can roughly double the amount of time taken by the change over the time required for a simple pointer assignment. But this extra time can be made up for by the increased overall performance of generational collection.

If the compiler knows about the collection algorithm, it can also get rid of many common cases by detecting inter-generational assignments at compile time. This won’t always work, of course, but if you know, for example, that the ephemeral generation always gets collected when a function returns (i.e., it serves as the "stack"), then you can detect cases where a new object is automatically assigned to something other than a local variable and automatically new up the new object in the same generation as the object that is going to point to it. This can cut down a lot on the amount we have to check for inter-generational references.

Even with this drawback, we can thus improve on many of our performance and memory usage problems by using a generational algorithm. A typical implementation might include an ephemeral genreration, one or more generations where objects are aged, and a final, "eldest," generation where objects end up if they persist long enough. You can cut out wasted memory by making the eldest generation use mark-sweep or mark-compact collection all the time.

Incremental garbage collection

One problem all of these approaches share is the fact that the garbage collector can’t run continuously in a separate thread while the rest of the program is doing its thing, and you can’t stop the collection in the middle (at least not without having to start over from scratch the next time). Generational collection can minimize the time wasted for any given collection pass, but you still have to stop and collect (and you still have to collect the whole heap from time to time).

There are modifications for both of the classical techniques that allow them to work in a multithreaded environment. The sweep pass of mark-sweep collection can happen while the program is doing useful work. The mark pass can also be stopped and restarted, but only at a cost. If you change any object that’s already been marked, you have to catch that change and add the object back into the list of objects that have to be traced. This again can add to the cost of pointer assignment, but here, if you have cooperative hardware, you can get the hardware to help you. You’d just mark each object that’s been marked as protected memory and then trap the protection violation that occurs when the object in question is written to.

Copying collection is harder. The same techniques used for mark-sweep collection can work, but they cost more. An alternative is to break available memory into three partitions, one of which is garbage, one of which is where new objects go, and the third of which contains stale references. This works with no extra cost on pointer assignments, but wastes lots of memory.

Low-level issues

So far, I’ve glossed over some of the low-level issues. Let’s take a few minutes to examine them now.

Root set. One is that the root set can potentially be large. In a garbage-collected language like Dylan, however, this isn’t necessarily true. Instead of having separate stack and static areas, blocks of local and global variables can be allocated on the heap just like user-created objects. The root set, then, is just a couple of registers pointing to the local and global variables (which may themselves be linked lists of variable blocks ["stack frames"]). This technique can probably be used in Java as well.

What’s a pointer? Another issue we’ve glossed over so far is the question of how the garbage collector can tell the difference between a pointer and some other kind of data. There are two ways of doing this: In some languages, such as Dylan, any type of object can be stored in any variable (variables can be "specialized," which allows compile-time type checking too, but support for untyped variables must be there). In this case, every variable is the same size, so you know where all the variables are, restricting your problem to distinguishing between pointers and immediate values. The language must use some kind of encoding to enable the garbage collector to distinguish an immediate value from a pointer simply by looking at the value itself. For example, you can take advantage of the fact that all objects (in most memory models) are allocated at long-word boundaries. If a pointer variable is a regular machine pointer, that means (on most architectures), that the two least significant bits of all pointers must be 0. So if either of those bits is 1, you know you’re dealing with an immediate value. This can slow arithmetic down a bit, and means integer variables have one bit less of precision, but these drawbacks can often be optimized out at compile time.

Java doesn’t have this problem because Java doesn’t have untyped variables, and unlike Dylan, there’s a hard-and-fast difference between an "object" and a "primitive type." Only primitive types are stored as immediate values. All objects, even small ones like Integer, are stored as pointers. So if the compiler knows the language is garbage collected, it can supply extra information that is available at run time (probably in a class definition somewhere) that the garbage collector can use to determine which variables in an object are pointers that need to be followed.

C++ poses an additional problem. Because the language doesn’t require garbage collection, a garbage collector in a C++ program is usually a bolted-on third-party library. This means the compiler (most of the time) can’t supply extra information that’s available to the garbage collector at runtime. Garbage collectors for C++, therefore, usually use a technique called conservative garbage collection. The basic idea behind conservative collection is to make as few assumptions as possible about the data you’re looking at. If there’s any doubt whatsoever about whether a particular value is a pointer or something else, assume it’s a pointer.

This, of course, imposes some restrictions on what the programmer can do. If a pointer is "disguised" in some way, the collector can’t see it. You also can’t calculate addresses of objects from the addresses of other objects or "poke" arbitrarily into memory. Generally, these aren’t good ideas in any case, but you have to be aware of them.

More importantly, there are significant restrictions on how the collector can operate. If you try to follow something that isn’t really a pointer, you may generate a machine exception, so you have to be sure to catch machine exceptions. Copying collection, or other schemes that move blocks around in memory, generally isn’t feasible because you can never know for sure that the thing you’re looking at is really a pointer that can be changed (and which points to something that can be moved). Mark-sweep collection also has to be implemented differently because you can’t rely on the presence of a block header (you don’t know for sure you’ve followed a real pointer). If you hook into malloc() and free() and replace them with your own allocation schemes, you can control some of these problems by ensuring that memory is allocated only from well-defined zones and by putting some kind of sentinel value in your block headers so that the collector can tell for sure (or at least with more certainty) what’s a pointer and what isn’t. And of course, the whole process goes slower because you follow more dead ends and have to perform more checks. Even so, there have been tests that show a good conservative collector can perform better than manual memory management for most C++ programs [find citation].

Memory used by the collection process. Another issue we’ve glossed over is the fact that the garbage collector itself requires memory to do its job. A simple collection implementation would use recursion to follow all the pointers. This basically means the collector is using the machine stack for its temporary scratch space, raising the danger of stack overflow.

Of course, on many architectures, you can use memory protection to detect stack overflow before anything is messed up, but you’d need some kind of backup plan for handling stack overflow. Another option is for the garbage collector itself to squirrel away some memory it can use to do its job. Again, though, if you run into a situation where there are more live blocks than you planned for when you allocated the scratch space, you’re in trouble. You could also have the collector use regular heap space itself, but since the collector is usually supposed to run in low-memory conditions, this doesn’t seem like a good idea.

There are a number of schemes for keeping track of which blocks need to be checked in place, or through the use of an extra scratch variable in each block’s block header. This eases the memory problem, but at a considerable cost in performance. In any event, most garbage collectors don’t use regular machine-stack space, or use it only in combination with some kind of more-robust backup scheme.


In C++, when an object is destroyed and its memory reclaimed, its destructor is called. The main use of destructors is to delete subordinate objects and reclaim their memory, and of course, if you’re using garbage collection, this isn’t necessary. But often other things are done in an object’s destructor: operating-system data structures are reclaimed, files are closed, counts or other internal statistics are adjusted, and so on. In a garbage-collected language, these types of features are enabled through the use of something called finalization. Basically, you can specify a method that is to be called before the garbage collector deletes an object. In Java, you do this by defining a method called finalize() for the object.

Finalization can’t really be implemented without some cooperation from the garbage collector. The basic idea is that every live object in the runtime that has finalization enabled is stored in a list that is maintained by the memory manager. We’ll call it the "finalization enabled" list. At the beginning of a garbage-collection pass, the objects in this list are moved to a second list, the "to be finalized" list. As each object is encountered during garbage collection, it’s moved back to the "finalization enabled" list. At the end of the pass, anything still in the "to be finalized" list is garbage. Naturally, the finalization lists are not considered part of the root set by the garbage collector, so the actual process of garbage collection is not affected.

Once you’ve finished a collection pass, you have a list of objects that need to be finalized. One by one, each object’s finalization method is called, after which the object can be treated as garbage. The only catch here is that the object’s finalization method can "resurrect" it: it could fix things so the object is once again reachable from the root set. A simple way to handle this problem is to declare resurrecting an object illegal. But this is generally impractical. A better way to avoid this is to have a "finalization method called" flag in the object’s block header. This flag is automatically cleared when the object is encountered during a collection pass, and set by the memory manager after the finalization method has been called. If, during the pass that calls the finalization method, an object’s finalization method hasn’t been called, the finalization method is called, the flag is set, and the object is returned to the "finalization enabled" list. If the flag is set, on the other hand, the object can be removed from the finalization lists and reclaimed safely.

The two finalization lists— the "finalization enabled" and "to be finalized" lists— can be conceptual rather than real. With a mark-sweep collector, you can just call the finalization methods on the sweep pass. You don’t need to keep track separately of which objects have finalization enabled at all. With copying collection, you can get by with just the "finalization enabled" list. Just go through it and call the finalization method for any object that hasn’t been copied into the new heap zone (i.e., that hasn’t been replaced by a forwarding address in its old location). The problem here is that you either have to go back and account for resurrection by copying the other objects into the new heap in a separate pass, or you have to run another garbage-collection pass right away.

Of course, one problem with finalization is that you have no guarantees of when the garbage collector is going to run or when the finalization methods are going to be called, meaning it’s possible that there could be a fairly long lapse between when you’re actually done with an object and when its finalization method is called. If the finalization method is supposed to release a monitor lock, for example, this is a problem. There are a couple of solutions to this.

Controlling garbage collection

Many programmers complain about the lack of control they have over the performance characteristics of garbage collection. What if the garbage collector decides to run right in the middle of a time-critical operation? It’s worth pointing out again that with manual memory management, you generally have little control over when exactly you’re going to pay the price to delete objects either, unless you specifically take measures to avoid this.

Typically, garbage-collected languages do provide ways of controlling the garbage collection. Almost all languages using garbage collection provide a way for you to force a collection pass, or force finalization methods to be called. In Java, you can do this with System.gc(). If you’re holding onto system resources that are released in finalization methods, this is one way of avoiding holding onto them longer than necessary. It’s also a way to allowing you to take time for a collection when it’s convenient, rather than just waiting for memory to get low or the collector to otherwise decide on its own that it’s time to do a collection.

Conversely, there’s usually a way to suspend garbage collection for a time during operations that really can’t be interrupted. You have to be careful with this feature, of course, since if you run out of memory, you’re stuck, but it’s generally there if you absolutely need it. Java, inexplicably, doesn’t currently provide you with a way to do this, although other languages do.

What garbage collection doesn’t do for you

Garbage collection isn’t a panacea. First of all, it doesn’t necessarily eliminate all memory-management bugs—it just renders them more benign. If the collector is working right, you generally don’t have to worry about memory leaks, and you don’t have to worry about deleting objects twice or accessing raw memory through old pointers. But you do have to worry about who should see changes to a given object. Sometimes you need objects or functions to all have their own copies of a particular object, and sometimes you need objects or functions to all point to the same object. Doing either wrong leads to erroneous results that can be challenging to track down, but at least the application won’t crash on you. There are also many times when you have to explicitly let go of an object by setting the variable pointing to it to NULL. This is particularly a problem with global variables. Failing to do this in a timely manner could cause you to hang onto system resources too long, or simply to use more memory than is really necessary (a more benign form of memory leak). Garbage collection doesn’t mean you don’t have to think about memory management issues; it just makes them easier to handle.

More importantly, garbage collection helps you only to manage memory, not to manage any of the other resources the system provides to you. In C++, an object can release everything it uses in its destructor, not just subordinate data structures. In a garbage-collected language, you could theoretically do the same thing with finalization, but since you typically don’t have any control over when collection occurs, you’ll likely hold onto resources longer than is strictly necessary. For some resources, this is fine, but for most, it’s a serious problem.

In C++, it’s a common idiom to wrap all system resources in objects that take ownership in a constructor and release ownership in the destructor. When a function uses one of these resources, it allocates the appropriate wrapper object on the stack, and the resource is automatically released when the wrapper object goes out of scope. This idiom doesn’t work with garbage collection.

Languages such as Dylan deal with the same problem using a different idiom. Instead of wrapping a system resource in a stack-allocated object, you use the language’s macro facility to acquire and release the resource. Instead of

    CFile theFile(filename);
    // do stuff with theFile...
// file is closed when theFile goes out of scope

in C++, in Dylan, you’d do something like this:

with-open-file(filename, the-file)
    // do stuff with the-file
end with-open-file;

The with-open-file macro would take care of wrapping the statement body (the stuff between "with-open-file" and "end") with the appropriate code to open and close the file. If that code is too complicated to compile inline, you can just have the macro call appropriate methods.

Of course, if you’re actually acquiring and releasing system resources in the contructor and destructor of a persistent object in C++, you either have to use finalization in Java or recast your algorithms to allow more timely acquisition and release of the resources in question. This may mean the macro trick can’t be used and you have to make all the appropriate calls yourself.

And unfortunately, you always have to make the calls manually in Java, because Java doesn’t have a macro language or any similar feature to make handling of non-memory resources easier. In my opinion, this is an annoying drawback of Java, but sometimes you have to give up something to get something.


Well, it’s been a long journey, but I hope I’ve succeeded in my mission here, to help explain how garbage collection really works and what kinds of tradeoffs are really involved in using a garbage-collected language such as Java versus a language such as C++ that uses manual memory management. My aim here is not to campaign for garbage collection (or for Java), but simply to inject a little more reality into the C++-Java debate. I hope this has helped.

I’d also like to point out that my aim here was not to provide a completely exhaustive treatment of the subject area, which is easily large enough to fill several books (see below). My aim was simply to provide a fairly detailed overview of the subject with some possible solutions to some of the problems. I have to confess here that I’m no expert in this stuff myself—I’m simply a curious programmer who did a little research. I had two primary sources for this article: Garbage Collection: Algorithms for Automatic Dynamic Memory Management by Richard Jones and Rafael Lins (John Wiley and sons, 1996) is the definitive textbook on automatic memory management and is a must-read for anyone implementing a garbage collector or interested in the subject. The Harlequin memory management page ( is a Web site with basic descriptions of various memory management algorithms and a very extensive bibliography of papers on different aspects of both manual and automatic memory management. It’s also got a pointer to Richard Jones’s Web site, which has an even bigger bibliography (probably the most exhaustive one on this subject). Both are also well worth checking out.

If you find this article to be well-written and technically accurate, the lion’s share of the credit should go to these two sources. If you find errors, they’re mine. And as always, I’m interested in continuing my own education on this subject, so if you find errors or areas that are incomplete, please write me.